{"repo_id":"CREMA","entity_id":"py:setup","uri":"program://CREMA/module/setup#L1-L36","kind":"module","name":"setup","path":"setup.py","language":"python","start_line":1,"end_line":36,"context_start_line":1,"context_end_line":36,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom setuptools import setup, find_namespace_packages\nimport platform\n\nDEPENDENCY_LINKS = []\nif platform.system() == \"Windows\":\n DEPENDENCY_LINKS.append(\"https://download.pytorch.org/whl/torch_stable.html\")\n\n\ndef fetch_requirements(filename):\n with open(filename) as f:\n return [ln.strip() for ln in f.read().split(\"\\n\")]\n\n\nsetup(\n name=\"salesforce-lavis\",\n version=\"1.0.0.dev1\",\n author=\"Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, Steven C.H. Hoi\",\n description=\"LAVIS - A One-stop Library for Language-Vision Intelligence\",\n long_description=open(\"README.md\", \"r\", encoding=\"utf-8\").read(),\n long_description_content_type=\"text/markdown\",\n keywords=\"Vision-Language, Multimodal, Image Captioning, Generative AI, Deep Learning, Library, PyTorch\",\n license=\"3-Clause BSD\",\n packages=find_namespace_packages(include=\"lavis.*\"),\n install_requires=fetch_requirements(\"requirements.txt\"),\n python_requires=\">=3.7.0\",\n include_package_data=True,\n dependency_links=DEPENDENCY_LINKS,\n zip_safe=False,\n)","source_hash":"c86ce9e47abd2dcdde774e01b93f9513338d6744c4d74b1c8a638036f53fae31","truncated":false} {"repo_id":"CREMA","entity_id":"py:setup.fetch_requirements","uri":"program://CREMA/function/setup.fetch_requirements#L16-L18","kind":"function","name":"fetch_requirements","path":"setup.py","language":"python","start_line":16,"end_line":18,"context_start_line":1,"context_end_line":36,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom setuptools import setup, find_namespace_packages\nimport platform\n\nDEPENDENCY_LINKS = []\nif platform.system() == \"Windows\":\n DEPENDENCY_LINKS.append(\"https://download.pytorch.org/whl/torch_stable.html\")\n\n\ndef fetch_requirements(filename):\n with open(filename) as f:\n return [ln.strip() for ln in f.read().split(\"\\n\")]\n\n\nsetup(\n name=\"salesforce-lavis\",\n version=\"1.0.0.dev1\",\n author=\"Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, Steven C.H. Hoi\",\n description=\"LAVIS - A One-stop Library for Language-Vision Intelligence\",\n long_description=open(\"README.md\", \"r\", encoding=\"utf-8\").read(),\n long_description_content_type=\"text/markdown\",\n keywords=\"Vision-Language, Multimodal, Image Captioning, Generative AI, Deep Learning, Library, PyTorch\",\n license=\"3-Clause BSD\",\n packages=find_namespace_packages(include=\"lavis.*\"),\n install_requires=fetch_requirements(\"requirements.txt\"),\n python_requires=\">=3.7.0\",\n include_package_data=True,\n dependency_links=DEPENDENCY_LINKS,\n zip_safe=False,\n)","source_hash":"c86ce9e47abd2dcdde774e01b93f9513338d6744c4d74b1c8a638036f53fae31","truncated":false} {"repo_id":"CREMA","entity_id":"py:train","uri":"program://CREMA/module/train#L1-L103","kind":"module","name":"train","path":"train.py","language":"python","start_line":1,"end_line":103,"context_start_line":1,"context_end_line":103,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport argparse\nimport os\nimport random\n\nimport numpy as np\nimport torch\nimport torch.backends.cudnn as cudnn\n\nimport lavis.tasks as tasks\nfrom lavis.common.config import Config\nfrom lavis.common.dist_utils import get_rank, init_distributed_mode\nfrom lavis.common.logger import setup_logger\nfrom lavis.common.optims import (\n LinearWarmupCosineLRScheduler,\n LinearWarmupStepLRScheduler,\n)\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import now\n\n# imports modules for registration\nfrom lavis.datasets.builders import *\nfrom lavis.models import *\nfrom lavis.processors import *\nfrom lavis.runners import *\nfrom lavis.tasks import *\n\n\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef get_runner_class(cfg):\n \"\"\"\n Get runner class from config. Default to epoch-based runner.\n \"\"\"\n runner_cls = registry.get_runner_class(cfg.run_cfg.get(\"runner\", \"runner_base\"))\n\n return runner_cls\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n\n task = tasks.setup_task(cfg)\n datasets = task.build_datasets(cfg)\n model = task.build_model(cfg)\n\n runner = get_runner_class(cfg)(\n cfg=cfg, job_id=job_id, task=task, model=model, datasets=datasets\n )\n runner.train()\n\n\nif __name__ == \"__main__\":\n main()","source_hash":"8d5098dd4bed5904b228ef010a65589538ac6ffaf690be7673b4f742b5435048","truncated":false} {"repo_id":"CREMA","entity_id":"py:train.parse_args","uri":"program://CREMA/function/train.parse_args#L35-L51","kind":"function","name":"parse_args","path":"train.py","language":"python","start_line":35,"end_line":51,"context_start_line":15,"context_end_line":71,"code":"\nimport lavis.tasks as tasks\nfrom lavis.common.config import Config\nfrom lavis.common.dist_utils import get_rank, init_distributed_mode\nfrom lavis.common.logger import setup_logger\nfrom lavis.common.optims import (\n LinearWarmupCosineLRScheduler,\n LinearWarmupStepLRScheduler,\n)\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import now\n\n# imports modules for registration\nfrom lavis.datasets.builders import *\nfrom lavis.models import *\nfrom lavis.processors import *\nfrom lavis.runners import *\nfrom lavis.tasks import *\n\n\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef get_runner_class(cfg):\n \"\"\"\n Get runner class from config. Default to epoch-based runner.\n \"\"\"\n runner_cls = registry.get_runner_class(cfg.run_cfg.get(\"runner\", \"runner_base\"))\n\n return runner_cls","source_hash":"8d5098dd4bed5904b228ef010a65589538ac6ffaf690be7673b4f742b5435048","truncated":false} {"repo_id":"CREMA","entity_id":"py:train.setup_seeds","uri":"program://CREMA/function/train.setup_seeds#L54-L62","kind":"function","name":"setup_seeds","path":"train.py","language":"python","start_line":54,"end_line":62,"context_start_line":34,"context_end_line":82,"code":"\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef get_runner_class(cfg):\n \"\"\"\n Get runner class from config. Default to epoch-based runner.\n \"\"\"\n runner_cls = registry.get_runner_class(cfg.run_cfg.get(\"runner\", \"runner_base\"))\n\n return runner_cls\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n","source_hash":"8d5098dd4bed5904b228ef010a65589538ac6ffaf690be7673b4f742b5435048","truncated":false} {"repo_id":"CREMA","entity_id":"py:train.get_runner_class","uri":"program://CREMA/function/train.get_runner_class#L65-L71","kind":"function","name":"get_runner_class","path":"train.py","language":"python","start_line":65,"end_line":71,"context_start_line":45,"context_end_line":91,"code":" )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef get_runner_class(cfg):\n \"\"\"\n Get runner class from config. Default to epoch-based runner.\n \"\"\"\n runner_cls = registry.get_runner_class(cfg.run_cfg.get(\"runner\", \"runner_base\"))\n\n return runner_cls\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n","source_hash":"8d5098dd4bed5904b228ef010a65589538ac6ffaf690be7673b4f742b5435048","truncated":false} {"repo_id":"CREMA","entity_id":"py:train.main","uri":"program://CREMA/function/train.main#L74-L99","kind":"function","name":"main","path":"train.py","language":"python","start_line":74,"end_line":99,"context_start_line":54,"context_end_line":103,"code":"def setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef get_runner_class(cfg):\n \"\"\"\n Get runner class from config. Default to epoch-based runner.\n \"\"\"\n runner_cls = registry.get_runner_class(cfg.run_cfg.get(\"runner\", \"runner_base\"))\n\n return runner_cls\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n\n task = tasks.setup_task(cfg)\n datasets = task.build_datasets(cfg)\n model = task.build_model(cfg)\n\n runner = get_runner_class(cfg)(\n cfg=cfg, job_id=job_id, task=task, model=model, datasets=datasets\n )\n runner.train()\n\n\nif __name__ == \"__main__\":\n main()","source_hash":"8d5098dd4bed5904b228ef010a65589538ac6ffaf690be7673b4f742b5435048","truncated":false} {"repo_id":"CREMA","entity_id":"py:evaluate","uri":"program://CREMA/module/evaluate#L1-L92","kind":"module","name":"evaluate","path":"evaluate.py","language":"python","start_line":1,"end_line":92,"context_start_line":1,"context_end_line":92,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport argparse\nimport random\n\nimport numpy as np\nimport torch\nimport torch.backends.cudnn as cudnn\n\nimport lavis.tasks as tasks\nfrom lavis.common.config import Config\nfrom lavis.common.dist_utils import get_rank, init_distributed_mode\nfrom lavis.common.logger import setup_logger\nfrom lavis.common.optims import (\n LinearWarmupCosineLRScheduler,\n LinearWarmupStepLRScheduler,\n)\nfrom lavis.common.utils import now\n\n# imports modules for registration\nfrom lavis.datasets.builders import *\nfrom lavis.models import *\nfrom lavis.processors import *\nfrom lavis.runners.runner_base import RunnerBase\nfrom lavis.tasks import *\n\n\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n\n task = tasks.setup_task(cfg)\n datasets = task.build_datasets(cfg)\n model = task.build_model(cfg)\n\n runner = RunnerBase(\n cfg=cfg, job_id=job_id, task=task, model=model, datasets=datasets\n )\n runner.evaluate(skip_reload=True)\n\n\nif __name__ == \"__main__\":\n main()","source_hash":"ccd3f5de75727a63d59035a6e356d939b0acf9bcbe77d7d905a70fa4b4bdb1db","truncated":false} {"repo_id":"CREMA","entity_id":"py:evaluate.parse_args","uri":"program://CREMA/function/evaluate.parse_args#L33-L49","kind":"function","name":"parse_args","path":"evaluate.py","language":"python","start_line":33,"end_line":49,"context_start_line":13,"context_end_line":69,"code":"import torch.backends.cudnn as cudnn\n\nimport lavis.tasks as tasks\nfrom lavis.common.config import Config\nfrom lavis.common.dist_utils import get_rank, init_distributed_mode\nfrom lavis.common.logger import setup_logger\nfrom lavis.common.optims import (\n LinearWarmupCosineLRScheduler,\n LinearWarmupStepLRScheduler,\n)\nfrom lavis.common.utils import now\n\n# imports modules for registration\nfrom lavis.datasets.builders import *\nfrom lavis.models import *\nfrom lavis.processors import *\nfrom lavis.runners.runner_base import RunnerBase\nfrom lavis.tasks import *\n\n\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n","source_hash":"ccd3f5de75727a63d59035a6e356d939b0acf9bcbe77d7d905a70fa4b4bdb1db","truncated":false} {"repo_id":"CREMA","entity_id":"py:evaluate.setup_seeds","uri":"program://CREMA/function/evaluate.setup_seeds#L52-L60","kind":"function","name":"setup_seeds","path":"evaluate.py","language":"python","start_line":52,"end_line":60,"context_start_line":32,"context_end_line":80,"code":"\ndef parse_args():\n parser = argparse.ArgumentParser(description=\"Training\")\n\n parser.add_argument(\"--cfg-path\", required=True, help=\"path to configuration file.\")\n parser.add_argument(\n \"--options\",\n nargs=\"+\",\n help=\"override some settings in the used config, the key-value pair \"\n \"in xxx=yyy format will be merged into config file (deprecate), \"\n \"change to --cfg-options instead.\",\n )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n","source_hash":"ccd3f5de75727a63d59035a6e356d939b0acf9bcbe77d7d905a70fa4b4bdb1db","truncated":false} {"repo_id":"CREMA","entity_id":"py:evaluate.main","uri":"program://CREMA/function/evaluate.main#L63-L88","kind":"function","name":"main","path":"evaluate.py","language":"python","start_line":63,"end_line":88,"context_start_line":43,"context_end_line":92,"code":" )\n\n args = parser.parse_args()\n # if 'LOCAL_RANK' not in os.environ:\n # os.environ['LOCAL_RANK'] = str(args.local_rank)\n\n return args\n\n\ndef setup_seeds(config):\n seed = config.run_cfg.seed + get_rank()\n\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n\n cudnn.benchmark = False\n cudnn.deterministic = True\n\n\ndef main():\n # allow auto-dl completes on main process without timeout when using NCCL backend.\n # os.environ[\"NCCL_BLOCKING_WAIT\"] = \"1\"\n\n # set before init_distributed_mode() to ensure the same job_id shared across all ranks.\n job_id = now()\n\n cfg = Config(parse_args())\n\n init_distributed_mode(cfg.run_cfg)\n\n setup_seeds(cfg)\n\n # set after init_distributed_mode() to only log on master.\n setup_logger()\n\n cfg.pretty_print()\n\n task = tasks.setup_task(cfg)\n datasets = task.build_datasets(cfg)\n model = task.build_model(cfg)\n\n runner = RunnerBase(\n cfg=cfg, job_id=job_id, task=task, model=model, datasets=datasets\n )\n runner.evaluate(skip_reload=True)\n\n\nif __name__ == \"__main__\":\n main()","source_hash":"ccd3f5de75727a63d59035a6e356d939b0acf9bcbe77d7d905a70fa4b4bdb1db","truncated":false} {"repo_id":"CREMA","entity_id":"py:docs.conf","uri":"program://CREMA/module/docs.conf#L1-L56","kind":"module","name":"docs.conf","path":"docs/conf.py","language":"python","start_line":1,"end_line":56,"context_start_line":1,"context_end_line":56,"code":"# Configuration file for the Sphinx documentation builder.\n#\n# This file only contains a selection of the most common options. For a full\n# list see the documentation:\n# https://www.sphinx-doc.org/en/master/usage/configuration.html\n\n# -- Path setup --------------------------------------------------------------\n\n# If extensions (or modules to document with autodoc) are in another directory,\n# add these directories to sys.path here. If the directory is relative to the\n# documentation root, use os.path.abspath to make it absolute, like shown here.\n#\n# import os\n# import sys\n# sys.path.insert(0, os.path.abspath('.'))\n\n\n# -- Project information -----------------------------------------------------\n\nproject = \"LAVIS\"\ncopyright = \"2022, salesforce.com inc.\"\nauthor = (\n \"Dongxu Li, Junnan Li, Hung Le, Guangsen Wang, Silvio Savarese, Steven C.H. Hoi\"\n)\n\n\n# -- General configuration ---------------------------------------------------\n\n# Add any Sphinx extension module names here, as strings. They can be\n# extensions coming with Sphinx (named 'sphinx.ext.*') or your custom\n# ones.\nextensions = [\"nbsphinx\"]\n\n# Add any paths that contain templates here, relative to this directory.\ntemplates_path = [\"_templates\"]\n\n# List of patterns, relative to source directory, that match files and\n# directories to ignore when looking for source files.\n# This pattern also affects html_static_path and html_extra_path.\nexclude_patterns = []\n\n\n# -- Options for HTML output -------------------------------------------------\n\n# The theme to use for HTML and HTML Help pages. See the documentation for\n# a list of builtin themes.\n#\n# html_theme = \"alabaster\"\nhtml_theme = \"sphinx_rtd_theme\"\n\n# Add any paths that contain custom static files (such as style sheets) here,\n# relative to this directory. They are copied after the builtin static files,\n# so a file named \"default.css\" will overwrite the builtin \"default.css\".\nhtml_static_path = [\"_static\"]\n\n# pygments_style = \"sphinx\"","source_hash":"77428b20b3cd46d5de8bd10ed20e7d6fc3c8bfa7af1c4d53f5431af6adcf7653","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit","uri":"program://CREMA/module/lavis.models.vit#L1-L527","kind":"module","name":"lavis.models.vit","path":"lavis/models/vit.py","language":"python","start_line":1,"end_line":527,"context_start_line":1,"context_end_line":527,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n \n Based on timm code base\n https://github.com/rwightman/pytorch-image-models/tree/master/timm\n\"\"\"\n\nimport math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom functools import partial\n\nfrom timm.models.vision_transformer import _cfg, PatchEmbed\nfrom timm.models.registry import register_model\nfrom timm.models.layers import trunc_normal_, DropPath\nfrom timm.models.helpers import named_apply, adapt_input_conv\n\nfrom fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper\nfrom lavis.models.base_model import BaseEncoder\n\n\nclass Mlp(nn.Module):\n \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n def __init__(\n self,\n in_features,\n hidden_features=None,\n out_features=None,\n act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n if register_hook:\n self.save_attention_map(attn)\n attn.register_hook(self.save_attn_gradients)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.0,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n use_grad_checkpointing=False,\n ):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(\n in_features=dim,\n hidden_features=mlp_hidden_dim,\n act_layer=act_layer,\n drop=drop,\n )\n\n if use_grad_checkpointing:\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, register_hook=False):\n x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformer\n A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` -\n https://arxiv.org/abs/2010.11929\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=True,\n qk_scale=None,\n representation_size=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.0,\n norm_layer=None,\n use_grad_checkpointing=False,\n ckpt_layer=0,\n ):\n \"\"\"\n Args:\n img_size (int, tuple): input image size\n patch_size (int, tuple): patch size\n in_chans (int): number of input channels\n num_classes (int): number of classes for classification head\n embed_dim (int): embedding dimension\n depth (int): depth of transformer\n num_heads (int): number of attention heads\n mlp_ratio (int): ratio of mlp hidden dim to embedding dim\n qkv_bias (bool): enable bias for qkv if True\n qk_scale (float): override default qk scale of head_dim ** -0.5 if set\n representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set\n drop_rate (float): dropout rate\n attn_drop_rate (float): attention dropout rate\n drop_path_rate (float): stochastic depth rate\n norm_layer: (nn.Module): normalization layer\n \"\"\"\n super().__init__()\n self.num_features = (\n self.embed_dim\n ) = embed_dim # num_features for consistency with other models\n norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)\n\n self.patch_embed = PatchEmbed(\n img_size=img_size,\n patch_size=patch_size,\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n dpr = [\n x.item() for x in torch.linspace(0, drop_path_rate, depth)\n ] # stochastic depth decay rule\n self.blocks = nn.ModuleList(\n [\n Block(\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= depth - ckpt_layer\n ),\n )\n for i in range(depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:\n w = w.transpose([1, 0])\n return torch.from_numpy(w)\n\n w = np.load(checkpoint_path)\n if not prefix and \"opt/target/embedding/kernel\" in w:\n prefix = \"opt/target/\"\n\n if hasattr(model.patch_embed, \"backbone\"):\n # hybrid\n backbone = model.patch_embed.backbone\n stem_only = not hasattr(backbone, \"stem\")\n stem = backbone if stem_only else backbone.stem\n stem.conv.weight.copy_(\n adapt_input_conv(\n stem.conv.weight.shape[1], _n2p(w[f\"{prefix}conv_root/kernel\"])\n )\n )\n stem.norm.weight.copy_(_n2p(w[f\"{prefix}gn_root/scale\"]))\n stem.norm.bias.copy_(_n2p(w[f\"{prefix}gn_root/bias\"]))\n if not stem_only:\n for i, stage in enumerate(backbone.stages):\n for j, block in enumerate(stage.blocks):\n bp = f\"{prefix}block{i + 1}/unit{j + 1}/\"\n for r in range(3):\n getattr(block, f\"conv{r + 1}\").weight.copy_(\n _n2p(w[f\"{bp}conv{r + 1}/kernel\"])\n )\n getattr(block, f\"norm{r + 1}\").weight.copy_(\n _n2p(w[f\"{bp}gn{r + 1}/scale\"])\n )\n getattr(block, f\"norm{r + 1}\").bias.copy_(\n _n2p(w[f\"{bp}gn{r + 1}/bias\"])\n )\n if block.downsample is not None:\n block.downsample.conv.weight.copy_(\n _n2p(w[f\"{bp}conv_proj/kernel\"])\n )\n block.downsample.norm.weight.copy_(\n _n2p(w[f\"{bp}gn_proj/scale\"])\n )\n block.downsample.norm.bias.copy_(_n2p(w[f\"{bp}gn_proj/bias\"]))\n embed_conv_w = _n2p(w[f\"{prefix}embedding/kernel\"])\n else:\n embed_conv_w = adapt_input_conv(\n model.patch_embed.proj.weight.shape[1], _n2p(w[f\"{prefix}embedding/kernel\"])\n )\n model.patch_embed.proj.weight.copy_(embed_conv_w)\n model.patch_embed.proj.bias.copy_(_n2p(w[f\"{prefix}embedding/bias\"]))\n model.cls_token.copy_(_n2p(w[f\"{prefix}cls\"], t=False))\n pos_embed_w = _n2p(w[f\"{prefix}Transformer/posembed_input/pos_embedding\"], t=False)\n if pos_embed_w.shape != model.pos_embed.shape:\n pos_embed_w = resize_pos_embed( # resize pos embedding when different size from pretrained weights\n pos_embed_w,\n model.pos_embed,\n getattr(model, \"num_tokens\", 1),\n model.patch_embed.grid_size,\n )\n model.pos_embed.copy_(pos_embed_w)\n model.norm.weight.copy_(_n2p(w[f\"{prefix}Transformer/encoder_norm/scale\"]))\n model.norm.bias.copy_(_n2p(w[f\"{prefix}Transformer/encoder_norm/bias\"]))\n # if isinstance(model.head, nn.Linear) and model.head.bias.shape[0] == w[f'{prefix}head/bias'].shape[-1]:\n # model.head.weight.copy_(_n2p(w[f'{prefix}head/kernel']))\n # model.head.bias.copy_(_n2p(w[f'{prefix}head/bias']))\n # if isinstance(getattr(model.pre_logits, 'fc', None), nn.Linear) and f'{prefix}pre_logits/bias' in w:\n # model.pre_logits.fc.weight.copy_(_n2p(w[f'{prefix}pre_logits/kernel']))\n # model.pre_logits.fc.bias.copy_(_n2p(w[f'{prefix}pre_logits/bias']))\n for i, block in enumerate(model.blocks.children()):\n block_prefix = f\"{prefix}Transformer/encoderblock_{i}/\"\n mha_prefix = block_prefix + \"MultiHeadDotProductAttention_1/\"\n block.norm1.weight.copy_(_n2p(w[f\"{block_prefix}LayerNorm_0/scale\"]))\n block.norm1.bias.copy_(_n2p(w[f\"{block_prefix}LayerNorm_0/bias\"]))\n block.attn.qkv.weight.copy_(\n torch.cat(\n [\n _n2p(w[f\"{mha_prefix}{n}/kernel\"], t=False).flatten(1).T\n for n in (\"query\", \"key\", \"value\")\n ]\n )\n )\n block.attn.qkv.bias.copy_(\n torch.cat(\n [\n _n2p(w[f\"{mha_prefix}{n}/bias\"], t=False).reshape(-1)\n for n in (\"query\", \"key\", \"value\")\n ]\n )\n )\n block.attn.proj.weight.copy_(_n2p(w[f\"{mha_prefix}out/kernel\"]).flatten(1))\n block.attn.proj.bias.copy_(_n2p(w[f\"{mha_prefix}out/bias\"]))\n for r in range(2):\n getattr(block.mlp, f\"fc{r + 1}\").weight.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/kernel\"])\n )\n getattr(block.mlp, f\"fc{r + 1}\").bias.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/bias\"])\n )\n block.norm2.weight.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/scale\"]))\n block.norm2.bias.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/bias\"]))\n\n\ndef resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()):\n # Rescale the grid of position embeddings when loading from state_dict. Adapted from\n # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224\n print(\"Resized position embedding: %s to %s\", posemb.shape, posemb_new.shape)\n ntok_new = posemb_new.shape[1]\n if num_tokens:\n posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]\n ntok_new -= num_tokens\n else:\n posemb_tok, posemb_grid = posemb[:, :0], posemb[0]\n gs_old = int(math.sqrt(len(posemb_grid)))\n if not len(gs_new): # backwards compatibility\n gs_new = [int(math.sqrt(ntok_new))] * 2\n assert len(gs_new) >= 2\n print(\"Position embedding grid-size from %s to %s\", [gs_old, gs_old], gs_new)\n posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)\n posemb_grid = F.interpolate(\n posemb_grid, size=gs_new, mode=\"bicubic\", align_corners=False\n )\n posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1)\n posemb = torch.cat([posemb_tok, posemb_grid], dim=1)\n return\n\n\ndef interpolate_pos_embed(pos_embed_checkpoint, visual_encoder):\n # interpolate position embedding\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = visual_encoder.patch_embed.num_patches\n num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches**0.5)\n\n if orig_size != new_size:\n # class_token and dist_token are kept unchanged\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(\n -1, orig_size, orig_size, embedding_size\n ).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode=\"bicubic\", align_corners=False\n )\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n print(\n \"reshape position embedding from %d to %d\" % (orig_size**2, new_size**2)\n )\n\n return new_pos_embed\n else:\n return pos_embed_checkpoint\n\n\nclass VisionTransformerEncoder(VisionTransformer, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n vit_type = cfg.get(\"vit_type\", \"base\")\n image_size = cfg.get(\"image_size\", 384)\n ckpt_layer = cfg.get(\"vit_ckpt_layer\", 0)\n drop_path_rate = cfg.get(\"vit_drop_path_rate\", 0)\n norm_layer_eps = cfg.get(\"vit_layer_norm_epsilon\", -1)\n use_grad_checkpointing = cfg.get(\"vit_grad_ckpt\", False)\n\n if norm_layer_eps == -1:\n norm_layer = None\n else:\n norm_layer = partial(nn.LayerNorm, eps=norm_layer_eps)\n\n # norm_layer=partial(nn.LayerNorm, eps=1e-6),\n assert vit_type in [\"base\", \"large\"], \"vit parameter must be base or large\"\n if vit_type == \"base\":\n vision_width = 768\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=12,\n num_heads=12,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0 or drop_path_rate,\n norm_layer=norm_layer,\n )\n\n if from_pretrained:\n checkpoint = torch.hub.load_state_dict_from_url(\n url=\"https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth\",\n map_location=\"cpu\",\n check_hash=True,\n )\n state_dict = checkpoint[\"model\"]\n state_dict[\"pos_embed\"] = interpolate_pos_embed(\n state_dict[\"pos_embed\"], visual_encoder\n )\n msg = visual_encoder.load_state_dict(state_dict, strict=False)\n\n elif vit_type == \"large\":\n vision_width = 1024\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=24,\n num_heads=16,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0.1 or drop_path_rate,\n norm_layer=norm_layer,\n )\n if from_pretrained:\n from timm.models.helpers import load_custom_pretrained\n from timm.models.vision_transformer import default_cfgs\n\n load_custom_pretrained(\n visual_encoder, default_cfgs[\"vit_large_patch16_224_in21k\"]\n )\n\n visual_encoder.vision_width = vision_width\n return visual_encoder\n\n def forward_features(self, x, register_blk=-1):\n return super().forward(x, register_blk)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.Mlp","uri":"program://CREMA/class/lavis.models.vit.Mlp#L26-L51","kind":"class","name":"Mlp","path":"lavis/models/vit.py","language":"python","start_line":26,"end_line":51,"context_start_line":6,"context_end_line":71,"code":" \n Based on timm code base\n https://github.com/rwightman/pytorch-image-models/tree/master/timm\n\"\"\"\n\nimport math\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom functools import partial\n\nfrom timm.models.vision_transformer import _cfg, PatchEmbed\nfrom timm.models.registry import register_model\nfrom timm.models.layers import trunc_normal_, DropPath\nfrom timm.models.helpers import named_apply, adapt_input_conv\n\nfrom fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper\nfrom lavis.models.base_model import BaseEncoder\n\n\nclass Mlp(nn.Module):\n \"\"\"MLP as used in Vision Transformer, MLP-Mixer and related networks\"\"\"\n\n def __init__(\n self,\n in_features,\n hidden_features=None,\n out_features=None,\n act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.Attention","uri":"program://CREMA/class/lavis.models.vit.Attention#L54-L112","kind":"class","name":"Attention","path":"lavis/models/vit.py","language":"python","start_line":54,"end_line":112,"context_start_line":34,"context_end_line":132,"code":" act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n if register_hook:\n self.save_attention_map(attn)\n attn.register_hook(self.save_attn_gradients)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.0,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n use_grad_checkpointing=False,\n ):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.Block","uri":"program://CREMA/class/lavis.models.vit.Block#L115-L158","kind":"class","name":"Block","path":"lavis/models/vit.py","language":"python","start_line":115,"end_line":158,"context_start_line":95,"context_end_line":178,"code":" q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n if register_hook:\n self.save_attention_map(attn)\n attn.register_hook(self.save_attn_gradients)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.0,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n use_grad_checkpointing=False,\n ):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(\n in_features=dim,\n hidden_features=mlp_hidden_dim,\n act_layer=act_layer,\n drop=drop,\n )\n\n if use_grad_checkpointing:\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, register_hook=False):\n x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformer\n A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` -\n https://arxiv.org/abs/2010.11929\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=True,\n qk_scale=None,","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.VisionTransformer","uri":"program://CREMA/class/lavis.models.vit.VisionTransformer#L161-L285","kind":"class","name":"VisionTransformer","path":"lavis/models/vit.py","language":"python","start_line":161,"end_line":285,"context_start_line":141,"context_end_line":305,"code":" self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(\n in_features=dim,\n hidden_features=mlp_hidden_dim,\n act_layer=act_layer,\n drop=drop,\n )\n\n if use_grad_checkpointing:\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, register_hook=False):\n x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformer\n A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` -\n https://arxiv.org/abs/2010.11929\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=True,\n qk_scale=None,\n representation_size=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.0,\n norm_layer=None,\n use_grad_checkpointing=False,\n ckpt_layer=0,\n ):\n \"\"\"\n Args:\n img_size (int, tuple): input image size\n patch_size (int, tuple): patch size\n in_chans (int): number of input channels\n num_classes (int): number of classes for classification head\n embed_dim (int): embedding dimension\n depth (int): depth of transformer\n num_heads (int): number of attention heads\n mlp_ratio (int): ratio of mlp hidden dim to embedding dim\n qkv_bias (bool): enable bias for qkv if True\n qk_scale (float): override default qk scale of head_dim ** -0.5 if set\n representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set\n drop_rate (float): dropout rate\n attn_drop_rate (float): attention dropout rate\n drop_path_rate (float): stochastic depth rate\n norm_layer: (nn.Module): normalization layer\n \"\"\"\n super().__init__()\n self.num_features = (\n self.embed_dim\n ) = embed_dim # num_features for consistency with other models\n norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)\n\n self.patch_embed = PatchEmbed(\n img_size=img_size,\n patch_size=patch_size,\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n dpr = [\n x.item() for x in torch.linspace(0, drop_path_rate, depth)\n ] # stochastic depth decay rule\n self.blocks = nn.ModuleList(\n [\n Block(\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= depth - ckpt_layer\n ),\n )\n for i in range(depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:\n w = w.transpose([1, 0])\n return torch.from_numpy(w)\n\n w = np.load(checkpoint_path)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit._load_weights","uri":"program://CREMA/function/lavis.models.vit._load_weights#L289-L399","kind":"function","name":"_load_weights","path":"lavis/models/vit.py","language":"python","start_line":289,"end_line":399,"context_start_line":269,"context_end_line":419,"code":" cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:\n w = w.transpose([1, 0])\n return torch.from_numpy(w)\n\n w = np.load(checkpoint_path)\n if not prefix and \"opt/target/embedding/kernel\" in w:\n prefix = \"opt/target/\"\n\n if hasattr(model.patch_embed, \"backbone\"):\n # hybrid\n backbone = model.patch_embed.backbone\n stem_only = not hasattr(backbone, \"stem\")\n stem = backbone if stem_only else backbone.stem\n stem.conv.weight.copy_(\n adapt_input_conv(\n stem.conv.weight.shape[1], _n2p(w[f\"{prefix}conv_root/kernel\"])\n )\n )\n stem.norm.weight.copy_(_n2p(w[f\"{prefix}gn_root/scale\"]))\n stem.norm.bias.copy_(_n2p(w[f\"{prefix}gn_root/bias\"]))\n if not stem_only:\n for i, stage in enumerate(backbone.stages):\n for j, block in enumerate(stage.blocks):\n bp = f\"{prefix}block{i + 1}/unit{j + 1}/\"\n for r in range(3):\n getattr(block, f\"conv{r + 1}\").weight.copy_(\n _n2p(w[f\"{bp}conv{r + 1}/kernel\"])\n )\n getattr(block, f\"norm{r + 1}\").weight.copy_(\n _n2p(w[f\"{bp}gn{r + 1}/scale\"])\n )\n getattr(block, f\"norm{r + 1}\").bias.copy_(\n _n2p(w[f\"{bp}gn{r + 1}/bias\"])\n )\n if block.downsample is not None:\n block.downsample.conv.weight.copy_(\n _n2p(w[f\"{bp}conv_proj/kernel\"])\n )\n block.downsample.norm.weight.copy_(\n _n2p(w[f\"{bp}gn_proj/scale\"])\n )\n block.downsample.norm.bias.copy_(_n2p(w[f\"{bp}gn_proj/bias\"]))\n embed_conv_w = _n2p(w[f\"{prefix}embedding/kernel\"])\n else:\n embed_conv_w = adapt_input_conv(\n model.patch_embed.proj.weight.shape[1], _n2p(w[f\"{prefix}embedding/kernel\"])\n )\n model.patch_embed.proj.weight.copy_(embed_conv_w)\n model.patch_embed.proj.bias.copy_(_n2p(w[f\"{prefix}embedding/bias\"]))\n model.cls_token.copy_(_n2p(w[f\"{prefix}cls\"], t=False))\n pos_embed_w = _n2p(w[f\"{prefix}Transformer/posembed_input/pos_embedding\"], t=False)\n if pos_embed_w.shape != model.pos_embed.shape:\n pos_embed_w = resize_pos_embed( # resize pos embedding when different size from pretrained weights\n pos_embed_w,\n model.pos_embed,\n getattr(model, \"num_tokens\", 1),\n model.patch_embed.grid_size,\n )\n model.pos_embed.copy_(pos_embed_w)\n model.norm.weight.copy_(_n2p(w[f\"{prefix}Transformer/encoder_norm/scale\"]))\n model.norm.bias.copy_(_n2p(w[f\"{prefix}Transformer/encoder_norm/bias\"]))\n # if isinstance(model.head, nn.Linear) and model.head.bias.shape[0] == w[f'{prefix}head/bias'].shape[-1]:\n # model.head.weight.copy_(_n2p(w[f'{prefix}head/kernel']))\n # model.head.bias.copy_(_n2p(w[f'{prefix}head/bias']))\n # if isinstance(getattr(model.pre_logits, 'fc', None), nn.Linear) and f'{prefix}pre_logits/bias' in w:\n # model.pre_logits.fc.weight.copy_(_n2p(w[f'{prefix}pre_logits/kernel']))\n # model.pre_logits.fc.bias.copy_(_n2p(w[f'{prefix}pre_logits/bias']))\n for i, block in enumerate(model.blocks.children()):\n block_prefix = f\"{prefix}Transformer/encoderblock_{i}/\"\n mha_prefix = block_prefix + \"MultiHeadDotProductAttention_1/\"\n block.norm1.weight.copy_(_n2p(w[f\"{block_prefix}LayerNorm_0/scale\"]))\n block.norm1.bias.copy_(_n2p(w[f\"{block_prefix}LayerNorm_0/bias\"]))\n block.attn.qkv.weight.copy_(\n torch.cat(\n [\n _n2p(w[f\"{mha_prefix}{n}/kernel\"], t=False).flatten(1).T\n for n in (\"query\", \"key\", \"value\")\n ]\n )\n )\n block.attn.qkv.bias.copy_(\n torch.cat(\n [\n _n2p(w[f\"{mha_prefix}{n}/bias\"], t=False).reshape(-1)\n for n in (\"query\", \"key\", \"value\")\n ]\n )\n )\n block.attn.proj.weight.copy_(_n2p(w[f\"{mha_prefix}out/kernel\"]).flatten(1))\n block.attn.proj.bias.copy_(_n2p(w[f\"{mha_prefix}out/bias\"]))\n for r in range(2):\n getattr(block.mlp, f\"fc{r + 1}\").weight.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/kernel\"])\n )\n getattr(block.mlp, f\"fc{r + 1}\").bias.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/bias\"])\n )\n block.norm2.weight.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/scale\"]))\n block.norm2.bias.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/bias\"]))\n\n\ndef resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()):\n # Rescale the grid of position embeddings when loading from state_dict. Adapted from\n # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224\n print(\"Resized position embedding: %s to %s\", posemb.shape, posemb_new.shape)\n ntok_new = posemb_new.shape[1]\n if num_tokens:\n posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]\n ntok_new -= num_tokens\n else:\n posemb_tok, posemb_grid = posemb[:, :0], posemb[0]\n gs_old = int(math.sqrt(len(posemb_grid)))\n if not len(gs_new): # backwards compatibility\n gs_new = [int(math.sqrt(ntok_new))] * 2\n assert len(gs_new) >= 2\n print(\"Position embedding grid-size from %s to %s\", [gs_old, gs_old], gs_new)\n posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)\n posemb_grid = F.interpolate(\n posemb_grid, size=gs_new, mode=\"bicubic\", align_corners=False","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.resize_pos_embed","uri":"program://CREMA/function/lavis.models.vit.resize_pos_embed#L402-L423","kind":"function","name":"resize_pos_embed","path":"lavis/models/vit.py","language":"python","start_line":402,"end_line":423,"context_start_line":382,"context_end_line":443,"code":" torch.cat(\n [\n _n2p(w[f\"{mha_prefix}{n}/bias\"], t=False).reshape(-1)\n for n in (\"query\", \"key\", \"value\")\n ]\n )\n )\n block.attn.proj.weight.copy_(_n2p(w[f\"{mha_prefix}out/kernel\"]).flatten(1))\n block.attn.proj.bias.copy_(_n2p(w[f\"{mha_prefix}out/bias\"]))\n for r in range(2):\n getattr(block.mlp, f\"fc{r + 1}\").weight.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/kernel\"])\n )\n getattr(block.mlp, f\"fc{r + 1}\").bias.copy_(\n _n2p(w[f\"{block_prefix}MlpBlock_3/Dense_{r}/bias\"])\n )\n block.norm2.weight.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/scale\"]))\n block.norm2.bias.copy_(_n2p(w[f\"{block_prefix}LayerNorm_2/bias\"]))\n\n\ndef resize_pos_embed(posemb, posemb_new, num_tokens=1, gs_new=()):\n # Rescale the grid of position embeddings when loading from state_dict. Adapted from\n # https://github.com/google-research/vision_transformer/blob/00883dd691c63a6830751563748663526e811cee/vit_jax/checkpoint.py#L224\n print(\"Resized position embedding: %s to %s\", posemb.shape, posemb_new.shape)\n ntok_new = posemb_new.shape[1]\n if num_tokens:\n posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]\n ntok_new -= num_tokens\n else:\n posemb_tok, posemb_grid = posemb[:, :0], posemb[0]\n gs_old = int(math.sqrt(len(posemb_grid)))\n if not len(gs_new): # backwards compatibility\n gs_new = [int(math.sqrt(ntok_new))] * 2\n assert len(gs_new) >= 2\n print(\"Position embedding grid-size from %s to %s\", [gs_old, gs_old], gs_new)\n posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)\n posemb_grid = F.interpolate(\n posemb_grid, size=gs_new, mode=\"bicubic\", align_corners=False\n )\n posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1)\n posemb = torch.cat([posemb_tok, posemb_grid], dim=1)\n return\n\n\ndef interpolate_pos_embed(pos_embed_checkpoint, visual_encoder):\n # interpolate position embedding\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = visual_encoder.patch_embed.num_patches\n num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches**0.5)\n\n if orig_size != new_size:\n # class_token and dist_token are kept unchanged\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(\n -1, orig_size, orig_size, embedding_size\n ).permute(0, 3, 1, 2)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.interpolate_pos_embed","uri":"program://CREMA/function/lavis.models.vit.interpolate_pos_embed#L426-L455","kind":"function","name":"interpolate_pos_embed","path":"lavis/models/vit.py","language":"python","start_line":426,"end_line":455,"context_start_line":406,"context_end_line":475,"code":" ntok_new = posemb_new.shape[1]\n if num_tokens:\n posemb_tok, posemb_grid = posemb[:, :num_tokens], posemb[0, num_tokens:]\n ntok_new -= num_tokens\n else:\n posemb_tok, posemb_grid = posemb[:, :0], posemb[0]\n gs_old = int(math.sqrt(len(posemb_grid)))\n if not len(gs_new): # backwards compatibility\n gs_new = [int(math.sqrt(ntok_new))] * 2\n assert len(gs_new) >= 2\n print(\"Position embedding grid-size from %s to %s\", [gs_old, gs_old], gs_new)\n posemb_grid = posemb_grid.reshape(1, gs_old, gs_old, -1).permute(0, 3, 1, 2)\n posemb_grid = F.interpolate(\n posemb_grid, size=gs_new, mode=\"bicubic\", align_corners=False\n )\n posemb_grid = posemb_grid.permute(0, 2, 3, 1).reshape(1, gs_new[0] * gs_new[1], -1)\n posemb = torch.cat([posemb_tok, posemb_grid], dim=1)\n return\n\n\ndef interpolate_pos_embed(pos_embed_checkpoint, visual_encoder):\n # interpolate position embedding\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = visual_encoder.patch_embed.num_patches\n num_extra_tokens = visual_encoder.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches**0.5)\n\n if orig_size != new_size:\n # class_token and dist_token are kept unchanged\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(\n -1, orig_size, orig_size, embedding_size\n ).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode=\"bicubic\", align_corners=False\n )\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n print(\n \"reshape position embedding from %d to %d\" % (orig_size**2, new_size**2)\n )\n\n return new_pos_embed\n else:\n return pos_embed_checkpoint\n\n\nclass VisionTransformerEncoder(VisionTransformer, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n vit_type = cfg.get(\"vit_type\", \"base\")\n image_size = cfg.get(\"image_size\", 384)\n ckpt_layer = cfg.get(\"vit_ckpt_layer\", 0)\n drop_path_rate = cfg.get(\"vit_drop_path_rate\", 0)\n norm_layer_eps = cfg.get(\"vit_layer_norm_epsilon\", -1)\n use_grad_checkpointing = cfg.get(\"vit_grad_ckpt\", False)\n\n if norm_layer_eps == -1:\n norm_layer = None\n else:\n norm_layer = partial(nn.LayerNorm, eps=norm_layer_eps)\n\n # norm_layer=partial(nn.LayerNorm, eps=1e-6),\n assert vit_type in [\"base\", \"large\"], \"vit parameter must be base or large\"","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.VisionTransformerEncoder","uri":"program://CREMA/class/lavis.models.vit.VisionTransformerEncoder#L458-L527","kind":"class","name":"VisionTransformerEncoder","path":"lavis/models/vit.py","language":"python","start_line":458,"end_line":527,"context_start_line":438,"context_end_line":527,"code":" extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(\n -1, orig_size, orig_size, embedding_size\n ).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode=\"bicubic\", align_corners=False\n )\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n print(\n \"reshape position embedding from %d to %d\" % (orig_size**2, new_size**2)\n )\n\n return new_pos_embed\n else:\n return pos_embed_checkpoint\n\n\nclass VisionTransformerEncoder(VisionTransformer, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n vit_type = cfg.get(\"vit_type\", \"base\")\n image_size = cfg.get(\"image_size\", 384)\n ckpt_layer = cfg.get(\"vit_ckpt_layer\", 0)\n drop_path_rate = cfg.get(\"vit_drop_path_rate\", 0)\n norm_layer_eps = cfg.get(\"vit_layer_norm_epsilon\", -1)\n use_grad_checkpointing = cfg.get(\"vit_grad_ckpt\", False)\n\n if norm_layer_eps == -1:\n norm_layer = None\n else:\n norm_layer = partial(nn.LayerNorm, eps=norm_layer_eps)\n\n # norm_layer=partial(nn.LayerNorm, eps=1e-6),\n assert vit_type in [\"base\", \"large\"], \"vit parameter must be base or large\"\n if vit_type == \"base\":\n vision_width = 768\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=12,\n num_heads=12,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0 or drop_path_rate,\n norm_layer=norm_layer,\n )\n\n if from_pretrained:\n checkpoint = torch.hub.load_state_dict_from_url(\n url=\"https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth\",\n map_location=\"cpu\",\n check_hash=True,\n )\n state_dict = checkpoint[\"model\"]\n state_dict[\"pos_embed\"] = interpolate_pos_embed(\n state_dict[\"pos_embed\"], visual_encoder\n )\n msg = visual_encoder.load_state_dict(state_dict, strict=False)\n\n elif vit_type == \"large\":\n vision_width = 1024\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=24,\n num_heads=16,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0.1 or drop_path_rate,\n norm_layer=norm_layer,\n )\n if from_pretrained:\n from timm.models.helpers import load_custom_pretrained\n from timm.models.vision_transformer import default_cfgs\n\n load_custom_pretrained(\n visual_encoder, default_cfgs[\"vit_large_patch16_224_in21k\"]\n )\n\n visual_encoder.vision_width = vision_width\n return visual_encoder\n\n def forward_features(self, x, register_blk=-1):\n return super().forward(x, register_blk)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.__init__","uri":"program://CREMA/function/lavis.models.vit.__init__#L167-L250","kind":"function","name":"__init__","path":"lavis/models/vit.py","language":"python","start_line":167,"end_line":250,"context_start_line":147,"context_end_line":270,"code":" act_layer=act_layer,\n drop=drop,\n )\n\n if use_grad_checkpointing:\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, register_hook=False):\n x = x + self.drop_path(self.attn(self.norm1(x), register_hook=register_hook))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformer\n A PyTorch impl of : `An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale` -\n https://arxiv.org/abs/2010.11929\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=True,\n qk_scale=None,\n representation_size=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.0,\n norm_layer=None,\n use_grad_checkpointing=False,\n ckpt_layer=0,\n ):\n \"\"\"\n Args:\n img_size (int, tuple): input image size\n patch_size (int, tuple): patch size\n in_chans (int): number of input channels\n num_classes (int): number of classes for classification head\n embed_dim (int): embedding dimension\n depth (int): depth of transformer\n num_heads (int): number of attention heads\n mlp_ratio (int): ratio of mlp hidden dim to embedding dim\n qkv_bias (bool): enable bias for qkv if True\n qk_scale (float): override default qk scale of head_dim ** -0.5 if set\n representation_size (Optional[int]): enable and set representation layer (pre-logits) to this value if set\n drop_rate (float): dropout rate\n attn_drop_rate (float): attention dropout rate\n drop_path_rate (float): stochastic depth rate\n norm_layer: (nn.Module): normalization layer\n \"\"\"\n super().__init__()\n self.num_features = (\n self.embed_dim\n ) = embed_dim # num_features for consistency with other models\n norm_layer = norm_layer or partial(nn.LayerNorm, eps=1e-6)\n\n self.patch_embed = PatchEmbed(\n img_size=img_size,\n patch_size=patch_size,\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n dpr = [\n x.item() for x in torch.linspace(0, drop_path_rate, depth)\n ] # stochastic depth decay rule\n self.blocks = nn.ModuleList(\n [\n Block(\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= depth - ckpt_layer\n ),\n )\n for i in range(depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.forward","uri":"program://CREMA/function/lavis.models.vit.forward#L265-L281","kind":"function","name":"forward","path":"lavis/models/vit.py","language":"python","start_line":265,"end_line":281,"context_start_line":245,"context_end_line":301,"code":" )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.save_attn_gradients","uri":"program://CREMA/function/lavis.models.vit.save_attn_gradients#L76-L77","kind":"function","name":"save_attn_gradients","path":"lavis/models/vit.py","language":"python","start_line":76,"end_line":77,"context_start_line":56,"context_end_line":97,"code":" self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.get_attn_gradients","uri":"program://CREMA/function/lavis.models.vit.get_attn_gradients#L79-L80","kind":"function","name":"get_attn_gradients","path":"lavis/models/vit.py","language":"python","start_line":79,"end_line":80,"context_start_line":59,"context_end_line":100,"code":" qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.save_attention_map","uri":"program://CREMA/function/lavis.models.vit.save_attention_map#L82-L83","kind":"function","name":"save_attention_map","path":"lavis/models/vit.py","language":"python","start_line":82,"end_line":83,"context_start_line":62,"context_end_line":103,"code":" proj_drop=0.0,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.get_attention_map","uri":"program://CREMA/function/lavis.models.vit.get_attention_map#L85-L86","kind":"function","name":"get_attention_map","path":"lavis/models/vit.py","language":"python","start_line":85,"end_line":86,"context_start_line":65,"context_end_line":106,"code":" self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim**-0.5\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_gradients = None\n self.attention_map = None\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def forward(self, x, register_hook=False):\n B, N, C = x.shape\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = (\n qkv[0],\n qkv[1],\n qkv[2],\n ) # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n if register_hook:\n self.save_attention_map(attn)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit._init_weights","uri":"program://CREMA/function/lavis.models.vit._init_weights#L252-L259","kind":"function","name":"_init_weights","path":"lavis/models/vit.py","language":"python","start_line":252,"end_line":259,"context_start_line":232,"context_end_line":279,"code":" mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= depth - ckpt_layer\n ),\n )\n for i in range(depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.no_weight_decay","uri":"program://CREMA/function/lavis.models.vit.no_weight_decay#L262-L263","kind":"function","name":"no_weight_decay","path":"lavis/models/vit.py","language":"python","start_line":262,"end_line":263,"context_start_line":242,"context_end_line":283,"code":" )\n for i in range(depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\"}\n\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.load_pretrained","uri":"program://CREMA/function/lavis.models.vit.load_pretrained#L284-L285","kind":"function","name":"load_pretrained","path":"lavis/models/vit.py","language":"python","start_line":284,"end_line":285,"context_start_line":264,"context_end_line":305,"code":"\n def forward(self, x, register_blk=-1):\n B = x.shape[0]\n x = self.patch_embed(x)\n\n cls_tokens = self.cls_token.expand(\n B, -1, -1\n ) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:\n w = w.transpose([1, 0])\n return torch.from_numpy(w)\n\n w = np.load(checkpoint_path)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit._n2p","uri":"program://CREMA/function/lavis.models.vit._n2p#L293-L303","kind":"function","name":"_n2p","path":"lavis/models/vit.py","language":"python","start_line":293,"end_line":303,"context_start_line":273,"context_end_line":323,"code":"\n x = x + self.pos_embed[:, : x.size(1), :]\n x = self.pos_drop(x)\n\n for i, blk in enumerate(self.blocks):\n x = blk(x, register_blk == i)\n x = self.norm(x)\n\n return x\n\n @torch.jit.ignore()\n def load_pretrained(self, checkpoint_path, prefix=\"\"):\n _load_weights(self, checkpoint_path, prefix)\n\n\n@torch.no_grad()\ndef _load_weights(model: VisionTransformer, checkpoint_path: str, prefix: str = \"\"):\n \"\"\"Load weights from .npz checkpoints for official Google Brain Flax implementation\"\"\"\n import numpy as np\n\n def _n2p(w, t=True):\n if w.ndim == 4 and w.shape[0] == w.shape[1] == w.shape[2] == 1:\n w = w.flatten()\n if t:\n if w.ndim == 4:\n w = w.transpose([3, 2, 0, 1])\n elif w.ndim == 3:\n w = w.transpose([2, 0, 1])\n elif w.ndim == 2:\n w = w.transpose([1, 0])\n return torch.from_numpy(w)\n\n w = np.load(checkpoint_path)\n if not prefix and \"opt/target/embedding/kernel\" in w:\n prefix = \"opt/target/\"\n\n if hasattr(model.patch_embed, \"backbone\"):\n # hybrid\n backbone = model.patch_embed.backbone\n stem_only = not hasattr(backbone, \"stem\")\n stem = backbone if stem_only else backbone.stem\n stem.conv.weight.copy_(\n adapt_input_conv(\n stem.conv.weight.shape[1], _n2p(w[f\"{prefix}conv_root/kernel\"])\n )\n )\n stem.norm.weight.copy_(_n2p(w[f\"{prefix}gn_root/scale\"]))\n stem.norm.bias.copy_(_n2p(w[f\"{prefix}gn_root/bias\"]))\n if not stem_only:\n for i, stage in enumerate(backbone.stages):\n for j, block in enumerate(stage.blocks):","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.from_config","uri":"program://CREMA/function/lavis.models.vit.from_config#L460-L524","kind":"function","name":"from_config","path":"lavis/models/vit.py","language":"python","start_line":460,"end_line":524,"context_start_line":440,"context_end_line":527,"code":" pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(\n -1, orig_size, orig_size, embedding_size\n ).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode=\"bicubic\", align_corners=False\n )\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n print(\n \"reshape position embedding from %d to %d\" % (orig_size**2, new_size**2)\n )\n\n return new_pos_embed\n else:\n return pos_embed_checkpoint\n\n\nclass VisionTransformerEncoder(VisionTransformer, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n vit_type = cfg.get(\"vit_type\", \"base\")\n image_size = cfg.get(\"image_size\", 384)\n ckpt_layer = cfg.get(\"vit_ckpt_layer\", 0)\n drop_path_rate = cfg.get(\"vit_drop_path_rate\", 0)\n norm_layer_eps = cfg.get(\"vit_layer_norm_epsilon\", -1)\n use_grad_checkpointing = cfg.get(\"vit_grad_ckpt\", False)\n\n if norm_layer_eps == -1:\n norm_layer = None\n else:\n norm_layer = partial(nn.LayerNorm, eps=norm_layer_eps)\n\n # norm_layer=partial(nn.LayerNorm, eps=1e-6),\n assert vit_type in [\"base\", \"large\"], \"vit parameter must be base or large\"\n if vit_type == \"base\":\n vision_width = 768\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=12,\n num_heads=12,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0 or drop_path_rate,\n norm_layer=norm_layer,\n )\n\n if from_pretrained:\n checkpoint = torch.hub.load_state_dict_from_url(\n url=\"https://dl.fbaipublicfiles.com/deit/deit_base_patch16_224-b5f2ef4d.pth\",\n map_location=\"cpu\",\n check_hash=True,\n )\n state_dict = checkpoint[\"model\"]\n state_dict[\"pos_embed\"] = interpolate_pos_embed(\n state_dict[\"pos_embed\"], visual_encoder\n )\n msg = visual_encoder.load_state_dict(state_dict, strict=False)\n\n elif vit_type == \"large\":\n vision_width = 1024\n visual_encoder = cls(\n img_size=image_size,\n patch_size=16,\n embed_dim=vision_width,\n depth=24,\n num_heads=16,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0.1 or drop_path_rate,\n norm_layer=norm_layer,\n )\n if from_pretrained:\n from timm.models.helpers import load_custom_pretrained\n from timm.models.vision_transformer import default_cfgs\n\n load_custom_pretrained(\n visual_encoder, default_cfgs[\"vit_large_patch16_224_in21k\"]\n )\n\n visual_encoder.vision_width = vision_width\n return visual_encoder\n\n def forward_features(self, x, register_blk=-1):\n return super().forward(x, register_blk)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.vit.forward_features","uri":"program://CREMA/function/lavis.models.vit.forward_features#L526-L527","kind":"function","name":"forward_features","path":"lavis/models/vit.py","language":"python","start_line":526,"end_line":527,"context_start_line":506,"context_end_line":527,"code":" patch_size=16,\n embed_dim=vision_width,\n depth=24,\n num_heads=16,\n use_grad_checkpointing=use_grad_checkpointing,\n ckpt_layer=ckpt_layer,\n drop_path_rate=0.1 or drop_path_rate,\n norm_layer=norm_layer,\n )\n if from_pretrained:\n from timm.models.helpers import load_custom_pretrained\n from timm.models.vision_transformer import default_cfgs\n\n load_custom_pretrained(\n visual_encoder, default_cfgs[\"vit_large_patch16_224_in21k\"]\n )\n\n visual_encoder.vision_width = vision_width\n return visual_encoder\n\n def forward_features(self, x, register_blk=-1):\n return super().forward(x, register_blk)","source_hash":"1bd9312ba2907fd505d918ef6829bf901e8b7e32e3110a8d1c611e58612a1abd","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit","uri":"program://CREMA/module/lavis.models.eva_vit#L1-L494","kind":"module","name":"lavis.models.eva_vit","path":"lavis/models/eva_vit.py","language":"python","start_line":1,"end_line":494,"context_start_line":1,"context_end_line":494,"code":"# Based on EVA, BEIT, timm and DeiT code bases\n# https://github.com/baaivision/EVA\n# https://github.com/rwightman/pytorch-image-models/tree/master/timm\n# https://github.com/microsoft/unilm/tree/master/beit\n# https://github.com/facebookresearch/deit/\n# https://github.com/facebookresearch/dino\n# --------------------------------------------------------'\nimport math\nfrom functools import partial\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils.checkpoint as checkpoint\nfrom timm.models.layers import drop_path, to_2tuple, trunc_normal_\nfrom timm.models.registry import register_model\n\nfrom lavis.common.dist_utils import download_cached_file\n\ndef _cfg(url='', **kwargs):\n return {\n 'url': url,\n 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,\n 'crop_pct': .9, 'interpolation': 'bicubic',\n 'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),\n **kwargs\n }\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n \"\"\"\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)\n \n def extra_repr(self) -> str:\n return 'p={}'.format(self.drop_prob)\n\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n # x = self.drop(x)\n # commit this for the orignal BERT implement \n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,\n proj_drop=0., window_size=None, attn_head_dim=None):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n if attn_head_dim is not None:\n head_dim = attn_head_dim\n all_head_dim = head_dim * self.num_heads\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)\n if qkv_bias:\n self.q_bias = nn.Parameter(torch.zeros(all_head_dim))\n self.v_bias = nn.Parameter(torch.zeros(all_head_dim))\n else:\n self.q_bias = None\n self.v_bias = None\n\n if window_size:\n self.window_size = window_size\n self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3\n self.relative_position_bias_table = nn.Parameter(\n torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH\n # cls to token & token 2 cls & cls to cls\n\n # get pair-wise relative position index for each token inside the window\n coords_h = torch.arange(window_size[0])\n coords_w = torch.arange(window_size[1])\n coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww\n coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww\n relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww\n relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2\n relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0\n relative_coords[:, :, 1] += window_size[1] - 1\n relative_coords[:, :, 0] *= 2 * window_size[1] - 1\n relative_position_index = \\\n torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)\n relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n else:\n self.window_size = None\n self.relative_position_bias_table = None\n self.relative_position_index = None\n\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(all_head_dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x, rel_pos_bias=None):\n B, N, C = x.shape\n qkv_bias = None\n if self.q_bias is not None:\n qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))\n # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)\n qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n\n q = q * self.scale\n attn = (q @ k.transpose(-2, -1))\n\n if self.relative_position_bias_table is not None:\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n attn = attn + relative_position_bias.unsqueeze(0)\n\n if rel_pos_bias is not None:\n attn = attn + rel_pos_bias\n \n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, -1)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,\n window_size=None, attn_head_dim=None):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,\n attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim)\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n if init_values is not None and init_values > 0:\n self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n else:\n self.gamma_1, self.gamma_2 = None, None\n\n def forward(self, x, rel_pos_bias=None):\n if self.gamma_1 is None:\n x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n else:\n x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\" Image to Patch Embedding\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)\n\n def forward(self, x, **kwargs):\n B, C, H, W = x.shape\n # FIXME look at relaxing size constraints\n assert H == self.img_size[0] and W == self.img_size[1], \\\n f\"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).\"\n x = self.proj(x).flatten(2).transpose(1, 2)\n return x\n\n\nclass RelativePositionBias(nn.Module):\n\n def __init__(self, window_size, num_heads):\n super().__init__()\n self.window_size = window_size\n self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3\n self.relative_position_bias_table = nn.Parameter(\n torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH\n # cls to token & token 2 cls & cls to cls\n\n # get pair-wise relative position index for each token inside the window\n coords_h = torch.arange(window_size[0])\n coords_w = torch.arange(window_size[1])\n coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww\n coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww\n relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww\n relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2\n relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0\n relative_coords[:, :, 1] += window_size[1] - 1\n relative_coords[:, :, 0] *= 2 * window_size[1] - 1\n relative_position_index = \\\n torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)\n relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n\n # trunc_normal_(self.relative_position_bias_table, std=.02)\n\n def forward(self):\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n\n\nclass VisionTransformer(nn.Module):\n \"\"\" Vision Transformer with support for patch or hybrid CNN input stage\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,\n num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,\n drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,\n use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,\n use_mean_pooling=True, init_scale=0.001, use_checkpoint=False):\n super().__init__()\n self.image_size = img_size\n self.num_classes = num_classes\n self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models\n\n self.patch_embed = PatchEmbed(\n img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n if use_abs_pos_emb:\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n else:\n self.pos_embed = None\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n if use_shared_rel_pos_bias:\n self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)\n else:\n self.rel_pos_bias = None\n self.use_checkpoint = use_checkpoint\n \n dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule\n self.use_rel_pos_bias = use_rel_pos_bias\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,\n init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)\n for i in range(depth)])\n# self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)\n# self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None\n# self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n\n# if self.fc_norm is not None:\n# t = x[:, 1:, :]\n# return self.fc_norm(t.mean(1))\n# else:\n# return x[:, 0]\n\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n\n return x, rel_pos_bias\n\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n features = []\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n x = blk(x, rel_pos_bias)\n features.append(x)\n\n return features\n \n \ndef interpolate_pos_embed(model, checkpoint_model):\n if 'pos_embed' in checkpoint_model:\n pos_embed_checkpoint = checkpoint_model['pos_embed'].float()\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = model.patch_embed.num_patches\n num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n checkpoint_model['pos_embed'] = new_pos_embed\n \n \ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n# if isinstance(l, (nn.MultiheadAttention, Attention)):\n# for attr in [*[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]], \"in_proj_bias\", \"bias_k\", \"bias_v\"]:\n# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n\ndef add_additional_channels_old(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n if num_additional_channels != 0:\n \n new_conv_weight = torch.zeros(1408, 3+num_additional_channels, 14, 14 )\n\n for key,value in state_dict.items():\n if key == \"patch_embed.proj.weight\":\n old_conv_weight = value\n new_conv_weight[:,0:3,:,:] = old_conv_weight\n state_dict[key] = new_conv_weight\n \ndef add_additional_channels(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n assert num_additional_channels % 3 == 0, \"num_additional_channels should be a multiple of 3\"\n \n if num_additional_channels != 0:\n old_conv_weight = state_dict[\"patch_embed.proj.weight\"]\n new_conv_weight = torch.zeros(1408, 3 + num_additional_channels, 14, 14)\n new_conv_\n# ... truncated ...","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit._cfg","uri":"program://CREMA/function/lavis.models.eva_vit._cfg#L20-L27","kind":"function","name":"_cfg","path":"lavis/models/eva_vit.py","language":"python","start_line":20,"end_line":27,"context_start_line":1,"context_end_line":47,"code":"# Based on EVA, BEIT, timm and DeiT code bases\n# https://github.com/baaivision/EVA\n# https://github.com/rwightman/pytorch-image-models/tree/master/timm\n# https://github.com/microsoft/unilm/tree/master/beit\n# https://github.com/facebookresearch/deit/\n# https://github.com/facebookresearch/dino\n# --------------------------------------------------------'\nimport math\nfrom functools import partial\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils.checkpoint as checkpoint\nfrom timm.models.layers import drop_path, to_2tuple, trunc_normal_\nfrom timm.models.registry import register_model\n\nfrom lavis.common.dist_utils import download_cached_file\n\ndef _cfg(url='', **kwargs):\n return {\n 'url': url,\n 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,\n 'crop_pct': .9, 'interpolation': 'bicubic',\n 'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),\n **kwargs\n }\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n \"\"\"\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)\n \n def extra_repr(self) -> str:\n return 'p={}'.format(self.drop_prob)\n\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.DropPath","uri":"program://CREMA/class/lavis.models.eva_vit.DropPath#L30-L41","kind":"class","name":"DropPath","path":"lavis/models/eva_vit.py","language":"python","start_line":30,"end_line":41,"context_start_line":10,"context_end_line":61,"code":"\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils.checkpoint as checkpoint\nfrom timm.models.layers import drop_path, to_2tuple, trunc_normal_\nfrom timm.models.registry import register_model\n\nfrom lavis.common.dist_utils import download_cached_file\n\ndef _cfg(url='', **kwargs):\n return {\n 'url': url,\n 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,\n 'crop_pct': .9, 'interpolation': 'bicubic',\n 'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),\n **kwargs\n }\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n \"\"\"\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)\n \n def extra_repr(self) -> str:\n return 'p={}'.format(self.drop_prob)\n\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n # x = self.drop(x)\n # commit this for the orignal BERT implement \n x = self.fc2(x)\n x = self.drop(x)\n return x","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.Mlp","uri":"program://CREMA/class/lavis.models.eva_vit.Mlp#L44-L61","kind":"class","name":"Mlp","path":"lavis/models/eva_vit.py","language":"python","start_line":44,"end_line":61,"context_start_line":24,"context_end_line":81,"code":" 'crop_pct': .9, 'interpolation': 'bicubic',\n 'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),\n **kwargs\n }\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n \"\"\"\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)\n \n def extra_repr(self) -> str:\n return 'p={}'.format(self.drop_prob)\n\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n # x = self.drop(x)\n # commit this for the orignal BERT implement \n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,\n proj_drop=0., window_size=None, attn_head_dim=None):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n if attn_head_dim is not None:\n head_dim = attn_head_dim\n all_head_dim = head_dim * self.num_heads\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)\n if qkv_bias:\n self.q_bias = nn.Parameter(torch.zeros(all_head_dim))\n self.v_bias = nn.Parameter(torch.zeros(all_head_dim))\n else:\n self.q_bias = None","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.Attention","uri":"program://CREMA/class/lavis.models.eva_vit.Attention#L64-L148","kind":"class","name":"Attention","path":"lavis/models/eva_vit.py","language":"python","start_line":64,"end_line":148,"context_start_line":44,"context_end_line":168,"code":"class Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n # x = self.drop(x)\n # commit this for the orignal BERT implement \n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0.,\n proj_drop=0., window_size=None, attn_head_dim=None):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n if attn_head_dim is not None:\n head_dim = attn_head_dim\n all_head_dim = head_dim * self.num_heads\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, all_head_dim * 3, bias=False)\n if qkv_bias:\n self.q_bias = nn.Parameter(torch.zeros(all_head_dim))\n self.v_bias = nn.Parameter(torch.zeros(all_head_dim))\n else:\n self.q_bias = None\n self.v_bias = None\n\n if window_size:\n self.window_size = window_size\n self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3\n self.relative_position_bias_table = nn.Parameter(\n torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH\n # cls to token & token 2 cls & cls to cls\n\n # get pair-wise relative position index for each token inside the window\n coords_h = torch.arange(window_size[0])\n coords_w = torch.arange(window_size[1])\n coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww\n coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww\n relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww\n relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2\n relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0\n relative_coords[:, :, 1] += window_size[1] - 1\n relative_coords[:, :, 0] *= 2 * window_size[1] - 1\n relative_position_index = \\\n torch.zeros(size=(window_size[0] * window_size[1] + 1, ) * 2, dtype=relative_coords.dtype)\n relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n else:\n self.window_size = None\n self.relative_position_bias_table = None\n self.relative_position_index = None\n\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(all_head_dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x, rel_pos_bias=None):\n B, N, C = x.shape\n qkv_bias = None\n if self.q_bias is not None:\n qkv_bias = torch.cat((self.q_bias, torch.zeros_like(self.v_bias, requires_grad=False), self.v_bias))\n # qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n qkv = F.linear(input=x, weight=self.qkv.weight, bias=qkv_bias)\n qkv = qkv.reshape(B, N, 3, self.num_heads, -1).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n\n q = q * self.scale\n attn = (q @ k.transpose(-2, -1))\n\n if self.relative_position_bias_table is not None:\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n attn = attn + relative_position_bias.unsqueeze(0)\n\n if rel_pos_bias is not None:\n attn = attn + rel_pos_bias\n \n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, -1)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,\n window_size=None, attn_head_dim=None):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,\n attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim)\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n if init_values is not None and init_values > 0:\n self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.Block","uri":"program://CREMA/class/lavis.models.eva_vit.Block#L151-L180","kind":"class","name":"Block","path":"lavis/models/eva_vit.py","language":"python","start_line":151,"end_line":180,"context_start_line":131,"context_end_line":200,"code":" if self.relative_position_bias_table is not None:\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n attn = attn + relative_position_bias.unsqueeze(0)\n\n if rel_pos_bias is not None:\n attn = attn + rel_pos_bias\n \n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, -1)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., init_values=None, act_layer=nn.GELU, norm_layer=nn.LayerNorm,\n window_size=None, attn_head_dim=None):\n super().__init__()\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale,\n attn_drop=attn_drop, proj_drop=drop, window_size=window_size, attn_head_dim=attn_head_dim)\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n if init_values is not None and init_values > 0:\n self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n else:\n self.gamma_1, self.gamma_2 = None, None\n\n def forward(self, x, rel_pos_bias=None):\n if self.gamma_1 is None:\n x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n else:\n x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\" Image to Patch Embedding\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)\n\n def forward(self, x, **kwargs):\n B, C, H, W = x.shape\n # FIXME look at relaxing size constraints","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.PatchEmbed","uri":"program://CREMA/class/lavis.models.eva_vit.PatchEmbed#L183-L204","kind":"class","name":"PatchEmbed","path":"lavis/models/eva_vit.py","language":"python","start_line":183,"end_line":204,"context_start_line":163,"context_end_line":224,"code":" self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n if init_values is not None and init_values > 0:\n self.gamma_1 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n self.gamma_2 = nn.Parameter(init_values * torch.ones((dim)),requires_grad=True)\n else:\n self.gamma_1, self.gamma_2 = None, None\n\n def forward(self, x, rel_pos_bias=None):\n if self.gamma_1 is None:\n x = x + self.drop_path(self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n else:\n x = x + self.drop_path(self.gamma_1 * self.attn(self.norm1(x), rel_pos_bias=rel_pos_bias))\n x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\" Image to Patch Embedding\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)\n\n def forward(self, x, **kwargs):\n B, C, H, W = x.shape\n # FIXME look at relaxing size constraints\n assert H == self.img_size[0] and W == self.img_size[1], \\\n f\"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).\"\n x = self.proj(x).flatten(2).transpose(1, 2)\n return x\n\n\nclass RelativePositionBias(nn.Module):\n\n def __init__(self, window_size, num_heads):\n super().__init__()\n self.window_size = window_size\n self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3\n self.relative_position_bias_table = nn.Parameter(\n torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH\n # cls to token & token 2 cls & cls to cls\n\n # get pair-wise relative position index for each token inside the window\n coords_h = torch.arange(window_size[0])\n coords_w = torch.arange(window_size[1])\n coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww\n coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww\n relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww\n relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2\n relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.RelativePositionBias","uri":"program://CREMA/class/lavis.models.eva_vit.RelativePositionBias#L207-L243","kind":"class","name":"RelativePositionBias","path":"lavis/models/eva_vit.py","language":"python","start_line":207,"end_line":243,"context_start_line":187,"context_end_line":263,"code":" super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.patch_shape = (img_size[0] // patch_size[0], img_size[1] // patch_size[1])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)\n\n def forward(self, x, **kwargs):\n B, C, H, W = x.shape\n # FIXME look at relaxing size constraints\n assert H == self.img_size[0] and W == self.img_size[1], \\\n f\"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]}).\"\n x = self.proj(x).flatten(2).transpose(1, 2)\n return x\n\n\nclass RelativePositionBias(nn.Module):\n\n def __init__(self, window_size, num_heads):\n super().__init__()\n self.window_size = window_size\n self.num_relative_distance = (2 * window_size[0] - 1) * (2 * window_size[1] - 1) + 3\n self.relative_position_bias_table = nn.Parameter(\n torch.zeros(self.num_relative_distance, num_heads)) # 2*Wh-1 * 2*Ww-1, nH\n # cls to token & token 2 cls & cls to cls\n\n # get pair-wise relative position index for each token inside the window\n coords_h = torch.arange(window_size[0])\n coords_w = torch.arange(window_size[1])\n coords = torch.stack(torch.meshgrid([coords_h, coords_w])) # 2, Wh, Ww\n coords_flatten = torch.flatten(coords, 1) # 2, Wh*Ww\n relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] # 2, Wh*Ww, Wh*Ww\n relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2\n relative_coords[:, :, 0] += window_size[0] - 1 # shift to start from 0\n relative_coords[:, :, 1] += window_size[1] - 1\n relative_coords[:, :, 0] *= 2 * window_size[1] - 1\n relative_position_index = \\\n torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)\n relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n\n # trunc_normal_(self.relative_position_bias_table, std=.02)\n\n def forward(self):\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n\n\nclass VisionTransformer(nn.Module):\n \"\"\" Vision Transformer with support for patch or hybrid CNN input stage\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,\n num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,\n drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,\n use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,\n use_mean_pooling=True, init_scale=0.001, use_checkpoint=False):\n super().__init__()\n self.image_size = img_size\n self.num_classes = num_classes\n self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models\n\n self.patch_embed = PatchEmbed(\n img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.VisionTransformer","uri":"program://CREMA/class/lavis.models.eva_vit.VisionTransformer#L246-L384","kind":"class","name":"VisionTransformer","path":"lavis/models/eva_vit.py","language":"python","start_line":246,"end_line":384,"context_start_line":226,"context_end_line":404,"code":" relative_coords[:, :, 0] *= 2 * window_size[1] - 1\n relative_position_index = \\\n torch.zeros(size=(window_size[0] * window_size[1] + 1,) * 2, dtype=relative_coords.dtype)\n relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n\n # trunc_normal_(self.relative_position_bias_table, std=.02)\n\n def forward(self):\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n\n\nclass VisionTransformer(nn.Module):\n \"\"\" Vision Transformer with support for patch or hybrid CNN input stage\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,\n num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,\n drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,\n use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,\n use_mean_pooling=True, init_scale=0.001, use_checkpoint=False):\n super().__init__()\n self.image_size = img_size\n self.num_classes = num_classes\n self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models\n\n self.patch_embed = PatchEmbed(\n img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n if use_abs_pos_emb:\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n else:\n self.pos_embed = None\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n if use_shared_rel_pos_bias:\n self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)\n else:\n self.rel_pos_bias = None\n self.use_checkpoint = use_checkpoint\n \n dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule\n self.use_rel_pos_bias = use_rel_pos_bias\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,\n init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)\n for i in range(depth)])\n# self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)\n# self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None\n# self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n\n# if self.fc_norm is not None:\n# t = x[:, 1:, :]\n# return self.fc_norm(t.mean(1))\n# else:\n# return x[:, 0]\n\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n\n return x, rel_pos_bias\n\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n features = []\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n x = blk(x, rel_pos_bias)\n features.append(x)\n\n return features\n \n \ndef interpolate_pos_embed(model, checkpoint_model):\n if 'pos_embed' in checkpoint_model:\n pos_embed_checkpoint = checkpoint_model['pos_embed'].float()\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = model.patch_embed.num_patches\n num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.interpolate_pos_embed","uri":"program://CREMA/function/lavis.models.eva_vit.interpolate_pos_embed#L387-L408","kind":"function","name":"interpolate_pos_embed","path":"lavis/models/eva_vit.py","language":"python","start_line":387,"end_line":408,"context_start_line":367,"context_end_line":428,"code":"\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n features = []\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n x = blk(x, rel_pos_bias)\n features.append(x)\n\n return features\n \n \ndef interpolate_pos_embed(model, checkpoint_model):\n if 'pos_embed' in checkpoint_model:\n pos_embed_checkpoint = checkpoint_model['pos_embed'].float()\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = model.patch_embed.num_patches\n num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n checkpoint_model['pos_embed'] = new_pos_embed\n \n \ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n# if isinstance(l, (nn.MultiheadAttention, Attention)):\n# for attr in [*[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]], \"in_proj_bias\", \"bias_k\", \"bias_v\"]:\n# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.convert_weights_to_fp16","uri":"program://CREMA/function/lavis.models.eva_vit.convert_weights_to_fp16#L411-L426","kind":"function","name":"convert_weights_to_fp16","path":"lavis/models/eva_vit.py","language":"python","start_line":411,"end_line":426,"context_start_line":391,"context_end_line":446,"code":" num_patches = model.patch_embed.num_patches\n num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n checkpoint_model['pos_embed'] = new_pos_embed\n \n \ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n# if isinstance(l, (nn.MultiheadAttention, Attention)):\n# for attr in [*[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]], \"in_proj_bias\", \"bias_k\", \"bias_v\"]:\n# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n\ndef add_additional_channels_old(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n if num_additional_channels != 0:\n \n new_conv_weight = torch.zeros(1408, 3+num_additional_channels, 14, 14 )\n\n for key,value in state_dict.items():\n if key == \"patch_embed.proj.weight\":\n old_conv_weight = value\n new_conv_weight[:,0:3,:,:] = old_conv_weight\n state_dict[key] = new_conv_weight\n \ndef add_additional_channels(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n assert num_additional_channels % 3 == 0, \"num_additional_channels should be a multiple of 3\"\n ","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.add_additional_channels_old","uri":"program://CREMA/function/lavis.models.eva_vit.add_additional_channels_old#L429-L440","kind":"function","name":"add_additional_channels_old","path":"lavis/models/eva_vit.py","language":"python","start_line":429,"end_line":440,"context_start_line":409,"context_end_line":460,"code":" \n \ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n# if isinstance(l, (nn.MultiheadAttention, Attention)):\n# for attr in [*[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]], \"in_proj_bias\", \"bias_k\", \"bias_v\"]:\n# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n\ndef add_additional_channels_old(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n if num_additional_channels != 0:\n \n new_conv_weight = torch.zeros(1408, 3+num_additional_channels, 14, 14 )\n\n for key,value in state_dict.items():\n if key == \"patch_embed.proj.weight\":\n old_conv_weight = value\n new_conv_weight[:,0:3,:,:] = old_conv_weight\n state_dict[key] = new_conv_weight\n \ndef add_additional_channels(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n assert num_additional_channels % 3 == 0, \"num_additional_channels should be a multiple of 3\"\n \n if num_additional_channels != 0:\n old_conv_weight = state_dict[\"patch_embed.proj.weight\"]\n new_conv_weight = torch.zeros(1408, 3 + num_additional_channels, 14, 14)\n new_conv_weight[:, :3, :, :] = old_conv_weight.clone()\n \n for i in range(num_additional_channels // 3):\n start_idx = 3 * (i + 1)\n new_conv_weight[:, start_idx:start_idx + 3, :, :] = old_conv_weight.clone()\n \n state_dict[\"patch_embed.proj.weight\"] = new_conv_weight\n\ndef create_eva_vit_g(img_size=224,drop_path_rate=0.4,use_checkpoint=False,precision=\"fp16\", in_chans=3):\n model = VisionTransformer(\n img_size=img_size,","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.add_additional_channels","uri":"program://CREMA/function/lavis.models.eva_vit.add_additional_channels#L442-L456","kind":"function","name":"add_additional_channels","path":"lavis/models/eva_vit.py","language":"python","start_line":442,"end_line":456,"context_start_line":422,"context_end_line":476,"code":"# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n\ndef add_additional_channels_old(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n if num_additional_channels != 0:\n \n new_conv_weight = torch.zeros(1408, 3+num_additional_channels, 14, 14 )\n\n for key,value in state_dict.items():\n if key == \"patch_embed.proj.weight\":\n old_conv_weight = value\n new_conv_weight[:,0:3,:,:] = old_conv_weight\n state_dict[key] = new_conv_weight\n \ndef add_additional_channels(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n assert num_additional_channels % 3 == 0, \"num_additional_channels should be a multiple of 3\"\n \n if num_additional_channels != 0:\n old_conv_weight = state_dict[\"patch_embed.proj.weight\"]\n new_conv_weight = torch.zeros(1408, 3 + num_additional_channels, 14, 14)\n new_conv_weight[:, :3, :, :] = old_conv_weight.clone()\n \n for i in range(num_additional_channels // 3):\n start_idx = 3 * (i + 1)\n new_conv_weight[:, start_idx:start_idx + 3, :, :] = old_conv_weight.clone()\n \n state_dict[\"patch_embed.proj.weight\"] = new_conv_weight\n\ndef create_eva_vit_g(img_size=224,drop_path_rate=0.4,use_checkpoint=False,precision=\"fp16\", in_chans=3):\n model = VisionTransformer(\n img_size=img_size,\n patch_size=14,\n use_mean_pooling=False,\n embed_dim=1408,\n in_chans=in_chans,\n depth=39,\n num_heads=1408//88,\n mlp_ratio=4.3637,\n qkv_bias=True,\n drop_path_rate=drop_path_rate,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n use_checkpoint=use_checkpoint,\n ) \n url = \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth\"\n cached_file = download_cached_file(\n url, check_hash=False, progress=True\n )","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.create_eva_vit_g","uri":"program://CREMA/function/lavis.models.eva_vit.create_eva_vit_g#L458-L494","kind":"function","name":"create_eva_vit_g","path":"lavis/models/eva_vit.py","language":"python","start_line":458,"end_line":494,"context_start_line":438,"context_end_line":494,"code":" old_conv_weight = value\n new_conv_weight[:,0:3,:,:] = old_conv_weight\n state_dict[key] = new_conv_weight\n \ndef add_additional_channels(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n assert num_additional_channels % 3 == 0, \"num_additional_channels should be a multiple of 3\"\n \n if num_additional_channels != 0:\n old_conv_weight = state_dict[\"patch_embed.proj.weight\"]\n new_conv_weight = torch.zeros(1408, 3 + num_additional_channels, 14, 14)\n new_conv_weight[:, :3, :, :] = old_conv_weight.clone()\n \n for i in range(num_additional_channels // 3):\n start_idx = 3 * (i + 1)\n new_conv_weight[:, start_idx:start_idx + 3, :, :] = old_conv_weight.clone()\n \n state_dict[\"patch_embed.proj.weight\"] = new_conv_weight\n\ndef create_eva_vit_g(img_size=224,drop_path_rate=0.4,use_checkpoint=False,precision=\"fp16\", in_chans=3):\n model = VisionTransformer(\n img_size=img_size,\n patch_size=14,\n use_mean_pooling=False,\n embed_dim=1408,\n in_chans=in_chans,\n depth=39,\n num_heads=1408//88,\n mlp_ratio=4.3637,\n qkv_bias=True,\n drop_path_rate=drop_path_rate,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n use_checkpoint=use_checkpoint,\n ) \n url = \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/models/BLIP2/eva_vit_g.pth\"\n cached_file = download_cached_file(\n url, check_hash=False, progress=True\n )\n\n state_dict = torch.load(cached_file, map_location=\"cpu\") \n\n if in_chans > 3:\n num_additional_channels = in_chans - 3\n print('num_additional_channels', num_additional_channels)\n add_additional_channels(state_dict, num_additional_channels)\n\n interpolate_pos_embed(model, state_dict)\n \n incompatible_keys = model.load_state_dict(state_dict, strict=False)\n# print(incompatible_keys)\n \n if precision == \"fp16\":\n# model.to(\"cuda\") \n\n convert_weights_to_fp16(model)\n return model","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.__init__","uri":"program://CREMA/function/lavis.models.eva_vit.__init__#L249-L295","kind":"function","name":"__init__","path":"lavis/models/eva_vit.py","language":"python","start_line":249,"end_line":295,"context_start_line":229,"context_end_line":315,"code":" relative_position_index[1:, 1:] = relative_coords.sum(-1) # Wh*Ww, Wh*Ww\n relative_position_index[0, 0:] = self.num_relative_distance - 3\n relative_position_index[0:, 0] = self.num_relative_distance - 2\n relative_position_index[0, 0] = self.num_relative_distance - 1\n\n self.register_buffer(\"relative_position_index\", relative_position_index)\n\n # trunc_normal_(self.relative_position_bias_table, std=.02)\n\n def forward(self):\n relative_position_bias = \\\n self.relative_position_bias_table[self.relative_position_index.view(-1)].view(\n self.window_size[0] * self.window_size[1] + 1,\n self.window_size[0] * self.window_size[1] + 1, -1) # Wh*Ww,Wh*Ww,nH\n return relative_position_bias.permute(2, 0, 1).contiguous() # nH, Wh*Ww, Wh*Ww\n\n\nclass VisionTransformer(nn.Module):\n \"\"\" Vision Transformer with support for patch or hybrid CNN input stage\n \"\"\"\n def __init__(self, img_size=224, patch_size=16, in_chans=3, num_classes=1000, embed_dim=768, depth=12,\n num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop_rate=0., attn_drop_rate=0.,\n drop_path_rate=0., norm_layer=nn.LayerNorm, init_values=None,\n use_abs_pos_emb=True, use_rel_pos_bias=False, use_shared_rel_pos_bias=False,\n use_mean_pooling=True, init_scale=0.001, use_checkpoint=False):\n super().__init__()\n self.image_size = img_size\n self.num_classes = num_classes\n self.num_features = self.embed_dim = embed_dim # num_features for consistency with other models\n\n self.patch_embed = PatchEmbed(\n img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim)\n num_patches = self.patch_embed.num_patches\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n if use_abs_pos_emb:\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n else:\n self.pos_embed = None\n self.pos_drop = nn.Dropout(p=drop_rate)\n\n if use_shared_rel_pos_bias:\n self.rel_pos_bias = RelativePositionBias(window_size=self.patch_embed.patch_shape, num_heads=num_heads)\n else:\n self.rel_pos_bias = None\n self.use_checkpoint = use_checkpoint\n \n dpr = [x.item() for x in torch.linspace(0, drop_path_rate, depth)] # stochastic depth decay rule\n self.use_rel_pos_bias = use_rel_pos_bias\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,\n init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)\n for i in range(depth)])\n# self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)\n# self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None\n# self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.forward","uri":"program://CREMA/function/lavis.models.eva_vit.forward#L349-L352","kind":"function","name":"forward","path":"lavis/models/eva_vit.py","language":"python","start_line":349,"end_line":352,"context_start_line":329,"context_end_line":372,"code":" x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n\n# if self.fc_norm is not None:\n# t = x[:, 1:, :]\n# return self.fc_norm(t.mean(1))\n# else:\n# return x[:, 0]\n\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n\n return x, rel_pos_bias\n\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.extra_repr","uri":"program://CREMA/function/lavis.models.eva_vit.extra_repr#L40-L41","kind":"function","name":"extra_repr","path":"lavis/models/eva_vit.py","language":"python","start_line":40,"end_line":41,"context_start_line":20,"context_end_line":61,"code":"def _cfg(url='', **kwargs):\n return {\n 'url': url,\n 'num_classes': 1000, 'input_size': (3, 224, 224), 'pool_size': None,\n 'crop_pct': .9, 'interpolation': 'bicubic',\n 'mean': (0.5, 0.5, 0.5), 'std': (0.5, 0.5, 0.5),\n **kwargs\n }\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n \"\"\"\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)\n \n def extra_repr(self) -> str:\n return 'p={}'.format(self.drop_prob)\n\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n # x = self.drop(x)\n # commit this for the orignal BERT implement \n x = self.fc2(x)\n x = self.drop(x)\n return x","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.fix_init_weight","uri":"program://CREMA/function/lavis.models.eva_vit.fix_init_weight#L300-L306","kind":"function","name":"fix_init_weight","path":"lavis/models/eva_vit.py","language":"python","start_line":300,"end_line":306,"context_start_line":280,"context_end_line":326,"code":" dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,\n init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)\n for i in range(depth)])\n# self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)\n# self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None\n# self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit._init_weights","uri":"program://CREMA/function/lavis.models.eva_vit._init_weights#L308-L315","kind":"function","name":"_init_weights","path":"lavis/models/eva_vit.py","language":"python","start_line":308,"end_line":315,"context_start_line":288,"context_end_line":335,"code":" if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.get_classifier","uri":"program://CREMA/function/lavis.models.eva_vit.get_classifier#L317-L318","kind":"function","name":"get_classifier","path":"lavis/models/eva_vit.py","language":"python","start_line":317,"end_line":318,"context_start_line":297,"context_end_line":338,"code":"# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.reset_classifier","uri":"program://CREMA/function/lavis.models.eva_vit.reset_classifier#L320-L322","kind":"function","name":"reset_classifier","path":"lavis/models/eva_vit.py","language":"python","start_line":320,"end_line":322,"context_start_line":300,"context_end_line":342,"code":" def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.forward_features","uri":"program://CREMA/function/lavis.models.eva_vit.forward_features#L324-L340","kind":"function","name":"forward_features","path":"lavis/models/eva_vit.py","language":"python","start_line":324,"end_line":340,"context_start_line":304,"context_end_line":360,"code":" for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n def forward_features(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n\n# if self.fc_norm is not None:\n# t = x[:, 1:, :]\n# return self.fc_norm(t.mean(1))\n# else:\n# return x[:, 0]\n\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.forward_before_blk","uri":"program://CREMA/function/lavis.models.eva_vit.forward_before_blk#L354-L366","kind":"function","name":"forward_before_blk","path":"lavis/models/eva_vit.py","language":"python","start_line":354,"end_line":366,"context_start_line":334,"context_end_line":386,"code":" rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n if self.use_checkpoint:\n x = checkpoint.checkpoint(blk, x, rel_pos_bias)\n else:\n x = blk(x, rel_pos_bias)\n return x\n# x = self.norm(x)\n\n# if self.fc_norm is not None:\n# t = x[:, 1:, :]\n# return self.fc_norm(t.mean(1))\n# else:\n# return x[:, 0]\n\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n\n return x, rel_pos_bias\n\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n features = []\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n x = blk(x, rel_pos_bias)\n features.append(x)\n\n return features\n \n ","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.get_intermediate_layers","uri":"program://CREMA/function/lavis.models.eva_vit.get_intermediate_layers#L368-L384","kind":"function","name":"get_intermediate_layers","path":"lavis/models/eva_vit.py","language":"python","start_line":368,"end_line":384,"context_start_line":348,"context_end_line":404,"code":"\n def forward(self, x):\n x = self.forward_features(x)\n# x = self.head(x)\n return x\n\n def forward_before_blk(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n\n return x, rel_pos_bias\n\n def get_intermediate_layers(self, x):\n x = self.patch_embed(x)\n batch_size, seq_len, _ = x.size()\n\n cls_tokens = self.cls_token.expand(batch_size, -1, -1) # stole cls_tokens impl from Phil Wang, thanks\n x = torch.cat((cls_tokens, x), dim=1)\n if self.pos_embed is not None:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n features = []\n rel_pos_bias = self.rel_pos_bias() if self.rel_pos_bias is not None else None\n for blk in self.blocks:\n x = blk(x, rel_pos_bias)\n features.append(x)\n\n return features\n \n \ndef interpolate_pos_embed(model, checkpoint_model):\n if 'pos_embed' in checkpoint_model:\n pos_embed_checkpoint = checkpoint_model['pos_embed'].float()\n embedding_size = pos_embed_checkpoint.shape[-1]\n num_patches = model.patch_embed.num_patches\n num_extra_tokens = model.pos_embed.shape[-2] - num_patches\n # height (== width) for the checkpoint position embedding\n orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit._convert_weights_to_fp16","uri":"program://CREMA/function/lavis.models.eva_vit._convert_weights_to_fp16#L414-L418","kind":"function","name":"_convert_weights_to_fp16","path":"lavis/models/eva_vit.py","language":"python","start_line":414,"end_line":418,"context_start_line":394,"context_end_line":438,"code":" orig_size = int((pos_embed_checkpoint.shape[-2] - num_extra_tokens) ** 0.5)\n # height (== width) for the new position embedding\n new_size = int(num_patches ** 0.5)\n # class_token and dist_token are kept unchanged\n if orig_size != new_size:\n print(\"Position interpolate from %dx%d to %dx%d\" % (orig_size, orig_size, new_size, new_size))\n extra_tokens = pos_embed_checkpoint[:, :num_extra_tokens]\n # only the position tokens are interpolated\n pos_tokens = pos_embed_checkpoint[:, num_extra_tokens:]\n pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)\n pos_tokens = torch.nn.functional.interpolate(\n pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)\n pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(1, 2)\n new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=1)\n checkpoint_model['pos_embed'] = new_pos_embed\n \n \ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n# if isinstance(l, (nn.MultiheadAttention, Attention)):\n# for attr in [*[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]], \"in_proj_bias\", \"bias_k\", \"bias_v\"]:\n# tensor = getattr(l, attr)\n# if tensor is not None:\n# tensor.data = tensor.data.half()\n\n model.apply(_convert_weights_to_fp16)\n \n\ndef add_additional_channels_old(state_dict, num_additional_channels):\n \"state_dict should be just from unet model, not the entire SD or GLIGEN\"\n\n if num_additional_channels != 0:\n \n new_conv_weight = torch.zeros(1408, 3+num_additional_channels, 14, 14 )\n\n for key,value in state_dict.items():\n if key == \"patch_embed.proj.weight\":\n old_conv_weight = value","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.eva_vit.rescale","uri":"program://CREMA/function/lavis.models.eva_vit.rescale#L301-L302","kind":"function","name":"rescale","path":"lavis/models/eva_vit.py","language":"python","start_line":301,"end_line":302,"context_start_line":281,"context_end_line":322,"code":" drop=drop_rate, attn_drop=attn_drop_rate, drop_path=dpr[i], norm_layer=norm_layer,\n init_values=init_values, window_size=self.patch_embed.patch_shape if use_rel_pos_bias else None)\n for i in range(depth)])\n# self.norm = nn.Identity() if use_mean_pooling else norm_layer(embed_dim)\n# self.fc_norm = norm_layer(embed_dim) if use_mean_pooling else None\n# self.head = nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n\n if self.pos_embed is not None:\n trunc_normal_(self.pos_embed, std=.02)\n trunc_normal_(self.cls_token, std=.02)\n # trunc_normal_(self.mask_token, std=.02)\n# if isinstance(self.head, nn.Linear):\n# trunc_normal_(self.head.weight, std=.02)\n self.apply(self._init_weights)\n self.fix_init_weight()\n# if isinstance(self.head, nn.Linear):\n# self.head.weight.data.mul_(init_scale)\n# self.head.bias.data.mul_(init_scale)\n\n def fix_init_weight(self):\n def rescale(param, layer_id):\n param.div_(math.sqrt(2.0 * layer_id))\n\n for layer_id, layer in enumerate(self.blocks):\n rescale(layer.attn.proj.weight.data, layer_id + 1)\n rescale(layer.mlp.fc2.weight.data, layer_id + 1)\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=''):\n self.num_classes = num_classes\n self.head = nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()","source_hash":"94c6531e6281f8f1e5a0a4a80d6daf47d1ac4d4e5fcfab01a577aac3e740d7fc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model","uri":"program://CREMA/module/lavis.models.base_model#L1-L247","kind":"module","name":"lavis.models.base_model","path":"lavis/models/base_model.py","language":"python","start_line":1,"end_line":247,"context_start_line":1,"context_end_line":247,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nfrom lavis.common.dist_utils import download_cached_file, is_dist_avail_and_initialized\nfrom lavis.common.utils import get_abs_path, is_url\nfrom omegaconf import OmegaConf\n\n\nclass BaseModel(nn.Module):\n \"\"\"Base class for models.\"\"\"\n\n def __init__(self):\n super().__init__()\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n def load_checkpoint(self, url_or_filename):\n \"\"\"\n Load from a finetuned checkpoint.\n\n This should expect no mismatch in the model keys and the checkpoint keys.\n \"\"\"\n\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint.keys():\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n\n @classmethod\n def from_pretrained(cls, model_type):\n \"\"\"\n Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}\".format(model_type)\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n def load_checkpoint_from_config(self, cfg, **kwargs):\n \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"\n self.load_from_pretrained(url_or_filename=pretrain_path, **kwargs)\n\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)\n return all_gradients[torch.distributed.get_rank()]\n\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)\n tensor_all = GatherLayer.apply(tensors)\n\n return torch.cat(tensor_all, dim=0)\n\n\n@torch.no_grad()\ndef concat_all_gather(tensor):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n *** Warning ***: torch.distributed.all_gather has no gradient.\n \"\"\"\n # if use distributed training\n if not is_dist_avail_and_initialized():\n return tensor\n\n tensors_gather = [\n torch.ones_like(tensor) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(tensors_gather, tensor, async_op=False)\n\n output = torch.cat(tensors_gather, dim=0)\n return output\n\n\ndef tile(x, dim, n_tile):\n init_dim = x.size(dim)\n repeat_idx = [1] * x.dim()\n repeat_idx[dim] = n_tile\n x = x.repeat(*(repeat_idx))\n order_index = torch.LongTensor(\n np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)])\n )\n return torch.index_select(x, dim, order_index.to(x.device))","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.BaseModel","uri":"program://CREMA/class/lavis.models.base_model.BaseModel#L19-L118","kind":"class","name":"BaseModel","path":"lavis/models/base_model.py","language":"python","start_line":19,"end_line":118,"context_start_line":1,"context_end_line":138,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nfrom lavis.common.dist_utils import download_cached_file, is_dist_avail_and_initialized\nfrom lavis.common.utils import get_abs_path, is_url\nfrom omegaconf import OmegaConf\n\n\nclass BaseModel(nn.Module):\n \"\"\"Base class for models.\"\"\"\n\n def __init__(self):\n super().__init__()\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n def load_checkpoint(self, url_or_filename):\n \"\"\"\n Load from a finetuned checkpoint.\n\n This should expect no mismatch in the model keys and the checkpoint keys.\n \"\"\"\n\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint.keys():\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n\n @classmethod\n def from_pretrained(cls, model_type):\n \"\"\"\n Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}\".format(model_type)\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n def load_checkpoint_from_config(self, cfg, **kwargs):\n \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"\n self.load_from_pretrained(url_or_filename=pretrain_path, **kwargs)\n\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.BaseEncoder","uri":"program://CREMA/class/lavis.models.base_model.BaseEncoder#L121-L134","kind":"class","name":"BaseEncoder","path":"lavis/models/base_model.py","language":"python","start_line":121,"end_line":134,"context_start_line":101,"context_end_line":154,"code":"\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.SharedQueueMixin","uri":"program://CREMA/class/lavis.models.base_model.SharedQueueMixin#L137-L158","kind":"class","name":"SharedQueueMixin","path":"lavis/models/base_model.py","language":"python","start_line":137,"end_line":158,"context_start_line":117,"context_end_line":178,"code":" else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.MomentumDistilationMixin","uri":"program://CREMA/class/lavis.models.base_model.MomentumDistilationMixin#L161-L179","kind":"class","name":"MomentumDistilationMixin","path":"lavis/models/base_model.py","language":"python","start_line":161,"end_line":179,"context_start_line":141,"context_end_line":199,"code":" image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.GatherLayer","uri":"program://CREMA/class/lavis.models.base_model.GatherLayer#L182-L200","kind":"class","name":"GatherLayer","path":"lavis/models/base_model.py","language":"python","start_line":182,"end_line":200,"context_start_line":162,"context_end_line":220,"code":" @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)\n return all_gradients[torch.distributed.get_rank()]\n\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)\n tensor_all = GatherLayer.apply(tensors)\n\n return torch.cat(tensor_all, dim=0)\n\n\n@torch.no_grad()","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.all_gather_with_grad","uri":"program://CREMA/function/lavis.models.base_model.all_gather_with_grad#L203-L217","kind":"function","name":"all_gather_with_grad","path":"lavis/models/base_model.py","language":"python","start_line":203,"end_line":217,"context_start_line":183,"context_end_line":237,"code":" \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)\n return all_gradients[torch.distributed.get_rank()]\n\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)\n tensor_all = GatherLayer.apply(tensors)\n\n return torch.cat(tensor_all, dim=0)\n\n\n@torch.no_grad()\ndef concat_all_gather(tensor):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n *** Warning ***: torch.distributed.all_gather has no gradient.\n \"\"\"\n # if use distributed training\n if not is_dist_avail_and_initialized():\n return tensor\n\n tensors_gather = [\n torch.ones_like(tensor) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(tensors_gather, tensor, async_op=False)\n\n output = torch.cat(tensors_gather, dim=0)\n return output\n","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.concat_all_gather","uri":"program://CREMA/function/lavis.models.base_model.concat_all_gather#L221-L236","kind":"function","name":"concat_all_gather","path":"lavis/models/base_model.py","language":"python","start_line":221,"end_line":236,"context_start_line":201,"context_end_line":247,"code":"\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)\n tensor_all = GatherLayer.apply(tensors)\n\n return torch.cat(tensor_all, dim=0)\n\n\n@torch.no_grad()\ndef concat_all_gather(tensor):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n *** Warning ***: torch.distributed.all_gather has no gradient.\n \"\"\"\n # if use distributed training\n if not is_dist_avail_and_initialized():\n return tensor\n\n tensors_gather = [\n torch.ones_like(tensor) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(tensors_gather, tensor, async_op=False)\n\n output = torch.cat(tensors_gather, dim=0)\n return output\n\n\ndef tile(x, dim, n_tile):\n init_dim = x.size(dim)\n repeat_idx = [1] * x.dim()\n repeat_idx[dim] = n_tile\n x = x.repeat(*(repeat_idx))\n order_index = torch.LongTensor(\n np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)])\n )\n return torch.index_select(x, dim, order_index.to(x.device))","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.tile","uri":"program://CREMA/function/lavis.models.base_model.tile#L239-L247","kind":"function","name":"tile","path":"lavis/models/base_model.py","language":"python","start_line":239,"end_line":247,"context_start_line":219,"context_end_line":247,"code":"\n@torch.no_grad()\ndef concat_all_gather(tensor):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n *** Warning ***: torch.distributed.all_gather has no gradient.\n \"\"\"\n # if use distributed training\n if not is_dist_avail_and_initialized():\n return tensor\n\n tensors_gather = [\n torch.ones_like(tensor) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(tensors_gather, tensor, async_op=False)\n\n output = torch.cat(tensors_gather, dim=0)\n return output\n\n\ndef tile(x, dim, n_tile):\n init_dim = x.size(dim)\n repeat_idx = [1] * x.dim()\n repeat_idx[dim] = n_tile\n x = x.repeat(*(repeat_idx))\n order_index = torch.LongTensor(\n np.concatenate([init_dim * np.arange(n_tile) + i for i in range(init_dim)])\n )\n return torch.index_select(x, dim, order_index.to(x.device))","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.__init__","uri":"program://CREMA/function/lavis.models.base_model.__init__#L126-L127","kind":"function","name":"__init__","path":"lavis/models/base_model.py","language":"python","start_line":126,"end_line":127,"context_start_line":106,"context_end_line":147,"code":" tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.device","uri":"program://CREMA/function/lavis.models.base_model.device#L133-L134","kind":"function","name":"device","path":"lavis/models/base_model.py","language":"python","start_line":133,"end_line":134,"context_start_line":113,"context_end_line":154,"code":" if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.load_checkpoint","uri":"program://CREMA/function/lavis.models.base_model.load_checkpoint#L29-L56","kind":"function","name":"load_checkpoint","path":"lavis/models/base_model.py","language":"python","start_line":29,"end_line":56,"context_start_line":9,"context_end_line":76,"code":"import os\n\nimport numpy as np\nimport torch\nimport torch.nn as nn\nfrom lavis.common.dist_utils import download_cached_file, is_dist_avail_and_initialized\nfrom lavis.common.utils import get_abs_path, is_url\nfrom omegaconf import OmegaConf\n\n\nclass BaseModel(nn.Module):\n \"\"\"Base class for models.\"\"\"\n\n def __init__(self):\n super().__init__()\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n def load_checkpoint(self, url_or_filename):\n \"\"\"\n Load from a finetuned checkpoint.\n\n This should expect no mismatch in the model keys and the checkpoint keys.\n \"\"\"\n\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint.keys():\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n\n @classmethod\n def from_pretrained(cls, model_type):\n \"\"\"\n Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.from_pretrained","uri":"program://CREMA/function/lavis.models.base_model.from_pretrained#L59-L72","kind":"function","name":"from_pretrained","path":"lavis/models/base_model.py","language":"python","start_line":59,"end_line":72,"context_start_line":39,"context_end_line":92,"code":" )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint.keys():\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n\n @classmethod\n def from_pretrained(cls, model_type):\n \"\"\"\n Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}\".format(model_type)\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n def load_checkpoint_from_config(self, cfg, **kwargs):\n \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.default_config_path","uri":"program://CREMA/function/lavis.models.base_model.default_config_path#L75-L79","kind":"function","name":"default_config_path","path":"lavis/models/base_model.py","language":"python","start_line":75,"end_line":79,"context_start_line":55,"context_end_line":99,"code":"\n return msg\n\n @classmethod\n def from_pretrained(cls, model_type):\n \"\"\"\n Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}\".format(model_type)\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n def load_checkpoint_from_config(self, cfg, **kwargs):\n \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.load_checkpoint_from_config","uri":"program://CREMA/function/lavis.models.base_model.load_checkpoint_from_config#L81-L100","kind":"function","name":"load_checkpoint_from_config","path":"lavis/models/base_model.py","language":"python","start_line":81,"end_line":100,"context_start_line":61,"context_end_line":120,"code":" Build a pretrained model from default configuration file, specified by model_type.\n\n Args:\n - model_type (str): model type, specifying architecture and checkpoints.\n\n Returns:\n - model (nn.Module): pretrained or finetuned model, depending on the configuration.\n \"\"\"\n model_cfg = OmegaConf.load(cls.default_config_path(model_type)).model\n model = cls.from_config(model_cfg)\n\n return model\n\n @classmethod\n def default_config_path(cls, model_type):\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}\".format(model_type)\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n def load_checkpoint_from_config(self, cfg, **kwargs):\n \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"\n self.load_from_pretrained(url_or_filename=pretrain_path, **kwargs)\n\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.before_evaluation","uri":"program://CREMA/function/lavis.models.base_model.before_evaluation#L102-L103","kind":"function","name":"before_evaluation","path":"lavis/models/base_model.py","language":"python","start_line":102,"end_line":103,"context_start_line":82,"context_end_line":123,"code":" \"\"\"\n Load checkpoint as specified in the config file.\n\n If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"\n self.load_from_pretrained(url_or_filename=pretrain_path, **kwargs)\n\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.show_n_params","uri":"program://CREMA/function/lavis.models.base_model.show_n_params#L105-L118","kind":"function","name":"show_n_params","path":"lavis/models/base_model.py","language":"python","start_line":105,"end_line":118,"context_start_line":85,"context_end_line":138,"code":" If load_finetuned is True, load the finetuned model; otherwise, load the pretrained model.\n When loading the pretrained model, each task-specific architecture may define their\n own load_from_pretrained() method.\n \"\"\"\n load_finetuned = cfg.get(\"load_finetuned\", True)\n if load_finetuned:\n finetune_path = cfg.get(\"finetuned\", None)\n assert (\n finetune_path is not None\n ), \"Found load_finetuned is True, but finetune_path is None.\"\n self.load_checkpoint(url_or_filename=finetune_path)\n else:\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n assert \"Found load_finetuned is False, but pretrain_path is None.\"\n self.load_from_pretrained(url_or_filename=pretrain_path, **kwargs)\n\n def before_evaluation(self, **kwargs):\n pass\n\n def show_n_params(self, return_str=True):\n tot = 0\n for p in self.parameters():\n w = 1\n for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.forward_features","uri":"program://CREMA/function/lavis.models.base_model.forward_features#L129-L130","kind":"function","name":"forward_features","path":"lavis/models/base_model.py","language":"python","start_line":129,"end_line":130,"context_start_line":109,"context_end_line":150,"code":" for x in p.shape:\n w *= x\n tot += w\n if return_str:\n if tot >= 1e6:\n return \"{:.1f}M\".format(tot / 1e6)\n else:\n return \"{:.1f}K\".format(tot / 1e3)\n else:\n return tot\n\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model._dequeue_and_enqueue","uri":"program://CREMA/function/lavis.models.base_model._dequeue_and_enqueue#L139-L158","kind":"function","name":"_dequeue_and_enqueue","path":"lavis/models/base_model.py","language":"python","start_line":139,"end_line":158,"context_start_line":119,"context_end_line":178,"code":"\n\nclass BaseEncoder(nn.Module):\n \"\"\"\n Base class for primitive encoders, such as ViT, TimeSformer, etc.\n \"\"\"\n\n def __init__(self):\n super().__init__()\n\n def forward_features(self, samples, **kwargs):\n raise NotImplementedError\n\n @property\n def device(self):\n return list(self.parameters())[0].device\n\n\nclass SharedQueueMixin:\n @torch.no_grad()\n def _dequeue_and_enqueue(self, image_feat, text_feat, idxs=None):\n # gather keys before updating queue\n image_feats = concat_all_gather(image_feat)\n text_feats = concat_all_gather(text_feat)\n\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.copy_params","uri":"program://CREMA/function/lavis.models.base_model.copy_params#L163-L169","kind":"function","name":"copy_params","path":"lavis/models/base_model.py","language":"python","start_line":163,"end_line":169,"context_start_line":143,"context_end_line":189,"code":"\n batch_size = image_feats.shape[0]\n\n ptr = int(self.queue_ptr)\n assert self.queue_size % batch_size == 0 # for simplicity\n\n # replace the keys at ptr (dequeue and enqueue)\n self.image_queue[:, ptr : ptr + batch_size] = image_feats.T\n self.text_queue[:, ptr : ptr + batch_size] = text_feats.T\n\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model._momentum_update","uri":"program://CREMA/function/lavis.models.base_model._momentum_update#L172-L179","kind":"function","name":"_momentum_update","path":"lavis/models/base_model.py","language":"python","start_line":172,"end_line":179,"context_start_line":152,"context_end_line":199,"code":"\n if idxs is not None:\n idxs = concat_all_gather(idxs)\n self.idx_queue[:, ptr : ptr + batch_size] = idxs.T\n\n ptr = (ptr + batch_size) % self.queue_size # move pointer\n self.queue_ptr[0] = ptr\n\n\nclass MomentumDistilationMixin:\n @torch.no_grad()\n def copy_params(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data.copy_(param.data) # initialize\n param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.forward","uri":"program://CREMA/function/lavis.models.base_model.forward#L189-L194","kind":"function","name":"forward","path":"lavis/models/base_model.py","language":"python","start_line":189,"end_line":194,"context_start_line":169,"context_end_line":214,"code":" param_m.requires_grad = False # not update by gradient\n\n @torch.no_grad()\n def _momentum_update(self):\n for model_pair in self.model_pairs:\n for param, param_m in zip(\n model_pair[0].parameters(), model_pair[1].parameters()\n ):\n param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)\n return all_gradients[torch.distributed.get_rank()]\n\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.base_model.backward","uri":"program://CREMA/function/lavis.models.base_model.backward#L197-L200","kind":"function","name":"backward","path":"lavis/models/base_model.py","language":"python","start_line":197,"end_line":200,"context_start_line":177,"context_end_line":220,"code":" param_m.data = param_m.data * self.momentum + param.data * (\n 1.0 - self.momentum\n )\n\n\nclass GatherLayer(torch.autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [\n torch.zeros_like(x) for _ in range(torch.distributed.get_world_size())\n ]\n torch.distributed.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n torch.distributed.all_reduce(all_gradients)\n return all_gradients[torch.distributed.get_rank()]\n\n\ndef all_gather_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = torch.distributed.get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n\n # tensor_all = GatherLayer.apply(tensors)\n tensor_all = GatherLayer.apply(tensors)\n\n return torch.cat(tensor_all, dim=0)\n\n\n@torch.no_grad()","source_hash":"0671eb3815f0c91f9d61649cc647db88b34df75381585cea18596a09dcca224c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk","uri":"program://CREMA/module/lavis.models.topk#L1-L339","kind":"module","name":"lavis.models.topk","path":"lavis/models/topk.py","language":"python","start_line":1,"end_line":339,"context_start_line":1,"context_end_line":339,"code":"# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nDETR model and criterion classes.\n\"\"\"\n\nimport math\nimport torch\nimport copy\nimport einops\nimport torch.nn.functional as F\nfrom torch import nn\n\nfrom dataclasses import dataclass\nfrom typing import Optional\nfrom enum import IntEnum\nfrom einops import rearrange\n\nclass PerturbedTopK(nn.Module):\n def __init__(self, k: int, num_samples: int = 1000):\n super(PerturbedTopK, self).__init__()\n self.num_samples = num_samples\n self.k = k\n\n def __call__(self, x, sigma):\n return PerturbedTopKFunction.apply(x, self.k, self.num_samples, sigma)\n\n\nclass PerturbedTopKFunction(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, k: int, num_samples: int = 1000, sigma: float = 0.05):\n #print('x', x.shape)\n b, d = x.shape\n # for Gaussian: noise and gradient are the same.\n noise = torch.normal(mean=0.0, std=1.0, size=(b, num_samples, d)).to(x.device)\n perturbed_x = x[:, None, :] + noise * sigma # b, nS, d\n #print('perturbed_x', perturbed_x.shape)\n topk_results = torch.topk(perturbed_x, k=k, dim=-1, sorted=False)\n #print('topk_results',topk_results)\n\n indices = topk_results.indices # b, nS, k\n indices = torch.sort(indices, dim=-1).values # b, nS, k\n # print('indices', indices.shape ,indices[0,0,0])\n\n perturbed_output = torch.nn.functional.one_hot(indices, num_classes=d).float()\n indicators = perturbed_output.mean(dim=1) # b, k, d\n # print('perturbed_output', perturbed_output.shape, perturbed_output[0,indices[0,0,0],0,0])\n\n # constants for backward\n ctx.k = k\n ctx.num_samples = num_samples\n ctx.sigma = sigma\n\n # tensors for backward\n ctx.perturbed_output = perturbed_output\n ctx.noise = noise\n return indicators\n\n @staticmethod\n def backward(ctx, grad_output):\n if grad_output is None:\n return tuple([None] * 5)\n\n noise_gradient = ctx.noise\n if ctx.sigma <= 1e-20:\n b, _, k, d = ctx.perturbed_output.size()\n expected_gradient = torch.zeros(b, k, d).to(grad_output.device)\n else:\n expected_gradient = (\n torch.einsum(\"bnkd,bnd->bkd\", ctx.perturbed_output, noise_gradient)\n / ctx.num_samples\n / (ctx.sigma)\n )\n\n grad_input = torch.einsum(\"bkd,bkd->bd\", grad_output, expected_gradient)\n\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1\n TEXT_UNUSED = 2 # ignore\n VISUAL_EMBEDDING = 3\n VISUAL_UNUSED = 4 # ignore\n\nclass ModalityEmbeddings(nn.Module):\n \"\"\"\n Provides embeddings that indicate type of modality; for use with multimodal inputs for ATP. See atp.py for usage.\n \"\"\"\n def __init__(self,\n d_model: int,\n use_text_query: bool = False,\n use_text_cands: bool = False,\n n_cands: int = 5):\n \"\"\"\n Details for each of these arguments are provided in ATPConfig.\n \"\"\"\n super().__init__()\n self.d_model = d_model\n self.embedding = nn.Embedding(num_embeddings=len(ModalityEmbeddingsID),\n embedding_dim=d_model)\n\n self.use_text_query = use_text_query\n self.use_text_cands = use_text_cands\n self.n_cands = n_cands if use_text_cands else 0\n self.n_text_feats = 1 if use_text_query else 0\n if use_text_cands:\n self.n_text_feats += n_cands\n\n def forward(self, x, num_frame):\n \"\"\"\n x: torch.tensor of size (L, N, D)\n returns modality embeddings for x of size (L, *, D)\n \"\"\"\n L, N, D = x.size() # (sequence_length, batch_size, feature_dim)\n num_txt = L - num_frame\n \n # assemble the IDs for the modality encodings, language inputs then vision inputs\n class_ids = []\n if self.use_text_query:\n class_ids.extend([ModalityEmbeddingsID.TEXT_QUESTION,] * num_txt)\n # if self.use_text_cands:\n # class_ids.extend([ModalityEmbeddingsID.TEXT_EMBEDDING,] * self.n_cands)\n class_ids.extend([ModalityEmbeddingsID.VISUAL_EMBEDDING,] * num_frame)\n \n class_ids = torch.tensor(\n class_ids,\n dtype=torch.long,\n device=x.device\n ).unsqueeze(-1)\n \n # return modality embeddings\n return self.embedding(class_ids)\n\n@dataclass\nclass ATPConfig:\n '''\n ATPConfig contains the parameters needed for the ATPSelectorModel (and its ATPEncoder).\n '''\n # ATPEncoder params\n n_layers: int = 6\n n_heads: int = 4\n d_model: int = 256\n d_input_t: int = 2048\n d_input_v: int = 1408\n d_model_ff: int = 256\n enc_dropout: float = 0.1\n use_text_query: bool = True # at least one use_text_* needs to be true for ATP to be multimodal\n use_text_cands: bool = False # ^ see above. (note: if both are false, ATP is vision-only)\n n_cands: int = 5 # only relevant when use_text_cands is set to true\n # ATPSelector params\n use_ste: bool = True # controls type of selector during ATP training; see ATPSelectorModel.forward\n sel_dropout: float = 0.0\n d_input: int = 512 # size of the input vision-language embeddings (e.g. CLIP-ViT-B32 is size 512)\n \n def default_args(cls):\n return cls(n_layers = 6,\n n_heads = 4,\n d_model = 256,\n d_input_t = 2048,\n d_input_v = 1408,\n d_model_ff = 256,\n enc_dropout = 0.1,\n use_text_query = True,\n use_text_cands = False,\n n_cands = 5,\n use_ste = True,\n sel_dropout = 0.0,\n d_input = 512)\n\n @classmethod\n def from_args(cls, args):\n return cls(n_layers = args.n_layers,\n n_heads = args.n_heads,\n d_model = args.d_model,\n d_model_ff = args.d_model_ff,\n enc_dropout = args.enc_dropout,\n use_text_query = args.use_text_query,\n use_text_cands = args.use_text_cands,\n n_cands = args.n_cands,\n use_ste = args.use_ste,\n sel_dropout = args.sel_dropout,\n d_input = args.d_input)\n\nclass ATPEncoder(nn.Module):\n \"\"\"\n The multimodal transformer encoder for the ATP model. For analysis purposes, the ATP encoder\n does not use any positional information (no positional encodings + transformer / self-attention)\n and is generally kept low-capacity. If the goal is raw accuracy (not analysis), you can relax these constraints.\n \"\"\"\n def __init__(self, config: ATPConfig):\n \"\"\"\n config: ATPConfig with parameters for the (transformer-based, atemporal) encoder for ATP.\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.d_model = config.d_model\n\n self.dropout = nn.Dropout(p=config.enc_dropout)\n\n\n self.modality_encoding = ModalityEmbeddings(d_model=self.d_model,\n use_text_query=config.use_text_query,\n use_text_cands=config.use_text_cands,\n n_cands=config.n_cands)\n \n atp_encoder_layer = nn.TransformerEncoderLayer(\n d_model=self.d_model,\n nhead=config.n_heads,\n dim_feedforward=config.d_model_ff,\n dropout=config.enc_dropout,\n activation='relu'\n )\n\n self.transformer_encoder = nn.TransformerEncoder(atp_encoder_layer, config.n_layers)\n\n def forward(self, x_inputs: torch.tensor, vis_L):\n \"\"\"\n x_inputs: torch.tensor of shape (L, N, D)\n \"\"\"\n L, N, D = x_inputs.size() # (sequence_length, batch_size, d_model)\n assert D == self.d_model, \"inputs dimension mismatch\"\n x_encoded = x_inputs * math.sqrt(self.d_model)\n x_encoded += self.modality_encoding(x_encoded, vis_L)\n x_encoded = self.dropout(x_encoded)\n x_encoded = self.transformer_encoder(x_encoded)\n\n return x_encoded\n\nclass TopK_Selector(nn.Module):\n \"\"\"\n The Atemporal Probe (ATP) selector model. Takes as input a sequence of image-language \n encoding and outputs a (discrete) selection over the input frames, to help analyze \n downstream discriminative video-language tasks.\n \"\"\"\n \n def __init__(self, config=ATPConfig, num_select=4):\n \"\"\"\n config: ATPConfig with parameters for initializing the ATPSelectorModel (and its encoder).\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.config = config\n self.t_embedding = nn.Linear(config.d_input_t, config.d_input)\n self.v_embedding = nn.Linear(config.d_input_v, config.d_input)\n self.embedding = nn.Linear(config.d_input, config.d_model)\n self.atp_encoder = ATPEncoder(config)\n self.dropout = nn.Dropout(p=config.sel_dropout)\n self.logits = nn.Linear(config.d_model, 1)\n self.num_select = num_select\n self.sigma = 0.1\n\n def forward(self,\n x_vis, # [b, t, d]\n x_txt, # [b, n, d]\n **kwargs):\n \"\"\"\n \"\"\"\n x_vis_cls = x_vis[:, :, 0, :] # b t n c\n N, vis_L, D = x_vis_cls.size() # (batch_size, sequence_length, feature_dimension)\n # embed the input sequence to the (smaller) model dimension (d_model) with modality encodings.\n x_vis_cls = self.v_embedding(self.dropout(x_vis_cls))\n x_txt = self.t_embedding(self.dropout(x_txt))\n x_inputs = []\n x_vis_cls = x_vis_cls.permute(1, 0, 2)\n x_inputs.append(x_txt.permute(1,0,2)) # (n, b, d)\n x_inputs.append(x_vis_cls)\n x_inputs = torch.cat(x_inputs, dim=0)\n x_encoded = self.embedding(self.dropout(x_inputs))\n x_atp_encoded = self.atp_encoder(x_encoded, vis_L)\n x_atp_encoded = x_atp_encoded.permute(1, 0, 2)\n x_encoded_v = x_atp_encoded[:, -vis_L: , :]\n # obtain selection scores (logits)\n x_logits = self.logits(self.dropout(x_encoded_v)).squeeze()\n #print('x_logits', x_logits.shape)\n\n if self.training:\n indices = PerturbedTopKFunction.apply(x_logits, self.num_select)\n #print('indices', indices.shape)\n indices = einops.rearrange(indices, \"b k d -> b d k\")\n\n if indices is not None:\n qa_frames = extract_frames_from_indicators(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n else:\n indices = HardTopK(self.num_select, x_logits)\n if indices is not None:\n qa_frames = extract_frames_from_indices(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n\n\n return qa_frames\n\nif __name__ == \"__main__\":\n selector_config = ATPConfig.default_args\n\n Selector = TopK_Selector(num_select=4) #.eval()\n\n x_vis = torch.rand([2, 8, 257, 1408])\n x_txt = torch.rand([2, 68, 2048])\n\n out = Selector(x_vis, x_txt)\n print(out.shape)\n\n","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.PerturbedTopK","uri":"program://CREMA/class/lavis.models.topk.PerturbedTopK#L18-L25","kind":"class","name":"PerturbedTopK","path":"lavis/models/topk.py","language":"python","start_line":18,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved\n\"\"\"\nDETR model and criterion classes.\n\"\"\"\n\nimport math\nimport torch\nimport copy\nimport einops\nimport torch.nn.functional as F\nfrom torch import nn\n\nfrom dataclasses import dataclass\nfrom typing import Optional\nfrom enum import IntEnum\nfrom einops import rearrange\n\nclass PerturbedTopK(nn.Module):\n def __init__(self, k: int, num_samples: int = 1000):\n super(PerturbedTopK, self).__init__()\n self.num_samples = num_samples\n self.k = k\n\n def __call__(self, x, sigma):\n return PerturbedTopKFunction.apply(x, self.k, self.num_samples, sigma)\n\n\nclass PerturbedTopKFunction(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, k: int, num_samples: int = 1000, sigma: float = 0.05):\n #print('x', x.shape)\n b, d = x.shape\n # for Gaussian: noise and gradient are the same.\n noise = torch.normal(mean=0.0, std=1.0, size=(b, num_samples, d)).to(x.device)\n perturbed_x = x[:, None, :] + noise * sigma # b, nS, d\n #print('perturbed_x', perturbed_x.shape)\n topk_results = torch.topk(perturbed_x, k=k, dim=-1, sorted=False)\n #print('topk_results',topk_results)\n\n indices = topk_results.indices # b, nS, k\n indices = torch.sort(indices, dim=-1).values # b, nS, k\n # print('indices', indices.shape ,indices[0,0,0])\n\n perturbed_output = torch.nn.functional.one_hot(indices, num_classes=d).float()\n indicators = perturbed_output.mean(dim=1) # b, k, d","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.PerturbedTopKFunction","uri":"program://CREMA/class/lavis.models.topk.PerturbedTopKFunction#L28-L76","kind":"class","name":"PerturbedTopKFunction","path":"lavis/models/topk.py","language":"python","start_line":28,"end_line":76,"context_start_line":8,"context_end_line":96,"code":"import copy\nimport einops\nimport torch.nn.functional as F\nfrom torch import nn\n\nfrom dataclasses import dataclass\nfrom typing import Optional\nfrom enum import IntEnum\nfrom einops import rearrange\n\nclass PerturbedTopK(nn.Module):\n def __init__(self, k: int, num_samples: int = 1000):\n super(PerturbedTopK, self).__init__()\n self.num_samples = num_samples\n self.k = k\n\n def __call__(self, x, sigma):\n return PerturbedTopKFunction.apply(x, self.k, self.num_samples, sigma)\n\n\nclass PerturbedTopKFunction(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, k: int, num_samples: int = 1000, sigma: float = 0.05):\n #print('x', x.shape)\n b, d = x.shape\n # for Gaussian: noise and gradient are the same.\n noise = torch.normal(mean=0.0, std=1.0, size=(b, num_samples, d)).to(x.device)\n perturbed_x = x[:, None, :] + noise * sigma # b, nS, d\n #print('perturbed_x', perturbed_x.shape)\n topk_results = torch.topk(perturbed_x, k=k, dim=-1, sorted=False)\n #print('topk_results',topk_results)\n\n indices = topk_results.indices # b, nS, k\n indices = torch.sort(indices, dim=-1).values # b, nS, k\n # print('indices', indices.shape ,indices[0,0,0])\n\n perturbed_output = torch.nn.functional.one_hot(indices, num_classes=d).float()\n indicators = perturbed_output.mean(dim=1) # b, k, d\n # print('perturbed_output', perturbed_output.shape, perturbed_output[0,indices[0,0,0],0,0])\n\n # constants for backward\n ctx.k = k\n ctx.num_samples = num_samples\n ctx.sigma = sigma\n\n # tensors for backward\n ctx.perturbed_output = perturbed_output\n ctx.noise = noise\n return indicators\n\n @staticmethod\n def backward(ctx, grad_output):\n if grad_output is None:\n return tuple([None] * 5)\n\n noise_gradient = ctx.noise\n if ctx.sigma <= 1e-20:\n b, _, k, d = ctx.perturbed_output.size()\n expected_gradient = torch.zeros(b, k, d).to(grad_output.device)\n else:\n expected_gradient = (\n torch.einsum(\"bnkd,bnd->bkd\", ctx.perturbed_output, noise_gradient)\n / ctx.num_samples\n / (ctx.sigma)\n )\n\n grad_input = torch.einsum(\"bkd,bkd->bd\", grad_output, expected_gradient)\n\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.HardTopK","uri":"program://CREMA/function/lavis.models.topk.HardTopK#L78-L82","kind":"function","name":"HardTopK","path":"lavis/models/topk.py","language":"python","start_line":78,"end_line":82,"context_start_line":58,"context_end_line":102,"code":" @staticmethod\n def backward(ctx, grad_output):\n if grad_output is None:\n return tuple([None] * 5)\n\n noise_gradient = ctx.noise\n if ctx.sigma <= 1e-20:\n b, _, k, d = ctx.perturbed_output.size()\n expected_gradient = torch.zeros(b, k, d).to(grad_output.device)\n else:\n expected_gradient = (\n torch.einsum(\"bnkd,bnd->bkd\", ctx.perturbed_output, noise_gradient)\n / ctx.num_samples\n / (ctx.sigma)\n )\n\n grad_input = torch.einsum(\"bkd,bkd->bd\", grad_output, expected_gradient)\n\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.batched_index_select","uri":"program://CREMA/function/lavis.models.topk.batched_index_select#L85-L93","kind":"function","name":"batched_index_select","path":"lavis/models/topk.py","language":"python","start_line":85,"end_line":93,"context_start_line":65,"context_end_line":113,"code":" b, _, k, d = ctx.perturbed_output.size()\n expected_gradient = torch.zeros(b, k, d).to(grad_output.device)\n else:\n expected_gradient = (\n torch.einsum(\"bnkd,bnd->bkd\", ctx.perturbed_output, noise_gradient)\n / ctx.num_samples\n / (ctx.sigma)\n )\n\n grad_input = torch.einsum(\"bkd,bkd->bd\", grad_output, expected_gradient)\n\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.extract_frames_from_indices","uri":"program://CREMA/function/lavis.models.topk.extract_frames_from_indices#L95-L101","kind":"function","name":"extract_frames_from_indices","path":"lavis/models/topk.py","language":"python","start_line":95,"end_line":101,"context_start_line":75,"context_end_line":121,"code":"\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1\n TEXT_UNUSED = 2 # ignore\n VISUAL_EMBEDDING = 3\n VISUAL_UNUSED = 4 # ignore\n\nclass ModalityEmbeddings(nn.Module):\n \"\"\"\n Provides embeddings that indicate type of modality; for use with multimodal inputs for ATP. See atp.py for usage.\n \"\"\"","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.extract_frames_from_indicators","uri":"program://CREMA/function/lavis.models.topk.extract_frames_from_indicators#L104-L108","kind":"function","name":"extract_frames_from_indicators","path":"lavis/models/topk.py","language":"python","start_line":104,"end_line":108,"context_start_line":84,"context_end_line":128,"code":"\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1\n TEXT_UNUSED = 2 # ignore\n VISUAL_EMBEDDING = 3\n VISUAL_UNUSED = 4 # ignore\n\nclass ModalityEmbeddings(nn.Module):\n \"\"\"\n Provides embeddings that indicate type of modality; for use with multimodal inputs for ATP. See atp.py for usage.\n \"\"\"\n def __init__(self,\n d_model: int,\n use_text_query: bool = False,\n use_text_cands: bool = False,\n n_cands: int = 5):\n \"\"\"\n Details for each of these arguments are provided in ATPConfig.","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.ModalityEmbeddingsID","uri":"program://CREMA/class/lavis.models.topk.ModalityEmbeddingsID#L111-L116","kind":"class","name":"ModalityEmbeddingsID","path":"lavis/models/topk.py","language":"python","start_line":111,"end_line":116,"context_start_line":91,"context_end_line":136,"code":" expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape\n k = indices.shape[-1]\n all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1\n TEXT_UNUSED = 2 # ignore\n VISUAL_EMBEDDING = 3\n VISUAL_UNUSED = 4 # ignore\n\nclass ModalityEmbeddings(nn.Module):\n \"\"\"\n Provides embeddings that indicate type of modality; for use with multimodal inputs for ATP. See atp.py for usage.\n \"\"\"\n def __init__(self,\n d_model: int,\n use_text_query: bool = False,\n use_text_cands: bool = False,\n n_cands: int = 5):\n \"\"\"\n Details for each of these arguments are provided in ATPConfig.\n \"\"\"\n super().__init__()\n self.d_model = d_model\n self.embedding = nn.Embedding(num_embeddings=len(ModalityEmbeddingsID),\n embedding_dim=d_model)\n\n self.use_text_query = use_text_query\n self.use_text_cands = use_text_cands","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.ModalityEmbeddings","uri":"program://CREMA/class/lavis.models.topk.ModalityEmbeddings#L118-L165","kind":"class","name":"ModalityEmbeddings","path":"lavis/models/topk.py","language":"python","start_line":118,"end_line":165,"context_start_line":98,"context_end_line":185,"code":" all_frame = x\n frames = batched_index_select(all_frame, 1, indices)\n frames = frames.contiguous().view(batch_size, k, n, channels)\n return frames\n\n\ndef extract_frames_from_indicators(x, indicators):\n indicators = rearrange(indicators, \"b d k -> b k d\")\n frames = torch.einsum(\"b k d, b d n c-> b k n c\",\n indicators, x)\n return frames\n\n\nclass ModalityEmbeddingsID(IntEnum):\n TEXT_QUESTION = 0\n TEXT_EMBEDDING = 1\n TEXT_UNUSED = 2 # ignore\n VISUAL_EMBEDDING = 3\n VISUAL_UNUSED = 4 # ignore\n\nclass ModalityEmbeddings(nn.Module):\n \"\"\"\n Provides embeddings that indicate type of modality; for use with multimodal inputs for ATP. See atp.py for usage.\n \"\"\"\n def __init__(self,\n d_model: int,\n use_text_query: bool = False,\n use_text_cands: bool = False,\n n_cands: int = 5):\n \"\"\"\n Details for each of these arguments are provided in ATPConfig.\n \"\"\"\n super().__init__()\n self.d_model = d_model\n self.embedding = nn.Embedding(num_embeddings=len(ModalityEmbeddingsID),\n embedding_dim=d_model)\n\n self.use_text_query = use_text_query\n self.use_text_cands = use_text_cands\n self.n_cands = n_cands if use_text_cands else 0\n self.n_text_feats = 1 if use_text_query else 0\n if use_text_cands:\n self.n_text_feats += n_cands\n\n def forward(self, x, num_frame):\n \"\"\"\n x: torch.tensor of size (L, N, D)\n returns modality embeddings for x of size (L, *, D)\n \"\"\"\n L, N, D = x.size() # (sequence_length, batch_size, feature_dim)\n num_txt = L - num_frame\n \n # assemble the IDs for the modality encodings, language inputs then vision inputs\n class_ids = []\n if self.use_text_query:\n class_ids.extend([ModalityEmbeddingsID.TEXT_QUESTION,] * num_txt)\n # if self.use_text_cands:\n # class_ids.extend([ModalityEmbeddingsID.TEXT_EMBEDDING,] * self.n_cands)\n class_ids.extend([ModalityEmbeddingsID.VISUAL_EMBEDDING,] * num_frame)\n \n class_ids = torch.tensor(\n class_ids,\n dtype=torch.long,\n device=x.device\n ).unsqueeze(-1)\n \n # return modality embeddings\n return self.embedding(class_ids)\n\n@dataclass\nclass ATPConfig:\n '''\n ATPConfig contains the parameters needed for the ATPSelectorModel (and its ATPEncoder).\n '''\n # ATPEncoder params\n n_layers: int = 6\n n_heads: int = 4\n d_model: int = 256\n d_input_t: int = 2048\n d_input_v: int = 1408\n d_model_ff: int = 256\n enc_dropout: float = 0.1\n use_text_query: bool = True # at least one use_text_* needs to be true for ATP to be multimodal\n use_text_cands: bool = False # ^ see above. (note: if both are false, ATP is vision-only)\n n_cands: int = 5 # only relevant when use_text_cands is set to true\n # ATPSelector params\n use_ste: bool = True # controls type of selector during ATP training; see ATPSelectorModel.forward\n sel_dropout: float = 0.0","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.ATPConfig","uri":"program://CREMA/class/lavis.models.topk.ATPConfig#L168-L215","kind":"class","name":"ATPConfig","path":"lavis/models/topk.py","language":"python","start_line":168,"end_line":215,"context_start_line":148,"context_end_line":235,"code":" num_txt = L - num_frame\n \n # assemble the IDs for the modality encodings, language inputs then vision inputs\n class_ids = []\n if self.use_text_query:\n class_ids.extend([ModalityEmbeddingsID.TEXT_QUESTION,] * num_txt)\n # if self.use_text_cands:\n # class_ids.extend([ModalityEmbeddingsID.TEXT_EMBEDDING,] * self.n_cands)\n class_ids.extend([ModalityEmbeddingsID.VISUAL_EMBEDDING,] * num_frame)\n \n class_ids = torch.tensor(\n class_ids,\n dtype=torch.long,\n device=x.device\n ).unsqueeze(-1)\n \n # return modality embeddings\n return self.embedding(class_ids)\n\n@dataclass\nclass ATPConfig:\n '''\n ATPConfig contains the parameters needed for the ATPSelectorModel (and its ATPEncoder).\n '''\n # ATPEncoder params\n n_layers: int = 6\n n_heads: int = 4\n d_model: int = 256\n d_input_t: int = 2048\n d_input_v: int = 1408\n d_model_ff: int = 256\n enc_dropout: float = 0.1\n use_text_query: bool = True # at least one use_text_* needs to be true for ATP to be multimodal\n use_text_cands: bool = False # ^ see above. (note: if both are false, ATP is vision-only)\n n_cands: int = 5 # only relevant when use_text_cands is set to true\n # ATPSelector params\n use_ste: bool = True # controls type of selector during ATP training; see ATPSelectorModel.forward\n sel_dropout: float = 0.0\n d_input: int = 512 # size of the input vision-language embeddings (e.g. CLIP-ViT-B32 is size 512)\n \n def default_args(cls):\n return cls(n_layers = 6,\n n_heads = 4,\n d_model = 256,\n d_input_t = 2048,\n d_input_v = 1408,\n d_model_ff = 256,\n enc_dropout = 0.1,\n use_text_query = True,\n use_text_cands = False,\n n_cands = 5,\n use_ste = True,\n sel_dropout = 0.0,\n d_input = 512)\n\n @classmethod\n def from_args(cls, args):\n return cls(n_layers = args.n_layers,\n n_heads = args.n_heads,\n d_model = args.d_model,\n d_model_ff = args.d_model_ff,\n enc_dropout = args.enc_dropout,\n use_text_query = args.use_text_query,\n use_text_cands = args.use_text_cands,\n n_cands = args.n_cands,\n use_ste = args.use_ste,\n sel_dropout = args.sel_dropout,\n d_input = args.d_input)\n\nclass ATPEncoder(nn.Module):\n \"\"\"\n The multimodal transformer encoder for the ATP model. For analysis purposes, the ATP encoder\n does not use any positional information (no positional encodings + transformer / self-attention)\n and is generally kept low-capacity. If the goal is raw accuracy (not analysis), you can relax these constraints.\n \"\"\"\n def __init__(self, config: ATPConfig):\n \"\"\"\n config: ATPConfig with parameters for the (transformer-based, atemporal) encoder for ATP.\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.d_model = config.d_model\n\n self.dropout = nn.Dropout(p=config.enc_dropout)\n\n\n self.modality_encoding = ModalityEmbeddings(d_model=self.d_model,\n use_text_query=config.use_text_query,","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.ATPEncoder","uri":"program://CREMA/class/lavis.models.topk.ATPEncoder#L217-L260","kind":"class","name":"ATPEncoder","path":"lavis/models/topk.py","language":"python","start_line":217,"end_line":260,"context_start_line":197,"context_end_line":280,"code":" use_text_cands = False,\n n_cands = 5,\n use_ste = True,\n sel_dropout = 0.0,\n d_input = 512)\n\n @classmethod\n def from_args(cls, args):\n return cls(n_layers = args.n_layers,\n n_heads = args.n_heads,\n d_model = args.d_model,\n d_model_ff = args.d_model_ff,\n enc_dropout = args.enc_dropout,\n use_text_query = args.use_text_query,\n use_text_cands = args.use_text_cands,\n n_cands = args.n_cands,\n use_ste = args.use_ste,\n sel_dropout = args.sel_dropout,\n d_input = args.d_input)\n\nclass ATPEncoder(nn.Module):\n \"\"\"\n The multimodal transformer encoder for the ATP model. For analysis purposes, the ATP encoder\n does not use any positional information (no positional encodings + transformer / self-attention)\n and is generally kept low-capacity. If the goal is raw accuracy (not analysis), you can relax these constraints.\n \"\"\"\n def __init__(self, config: ATPConfig):\n \"\"\"\n config: ATPConfig with parameters for the (transformer-based, atemporal) encoder for ATP.\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.d_model = config.d_model\n\n self.dropout = nn.Dropout(p=config.enc_dropout)\n\n\n self.modality_encoding = ModalityEmbeddings(d_model=self.d_model,\n use_text_query=config.use_text_query,\n use_text_cands=config.use_text_cands,\n n_cands=config.n_cands)\n \n atp_encoder_layer = nn.TransformerEncoderLayer(\n d_model=self.d_model,\n nhead=config.n_heads,\n dim_feedforward=config.d_model_ff,\n dropout=config.enc_dropout,\n activation='relu'\n )\n\n self.transformer_encoder = nn.TransformerEncoder(atp_encoder_layer, config.n_layers)\n\n def forward(self, x_inputs: torch.tensor, vis_L):\n \"\"\"\n x_inputs: torch.tensor of shape (L, N, D)\n \"\"\"\n L, N, D = x_inputs.size() # (sequence_length, batch_size, d_model)\n assert D == self.d_model, \"inputs dimension mismatch\"\n x_encoded = x_inputs * math.sqrt(self.d_model)\n x_encoded += self.modality_encoding(x_encoded, vis_L)\n x_encoded = self.dropout(x_encoded)\n x_encoded = self.transformer_encoder(x_encoded)\n\n return x_encoded\n\nclass TopK_Selector(nn.Module):\n \"\"\"\n The Atemporal Probe (ATP) selector model. Takes as input a sequence of image-language \n encoding and outputs a (discrete) selection over the input frames, to help analyze \n downstream discriminative video-language tasks.\n \"\"\"\n \n def __init__(self, config=ATPConfig, num_select=4):\n \"\"\"\n config: ATPConfig with parameters for initializing the ATPSelectorModel (and its encoder).\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.config = config\n self.t_embedding = nn.Linear(config.d_input_t, config.d_input)\n self.v_embedding = nn.Linear(config.d_input_v, config.d_input)\n self.embedding = nn.Linear(config.d_input, config.d_model)\n self.atp_encoder = ATPEncoder(config)\n self.dropout = nn.Dropout(p=config.sel_dropout)","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.TopK_Selector","uri":"program://CREMA/class/lavis.models.topk.TopK_Selector#L262-L326","kind":"class","name":"TopK_Selector","path":"lavis/models/topk.py","language":"python","start_line":262,"end_line":326,"context_start_line":242,"context_end_line":339,"code":" dim_feedforward=config.d_model_ff,\n dropout=config.enc_dropout,\n activation='relu'\n )\n\n self.transformer_encoder = nn.TransformerEncoder(atp_encoder_layer, config.n_layers)\n\n def forward(self, x_inputs: torch.tensor, vis_L):\n \"\"\"\n x_inputs: torch.tensor of shape (L, N, D)\n \"\"\"\n L, N, D = x_inputs.size() # (sequence_length, batch_size, d_model)\n assert D == self.d_model, \"inputs dimension mismatch\"\n x_encoded = x_inputs * math.sqrt(self.d_model)\n x_encoded += self.modality_encoding(x_encoded, vis_L)\n x_encoded = self.dropout(x_encoded)\n x_encoded = self.transformer_encoder(x_encoded)\n\n return x_encoded\n\nclass TopK_Selector(nn.Module):\n \"\"\"\n The Atemporal Probe (ATP) selector model. Takes as input a sequence of image-language \n encoding and outputs a (discrete) selection over the input frames, to help analyze \n downstream discriminative video-language tasks.\n \"\"\"\n \n def __init__(self, config=ATPConfig, num_select=4):\n \"\"\"\n config: ATPConfig with parameters for initializing the ATPSelectorModel (and its encoder).\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.config = config\n self.t_embedding = nn.Linear(config.d_input_t, config.d_input)\n self.v_embedding = nn.Linear(config.d_input_v, config.d_input)\n self.embedding = nn.Linear(config.d_input, config.d_model)\n self.atp_encoder = ATPEncoder(config)\n self.dropout = nn.Dropout(p=config.sel_dropout)\n self.logits = nn.Linear(config.d_model, 1)\n self.num_select = num_select\n self.sigma = 0.1\n\n def forward(self,\n x_vis, # [b, t, d]\n x_txt, # [b, n, d]\n **kwargs):\n \"\"\"\n \"\"\"\n x_vis_cls = x_vis[:, :, 0, :] # b t n c\n N, vis_L, D = x_vis_cls.size() # (batch_size, sequence_length, feature_dimension)\n # embed the input sequence to the (smaller) model dimension (d_model) with modality encodings.\n x_vis_cls = self.v_embedding(self.dropout(x_vis_cls))\n x_txt = self.t_embedding(self.dropout(x_txt))\n x_inputs = []\n x_vis_cls = x_vis_cls.permute(1, 0, 2)\n x_inputs.append(x_txt.permute(1,0,2)) # (n, b, d)\n x_inputs.append(x_vis_cls)\n x_inputs = torch.cat(x_inputs, dim=0)\n x_encoded = self.embedding(self.dropout(x_inputs))\n x_atp_encoded = self.atp_encoder(x_encoded, vis_L)\n x_atp_encoded = x_atp_encoded.permute(1, 0, 2)\n x_encoded_v = x_atp_encoded[:, -vis_L: , :]\n # obtain selection scores (logits)\n x_logits = self.logits(self.dropout(x_encoded_v)).squeeze()\n #print('x_logits', x_logits.shape)\n\n if self.training:\n indices = PerturbedTopKFunction.apply(x_logits, self.num_select)\n #print('indices', indices.shape)\n indices = einops.rearrange(indices, \"b k d -> b d k\")\n\n if indices is not None:\n qa_frames = extract_frames_from_indicators(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n else:\n indices = HardTopK(self.num_select, x_logits)\n if indices is not None:\n qa_frames = extract_frames_from_indices(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n\n\n return qa_frames\n\nif __name__ == \"__main__\":\n selector_config = ATPConfig.default_args\n\n Selector = TopK_Selector(num_select=4) #.eval()\n\n x_vis = torch.rand([2, 8, 257, 1408])\n x_txt = torch.rand([2, 68, 2048])\n\n out = Selector(x_vis, x_txt)\n print(out.shape)\n\n","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.__init__","uri":"program://CREMA/function/lavis.models.topk.__init__#L269-L283","kind":"function","name":"__init__","path":"lavis/models/topk.py","language":"python","start_line":269,"end_line":283,"context_start_line":249,"context_end_line":303,"code":" def forward(self, x_inputs: torch.tensor, vis_L):\n \"\"\"\n x_inputs: torch.tensor of shape (L, N, D)\n \"\"\"\n L, N, D = x_inputs.size() # (sequence_length, batch_size, d_model)\n assert D == self.d_model, \"inputs dimension mismatch\"\n x_encoded = x_inputs * math.sqrt(self.d_model)\n x_encoded += self.modality_encoding(x_encoded, vis_L)\n x_encoded = self.dropout(x_encoded)\n x_encoded = self.transformer_encoder(x_encoded)\n\n return x_encoded\n\nclass TopK_Selector(nn.Module):\n \"\"\"\n The Atemporal Probe (ATP) selector model. Takes as input a sequence of image-language \n encoding and outputs a (discrete) selection over the input frames, to help analyze \n downstream discriminative video-language tasks.\n \"\"\"\n \n def __init__(self, config=ATPConfig, num_select=4):\n \"\"\"\n config: ATPConfig with parameters for initializing the ATPSelectorModel (and its encoder).\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.config = config\n self.t_embedding = nn.Linear(config.d_input_t, config.d_input)\n self.v_embedding = nn.Linear(config.d_input_v, config.d_input)\n self.embedding = nn.Linear(config.d_input, config.d_model)\n self.atp_encoder = ATPEncoder(config)\n self.dropout = nn.Dropout(p=config.sel_dropout)\n self.logits = nn.Linear(config.d_model, 1)\n self.num_select = num_select\n self.sigma = 0.1\n\n def forward(self,\n x_vis, # [b, t, d]\n x_txt, # [b, n, d]\n **kwargs):\n \"\"\"\n \"\"\"\n x_vis_cls = x_vis[:, :, 0, :] # b t n c\n N, vis_L, D = x_vis_cls.size() # (batch_size, sequence_length, feature_dimension)\n # embed the input sequence to the (smaller) model dimension (d_model) with modality encodings.\n x_vis_cls = self.v_embedding(self.dropout(x_vis_cls))\n x_txt = self.t_embedding(self.dropout(x_txt))\n x_inputs = []\n x_vis_cls = x_vis_cls.permute(1, 0, 2)\n x_inputs.append(x_txt.permute(1,0,2)) # (n, b, d)\n x_inputs.append(x_vis_cls)\n x_inputs = torch.cat(x_inputs, dim=0)\n x_encoded = self.embedding(self.dropout(x_inputs))\n x_atp_encoded = self.atp_encoder(x_encoded, vis_L)\n x_atp_encoded = x_atp_encoded.permute(1, 0, 2)","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.__call__","uri":"program://CREMA/function/lavis.models.topk.__call__#L24-L25","kind":"function","name":"__call__","path":"lavis/models/topk.py","language":"python","start_line":24,"end_line":25,"context_start_line":4,"context_end_line":45,"code":"\"\"\"\n\nimport math\nimport torch\nimport copy\nimport einops\nimport torch.nn.functional as F\nfrom torch import nn\n\nfrom dataclasses import dataclass\nfrom typing import Optional\nfrom enum import IntEnum\nfrom einops import rearrange\n\nclass PerturbedTopK(nn.Module):\n def __init__(self, k: int, num_samples: int = 1000):\n super(PerturbedTopK, self).__init__()\n self.num_samples = num_samples\n self.k = k\n\n def __call__(self, x, sigma):\n return PerturbedTopKFunction.apply(x, self.k, self.num_samples, sigma)\n\n\nclass PerturbedTopKFunction(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, k: int, num_samples: int = 1000, sigma: float = 0.05):\n #print('x', x.shape)\n b, d = x.shape\n # for Gaussian: noise and gradient are the same.\n noise = torch.normal(mean=0.0, std=1.0, size=(b, num_samples, d)).to(x.device)\n perturbed_x = x[:, None, :] + noise * sigma # b, nS, d\n #print('perturbed_x', perturbed_x.shape)\n topk_results = torch.topk(perturbed_x, k=k, dim=-1, sorted=False)\n #print('topk_results',topk_results)\n\n indices = topk_results.indices # b, nS, k\n indices = torch.sort(indices, dim=-1).values # b, nS, k\n # print('indices', indices.shape ,indices[0,0,0])\n\n perturbed_output = torch.nn.functional.one_hot(indices, num_classes=d).float()\n indicators = perturbed_output.mean(dim=1) # b, k, d","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.forward","uri":"program://CREMA/function/lavis.models.topk.forward#L285-L326","kind":"function","name":"forward","path":"lavis/models/topk.py","language":"python","start_line":285,"end_line":326,"context_start_line":265,"context_end_line":339,"code":" encoding and outputs a (discrete) selection over the input frames, to help analyze \n downstream discriminative video-language tasks.\n \"\"\"\n \n def __init__(self, config=ATPConfig, num_select=4):\n \"\"\"\n config: ATPConfig with parameters for initializing the ATPSelectorModel (and its encoder).\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.config = config\n self.t_embedding = nn.Linear(config.d_input_t, config.d_input)\n self.v_embedding = nn.Linear(config.d_input_v, config.d_input)\n self.embedding = nn.Linear(config.d_input, config.d_model)\n self.atp_encoder = ATPEncoder(config)\n self.dropout = nn.Dropout(p=config.sel_dropout)\n self.logits = nn.Linear(config.d_model, 1)\n self.num_select = num_select\n self.sigma = 0.1\n\n def forward(self,\n x_vis, # [b, t, d]\n x_txt, # [b, n, d]\n **kwargs):\n \"\"\"\n \"\"\"\n x_vis_cls = x_vis[:, :, 0, :] # b t n c\n N, vis_L, D = x_vis_cls.size() # (batch_size, sequence_length, feature_dimension)\n # embed the input sequence to the (smaller) model dimension (d_model) with modality encodings.\n x_vis_cls = self.v_embedding(self.dropout(x_vis_cls))\n x_txt = self.t_embedding(self.dropout(x_txt))\n x_inputs = []\n x_vis_cls = x_vis_cls.permute(1, 0, 2)\n x_inputs.append(x_txt.permute(1,0,2)) # (n, b, d)\n x_inputs.append(x_vis_cls)\n x_inputs = torch.cat(x_inputs, dim=0)\n x_encoded = self.embedding(self.dropout(x_inputs))\n x_atp_encoded = self.atp_encoder(x_encoded, vis_L)\n x_atp_encoded = x_atp_encoded.permute(1, 0, 2)\n x_encoded_v = x_atp_encoded[:, -vis_L: , :]\n # obtain selection scores (logits)\n x_logits = self.logits(self.dropout(x_encoded_v)).squeeze()\n #print('x_logits', x_logits.shape)\n\n if self.training:\n indices = PerturbedTopKFunction.apply(x_logits, self.num_select)\n #print('indices', indices.shape)\n indices = einops.rearrange(indices, \"b k d -> b d k\")\n\n if indices is not None:\n qa_frames = extract_frames_from_indicators(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n else:\n indices = HardTopK(self.num_select, x_logits)\n if indices is not None:\n qa_frames = extract_frames_from_indices(x_vis, indices)\n else:\n raise RuntimeError(\"Empty indices!\")\n\n\n return qa_frames\n\nif __name__ == \"__main__\":\n selector_config = ATPConfig.default_args\n\n Selector = TopK_Selector(num_select=4) #.eval()\n\n x_vis = torch.rand([2, 8, 257, 1408])\n x_txt = torch.rand([2, 68, 2048])\n\n out = Selector(x_vis, x_txt)\n print(out.shape)\n\n","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.backward","uri":"program://CREMA/function/lavis.models.topk.backward#L59-L76","kind":"function","name":"backward","path":"lavis/models/topk.py","language":"python","start_line":59,"end_line":76,"context_start_line":39,"context_end_line":96,"code":"\n indices = topk_results.indices # b, nS, k\n indices = torch.sort(indices, dim=-1).values # b, nS, k\n # print('indices', indices.shape ,indices[0,0,0])\n\n perturbed_output = torch.nn.functional.one_hot(indices, num_classes=d).float()\n indicators = perturbed_output.mean(dim=1) # b, k, d\n # print('perturbed_output', perturbed_output.shape, perturbed_output[0,indices[0,0,0],0,0])\n\n # constants for backward\n ctx.k = k\n ctx.num_samples = num_samples\n ctx.sigma = sigma\n\n # tensors for backward\n ctx.perturbed_output = perturbed_output\n ctx.noise = noise\n return indicators\n\n @staticmethod\n def backward(ctx, grad_output):\n if grad_output is None:\n return tuple([None] * 5)\n\n noise_gradient = ctx.noise\n if ctx.sigma <= 1e-20:\n b, _, k, d = ctx.perturbed_output.size()\n expected_gradient = torch.zeros(b, k, d).to(grad_output.device)\n else:\n expected_gradient = (\n torch.einsum(\"bnkd,bnd->bkd\", ctx.perturbed_output, noise_gradient)\n / ctx.num_samples\n / (ctx.sigma)\n )\n\n grad_input = torch.einsum(\"bkd,bkd->bd\", grad_output, expected_gradient)\n\n return (grad_input,) + tuple([None] * 5)\n\ndef HardTopK(k, x):\n topk_results = torch.topk(x, k=k, dim=-1, sorted=False)\n indices = topk_results.indices # b, k\n indices = torch.sort(indices, dim=-1).values\n return indices\n\n\ndef batched_index_select(input, dim, index):\n for i in range(1, len(input.shape)):\n if i != dim:\n index = index.unsqueeze(i)\n expanse = list(input.shape)\n expanse[0] = -1\n expanse[dim] = -1\n index = index.expand(expanse)\n return torch.gather(input, dim, index)\n\ndef extract_frames_from_indices(x, indices):\n batch_size, _, n, channels = x.shape","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.default_args","uri":"program://CREMA/function/lavis.models.topk.default_args#L188-L201","kind":"function","name":"default_args","path":"lavis/models/topk.py","language":"python","start_line":188,"end_line":201,"context_start_line":168,"context_end_line":221,"code":"class ATPConfig:\n '''\n ATPConfig contains the parameters needed for the ATPSelectorModel (and its ATPEncoder).\n '''\n # ATPEncoder params\n n_layers: int = 6\n n_heads: int = 4\n d_model: int = 256\n d_input_t: int = 2048\n d_input_v: int = 1408\n d_model_ff: int = 256\n enc_dropout: float = 0.1\n use_text_query: bool = True # at least one use_text_* needs to be true for ATP to be multimodal\n use_text_cands: bool = False # ^ see above. (note: if both are false, ATP is vision-only)\n n_cands: int = 5 # only relevant when use_text_cands is set to true\n # ATPSelector params\n use_ste: bool = True # controls type of selector during ATP training; see ATPSelectorModel.forward\n sel_dropout: float = 0.0\n d_input: int = 512 # size of the input vision-language embeddings (e.g. CLIP-ViT-B32 is size 512)\n \n def default_args(cls):\n return cls(n_layers = 6,\n n_heads = 4,\n d_model = 256,\n d_input_t = 2048,\n d_input_v = 1408,\n d_model_ff = 256,\n enc_dropout = 0.1,\n use_text_query = True,\n use_text_cands = False,\n n_cands = 5,\n use_ste = True,\n sel_dropout = 0.0,\n d_input = 512)\n\n @classmethod\n def from_args(cls, args):\n return cls(n_layers = args.n_layers,\n n_heads = args.n_heads,\n d_model = args.d_model,\n d_model_ff = args.d_model_ff,\n enc_dropout = args.enc_dropout,\n use_text_query = args.use_text_query,\n use_text_cands = args.use_text_cands,\n n_cands = args.n_cands,\n use_ste = args.use_ste,\n sel_dropout = args.sel_dropout,\n d_input = args.d_input)\n\nclass ATPEncoder(nn.Module):\n \"\"\"\n The multimodal transformer encoder for the ATP model. For analysis purposes, the ATP encoder\n does not use any positional information (no positional encodings + transformer / self-attention)\n and is generally kept low-capacity. If the goal is raw accuracy (not analysis), you can relax these constraints.","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.topk.from_args","uri":"program://CREMA/function/lavis.models.topk.from_args#L204-L215","kind":"function","name":"from_args","path":"lavis/models/topk.py","language":"python","start_line":204,"end_line":215,"context_start_line":184,"context_end_line":235,"code":" use_ste: bool = True # controls type of selector during ATP training; see ATPSelectorModel.forward\n sel_dropout: float = 0.0\n d_input: int = 512 # size of the input vision-language embeddings (e.g. CLIP-ViT-B32 is size 512)\n \n def default_args(cls):\n return cls(n_layers = 6,\n n_heads = 4,\n d_model = 256,\n d_input_t = 2048,\n d_input_v = 1408,\n d_model_ff = 256,\n enc_dropout = 0.1,\n use_text_query = True,\n use_text_cands = False,\n n_cands = 5,\n use_ste = True,\n sel_dropout = 0.0,\n d_input = 512)\n\n @classmethod\n def from_args(cls, args):\n return cls(n_layers = args.n_layers,\n n_heads = args.n_heads,\n d_model = args.d_model,\n d_model_ff = args.d_model_ff,\n enc_dropout = args.enc_dropout,\n use_text_query = args.use_text_query,\n use_text_cands = args.use_text_cands,\n n_cands = args.n_cands,\n use_ste = args.use_ste,\n sel_dropout = args.sel_dropout,\n d_input = args.d_input)\n\nclass ATPEncoder(nn.Module):\n \"\"\"\n The multimodal transformer encoder for the ATP model. For analysis purposes, the ATP encoder\n does not use any positional information (no positional encodings + transformer / self-attention)\n and is generally kept low-capacity. If the goal is raw accuracy (not analysis), you can relax these constraints.\n \"\"\"\n def __init__(self, config: ATPConfig):\n \"\"\"\n config: ATPConfig with parameters for the (transformer-based, atemporal) encoder for ATP.\n See ATPConfig documentation for details.\n \"\"\"\n super().__init__()\n self.d_model = config.d_model\n\n self.dropout = nn.Dropout(p=config.enc_dropout)\n\n\n self.modality_encoding = ModalityEmbeddings(d_model=self.d_model,\n use_text_query=config.use_text_query,","source_hash":"1d245a1cf859112c5bb44cbe7be532b7e16359190a7e06d4257e63baaedd8bf3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med","uri":"program://CREMA/module/lavis.models.med#L1-L1416","kind":"module","name":"lavis.models.med","path":"lavis/models/med.py","language":"python","start_line":1,"end_line":1416,"context_start_line":1,"context_end_line":1416,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n \n Based on huggingface code base\n https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert\n\"\"\"\n\nimport math\nimport os\nimport warnings\nfrom dataclasses import dataclass\nfrom typing import Optional, Tuple\n\nimport torch\nfrom torch import Tensor, device\nimport torch.utils.checkpoint\nfrom torch import nn\nfrom torch.nn import CrossEntropyLoss\nimport torch.nn.functional as F\nfrom transformers import BatchEncoding, PreTrainedTokenizer\n\nfrom transformers.activations import ACT2FN\nfrom transformers.file_utils import (\n ModelOutput,\n)\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPastAndCrossAttentions,\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n MaskedLMOutput,\n MultipleChoiceModelOutput,\n NextSentencePredictorOutput,\n QuestionAnsweringModelOutput,\n SequenceClassifierOutput,\n TokenClassifierOutput,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.utils import logging\nfrom transformers.models.bert.configuration_bert import BertConfig\nfrom lavis.common.utils import get_abs_path\n\nfrom lavis.models.base_model import BaseEncoder\n\nlogging.set_verbosity_error()\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n if config.add_type_embeddings:\n self.token_type_embeddings = nn.Embedding(\n config.type_vocab_size, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n token_type_ids=None,\n position_ids=None,\n inputs_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n input_shape = input_ids.size()\n else:\n input_shape = inputs_embeds.size()[:-1]\n\n seq_length = input_shape[1]\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n if token_type_ids is not None:\n token_type_embeddings = self.token_type_embeddings(token_type_ids)\n\n embeddings = inputs_embeds + token_type_embeddings\n else:\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n elif past_key_value is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n\n # compatibility for ALBEF and BLIP\n try:\n # ALBEF & ALPRO\n fusion_layer = self.config.fusion_layer\n add_cross_attention = (\n fusion_layer <= layer_num and self.config.add_cross_attention\n )\n\n self.fusion_layer = fusion_layer\n except AttributeError:\n # BLIP\n self.fusion_layer = self.config.num_hidden_layers\n add_cross_attention = self.config.add_cross_attention\n\n # if self.config.add_cross_attention:\n if add_cross_attention:\n self.crossattention = BertAttention(\n config, is_cross_attention=self.config.add_cross_attention\n )\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n mode=None,\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n )\n attention_output = self_attention_outputs[0]\n\n outputs = self_attention_outputs[1:-1]\n present_key_value = self_attention_outputs[-1]\n\n # TODO line 482 in albef/models/xbert.py\n # compatibility for ALBEF and BLIP\n if mode in [\"multimodal\", \"fusion\"] and hasattr(self, \"crossattention\"):\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n\n if isinstance(encoder_hidden_states, list):\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states[\n (self.layer_num - self.fusion_layer)\n % len(encoder_hidden_states)\n ],\n encoder_attention_mask[\n (self.layer_num - self.fusion_layer)\n % len(encoder_hidden_states)\n ],\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = outputs + cross_attention_outputs[1:-1]\n\n else:\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,\n output_hidden_states=False,\n return_dict=True,\n mode=\"multimodal\",\n ):\n all_hidden_states = () if output_hidden_states else None\n all_self_attentions = () if output_attentions else None\n all_cross_attentions = (\n () if output_attentions and self.config.add_cross_attention else None\n )\n\n next_decoder_cache = () if use_cache else None\n\n try:\n # ALBEF\n fusion_layer = self.config.fusion_layer\n except AttributeError:\n # BLIP\n fusion_layer = self.config.num_hidden_layers\n\n if mode == \"text\":\n# ... truncated ...","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertEmbeddings","uri":"program://CREMA/class/lavis.models.med.BertEmbeddings#L56-L123","kind":"class","name":"BertEmbeddings","path":"lavis/models/med.py","language":"python","start_line":56,"end_line":123,"context_start_line":36,"context_end_line":143,"code":" QuestionAnsweringModelOutput,\n SequenceClassifierOutput,\n TokenClassifierOutput,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.utils import logging\nfrom transformers.models.bert.configuration_bert import BertConfig\nfrom lavis.common.utils import get_abs_path\n\nfrom lavis.models.base_model import BaseEncoder\n\nlogging.set_verbosity_error()\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n if config.add_type_embeddings:\n self.token_type_embeddings = nn.Embedding(\n config.type_vocab_size, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n token_type_ids=None,\n position_ids=None,\n inputs_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n input_shape = input_ids.size()\n else:\n input_shape = inputs_embeds.size()[:-1]\n\n seq_length = input_shape[1]\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n if token_type_ids is not None:\n token_type_embeddings = self.token_type_embeddings(token_type_ids)\n\n embeddings = inputs_embeds + token_type_embeddings\n else:\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertSelfAttention","uri":"program://CREMA/class/lavis.models.med.BertSelfAttention#L126-L289","kind":"class","name":"BertSelfAttention","path":"lavis/models/med.py","language":"python","start_line":126,"end_line":289,"context_start_line":106,"context_end_line":309,"code":" ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n if token_type_ids is not None:\n token_type_embeddings = self.token_type_embeddings(token_type_ids)\n\n embeddings = inputs_embeds + token_type_embeddings\n else:\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n elif past_key_value is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertSelfOutput","uri":"program://CREMA/class/lavis.models.med.BertSelfOutput#L292-L303","kind":"class","name":"BertSelfOutput","path":"lavis/models/med.py","language":"python","start_line":292,"end_line":303,"context_start_line":272,"context_end_line":323,"code":" attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertAttention","uri":"program://CREMA/class/lavis.models.med.BertAttention#L306-L359","kind":"class","name":"BertAttention","path":"lavis/models/med.py","language":"python","start_line":306,"end_line":359,"context_start_line":286,"context_end_line":379,"code":" )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertIntermediate","uri":"program://CREMA/class/lavis.models.med.BertIntermediate#L362-L374","kind":"class","name":"BertIntermediate","path":"lavis/models/med.py","language":"python","start_line":362,"end_line":374,"context_start_line":342,"context_end_line":394,"code":" encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertOutput","uri":"program://CREMA/class/lavis.models.med.BertOutput#L377-L388","kind":"class","name":"BertOutput","path":"lavis/models/med.py","language":"python","start_line":377,"end_line":388,"context_start_line":357,"context_end_line":408,"code":" 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n\n # compatibility for ALBEF and BLIP\n try:\n # ALBEF & ALPRO\n fusion_layer = self.config.fusion_layer\n add_cross_attention = (\n fusion_layer <= layer_num and self.config.add_cross_attention\n )\n\n self.fusion_layer = fusion_layer","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertLayer","uri":"program://CREMA/class/lavis.models.med.BertLayer#L391-L502","kind":"class","name":"BertLayer","path":"lavis/models/med.py","language":"python","start_line":391,"end_line":502,"context_start_line":371,"context_end_line":522,"code":" def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n\n # compatibility for ALBEF and BLIP\n try:\n # ALBEF & ALPRO\n fusion_layer = self.config.fusion_layer\n add_cross_attention = (\n fusion_layer <= layer_num and self.config.add_cross_attention\n )\n\n self.fusion_layer = fusion_layer\n except AttributeError:\n # BLIP\n self.fusion_layer = self.config.num_hidden_layers\n add_cross_attention = self.config.add_cross_attention\n\n # if self.config.add_cross_attention:\n if add_cross_attention:\n self.crossattention = BertAttention(\n config, is_cross_attention=self.config.add_cross_attention\n )\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n mode=None,\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n )\n attention_output = self_attention_outputs[0]\n\n outputs = self_attention_outputs[1:-1]\n present_key_value = self_attention_outputs[-1]\n\n # TODO line 482 in albef/models/xbert.py\n # compatibility for ALBEF and BLIP\n if mode in [\"multimodal\", \"fusion\"] and hasattr(self, \"crossattention\"):\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n\n if isinstance(encoder_hidden_states, list):\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states[\n (self.layer_num - self.fusion_layer)\n % len(encoder_hidden_states)\n ],\n encoder_attention_mask[\n (self.layer_num - self.fusion_layer)\n % len(encoder_hidden_states)\n ],\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = outputs + cross_attention_outputs[1:-1]\n\n else:\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertEncoder","uri":"program://CREMA/class/lavis.models.med.BertEncoder#L505-L630","kind":"class","name":"BertEncoder","path":"lavis/models/med.py","language":"python","start_line":505,"end_line":630,"context_start_line":485,"context_end_line":650,"code":" outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,\n output_hidden_states=False,\n return_dict=True,\n mode=\"multimodal\",\n ):\n all_hidden_states = () if output_hidden_states else None\n all_self_attentions = () if output_attentions else None\n all_cross_attentions = (\n () if output_attentions and self.config.add_cross_attention else None\n )\n\n next_decoder_cache = () if use_cache else None\n\n try:\n # ALBEF\n fusion_layer = self.config.fusion_layer\n except AttributeError:\n # BLIP\n fusion_layer = self.config.num_hidden_layers\n\n if mode == \"text\":\n start_layer = 0\n # output_layer = self.config.fusion_layer\n output_layer = fusion_layer\n\n elif mode == \"fusion\":\n # start_layer = self.config.fusion_layer\n start_layer = fusion_layer\n output_layer = self.config.num_hidden_layers\n\n elif mode == \"multimodal\":\n start_layer = 0\n output_layer = self.config.num_hidden_layers\n\n # compatibility for ALBEF and BLIP\n # for i in range(self.config.num_hidden_layers):\n for i in range(start_layer, output_layer):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n # TODO pay attention to this.\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n mode=mode,\n )\n\n hidden_states = layer_outputs[0]\n if use_cache:\n next_decoder_cache += (layer_outputs[-1],)\n if output_attentions:\n all_self_attentions = all_self_attentions + (layer_outputs[1],)\n\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if not return_dict:\n return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertPooler","uri":"program://CREMA/class/lavis.models.med.BertPooler#L633-L645","kind":"class","name":"BertPooler","path":"lavis/models/med.py","language":"python","start_line":633,"end_line":645,"context_start_line":613,"context_end_line":665,"code":" return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertPredictionHeadTransform","uri":"program://CREMA/class/lavis.models.med.BertPredictionHeadTransform#L648-L662","kind":"class","name":"BertPredictionHeadTransform","path":"lavis/models/med.py","language":"python","start_line":648,"end_line":662,"context_start_line":628,"context_end_line":682,"code":" attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertLMPredictionHead","uri":"program://CREMA/class/lavis.models.med.BertLMPredictionHead#L665-L682","kind":"class","name":"BertLMPredictionHead","path":"lavis/models/med.py","language":"python","start_line":665,"end_line":682,"context_start_line":645,"context_end_line":702,"code":" return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertOnlyMLMHead","uri":"program://CREMA/class/lavis.models.med.BertOnlyMLMHead#L685-L692","kind":"class","name":"BertOnlyMLMHead","path":"lavis/models/med.py","language":"python","start_line":685,"end_line":692,"context_start_line":665,"context_end_line":712,"code":"class BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertPreTrainedModel","uri":"program://CREMA/class/lavis.models.med.BertPreTrainedModel#L695-L715","kind":"class","name":"BertPreTrainedModel","path":"lavis/models/med.py","language":"python","start_line":695,"end_line":715,"context_start_line":675,"context_end_line":735,"code":"\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertModel","uri":"program://CREMA/class/lavis.models.med.BertModel#L718-L1002","kind":"class","name":"BertModel","path":"lavis/models/med.py","language":"python","start_line":718,"end_line":1002,"context_start_line":698,"context_end_line":1022,"code":" models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n # in case past_key_values are used we need to add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, seq_length, prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n token_type_ids=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n \"\"\"\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n if is_decoder:\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n else:\n use_cache = False\n\n if input_ids is not None and inputs_embeds is not None:\n raise ValueError(\n \"You cannot specify both input_ids and inputs_embeds at the same time\"\n )\n elif input_ids is not None:\n input_shape = input_ids.size()\n batch_size, seq_length = input_shape\n device = input_ids.device\n elif inputs_embeds is not None:\n input_shape = inputs_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = inputs_embeds.device\n elif encoder_embeds is not None:\n input_shape = encoder_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = encoder_embeds.device\n else:\n raise ValueError(\n \"You have to specify either input_ids or inputs_embeds or encoder_embeds\"\n )\n\n # past_key_values_length\n past_key_values_length = (\n past_key_values[0][0].shape[2] if past_key_values is not None else 0\n )\n\n if attention_mask is None:\n attention_mask = torch.ones(\n ((batch_size, seq_length + past_key_values_length)), device=device\n )\n\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(\n attention_mask, input_shape, device, is_decoder\n )\n\n # If a 2D or 3D attention mask is provided for the cross-attention\n # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if encoder_hidden_states is not None:\n if type(encoder_hidden_states) == list:\n encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[\n 0\n ].size()\n else:\n (\n encoder_batch_size,\n encoder_sequence_length,\n _,\n ) = encoder_hidden_states.size()\n encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)\n\n if type(encoder_attention_mask) == list:\n encoder_extended_attention_mask = [\n self.invert_attention_mask(mask) for mask in encoder_attention_mask\n ]\n elif encoder_attention_mask is None:\n encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = None\n\n # Prepare head mask if needed\n # 1.0 in head_mask indicate we keep the head\n # attention_probs has shape bsz x n_heads x N x N\n # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]\n # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]\n head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)\n\n if encoder_embeds is None:\n embedding_output = self.embeddings(\n input_ids=input_ids,\n position_ids=position_ids,\n token_type_ids=token_type_ids,\n inputs_embeds=inputs_embeds,\n past_key_values_length=past_key_values_length,\n )\n else:\n embedding_output = encoder_embeds\n\n encoder_outputs = self.encoder(\n embedding_output,\n attention_mask=extended_attention_mask,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n mode=mode,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertForMaskedLM","uri":"program://CREMA/class/lavis.models.med.BertForMaskedLM#L1005-L1128","kind":"class","name":"BertForMaskedLM","path":"lavis/models/med.py","language":"python","start_line":1005,"end_line":1128,"context_start_line":985,"context_end_line":1148,"code":" mode=mode,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n # token_type_ids=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n soft_labels=None,\n alpha=0,\n return_logits=False,\n ):\n r\"\"\"\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,\n config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored\n (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``\n \"\"\"\n\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n # token_type_ids=token_type_ids,\n position_ids=position_ids,\n head_mask=head_mask,\n inputs_embeds=inputs_embeds,\n encoder_embeds=encoder_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n mode=mode,\n )\n\n sequence_output = outputs[0]\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores\n\n masked_lm_loss = None\n if labels is not None:\n loss_fct = CrossEntropyLoss() # -100 index = padding token\n masked_lm_loss = loss_fct(\n prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)\n )\n\n if soft_labels is not None:\n loss_distill = -torch.sum(\n F.log_softmax(prediction_scores, dim=-1) * soft_labels, dim=-1\n )\n loss_distill = loss_distill[labels != -100].mean()\n masked_lm_loss = (1 - alpha) * masked_lm_loss + alpha * loss_distill\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return (\n ((masked_lm_loss,) + output) if masked_lm_loss is not None else output\n )\n\n return MaskedLMOutput(\n loss=masked_lm_loss,\n logits=prediction_scores,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, attention_mask=None, **model_kwargs\n ):\n input_shape = input_ids.shape\n effective_batch_size = input_shape[0]\n\n # add a dummy token\n assert (\n self.config.pad_token_id is not None\n ), \"The PAD token should be defined for generation\"\n attention_mask = torch.cat(\n [attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))],\n dim=-1,\n )\n dummy_token = torch.full(\n (effective_batch_size, 1),\n self.config.pad_token_id,\n dtype=torch.long,\n device=input_ids.device,\n )\n input_ids = torch.cat([input_ids, dummy_token], dim=1)\n\n return {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.BertLMHeadModel","uri":"program://CREMA/class/lavis.models.med.BertLMHeadModel#L1131-L1295","kind":"class","name":"BertLMHeadModel","path":"lavis/models/med.py","language":"python","start_line":1131,"end_line":1295,"context_start_line":1111,"context_end_line":1315,"code":"\n # add a dummy token\n assert (\n self.config.pad_token_id is not None\n ), \"The PAD token should be defined for generation\"\n attention_mask = torch.cat(\n [attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))],\n dim=-1,\n )\n dummy_token = torch.full(\n (effective_batch_size, 1),\n self.config.pad_token_id,\n dtype=torch.long,\n device=input_ids.device,\n )\n input_ids = torch.cat([input_ids, dummy_token], dim=1)\n\n return {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=True,\n reduction=\"mean\",\n mode=\"multimodal\",\n soft_labels=None,\n alpha=0,\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in\n ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are\n ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n Returns:\n Example::\n >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig\n >>> import torch\n >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\n >>> config = BertConfig.from_pretrained(\"bert-base-cased\")\n >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)\n >>> inputs = tokenizer(\"Hello, my dog is cute\", return_tensors=\"pt\")\n >>> outputs = model(**inputs)\n >>> prediction_logits = outputs.logits\n \"\"\"\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n if labels is not None:\n use_cache = False\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n position_ids=position_ids,\n head_mask=head_mask,\n inputs_embeds=inputs_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n mode=mode,\n )\n\n sequence_output = outputs[0]\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores[:, :-1, :].contiguous()\n\n lm_loss = None\n if labels is not None:\n # we are doing next-token prediction; shift prediction scores and input ids by one\n shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()\n labels = labels[:, 1:].contiguous()\n loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)\n lm_loss = loss_fct(\n shifted_prediction_scores.view(-1, self.config.vocab_size),\n labels.view(-1),\n )\n if reduction == \"none\":\n lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)\n\n if soft_labels is not None:\n loss_distill = -torch.sum(\n F.log_softmax(shifted_prediction_scores, dim=-1) * soft_labels, dim=-1\n )\n loss_distill = (loss_distill * (labels != -100)).sum(1)\n lm_loss = (1 - alpha) * lm_loss + alpha * loss_distill\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return ((lm_loss,) + output) if lm_loss is not None else output\n\n return CausalLMOutputWithCrossAttentions(\n loss=lm_loss,\n logits=prediction_scores,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n cross_attentions=outputs.cross_attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, past=None, attention_mask=None, **model_kwargs\n ):\n input_shape = input_ids.shape\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_shape)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass XBertLMHeadDecoder(BertLMHeadModel):\n \"\"\"\n This class decouples the decoder forward logic from the VL model.\n In this way, different VL models can share this decoder as long as\n they feed encoder_embeds as required.\n \"\"\"\n\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\"bert-base-uncased\", config=med_config)\n else:\n return cls(config=med_config)\n","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.XBertLMHeadDecoder","uri":"program://CREMA/class/lavis.models.med.XBertLMHeadDecoder#L1298-L1371","kind":"class","name":"XBertLMHeadDecoder","path":"lavis/models/med.py","language":"python","start_line":1298,"end_line":1371,"context_start_line":1278,"context_end_line":1391,"code":" return {\n \"input_ids\": input_ids,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass XBertLMHeadDecoder(BertLMHeadModel):\n \"\"\"\n This class decouples the decoder forward logic from the VL model.\n In this way, different VL models can share this decoder as long as\n they feed encoder_embeds as required.\n \"\"\"\n\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\"bert-base-uncased\", config=med_config)\n else:\n return cls(config=med_config)\n\n def generate_from_encoder(\n self,\n tokenized_prompt,\n visual_embeds,\n sep_token_id,\n pad_token_id,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n **kwargs\n ):\n\n if not use_nucleus_sampling:\n num_beams = num_beams\n visual_embeds = visual_embeds.repeat_interleave(num_beams, dim=0)\n\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": visual_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n if use_nucleus_sampling:\n # nucleus sampling\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n do_sample=True,\n top_p=top_p,\n num_return_sequences=1,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=1.1,\n **model_kwargs\n )\n else:\n # beam search\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n return outputs\n\n\nclass XBertEncoder(BertModel, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\n \"bert-base-uncased\", config=med_config, add_pooling_layer=False\n )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.XBertEncoder","uri":"program://CREMA/class/lavis.models.med.XBertEncoder#L1374-L1416","kind":"class","name":"XBertEncoder","path":"lavis/models/med.py","language":"python","start_line":1374,"end_line":1416,"context_start_line":1354,"context_end_line":1416,"code":" pad_token_id=pad_token_id,\n repetition_penalty=1.1,\n **model_kwargs\n )\n else:\n # beam search\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n return outputs\n\n\nclass XBertEncoder(BertModel, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\n \"bert-base-uncased\", config=med_config, add_pooling_layer=False\n )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = tokenized_text\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=visual_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n return text_output\n\n def forward_text(self, tokenized_text, **kwargs):\n text = tokenized_text\n token_type_ids = kwargs.get(\"token_type_ids\", None)\n\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n token_type_ids=token_type_ids,\n return_dict=True,\n mode=\"text\",\n )\n\n return text_output","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.__init__","uri":"program://CREMA/function/lavis.models.med.__init__#L1136-L1142","kind":"function","name":"__init__","path":"lavis/models/med.py","language":"python","start_line":1136,"end_line":1142,"context_start_line":1116,"context_end_line":1162,"code":" attention_mask = torch.cat(\n [attention_mask, attention_mask.new_zeros((attention_mask.shape[0], 1))],\n dim=-1,\n )\n dummy_token = torch.full(\n (effective_batch_size, 1),\n self.config.pad_token_id,\n dtype=torch.long,\n device=input_ids.device,\n )\n input_ids = torch.cat([input_ids, dummy_token], dim=1)\n\n return {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.forward","uri":"program://CREMA/function/lavis.models.med.forward#L1150-L1264","kind":"function","name":"forward","path":"lavis/models/med.py","language":"python","start_line":1150,"end_line":1264,"context_start_line":1130,"context_end_line":1284,"code":"\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=True,\n reduction=\"mean\",\n mode=\"multimodal\",\n soft_labels=None,\n alpha=0,\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in\n ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are\n ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n Returns:\n Example::\n >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig\n >>> import torch\n >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\n >>> config = BertConfig.from_pretrained(\"bert-base-cased\")\n >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)\n >>> inputs = tokenizer(\"Hello, my dog is cute\", return_tensors=\"pt\")\n >>> outputs = model(**inputs)\n >>> prediction_logits = outputs.logits\n \"\"\"\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n if labels is not None:\n use_cache = False\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n position_ids=position_ids,\n head_mask=head_mask,\n inputs_embeds=inputs_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n mode=mode,\n )\n\n sequence_output = outputs[0]\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores[:, :-1, :].contiguous()\n\n lm_loss = None\n if labels is not None:\n # we are doing next-token prediction; shift prediction scores and input ids by one\n shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()\n labels = labels[:, 1:].contiguous()\n loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)\n lm_loss = loss_fct(\n shifted_prediction_scores.view(-1, self.config.vocab_size),\n labels.view(-1),\n )\n if reduction == \"none\":\n lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)\n\n if soft_labels is not None:\n loss_distill = -torch.sum(\n F.log_softmax(shifted_prediction_scores, dim=-1) * soft_labels, dim=-1\n )\n loss_distill = (loss_distill * (labels != -100)).sum(1)\n lm_loss = (1 - alpha) * lm_loss + alpha * loss_distill\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return ((lm_loss,) + output) if lm_loss is not None else output\n\n return CausalLMOutputWithCrossAttentions(\n loss=lm_loss,\n logits=prediction_scores,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n cross_attentions=outputs.cross_attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, past=None, attention_mask=None, **model_kwargs\n ):\n input_shape = input_ids.shape\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_shape)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.save_attn_gradients","uri":"program://CREMA/function/lavis.models.med.save_attn_gradients#L164-L165","kind":"function","name":"save_attn_gradients","path":"lavis/models/med.py","language":"python","start_line":164,"end_line":165,"context_start_line":144,"context_end_line":185,"code":" self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.get_attn_gradients","uri":"program://CREMA/function/lavis.models.med.get_attn_gradients#L167-L168","kind":"function","name":"get_attn_gradients","path":"lavis/models/med.py","language":"python","start_line":167,"end_line":168,"context_start_line":147,"context_end_line":188,"code":" self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.save_attention_map","uri":"program://CREMA/function/lavis.models.med.save_attention_map#L170-L171","kind":"function","name":"save_attention_map","path":"lavis/models/med.py","language":"python","start_line":170,"end_line":171,"context_start_line":150,"context_end_line":191,"code":" self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.get_attention_map","uri":"program://CREMA/function/lavis.models.med.get_attention_map#L173-L174","kind":"function","name":"get_attention_map","path":"lavis/models/med.py","language":"python","start_line":173,"end_line":174,"context_start_line":153,"context_end_line":194,"code":" )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.transpose_for_scores","uri":"program://CREMA/function/lavis.models.med.transpose_for_scores#L176-L182","kind":"function","name":"transpose_for_scores","path":"lavis/models/med.py","language":"python","start_line":176,"end_line":182,"context_start_line":156,"context_end_line":202,"code":" or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.prune_heads","uri":"program://CREMA/function/lavis.models.med.prune_heads#L313-L334","kind":"function","name":"prune_heads","path":"lavis/models/med.py","language":"python","start_line":313,"end_line":334,"context_start_line":293,"context_end_line":354,"code":" def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.feed_forward_chunk","uri":"program://CREMA/function/lavis.models.med.feed_forward_chunk#L499-L502","kind":"function","name":"feed_forward_chunk","path":"lavis/models/med.py","language":"python","start_line":499,"end_line":502,"context_start_line":479,"context_end_line":522,"code":" encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med._init_weights","uri":"program://CREMA/function/lavis.models.med._init_weights#L705-L715","kind":"function","name":"_init_weights","path":"lavis/models/med.py","language":"python","start_line":705,"end_line":715,"context_start_line":685,"context_end_line":735,"code":"class BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.get_input_embeddings","uri":"program://CREMA/function/lavis.models.med.get_input_embeddings#L740-L741","kind":"function","name":"get_input_embeddings","path":"lavis/models/med.py","language":"python","start_line":740,"end_line":741,"context_start_line":720,"context_end_line":761,"code":" The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.set_input_embeddings","uri":"program://CREMA/function/lavis.models.med.set_input_embeddings#L743-L744","kind":"function","name":"set_input_embeddings","path":"lavis/models/med.py","language":"python","start_line":743,"end_line":744,"context_start_line":723,"context_end_line":764,"code":" Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med._prune_heads","uri":"program://CREMA/function/lavis.models.med._prune_heads#L746-L752","kind":"function","name":"_prune_heads","path":"lavis/models/med.py","language":"python","start_line":746,"end_line":752,"context_start_line":726,"context_end_line":772,"code":" \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.get_extended_attention_mask","uri":"program://CREMA/function/lavis.models.med.get_extended_attention_mask#L754-L830","kind":"function","name":"get_extended_attention_mask","path":"lavis/models/med.py","language":"python","start_line":754,"end_line":830,"context_start_line":734,"context_end_line":850,"code":" self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n # in case past_key_values are used we need to add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, seq_length, prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n token_type_ids=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n ):","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.get_output_embeddings","uri":"program://CREMA/function/lavis.models.med.get_output_embeddings#L1144-L1145","kind":"function","name":"get_output_embeddings","path":"lavis/models/med.py","language":"python","start_line":1144,"end_line":1145,"context_start_line":1124,"context_end_line":1165,"code":" device=input_ids.device,\n )\n input_ids = torch.cat([input_ids, dummy_token], dim=1)\n\n return {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.set_output_embeddings","uri":"program://CREMA/function/lavis.models.med.set_output_embeddings#L1147-L1148","kind":"function","name":"set_output_embeddings","path":"lavis/models/med.py","language":"python","start_line":1147,"end_line":1148,"context_start_line":1127,"context_end_line":1168,"code":"\n return {\"input_ids\": input_ids, \"attention_mask\": attention_mask}\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=True,\n reduction=\"mean\",\n mode=\"multimodal\",","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.prepare_inputs_for_generation","uri":"program://CREMA/function/lavis.models.med.prepare_inputs_for_generation#L1266-L1285","kind":"function","name":"prepare_inputs_for_generation","path":"lavis/models/med.py","language":"python","start_line":1266,"end_line":1285,"context_start_line":1246,"context_end_line":1305,"code":" if soft_labels is not None:\n loss_distill = -torch.sum(\n F.log_softmax(shifted_prediction_scores, dim=-1) * soft_labels, dim=-1\n )\n loss_distill = (loss_distill * (labels != -100)).sum(1)\n lm_loss = (1 - alpha) * lm_loss + alpha * loss_distill\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return ((lm_loss,) + output) if lm_loss is not None else output\n\n return CausalLMOutputWithCrossAttentions(\n loss=lm_loss,\n logits=prediction_scores,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n cross_attentions=outputs.cross_attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, past=None, attention_mask=None, **model_kwargs\n ):\n input_shape = input_ids.shape\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_shape)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass XBertLMHeadDecoder(BertLMHeadModel):\n \"\"\"\n This class decouples the decoder forward logic from the VL model.\n In this way, different VL models can share this decoder as long as\n they feed encoder_embeds as required.\n \"\"\"\n\n @classmethod","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med._reorder_cache","uri":"program://CREMA/function/lavis.models.med._reorder_cache#L1287-L1295","kind":"function","name":"_reorder_cache","path":"lavis/models/med.py","language":"python","start_line":1287,"end_line":1295,"context_start_line":1267,"context_end_line":1315,"code":" self, input_ids, past=None, attention_mask=None, **model_kwargs\n ):\n input_shape = input_ids.shape\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_shape)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass XBertLMHeadDecoder(BertLMHeadModel):\n \"\"\"\n This class decouples the decoder forward logic from the VL model.\n In this way, different VL models can share this decoder as long as\n they feed encoder_embeds as required.\n \"\"\"\n\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\"bert-base-uncased\", config=med_config)\n else:\n return cls(config=med_config)\n","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.from_config","uri":"program://CREMA/function/lavis.models.med.from_config#L1376-L1386","kind":"function","name":"from_config","path":"lavis/models/med.py","language":"python","start_line":1376,"end_line":1386,"context_start_line":1356,"context_end_line":1406,"code":" **model_kwargs\n )\n else:\n # beam search\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n return outputs\n\n\nclass XBertEncoder(BertModel, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\n \"bert-base-uncased\", config=med_config, add_pooling_layer=False\n )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = tokenized_text\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=visual_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n return text_output\n\n def forward_text(self, tokenized_text, **kwargs):\n text = tokenized_text\n token_type_ids = kwargs.get(\"token_type_ids\", None)","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.generate_from_encoder","uri":"program://CREMA/function/lavis.models.med.generate_from_encoder#L1316-L1371","kind":"function","name":"generate_from_encoder","path":"lavis/models/med.py","language":"python","start_line":1316,"end_line":1371,"context_start_line":1296,"context_end_line":1391,"code":"\n\nclass XBertLMHeadDecoder(BertLMHeadModel):\n \"\"\"\n This class decouples the decoder forward logic from the VL model.\n In this way, different VL models can share this decoder as long as\n they feed encoder_embeds as required.\n \"\"\"\n\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\"bert-base-uncased\", config=med_config)\n else:\n return cls(config=med_config)\n\n def generate_from_encoder(\n self,\n tokenized_prompt,\n visual_embeds,\n sep_token_id,\n pad_token_id,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n **kwargs\n ):\n\n if not use_nucleus_sampling:\n num_beams = num_beams\n visual_embeds = visual_embeds.repeat_interleave(num_beams, dim=0)\n\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": visual_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n if use_nucleus_sampling:\n # nucleus sampling\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n do_sample=True,\n top_p=top_p,\n num_return_sequences=1,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=1.1,\n **model_kwargs\n )\n else:\n # beam search\n outputs = self.generate(\n input_ids=tokenized_prompt.input_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=sep_token_id,\n pad_token_id=pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n return outputs\n\n\nclass XBertEncoder(BertModel, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\n \"bert-base-uncased\", config=med_config, add_pooling_layer=False\n )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.forward_automask","uri":"program://CREMA/function/lavis.models.med.forward_automask#L1388-L1402","kind":"function","name":"forward_automask","path":"lavis/models/med.py","language":"python","start_line":1388,"end_line":1402,"context_start_line":1368,"context_end_line":1416,"code":" **model_kwargs\n )\n\n return outputs\n\n\nclass XBertEncoder(BertModel, BaseEncoder):\n @classmethod\n def from_config(cls, cfg, from_pretrained=False):\n\n med_config_path = get_abs_path(cfg.get(\"med_config_path\"))\n med_config = BertConfig.from_json_file(med_config_path)\n\n if from_pretrained:\n return cls.from_pretrained(\n \"bert-base-uncased\", config=med_config, add_pooling_layer=False\n )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = tokenized_text\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=visual_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n return text_output\n\n def forward_text(self, tokenized_text, **kwargs):\n text = tokenized_text\n token_type_ids = kwargs.get(\"token_type_ids\", None)\n\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n token_type_ids=token_type_ids,\n return_dict=True,\n mode=\"text\",\n )\n\n return text_output","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.forward_text","uri":"program://CREMA/function/lavis.models.med.forward_text#L1404-L1416","kind":"function","name":"forward_text","path":"lavis/models/med.py","language":"python","start_line":1404,"end_line":1416,"context_start_line":1384,"context_end_line":1416,"code":" )\n else:\n return cls(config=med_config, add_pooling_layer=False)\n\n def forward_automask(self, tokenized_text, visual_embeds, **kwargs):\n image_atts = torch.ones(visual_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = tokenized_text\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=visual_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n return text_output\n\n def forward_text(self, tokenized_text, **kwargs):\n text = tokenized_text\n token_type_ids = kwargs.get(\"token_type_ids\", None)\n\n text_output = super().forward(\n text.input_ids,\n attention_mask=text.attention_mask,\n token_type_ids=token_type_ids,\n return_dict=True,\n mode=\"text\",\n )\n\n return text_output","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.create_custom_forward","uri":"program://CREMA/function/lavis.models.med.create_custom_forward#L576-L580","kind":"function","name":"create_custom_forward","path":"lavis/models/med.py","language":"python","start_line":576,"end_line":580,"context_start_line":556,"context_end_line":600,"code":"\n # compatibility for ALBEF and BLIP\n # for i in range(self.config.num_hidden_layers):\n for i in range(start_layer, output_layer):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n # TODO pay attention to this.\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n mode=mode,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.med.custom_forward","uri":"program://CREMA/function/lavis.models.med.custom_forward#L577-L578","kind":"function","name":"custom_forward","path":"lavis/models/med.py","language":"python","start_line":577,"end_line":578,"context_start_line":557,"context_end_line":598,"code":" # compatibility for ALBEF and BLIP\n # for i in range(self.config.num_hidden_layers):\n for i in range(start_layer, output_layer):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n # TODO pay attention to this.\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,","source_hash":"5857ec781f3e58e6d8c463d37902817b07fccc87d78817c84b6d1045337ab3d2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats_encoder","uri":"program://CREMA/module/lavis.models.beats_encoder#L1-L47","kind":"module","name":"lavis.models.beats_encoder","path":"lavis/models/beats_encoder.py","language":"python","start_line":1,"end_line":47,"context_start_line":1,"context_end_line":47,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.models.base_model import BaseEncoder\nfrom lavis.models.beats.BEATs import BEATs, BEATsConfig\nimport torch \nfrom lavis.common.utils import is_url\nfrom lavis.common.dist_utils import download_cached_file\nimport os \n\n\n# ckp_path = \"https://valle.blob.core.windows.net/share/BEATs/BEATs_iter3_plus_AS2M.pt?sv=2020-08-04&st=2023-03-01T07%3A51%3A05Z&se=2033-03-02T07%3A51%3A00Z&sr=c&sp=rl&sig=QJXmSJG9DbMKf48UDIU1MfzIro8HQOf3sqlNXiflY1I%3D\"\n\nckp_path = '/nas-hdd/shoubin/pretrained_model/BEATs_iter3_plus_AS2M.pt'\n\nclass BeatsEncoder(BaseEncoder):\n def __init__(self, checkpoint_path=ckp_path):\n super().__init__()\n \n # load the pre-trained checkpoints\n if is_url(checkpoint_path):\n cached_file = download_cached_file(\n checkpoint_path, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file)\n elif os.path.isfile(checkpoint_path):\n print('loading checkpoint for BEATs Encoder')\n checkpoint = torch.load(checkpoint_path)\n\n cfg = BEATsConfig(checkpoint['cfg'])\n self.num_features = cfg.encoder_embed_dim\n self.model = BEATs(cfg)\n self.model.load_state_dict(checkpoint['model'])\n self.model.eval()\n\n @classmethod\n def from_config(cls, cfg):\n checkpoint_path = cfg.get(\"checkpoint_path\", ckp_path)\n return cls(checkpoint_path)\n\n def forward(self, x):\n with torch.no_grad():\n return self.model.extract_features(x.squeeze(1))[0]","source_hash":"dbcbada9277a164e24274b6cf22eab015aad1c050161d8e89de8ef68936c07cb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats_encoder.BeatsEncoder","uri":"program://CREMA/class/lavis.models.beats_encoder.BeatsEncoder#L20-L47","kind":"class","name":"BeatsEncoder","path":"lavis/models/beats_encoder.py","language":"python","start_line":20,"end_line":47,"context_start_line":1,"context_end_line":47,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.models.base_model import BaseEncoder\nfrom lavis.models.beats.BEATs import BEATs, BEATsConfig\nimport torch \nfrom lavis.common.utils import is_url\nfrom lavis.common.dist_utils import download_cached_file\nimport os \n\n\n# ckp_path = \"https://valle.blob.core.windows.net/share/BEATs/BEATs_iter3_plus_AS2M.pt?sv=2020-08-04&st=2023-03-01T07%3A51%3A05Z&se=2033-03-02T07%3A51%3A00Z&sr=c&sp=rl&sig=QJXmSJG9DbMKf48UDIU1MfzIro8HQOf3sqlNXiflY1I%3D\"\n\nckp_path = '/nas-hdd/shoubin/pretrained_model/BEATs_iter3_plus_AS2M.pt'\n\nclass BeatsEncoder(BaseEncoder):\n def __init__(self, checkpoint_path=ckp_path):\n super().__init__()\n \n # load the pre-trained checkpoints\n if is_url(checkpoint_path):\n cached_file = download_cached_file(\n checkpoint_path, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file)\n elif os.path.isfile(checkpoint_path):\n print('loading checkpoint for BEATs Encoder')\n checkpoint = torch.load(checkpoint_path)\n\n cfg = BEATsConfig(checkpoint['cfg'])\n self.num_features = cfg.encoder_embed_dim\n self.model = BEATs(cfg)\n self.model.load_state_dict(checkpoint['model'])\n self.model.eval()\n\n @classmethod\n def from_config(cls, cfg):\n checkpoint_path = cfg.get(\"checkpoint_path\", ckp_path)\n return cls(checkpoint_path)\n\n def forward(self, x):\n with torch.no_grad():\n return self.model.extract_features(x.squeeze(1))[0]","source_hash":"dbcbada9277a164e24274b6cf22eab015aad1c050161d8e89de8ef68936c07cb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats_encoder.__init__","uri":"program://CREMA/function/lavis.models.beats_encoder.__init__#L21-L38","kind":"function","name":"__init__","path":"lavis/models/beats_encoder.py","language":"python","start_line":21,"end_line":38,"context_start_line":1,"context_end_line":47,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.models.base_model import BaseEncoder\nfrom lavis.models.beats.BEATs import BEATs, BEATsConfig\nimport torch \nfrom lavis.common.utils import is_url\nfrom lavis.common.dist_utils import download_cached_file\nimport os \n\n\n# ckp_path = \"https://valle.blob.core.windows.net/share/BEATs/BEATs_iter3_plus_AS2M.pt?sv=2020-08-04&st=2023-03-01T07%3A51%3A05Z&se=2033-03-02T07%3A51%3A00Z&sr=c&sp=rl&sig=QJXmSJG9DbMKf48UDIU1MfzIro8HQOf3sqlNXiflY1I%3D\"\n\nckp_path = '/nas-hdd/shoubin/pretrained_model/BEATs_iter3_plus_AS2M.pt'\n\nclass BeatsEncoder(BaseEncoder):\n def __init__(self, checkpoint_path=ckp_path):\n super().__init__()\n \n # load the pre-trained checkpoints\n if is_url(checkpoint_path):\n cached_file = download_cached_file(\n checkpoint_path, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file)\n elif os.path.isfile(checkpoint_path):\n print('loading checkpoint for BEATs Encoder')\n checkpoint = torch.load(checkpoint_path)\n\n cfg = BEATsConfig(checkpoint['cfg'])\n self.num_features = cfg.encoder_embed_dim\n self.model = BEATs(cfg)\n self.model.load_state_dict(checkpoint['model'])\n self.model.eval()\n\n @classmethod\n def from_config(cls, cfg):\n checkpoint_path = cfg.get(\"checkpoint_path\", ckp_path)\n return cls(checkpoint_path)\n\n def forward(self, x):\n with torch.no_grad():\n return self.model.extract_features(x.squeeze(1))[0]","source_hash":"dbcbada9277a164e24274b6cf22eab015aad1c050161d8e89de8ef68936c07cb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats_encoder.from_config","uri":"program://CREMA/function/lavis.models.beats_encoder.from_config#L41-L43","kind":"function","name":"from_config","path":"lavis/models/beats_encoder.py","language":"python","start_line":41,"end_line":43,"context_start_line":21,"context_end_line":47,"code":" def __init__(self, checkpoint_path=ckp_path):\n super().__init__()\n \n # load the pre-trained checkpoints\n if is_url(checkpoint_path):\n cached_file = download_cached_file(\n checkpoint_path, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file)\n elif os.path.isfile(checkpoint_path):\n print('loading checkpoint for BEATs Encoder')\n checkpoint = torch.load(checkpoint_path)\n\n cfg = BEATsConfig(checkpoint['cfg'])\n self.num_features = cfg.encoder_embed_dim\n self.model = BEATs(cfg)\n self.model.load_state_dict(checkpoint['model'])\n self.model.eval()\n\n @classmethod\n def from_config(cls, cfg):\n checkpoint_path = cfg.get(\"checkpoint_path\", ckp_path)\n return cls(checkpoint_path)\n\n def forward(self, x):\n with torch.no_grad():\n return self.model.extract_features(x.squeeze(1))[0]","source_hash":"dbcbada9277a164e24274b6cf22eab015aad1c050161d8e89de8ef68936c07cb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats_encoder.forward","uri":"program://CREMA/function/lavis.models.beats_encoder.forward#L45-L47","kind":"function","name":"forward","path":"lavis/models/beats_encoder.py","language":"python","start_line":45,"end_line":47,"context_start_line":25,"context_end_line":47,"code":" if is_url(checkpoint_path):\n cached_file = download_cached_file(\n checkpoint_path, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file)\n elif os.path.isfile(checkpoint_path):\n print('loading checkpoint for BEATs Encoder')\n checkpoint = torch.load(checkpoint_path)\n\n cfg = BEATsConfig(checkpoint['cfg'])\n self.num_features = cfg.encoder_embed_dim\n self.model = BEATs(cfg)\n self.model.load_state_dict(checkpoint['model'])\n self.model.eval()\n\n @classmethod\n def from_config(cls, cfg):\n checkpoint_path = cfg.get(\"checkpoint_path\", ckp_path)\n return cls(checkpoint_path)\n\n def forward(self, x):\n with torch.no_grad():\n return self.model.extract_features(x.squeeze(1))[0]","source_hash":"dbcbada9277a164e24274b6cf22eab015aad1c050161d8e89de8ef68936c07cb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone","uri":"program://CREMA/module/lavis.models.beats.backbone#L1-L783","kind":"module","name":"lavis.models.beats.backbone","path":"lavis/models/beats/backbone.py","language":"python","start_line":1,"end_line":783,"context_start_line":1,"context_end_line":783,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\nimport math\nimport numpy as np\nfrom typing import Dict, Optional, Tuple\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\nfrom torch.nn import LayerNorm, Parameter\nfrom lavis.models.beats.modules import (\n GradMultiply,\n SamePad,\n get_activation_fn,\n GLU_Linear,\n quant_noise,\n)\n\n\nclass TransformerEncoder(nn.Module):\n def __init__(self, args):\n super().__init__()\n\n self.dropout = args.dropout\n self.embedding_dim = args.encoder_embed_dim\n\n self.pos_conv = nn.Conv1d(\n self.embedding_dim,\n self.embedding_dim,\n kernel_size=args.conv_pos,\n padding=args.conv_pos // 2,\n groups=args.conv_pos_groups,\n )\n dropout = 0\n std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim))\n nn.init.normal_(self.pos_conv.weight, mean=0, std=std)\n nn.init.constant_(self.pos_conv.bias, 0)\n\n self.pos_conv = nn.utils.weight_norm(self.pos_conv, name=\"weight\", dim=2)\n self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU())\n\n if hasattr(args, \"relative_position_embedding\"):\n self.relative_position_embedding = args.relative_position_embedding\n self.num_buckets = args.num_buckets\n self.max_distance = args.max_distance\n else:\n self.relative_position_embedding = False\n self.num_buckets = 0\n self.max_distance = 0\n\n self.layers = nn.ModuleList(\n [\n TransformerSentenceEncoderLayer(\n embedding_dim=self.embedding_dim,\n ffn_embedding_dim=args.encoder_ffn_embed_dim,\n num_attention_heads=args.encoder_attention_heads,\n dropout=self.dropout,\n attention_dropout=args.attention_dropout,\n activation_dropout=args.activation_dropout,\n activation_fn=args.activation_fn,\n layer_norm_first=args.layer_norm_first,\n deep_norm=args.deep_norm,\n has_relative_attention_bias=self.relative_position_embedding,\n num_buckets=self.num_buckets,\n max_distance=self.max_distance,\n gru_rel_pos=args.gru_rel_pos,\n encoder_layers=args.encoder_layers,\n )\n for i in range(args.encoder_layers)\n ]\n )\n if self.relative_position_embedding:\n for i in range(1, args.encoder_layers):\n del self.layers[i].self_attn.relative_attention_bias\n self.layers[i].self_attn.relative_attention_bias = self.layers[0].self_attn.relative_attention_bias\n\n self.layer_norm_first = args.layer_norm_first\n self.layer_norm = LayerNorm(self.embedding_dim)\n self.layerdrop = args.encoder_layerdrop\n\n self.apply(init_bert_params)\n\n if args.deep_norm:\n deep_norm_beta = math.pow(8 * args.encoder_layers, -1 / 4)\n for i in range(args.encoder_layers):\n nn.init.xavier_normal_(self.layers[i].self_attn.k_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.v_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].self_attn.q_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.out_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc1.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc2.weight, gain=deep_norm_beta)\n\n self.layer_wise_gradient_decay_ratio = getattr(args, \"layer_wise_gradient_decay_ratio\", 1)\n\n def forward(self, x, padding_mask=None, layer=None):\n x, layer_results = self.extract_features(x, padding_mask, layer)\n\n if self.layer_norm_first and layer is None:\n x = self.layer_norm(x)\n\n return x, layer_results\n\n def extract_features(self, x, padding_mask=None, tgt_layer=None):\n\n if padding_mask is not None:\n x[padding_mask] = 0\n\n x_conv = self.pos_conv(x.transpose(1, 2))\n x_conv = x_conv.transpose(1, 2)\n x = x + x_conv\n\n if not self.layer_norm_first:\n x = self.layer_norm(x)\n\n x = F.dropout(x, p=self.dropout, training=self.training)\n\n # B x T x C -> T x B x C\n x = x.transpose(0, 1)\n\n layer_results = []\n z = None\n if tgt_layer is not None:\n layer_results.append((x, z))\n r = None\n pos_bias = None\n for i, layer in enumerate(self.layers):\n if self.layer_wise_gradient_decay_ratio != 1.0:\n x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)\n dropout_probability = np.random.random()\n if not self.training or (dropout_probability > self.layerdrop):\n x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)\n if tgt_layer is not None:\n layer_results.append((x, z))\n if i == tgt_layer:\n r = x\n break\n\n if r is not None:\n x = r\n\n # T x B x C -> B x T x C\n x = x.transpose(0, 1)\n\n return x, layer_results\n\n\nclass TransformerSentenceEncoderLayer(nn.Module):\n def __init__(\n self,\n embedding_dim: float = 768,\n ffn_embedding_dim: float = 3072,\n num_attention_heads: float = 8,\n dropout: float = 0.1,\n attention_dropout: float = 0.1,\n activation_dropout: float = 0.1,\n activation_fn: str = \"relu\",\n layer_norm_first: bool = False,\n deep_norm: bool = False,\n has_relative_attention_bias: bool = False,\n num_buckets: int = 0,\n max_distance: int = 0,\n rescale_init: bool = False,\n gru_rel_pos: bool = False,\n encoder_layers: int = 0,\n ) -> None:\n\n super().__init__()\n self.embedding_dim = embedding_dim\n self.dropout = dropout\n self.activation_dropout = activation_dropout\n\n self.activation_name = activation_fn\n self.activation_fn = get_activation_fn(activation_fn)\n self.self_attn = MultiheadAttention(\n self.embedding_dim,\n num_attention_heads,\n dropout=attention_dropout,\n self_attention=True,\n has_relative_attention_bias=has_relative_attention_bias,\n num_buckets=num_buckets,\n max_distance=max_distance,\n rescale_init=rescale_init,\n gru_rel_pos=gru_rel_pos,\n )\n\n self.dropout1 = nn.Dropout(dropout)\n self.dropout2 = nn.Dropout(self.activation_dropout)\n self.dropout3 = nn.Dropout(dropout)\n\n self.layer_norm_first = layer_norm_first\n\n self.self_attn_layer_norm = LayerNorm(self.embedding_dim)\n\n if self.activation_name == \"glu\":\n self.fc1 = GLU_Linear(self.embedding_dim, ffn_embedding_dim, \"swish\")\n else:\n self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)\n self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)\n\n self.final_layer_norm = LayerNorm(self.embedding_dim)\n\n self.deep_norm = deep_norm\n if self.deep_norm:\n self.deep_norm_alpha = math.pow(2 * encoder_layers, 1 / 4)\n else:\n self.deep_norm_alpha = 1\n\n def forward(\n self,\n x: torch.Tensor,\n self_attn_mask: torch.Tensor = None,\n self_attn_padding_mask: torch.Tensor = None,\n need_weights: bool = False,\n pos_bias=None\n ):\n residual = x\n\n if self.layer_norm_first:\n x = self.self_attn_layer_norm(x)\n x, attn, pos_bias = self.self_attn(\n query=x,\n key=x,\n value=x,\n key_padding_mask=self_attn_padding_mask,\n need_weights=False,\n attn_mask=self_attn_mask,\n position_bias=pos_bias\n )\n x = self.dropout1(x)\n x = residual + x\n\n residual = x\n x = self.final_layer_norm(x)\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual + x\n else:\n x, attn, pos_bias = self.self_attn(\n query=x,\n key=x,\n value=x,\n key_padding_mask=self_attn_padding_mask,\n need_weights=need_weights,\n attn_mask=self_attn_mask,\n position_bias=pos_bias\n )\n\n x = self.dropout1(x)\n x = residual * self.deep_norm_alpha + x\n\n x = self.self_attn_layer_norm(x)\n\n residual = x\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual * self.deep_norm_alpha + x\n x = self.final_layer_norm(x)\n\n return x, attn, pos_bias\n\n\nclass MultiheadAttention(nn.Module):\n \"\"\"Multi-headed attention.\n\n See \"Attention Is All You Need\" for more details.\n \"\"\"\n\n def __init__(\n self,\n embed_dim,\n num_heads,\n kdim=None,\n vdim=None,\n dropout=0.0,\n bias=True,\n add_bias_kv=False,\n add_zero_attn=False,\n self_attention=False,\n encoder_decoder_attention=False,\n q_noise=0.0,\n qn_block_size=8,\n has_relative_attention_bias=False,\n num_buckets=32,\n max_distance=128,\n gru_rel_pos=False,\n rescale_init=False,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.kdim = kdim if kdim is not None else embed_dim\n self.vdim = vdim if vdim is not None else embed_dim\n self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim\n\n self.num_heads = num_heads\n self.dropout_module = nn.Dropout(dropout)\n\n self.has_relative_attention_bias = has_relative_attention_bias\n self.num_buckets = num_buckets\n self.max_distance = max_distance\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)\n\n self.head_dim = embed_dim // num_heads\n self.q_head_dim = self.head_dim\n self.k_head_dim = self.head_dim\n assert (\n self.head_dim * num_heads == self.embed_dim\n ), \"embed_dim must be divisible by num_heads\"\n self.scaling = self.head_dim ** -0.5\n\n self.self_attention = self_attention\n self.encoder_decoder_attention = encoder_decoder_attention\n\n assert not self.self_attention or self.qkv_same_dim, (\n \"Self-attention requires query, key and \" \"value to be of the same size\"\n )\n\n k_bias = True\n if rescale_init:\n k_bias = False\n\n k_embed_dim = embed_dim\n q_embed_dim = embed_dim\n\n self.k_proj = quant_noise(\n nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size\n )\n self.v_proj = quant_noise(\n nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n self.q_proj = quant_noise(\n nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n self.out_proj = quant_noise(\n nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n if add_bias_kv:\n self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))\n self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))\n else:\n self.bias_k = self.bias_v = None\n\n self.add_zero_attn = add_zero_attn\n\n self.gru_rel_pos = gru_rel_pos\n if self.gru_rel_pos:\n self.grep_linear = nn.Linear(self.q_head_dim, 8)\n self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))\n\n self.reset_parameters()\n\n def reset_parameters(self):\n if self.qkv_same_dim:\n # Empirically observed the convergence to be much better with\n # the scaled initialization\n nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))\n else:\n nn.init.xavier_uniform_(self.k_proj.weight)\n nn.init.xavier_uniform_(self.v_proj.weight)\n nn.init.xavier_uniform_(self.q_proj.weight)\n\n nn.init.xavier_uniform_(self.out_proj.weight)\n if self.out_proj.bias is not None:\n nn.init.constant_(self.out_proj.bias, 0.0)\n if self.bias_k is not None:\n nn.init.xavier_normal_(self.bias_k)\n if self.bias_v is not None:\n nn.init.xavier_normal_(self.bias_v)\n if self.has_relative_attention_bias:\n nn.init.xavier_normal_(self.relative_attention_bias.weight)\n\n def _relative_positions_bucket(self, relative_positions, bidirectional=True):\n num_buckets = self.num_buckets\n max_distance = self.max_distance\n relative_buckets = 0\n\n if bidirectional:\n num_buckets = num_buckets // 2\n relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets\n relative_positions = torch.abs(relative_positions)\n else:\n relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))\n\n max_exact = num_buckets // 2\n is_small = relative_positions < max_exact\n\n relative_postion_if_large = max_exact + (\n torch.log(relative_positions.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_postion_if_large = torch.min(\n relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)\n )\n\n relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)\n return relative_buckets\n\n def compute_bias(self, query_length, key_length):\n context_position = torch.arange(query_length, dtype=torch.long)[:, None]\n memory_position = torch.arange(key_length, dtype=torch.long)[None, :]\n relative_position = memory_position - context_position\n relative_position_bucket = self._relative_positions_bucket(\n relative_position,\n bidirectional=True\n )\n relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)\n values = self.relative_attention_bias(relative_position_bucket)\n values = values.permute([2, 0, 1])\n return values\n\n def forward(\n self,\n query,\n key: Optional[Tensor],\n value: Optional[Tensor],\n key_padding_mask: Optional[Tensor] = None,\n incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,\n need_weights: bool = True,\n static_kv: bool = False,\n attn_mask: Optional[Tensor] = None,\n before_softmax: bool = False,\n need_head_weights: bool = False,\n position_bias: Optional[Tensor] = None\n ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:\n \"\"\"Input shape: Time x Batch x Channel\n\n Args:\n key_padding_mask (ByteTensor, optional): mask to exclude\n keys that are pads, of shape `(batch, src_len)`, where\n padding elements are indicated by 1s.\n need_weights (bool, optional): return the attention weights,\n averaged over heads (default: False).\n attn_mask (ByteTensor, optional): typically used to\n implement causal attention, where the mask prevents the\n attention from looking forward in time (default: None).\n before_softmax (bool, optional): return the raw attention\n weights and values before the attention softmax.\n need_head_weights (bool, optional): return the attention\n weights for each head. Implies *need_weights*. Default:\n return the average attention weights over all heads.\n \"\"\"\n if need_head_weights:\n need_weights = True\n\n is_tpu = query.device.type == \"xla\"\n\n tgt_len, bsz, embed_dim = query.size()\n src_len = tgt_len\n assert embed_dim == self.embed_dim\n assert list(query.size()) == [tgt_len, bsz, embed_dim]\n if key is not None:\n src_len, key_bsz, _ = key.size()\n if not torch.jit.is_scripting():\n assert key_bsz == bsz\n assert value is not None\n assert src_len, bsz == value.shape[:2]\n\n if self.has_relative_attention_bias and position_bias is None:\n position_bias = self.compute_bias(tgt_len, src_len)\n position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)\n\n if incremental_state is not None:\n saved_state = self._get_input_buffer(incremental_state)\n if saved_state is not None and \"prev_key\" in saved_state:\n # previous time steps are cached - no need to recompute\n # key and value if they are static\n if static_kv:\n assert self.encoder_decoder_attention and not self.self_attention\n key = value = None\n else:\n saved_state = None\n\n if self.self_attention:\n q = self.q_proj(query)\n k = self.k_proj(query)\n v = self.v_proj(query)\n elif self.encoder_decoder_attention:\n # encoder-decoder attention\n q = self.q_proj(query)\n if key is None:\n assert value is None\n k = v = None\n else:\n k = self.k_proj(key)\n v = self.v_proj(key)\n\n else:\n assert key is not None and value is not None\n q = self.q_proj(query)\n k = self.k_proj(key)\n v = self.v_proj(value)\n q *= self.scaling\n alpha = 32\n q *= 1 / alpha\n\n if self.bias_k is not None:\n assert self.bias_v is not None\n k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])\n v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])\n if attn_mask is not None:\n attn_mask = torch.cat(\n [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1\n )\n if key_padding_mask is not None:\n key_padding_mask = torch.cat(\n [\n key_padding_mask,\n key_padding_mask.new_zeros(key_padding_mask.size(0), 1),\n ],\n dim=1,\n )\n\n q = (\n# ... truncated ...","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.TransformerEncoder","uri":"program://CREMA/class/lavis.models.beats.backbone.TransformerEncoder#L26-L150","kind":"class","name":"TransformerEncoder","path":"lavis/models/beats/backbone.py","language":"python","start_line":26,"end_line":150,"context_start_line":6,"context_end_line":170,"code":"# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\nimport math\nimport numpy as np\nfrom typing import Dict, Optional, Tuple\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\nfrom torch.nn import LayerNorm, Parameter\nfrom lavis.models.beats.modules import (\n GradMultiply,\n SamePad,\n get_activation_fn,\n GLU_Linear,\n quant_noise,\n)\n\n\nclass TransformerEncoder(nn.Module):\n def __init__(self, args):\n super().__init__()\n\n self.dropout = args.dropout\n self.embedding_dim = args.encoder_embed_dim\n\n self.pos_conv = nn.Conv1d(\n self.embedding_dim,\n self.embedding_dim,\n kernel_size=args.conv_pos,\n padding=args.conv_pos // 2,\n groups=args.conv_pos_groups,\n )\n dropout = 0\n std = math.sqrt((4 * (1.0 - dropout)) / (args.conv_pos * self.embedding_dim))\n nn.init.normal_(self.pos_conv.weight, mean=0, std=std)\n nn.init.constant_(self.pos_conv.bias, 0)\n\n self.pos_conv = nn.utils.weight_norm(self.pos_conv, name=\"weight\", dim=2)\n self.pos_conv = nn.Sequential(self.pos_conv, SamePad(args.conv_pos), nn.GELU())\n\n if hasattr(args, \"relative_position_embedding\"):\n self.relative_position_embedding = args.relative_position_embedding\n self.num_buckets = args.num_buckets\n self.max_distance = args.max_distance\n else:\n self.relative_position_embedding = False\n self.num_buckets = 0\n self.max_distance = 0\n\n self.layers = nn.ModuleList(\n [\n TransformerSentenceEncoderLayer(\n embedding_dim=self.embedding_dim,\n ffn_embedding_dim=args.encoder_ffn_embed_dim,\n num_attention_heads=args.encoder_attention_heads,\n dropout=self.dropout,\n attention_dropout=args.attention_dropout,\n activation_dropout=args.activation_dropout,\n activation_fn=args.activation_fn,\n layer_norm_first=args.layer_norm_first,\n deep_norm=args.deep_norm,\n has_relative_attention_bias=self.relative_position_embedding,\n num_buckets=self.num_buckets,\n max_distance=self.max_distance,\n gru_rel_pos=args.gru_rel_pos,\n encoder_layers=args.encoder_layers,\n )\n for i in range(args.encoder_layers)\n ]\n )\n if self.relative_position_embedding:\n for i in range(1, args.encoder_layers):\n del self.layers[i].self_attn.relative_attention_bias\n self.layers[i].self_attn.relative_attention_bias = self.layers[0].self_attn.relative_attention_bias\n\n self.layer_norm_first = args.layer_norm_first\n self.layer_norm = LayerNorm(self.embedding_dim)\n self.layerdrop = args.encoder_layerdrop\n\n self.apply(init_bert_params)\n\n if args.deep_norm:\n deep_norm_beta = math.pow(8 * args.encoder_layers, -1 / 4)\n for i in range(args.encoder_layers):\n nn.init.xavier_normal_(self.layers[i].self_attn.k_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.v_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].self_attn.q_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.out_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc1.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc2.weight, gain=deep_norm_beta)\n\n self.layer_wise_gradient_decay_ratio = getattr(args, \"layer_wise_gradient_decay_ratio\", 1)\n\n def forward(self, x, padding_mask=None, layer=None):\n x, layer_results = self.extract_features(x, padding_mask, layer)\n\n if self.layer_norm_first and layer is None:\n x = self.layer_norm(x)\n\n return x, layer_results\n\n def extract_features(self, x, padding_mask=None, tgt_layer=None):\n\n if padding_mask is not None:\n x[padding_mask] = 0\n\n x_conv = self.pos_conv(x.transpose(1, 2))\n x_conv = x_conv.transpose(1, 2)\n x = x + x_conv\n\n if not self.layer_norm_first:\n x = self.layer_norm(x)\n\n x = F.dropout(x, p=self.dropout, training=self.training)\n\n # B x T x C -> T x B x C\n x = x.transpose(0, 1)\n\n layer_results = []\n z = None\n if tgt_layer is not None:\n layer_results.append((x, z))\n r = None\n pos_bias = None\n for i, layer in enumerate(self.layers):\n if self.layer_wise_gradient_decay_ratio != 1.0:\n x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)\n dropout_probability = np.random.random()\n if not self.training or (dropout_probability > self.layerdrop):\n x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)\n if tgt_layer is not None:\n layer_results.append((x, z))\n if i == tgt_layer:\n r = x\n break\n\n if r is not None:\n x = r\n\n # T x B x C -> B x T x C\n x = x.transpose(0, 1)\n\n return x, layer_results\n\n\nclass TransformerSentenceEncoderLayer(nn.Module):\n def __init__(\n self,\n embedding_dim: float = 768,\n ffn_embedding_dim: float = 3072,\n num_attention_heads: float = 8,\n dropout: float = 0.1,\n attention_dropout: float = 0.1,\n activation_dropout: float = 0.1,\n activation_fn: str = \"relu\",\n layer_norm_first: bool = False,\n deep_norm: bool = False,\n has_relative_attention_bias: bool = False,\n num_buckets: int = 0,\n max_distance: int = 0,\n rescale_init: bool = False,\n gru_rel_pos: bool = False,\n encoder_layers: int = 0,","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.TransformerSentenceEncoderLayer","uri":"program://CREMA/class/lavis.models.beats.backbone.TransformerSentenceEncoderLayer#L153-L275","kind":"class","name":"TransformerSentenceEncoderLayer","path":"lavis/models/beats/backbone.py","language":"python","start_line":153,"end_line":275,"context_start_line":133,"context_end_line":295,"code":" if self.layer_wise_gradient_decay_ratio != 1.0:\n x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)\n dropout_probability = np.random.random()\n if not self.training or (dropout_probability > self.layerdrop):\n x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)\n if tgt_layer is not None:\n layer_results.append((x, z))\n if i == tgt_layer:\n r = x\n break\n\n if r is not None:\n x = r\n\n # T x B x C -> B x T x C\n x = x.transpose(0, 1)\n\n return x, layer_results\n\n\nclass TransformerSentenceEncoderLayer(nn.Module):\n def __init__(\n self,\n embedding_dim: float = 768,\n ffn_embedding_dim: float = 3072,\n num_attention_heads: float = 8,\n dropout: float = 0.1,\n attention_dropout: float = 0.1,\n activation_dropout: float = 0.1,\n activation_fn: str = \"relu\",\n layer_norm_first: bool = False,\n deep_norm: bool = False,\n has_relative_attention_bias: bool = False,\n num_buckets: int = 0,\n max_distance: int = 0,\n rescale_init: bool = False,\n gru_rel_pos: bool = False,\n encoder_layers: int = 0,\n ) -> None:\n\n super().__init__()\n self.embedding_dim = embedding_dim\n self.dropout = dropout\n self.activation_dropout = activation_dropout\n\n self.activation_name = activation_fn\n self.activation_fn = get_activation_fn(activation_fn)\n self.self_attn = MultiheadAttention(\n self.embedding_dim,\n num_attention_heads,\n dropout=attention_dropout,\n self_attention=True,\n has_relative_attention_bias=has_relative_attention_bias,\n num_buckets=num_buckets,\n max_distance=max_distance,\n rescale_init=rescale_init,\n gru_rel_pos=gru_rel_pos,\n )\n\n self.dropout1 = nn.Dropout(dropout)\n self.dropout2 = nn.Dropout(self.activation_dropout)\n self.dropout3 = nn.Dropout(dropout)\n\n self.layer_norm_first = layer_norm_first\n\n self.self_attn_layer_norm = LayerNorm(self.embedding_dim)\n\n if self.activation_name == \"glu\":\n self.fc1 = GLU_Linear(self.embedding_dim, ffn_embedding_dim, \"swish\")\n else:\n self.fc1 = nn.Linear(self.embedding_dim, ffn_embedding_dim)\n self.fc2 = nn.Linear(ffn_embedding_dim, self.embedding_dim)\n\n self.final_layer_norm = LayerNorm(self.embedding_dim)\n\n self.deep_norm = deep_norm\n if self.deep_norm:\n self.deep_norm_alpha = math.pow(2 * encoder_layers, 1 / 4)\n else:\n self.deep_norm_alpha = 1\n\n def forward(\n self,\n x: torch.Tensor,\n self_attn_mask: torch.Tensor = None,\n self_attn_padding_mask: torch.Tensor = None,\n need_weights: bool = False,\n pos_bias=None\n ):\n residual = x\n\n if self.layer_norm_first:\n x = self.self_attn_layer_norm(x)\n x, attn, pos_bias = self.self_attn(\n query=x,\n key=x,\n value=x,\n key_padding_mask=self_attn_padding_mask,\n need_weights=False,\n attn_mask=self_attn_mask,\n position_bias=pos_bias\n )\n x = self.dropout1(x)\n x = residual + x\n\n residual = x\n x = self.final_layer_norm(x)\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual + x\n else:\n x, attn, pos_bias = self.self_attn(\n query=x,\n key=x,\n value=x,\n key_padding_mask=self_attn_padding_mask,\n need_weights=need_weights,\n attn_mask=self_attn_mask,\n position_bias=pos_bias\n )\n\n x = self.dropout1(x)\n x = residual * self.deep_norm_alpha + x\n\n x = self.self_attn_layer_norm(x)\n\n residual = x\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual * self.deep_norm_alpha + x\n x = self.final_layer_norm(x)\n\n return x, attn, pos_bias\n\n\nclass MultiheadAttention(nn.Module):\n \"\"\"Multi-headed attention.\n\n See \"Attention Is All You Need\" for more details.\n \"\"\"\n\n def __init__(\n self,\n embed_dim,\n num_heads,\n kdim=None,\n vdim=None,\n dropout=0.0,\n bias=True,\n add_bias_kv=False,\n add_zero_attn=False,\n self_attention=False,\n encoder_decoder_attention=False,","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.MultiheadAttention","uri":"program://CREMA/class/lavis.models.beats.backbone.MultiheadAttention#L278-L748","kind":"class","name":"MultiheadAttention","path":"lavis/models/beats/backbone.py","language":"python","start_line":278,"end_line":748,"context_start_line":258,"context_end_line":768,"code":"\n x = self.dropout1(x)\n x = residual * self.deep_norm_alpha + x\n\n x = self.self_attn_layer_norm(x)\n\n residual = x\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual * self.deep_norm_alpha + x\n x = self.final_layer_norm(x)\n\n return x, attn, pos_bias\n\n\nclass MultiheadAttention(nn.Module):\n \"\"\"Multi-headed attention.\n\n See \"Attention Is All You Need\" for more details.\n \"\"\"\n\n def __init__(\n self,\n embed_dim,\n num_heads,\n kdim=None,\n vdim=None,\n dropout=0.0,\n bias=True,\n add_bias_kv=False,\n add_zero_attn=False,\n self_attention=False,\n encoder_decoder_attention=False,\n q_noise=0.0,\n qn_block_size=8,\n has_relative_attention_bias=False,\n num_buckets=32,\n max_distance=128,\n gru_rel_pos=False,\n rescale_init=False,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.kdim = kdim if kdim is not None else embed_dim\n self.vdim = vdim if vdim is not None else embed_dim\n self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim\n\n self.num_heads = num_heads\n self.dropout_module = nn.Dropout(dropout)\n\n self.has_relative_attention_bias = has_relative_attention_bias\n self.num_buckets = num_buckets\n self.max_distance = max_distance\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)\n\n self.head_dim = embed_dim // num_heads\n self.q_head_dim = self.head_dim\n self.k_head_dim = self.head_dim\n assert (\n self.head_dim * num_heads == self.embed_dim\n ), \"embed_dim must be divisible by num_heads\"\n self.scaling = self.head_dim ** -0.5\n\n self.self_attention = self_attention\n self.encoder_decoder_attention = encoder_decoder_attention\n\n assert not self.self_attention or self.qkv_same_dim, (\n \"Self-attention requires query, key and \" \"value to be of the same size\"\n )\n\n k_bias = True\n if rescale_init:\n k_bias = False\n\n k_embed_dim = embed_dim\n q_embed_dim = embed_dim\n\n self.k_proj = quant_noise(\n nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size\n )\n self.v_proj = quant_noise(\n nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n self.q_proj = quant_noise(\n nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n self.out_proj = quant_noise(\n nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n if add_bias_kv:\n self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))\n self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))\n else:\n self.bias_k = self.bias_v = None\n\n self.add_zero_attn = add_zero_attn\n\n self.gru_rel_pos = gru_rel_pos\n if self.gru_rel_pos:\n self.grep_linear = nn.Linear(self.q_head_dim, 8)\n self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))\n\n self.reset_parameters()\n\n def reset_parameters(self):\n if self.qkv_same_dim:\n # Empirically observed the convergence to be much better with\n # the scaled initialization\n nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))\n else:\n nn.init.xavier_uniform_(self.k_proj.weight)\n nn.init.xavier_uniform_(self.v_proj.weight)\n nn.init.xavier_uniform_(self.q_proj.weight)\n\n nn.init.xavier_uniform_(self.out_proj.weight)\n if self.out_proj.bias is not None:\n nn.init.constant_(self.out_proj.bias, 0.0)\n if self.bias_k is not None:\n nn.init.xavier_normal_(self.bias_k)\n if self.bias_v is not None:\n nn.init.xavier_normal_(self.bias_v)\n if self.has_relative_attention_bias:\n nn.init.xavier_normal_(self.relative_attention_bias.weight)\n\n def _relative_positions_bucket(self, relative_positions, bidirectional=True):\n num_buckets = self.num_buckets\n max_distance = self.max_distance\n relative_buckets = 0\n\n if bidirectional:\n num_buckets = num_buckets // 2\n relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets\n relative_positions = torch.abs(relative_positions)\n else:\n relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))\n\n max_exact = num_buckets // 2\n is_small = relative_positions < max_exact\n\n relative_postion_if_large = max_exact + (\n torch.log(relative_positions.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_postion_if_large = torch.min(\n relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)\n )\n\n relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)\n return relative_buckets\n\n def compute_bias(self, query_length, key_length):\n context_position = torch.arange(query_length, dtype=torch.long)[:, None]\n memory_position = torch.arange(key_length, dtype=torch.long)[None, :]\n relative_position = memory_position - context_position\n relative_position_bucket = self._relative_positions_bucket(\n relative_position,\n bidirectional=True\n )\n relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)\n values = self.relative_attention_bias(relative_position_bucket)\n values = values.permute([2, 0, 1])\n return values\n\n def forward(\n self,\n query,\n key: Optional[Tensor],\n value: Optional[Tensor],\n key_padding_mask: Optional[Tensor] = None,\n incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,\n need_weights: bool = True,\n static_kv: bool = False,\n attn_mask: Optional[Tensor] = None,\n before_softmax: bool = False,\n need_head_weights: bool = False,\n position_bias: Optional[Tensor] = None\n ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:\n \"\"\"Input shape: Time x Batch x Channel\n\n Args:\n key_padding_mask (ByteTensor, optional): mask to exclude\n keys that are pads, of shape `(batch, src_len)`, where\n padding elements are indicated by 1s.\n need_weights (bool, optional): return the attention weights,\n averaged over heads (default: False).\n attn_mask (ByteTensor, optional): typically used to\n implement causal attention, where the mask prevents the\n attention from looking forward in time (default: None).\n before_softmax (bool, optional): return the raw attention\n weights and values before the attention softmax.\n need_head_weights (bool, optional): return the attention\n weights for each head. Implies *need_weights*. Default:\n return the average attention weights over all heads.\n \"\"\"\n if need_head_weights:\n need_weights = True\n\n is_tpu = query.device.type == \"xla\"\n\n tgt_len, bsz, embed_dim = query.size()\n src_len = tgt_len\n assert embed_dim == self.embed_dim\n assert list(query.size()) == [tgt_len, bsz, embed_dim]\n if key is not None:\n src_len, key_bsz, _ = key.size()\n if not torch.jit.is_scripting():\n assert key_bsz == bsz\n assert value is not None\n assert src_len, bsz == value.shape[:2]\n\n if self.has_relative_attention_bias and position_bias is None:\n position_bias = self.compute_bias(tgt_len, src_len)\n position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)\n\n if incremental_state is not None:\n saved_state = self._get_input_buffer(incremental_state)\n if saved_state is not None and \"prev_key\" in saved_state:\n # previous time steps are cached - no need to recompute\n # key and value if they are static\n if static_kv:\n assert self.encoder_decoder_attention and not self.self_attention\n key = value = None\n else:\n saved_state = None\n\n if self.self_attention:\n q = self.q_proj(query)\n k = self.k_proj(query)\n v = self.v_proj(query)\n elif self.encoder_decoder_attention:\n # encoder-decoder attention\n q = self.q_proj(query)\n if key is None:\n assert value is None\n k = v = None\n else:\n k = self.k_proj(key)\n v = self.v_proj(key)\n\n else:\n assert key is not None and value is not None\n q = self.q_proj(query)\n k = self.k_proj(key)\n v = self.v_proj(value)\n q *= self.scaling\n alpha = 32\n q *= 1 / alpha\n\n if self.bias_k is not None:\n assert self.bias_v is not None\n k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])\n v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])\n if attn_mask is not None:\n attn_mask = torch.cat(\n [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1\n )\n if key_padding_mask is not None:\n key_padding_mask = torch.cat(\n [\n key_padding_mask,\n key_padding_mask.new_zeros(key_padding_mask.size(0), 1),\n ],\n dim=1,\n )\n\n q = (\n q.contiguous()\n .view(tgt_len, bsz * self.num_heads, self.q_head_dim)\n .transpose(0, 1)\n )\n if k is not None:\n k = (\n k.contiguous()\n .view(-1, bsz * self.num_heads, self.k_head_dim)\n .transpose(0, 1)\n )\n if v is not None:\n v = (\n v.contiguous()\n .view(-1, bsz * self.num_heads, self.head_dim)\n .transpose(0, 1)\n )\n\n if saved_state is not None:\n # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)\n if \"prev_key\" in saved_state:\n _prev_key = saved_state[\"prev_key\"]\n assert _prev_key is not None\n prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)\n if static_kv:\n k = prev_key\n else:\n assert k is not None\n k = torch.cat([prev_key, k], dim=1)\n src_len = k.size(1)\n if \"prev_value\" in saved_state:\n _prev_value = saved_state[\"prev_value\"]\n assert _prev_value is not None\n prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)\n if static_kv:\n v = prev_value\n else:\n assert v is not None\n v = torch.cat([prev_value, v], dim=1)\n prev_key_padding_mask: Optional[Tensor] = None\n if \"prev_key_padding_mask\" in saved_state:\n prev_key_padding_mask = saved_state[\"prev_key_padding_mask\"]\n assert k is not None and v is not None\n key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(\n key_padding_mask=key_padding_mask,\n prev_key_padding_mask=prev_key_padding_mask,\n batch_size=bsz,\n src_len=k.size(1),\n static_kv=static_kv,\n )\n\n saved_state[\"prev_key\"] = k.view(bsz, self.num_heads, -1, self.head_dim)\n saved_state[\"prev_value\"] = v.view(bsz, self.num_heads, -1, self.head_dim)\n saved_state[\"prev_key_padding_mask\"] = key_padding_mask\n # In this branch incremental_state is never None\n assert incremental_state is not None\n incremental_state = self._set_input_buffer(incremental_state, saved_state)\n assert k is not None\n assert k.size(1) == src_len\n\n # This is part of a workaround to get around fork/join parallelism\n # not supporting Optional types.\n if key_padding_mask is not None and key_padding_mask.dim() == 0:\n key_padding_mask = None\n\n if key_padding_mask is not None:\n assert key_padding_mask.size(0) == bsz\n assert key_padding_mask.size(1) == src_len\n\n if self.add_zero_attn:\n assert v is not None\n src_len += 1\n k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)\n v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)\n if attn_mask is not None:\n attn_mask = torch.cat(\n [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1\n )\n if key_padding_mask is not None:\n key_padding_mask = torch.cat(\n [\n key_padding_mask,\n torch.zeros(key_padding_mask.size(0), 1).type_as(\n key_padding_mask\n ),\n ],\n dim=1,\n )\n\n attn_weights = torch.bmm(q, k.transpose(1, 2))\n attn_weights = (attn_weights - attn_weights.max(dim=-1, keepdim=True)[0]) * alpha\n attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)\n\n assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]\n\n if attn_mask is not None:\n attn_mask = attn_mask.unsqueeze(0)\n attn_weights += attn_mask\n\n if key_padding_mask is not None:\n # don't attend to padding symbols\n attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)\n if not is_tpu:\n attn_weights = attn_weights.masked_fill(\n key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),\n float(\"-inf\"),\n )\n else:\n attn_weights = attn_weights.transpose(0, 2)\n attn_weights = attn_weights.masked_fill(key_padding_mask, float(\"-inf\"))\n attn_weights = attn_weights.transpose(0, 2)\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n if before_softmax:\n return attn_weights, v, position_bias\n\n if position_bias is not None:\n attn_mask_rel_pos = position_bias\n if self.gru_rel_pos == 1:\n query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim) * alpha / self.scaling\n _B, _H, _L, __ = query_layer.size()\n gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(\n _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)\n gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0\n attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, tgt_len, 1) * position_bias\n\n attn_mask_rel_pos = attn_mask_rel_pos.view(attn_weights.size())\n\n attn_weights = attn_weights + attn_mask_rel_pos\n\n attn_weights_float = F.softmax(\n attn_weights, dim=-1\n )\n attn_weights = attn_weights_float.type_as(attn_weights)\n attn_probs = self.dropout_module(attn_weights)\n\n assert v is not None\n attn = torch.bmm(attn_probs, v)\n assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]\n attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)\n attn = self.out_proj(attn)\n attn_weights: Optional[Tensor] = None\n if need_weights:\n attn_weights = attn_weights_float.view(\n bsz, self.num_heads, tgt_len, src_len\n ).transpose(1, 0)\n if not need_head_weights:\n # average attention weights over heads\n attn_weights = attn_weights.mean(dim=0)\n\n return attn, attn_weights, position_bias\n\n @staticmethod\n def _append_prev_key_padding_mask(\n key_padding_mask: Optional[Tensor],\n prev_key_padding_mask: Optional[Tensor],\n batch_size: int,\n src_len: int,\n static_kv: bool,\n ) -> Optional[Tensor]:\n # saved key padding masks have shape (bsz, seq_len)\n if prev_key_padding_mask is not None and static_kv:\n new_key_padding_mask = prev_key_padding_mask\n elif prev_key_padding_mask is not None and key_padding_mask is not None:\n new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1\n )\n # During incremental decoding, as the padding token enters and\n # leaves the frame, there will be a time when prev or current\n # is None\n elif prev_key_padding_mask is not None:\n if src_len > prev_key_padding_mask.size(1):\n filler = torch.zeros(\n (batch_size, src_len - prev_key_padding_mask.size(1)),\n device=prev_key_padding_mask.device,\n )\n new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), filler.float()], dim=1\n )\n else:\n new_key_padding_mask = prev_key_padding_mask.float()\n elif key_padding_mask is not None:\n if src_len > key_padding_mask.size(1):\n filler = torch.zeros(\n (batch_size, src_len - key_padding_mask.size(1)),\n device=key_padding_mask.device,\n )\n new_key_padding_mask = torch.cat(\n [filler.float(), key_padding_mask.float()], dim=1\n )\n else:\n new_key_padding_mask = key_padding_mask.float()\n else:\n new_key_padding_mask = prev_key_padding_mask\n return new_key_padding_mask\n\n def _get_input_buffer(\n self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n# ... truncated ...","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.init_bert_params","uri":"program://CREMA/function/lavis.models.beats.backbone.init_bert_params#L751-L783","kind":"function","name":"init_bert_params","path":"lavis/models/beats/backbone.py","language":"python","start_line":751,"end_line":783,"context_start_line":731,"context_end_line":783,"code":" self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n This overrides the default initializations depending on the specified arguments.\n 1. If normal_init_linear_weights is set then weights of linear\n layer will be initialized using the normal distribution and\n bais will be set to the specified value.\n 2. If normal_init_embed_weights is set then weights of embedding\n layer will be initialized using the normal distribution.\n 3. If normal_init_proj_weights is set then weights of\n in_project_weight for MultiHeadAttention initialized using\n the normal distribution (to be validated).\n \"\"\"\n\n def normal_(data):\n # with FSDP, module params will be on CUDA, so we cast them back to CPU\n # so that the RNG is consistent with and without FSDP\n data.copy_(\n data.cpu().normal_(mean=0.0, std=0.02).to(data.device)\n )\n\n if isinstance(module, nn.Linear):\n normal_(module.weight.data)\n if module.bias is not None:\n module.bias.data.zero_()\n if isinstance(module, nn.Embedding):\n normal_(module.weight.data)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n if isinstance(module, MultiheadAttention):\n normal_(module.q_proj.weight.data)\n normal_(module.k_proj.weight.data)\n normal_(module.v_proj.weight.data)","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.__init__","uri":"program://CREMA/function/lavis.models.beats.backbone.__init__#L284-L368","kind":"function","name":"__init__","path":"lavis/models/beats/backbone.py","language":"python","start_line":284,"end_line":368,"context_start_line":264,"context_end_line":388,"code":" residual = x\n if self.activation_name == \"glu\":\n x = self.fc1(x)\n else:\n x = self.activation_fn(self.fc1(x))\n x = self.dropout2(x)\n x = self.fc2(x)\n x = self.dropout3(x)\n x = residual * self.deep_norm_alpha + x\n x = self.final_layer_norm(x)\n\n return x, attn, pos_bias\n\n\nclass MultiheadAttention(nn.Module):\n \"\"\"Multi-headed attention.\n\n See \"Attention Is All You Need\" for more details.\n \"\"\"\n\n def __init__(\n self,\n embed_dim,\n num_heads,\n kdim=None,\n vdim=None,\n dropout=0.0,\n bias=True,\n add_bias_kv=False,\n add_zero_attn=False,\n self_attention=False,\n encoder_decoder_attention=False,\n q_noise=0.0,\n qn_block_size=8,\n has_relative_attention_bias=False,\n num_buckets=32,\n max_distance=128,\n gru_rel_pos=False,\n rescale_init=False,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.kdim = kdim if kdim is not None else embed_dim\n self.vdim = vdim if vdim is not None else embed_dim\n self.qkv_same_dim = self.kdim == embed_dim and self.vdim == embed_dim\n\n self.num_heads = num_heads\n self.dropout_module = nn.Dropout(dropout)\n\n self.has_relative_attention_bias = has_relative_attention_bias\n self.num_buckets = num_buckets\n self.max_distance = max_distance\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(num_buckets, num_heads)\n\n self.head_dim = embed_dim // num_heads\n self.q_head_dim = self.head_dim\n self.k_head_dim = self.head_dim\n assert (\n self.head_dim * num_heads == self.embed_dim\n ), \"embed_dim must be divisible by num_heads\"\n self.scaling = self.head_dim ** -0.5\n\n self.self_attention = self_attention\n self.encoder_decoder_attention = encoder_decoder_attention\n\n assert not self.self_attention or self.qkv_same_dim, (\n \"Self-attention requires query, key and \" \"value to be of the same size\"\n )\n\n k_bias = True\n if rescale_init:\n k_bias = False\n\n k_embed_dim = embed_dim\n q_embed_dim = embed_dim\n\n self.k_proj = quant_noise(\n nn.Linear(self.kdim, k_embed_dim, bias=k_bias), q_noise, qn_block_size\n )\n self.v_proj = quant_noise(\n nn.Linear(self.vdim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n self.q_proj = quant_noise(\n nn.Linear(embed_dim, q_embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n self.out_proj = quant_noise(\n nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n if add_bias_kv:\n self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))\n self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))\n else:\n self.bias_k = self.bias_v = None\n\n self.add_zero_attn = add_zero_attn\n\n self.gru_rel_pos = gru_rel_pos\n if self.gru_rel_pos:\n self.grep_linear = nn.Linear(self.q_head_dim, 8)\n self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))\n\n self.reset_parameters()\n\n def reset_parameters(self):\n if self.qkv_same_dim:\n # Empirically observed the convergence to be much better with\n # the scaled initialization\n nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))\n else:\n nn.init.xavier_uniform_(self.k_proj.weight)\n nn.init.xavier_uniform_(self.v_proj.weight)\n nn.init.xavier_uniform_(self.q_proj.weight)\n\n nn.init.xavier_uniform_(self.out_proj.weight)\n if self.out_proj.bias is not None:\n nn.init.constant_(self.out_proj.bias, 0.0)\n if self.bias_k is not None:\n nn.init.xavier_normal_(self.bias_k)\n if self.bias_v is not None:\n nn.init.xavier_normal_(self.bias_v)","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.forward","uri":"program://CREMA/function/lavis.models.beats.backbone.forward#L432-L684","kind":"function","name":"forward","path":"lavis/models/beats/backbone.py","language":"python","start_line":432,"end_line":684,"context_start_line":412,"context_end_line":704,"code":" relative_postion_if_large = torch.min(\n relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)\n )\n\n relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)\n return relative_buckets\n\n def compute_bias(self, query_length, key_length):\n context_position = torch.arange(query_length, dtype=torch.long)[:, None]\n memory_position = torch.arange(key_length, dtype=torch.long)[None, :]\n relative_position = memory_position - context_position\n relative_position_bucket = self._relative_positions_bucket(\n relative_position,\n bidirectional=True\n )\n relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)\n values = self.relative_attention_bias(relative_position_bucket)\n values = values.permute([2, 0, 1])\n return values\n\n def forward(\n self,\n query,\n key: Optional[Tensor],\n value: Optional[Tensor],\n key_padding_mask: Optional[Tensor] = None,\n incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,\n need_weights: bool = True,\n static_kv: bool = False,\n attn_mask: Optional[Tensor] = None,\n before_softmax: bool = False,\n need_head_weights: bool = False,\n position_bias: Optional[Tensor] = None\n ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:\n \"\"\"Input shape: Time x Batch x Channel\n\n Args:\n key_padding_mask (ByteTensor, optional): mask to exclude\n keys that are pads, of shape `(batch, src_len)`, where\n padding elements are indicated by 1s.\n need_weights (bool, optional): return the attention weights,\n averaged over heads (default: False).\n attn_mask (ByteTensor, optional): typically used to\n implement causal attention, where the mask prevents the\n attention from looking forward in time (default: None).\n before_softmax (bool, optional): return the raw attention\n weights and values before the attention softmax.\n need_head_weights (bool, optional): return the attention\n weights for each head. Implies *need_weights*. Default:\n return the average attention weights over all heads.\n \"\"\"\n if need_head_weights:\n need_weights = True\n\n is_tpu = query.device.type == \"xla\"\n\n tgt_len, bsz, embed_dim = query.size()\n src_len = tgt_len\n assert embed_dim == self.embed_dim\n assert list(query.size()) == [tgt_len, bsz, embed_dim]\n if key is not None:\n src_len, key_bsz, _ = key.size()\n if not torch.jit.is_scripting():\n assert key_bsz == bsz\n assert value is not None\n assert src_len, bsz == value.shape[:2]\n\n if self.has_relative_attention_bias and position_bias is None:\n position_bias = self.compute_bias(tgt_len, src_len)\n position_bias = position_bias.unsqueeze(0).repeat(bsz, 1, 1, 1).view(bsz * self.num_heads, tgt_len, src_len)\n\n if incremental_state is not None:\n saved_state = self._get_input_buffer(incremental_state)\n if saved_state is not None and \"prev_key\" in saved_state:\n # previous time steps are cached - no need to recompute\n # key and value if they are static\n if static_kv:\n assert self.encoder_decoder_attention and not self.self_attention\n key = value = None\n else:\n saved_state = None\n\n if self.self_attention:\n q = self.q_proj(query)\n k = self.k_proj(query)\n v = self.v_proj(query)\n elif self.encoder_decoder_attention:\n # encoder-decoder attention\n q = self.q_proj(query)\n if key is None:\n assert value is None\n k = v = None\n else:\n k = self.k_proj(key)\n v = self.v_proj(key)\n\n else:\n assert key is not None and value is not None\n q = self.q_proj(query)\n k = self.k_proj(key)\n v = self.v_proj(value)\n q *= self.scaling\n alpha = 32\n q *= 1 / alpha\n\n if self.bias_k is not None:\n assert self.bias_v is not None\n k = torch.cat([k, self.bias_k.repeat(1, bsz, 1)])\n v = torch.cat([v, self.bias_v.repeat(1, bsz, 1)])\n if attn_mask is not None:\n attn_mask = torch.cat(\n [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1\n )\n if key_padding_mask is not None:\n key_padding_mask = torch.cat(\n [\n key_padding_mask,\n key_padding_mask.new_zeros(key_padding_mask.size(0), 1),\n ],\n dim=1,\n )\n\n q = (\n q.contiguous()\n .view(tgt_len, bsz * self.num_heads, self.q_head_dim)\n .transpose(0, 1)\n )\n if k is not None:\n k = (\n k.contiguous()\n .view(-1, bsz * self.num_heads, self.k_head_dim)\n .transpose(0, 1)\n )\n if v is not None:\n v = (\n v.contiguous()\n .view(-1, bsz * self.num_heads, self.head_dim)\n .transpose(0, 1)\n )\n\n if saved_state is not None:\n # saved states are stored with shape (bsz, num_heads, seq_len, head_dim)\n if \"prev_key\" in saved_state:\n _prev_key = saved_state[\"prev_key\"]\n assert _prev_key is not None\n prev_key = _prev_key.view(bsz * self.num_heads, -1, self.head_dim)\n if static_kv:\n k = prev_key\n else:\n assert k is not None\n k = torch.cat([prev_key, k], dim=1)\n src_len = k.size(1)\n if \"prev_value\" in saved_state:\n _prev_value = saved_state[\"prev_value\"]\n assert _prev_value is not None\n prev_value = _prev_value.view(bsz * self.num_heads, -1, self.head_dim)\n if static_kv:\n v = prev_value\n else:\n assert v is not None\n v = torch.cat([prev_value, v], dim=1)\n prev_key_padding_mask: Optional[Tensor] = None\n if \"prev_key_padding_mask\" in saved_state:\n prev_key_padding_mask = saved_state[\"prev_key_padding_mask\"]\n assert k is not None and v is not None\n key_padding_mask = MultiheadAttention._append_prev_key_padding_mask(\n key_padding_mask=key_padding_mask,\n prev_key_padding_mask=prev_key_padding_mask,\n batch_size=bsz,\n src_len=k.size(1),\n static_kv=static_kv,\n )\n\n saved_state[\"prev_key\"] = k.view(bsz, self.num_heads, -1, self.head_dim)\n saved_state[\"prev_value\"] = v.view(bsz, self.num_heads, -1, self.head_dim)\n saved_state[\"prev_key_padding_mask\"] = key_padding_mask\n # In this branch incremental_state is never None\n assert incremental_state is not None\n incremental_state = self._set_input_buffer(incremental_state, saved_state)\n assert k is not None\n assert k.size(1) == src_len\n\n # This is part of a workaround to get around fork/join parallelism\n # not supporting Optional types.\n if key_padding_mask is not None and key_padding_mask.dim() == 0:\n key_padding_mask = None\n\n if key_padding_mask is not None:\n assert key_padding_mask.size(0) == bsz\n assert key_padding_mask.size(1) == src_len\n\n if self.add_zero_attn:\n assert v is not None\n src_len += 1\n k = torch.cat([k, k.new_zeros((k.size(0), 1) + k.size()[2:])], dim=1)\n v = torch.cat([v, v.new_zeros((v.size(0), 1) + v.size()[2:])], dim=1)\n if attn_mask is not None:\n attn_mask = torch.cat(\n [attn_mask, attn_mask.new_zeros(attn_mask.size(0), 1)], dim=1\n )\n if key_padding_mask is not None:\n key_padding_mask = torch.cat(\n [\n key_padding_mask,\n torch.zeros(key_padding_mask.size(0), 1).type_as(\n key_padding_mask\n ),\n ],\n dim=1,\n )\n\n attn_weights = torch.bmm(q, k.transpose(1, 2))\n attn_weights = (attn_weights - attn_weights.max(dim=-1, keepdim=True)[0]) * alpha\n attn_weights = self.apply_sparse_mask(attn_weights, tgt_len, src_len, bsz)\n\n assert list(attn_weights.size()) == [bsz * self.num_heads, tgt_len, src_len]\n\n if attn_mask is not None:\n attn_mask = attn_mask.unsqueeze(0)\n attn_weights += attn_mask\n\n if key_padding_mask is not None:\n # don't attend to padding symbols\n attn_weights = attn_weights.view(bsz, self.num_heads, tgt_len, src_len)\n if not is_tpu:\n attn_weights = attn_weights.masked_fill(\n key_padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),\n float(\"-inf\"),\n )\n else:\n attn_weights = attn_weights.transpose(0, 2)\n attn_weights = attn_weights.masked_fill(key_padding_mask, float(\"-inf\"))\n attn_weights = attn_weights.transpose(0, 2)\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n if before_softmax:\n return attn_weights, v, position_bias\n\n if position_bias is not None:\n attn_mask_rel_pos = position_bias\n if self.gru_rel_pos == 1:\n query_layer = q.view(bsz, self.num_heads, tgt_len, self.q_head_dim) * alpha / self.scaling\n _B, _H, _L, __ = query_layer.size()\n gate_a, gate_b = torch.sigmoid(self.grep_linear(query_layer).view(\n _B, _H, _L, 2, 4).sum(-1, keepdim=False)).chunk(2, dim=-1)\n gate_a_1 = gate_a * (gate_b * self.grep_a - 1.0) + 2.0\n attn_mask_rel_pos = gate_a_1.view(bsz * self.num_heads, tgt_len, 1) * position_bias\n\n attn_mask_rel_pos = attn_mask_rel_pos.view(attn_weights.size())\n\n attn_weights = attn_weights + attn_mask_rel_pos\n\n attn_weights_float = F.softmax(\n attn_weights, dim=-1\n )\n attn_weights = attn_weights_float.type_as(attn_weights)\n attn_probs = self.dropout_module(attn_weights)\n\n assert v is not None\n attn = torch.bmm(attn_probs, v)\n assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]\n attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)\n attn = self.out_proj(attn)\n attn_weights: Optional[Tensor] = None\n if need_weights:\n attn_weights = attn_weights_float.view(\n bsz, self.num_heads, tgt_len, src_len\n ).transpose(1, 0)\n if not need_head_weights:\n # average attention weights over heads\n attn_weights = attn_weights.mean(dim=0)\n\n return attn, attn_weights, position_bias\n\n @staticmethod\n def _append_prev_key_padding_mask(\n key_padding_mask: Optional[Tensor],\n prev_key_padding_mask: Optional[Tensor],\n batch_size: int,\n src_len: int,\n static_kv: bool,\n ) -> Optional[Tensor]:\n # saved key padding masks have shape (bsz, seq_len)\n if prev_key_padding_mask is not None and static_kv:\n new_key_padding_mask = prev_key_padding_mask\n elif prev_key_padding_mask is not None and key_padding_mask is not None:\n new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1\n )\n # During incremental decoding, as the padding token enters and\n # leaves the frame, there will be a time when prev or current\n # is None\n elif prev_key_padding_mask is not None:","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.extract_features","uri":"program://CREMA/function/lavis.models.beats.backbone.extract_features#L109-L150","kind":"function","name":"extract_features","path":"lavis/models/beats/backbone.py","language":"python","start_line":109,"end_line":150,"context_start_line":89,"context_end_line":170,"code":" if args.deep_norm:\n deep_norm_beta = math.pow(8 * args.encoder_layers, -1 / 4)\n for i in range(args.encoder_layers):\n nn.init.xavier_normal_(self.layers[i].self_attn.k_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.v_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].self_attn.q_proj.weight, gain=1)\n nn.init.xavier_normal_(self.layers[i].self_attn.out_proj.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc1.weight, gain=deep_norm_beta)\n nn.init.xavier_normal_(self.layers[i].fc2.weight, gain=deep_norm_beta)\n\n self.layer_wise_gradient_decay_ratio = getattr(args, \"layer_wise_gradient_decay_ratio\", 1)\n\n def forward(self, x, padding_mask=None, layer=None):\n x, layer_results = self.extract_features(x, padding_mask, layer)\n\n if self.layer_norm_first and layer is None:\n x = self.layer_norm(x)\n\n return x, layer_results\n\n def extract_features(self, x, padding_mask=None, tgt_layer=None):\n\n if padding_mask is not None:\n x[padding_mask] = 0\n\n x_conv = self.pos_conv(x.transpose(1, 2))\n x_conv = x_conv.transpose(1, 2)\n x = x + x_conv\n\n if not self.layer_norm_first:\n x = self.layer_norm(x)\n\n x = F.dropout(x, p=self.dropout, training=self.training)\n\n # B x T x C -> T x B x C\n x = x.transpose(0, 1)\n\n layer_results = []\n z = None\n if tgt_layer is not None:\n layer_results.append((x, z))\n r = None\n pos_bias = None\n for i, layer in enumerate(self.layers):\n if self.layer_wise_gradient_decay_ratio != 1.0:\n x = GradMultiply.apply(x, self.layer_wise_gradient_decay_ratio)\n dropout_probability = np.random.random()\n if not self.training or (dropout_probability > self.layerdrop):\n x, z, pos_bias = layer(x, self_attn_padding_mask=padding_mask, need_weights=False, pos_bias=pos_bias)\n if tgt_layer is not None:\n layer_results.append((x, z))\n if i == tgt_layer:\n r = x\n break\n\n if r is not None:\n x = r\n\n # T x B x C -> B x T x C\n x = x.transpose(0, 1)\n\n return x, layer_results\n\n\nclass TransformerSentenceEncoderLayer(nn.Module):\n def __init__(\n self,\n embedding_dim: float = 768,\n ffn_embedding_dim: float = 3072,\n num_attention_heads: float = 8,\n dropout: float = 0.1,\n attention_dropout: float = 0.1,\n activation_dropout: float = 0.1,\n activation_fn: str = \"relu\",\n layer_norm_first: bool = False,\n deep_norm: bool = False,\n has_relative_attention_bias: bool = False,\n num_buckets: int = 0,\n max_distance: int = 0,\n rescale_init: bool = False,\n gru_rel_pos: bool = False,\n encoder_layers: int = 0,","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.reset_parameters","uri":"program://CREMA/function/lavis.models.beats.backbone.reset_parameters#L370-L390","kind":"function","name":"reset_parameters","path":"lavis/models/beats/backbone.py","language":"python","start_line":370,"end_line":390,"context_start_line":350,"context_end_line":410,"code":"\n self.out_proj = quant_noise(\n nn.Linear(embed_dim, embed_dim, bias=bias), q_noise, qn_block_size\n )\n\n if add_bias_kv:\n self.bias_k = Parameter(torch.Tensor(1, 1, embed_dim))\n self.bias_v = Parameter(torch.Tensor(1, 1, embed_dim))\n else:\n self.bias_k = self.bias_v = None\n\n self.add_zero_attn = add_zero_attn\n\n self.gru_rel_pos = gru_rel_pos\n if self.gru_rel_pos:\n self.grep_linear = nn.Linear(self.q_head_dim, 8)\n self.grep_a = nn.Parameter(torch.ones(1, num_heads, 1, 1))\n\n self.reset_parameters()\n\n def reset_parameters(self):\n if self.qkv_same_dim:\n # Empirically observed the convergence to be much better with\n # the scaled initialization\n nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))\n else:\n nn.init.xavier_uniform_(self.k_proj.weight)\n nn.init.xavier_uniform_(self.v_proj.weight)\n nn.init.xavier_uniform_(self.q_proj.weight)\n\n nn.init.xavier_uniform_(self.out_proj.weight)\n if self.out_proj.bias is not None:\n nn.init.constant_(self.out_proj.bias, 0.0)\n if self.bias_k is not None:\n nn.init.xavier_normal_(self.bias_k)\n if self.bias_v is not None:\n nn.init.xavier_normal_(self.bias_v)\n if self.has_relative_attention_bias:\n nn.init.xavier_normal_(self.relative_attention_bias.weight)\n\n def _relative_positions_bucket(self, relative_positions, bidirectional=True):\n num_buckets = self.num_buckets\n max_distance = self.max_distance\n relative_buckets = 0\n\n if bidirectional:\n num_buckets = num_buckets // 2\n relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets\n relative_positions = torch.abs(relative_positions)\n else:\n relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))\n\n max_exact = num_buckets // 2\n is_small = relative_positions < max_exact\n\n relative_postion_if_large = max_exact + (\n torch.log(relative_positions.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone._relative_positions_bucket","uri":"program://CREMA/function/lavis.models.beats.backbone._relative_positions_bucket#L392-L417","kind":"function","name":"_relative_positions_bucket","path":"lavis/models/beats/backbone.py","language":"python","start_line":392,"end_line":417,"context_start_line":372,"context_end_line":437,"code":" # Empirically observed the convergence to be much better with\n # the scaled initialization\n nn.init.xavier_uniform_(self.k_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.v_proj.weight, gain=1 / math.sqrt(2))\n nn.init.xavier_uniform_(self.q_proj.weight, gain=1 / math.sqrt(2))\n else:\n nn.init.xavier_uniform_(self.k_proj.weight)\n nn.init.xavier_uniform_(self.v_proj.weight)\n nn.init.xavier_uniform_(self.q_proj.weight)\n\n nn.init.xavier_uniform_(self.out_proj.weight)\n if self.out_proj.bias is not None:\n nn.init.constant_(self.out_proj.bias, 0.0)\n if self.bias_k is not None:\n nn.init.xavier_normal_(self.bias_k)\n if self.bias_v is not None:\n nn.init.xavier_normal_(self.bias_v)\n if self.has_relative_attention_bias:\n nn.init.xavier_normal_(self.relative_attention_bias.weight)\n\n def _relative_positions_bucket(self, relative_positions, bidirectional=True):\n num_buckets = self.num_buckets\n max_distance = self.max_distance\n relative_buckets = 0\n\n if bidirectional:\n num_buckets = num_buckets // 2\n relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets\n relative_positions = torch.abs(relative_positions)\n else:\n relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))\n\n max_exact = num_buckets // 2\n is_small = relative_positions < max_exact\n\n relative_postion_if_large = max_exact + (\n torch.log(relative_positions.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_postion_if_large = torch.min(\n relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)\n )\n\n relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)\n return relative_buckets\n\n def compute_bias(self, query_length, key_length):\n context_position = torch.arange(query_length, dtype=torch.long)[:, None]\n memory_position = torch.arange(key_length, dtype=torch.long)[None, :]\n relative_position = memory_position - context_position\n relative_position_bucket = self._relative_positions_bucket(\n relative_position,\n bidirectional=True\n )\n relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)\n values = self.relative_attention_bias(relative_position_bucket)\n values = values.permute([2, 0, 1])\n return values\n\n def forward(\n self,\n query,\n key: Optional[Tensor],\n value: Optional[Tensor],\n key_padding_mask: Optional[Tensor] = None,","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.compute_bias","uri":"program://CREMA/function/lavis.models.beats.backbone.compute_bias#L419-L430","kind":"function","name":"compute_bias","path":"lavis/models/beats/backbone.py","language":"python","start_line":419,"end_line":430,"context_start_line":399,"context_end_line":450,"code":" relative_buckets += (relative_positions > 0).to(torch.long) * num_buckets\n relative_positions = torch.abs(relative_positions)\n else:\n relative_positions = -torch.min(relative_positions, torch.zeros_like(relative_positions))\n\n max_exact = num_buckets // 2\n is_small = relative_positions < max_exact\n\n relative_postion_if_large = max_exact + (\n torch.log(relative_positions.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_postion_if_large = torch.min(\n relative_postion_if_large, torch.full_like(relative_postion_if_large, num_buckets - 1)\n )\n\n relative_buckets += torch.where(is_small, relative_positions, relative_postion_if_large)\n return relative_buckets\n\n def compute_bias(self, query_length, key_length):\n context_position = torch.arange(query_length, dtype=torch.long)[:, None]\n memory_position = torch.arange(key_length, dtype=torch.long)[None, :]\n relative_position = memory_position - context_position\n relative_position_bucket = self._relative_positions_bucket(\n relative_position,\n bidirectional=True\n )\n relative_position_bucket = relative_position_bucket.to(self.relative_attention_bias.weight.device)\n values = self.relative_attention_bias(relative_position_bucket)\n values = values.permute([2, 0, 1])\n return values\n\n def forward(\n self,\n query,\n key: Optional[Tensor],\n value: Optional[Tensor],\n key_padding_mask: Optional[Tensor] = None,\n incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]] = None,\n need_weights: bool = True,\n static_kv: bool = False,\n attn_mask: Optional[Tensor] = None,\n before_softmax: bool = False,\n need_head_weights: bool = False,\n position_bias: Optional[Tensor] = None\n ) -> Tuple[Tensor, Optional[Tensor], Optional[Tensor]]:\n \"\"\"Input shape: Time x Batch x Channel\n\n Args:\n key_padding_mask (ByteTensor, optional): mask to exclude\n keys that are pads, of shape `(batch, src_len)`, where","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone._append_prev_key_padding_mask","uri":"program://CREMA/function/lavis.models.beats.backbone._append_prev_key_padding_mask#L687-L728","kind":"function","name":"_append_prev_key_padding_mask","path":"lavis/models/beats/backbone.py","language":"python","start_line":687,"end_line":728,"context_start_line":667,"context_end_line":748,"code":" attn_weights = attn_weights_float.type_as(attn_weights)\n attn_probs = self.dropout_module(attn_weights)\n\n assert v is not None\n attn = torch.bmm(attn_probs, v)\n assert list(attn.size()) == [bsz * self.num_heads, tgt_len, self.head_dim]\n attn = attn.transpose(0, 1).contiguous().view(tgt_len, bsz, embed_dim)\n attn = self.out_proj(attn)\n attn_weights: Optional[Tensor] = None\n if need_weights:\n attn_weights = attn_weights_float.view(\n bsz, self.num_heads, tgt_len, src_len\n ).transpose(1, 0)\n if not need_head_weights:\n # average attention weights over heads\n attn_weights = attn_weights.mean(dim=0)\n\n return attn, attn_weights, position_bias\n\n @staticmethod\n def _append_prev_key_padding_mask(\n key_padding_mask: Optional[Tensor],\n prev_key_padding_mask: Optional[Tensor],\n batch_size: int,\n src_len: int,\n static_kv: bool,\n ) -> Optional[Tensor]:\n # saved key padding masks have shape (bsz, seq_len)\n if prev_key_padding_mask is not None and static_kv:\n new_key_padding_mask = prev_key_padding_mask\n elif prev_key_padding_mask is not None and key_padding_mask is not None:\n new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), key_padding_mask.float()], dim=1\n )\n # During incremental decoding, as the padding token enters and\n # leaves the frame, there will be a time when prev or current\n # is None\n elif prev_key_padding_mask is not None:\n if src_len > prev_key_padding_mask.size(1):\n filler = torch.zeros(\n (batch_size, src_len - prev_key_padding_mask.size(1)),\n device=prev_key_padding_mask.device,\n )\n new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), filler.float()], dim=1\n )\n else:\n new_key_padding_mask = prev_key_padding_mask.float()\n elif key_padding_mask is not None:\n if src_len > key_padding_mask.size(1):\n filler = torch.zeros(\n (batch_size, src_len - key_padding_mask.size(1)),\n device=key_padding_mask.device,\n )\n new_key_padding_mask = torch.cat(\n [filler.float(), key_padding_mask.float()], dim=1\n )\n else:\n new_key_padding_mask = key_padding_mask.float()\n else:\n new_key_padding_mask = prev_key_padding_mask\n return new_key_padding_mask\n\n def _get_input_buffer(\n self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone._get_input_buffer","uri":"program://CREMA/function/lavis.models.beats.backbone._get_input_buffer#L730-L738","kind":"function","name":"_get_input_buffer","path":"lavis/models/beats/backbone.py","language":"python","start_line":730,"end_line":738,"context_start_line":710,"context_end_line":758,"code":" new_key_padding_mask = torch.cat(\n [prev_key_padding_mask.float(), filler.float()], dim=1\n )\n else:\n new_key_padding_mask = prev_key_padding_mask.float()\n elif key_padding_mask is not None:\n if src_len > key_padding_mask.size(1):\n filler = torch.zeros(\n (batch_size, src_len - key_padding_mask.size(1)),\n device=key_padding_mask.device,\n )\n new_key_padding_mask = torch.cat(\n [filler.float(), key_padding_mask.float()], dim=1\n )\n else:\n new_key_padding_mask = key_padding_mask.float()\n else:\n new_key_padding_mask = prev_key_padding_mask\n return new_key_padding_mask\n\n def _get_input_buffer(\n self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n This overrides the default initializations depending on the specified arguments.\n 1. If normal_init_linear_weights is set then weights of linear\n layer will be initialized using the normal distribution and\n bais will be set to the specified value.\n 2. If normal_init_embed_weights is set then weights of embedding","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone._set_input_buffer","uri":"program://CREMA/function/lavis.models.beats.backbone._set_input_buffer#L740-L745","kind":"function","name":"_set_input_buffer","path":"lavis/models/beats/backbone.py","language":"python","start_line":740,"end_line":745,"context_start_line":720,"context_end_line":765,"code":" )\n new_key_padding_mask = torch.cat(\n [filler.float(), key_padding_mask.float()], dim=1\n )\n else:\n new_key_padding_mask = key_padding_mask.float()\n else:\n new_key_padding_mask = prev_key_padding_mask\n return new_key_padding_mask\n\n def _get_input_buffer(\n self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n This overrides the default initializations depending on the specified arguments.\n 1. If normal_init_linear_weights is set then weights of linear\n layer will be initialized using the normal distribution and\n bais will be set to the specified value.\n 2. If normal_init_embed_weights is set then weights of embedding\n layer will be initialized using the normal distribution.\n 3. If normal_init_proj_weights is set then weights of\n in_project_weight for MultiHeadAttention initialized using\n the normal distribution (to be validated).\n \"\"\"\n\n def normal_(data):","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.apply_sparse_mask","uri":"program://CREMA/function/lavis.models.beats.backbone.apply_sparse_mask#L747-L748","kind":"function","name":"apply_sparse_mask","path":"lavis/models/beats/backbone.py","language":"python","start_line":747,"end_line":748,"context_start_line":727,"context_end_line":768,"code":" new_key_padding_mask = prev_key_padding_mask\n return new_key_padding_mask\n\n def _get_input_buffer(\n self, incremental_state: Optional[Dict[str, Dict[str, Optional[Tensor]]]]\n ) -> Dict[str, Optional[Tensor]]:\n result = self.get_incremental_state(incremental_state, \"attn_state\")\n if result is not None:\n return result\n else:\n empty_result: Dict[str, Optional[Tensor]] = {}\n return empty_result\n\n def _set_input_buffer(\n self,\n incremental_state: Dict[str, Dict[str, Optional[Tensor]]],\n buffer: Dict[str, Optional[Tensor]],\n ):\n return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n This overrides the default initializations depending on the specified arguments.\n 1. If normal_init_linear_weights is set then weights of linear\n layer will be initialized using the normal distribution and\n bais will be set to the specified value.\n 2. If normal_init_embed_weights is set then weights of embedding\n layer will be initialized using the normal distribution.\n 3. If normal_init_proj_weights is set then weights of\n in_project_weight for MultiHeadAttention initialized using\n the normal distribution (to be validated).\n \"\"\"\n\n def normal_(data):\n # with FSDP, module params will be on CUDA, so we cast them back to CPU\n # so that the RNG is consistent with and without FSDP\n data.copy_(","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.backbone.normal_","uri":"program://CREMA/function/lavis.models.beats.backbone.normal_#L765-L770","kind":"function","name":"normal_","path":"lavis/models/beats/backbone.py","language":"python","start_line":765,"end_line":770,"context_start_line":745,"context_end_line":783,"code":" return self.set_incremental_state(incremental_state, \"attn_state\", buffer)\n\n def apply_sparse_mask(self, attn_weights, tgt_len: int, src_len: int, bsz: int):\n return attn_weights\n\n\ndef init_bert_params(module):\n \"\"\"\n Initialize the weights specific to the BERT Model.\n This overrides the default initializations depending on the specified arguments.\n 1. If normal_init_linear_weights is set then weights of linear\n layer will be initialized using the normal distribution and\n bais will be set to the specified value.\n 2. If normal_init_embed_weights is set then weights of embedding\n layer will be initialized using the normal distribution.\n 3. If normal_init_proj_weights is set then weights of\n in_project_weight for MultiHeadAttention initialized using\n the normal distribution (to be validated).\n \"\"\"\n\n def normal_(data):\n # with FSDP, module params will be on CUDA, so we cast them back to CPU\n # so that the RNG is consistent with and without FSDP\n data.copy_(\n data.cpu().normal_(mean=0.0, std=0.02).to(data.device)\n )\n\n if isinstance(module, nn.Linear):\n normal_(module.weight.data)\n if module.bias is not None:\n module.bias.data.zero_()\n if isinstance(module, nn.Embedding):\n normal_(module.weight.data)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n if isinstance(module, MultiheadAttention):\n normal_(module.q_proj.weight.data)\n normal_(module.k_proj.weight.data)\n normal_(module.v_proj.weight.data)","source_hash":"cdc0427161b55558e9746d77954e88e78de5a11e886d26102b6f55d641ac4269","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers","uri":"program://CREMA/module/lavis.models.beats.Tokenizers#L1-L173","kind":"module","name":"lavis.models.beats.Tokenizers","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":1,"end_line":173,"context_start_line":1,"context_end_line":173,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import LayerNorm\nimport torchaudio.compliance.kaldi as ta_kaldi\n\nfrom lavis.models.beats.backbone import (\n TransformerEncoder,\n)\nfrom lavis.models.beats.quantizer import (\n NormEMAVectorQuantizer,\n)\n\nimport logging\nfrom typing import Optional\n\nlogger = logging.getLogger(__name__)\n\n\nclass TokenizersConfig:\n def __init__(self, cfg=None):\n self.input_patch_size: int = -1 # path size of patch embedding\n self.embed_dim: int = 512 # patch embedding dimension\n self.conv_bias: bool = False # include bias in conv encoder\n\n self.encoder_layers: int = 12 # num encoder layers in the transformer\n self.encoder_embed_dim: int = 768 # encoder embedding dimension\n self.encoder_ffn_embed_dim: int = 3072 # encoder embedding dimension for FFN\n self.encoder_attention_heads: int = 12 # num encoder attention heads\n self.activation_fn: str = \"gelu\" # activation function to use\n\n self.layer_norm_first: bool = False # apply layernorm first in the transformer\n self.deep_norm: bool = False # apply deep_norm first in the transformer\n\n # dropouts\n self.dropout: float = 0.1 # dropout probability for the transformer\n self.attention_dropout: float = 0.1 # dropout probability for attention weights\n self.activation_dropout: float = 0.0 # dropout probability after activation in FFN\n self.encoder_layerdrop: float = 0.0 # probability of dropping a tarnsformer layer\n self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # quantizer\n self.quant_n: int = 1024 # codebook number in quantizer\n self.quant_dim: int = 256 # codebook dimension in quantizer\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass Tokenizers(nn.Module):\n def __init__(\n self,\n cfg: TokenizersConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"Tokenizers Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n self.quantize = NormEMAVectorQuantizer(\n n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,\n )\n self.quant_n = cfg.quant_n\n self.quantize_layer = nn.Sequential(\n nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),\n nn.Tanh(),\n nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim) # for quantize\n )\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_labels(\n self,\n source: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n quantize_input = self.quantize_layer(x)\n quantize_feature, embed_loss, embed_ind = self.quantize(quantize_input)\n\n return embed_ind\n","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.TokenizersConfig","uri":"program://CREMA/class/lavis.models.beats.Tokenizers.TokenizersConfig#L29-L69","kind":"class","name":"TokenizersConfig","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":29,"end_line":69,"context_start_line":9,"context_end_line":89,"code":"\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import LayerNorm\nimport torchaudio.compliance.kaldi as ta_kaldi\n\nfrom lavis.models.beats.backbone import (\n TransformerEncoder,\n)\nfrom lavis.models.beats.quantizer import (\n NormEMAVectorQuantizer,\n)\n\nimport logging\nfrom typing import Optional\n\nlogger = logging.getLogger(__name__)\n\n\nclass TokenizersConfig:\n def __init__(self, cfg=None):\n self.input_patch_size: int = -1 # path size of patch embedding\n self.embed_dim: int = 512 # patch embedding dimension\n self.conv_bias: bool = False # include bias in conv encoder\n\n self.encoder_layers: int = 12 # num encoder layers in the transformer\n self.encoder_embed_dim: int = 768 # encoder embedding dimension\n self.encoder_ffn_embed_dim: int = 3072 # encoder embedding dimension for FFN\n self.encoder_attention_heads: int = 12 # num encoder attention heads\n self.activation_fn: str = \"gelu\" # activation function to use\n\n self.layer_norm_first: bool = False # apply layernorm first in the transformer\n self.deep_norm: bool = False # apply deep_norm first in the transformer\n\n # dropouts\n self.dropout: float = 0.1 # dropout probability for the transformer\n self.attention_dropout: float = 0.1 # dropout probability for attention weights\n self.activation_dropout: float = 0.0 # dropout probability after activation in FFN\n self.encoder_layerdrop: float = 0.0 # probability of dropping a tarnsformer layer\n self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # quantizer\n self.quant_n: int = 1024 # codebook number in quantizer\n self.quant_dim: int = 256 # codebook dimension in quantizer\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass Tokenizers(nn.Module):\n def __init__(\n self,\n cfg: TokenizersConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"Tokenizers Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.Tokenizers","uri":"program://CREMA/class/lavis.models.beats.Tokenizers.Tokenizers#L72-L172","kind":"class","name":"Tokenizers","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":72,"end_line":172,"context_start_line":52,"context_end_line":173,"code":" self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # quantizer\n self.quant_n: int = 1024 # codebook number in quantizer\n self.quant_dim: int = 256 # codebook dimension in quantizer\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass Tokenizers(nn.Module):\n def __init__(\n self,\n cfg: TokenizersConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"Tokenizers Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n self.quantize = NormEMAVectorQuantizer(\n n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,\n )\n self.quant_n = cfg.quant_n\n self.quantize_layer = nn.Sequential(\n nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),\n nn.Tanh(),\n nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim) # for quantize\n )\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_labels(\n self,\n source: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n quantize_input = self.quantize_layer(x)\n quantize_feature, embed_loss, embed_ind = self.quantize(quantize_input)\n\n return embed_ind\n","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.__init__","uri":"program://CREMA/function/lavis.models.beats.Tokenizers.__init__#L73-L107","kind":"function","name":"__init__","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":73,"end_line":107,"context_start_line":53,"context_end_line":127,"code":" self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # quantizer\n self.quant_n: int = 1024 # codebook number in quantizer\n self.quant_dim: int = 256 # codebook dimension in quantizer\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass Tokenizers(nn.Module):\n def __init__(\n self,\n cfg: TokenizersConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"Tokenizers Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n self.quantize = NormEMAVectorQuantizer(\n n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,\n )\n self.quant_n = cfg.quant_n\n self.quantize_layer = nn.Sequential(\n nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),\n nn.Tanh(),\n nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim) # for quantize\n )\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.update","uri":"program://CREMA/function/lavis.models.beats.Tokenizers.update#L68-L69","kind":"function","name":"update","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":68,"end_line":69,"context_start_line":48,"context_end_line":89,"code":" self.encoder_layerdrop: float = 0.0 # probability of dropping a tarnsformer layer\n self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # quantizer\n self.quant_n: int = 1024 # codebook number in quantizer\n self.quant_dim: int = 256 # codebook dimension in quantizer\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass Tokenizers(nn.Module):\n def __init__(\n self,\n cfg: TokenizersConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"Tokenizers Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.forward_padding_mask","uri":"program://CREMA/function/lavis.models.beats.Tokenizers.forward_padding_mask#L109-L121","kind":"function","name":"forward_padding_mask","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":109,"end_line":121,"context_start_line":89,"context_end_line":141,"code":" self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n self.quantize = NormEMAVectorQuantizer(\n n_embed=cfg.quant_n, embedding_dim=cfg.quant_dim, beta=1.0, kmeans_init=True, decay=0.99,\n )\n self.quant_n = cfg.quant_n\n self.quantize_layer = nn.Sequential(\n nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),\n nn.Tanh(),\n nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim) # for quantize\n )\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_labels(\n self,\n source: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.preprocess","uri":"program://CREMA/function/lavis.models.beats.Tokenizers.preprocess#L123-L136","kind":"function","name":"preprocess","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":123,"end_line":136,"context_start_line":103,"context_end_line":156,"code":" self.quantize_layer = nn.Sequential(\n nn.Linear(cfg.encoder_embed_dim, cfg.encoder_embed_dim),\n nn.Tanh(),\n nn.Linear(cfg.encoder_embed_dim, cfg.quant_dim) # for quantize\n )\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_labels(\n self,\n source: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.Tokenizers.extract_labels","uri":"program://CREMA/function/lavis.models.beats.Tokenizers.extract_labels#L138-L172","kind":"function","name":"extract_labels","path":"lavis/models/beats/Tokenizers.py","language":"python","start_line":138,"end_line":172,"context_start_line":118,"context_end_line":173,"code":" padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_labels(\n self,\n source: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n quantize_input = self.quantize_layer(x)\n quantize_feature, embed_loss, embed_ind = self.quantize(quantize_input)\n\n return embed_ind\n","source_hash":"f668340bfadad3739c7fffd1ace2ceb917cd46da38a2b197a617de6b7e6f2fa3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs","uri":"program://CREMA/module/lavis.models.beats.BEATs#L1-L180","kind":"module","name":"lavis.models.beats.BEATs","path":"lavis/models/beats/BEATs.py","language":"python","start_line":1,"end_line":180,"context_start_line":1,"context_end_line":180,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import LayerNorm\nimport torchaudio.compliance.kaldi as ta_kaldi\n\nfrom lavis.models.beats.backbone import (\n TransformerEncoder,\n)\n\nimport logging\nfrom typing import Optional\n\nlogger = logging.getLogger(__name__)\n\n\nclass BEATsConfig:\n def __init__(self, cfg=None):\n self.input_patch_size: int = -1 # path size of patch embedding\n self.embed_dim: int = 512 # patch embedding dimension\n self.conv_bias: bool = False # include bias in conv encoder\n\n self.encoder_layers: int = 12 # num encoder layers in the transformer\n self.encoder_embed_dim: int = 768 # encoder embedding dimension\n self.encoder_ffn_embed_dim: int = 3072 # encoder embedding dimension for FFN\n self.encoder_attention_heads: int = 12 # num encoder attention heads\n self.activation_fn: str = \"gelu\" # activation function to use\n\n self.layer_wise_gradient_decay_ratio: float = 1.0 # ratio for layer-wise gradient decay\n self.layer_norm_first: bool = False # apply layernorm first in the transformer\n self.deep_norm: bool = False # apply deep_norm first in the transformer\n\n # dropouts\n self.dropout: float = 0.1 # dropout probability for the transformer\n self.attention_dropout: float = 0.1 # dropout probability for attention weights\n self.activation_dropout: float = 0.0 # dropout probability after activation in FFN\n self.encoder_layerdrop: float = 0.0 # probability of dropping a tarnsformer layer\n self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # label predictor\n self.finetuned_model: bool = False # whether the model is a fine-tuned model.\n self.predictor_dropout: float = 0.1 # dropout probability for the predictor\n self.predictor_class: int = 527 # target class number for the predictor\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass BEATs(nn.Module):\n def __init__(\n self,\n cfg: BEATsConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"BEATs Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n if cfg.finetuned_model:\n self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)\n self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)\n else:\n self.predictor = None\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_features(\n self,\n fbank: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n ## NOTE: preprocessing is done separately in lavis.processsors.audio_processors.BeatsAudioProcessor\n # fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n if self.predictor is not None:\n x = self.predictor_dropout(x)\n logits = self.predictor(x)\n\n if padding_mask is not None and padding_mask.any():\n logits[padding_mask] = 0\n logits = logits.sum(dim=1)\n logits = logits / (~padding_mask).sum(dim=1).unsqueeze(-1).expand_as(logits)\n else:\n logits = logits.mean(dim=1)\n\n lprobs = torch.sigmoid(logits)\n\n return lprobs, padding_mask\n else:\n return x, padding_mask","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.BEATsConfig","uri":"program://CREMA/class/lavis.models.beats.BEATs.BEATsConfig#L26-L68","kind":"class","name":"BEATsConfig","path":"lavis/models/beats/BEATs.py","language":"python","start_line":26,"end_line":68,"context_start_line":6,"context_end_line":88,"code":"# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\n\nimport torch\nimport torch.nn as nn\nfrom torch.nn import LayerNorm\nimport torchaudio.compliance.kaldi as ta_kaldi\n\nfrom lavis.models.beats.backbone import (\n TransformerEncoder,\n)\n\nimport logging\nfrom typing import Optional\n\nlogger = logging.getLogger(__name__)\n\n\nclass BEATsConfig:\n def __init__(self, cfg=None):\n self.input_patch_size: int = -1 # path size of patch embedding\n self.embed_dim: int = 512 # patch embedding dimension\n self.conv_bias: bool = False # include bias in conv encoder\n\n self.encoder_layers: int = 12 # num encoder layers in the transformer\n self.encoder_embed_dim: int = 768 # encoder embedding dimension\n self.encoder_ffn_embed_dim: int = 3072 # encoder embedding dimension for FFN\n self.encoder_attention_heads: int = 12 # num encoder attention heads\n self.activation_fn: str = \"gelu\" # activation function to use\n\n self.layer_wise_gradient_decay_ratio: float = 1.0 # ratio for layer-wise gradient decay\n self.layer_norm_first: bool = False # apply layernorm first in the transformer\n self.deep_norm: bool = False # apply deep_norm first in the transformer\n\n # dropouts\n self.dropout: float = 0.1 # dropout probability for the transformer\n self.attention_dropout: float = 0.1 # dropout probability for attention weights\n self.activation_dropout: float = 0.0 # dropout probability after activation in FFN\n self.encoder_layerdrop: float = 0.0 # probability of dropping a tarnsformer layer\n self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # label predictor\n self.finetuned_model: bool = False # whether the model is a fine-tuned model.\n self.predictor_dropout: float = 0.1 # dropout probability for the predictor\n self.predictor_class: int = 527 # target class number for the predictor\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass BEATs(nn.Module):\n def __init__(\n self,\n cfg: BEATsConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"BEATs Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.BEATs","uri":"program://CREMA/class/lavis.models.beats.BEATs.BEATs#L71-L180","kind":"class","name":"BEATs","path":"lavis/models/beats/BEATs.py","language":"python","start_line":71,"end_line":180,"context_start_line":51,"context_end_line":180,"code":" self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # label predictor\n self.finetuned_model: bool = False # whether the model is a fine-tuned model.\n self.predictor_dropout: float = 0.1 # dropout probability for the predictor\n self.predictor_class: int = 527 # target class number for the predictor\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass BEATs(nn.Module):\n def __init__(\n self,\n cfg: BEATsConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"BEATs Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n if cfg.finetuned_model:\n self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)\n self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)\n else:\n self.predictor = None\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_features(\n self,\n fbank: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n ## NOTE: preprocessing is done separately in lavis.processsors.audio_processors.BeatsAudioProcessor\n # fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n if self.predictor is not None:\n x = self.predictor_dropout(x)\n logits = self.predictor(x)\n\n if padding_mask is not None and padding_mask.any():\n logits[padding_mask] = 0\n logits = logits.sum(dim=1)\n logits = logits / (~padding_mask).sum(dim=1).unsqueeze(-1).expand_as(logits)\n else:\n logits = logits.mean(dim=1)\n\n lprobs = torch.sigmoid(logits)\n\n return lprobs, padding_mask\n else:\n return x, padding_mask","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.__init__","uri":"program://CREMA/function/lavis.models.beats.BEATs.__init__#L72-L102","kind":"function","name":"__init__","path":"lavis/models/beats/BEATs.py","language":"python","start_line":72,"end_line":102,"context_start_line":52,"context_end_line":122,"code":"\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # label predictor\n self.finetuned_model: bool = False # whether the model is a fine-tuned model.\n self.predictor_dropout: float = 0.1 # dropout probability for the predictor\n self.predictor_class: int = 527 # target class number for the predictor\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass BEATs(nn.Module):\n def __init__(\n self,\n cfg: BEATsConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"BEATs Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n if cfg.finetuned_model:\n self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)\n self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)\n else:\n self.predictor = None\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.update","uri":"program://CREMA/function/lavis.models.beats.BEATs.update#L67-L68","kind":"function","name":"update","path":"lavis/models/beats/BEATs.py","language":"python","start_line":67,"end_line":68,"context_start_line":47,"context_end_line":88,"code":" self.dropout_input: float = 0.0 # dropout to apply to the input (after feat extr)\n\n # positional embeddings\n self.conv_pos: int = 128 # number of filters for convolutional positional embeddings\n self.conv_pos_groups: int = 16 # number of groups for convolutional positional embedding\n\n # relative position embedding\n self.relative_position_embedding: bool = False # apply relative position embedding\n self.num_buckets: int = 320 # number of buckets for relative position embedding\n self.max_distance: int = 1280 # maximum distance for relative position embedding\n self.gru_rel_pos: bool = False # apply gated relative position embedding\n\n # label predictor\n self.finetuned_model: bool = False # whether the model is a fine-tuned model.\n self.predictor_dropout: float = 0.1 # dropout probability for the predictor\n self.predictor_class: int = 527 # target class number for the predictor\n\n if cfg is not None:\n self.update(cfg)\n\n def update(self, cfg: dict):\n self.__dict__.update(cfg)\n\n\nclass BEATs(nn.Module):\n def __init__(\n self,\n cfg: BEATsConfig,\n ) -> None:\n super().__init__()\n logger.info(f\"BEATs Config: {cfg.__dict__}\")\n\n self.cfg = cfg\n\n self.embed = cfg.embed_dim\n self.post_extract_proj = (\n nn.Linear(self.embed, cfg.encoder_embed_dim)\n if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.forward_padding_mask","uri":"program://CREMA/function/lavis.models.beats.BEATs.forward_padding_mask#L104-L116","kind":"function","name":"forward_padding_mask","path":"lavis/models/beats/BEATs.py","language":"python","start_line":104,"end_line":116,"context_start_line":84,"context_end_line":136,"code":" if self.embed != cfg.encoder_embed_dim\n else None\n )\n\n self.input_patch_size = cfg.input_patch_size\n self.patch_embedding = nn.Conv2d(1, self.embed, kernel_size=self.input_patch_size, stride=self.input_patch_size,\n bias=cfg.conv_bias)\n\n self.dropout_input = nn.Dropout(cfg.dropout_input)\n\n assert not cfg.deep_norm or not cfg.layer_norm_first\n self.encoder = TransformerEncoder(cfg)\n self.layer_norm = LayerNorm(self.embed)\n\n if cfg.finetuned_model:\n self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)\n self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)\n else:\n self.predictor = None\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_features(\n self,\n fbank: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.preprocess","uri":"program://CREMA/function/lavis.models.beats.BEATs.preprocess#L118-L131","kind":"function","name":"preprocess","path":"lavis/models/beats/BEATs.py","language":"python","start_line":118,"end_line":131,"context_start_line":98,"context_end_line":151,"code":" if cfg.finetuned_model:\n self.predictor_dropout = nn.Dropout(cfg.predictor_dropout)\n self.predictor = nn.Linear(cfg.encoder_embed_dim, cfg.predictor_class)\n else:\n self.predictor = None\n\n def forward_padding_mask(\n self,\n features: torch.Tensor,\n padding_mask: torch.Tensor,\n ) -> torch.Tensor:\n extra = padding_mask.size(1) % features.size(1)\n if extra > 0:\n padding_mask = padding_mask[:, :-extra]\n padding_mask = padding_mask.view(\n padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_features(\n self,\n fbank: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n ## NOTE: preprocessing is done separately in lavis.processsors.audio_processors.BeatsAudioProcessor\n # fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.BEATs.extract_features","uri":"program://CREMA/function/lavis.models.beats.BEATs.extract_features#L133-L180","kind":"function","name":"extract_features","path":"lavis/models/beats/BEATs.py","language":"python","start_line":133,"end_line":180,"context_start_line":113,"context_end_line":180,"code":" padding_mask.size(0), features.size(1), -1\n )\n padding_mask = padding_mask.all(-1)\n return padding_mask\n\n def preprocess(\n self,\n source: torch.Tensor,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ) -> torch.Tensor:\n fbanks = []\n for waveform in source:\n waveform = waveform.unsqueeze(0) * 2 ** 15\n fbank = ta_kaldi.fbank(waveform, num_mel_bins=128, sample_frequency=16000, frame_length=25, frame_shift=10)\n fbanks.append(fbank)\n fbank = torch.stack(fbanks, dim=0)\n fbank = (fbank - fbank_mean) / (2 * fbank_std)\n return fbank\n\n def extract_features(\n self,\n fbank: torch.Tensor,\n padding_mask: Optional[torch.Tensor] = None,\n fbank_mean: float = 15.41663,\n fbank_std: float = 6.55582,\n ):\n ## NOTE: preprocessing is done separately in lavis.processsors.audio_processors.BeatsAudioProcessor\n # fbank = self.preprocess(source, fbank_mean=fbank_mean, fbank_std=fbank_std)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(fbank, padding_mask)\n\n fbank = fbank.unsqueeze(1)\n features = self.patch_embedding(fbank)\n features = features.reshape(features.shape[0], features.shape[1], -1)\n features = features.transpose(1, 2)\n features = self.layer_norm(features)\n\n if padding_mask is not None:\n padding_mask = self.forward_padding_mask(features, padding_mask)\n\n if self.post_extract_proj is not None:\n features = self.post_extract_proj(features)\n\n x = self.dropout_input(features)\n\n x, layer_results = self.encoder(\n x,\n padding_mask=padding_mask,\n )\n\n if self.predictor is not None:\n x = self.predictor_dropout(x)\n logits = self.predictor(x)\n\n if padding_mask is not None and padding_mask.any():\n logits[padding_mask] = 0\n logits = logits.sum(dim=1)\n logits = logits / (~padding_mask).sum(dim=1).unsqueeze(-1).expand_as(logits)\n else:\n logits = logits.mean(dim=1)\n\n lprobs = torch.sigmoid(logits)\n\n return lprobs, padding_mask\n else:\n return x, padding_mask","source_hash":"695b52030c8ecf57b04b8171928733064420e97c5e4afe867d35d63633a691f2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer","uri":"program://CREMA/module/lavis.models.beats.quantizer#L1-L215","kind":"module","name":"lavis.models.beats.quantizer","path":"lavis/models/beats/quantizer.py","language":"python","start_line":1,"end_line":215,"context_start_line":1,"context_end_line":215,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on VQGAN code bases\n# https://github.com/CompVis/taming-transformers\n# --------------------------------------------------------'\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.distributed as distributed\n\ntry:\n from einops import rearrange, repeat\nexcept ImportError:\n pass\n\n\ndef l2norm(t):\n return F.normalize(t, p=2, dim=-1)\n\n\ndef ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n\n\ndef sample_vectors(samples, num):\n num_samples, device = samples.shape[0], samples.device\n\n if num_samples >= num:\n indices = torch.randperm(num_samples, device=device)[:num]\n else:\n indices = torch.randint(0, num_samples, (num,), device=device)\n\n return samples[indices]\n\n\ndef kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):\n dim, dtype, device = samples.shape[-1], samples.dtype, samples.device\n\n means = sample_vectors(samples, num_clusters)\n\n for _ in range(num_iters):\n if use_cosine_sim:\n dists = samples @ means.t()\n else:\n diffs = rearrange(samples, 'n d -> n () d') \\\n - rearrange(means, 'c d -> () c d')\n dists = -(diffs ** 2).sum(dim=-1)\n\n buckets = dists.max(dim=-1).indices\n bins = torch.bincount(buckets, minlength=num_clusters)\n zero_mask = bins == 0\n bins_min_clamped = bins.masked_fill(zero_mask, 1)\n\n new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype)\n new_means.scatter_add_(0, repeat(buckets, 'n -> n d', d=dim), samples)\n new_means = new_means / bins_min_clamped[..., None]\n\n if use_cosine_sim:\n new_means = l2norm(new_means)\n\n means = torch.where(zero_mask[..., None], means, new_means)\n\n return means, bins\n\n\nclass EmbeddingEMA(nn.Module):\n def __init__(self, num_tokens, codebook_dim, decay=0.99, eps=1e-5, kmeans_init=True, codebook_init_path=''):\n super().__init__()\n self.num_tokens = num_tokens\n self.codebook_dim = codebook_dim\n self.decay = decay\n self.eps = eps\n if codebook_init_path == '':\n if not kmeans_init:\n weight = torch.randn(num_tokens, codebook_dim)\n weight = l2norm(weight)\n else:\n weight = torch.zeros(num_tokens, codebook_dim)\n self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n else:\n print(f\"load init codebook weight from {codebook_init_path}\")\n codebook_ckpt_weight = torch.load(codebook_init_path, map_location='cpu')\n weight = codebook_ckpt_weight.clone()\n self.register_buffer('initted', torch.Tensor([True]))\n\n self.weight = nn.Parameter(weight, requires_grad=False)\n self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False)\n self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)\n # self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n self.update = True\n\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce\n else:\n self.all_reduce_fn = nn.Identity()\n\n def reset_cluster_size(self, device):\n if self.statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(self.num_tokens))\n self.cluster_size = self.cluster_size.to(device)\n\n def forward(self, z):\n # reshape z -> (batch, height, width, channel) and flatten\n # z, 'b c h w -> b h w c'\n # z = rearrange(z, 'b c h w -> b h w c')\n # z = z.transpose(1, 2)\n z = l2norm(z)\n z_flattened = z.reshape(-1, self.codebook_dim)\n\n self.embedding.init_embed_(z_flattened)\n\n d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \\\n self.embedding.weight.pow(2).sum(dim=1) - 2 * \\\n torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight) # 'n d -> d n'\n\n encoding_indices = torch.argmin(d, dim=1)\n\n z_q = self.embedding(encoding_indices).view(z.shape)\n\n encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype)\n\n if not self.training:\n with torch.no_grad():\n cluster_size = encodings.sum(0)\n self.all_reduce_fn(cluster_size)\n ema_inplace(self.cluster_size, cluster_size, self.decay)\n\n if self.training and self.embedding.update:\n # EMA cluster size\n\n bins = encodings.sum(0)\n self.all_reduce_fn(bins)\n\n # self.embedding.cluster_size_ema_update(bins)\n ema_inplace(self.cluster_size, bins, self.decay)\n\n zero_mask = (bins == 0)\n bins = bins.masked_fill(zero_mask, 1.)\n\n embed_sum = z_flattened.t() @ encodings\n self.all_reduce_fn(embed_sum)\n\n embed_normalized = (embed_sum / bins.unsqueeze(0)).t()\n embed_normalized = l2norm(embed_normalized)\n\n embed_normalized = torch.where(zero_mask[..., None], self.embedding.weight,\n embed_normalized)\n norm_ema_inplace(self.embedding.weight, embed_normalized, self.decay)\n\n # compute loss for embedding\n loss = self.beta * F.mse_loss(z_q.detach(), z)\n\n # preserve gradients\n z_q = z + (z_q - z).detach()\n\n # reshape back to match original input shape\n # z_q, 'b h w c -> b c h w'\n # z_q = rearrange(z_q, 'b h w c -> b c h w')\n # z_q = z_q.transpose(1, 2)\n return z_q, loss, encoding_indices","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.l2norm","uri":"program://CREMA/function/lavis.models.beats.quantizer.l2norm#L21-L22","kind":"function","name":"l2norm","path":"lavis/models/beats/quantizer.py","language":"python","start_line":21,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on VQGAN code bases\n# https://github.com/CompVis/taming-transformers\n# --------------------------------------------------------'\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.distributed as distributed\n\ntry:\n from einops import rearrange, repeat\nexcept ImportError:\n pass\n\n\ndef l2norm(t):\n return F.normalize(t, p=2, dim=-1)\n\n\ndef ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n\n\ndef sample_vectors(samples, num):\n num_samples, device = samples.shape[0], samples.device\n\n if num_samples >= num:\n indices = torch.randperm(num_samples, device=device)[:num]\n else:\n indices = torch.randint(0, num_samples, (num,), device=device)\n\n return samples[indices]\n\n\ndef kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):\n dim, dtype, device = samples.shape[-1], samples.dtype, samples.device\n","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.ema_inplace","uri":"program://CREMA/function/lavis.models.beats.quantizer.ema_inplace#L25-L26","kind":"function","name":"ema_inplace","path":"lavis/models/beats/quantizer.py","language":"python","start_line":25,"end_line":26,"context_start_line":5,"context_end_line":46,"code":"# Licensed under The MIT License [see LICENSE for details]\n# Based on VQGAN code bases\n# https://github.com/CompVis/taming-transformers\n# --------------------------------------------------------'\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.distributed as distributed\n\ntry:\n from einops import rearrange, repeat\nexcept ImportError:\n pass\n\n\ndef l2norm(t):\n return F.normalize(t, p=2, dim=-1)\n\n\ndef ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n\n\ndef sample_vectors(samples, num):\n num_samples, device = samples.shape[0], samples.device\n\n if num_samples >= num:\n indices = torch.randperm(num_samples, device=device)[:num]\n else:\n indices = torch.randint(0, num_samples, (num,), device=device)\n\n return samples[indices]\n\n\ndef kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):\n dim, dtype, device = samples.shape[-1], samples.dtype, samples.device\n\n means = sample_vectors(samples, num_clusters)\n\n for _ in range(num_iters):\n if use_cosine_sim:","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.sample_vectors","uri":"program://CREMA/function/lavis.models.beats.quantizer.sample_vectors#L29-L37","kind":"function","name":"sample_vectors","path":"lavis/models/beats/quantizer.py","language":"python","start_line":29,"end_line":37,"context_start_line":9,"context_end_line":57,"code":"\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.distributed as distributed\n\ntry:\n from einops import rearrange, repeat\nexcept ImportError:\n pass\n\n\ndef l2norm(t):\n return F.normalize(t, p=2, dim=-1)\n\n\ndef ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n\n\ndef sample_vectors(samples, num):\n num_samples, device = samples.shape[0], samples.device\n\n if num_samples >= num:\n indices = torch.randperm(num_samples, device=device)[:num]\n else:\n indices = torch.randint(0, num_samples, (num,), device=device)\n\n return samples[indices]\n\n\ndef kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):\n dim, dtype, device = samples.shape[-1], samples.dtype, samples.device\n\n means = sample_vectors(samples, num_clusters)\n\n for _ in range(num_iters):\n if use_cosine_sim:\n dists = samples @ means.t()\n else:\n diffs = rearrange(samples, 'n d -> n () d') \\\n - rearrange(means, 'c d -> () c d')\n dists = -(diffs ** 2).sum(dim=-1)\n\n buckets = dists.max(dim=-1).indices\n bins = torch.bincount(buckets, minlength=num_clusters)\n zero_mask = bins == 0\n bins_min_clamped = bins.masked_fill(zero_mask, 1)\n","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.kmeans","uri":"program://CREMA/function/lavis.models.beats.quantizer.kmeans#L40-L67","kind":"function","name":"kmeans","path":"lavis/models/beats/quantizer.py","language":"python","start_line":40,"end_line":67,"context_start_line":20,"context_end_line":87,"code":"\ndef l2norm(t):\n return F.normalize(t, p=2, dim=-1)\n\n\ndef ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n\n\ndef sample_vectors(samples, num):\n num_samples, device = samples.shape[0], samples.device\n\n if num_samples >= num:\n indices = torch.randperm(num_samples, device=device)[:num]\n else:\n indices = torch.randint(0, num_samples, (num,), device=device)\n\n return samples[indices]\n\n\ndef kmeans(samples, num_clusters, num_iters=10, use_cosine_sim=False):\n dim, dtype, device = samples.shape[-1], samples.dtype, samples.device\n\n means = sample_vectors(samples, num_clusters)\n\n for _ in range(num_iters):\n if use_cosine_sim:\n dists = samples @ means.t()\n else:\n diffs = rearrange(samples, 'n d -> n () d') \\\n - rearrange(means, 'c d -> () c d')\n dists = -(diffs ** 2).sum(dim=-1)\n\n buckets = dists.max(dim=-1).indices\n bins = torch.bincount(buckets, minlength=num_clusters)\n zero_mask = bins == 0\n bins_min_clamped = bins.masked_fill(zero_mask, 1)\n\n new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype)\n new_means.scatter_add_(0, repeat(buckets, 'n -> n d', d=dim), samples)\n new_means = new_means / bins_min_clamped[..., None]\n\n if use_cosine_sim:\n new_means = l2norm(new_means)\n\n means = torch.where(zero_mask[..., None], means, new_means)\n\n return means, bins\n\n\nclass EmbeddingEMA(nn.Module):\n def __init__(self, num_tokens, codebook_dim, decay=0.99, eps=1e-5, kmeans_init=True, codebook_init_path=''):\n super().__init__()\n self.num_tokens = num_tokens\n self.codebook_dim = codebook_dim\n self.decay = decay\n self.eps = eps\n if codebook_init_path == '':\n if not kmeans_init:\n weight = torch.randn(num_tokens, codebook_dim)\n weight = l2norm(weight)\n else:\n weight = torch.zeros(num_tokens, codebook_dim)\n self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n else:\n print(f\"load init codebook weight from {codebook_init_path}\")\n codebook_ckpt_weight = torch.load(codebook_init_path, map_location='cpu')\n weight = codebook_ckpt_weight.clone()","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.EmbeddingEMA","uri":"program://CREMA/class/lavis.models.beats.quantizer.EmbeddingEMA#L70-L123","kind":"class","name":"EmbeddingEMA","path":"lavis/models/beats/quantizer.py","language":"python","start_line":70,"end_line":123,"context_start_line":50,"context_end_line":143,"code":" - rearrange(means, 'c d -> () c d')\n dists = -(diffs ** 2).sum(dim=-1)\n\n buckets = dists.max(dim=-1).indices\n bins = torch.bincount(buckets, minlength=num_clusters)\n zero_mask = bins == 0\n bins_min_clamped = bins.masked_fill(zero_mask, 1)\n\n new_means = buckets.new_zeros(num_clusters, dim, dtype=dtype)\n new_means.scatter_add_(0, repeat(buckets, 'n -> n d', d=dim), samples)\n new_means = new_means / bins_min_clamped[..., None]\n\n if use_cosine_sim:\n new_means = l2norm(new_means)\n\n means = torch.where(zero_mask[..., None], means, new_means)\n\n return means, bins\n\n\nclass EmbeddingEMA(nn.Module):\n def __init__(self, num_tokens, codebook_dim, decay=0.99, eps=1e-5, kmeans_init=True, codebook_init_path=''):\n super().__init__()\n self.num_tokens = num_tokens\n self.codebook_dim = codebook_dim\n self.decay = decay\n self.eps = eps\n if codebook_init_path == '':\n if not kmeans_init:\n weight = torch.randn(num_tokens, codebook_dim)\n weight = l2norm(weight)\n else:\n weight = torch.zeros(num_tokens, codebook_dim)\n self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n else:\n print(f\"load init codebook weight from {codebook_init_path}\")\n codebook_ckpt_weight = torch.load(codebook_init_path, map_location='cpu')\n weight = codebook_ckpt_weight.clone()\n self.register_buffer('initted', torch.Tensor([True]))\n\n self.weight = nn.Parameter(weight, requires_grad=False)\n self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False)\n self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)\n # self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n self.update = True\n\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.norm_ema_inplace","uri":"program://CREMA/function/lavis.models.beats.quantizer.norm_ema_inplace#L126-L128","kind":"function","name":"norm_ema_inplace","path":"lavis/models/beats/quantizer.py","language":"python","start_line":126,"end_line":128,"context_start_line":106,"context_end_line":148,"code":" def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.NormEMAVectorQuantizer","uri":"program://CREMA/class/lavis.models.beats.quantizer.NormEMAVectorQuantizer#L131-L215","kind":"class","name":"NormEMAVectorQuantizer","path":"lavis/models/beats/quantizer.py","language":"python","start_line":131,"end_line":215,"context_start_line":111,"context_end_line":215,"code":"\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce\n else:\n self.all_reduce_fn = nn.Identity()\n\n def reset_cluster_size(self, device):\n if self.statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(self.num_tokens))\n self.cluster_size = self.cluster_size.to(device)\n\n def forward(self, z):\n # reshape z -> (batch, height, width, channel) and flatten\n # z, 'b c h w -> b h w c'\n # z = rearrange(z, 'b c h w -> b h w c')\n # z = z.transpose(1, 2)\n z = l2norm(z)\n z_flattened = z.reshape(-1, self.codebook_dim)\n\n self.embedding.init_embed_(z_flattened)\n\n d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \\\n self.embedding.weight.pow(2).sum(dim=1) - 2 * \\\n torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight) # 'n d -> d n'\n\n encoding_indices = torch.argmin(d, dim=1)\n\n z_q = self.embedding(encoding_indices).view(z.shape)\n\n encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype)\n\n if not self.training:\n with torch.no_grad():\n cluster_size = encodings.sum(0)\n self.all_reduce_fn(cluster_size)\n ema_inplace(self.cluster_size, cluster_size, self.decay)\n\n if self.training and self.embedding.update:\n # EMA cluster size\n\n bins = encodings.sum(0)\n self.all_reduce_fn(bins)\n\n # self.embedding.cluster_size_ema_update(bins)\n ema_inplace(self.cluster_size, bins, self.decay)\n\n zero_mask = (bins == 0)\n bins = bins.masked_fill(zero_mask, 1.)\n\n embed_sum = z_flattened.t() @ encodings\n self.all_reduce_fn(embed_sum)\n\n embed_normalized = (embed_sum / bins.unsqueeze(0)).t()\n embed_normalized = l2norm(embed_normalized)\n\n embed_normalized = torch.where(zero_mask[..., None], self.embedding.weight,\n embed_normalized)\n norm_ema_inplace(self.embedding.weight, embed_normalized, self.decay)\n\n # compute loss for embedding\n loss = self.beta * F.mse_loss(z_q.detach(), z)\n\n # preserve gradients\n z_q = z + (z_q - z).detach()\n\n # reshape back to match original input shape\n # z_q, 'b h w c -> b c h w'\n # z_q = rearrange(z_q, 'b h w c -> b c h w')\n # z_q = z_q.transpose(1, 2)\n return z_q, loss, encoding_indices","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.__init__","uri":"program://CREMA/function/lavis.models.beats.quantizer.__init__#L132-L150","kind":"function","name":"__init__","path":"lavis/models/beats/quantizer.py","language":"python","start_line":132,"end_line":150,"context_start_line":112,"context_end_line":170,"code":" def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce\n else:\n self.all_reduce_fn = nn.Identity()\n\n def reset_cluster_size(self, device):\n if self.statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(self.num_tokens))\n self.cluster_size = self.cluster_size.to(device)\n\n def forward(self, z):\n # reshape z -> (batch, height, width, channel) and flatten\n # z, 'b c h w -> b h w c'\n # z = rearrange(z, 'b c h w -> b h w c')\n # z = z.transpose(1, 2)\n z = l2norm(z)\n z_flattened = z.reshape(-1, self.codebook_dim)\n\n self.embedding.init_embed_(z_flattened)\n\n d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \\\n self.embedding.weight.pow(2).sum(dim=1) - 2 * \\\n torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight) # 'n d -> d n'\n","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.init_embed_","uri":"program://CREMA/function/lavis.models.beats.quantizer.init_embed_#L97-L104","kind":"function","name":"init_embed_","path":"lavis/models/beats/quantizer.py","language":"python","start_line":97,"end_line":104,"context_start_line":77,"context_end_line":124,"code":" if codebook_init_path == '':\n if not kmeans_init:\n weight = torch.randn(num_tokens, codebook_dim)\n weight = l2norm(weight)\n else:\n weight = torch.zeros(num_tokens, codebook_dim)\n self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n else:\n print(f\"load init codebook weight from {codebook_init_path}\")\n codebook_ckpt_weight = torch.load(codebook_init_path, map_location='cpu')\n weight = codebook_ckpt_weight.clone()\n self.register_buffer('initted', torch.Tensor([True]))\n\n self.weight = nn.Parameter(weight, requires_grad=False)\n self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False)\n self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)\n # self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n self.update = True\n\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.forward","uri":"program://CREMA/function/lavis.models.beats.quantizer.forward#L157-L215","kind":"function","name":"forward","path":"lavis/models/beats/quantizer.py","language":"python","start_line":157,"end_line":215,"context_start_line":137,"context_end_line":215,"code":" self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce\n else:\n self.all_reduce_fn = nn.Identity()\n\n def reset_cluster_size(self, device):\n if self.statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(self.num_tokens))\n self.cluster_size = self.cluster_size.to(device)\n\n def forward(self, z):\n # reshape z -> (batch, height, width, channel) and flatten\n # z, 'b c h w -> b h w c'\n # z = rearrange(z, 'b c h w -> b h w c')\n # z = z.transpose(1, 2)\n z = l2norm(z)\n z_flattened = z.reshape(-1, self.codebook_dim)\n\n self.embedding.init_embed_(z_flattened)\n\n d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \\\n self.embedding.weight.pow(2).sum(dim=1) - 2 * \\\n torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight) # 'n d -> d n'\n\n encoding_indices = torch.argmin(d, dim=1)\n\n z_q = self.embedding(encoding_indices).view(z.shape)\n\n encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype)\n\n if not self.training:\n with torch.no_grad():\n cluster_size = encodings.sum(0)\n self.all_reduce_fn(cluster_size)\n ema_inplace(self.cluster_size, cluster_size, self.decay)\n\n if self.training and self.embedding.update:\n # EMA cluster size\n\n bins = encodings.sum(0)\n self.all_reduce_fn(bins)\n\n # self.embedding.cluster_size_ema_update(bins)\n ema_inplace(self.cluster_size, bins, self.decay)\n\n zero_mask = (bins == 0)\n bins = bins.masked_fill(zero_mask, 1.)\n\n embed_sum = z_flattened.t() @ encodings\n self.all_reduce_fn(embed_sum)\n\n embed_normalized = (embed_sum / bins.unsqueeze(0)).t()\n embed_normalized = l2norm(embed_normalized)\n\n embed_normalized = torch.where(zero_mask[..., None], self.embedding.weight,\n embed_normalized)\n norm_ema_inplace(self.embedding.weight, embed_normalized, self.decay)\n\n # compute loss for embedding\n loss = self.beta * F.mse_loss(z_q.detach(), z)\n\n # preserve gradients\n z_q = z + (z_q - z).detach()\n\n # reshape back to match original input shape\n # z_q, 'b h w c -> b c h w'\n # z_q = rearrange(z_q, 'b h w c -> b c h w')\n # z_q = z_q.transpose(1, 2)\n return z_q, loss, encoding_indices","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.cluster_size_ema_update","uri":"program://CREMA/function/lavis.models.beats.quantizer.cluster_size_ema_update#L109-L110","kind":"function","name":"cluster_size_ema_update","path":"lavis/models/beats/quantizer.py","language":"python","start_line":109,"end_line":110,"context_start_line":89,"context_end_line":130,"code":"\n self.weight = nn.Parameter(weight, requires_grad=False)\n self.cluster_size = nn.Parameter(torch.zeros(num_tokens), requires_grad=False)\n self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)\n # self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n self.update = True\n\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.embed_avg_ema_update","uri":"program://CREMA/function/lavis.models.beats.quantizer.embed_avg_ema_update#L112-L113","kind":"function","name":"embed_avg_ema_update","path":"lavis/models/beats/quantizer.py","language":"python","start_line":112,"end_line":113,"context_start_line":92,"context_end_line":133,"code":" self.embed_avg = nn.Parameter(weight.clone(), requires_grad=False)\n # self.register_buffer('initted', torch.Tensor([not kmeans_init]))\n self.update = True\n\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.weight_update","uri":"program://CREMA/function/lavis.models.beats.quantizer.weight_update#L115-L123","kind":"function","name":"weight_update","path":"lavis/models/beats/quantizer.py","language":"python","start_line":115,"end_line":123,"context_start_line":95,"context_end_line":143,"code":"\n @torch.jit.ignore\n def init_embed_(self, data):\n if self.initted:\n return\n print(\"Performing Kemans init for codebook\")\n embed, cluster_size = kmeans(data, self.num_tokens, 10, use_cosine_sim=True)\n self.weight.data.copy_(embed)\n self.cluster_size.data.copy_(cluster_size)\n self.initted.data.copy_(torch.Tensor([True]))\n\n def forward(self, embed_id):\n return F.embedding(embed_id, self.weight)\n\n def cluster_size_ema_update(self, new_cluster_size):\n self.cluster_size.data.mul_(self.decay).add_(new_cluster_size, alpha=1 - self.decay)\n\n def embed_avg_ema_update(self, new_embed_avg):\n self.embed_avg.data.mul_(self.decay).add_(new_embed_avg, alpha=1 - self.decay)\n\n def weight_update(self, num_tokens):\n n = self.cluster_size.sum()\n smoothed_cluster_size = (\n (self.cluster_size + self.eps) / (n + num_tokens * self.eps) * n\n )\n # normalize embedding average with smoothed cluster size\n embed_normalized = self.embed_avg / smoothed_cluster_size.unsqueeze(1)\n # embed_normalized = l2norm(self.embed_avg / smoothed_cluster_size.unsqueeze(1))\n self.weight.data.copy_(embed_normalized)\n\n\ndef norm_ema_inplace(moving_avg, new, decay):\n moving_avg.data.mul_(decay).add_(new, alpha=(1 - decay))\n moving_avg.data.copy_(l2norm(moving_avg.data))\n\n\nclass NormEMAVectorQuantizer(nn.Module):\n def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.quantizer.reset_cluster_size","uri":"program://CREMA/function/lavis.models.beats.quantizer.reset_cluster_size#L152-L155","kind":"function","name":"reset_cluster_size","path":"lavis/models/beats/quantizer.py","language":"python","start_line":152,"end_line":155,"context_start_line":132,"context_end_line":175,"code":" def __init__(self, n_embed, embedding_dim, beta, decay=0.99, eps=1e-5,\n statistic_code_usage=True, kmeans_init=False, codebook_init_path=''):\n super().__init__()\n self.codebook_dim = embedding_dim\n self.num_tokens = n_embed\n self.beta = beta\n self.decay = decay\n\n # learnable = True if orthogonal_reg_weight > 0 else False\n self.embedding = EmbeddingEMA(self.num_tokens, self.codebook_dim, decay, eps, kmeans_init, codebook_init_path)\n\n self.statistic_code_usage = statistic_code_usage\n if statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(n_embed))\n if distributed.is_available() and distributed.is_initialized():\n print(\"ddp is enable, so use ddp_reduce to sync the statistic_code_usage for each gpu!\")\n self.all_reduce_fn = distributed.all_reduce\n else:\n self.all_reduce_fn = nn.Identity()\n\n def reset_cluster_size(self, device):\n if self.statistic_code_usage:\n self.register_buffer('cluster_size', torch.zeros(self.num_tokens))\n self.cluster_size = self.cluster_size.to(device)\n\n def forward(self, z):\n # reshape z -> (batch, height, width, channel) and flatten\n # z, 'b c h w -> b h w c'\n # z = rearrange(z, 'b c h w -> b h w c')\n # z = z.transpose(1, 2)\n z = l2norm(z)\n z_flattened = z.reshape(-1, self.codebook_dim)\n\n self.embedding.init_embed_(z_flattened)\n\n d = z_flattened.pow(2).sum(dim=1, keepdim=True) + \\\n self.embedding.weight.pow(2).sum(dim=1) - 2 * \\\n torch.einsum('bd,nd->bn', z_flattened, self.embedding.weight) # 'n d -> d n'\n\n encoding_indices = torch.argmin(d, dim=1)\n\n z_q = self.embedding(encoding_indices).view(z.shape)\n\n encodings = F.one_hot(encoding_indices, self.num_tokens).type(z.dtype)","source_hash":"8d15198f22618db773a5fed2ca416fbde67c7d56ce621822ab9d49d51e5534bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules","uri":"program://CREMA/module/lavis.models.beats.modules#L1-L219","kind":"module","name":"lavis.models.beats.modules","path":"lavis/models/beats/modules.py","language":"python","start_line":1,"end_line":219,"context_start_line":1,"context_end_line":219,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\nimport math\nimport warnings\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\n\n\nclass GradMultiply(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, scale):\n ctx.scale = scale\n res = x.new(x)\n return res\n\n @staticmethod\n def backward(ctx, grad):\n return grad * ctx.scale, None\n\n\nclass SamePad(nn.Module):\n def __init__(self, kernel_size, causal=False):\n super().__init__()\n if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()\n\n def forward(self, x):\n return x * self.act(x)\n\n\nclass GLU_Linear(nn.Module):\n def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()\n elif glu_type == \"swish\":\n self.glu_act = Swish()\n elif glu_type == \"relu\":\n self.glu_act = torch.nn.ReLU()\n elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)\n else:\n self.linear = nn.Linear(input_dim, output_dim * 2, False)\n\n def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":\n return gelu\n elif activation == \"gelu_fast\":\n warnings.warn(\n \"--activation-fn=gelu_fast has been renamed to gelu_accurate\"\n )\n return gelu_accurate\n elif activation == \"gelu_accurate\":\n return gelu_accurate\n elif activation == \"tanh\":\n return torch.tanh\n elif activation == \"linear\":\n return lambda x: x\n elif activation == \"glu\":\n return lambda x: x\n else:\n raise RuntimeError(\"--activation-fn {} not supported\".format(activation))\n\n\ndef quant_noise(module, p, block_size):\n \"\"\"\n Wraps modules and applies quantization noise to the weights for\n subsequent quantization with Iterative Product Quantization as\n described in \"Training with Quantization Noise for Extreme Model Compression\"\n\n Args:\n - module: nn.Module\n - p: amount of Quantization Noise\n - block_size: size of the blocks for subsequent quantization with iPQ\n\n Remarks:\n - Module weights must have the right sizes wrt the block size\n - Only Linear, Embedding and Conv2d modules are supported for the moment\n - For more detail on how to quantize by blocks with convolutional weights,\n see \"And the Bit Goes Down: Revisiting the Quantization of Neural Networks\"\n - We implement the simplest form of noise here as stated in the paper\n which consists in randomly dropping blocks\n \"\"\"\n\n # if no quantization noise, don't register hook\n if p <= 0:\n return module\n\n # supported modules\n assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))\n\n # test whether module.weight has the right sizes wrt block_size\n is_conv = module.weight.ndim == 4\n\n # 2D matrix\n if not is_conv:\n assert (\n module.weight.size(1) % block_size == 0\n ), \"Input features must be a multiple of block sizes\"\n\n # 4D matrix\n else:\n # 1x1 convolutions\n if module.kernel_size == (1, 1):\n assert (\n module.in_channels % block_size == 0\n ), \"Input channels must be a multiple of block sizes\"\n # regular convolutions\n else:\n k = module.kernel_size[0] * module.kernel_size[1]\n assert k % block_size == 0, \"Kernel size must be a multiple of block size\"\n\n def _forward_pre_hook(mod, input):\n # no noise for evaluation\n if mod.training:\n if not is_conv:\n # gather weight and sizes\n weight = mod.weight\n in_features = weight.size(1)\n out_features = weight.size(0)\n\n # split weight matrix into blocks and randomly drop selected blocks\n mask = torch.zeros(\n in_features // block_size * out_features, device=weight.device\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)\n\n else:\n # gather weight and sizes\n weight = mod.weight\n in_channels = mod.in_channels\n out_channels = mod.out_channels\n\n # split weight matrix into blocks and randomly drop selected blocks\n if mod.kernel_size == (1, 1):\n mask = torch.zeros(\n int(in_channels // block_size * out_channels),\n device=weight.device,\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)\n else:\n mask = torch.zeros(\n weight.size(0), weight.size(1), device=weight.device\n )\n mask.bernoulli_(p)\n mask = (\n mask.unsqueeze(2)\n .unsqueeze(3)\n .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])\n )\n\n # scale weights and apply mask\n mask = mask.to(\n torch.bool\n ) # x.bool() is not currently supported in TorchScript\n s = 1 / (1 - p)\n mod.weight.data = s * weight.masked_fill(mask, 0)\n\n module.register_forward_pre_hook(_forward_pre_hook)\n return module\n","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.GradMultiply","uri":"program://CREMA/class/lavis.models.beats.modules.GradMultiply#L17-L26","kind":"class","name":"GradMultiply","path":"lavis/models/beats/modules.py","language":"python","start_line":17,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"# --------------------------------------------------------\n# BEATs: Audio Pre-Training with Acoustic Tokenizers (https://arxiv.org/abs/2212.09058)\n# Github source: https://github.com/microsoft/unilm/tree/master/beats\n# Copyright (c) 2022 Microsoft\n# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\nimport math\nimport warnings\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\n\n\nclass GradMultiply(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, scale):\n ctx.scale = scale\n res = x.new(x)\n return res\n\n @staticmethod\n def backward(ctx, grad):\n return grad * ctx.scale, None\n\n\nclass SamePad(nn.Module):\n def __init__(self, kernel_size, causal=False):\n super().__init__()\n if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.SamePad","uri":"program://CREMA/class/lavis.models.beats.modules.SamePad#L29-L40","kind":"class","name":"SamePad","path":"lavis/models/beats/modules.py","language":"python","start_line":29,"end_line":40,"context_start_line":9,"context_end_line":60,"code":"\nimport math\nimport warnings\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\n\n\nclass GradMultiply(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, scale):\n ctx.scale = scale\n res = x.new(x)\n return res\n\n @staticmethod\n def backward(ctx, grad):\n return grad * ctx.scale, None\n\n\nclass SamePad(nn.Module):\n def __init__(self, kernel_size, causal=False):\n super().__init__()\n if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()\n\n def forward(self, x):\n return x * self.act(x)\n\n\nclass GLU_Linear(nn.Module):\n def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.Swish","uri":"program://CREMA/class/lavis.models.beats.modules.Swish#L43-L49","kind":"class","name":"Swish","path":"lavis/models/beats/modules.py","language":"python","start_line":43,"end_line":49,"context_start_line":23,"context_end_line":69,"code":"\n @staticmethod\n def backward(ctx, grad):\n return grad * ctx.scale, None\n\n\nclass SamePad(nn.Module):\n def __init__(self, kernel_size, causal=False):\n super().__init__()\n if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()\n\n def forward(self, x):\n return x * self.act(x)\n\n\nclass GLU_Linear(nn.Module):\n def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()\n elif glu_type == \"swish\":\n self.glu_act = Swish()\n elif glu_type == \"relu\":\n self.glu_act = torch.nn.ReLU()\n elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.GLU_Linear","uri":"program://CREMA/class/lavis.models.beats.modules.GLU_Linear#L52-L82","kind":"class","name":"GLU_Linear","path":"lavis/models/beats/modules.py","language":"python","start_line":52,"end_line":82,"context_start_line":32,"context_end_line":102,"code":" if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()\n\n def forward(self, x):\n return x * self.act(x)\n\n\nclass GLU_Linear(nn.Module):\n def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()\n elif glu_type == \"swish\":\n self.glu_act = Swish()\n elif glu_type == \"relu\":\n self.glu_act = torch.nn.ReLU()\n elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)\n else:\n self.linear = nn.Linear(input_dim, output_dim * 2, False)\n\n def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.gelu_accurate","uri":"program://CREMA/function/lavis.models.beats.modules.gelu_accurate#L85-L90","kind":"function","name":"gelu_accurate","path":"lavis/models/beats/modules.py","language":"python","start_line":85,"end_line":90,"context_start_line":65,"context_end_line":110,"code":" elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)\n else:\n self.linear = nn.Linear(input_dim, output_dim * 2, False)\n\n def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":\n return gelu\n elif activation == \"gelu_fast\":\n warnings.warn(\n \"--activation-fn=gelu_fast has been renamed to gelu_accurate\"\n )\n return gelu_accurate\n elif activation == \"gelu_accurate\":\n return gelu_accurate","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.gelu","uri":"program://CREMA/function/lavis.models.beats.modules.gelu#L93-L94","kind":"function","name":"gelu","path":"lavis/models/beats/modules.py","language":"python","start_line":93,"end_line":94,"context_start_line":73,"context_end_line":114,"code":" def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":\n return gelu\n elif activation == \"gelu_fast\":\n warnings.warn(\n \"--activation-fn=gelu_fast has been renamed to gelu_accurate\"\n )\n return gelu_accurate\n elif activation == \"gelu_accurate\":\n return gelu_accurate\n elif activation == \"tanh\":\n return torch.tanh\n elif activation == \"linear\":\n return lambda x: x","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.get_activation_fn","uri":"program://CREMA/function/lavis.models.beats.modules.get_activation_fn#L97-L118","kind":"function","name":"get_activation_fn","path":"lavis/models/beats/modules.py","language":"python","start_line":97,"end_line":118,"context_start_line":77,"context_end_line":138,"code":" if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":\n return gelu\n elif activation == \"gelu_fast\":\n warnings.warn(\n \"--activation-fn=gelu_fast has been renamed to gelu_accurate\"\n )\n return gelu_accurate\n elif activation == \"gelu_accurate\":\n return gelu_accurate\n elif activation == \"tanh\":\n return torch.tanh\n elif activation == \"linear\":\n return lambda x: x\n elif activation == \"glu\":\n return lambda x: x\n else:\n raise RuntimeError(\"--activation-fn {} not supported\".format(activation))\n\n\ndef quant_noise(module, p, block_size):\n \"\"\"\n Wraps modules and applies quantization noise to the weights for\n subsequent quantization with Iterative Product Quantization as\n described in \"Training with Quantization Noise for Extreme Model Compression\"\n\n Args:\n - module: nn.Module\n - p: amount of Quantization Noise\n - block_size: size of the blocks for subsequent quantization with iPQ\n\n Remarks:\n - Module weights must have the right sizes wrt the block size\n - Only Linear, Embedding and Conv2d modules are supported for the moment\n - For more detail on how to quantize by blocks with convolutional weights,\n see \"And the Bit Goes Down: Revisiting the Quantization of Neural Networks\"\n - We implement the simplest form of noise here as stated in the paper\n which consists in randomly dropping blocks","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.quant_noise","uri":"program://CREMA/function/lavis.models.beats.modules.quant_noise#L121-L218","kind":"function","name":"quant_noise","path":"lavis/models/beats/modules.py","language":"python","start_line":121,"end_line":218,"context_start_line":101,"context_end_line":219,"code":" return F.relu\n elif activation == \"gelu\":\n return gelu\n elif activation == \"gelu_fast\":\n warnings.warn(\n \"--activation-fn=gelu_fast has been renamed to gelu_accurate\"\n )\n return gelu_accurate\n elif activation == \"gelu_accurate\":\n return gelu_accurate\n elif activation == \"tanh\":\n return torch.tanh\n elif activation == \"linear\":\n return lambda x: x\n elif activation == \"glu\":\n return lambda x: x\n else:\n raise RuntimeError(\"--activation-fn {} not supported\".format(activation))\n\n\ndef quant_noise(module, p, block_size):\n \"\"\"\n Wraps modules and applies quantization noise to the weights for\n subsequent quantization with Iterative Product Quantization as\n described in \"Training with Quantization Noise for Extreme Model Compression\"\n\n Args:\n - module: nn.Module\n - p: amount of Quantization Noise\n - block_size: size of the blocks for subsequent quantization with iPQ\n\n Remarks:\n - Module weights must have the right sizes wrt the block size\n - Only Linear, Embedding and Conv2d modules are supported for the moment\n - For more detail on how to quantize by blocks with convolutional weights,\n see \"And the Bit Goes Down: Revisiting the Quantization of Neural Networks\"\n - We implement the simplest form of noise here as stated in the paper\n which consists in randomly dropping blocks\n \"\"\"\n\n # if no quantization noise, don't register hook\n if p <= 0:\n return module\n\n # supported modules\n assert isinstance(module, (nn.Linear, nn.Embedding, nn.Conv2d))\n\n # test whether module.weight has the right sizes wrt block_size\n is_conv = module.weight.ndim == 4\n\n # 2D matrix\n if not is_conv:\n assert (\n module.weight.size(1) % block_size == 0\n ), \"Input features must be a multiple of block sizes\"\n\n # 4D matrix\n else:\n # 1x1 convolutions\n if module.kernel_size == (1, 1):\n assert (\n module.in_channels % block_size == 0\n ), \"Input channels must be a multiple of block sizes\"\n # regular convolutions\n else:\n k = module.kernel_size[0] * module.kernel_size[1]\n assert k % block_size == 0, \"Kernel size must be a multiple of block size\"\n\n def _forward_pre_hook(mod, input):\n # no noise for evaluation\n if mod.training:\n if not is_conv:\n # gather weight and sizes\n weight = mod.weight\n in_features = weight.size(1)\n out_features = weight.size(0)\n\n # split weight matrix into blocks and randomly drop selected blocks\n mask = torch.zeros(\n in_features // block_size * out_features, device=weight.device\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)\n\n else:\n # gather weight and sizes\n weight = mod.weight\n in_channels = mod.in_channels\n out_channels = mod.out_channels\n\n # split weight matrix into blocks and randomly drop selected blocks\n if mod.kernel_size == (1, 1):\n mask = torch.zeros(\n int(in_channels // block_size * out_channels),\n device=weight.device,\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)\n else:\n mask = torch.zeros(\n weight.size(0), weight.size(1), device=weight.device\n )\n mask.bernoulli_(p)\n mask = (\n mask.unsqueeze(2)\n .unsqueeze(3)\n .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])\n )\n\n # scale weights and apply mask\n mask = mask.to(\n torch.bool\n ) # x.bool() is not currently supported in TorchScript\n s = 1 / (1 - p)\n mod.weight.data = s * weight.masked_fill(mask, 0)\n\n module.register_forward_pre_hook(_forward_pre_hook)\n return module\n","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.forward","uri":"program://CREMA/function/lavis.models.beats.modules.forward#L73-L82","kind":"function","name":"forward","path":"lavis/models/beats/modules.py","language":"python","start_line":73,"end_line":82,"context_start_line":53,"context_end_line":102,"code":" def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()\n elif glu_type == \"swish\":\n self.glu_act = Swish()\n elif glu_type == \"relu\":\n self.glu_act = torch.nn.ReLU()\n elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)\n else:\n self.linear = nn.Linear(input_dim, output_dim * 2, False)\n\n def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n\n\ndef gelu(x: torch.Tensor) -> torch.Tensor:\n return torch.nn.functional.gelu(x.float()).type_as(x)\n\n\ndef get_activation_fn(activation: str):\n \"\"\"Returns the activation function corresponding to `activation`\"\"\"\n\n if activation == \"relu\":\n return F.relu\n elif activation == \"gelu\":","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.backward","uri":"program://CREMA/function/lavis.models.beats.modules.backward#L25-L26","kind":"function","name":"backward","path":"lavis/models/beats/modules.py","language":"python","start_line":25,"end_line":26,"context_start_line":5,"context_end_line":46,"code":"# Licensed under The MIT License [see LICENSE for details]\n# Based on fairseq code bases\n# https://github.com/pytorch/fairseq\n# --------------------------------------------------------\n\nimport math\nimport warnings\nimport torch\nfrom torch import Tensor, nn\nimport torch.nn.functional as F\n\n\nclass GradMultiply(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x, scale):\n ctx.scale = scale\n res = x.new(x)\n return res\n\n @staticmethod\n def backward(ctx, grad):\n return grad * ctx.scale, None\n\n\nclass SamePad(nn.Module):\n def __init__(self, kernel_size, causal=False):\n super().__init__()\n if causal:\n self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules.__init__","uri":"program://CREMA/function/lavis.models.beats.modules.__init__#L53-L71","kind":"function","name":"__init__","path":"lavis/models/beats/modules.py","language":"python","start_line":53,"end_line":71,"context_start_line":33,"context_end_line":91,"code":" self.remove = kernel_size - 1\n else:\n self.remove = 1 if kernel_size % 2 == 0 else 0\n\n def forward(self, x):\n if self.remove > 0:\n x = x[:, :, : -self.remove]\n return x\n\n\nclass Swish(nn.Module):\n def __init__(self):\n super(Swish, self).__init__()\n self.act = torch.nn.Sigmoid()\n\n def forward(self, x):\n return x * self.act(x)\n\n\nclass GLU_Linear(nn.Module):\n def __init__(self, input_dim, output_dim, glu_type=\"sigmoid\", bias_in_glu=True):\n super(GLU_Linear, self).__init__()\n\n self.glu_type = glu_type\n self.output_dim = output_dim\n\n if glu_type == \"sigmoid\":\n self.glu_act = torch.nn.Sigmoid()\n elif glu_type == \"swish\":\n self.glu_act = Swish()\n elif glu_type == \"relu\":\n self.glu_act = torch.nn.ReLU()\n elif glu_type == \"gelu\":\n self.glu_act = torch.nn.GELU()\n\n if bias_in_glu:\n self.linear = nn.Linear(input_dim, output_dim * 2, True)\n else:\n self.linear = nn.Linear(input_dim, output_dim * 2, False)\n\n def forward(self, x):\n # to be consistent with GLU_Linear, we assume the input always has the #channel (#dim) in the last dimension of the tensor, so need to switch the dimension first for 1D-Conv case\n x = self.linear(x)\n\n if self.glu_type == \"bilinear\":\n x = (x[:, :, 0:self.output_dim] * x[:, :, self.output_dim:self.output_dim * 2])\n else:\n x = (x[:, :, 0:self.output_dim] * self.glu_act(x[:, :, self.output_dim:self.output_dim * 2]))\n\n return x\n\n\ndef gelu_accurate(x):\n if not hasattr(gelu_accurate, \"_a\"):\n gelu_accurate._a = math.sqrt(2 / math.pi)\n return (\n 0.5 * x * (1 + torch.tanh(gelu_accurate._a * (x + 0.044715 * torch.pow(x, 3))))\n )\n","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.beats.modules._forward_pre_hook","uri":"program://CREMA/function/lavis.models.beats.modules._forward_pre_hook#L169-L215","kind":"function","name":"_forward_pre_hook","path":"lavis/models/beats/modules.py","language":"python","start_line":169,"end_line":215,"context_start_line":149,"context_end_line":219,"code":" is_conv = module.weight.ndim == 4\n\n # 2D matrix\n if not is_conv:\n assert (\n module.weight.size(1) % block_size == 0\n ), \"Input features must be a multiple of block sizes\"\n\n # 4D matrix\n else:\n # 1x1 convolutions\n if module.kernel_size == (1, 1):\n assert (\n module.in_channels % block_size == 0\n ), \"Input channels must be a multiple of block sizes\"\n # regular convolutions\n else:\n k = module.kernel_size[0] * module.kernel_size[1]\n assert k % block_size == 0, \"Kernel size must be a multiple of block size\"\n\n def _forward_pre_hook(mod, input):\n # no noise for evaluation\n if mod.training:\n if not is_conv:\n # gather weight and sizes\n weight = mod.weight\n in_features = weight.size(1)\n out_features = weight.size(0)\n\n # split weight matrix into blocks and randomly drop selected blocks\n mask = torch.zeros(\n in_features // block_size * out_features, device=weight.device\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_features)\n\n else:\n # gather weight and sizes\n weight = mod.weight\n in_channels = mod.in_channels\n out_channels = mod.out_channels\n\n # split weight matrix into blocks and randomly drop selected blocks\n if mod.kernel_size == (1, 1):\n mask = torch.zeros(\n int(in_channels // block_size * out_channels),\n device=weight.device,\n )\n mask.bernoulli_(p)\n mask = mask.repeat_interleave(block_size, -1).view(-1, in_channels)\n else:\n mask = torch.zeros(\n weight.size(0), weight.size(1), device=weight.device\n )\n mask.bernoulli_(p)\n mask = (\n mask.unsqueeze(2)\n .unsqueeze(3)\n .repeat(1, 1, mod.kernel_size[0], mod.kernel_size[1])\n )\n\n # scale weights and apply mask\n mask = mask.to(\n torch.bool\n ) # x.bool() is not currently supported in TorchScript\n s = 1 / (1 - p)\n mod.weight.data = s * weight.masked_fill(mask, 0)\n\n module.register_forward_pre_hook(_forward_pre_hook)\n return module\n","source_hash":"edeb6b6cd6a784da749f932c3e0783c0bce556fc768d0d23a4d53d4b819eb424","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5","uri":"program://CREMA/module/lavis.models.blip2_models.modeling_t5#L1-L2066","kind":"module","name":"lavis.models.blip2_models.modeling_t5","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1,"end_line":2066,"context_start_line":1,"context_end_line":2066,"code":"# coding=utf-8\n# Copyright 2018 Mesh TensorFlow authors, T5 Authors and HuggingFace Inc. team.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\"\"\" PyTorch T5 model.\"\"\"\n\n\nimport copy\nimport math\nimport os\nimport warnings\nfrom typing import Optional, Tuple, Union\n\nimport torch\nfrom torch import nn\nfrom torch.nn import CrossEntropyLoss\nfrom torch.utils.checkpoint import checkpoint\n\nfrom transformers.activations import ACT2FN\nfrom transformers.modeling_outputs import (\n BaseModelOutput,\n BaseModelOutputWithPastAndCrossAttentions,\n Seq2SeqLMOutput,\n Seq2SeqModelOutput,\n)\nfrom transformers.modeling_utils import PreTrainedModel\nfrom transformers.pytorch_utils import (\n ALL_LAYERNORM_LAYERS,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.utils import (\n DUMMY_INPUTS,\n DUMMY_MASK,\n add_start_docstrings,\n add_start_docstrings_to_model_forward,\n is_torch_fx_proxy,\n logging,\n replace_return_docstrings,\n)\nfrom transformers.utils.model_parallel_utils import assert_device_map, get_device_map\nfrom transformers.models.t5.configuration_t5 import T5Config\n\n\nlogger = logging.get_logger(__name__)\n\n_CONFIG_FOR_DOC = \"T5Config\"\n_TOKENIZER_FOR_DOC = \"T5Tokenizer\"\n_CHECKPOINT_FOR_DOC = \"t5-small\"\n\n####################################################\n# This dict contains ids and associated url\n# for the pretrained weights provided with the models\n####################################################\nT5_PRETRAINED_MODEL_ARCHIVE_LIST = [\n \"t5-small\",\n \"t5-base\",\n \"t5-large\",\n \"t5-3b\",\n \"t5-11b\",\n # See all T5 models at https://huggingface.co/models?filter=t5\n]\n\n\n####################################################\n# This is a conversion method from TF 1.0 to PyTorch\n# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28\n####################################################\ndef load_tf_weights_in_t5(model, config, tf_checkpoint_path):\n \"\"\"Load tf checkpoints in a pytorch model.\"\"\"\n try:\n import re\n\n import numpy as np\n import tensorflow as tf\n except ImportError:\n logger.error(\n \"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see \"\n \"https://www.tensorflow.org/install/ for installation instructions.\"\n )\n raise\n tf_path = os.path.abspath(tf_checkpoint_path)\n logger.info(f\"Converting TensorFlow checkpoint from {tf_path}\")\n # Load weights from TF model\n init_vars = tf.train.list_variables(tf_path)\n names = []\n tf_weights = {}\n for name, shape in init_vars:\n logger.info(f\"Loading TF weight {name} with shape {shape}\")\n array = tf.train.load_variable(tf_path, name)\n names.append(name)\n tf_weights[name] = array\n\n for txt_name in names:\n name = txt_name.split(\"/\")\n # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v\n # which are not required for using pretrained model\n if any(\n n\n in [\n \"adam_v\",\n \"adam_m\",\n \"AdamWeightDecayOptimizer\",\n \"AdamWeightDecayOptimizer_1\",\n \"global_step\",\n ]\n for n in name\n ):\n logger.info(f\"Skipping {'/'.join(name)}\")\n tf_weights.pop(txt_name, None)\n continue\n if \"_slot_\" in name[-1]:\n logger.info(f\"Skipping {'/'.join(name)}\")\n tf_weights.pop(txt_name, None)\n continue\n pointer = model\n array = tf_weights[txt_name]\n\n for m_name in name:\n if re.fullmatch(r\"[A-Za-z]+_\\d+\", m_name):\n scope_names = re.split(r\"_(\\d+)\", m_name)\n else:\n scope_names = [m_name]\n if scope_names[0] in [\"kernel\", \"scale\", \"embedding\"]:\n pointer = getattr(pointer, \"weight\")\n elif scope_names[0] == \"self_attention\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[0]\n elif scope_names[0] == \"enc_dec_attention\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[1]\n elif scope_names[0] == \"dense_relu_dense\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[2]\n elif scope_names[0] == \"rms_norm\":\n if hasattr(pointer, \"layer_norm\"):\n pointer = getattr(pointer, \"layer_norm\")\n elif hasattr(pointer, \"final_layer_norm\"):\n pointer = getattr(pointer, \"final_layer_norm\")\n elif scope_names[0] == \"scale\":\n pointer = getattr(pointer, \"weight\")\n elif scope_names[0] == \"output_bias\" or scope_names[0] == \"beta\":\n pointer = getattr(pointer, \"bias\")\n elif scope_names[0] == \"squad\":\n pointer = getattr(pointer, \"classifier\")\n elif scope_names[0] == \"decoder\" and name[1] == \"logits\":\n continue\n elif scope_names[0] == \"logits\":\n pointer = getattr(pointer, \"lm_head\")\n elif (\n scope_names[0] == \"wi\"\n and len(scope_names) > 1\n and scope_names[1].isdigit()\n ):\n pointer = getattr(pointer, f\"wi_{scope_names[1]}\")\n continue\n else:\n try:\n pointer = getattr(pointer, scope_names[0])\n except AttributeError:\n logger.info(f\"Skipping {'/'.join(name)}\")\n continue\n if len(scope_names) >= 2:\n num = int(scope_names[1])\n pointer = pointer[num]\n if scope_names[0] not in [\"kernel\", \"scale\", \"embedding\"]:\n pointer = getattr(pointer, \"weight\")\n if scope_names[0] != \"embedding\":\n logger.info(f\"Transposing numpy weight of shape {array.shape} for {name}\")\n array = np.transpose(array)\n try:\n assert (\n pointer.shape == array.shape\n ), f\"Pointer shape {pointer.shape} and array shape {array.shape} mismatched\"\n except AssertionError as e:\n e.args += (pointer.shape, array.shape)\n raise\n logger.info(f\"Initialize PyTorch weight {name}\")\n pointer.data = torch.from_numpy(array.astype(np.float32))\n tf_weights.pop(txt_name, None)\n\n logger.info(f\"Weights not copied to PyTorch model: {', '.join(tf_weights.keys())}.\")\n return model\n\n\n####################################################\n# PyTorch Models are constructed by sub-classing\n# - torch.nn.Module for the layers and\n# - PreTrainedModel for the models (it-self a sub-class of nn.Module)\n####################################################\nPARALLELIZE_DOCSTRING = r\"\"\"\n This is an experimental feature and is a subject to change at a moment's notice.\n\n Uses a device map to distribute attention modules of the model across several devices. If no device map is given,\n it will evenly distribute blocks across all devices.\n\n Args:\n device_map (`Dict[int, list]`, optional, defaults to None):\n A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always\n automatically mapped to the first device (for esoteric reasons). That means that the first device should\n have fewer attention modules mapped to it than other devices. For reference, the t5 models have the\n following number of attention modules:\n\n - t5-small: 6\n - t5-base: 12\n - t5-large: 24\n - t5-3b: 24\n - t5-11b: 24\n\n Example:\n\n ```python\n # Here is an example of a device map on a machine with 4 GPUs using t5-3b, which has a total of 24 attention modules:\n model = T5ForConditionalGeneration.from_pretrained(\"t5-3b\")\n device_map = {\n 0: [0, 1, 2],\n 1: [3, 4, 5, 6, 7, 8, 9],\n 2: [10, 11, 12, 13, 14, 15, 16],\n 3: [17, 18, 19, 20, 21, 22, 23],\n }\n model.parallelize(device_map)\n ```\n\"\"\"\nDEPARALLELIZE_DOCSTRING = r\"\"\"\n Moves the model to cpu from a model parallel state.\n\n Example:\n\n ```python\n # On a 4 GPU machine with t5-3b:\n model = T5ForConditionalGeneration.from_pretrained(\"t5-3b\")\n device_map = {\n 0: [0, 1, 2],\n 1: [3, 4, 5, 6, 7, 8, 9],\n 2: [10, 11, 12, 13, 14, 15, 16],\n 3: [17, 18, 19, 20, 21, 22, 23],\n }\n model.parallelize(device_map) # Splits the model across several devices\n model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()\n ```\n\"\"\"\n\n\nclass T5LayerNorm(nn.Module):\n def __init__(self, hidden_size, eps=1e-6):\n \"\"\"\n Construct a layernorm module in the T5 style. No bias and no subtraction of mean.\n \"\"\"\n super().__init__()\n self.weight = nn.Parameter(torch.ones(hidden_size))\n self.variance_epsilon = eps\n\n def forward(self, hidden_states):\n\n # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean\n # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated\n # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for\n # half-precision inputs is done in fp32\n\n variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)\n hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)\n\n # convert into half-precision if necessary\n if self.weight.dtype in [torch.float16, torch.bfloat16]:\n hidden_states = hidden_states.to(self.weight.dtype)\n\n return self.weight * hidden_states\n\n\ntry:\n from apex.normalization import FusedRMSNorm\n\n T5LayerNorm = FusedRMSNorm # noqa\n\n logger.info(\n \"Discovered apex.normalization.FusedRMSNorm - will use it instead of T5LayerNorm\"\n )\nexcept ImportError:\n # using the normal T5LayerNorm\n pass\nexcept Exception:\n logger.warning(\"discovered apex but it failed to load, falling back to T5LayerNorm\")\n pass\n\nALL_LAYERNORM_LAYERS.append(T5LayerNorm)\n\n\nclass T5DenseActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_states = self.wi(hidden_states)\n hidden_states = self.act(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5DenseGatedActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_gelu = self.act(self.wi_0(hidden_states))\n hidden_linear = self.wi_1(hidden_states)\n hidden_states = hidden_gelu * hidden_linear\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5LayerFF(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n if config.is_gated_act:\n self.DenseReluDense = T5DenseGatedActDense(config)\n else:\n self.DenseReluDense = T5DenseActDense(config)\n\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(self, hidden_states):\n forwarded_states = self.layer_norm(hidden_states)\n forwarded_states = self.DenseReluDense(forwarded_states)\n hidden_states = hidden_states + self.dropout(forwarded_states)\n return hidden_states\n\n\nclass T5Attention(nn.Module):\n def __init__(self, config: T5Config, has_relative_attention_bias=False):\n super().__init__()\n self.is_decoder = config.is_decoder\n self.has_relative_attention_bias = has_relative_attention_bias\n self.relative_attention_num_buckets = config.relative_attention_num_buckets\n self.relative_attention_max_distance = config.relative_attention_max_distance\n self.d_model = config.d_model\n self.key_value_proj_dim = config.d_kv\n self.n_heads = config.num_heads\n self.dropout = config.dropout_rate\n self.inner_dim = self.n_heads * self.key_value_proj_dim\n\n # Mesh TensorFlow initialization to avoid scaling before softmax\n self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)\n\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(\n self.relative_attention_num_buckets, self.n_heads\n )\n self.pruned_heads = set()\n self.gradient_checkpointing = False\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads\n )\n # Prune linear layers\n self.q = prune_linear_layer(self.q, index)\n self.k = prune_linear_layer(self.k, index)\n self.v = prune_linear_layer(self.v, index)\n self.o = prune_linear_layer(self.o, index, dim=1)\n # Update hyper params\n self.n_heads = self.n_heads - len(heads)\n self.inner_dim = self.key_value_proj_dim * self.n_heads\n self.pruned_heads = self.pruned_heads.union(heads)\n\n @staticmethod\n def _relative_position_bucket(\n relative_position, bidirectional=True, num_buckets=32, max_distance=128\n ):\n \"\"\"\n Adapted from Mesh Tensorflow:\n https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593\n\n Translate relative position to a bucket number for relative attention. The relative position is defined as\n memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to\n position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for\n small absolute relative_position and larger buckets for larger absolute relative_positions. All relative\n positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.\n This should allow for more graceful generalization to longer sequences than the model has been trained on\n\n Args:\n relative_position: an int32 Tensor\n bidirectional: a boolean - whether the attention is bidirectional\n num_buckets: an integer\n max_distance: an integer\n\n Returns:\n a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)\n \"\"\"\n relative_buckets = 0\n if bidirectional:\n num_buckets //= 2\n relative_buckets += (relative_position > 0).to(torch.long) * num_buckets\n relative_position = torch.abs(relative_position)\n else:\n relative_position = -torch.min(\n relative_position, torch.zeros_like(relative_position)\n )\n # now relative_position is in the range [0, inf)\n\n # half of the buckets are for exact increments in positions\n max_exact = num_buckets // 2\n is_small = relative_position < max_exact\n\n # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance\n relative_position_if_large = max_exact + (\n torch.log(relative_position.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_position_if_large = torch.min(\n relative_position_if_large,\n torch.full_like(relative_position_if_large, num_buckets - 1),\n )\n\n relative_buckets += torch.where(\n is_small, relative_position, relative_position_if_large\n )\n return relative_buckets\n\n def compute_bias(self, query_length, key_length, device=None):\n \"\"\"Compute binned relative position bias\"\"\"\n if device is None:\n device = self.relative_attention_bias.weight.device\n context_position = torch.arange(query_length, dtype=torch.long, device=device)[\n :, None\n ]\n memory_position = torch.arange(key_length, dtype=torch.long, device=device)[\n None, :\n ]\n relative_position = (\n memory_position - context_position\n ) # shape (query_length, key_length)\n relative_position_bucket = self._relative_position_bucket(\n relative_position, # shape (query_length, key_length)\n bidirectional=(not self.is_decoder),\n num_buckets=self.relative_attention_num_buckets,\n max_distance=self.relative_attention_max_distance,\n )\n values = self.relative_attention_bias(\n relative_position_bucket\n ) # shape (query_length, key_length, num_heads)\n values = values.permute([2, 0, 1]).unsqueeze(\n 0\n ) # shape (1, num_heads, query_length, key_length)\n return values\n\n def forward(\n self,\n hidden_states,\n mask=None,\n key_value_states=None,\n position_bias=None,\n past_key_value=None,\n layer_head_mask=None,\n query_length=None,\n use_cache=False,\n output_attentions=False,\n ):\n \"\"\"\n Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).\n \"\"\"\n # Input is (batch_size, seq_length, dim)\n # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)\n # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)\n batch_size, seq_length = hidden_states.shape[:2]\n\n real_seq_length = seq_length\n\n if past_key_value is not None:\n assert (\n len(past_key_value) == 2\n ), f\"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states\"\n real_seq_length += (\n past_key_value[0].shape[2] if query_length is None else query_length\n )\n\n key_length = (\n real_seq_length if key_value_states is None else key_value_states.shape[1]\n )\n\n def shape(states):\n \"\"\"projection\"\"\"\n return states.view(\n batch_size, -1, self.n_heads, self.key_value_proj_dim\n ).transpose(1, 2)\n\n def uns\n# ... truncated ...","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.load_tf_weights_in_t5","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.load_tf_weights_in_t5#L79-L193","kind":"function","name":"load_tf_weights_in_t5","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":79,"end_line":193,"context_start_line":59,"context_end_line":213,"code":"_CHECKPOINT_FOR_DOC = \"t5-small\"\n\n####################################################\n# This dict contains ids and associated url\n# for the pretrained weights provided with the models\n####################################################\nT5_PRETRAINED_MODEL_ARCHIVE_LIST = [\n \"t5-small\",\n \"t5-base\",\n \"t5-large\",\n \"t5-3b\",\n \"t5-11b\",\n # See all T5 models at https://huggingface.co/models?filter=t5\n]\n\n\n####################################################\n# This is a conversion method from TF 1.0 to PyTorch\n# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28\n####################################################\ndef load_tf_weights_in_t5(model, config, tf_checkpoint_path):\n \"\"\"Load tf checkpoints in a pytorch model.\"\"\"\n try:\n import re\n\n import numpy as np\n import tensorflow as tf\n except ImportError:\n logger.error(\n \"Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see \"\n \"https://www.tensorflow.org/install/ for installation instructions.\"\n )\n raise\n tf_path = os.path.abspath(tf_checkpoint_path)\n logger.info(f\"Converting TensorFlow checkpoint from {tf_path}\")\n # Load weights from TF model\n init_vars = tf.train.list_variables(tf_path)\n names = []\n tf_weights = {}\n for name, shape in init_vars:\n logger.info(f\"Loading TF weight {name} with shape {shape}\")\n array = tf.train.load_variable(tf_path, name)\n names.append(name)\n tf_weights[name] = array\n\n for txt_name in names:\n name = txt_name.split(\"/\")\n # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v\n # which are not required for using pretrained model\n if any(\n n\n in [\n \"adam_v\",\n \"adam_m\",\n \"AdamWeightDecayOptimizer\",\n \"AdamWeightDecayOptimizer_1\",\n \"global_step\",\n ]\n for n in name\n ):\n logger.info(f\"Skipping {'/'.join(name)}\")\n tf_weights.pop(txt_name, None)\n continue\n if \"_slot_\" in name[-1]:\n logger.info(f\"Skipping {'/'.join(name)}\")\n tf_weights.pop(txt_name, None)\n continue\n pointer = model\n array = tf_weights[txt_name]\n\n for m_name in name:\n if re.fullmatch(r\"[A-Za-z]+_\\d+\", m_name):\n scope_names = re.split(r\"_(\\d+)\", m_name)\n else:\n scope_names = [m_name]\n if scope_names[0] in [\"kernel\", \"scale\", \"embedding\"]:\n pointer = getattr(pointer, \"weight\")\n elif scope_names[0] == \"self_attention\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[0]\n elif scope_names[0] == \"enc_dec_attention\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[1]\n elif scope_names[0] == \"dense_relu_dense\":\n pointer = getattr(pointer, \"layer\")\n pointer = pointer[2]\n elif scope_names[0] == \"rms_norm\":\n if hasattr(pointer, \"layer_norm\"):\n pointer = getattr(pointer, \"layer_norm\")\n elif hasattr(pointer, \"final_layer_norm\"):\n pointer = getattr(pointer, \"final_layer_norm\")\n elif scope_names[0] == \"scale\":\n pointer = getattr(pointer, \"weight\")\n elif scope_names[0] == \"output_bias\" or scope_names[0] == \"beta\":\n pointer = getattr(pointer, \"bias\")\n elif scope_names[0] == \"squad\":\n pointer = getattr(pointer, \"classifier\")\n elif scope_names[0] == \"decoder\" and name[1] == \"logits\":\n continue\n elif scope_names[0] == \"logits\":\n pointer = getattr(pointer, \"lm_head\")\n elif (\n scope_names[0] == \"wi\"\n and len(scope_names) > 1\n and scope_names[1].isdigit()\n ):\n pointer = getattr(pointer, f\"wi_{scope_names[1]}\")\n continue\n else:\n try:\n pointer = getattr(pointer, scope_names[0])\n except AttributeError:\n logger.info(f\"Skipping {'/'.join(name)}\")\n continue\n if len(scope_names) >= 2:\n num = int(scope_names[1])\n pointer = pointer[num]\n if scope_names[0] not in [\"kernel\", \"scale\", \"embedding\"]:\n pointer = getattr(pointer, \"weight\")\n if scope_names[0] != \"embedding\":\n logger.info(f\"Transposing numpy weight of shape {array.shape} for {name}\")\n array = np.transpose(array)\n try:\n assert (\n pointer.shape == array.shape\n ), f\"Pointer shape {pointer.shape} and array shape {array.shape} mismatched\"\n except AssertionError as e:\n e.args += (pointer.shape, array.shape)\n raise\n logger.info(f\"Initialize PyTorch weight {name}\")\n pointer.data = torch.from_numpy(array.astype(np.float32))\n tf_weights.pop(txt_name, None)\n\n logger.info(f\"Weights not copied to PyTorch model: {', '.join(tf_weights.keys())}.\")\n return model\n\n\n####################################################\n# PyTorch Models are constructed by sub-classing\n# - torch.nn.Module for the layers and\n# - PreTrainedModel for the models (it-self a sub-class of nn.Module)\n####################################################\nPARALLELIZE_DOCSTRING = r\"\"\"\n This is an experimental feature and is a subject to change at a moment's notice.\n\n Uses a device map to distribute attention modules of the model across several devices. If no device map is given,\n it will evenly distribute blocks across all devices.\n\n Args:\n device_map (`Dict[int, list]`, optional, defaults to None):\n A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always\n automatically mapped to the first device (for esoteric reasons). That means that the first device should\n have fewer attention modules mapped to it than other devices. For reference, the t5 models have the\n following number of attention modules:\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5LayerNorm","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5LayerNorm#L254-L277","kind":"class","name":"T5LayerNorm","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":254,"end_line":277,"context_start_line":234,"context_end_line":297,"code":"DEPARALLELIZE_DOCSTRING = r\"\"\"\n Moves the model to cpu from a model parallel state.\n\n Example:\n\n ```python\n # On a 4 GPU machine with t5-3b:\n model = T5ForConditionalGeneration.from_pretrained(\"t5-3b\")\n device_map = {\n 0: [0, 1, 2],\n 1: [3, 4, 5, 6, 7, 8, 9],\n 2: [10, 11, 12, 13, 14, 15, 16],\n 3: [17, 18, 19, 20, 21, 22, 23],\n }\n model.parallelize(device_map) # Splits the model across several devices\n model.deparallelize() # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()\n ```\n\"\"\"\n\n\nclass T5LayerNorm(nn.Module):\n def __init__(self, hidden_size, eps=1e-6):\n \"\"\"\n Construct a layernorm module in the T5 style. No bias and no subtraction of mean.\n \"\"\"\n super().__init__()\n self.weight = nn.Parameter(torch.ones(hidden_size))\n self.variance_epsilon = eps\n\n def forward(self, hidden_states):\n\n # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean\n # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated\n # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for\n # half-precision inputs is done in fp32\n\n variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)\n hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)\n\n # convert into half-precision if necessary\n if self.weight.dtype in [torch.float16, torch.bfloat16]:\n hidden_states = hidden_states.to(self.weight.dtype)\n\n return self.weight * hidden_states\n\n\ntry:\n from apex.normalization import FusedRMSNorm\n\n T5LayerNorm = FusedRMSNorm # noqa\n\n logger.info(\n \"Discovered apex.normalization.FusedRMSNorm - will use it instead of T5LayerNorm\"\n )\nexcept ImportError:\n # using the normal T5LayerNorm\n pass\nexcept Exception:\n logger.warning(\"discovered apex but it failed to load, falling back to T5LayerNorm\")\n pass\n\nALL_LAYERNORM_LAYERS.append(T5LayerNorm)\n\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5DenseActDense","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5DenseActDense#L298-L311","kind":"class","name":"T5DenseActDense","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":298,"end_line":311,"context_start_line":278,"context_end_line":331,"code":"\n\ntry:\n from apex.normalization import FusedRMSNorm\n\n T5LayerNorm = FusedRMSNorm # noqa\n\n logger.info(\n \"Discovered apex.normalization.FusedRMSNorm - will use it instead of T5LayerNorm\"\n )\nexcept ImportError:\n # using the normal T5LayerNorm\n pass\nexcept Exception:\n logger.warning(\"discovered apex but it failed to load, falling back to T5LayerNorm\")\n pass\n\nALL_LAYERNORM_LAYERS.append(T5LayerNorm)\n\n\nclass T5DenseActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_states = self.wi(hidden_states)\n hidden_states = self.act(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5DenseGatedActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_gelu = self.act(self.wi_0(hidden_states))\n hidden_linear = self.wi_1(hidden_states)\n hidden_states = hidden_gelu * hidden_linear\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5DenseGatedActDense","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5DenseGatedActDense#L314-L329","kind":"class","name":"T5DenseGatedActDense","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":314,"end_line":329,"context_start_line":294,"context_end_line":349,"code":"\nALL_LAYERNORM_LAYERS.append(T5LayerNorm)\n\n\nclass T5DenseActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_states = self.wi(hidden_states)\n hidden_states = self.act(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5DenseGatedActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_gelu = self.act(self.wi_0(hidden_states))\n hidden_linear = self.wi_1(hidden_states)\n hidden_states = hidden_gelu * hidden_linear\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5LayerFF(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n if config.is_gated_act:\n self.DenseReluDense = T5DenseGatedActDense(config)\n else:\n self.DenseReluDense = T5DenseActDense(config)\n\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(self, hidden_states):\n forwarded_states = self.layer_norm(hidden_states)\n forwarded_states = self.DenseReluDense(forwarded_states)\n hidden_states = hidden_states + self.dropout(forwarded_states)\n return hidden_states\n\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5LayerFF","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5LayerFF#L332-L347","kind":"class","name":"T5LayerFF","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":332,"end_line":347,"context_start_line":312,"context_end_line":367,"code":"\n\nclass T5DenseGatedActDense(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)\n self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)\n self.dropout = nn.Dropout(config.dropout_rate)\n self.act = ACT2FN[config.dense_act_fn]\n\n def forward(self, hidden_states):\n hidden_gelu = self.act(self.wi_0(hidden_states))\n hidden_linear = self.wi_1(hidden_states)\n hidden_states = hidden_gelu * hidden_linear\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.wo(hidden_states)\n return hidden_states\n\n\nclass T5LayerFF(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n if config.is_gated_act:\n self.DenseReluDense = T5DenseGatedActDense(config)\n else:\n self.DenseReluDense = T5DenseActDense(config)\n\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(self, hidden_states):\n forwarded_states = self.layer_norm(hidden_states)\n forwarded_states = self.DenseReluDense(forwarded_states)\n hidden_states = hidden_states + self.dropout(forwarded_states)\n return hidden_states\n\n\nclass T5Attention(nn.Module):\n def __init__(self, config: T5Config, has_relative_attention_bias=False):\n super().__init__()\n self.is_decoder = config.is_decoder\n self.has_relative_attention_bias = has_relative_attention_bias\n self.relative_attention_num_buckets = config.relative_attention_num_buckets\n self.relative_attention_max_distance = config.relative_attention_max_distance\n self.d_model = config.d_model\n self.key_value_proj_dim = config.d_kv\n self.n_heads = config.num_heads\n self.dropout = config.dropout_rate\n self.inner_dim = self.n_heads * self.key_value_proj_dim\n\n # Mesh TensorFlow initialization to avoid scaling before softmax\n self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5Attention","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5Attention#L350-L620","kind":"class","name":"T5Attention","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":350,"end_line":620,"context_start_line":330,"context_end_line":640,"code":"\n\nclass T5LayerFF(nn.Module):\n def __init__(self, config: T5Config):\n super().__init__()\n if config.is_gated_act:\n self.DenseReluDense = T5DenseGatedActDense(config)\n else:\n self.DenseReluDense = T5DenseActDense(config)\n\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(self, hidden_states):\n forwarded_states = self.layer_norm(hidden_states)\n forwarded_states = self.DenseReluDense(forwarded_states)\n hidden_states = hidden_states + self.dropout(forwarded_states)\n return hidden_states\n\n\nclass T5Attention(nn.Module):\n def __init__(self, config: T5Config, has_relative_attention_bias=False):\n super().__init__()\n self.is_decoder = config.is_decoder\n self.has_relative_attention_bias = has_relative_attention_bias\n self.relative_attention_num_buckets = config.relative_attention_num_buckets\n self.relative_attention_max_distance = config.relative_attention_max_distance\n self.d_model = config.d_model\n self.key_value_proj_dim = config.d_kv\n self.n_heads = config.num_heads\n self.dropout = config.dropout_rate\n self.inner_dim = self.n_heads * self.key_value_proj_dim\n\n # Mesh TensorFlow initialization to avoid scaling before softmax\n self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)\n\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(\n self.relative_attention_num_buckets, self.n_heads\n )\n self.pruned_heads = set()\n self.gradient_checkpointing = False\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads\n )\n # Prune linear layers\n self.q = prune_linear_layer(self.q, index)\n self.k = prune_linear_layer(self.k, index)\n self.v = prune_linear_layer(self.v, index)\n self.o = prune_linear_layer(self.o, index, dim=1)\n # Update hyper params\n self.n_heads = self.n_heads - len(heads)\n self.inner_dim = self.key_value_proj_dim * self.n_heads\n self.pruned_heads = self.pruned_heads.union(heads)\n\n @staticmethod\n def _relative_position_bucket(\n relative_position, bidirectional=True, num_buckets=32, max_distance=128\n ):\n \"\"\"\n Adapted from Mesh Tensorflow:\n https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593\n\n Translate relative position to a bucket number for relative attention. The relative position is defined as\n memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to\n position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for\n small absolute relative_position and larger buckets for larger absolute relative_positions. All relative\n positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.\n This should allow for more graceful generalization to longer sequences than the model has been trained on\n\n Args:\n relative_position: an int32 Tensor\n bidirectional: a boolean - whether the attention is bidirectional\n num_buckets: an integer\n max_distance: an integer\n\n Returns:\n a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)\n \"\"\"\n relative_buckets = 0\n if bidirectional:\n num_buckets //= 2\n relative_buckets += (relative_position > 0).to(torch.long) * num_buckets\n relative_position = torch.abs(relative_position)\n else:\n relative_position = -torch.min(\n relative_position, torch.zeros_like(relative_position)\n )\n # now relative_position is in the range [0, inf)\n\n # half of the buckets are for exact increments in positions\n max_exact = num_buckets // 2\n is_small = relative_position < max_exact\n\n # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance\n relative_position_if_large = max_exact + (\n torch.log(relative_position.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_position_if_large = torch.min(\n relative_position_if_large,\n torch.full_like(relative_position_if_large, num_buckets - 1),\n )\n\n relative_buckets += torch.where(\n is_small, relative_position, relative_position_if_large\n )\n return relative_buckets\n\n def compute_bias(self, query_length, key_length, device=None):\n \"\"\"Compute binned relative position bias\"\"\"\n if device is None:\n device = self.relative_attention_bias.weight.device\n context_position = torch.arange(query_length, dtype=torch.long, device=device)[\n :, None\n ]\n memory_position = torch.arange(key_length, dtype=torch.long, device=device)[\n None, :\n ]\n relative_position = (\n memory_position - context_position\n ) # shape (query_length, key_length)\n relative_position_bucket = self._relative_position_bucket(\n relative_position, # shape (query_length, key_length)\n bidirectional=(not self.is_decoder),\n num_buckets=self.relative_attention_num_buckets,\n max_distance=self.relative_attention_max_distance,\n )\n values = self.relative_attention_bias(\n relative_position_bucket\n ) # shape (query_length, key_length, num_heads)\n values = values.permute([2, 0, 1]).unsqueeze(\n 0\n ) # shape (1, num_heads, query_length, key_length)\n return values\n\n def forward(\n self,\n hidden_states,\n mask=None,\n key_value_states=None,\n position_bias=None,\n past_key_value=None,\n layer_head_mask=None,\n query_length=None,\n use_cache=False,\n output_attentions=False,\n ):\n \"\"\"\n Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).\n \"\"\"\n # Input is (batch_size, seq_length, dim)\n # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)\n # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)\n batch_size, seq_length = hidden_states.shape[:2]\n\n real_seq_length = seq_length\n\n if past_key_value is not None:\n assert (\n len(past_key_value) == 2\n ), f\"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states\"\n real_seq_length += (\n past_key_value[0].shape[2] if query_length is None else query_length\n )\n\n key_length = (\n real_seq_length if key_value_states is None else key_value_states.shape[1]\n )\n\n def shape(states):\n \"\"\"projection\"\"\"\n return states.view(\n batch_size, -1, self.n_heads, self.key_value_proj_dim\n ).transpose(1, 2)\n\n def unshape(states):\n \"\"\"reshape\"\"\"\n return (\n states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)\n )\n\n def project(hidden_states, proj_layer, key_value_states, past_key_value):\n \"\"\"projects hidden states correctly to key/query states\"\"\"\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(hidden_states))\n elif past_key_value is None:\n # cross-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(key_value_states))\n\n if past_key_value is not None:\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, key_length, dim_per_head)\n hidden_states = torch.cat([past_key_value, hidden_states], dim=2)\n else:\n # cross-attn\n hidden_states = past_key_value\n return hidden_states\n\n # get query states\n query_states = shape(\n self.q(hidden_states)\n ) # (batch_size, n_heads, seq_length, dim_per_head)\n\n # get key/value states\n key_states = project(\n hidden_states,\n self.k,\n key_value_states,\n past_key_value[0] if past_key_value is not None else None,\n )\n value_states = project(\n hidden_states,\n self.v,\n key_value_states,\n past_key_value[1] if past_key_value is not None else None,\n )\n\n # compute scores\n scores = torch.matmul(\n query_states, key_states.transpose(3, 2)\n ) # equivalent of torch.einsum(\"bnqd,bnkd->bnqk\", query_states, key_states), compatible with onnx op>9\n\n if position_bias is None:\n if not self.has_relative_attention_bias:\n position_bias = torch.zeros(\n (1, self.n_heads, real_seq_length, key_length),\n device=scores.device,\n dtype=scores.dtype,\n )\n if self.gradient_checkpointing and self.training:\n position_bias.requires_grad = True\n else:\n position_bias = self.compute_bias(\n real_seq_length, key_length, device=scores.device\n )\n\n # if key and values are already calculated\n # we want only the last query position bias\n if past_key_value is not None:\n position_bias = position_bias[:, :, -hidden_states.size(1) :, :]\n\n if mask is not None:\n position_bias = (\n position_bias + mask\n ) # (batch_size, n_heads, seq_length, key_length)\n\n if self.pruned_heads:\n mask = torch.ones(position_bias.shape[1])\n mask[list(self.pruned_heads)] = 0\n position_bias_masked = position_bias[:, mask.bool()]\n else:\n position_bias_masked = position_bias\n\n scores += position_bias_masked\n attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(\n scores\n ) # (batch_size, n_heads, seq_length, key_length)\n attn_weights = nn.functional.dropout(\n attn_weights, p=self.dropout, training=self.training\n ) # (batch_size, n_heads, seq_length, key_length)\n\n # Mask heads if we want to\n if layer_head_mask is not None:\n attn_weights = attn_weights * layer_head_mask\n\n attn_output = unshape(\n torch.matmul(attn_weights, value_states)\n ) # (batch_size, seq_length, dim)\n attn_output = self.o(attn_output)\n\n present_key_value_state = (\n (key_states, value_states) if (self.is_decoder and use_cache) else None\n )\n outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)\n\n if output_attentions:\n outputs = outputs + (attn_weights,)\n return outputs\n\n\nclass T5LayerSelfAttention(nn.Module):\n def __init__(self, config, has_relative_attention_bias=False):\n super().__init__()\n self.SelfAttention = T5Attention(\n config, has_relative_attention_bias=has_relative_attention_bias\n )\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n position_bias=None,\n layer_head_mask=None,\n past_key_value=None,\n use_cache=False,\n output_attentions=False,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5LayerSelfAttention","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5LayerSelfAttention#L623-L656","kind":"class","name":"T5LayerSelfAttention","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":623,"end_line":656,"context_start_line":603,"context_end_line":676,"code":"\n # Mask heads if we want to\n if layer_head_mask is not None:\n attn_weights = attn_weights * layer_head_mask\n\n attn_output = unshape(\n torch.matmul(attn_weights, value_states)\n ) # (batch_size, seq_length, dim)\n attn_output = self.o(attn_output)\n\n present_key_value_state = (\n (key_states, value_states) if (self.is_decoder and use_cache) else None\n )\n outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)\n\n if output_attentions:\n outputs = outputs + (attn_weights,)\n return outputs\n\n\nclass T5LayerSelfAttention(nn.Module):\n def __init__(self, config, has_relative_attention_bias=False):\n super().__init__()\n self.SelfAttention = T5Attention(\n config, has_relative_attention_bias=has_relative_attention_bias\n )\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n position_bias=None,\n layer_head_mask=None,\n past_key_value=None,\n use_cache=False,\n output_attentions=False,\n ):\n normed_hidden_states = self.layer_norm(hidden_states)\n attention_output = self.SelfAttention(\n normed_hidden_states,\n mask=attention_mask,\n position_bias=position_bias,\n layer_head_mask=layer_head_mask,\n past_key_value=past_key_value,\n use_cache=use_cache,\n output_attentions=output_attentions,\n )\n hidden_states = hidden_states + self.dropout(attention_output[0])\n outputs = (hidden_states,) + attention_output[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass T5LayerCrossAttention(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.EncDecAttention = T5Attention(config, has_relative_attention_bias=False)\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(\n self,\n hidden_states,\n key_value_states,\n attention_mask=None,\n position_bias=None,\n layer_head_mask=None,\n past_key_value=None,\n use_cache=False,\n query_length=None,\n output_attentions=False,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5LayerCrossAttention","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5LayerCrossAttention#L659-L694","kind":"class","name":"T5LayerCrossAttention","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":659,"end_line":694,"context_start_line":639,"context_end_line":714,"code":" use_cache=False,\n output_attentions=False,\n ):\n normed_hidden_states = self.layer_norm(hidden_states)\n attention_output = self.SelfAttention(\n normed_hidden_states,\n mask=attention_mask,\n position_bias=position_bias,\n layer_head_mask=layer_head_mask,\n past_key_value=past_key_value,\n use_cache=use_cache,\n output_attentions=output_attentions,\n )\n hidden_states = hidden_states + self.dropout(attention_output[0])\n outputs = (hidden_states,) + attention_output[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass T5LayerCrossAttention(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.EncDecAttention = T5Attention(config, has_relative_attention_bias=False)\n self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)\n self.dropout = nn.Dropout(config.dropout_rate)\n\n def forward(\n self,\n hidden_states,\n key_value_states,\n attention_mask=None,\n position_bias=None,\n layer_head_mask=None,\n past_key_value=None,\n use_cache=False,\n query_length=None,\n output_attentions=False,\n ):\n normed_hidden_states = self.layer_norm(hidden_states)\n attention_output = self.EncDecAttention(\n normed_hidden_states,\n mask=attention_mask,\n key_value_states=key_value_states,\n position_bias=position_bias,\n layer_head_mask=layer_head_mask,\n past_key_value=past_key_value,\n use_cache=use_cache,\n query_length=query_length,\n output_attentions=output_attentions,\n )\n layer_output = hidden_states + self.dropout(attention_output[0])\n outputs = (layer_output,) + attention_output[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass T5Block(nn.Module):\n def __init__(self, config, has_relative_attention_bias=False):\n super().__init__()\n self.is_decoder = config.is_decoder\n self.layer = nn.ModuleList()\n self.layer.append(\n T5LayerSelfAttention(\n config, has_relative_attention_bias=has_relative_attention_bias\n )\n )\n if self.is_decoder:\n self.layer.append(T5LayerCrossAttention(config))\n\n self.layer.append(T5LayerFF(config))\n\n def forward(\n self,\n hidden_states,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5Block","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5Block#L697-L826","kind":"class","name":"T5Block","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":697,"end_line":826,"context_start_line":677,"context_end_line":846,"code":" ):\n normed_hidden_states = self.layer_norm(hidden_states)\n attention_output = self.EncDecAttention(\n normed_hidden_states,\n mask=attention_mask,\n key_value_states=key_value_states,\n position_bias=position_bias,\n layer_head_mask=layer_head_mask,\n past_key_value=past_key_value,\n use_cache=use_cache,\n query_length=query_length,\n output_attentions=output_attentions,\n )\n layer_output = hidden_states + self.dropout(attention_output[0])\n outputs = (layer_output,) + attention_output[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass T5Block(nn.Module):\n def __init__(self, config, has_relative_attention_bias=False):\n super().__init__()\n self.is_decoder = config.is_decoder\n self.layer = nn.ModuleList()\n self.layer.append(\n T5LayerSelfAttention(\n config, has_relative_attention_bias=has_relative_attention_bias\n )\n )\n if self.is_decoder:\n self.layer.append(T5LayerCrossAttention(config))\n\n self.layer.append(T5LayerFF(config))\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n position_bias=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n encoder_decoder_position_bias=None,\n layer_head_mask=None,\n cross_attn_layer_head_mask=None,\n past_key_value=None,\n use_cache=False,\n output_attentions=False,\n return_dict=True,\n ):\n\n if past_key_value is not None:\n if not self.is_decoder:\n logger.warning(\n \"`past_key_values` is passed to the encoder. Please make sure this is intended.\"\n )\n expected_num_past_key_values = 2 if encoder_hidden_states is None else 4\n\n if len(past_key_value) != expected_num_past_key_values:\n raise ValueError(\n f\"There should be {expected_num_past_key_values} past states. \"\n f\"{'2 (past / key) for cross attention. ' if expected_num_past_key_values == 4 else ''}\"\n f\"Got {len(past_key_value)} past key / value states\"\n )\n\n self_attn_past_key_value = past_key_value[:2]\n cross_attn_past_key_value = past_key_value[2:]\n else:\n self_attn_past_key_value, cross_attn_past_key_value = None, None\n\n self_attention_outputs = self.layer[0](\n hidden_states,\n attention_mask=attention_mask,\n position_bias=position_bias,\n layer_head_mask=layer_head_mask,\n past_key_value=self_attn_past_key_value,\n use_cache=use_cache,\n output_attentions=output_attentions,\n )\n hidden_states, present_key_value_state = self_attention_outputs[:2]\n attention_outputs = self_attention_outputs[\n 2:\n ] # Keep self-attention outputs and relative position weights\n\n # clamp inf values to enable fp16 training\n if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():\n clamp_value = torch.finfo(hidden_states.dtype).max - 1000\n hidden_states = torch.clamp(\n hidden_states, min=-clamp_value, max=clamp_value\n )\n\n do_cross_attention = self.is_decoder and encoder_hidden_states is not None\n if do_cross_attention:\n # the actual query length is unknown for cross attention\n # if using past key value states. Need to inject it here\n if present_key_value_state is not None:\n query_length = present_key_value_state[0].shape[2]\n else:\n query_length = None\n\n cross_attention_outputs = self.layer[1](\n hidden_states,\n key_value_states=encoder_hidden_states,\n attention_mask=encoder_attention_mask,\n position_bias=encoder_decoder_position_bias,\n layer_head_mask=cross_attn_layer_head_mask,\n past_key_value=cross_attn_past_key_value,\n query_length=query_length,\n use_cache=use_cache,\n output_attentions=output_attentions,\n )\n hidden_states = cross_attention_outputs[0]\n\n # clamp inf values to enable fp16 training\n if (\n hidden_states.dtype == torch.float16\n and torch.isinf(hidden_states).any()\n ):\n clamp_value = torch.finfo(hidden_states.dtype).max - 1000\n hidden_states = torch.clamp(\n hidden_states, min=-clamp_value, max=clamp_value\n )\n\n # Combine self attn and cross attn key value states\n if present_key_value_state is not None:\n present_key_value_state = (\n present_key_value_state + cross_attention_outputs[1]\n )\n\n # Keep cross-attention outputs and relative position weights\n attention_outputs = attention_outputs + cross_attention_outputs[2:]\n\n # Apply Feed Forward layer\n hidden_states = self.layer[-1](hidden_states)\n\n # clamp inf values to enable fp16 training\n if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():\n clamp_value = torch.finfo(hidden_states.dtype).max - 1000\n hidden_states = torch.clamp(\n hidden_states, min=-clamp_value, max=clamp_value\n )\n\n outputs = (hidden_states,)\n\n if use_cache:\n outputs = outputs + (present_key_value_state,) + attention_outputs\n else:\n outputs = outputs + attention_outputs\n\n return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)\n\n\nclass T5PreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = T5Config\n load_tf_weights = load_tf_weights_in_t5\n base_model_prefix = \"transformer\"\n is_parallelizable = True\n supports_gradient_checkpointing = True\n _no_split_modules = [\"T5Block\"]\n\n @property\n def dummy_inputs(self):\n input_ids = torch.tensor(DUMMY_INPUTS)\n input_mask = torch.tensor(DUMMY_MASK)\n dummy_inputs = {","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5PreTrainedModel","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5PreTrainedModel#L829-L948","kind":"class","name":"T5PreTrainedModel","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":829,"end_line":948,"context_start_line":809,"context_end_line":968,"code":" # Apply Feed Forward layer\n hidden_states = self.layer[-1](hidden_states)\n\n # clamp inf values to enable fp16 training\n if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():\n clamp_value = torch.finfo(hidden_states.dtype).max - 1000\n hidden_states = torch.clamp(\n hidden_states, min=-clamp_value, max=clamp_value\n )\n\n outputs = (hidden_states,)\n\n if use_cache:\n outputs = outputs + (present_key_value_state,) + attention_outputs\n else:\n outputs = outputs + attention_outputs\n\n return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)\n\n\nclass T5PreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = T5Config\n load_tf_weights = load_tf_weights_in_t5\n base_model_prefix = \"transformer\"\n is_parallelizable = True\n supports_gradient_checkpointing = True\n _no_split_modules = [\"T5Block\"]\n\n @property\n def dummy_inputs(self):\n input_ids = torch.tensor(DUMMY_INPUTS)\n input_mask = torch.tensor(DUMMY_MASK)\n dummy_inputs = {\n \"decoder_input_ids\": input_ids,\n \"input_ids\": input_ids,\n \"decoder_attention_mask\": input_mask,\n }\n return dummy_inputs\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n factor = (\n self.config.initializer_factor\n ) # Used for testing weights initialization\n if isinstance(module, T5LayerNorm):\n module.weight.data.fill_(factor * 1.0)\n elif isinstance(module, (T5Model, T5ForConditionalGeneration, T5EncoderModel)):\n # Mesh TensorFlow embeddings initialization\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624\n module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)\n if hasattr(module, \"lm_head\") and not self.config.tie_word_embeddings:\n module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)\n elif isinstance(module, T5DenseActDense):\n # Mesh TensorFlow FF initialization\n # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56\n # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89\n module.wi.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi, \"bias\") and module.wi.bias is not None:\n module.wi.bias.data.zero_()\n module.wo.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_ff) ** -0.5)\n )\n if hasattr(module.wo, \"bias\") and module.wo.bias is not None:\n module.wo.bias.data.zero_()\n elif isinstance(module, T5DenseGatedActDense):\n module.wi_0.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi_0, \"bias\") and module.wi_0.bias is not None:\n module.wi_0.bias.data.zero_()\n module.wi_1.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi_1, \"bias\") and module.wi_1.bias is not None:\n module.wi_1.bias.data.zero_()\n module.wo.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_ff) ** -0.5)\n )\n if hasattr(module.wo, \"bias\") and module.wo.bias is not None:\n module.wo.bias.data.zero_()\n elif isinstance(module, T5Attention):\n # Mesh TensorFlow attention initialization to avoid scaling before softmax\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136\n d_model = self.config.d_model\n key_value_proj_dim = self.config.d_kv\n n_heads = self.config.num_heads\n module.q.weight.data.normal_(\n mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5)\n )\n module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.o.weight.data.normal_(\n mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5)\n )\n if module.has_relative_attention_bias:\n module.relative_attention_bias.weight.data.normal_(\n mean=0.0, std=factor * ((d_model) ** -0.5)\n )\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (T5Attention, T5Stack)):\n module.gradient_checkpointing = value\n\n def _shift_right(self, input_ids):\n decoder_start_token_id = self.config.decoder_start_token_id\n pad_token_id = self.config.pad_token_id\n\n assert decoder_start_token_id is not None, (\n \"self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id.\"\n \" See T5 docs for more information\"\n )\n\n # shift inputs to the right\n if is_torch_fx_proxy(input_ids):\n # Item assignment is not supported natively for proxies.\n shifted_input_ids = torch.full(\n input_ids.shape[:-1] + (1,), decoder_start_token_id\n )\n shifted_input_ids = torch.cat(\n [shifted_input_ids, input_ids[..., :-1]], dim=-1\n )\n else:\n shifted_input_ids = input_ids.new_zeros(input_ids.shape)\n shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()\n shifted_input_ids[..., 0] = decoder_start_token_id\n\n assert (\n pad_token_id is not None\n ), \"self.model.config.pad_token_id has to be defined.\"\n # replace possible -100 values in labels by `pad_token_id`\n shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)\n\n return shifted_input_ids\n\n\nclass T5Stack(T5PreTrainedModel):\n def __init__(self, config, embed_tokens=None):\n super().__init__(config)\n\n self.embed_tokens = embed_tokens\n self.is_decoder = config.is_decoder\n\n self.block = nn.ModuleList(\n [\n T5Block(config, has_relative_attention_bias=bool(i == 0))\n for i in range(config.num_layers)\n ]\n )\n self.final_layer_norm = T5LayerNorm(\n config.d_model, eps=config.layer_norm_epsilon\n )\n self.dropout = nn.Dropout(config.dropout_rate)\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5Stack","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5Stack#L951-L1282","kind":"class","name":"T5Stack","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":951,"end_line":1282,"context_start_line":931,"context_end_line":1302,"code":" shifted_input_ids = torch.full(\n input_ids.shape[:-1] + (1,), decoder_start_token_id\n )\n shifted_input_ids = torch.cat(\n [shifted_input_ids, input_ids[..., :-1]], dim=-1\n )\n else:\n shifted_input_ids = input_ids.new_zeros(input_ids.shape)\n shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()\n shifted_input_ids[..., 0] = decoder_start_token_id\n\n assert (\n pad_token_id is not None\n ), \"self.model.config.pad_token_id has to be defined.\"\n # replace possible -100 values in labels by `pad_token_id`\n shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)\n\n return shifted_input_ids\n\n\nclass T5Stack(T5PreTrainedModel):\n def __init__(self, config, embed_tokens=None):\n super().__init__(config)\n\n self.embed_tokens = embed_tokens\n self.is_decoder = config.is_decoder\n\n self.block = nn.ModuleList(\n [\n T5Block(config, has_relative_attention_bias=bool(i == 0))\n for i in range(config.num_layers)\n ]\n )\n self.final_layer_norm = T5LayerNorm(\n config.d_model, eps=config.layer_norm_epsilon\n )\n self.dropout = nn.Dropout(config.dropout_rate)\n\n # Initialize weights and apply final processing\n self.post_init()\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n self.gradient_checkpointing = False\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n # Check validity of device_map\n self.device_map = (\n get_device_map(len(self.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.block))\n self.model_parallel = True\n self.first_device = (\n \"cpu\"\n if \"cpu\" in self.device_map.keys()\n else \"cuda:\" + str(min(self.device_map.keys()))\n )\n self.last_device = \"cuda:\" + str(max(self.device_map.keys()))\n # Load onto devices\n for k, v in self.device_map.items():\n for layer in v:\n cuda_device = \"cuda:\" + str(k)\n self.block[layer] = self.block[layer].to(cuda_device)\n\n # Set embed_tokens to first layer\n self.embed_tokens = self.embed_tokens.to(self.first_device)\n # Set final layer norm to last device\n self.final_layer_norm = self.final_layer_norm.to(self.last_device)\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.model_parallel = False\n self.device_map = None\n self.first_device = \"cpu\"\n self.last_device = \"cpu\"\n for i in range(len(self.block)):\n self.block[i] = self.block[i].to(\"cpu\")\n self.embed_tokens = self.embed_tokens.to(\"cpu\")\n self.final_layer_norm = self.final_layer_norm.to(\"cpu\")\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.embed_tokens\n\n def set_input_embeddings(self, new_embeddings):\n self.embed_tokens = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n inputs_embeds=None,\n head_mask=None,\n cross_attn_head_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n ):\n # Model parallel\n if self.model_parallel:\n torch.cuda.set_device(self.first_device)\n self.embed_tokens = self.embed_tokens.to(self.first_device)\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n if input_ids is not None and inputs_embeds is not None:\n err_msg_prefix = \"decoder_\" if self.is_decoder else \"\"\n raise ValueError(\n f\"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time\"\n )\n elif input_ids is not None:\n input_shape = input_ids.size()\n input_ids = input_ids.view(-1, input_shape[-1])\n elif inputs_embeds is not None:\n input_shape = inputs_embeds.size()[:-1]\n else:\n err_msg_prefix = \"decoder_\" if self.is_decoder else \"\"\n raise ValueError(\n f\"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds\"\n )\n\n if inputs_embeds is None:\n assert (\n self.embed_tokens is not None\n ), \"You have to initialize the model with valid token embeddings\"\n inputs_embeds = self.embed_tokens(input_ids)\n\n batch_size, seq_length = input_shape\n\n # required mask seq length can be calculated via length of past\n mask_seq_length = (\n past_key_values[0][0].shape[2] + seq_length\n if past_key_values is not None\n else seq_length\n )\n\n if use_cache is True:\n assert (\n self.is_decoder\n ), f\"`use_cache` can only be set to `True` if {self} is used as a decoder\"\n\n if attention_mask is None:\n attention_mask = torch.ones(\n batch_size, mask_seq_length, device=inputs_embeds.device\n )\n if (\n self.is_decoder\n and encoder_attention_mask is None\n and encoder_hidden_states is not None\n ):\n encoder_seq_length = encoder_hidden_states.shape[1]\n encoder_attention_mask = torch.ones(\n batch_size,\n encoder_seq_length,\n device=inputs_embeds.device,\n dtype=torch.long,\n )\n\n # initialize past_key_values with `None` if past does not exist\n if past_key_values is None:\n past_key_values = [None] * len(self.block)\n\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n extended_attention_mask = self.get_extended_attention_mask(\n attention_mask, input_shape\n )\n\n # If a 2D or 3D attention mask is provided for the cross-attention\n # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if self.is_decoder and encoder_hidden_states is not None:\n (\n encoder_batch_size,\n encoder_sequence_length,\n _,\n ) = encoder_hidden_states.size()\n encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)\n if encoder_attention_mask is None:\n encoder_attention_mask = torch.ones(\n encoder_hidden_shape, device=inputs_embeds.device\n )\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = None\n\n # Prepare head mask if needed\n head_mask = self.get_head_mask(head_mask, self.config.num_layers)\n cross_attn_head_mask = self.get_head_mask(\n cross_attn_head_mask, self.config.num_layers\n )\n present_key_value_states = () if use_cache else None\n all_hidden_states = () if output_hidden_states else None\n all_attentions = () if output_attentions else None\n all_cross_attentions = () if (output_attentions and self.is_decoder) else None\n position_bias = None\n encoder_decoder_position_bias = None\n\n hidden_states = self.dropout(inputs_embeds)\n\n for i, (layer_module, past_key_value) in enumerate(\n zip(self.block, past_key_values)\n ):\n layer_head_mask = head_mask[i]\n cross_attn_layer_head_mask = cross_attn_head_mask[i]\n # Model parallel\n if self.model_parallel:\n torch.cuda.set_device(hidden_states.device)\n # Ensure that attention_mask is always on the same device as hidden_states\n if attention_mask is not None:\n attention_mask = attention_mask.to(hidden_states.device)\n if position_bias is not None:\n position_bias = position_bias.to(hidden_states.device)\n if encoder_hidden_states is not None:\n encoder_hidden_states = encoder_hidden_states.to(\n hidden_states.device\n )\n if encoder_extended_attention_mask is not None:\n encoder_extended_attention_mask = (\n encoder_extended_attention_mask.to(hidden_states.device)\n )\n if encoder_decoder_position_bias is not None:\n encoder_decoder_position_bias = encoder_decoder_position_bias.to(\n hidden_states.device\n )\n if layer_head_mask is not None:\n layer_head_mask = layer_head_mask.to(hidden_states.device)\n if cross_attn_layer_head_mask is not None:\n cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(\n hidden_states.device\n )\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if self.gradient_checkpointing and self.training:\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return tuple(module(*inputs, use_cache, output_attentions))\n\n return custom_forward\n\n layer_outputs = checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n extended_attention_mask,\n position_bias,\n encoder_hidden_states,\n encoder_extended_attention_mask,\n encoder_decoder_position_bias,\n layer_head_mask,\n cross_attn_layer_head_mask,\n None, # past_key_value is always None with gradient checkpointing\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask=extended_attention_mask,\n position_bias=position_bias,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,\n encoder_decoder_position_bias=encoder_decoder_position_bias,\n layer_head_mask=layer_head_mask,\n cross_attn_layer_head_mask=cross_attn_layer_head_mask,\n past_key_value=past_key_value,\n use_cache=use_cache,\n output_attentions=output_attentions,\n )\n\n # layer_outputs is a tuple with:\n # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)\n if use_cache is False:\n layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]\n\n hidden_states, present_key_value_state = layer_outputs[:2]\n\n # We share the position biases between the layers - the first layer store them\n # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),\n # (cross-attention position bias), (cross-attention weights)\n position_bias = layer_outputs[2]\n if self.is_decoder and encoder_hidden_states is not None:\n encoder_decoder_position_bias = layer_outputs[\n 4 if output_attentions else 3\n ]\n # append next layer key value states\n if use_cache:\n present_key_value_states = present_key_value_states + (\n present_key_value_state,\n )\n\n if output_attentions:\n all_attentions = all_attentions + (layer_outputs[3],)\n if self.is_decoder:\n all_cross_attentions = all_cross_attentions + (layer_outputs[5],)\n\n # Model Parallel: If it's the last layer for that device, put things on the next device\n if self.model_parallel:\n for k, v in self.device_map.items():\n if i == v[-1] and \"cuda:\" + str(k) != self.last_device:\n hidden_states = hidden_states.to(\"cuda:\" + str(k + 1))\n\n hidden_states = self.final_layer_norm(hidden_states)\n hidden_states = self.dropout(hidden_states)\n\n # Add last layer\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if not return_dict:\n return tuple(\n v\n for v in [\n hidden_states,\n present_key_value_states,\n all_hidden_states,\n all_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=present_key_value_states,\n hidden_states=all_hidden_states,\n attentions=all_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nT5_START_DOCSTRING = r\"\"\"\n\n The T5 model was proposed in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text\n Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan\n Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. It's an encoder decoder transformer pre-trained in a\n text-to-text denoising generative setting.\n\n This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the\n library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads\n etc.)\n\n This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.\n Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage\n and behavior.\n\n Parameters:\n config ([`T5Config`]): Model configuration class with all the parameters of the model.\n Initializing with a config file does not load the weights associated with the model, only the","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5Model","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5Model#L1449-L1643","kind":"class","name":"T5Model","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1449,"end_line":1643,"context_start_line":1429,"context_end_line":1663,"code":" output_hidden_states (`bool`, *optional*):\n Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for\n more detail.\n return_dict (`bool`, *optional*):\n Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.\n\"\"\"\n\n# Warning message for FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask\n__HEAD_MASK_WARNING_MSG = \"\"\"\nThe input argument `head_mask` was split into two arguments `head_mask` and `decoder_head_mask`. Currently,\n`decoder_head_mask` is set to copy `head_mask`, but this feature is deprecated and will be removed in future versions.\nIf you do not want to use any `decoder_head_mask` now, please set `decoder_head_mask = torch.ones(num_layers,\nnum_heads)`.\n\"\"\"\n\n\n@add_start_docstrings(\n \"The bare T5 Model transformer outputting raw hidden-states without any specific head on top.\",\n T5_START_DOCSTRING,\n)\nclass T5Model(T5PreTrainedModel):\n _keys_to_ignore_on_load_missing = [\n r\"encoder.embed_tokens.weight\",\n r\"decoder.embed_tokens.weight\",\n ]\n _keys_to_ignore_on_load_unexpected = [\n r\"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.is_decoder = False\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n\n decoder_config = copy.deepcopy(config)\n decoder_config.is_decoder = True\n decoder_config.is_encoder_decoder = False\n decoder_config.num_layers = config.num_decoder_layers\n self.decoder = T5Stack(decoder_config, self.shared)\n\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.decoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.decoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.decoder = self.decoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n self.decoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def get_decoder(self):\n return self.decoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n decoder_input_ids: Optional[torch.LongTensor] = None,\n decoder_attention_mask: Optional[torch.BoolTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n decoder_head_mask: Optional[torch.FloatTensor] = None,\n cross_attn_head_mask: Optional[torch.Tensor] = None,\n encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,\n past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,\n inputs_embeds: Optional[torch.Tensor] = None,\n decoder_inputs_embeds: Optional[torch.Tensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:\n r\"\"\"\n Returns:\n\n Example:\n\n ```python\n >>> from transformers import T5Tokenizer, T5Model\n\n >>> tokenizer = T5Tokenizer.from_pretrained(\"t5-small\")\n >>> model = T5Model.from_pretrained(\"t5-small\")\n\n >>> input_ids = tokenizer(\n ... \"Studies have been shown that owning a dog is good for you\", return_tensors=\"pt\"\n ... ).input_ids # Batch size 1\n >>> decoder_input_ids = tokenizer(\"Studies show that\", return_tensors=\"pt\").input_ids # Batch size 1\n\n >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for T5Model.\n >>> # This is not needed for torch's T5ForConditionalGeneration as it does this internally using labels arg.\n >>> decoder_input_ids = model._shift_right(decoder_input_ids)\n\n >>> # forward pass\n >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)\n >>> last_hidden_states = outputs.last_hidden_state\n ```\"\"\"\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask\n if head_mask is not None and decoder_head_mask is None:\n if self.config.num_layers == self.config.num_decoder_layers:\n warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)\n decoder_head_mask = head_mask\n\n # Encode if needed (training, first prediction pass)\n if encoder_outputs is None:\n encoder_outputs = self.encoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n inputs_embeds=inputs_embeds,\n head_mask=head_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):\n encoder_outputs = BaseModelOutput(\n last_hidden_state=encoder_outputs[0],\n hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,\n attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,\n )\n\n hidden_states = encoder_outputs[0]\n\n # Set device for model parallelism\n if self.model_parallel:\n torch.cuda.set_device(self.decoder.first_device)\n hidden_states = hidden_states.to(self.decoder.first_device)\n if decoder_input_ids is not None:\n decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)\n if attention_mask is not None:\n attention_mask = attention_mask.to(self.decoder.first_device)\n if decoder_attention_mask is not None:\n decoder_attention_mask = decoder_attention_mask.to(\n self.decoder.first_device\n )\n\n # Decode\n decoder_outputs = self.decoder(\n input_ids=decoder_input_ids,\n attention_mask=decoder_attention_mask,\n inputs_embeds=decoder_inputs_embeds,\n past_key_values=past_key_values,\n encoder_hidden_states=hidden_states,\n encoder_attention_mask=attention_mask,\n head_mask=decoder_head_mask,\n cross_attn_head_mask=cross_attn_head_mask,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n if not return_dict:\n return decoder_outputs + encoder_outputs\n\n return Seq2SeqModelOutput(\n last_hidden_state=decoder_outputs.last_hidden_state,\n past_key_values=decoder_outputs.past_key_values,\n decoder_hidden_states=decoder_outputs.hidden_states,\n decoder_attentions=decoder_outputs.attentions,\n cross_attentions=decoder_outputs.cross_attentions,\n encoder_last_hidden_state=encoder_outputs.last_hidden_state,\n encoder_hidden_states=encoder_outputs.hidden_states,\n encoder_attentions=encoder_outputs.attentions,\n )\n\n\n@add_start_docstrings(\n \"\"\"T5 Model with a `language modeling` head on top.\"\"\", T5_START_DOCSTRING\n)\nclass T5ForConditionalGeneration(T5PreTrainedModel):\n _keys_to_ignore_on_load_missing = [\n r\"encoder.embed_tokens.weight\",\n r\"decoder.embed_tokens.weight\",\n r\"lm_head.weight\",\n ]\n _keys_to_ignore_on_load_unexpected = [\n r\"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.model_dim = config.d_model\n\n self.shared = nn.Embedding(config.vocab_size, config.d_model)","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5ForConditionalGeneration","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5ForConditionalGeneration#L1649-L1957","kind":"class","name":"T5ForConditionalGeneration","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1649,"end_line":1957,"context_start_line":1629,"context_end_line":1977,"code":" )\n\n if not return_dict:\n return decoder_outputs + encoder_outputs\n\n return Seq2SeqModelOutput(\n last_hidden_state=decoder_outputs.last_hidden_state,\n past_key_values=decoder_outputs.past_key_values,\n decoder_hidden_states=decoder_outputs.hidden_states,\n decoder_attentions=decoder_outputs.attentions,\n cross_attentions=decoder_outputs.cross_attentions,\n encoder_last_hidden_state=encoder_outputs.last_hidden_state,\n encoder_hidden_states=encoder_outputs.hidden_states,\n encoder_attentions=encoder_outputs.attentions,\n )\n\n\n@add_start_docstrings(\n \"\"\"T5 Model with a `language modeling` head on top.\"\"\", T5_START_DOCSTRING\n)\nclass T5ForConditionalGeneration(T5PreTrainedModel):\n _keys_to_ignore_on_load_missing = [\n r\"encoder.embed_tokens.weight\",\n r\"decoder.embed_tokens.weight\",\n r\"lm_head.weight\",\n ]\n _keys_to_ignore_on_load_unexpected = [\n r\"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.model_dim = config.d_model\n\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.is_decoder = False\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n\n decoder_config = copy.deepcopy(config)\n decoder_config.is_decoder = True\n decoder_config.is_encoder_decoder = False\n decoder_config.num_layers = config.num_decoder_layers\n self.decoder = T5Stack(decoder_config, self.shared)\n\n self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)\n\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.decoder.parallelize(self.device_map)\n self.lm_head = self.lm_head.to(self.decoder.first_device)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.decoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.decoder = self.decoder.to(\"cpu\")\n self.lm_head = self.lm_head.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n self.decoder.set_input_embeddings(new_embeddings)\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def get_encoder(self):\n return self.encoder\n\n def get_decoder(self):\n return self.decoder\n\n @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n decoder_input_ids: Optional[torch.LongTensor] = None,\n decoder_attention_mask: Optional[torch.BoolTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n decoder_head_mask: Optional[torch.FloatTensor] = None,\n cross_attn_head_mask: Optional[torch.Tensor] = None,\n encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,\n past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n decoder_inputs_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n reduction: Optional[str] = \"mean\",\n ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:\n r\"\"\"\n labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):\n Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,\n config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for\n labels in `[0, ..., config.vocab_size]`\n\n Returns:\n\n Examples:\n\n ```python\n >>> from transformers import T5Tokenizer, T5ForConditionalGeneration\n\n >>> tokenizer = T5Tokenizer.from_pretrained(\"t5-small\")\n >>> model = T5ForConditionalGeneration.from_pretrained(\"t5-small\")\n\n >>> # training\n >>> input_ids = tokenizer(\"The walks in park\", return_tensors=\"pt\").input_ids\n >>> labels = tokenizer(\" cute dog the \", return_tensors=\"pt\").input_ids\n >>> outputs = model(input_ids=input_ids, labels=labels)\n >>> loss = outputs.loss\n >>> logits = outputs.logits\n\n >>> # inference\n >>> input_ids = tokenizer(\n ... \"summarize: studies have shown that owning a dog is good for you\", return_tensors=\"pt\"\n ... ).input_ids # Batch size 1\n >>> outputs = model.generate(input_ids)\n >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))\n >>> # studies have shown that owning a dog is good for you.\n ```\"\"\"\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask\n if head_mask is not None and decoder_head_mask is None:\n if self.config.num_layers == self.config.num_decoder_layers:\n warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)\n decoder_head_mask = head_mask\n\n # Encode if needed (training, first prediction pass)\n if encoder_outputs is None:\n # Convert encoder inputs in embeddings if needed\n encoder_outputs = self.encoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n inputs_embeds=inputs_embeds,\n head_mask=head_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):\n encoder_outputs = BaseModelOutput(\n last_hidden_state=encoder_outputs[0],\n hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,\n attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,\n )\n \n # print('encoder_outputs[last_hidden_state]', encoder_outputs.shape) \n # print('encoder_outputs[attentions]', encoder_outputs[2].shape)\n \n hidden_states = encoder_outputs[0]\n\n if self.model_parallel:\n torch.cuda.set_device(self.decoder.first_device)\n\n if (\n labels is not None\n and decoder_input_ids is None\n and decoder_inputs_embeds is None\n ):\n # get decoder inputs from shifting lm labels to the right\n decoder_input_ids = self._shift_right(labels)\n\n # Set device for model parallelism\n if self.model_parallel:\n torch.cuda.set_device(self.decoder.first_device)\n hidden_states = hidden_states.to(self.decoder.first_device)\n if decoder_input_ids is not None:\n decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)\n if attention_mask is not None:\n attention_mask = attention_mask.to(self.decoder.first_device)\n if decoder_attention_mask is not None:\n decoder_attention_mask = decoder_attention_mask.to(\n self.decoder.first_device\n )\n\n # Decode\n decoder_outputs = self.decoder(\n input_ids=decoder_input_ids,\n attention_mask=decoder_attention_mask,\n inputs_embeds=decoder_inputs_embeds,\n past_key_values=past_key_values,\n encoder_hidden_states=hidden_states,\n encoder_attention_mask=attention_mask,\n head_mask=decoder_head_mask,\n cross_attn_head_mask=cross_attn_head_mask,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n sequence_output = decoder_outputs[0]\n\n # Set device for model parallelism\n if self.model_parallel:\n torch.cuda.set_device(self.encoder.first_device)\n self.lm_head = self.lm_head.to(self.encoder.first_device)\n sequence_output = sequence_output.to(self.lm_head.weight.device)\n\n if self.config.tie_word_embeddings:\n # Rescale output before projecting on vocab\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586\n sequence_output = sequence_output * (self.model_dim**-0.5)\n\n lm_logits = self.lm_head(sequence_output)\n\n loss = None\n if labels is not None:\n loss_fct = CrossEntropyLoss(ignore_index=-100, reduction=reduction)\n loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))\n if reduction == \"none\":\n loss = loss.view(lm_logits.size(0), -1).sum(1)\n\n if not return_dict:\n output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs\n return ((loss,) + output) if loss is not None else output\n\n return Seq2SeqLMOutput(\n loss=loss,\n logits=lm_logits,\n past_key_values=decoder_outputs.past_key_values,\n decoder_hidden_states=decoder_outputs.hidden_states,\n decoder_attentions=decoder_outputs.attentions,\n cross_attentions=decoder_outputs.cross_attentions,\n encoder_last_hidden_state=encoder_outputs.last_hidden_state,\n encoder_hidden_states=encoder_outputs.hidden_states,\n encoder_attentions=encoder_outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self,\n input_ids,\n past=None,\n attention_mask=None,\n head_mask=None,\n decoder_head_mask=None,\n cross_attn_head_mask=None,\n use_cache=None,\n encoder_outputs=None,\n **kwargs,\n ):\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"decoder_input_ids\": input_ids,\n \"past_key_values\": past,\n \"encoder_outputs\": encoder_outputs,\n \"attention_mask\": attention_mask,\n \"head_mask\": head_mask,\n \"decoder_head_mask\": decoder_head_mask,\n \"cross_attn_head_mask\": cross_attn_head_mask,\n \"use_cache\": use_cache,\n }\n\n def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):\n return self._shift_right(labels)\n\n def _reorder_cache(self, past, beam_idx):\n # if decoder past is not included in output\n # speedy decoding is disabled and no need to reorder\n if past is None:\n logger.warning(\n \"You might want to consider setting `use_cache=True` to speed up decoding\"\n )\n return past\n\n reordered_decoder_past = ()\n for layer_past_states in past:\n # get the correct batch idx from layer past batch dim\n # batch dim of `past` is at 2nd position\n reordered_layer_past_states = ()\n for layer_past_state in layer_past_states:\n # need to set correct `past` for each of the four key / value states\n reordered_layer_past_states = reordered_layer_past_states + (\n layer_past_state.index_select(\n 0, beam_idx.to(layer_past_state.device)\n ),\n )\n\n assert reordered_layer_past_states[0].shape == layer_past_states[0].shape\n assert len(reordered_layer_past_states) == len(layer_past_states)\n\n reordered_decoder_past = reordered_decoder_past + (\n reordered_layer_past_states,\n )\n return reordered_decoder_past\n\n\n@add_start_docstrings(\n \"The bare T5 Model transformer outputting encoder's raw hidden-states without any specific head on top.\",\n T5_START_DOCSTRING,\n)\nclass T5EncoderModel(T5PreTrainedModel):\n authorized_missing_keys = [\n r\"encoder.embed_tokens.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.T5EncoderModel","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_t5.T5EncoderModel#L1964-L2066","kind":"class","name":"T5EncoderModel","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1964,"end_line":2066,"context_start_line":1944,"context_end_line":2066,"code":" # need to set correct `past` for each of the four key / value states\n reordered_layer_past_states = reordered_layer_past_states + (\n layer_past_state.index_select(\n 0, beam_idx.to(layer_past_state.device)\n ),\n )\n\n assert reordered_layer_past_states[0].shape == layer_past_states[0].shape\n assert len(reordered_layer_past_states) == len(layer_past_states)\n\n reordered_decoder_past = reordered_decoder_past + (\n reordered_layer_past_states,\n )\n return reordered_decoder_past\n\n\n@add_start_docstrings(\n \"The bare T5 Model transformer outputting encoder's raw hidden-states without any specific head on top.\",\n T5_START_DOCSTRING,\n)\nclass T5EncoderModel(T5PreTrainedModel):\n authorized_missing_keys = [\n r\"encoder.embed_tokens.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:\n r\"\"\"\n Returns:\n\n Example:\n\n ```python\n >>> from transformers import T5Tokenizer, T5EncoderModel\n\n >>> tokenizer = T5Tokenizer.from_pretrained(\"t5-small\")\n >>> model = T5EncoderModel.from_pretrained(\"t5-small\")\n >>> input_ids = tokenizer(\n ... \"Studies have been shown that owning a dog is good for you\", return_tensors=\"pt\"\n ... ).input_ids # Batch size 1\n >>> outputs = model(input_ids=input_ids)\n >>> last_hidden_states = outputs.last_hidden_state\n ```\"\"\"\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n encoder_outputs = self.encoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n inputs_embeds=inputs_embeds,\n head_mask=head_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n return encoder_outputs","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.__init__#L1969-L1983","kind":"function","name":"__init__","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1969,"end_line":1983,"context_start_line":1949,"context_end_line":2003,"code":" )\n\n assert reordered_layer_past_states[0].shape == layer_past_states[0].shape\n assert len(reordered_layer_past_states) == len(layer_past_states)\n\n reordered_decoder_past = reordered_decoder_past + (\n reordered_layer_past_states,\n )\n return reordered_decoder_past\n\n\n@add_start_docstrings(\n \"The bare T5 Model transformer outputting encoder's raw hidden-states without any specific head on top.\",\n T5_START_DOCSTRING,\n)\nclass T5EncoderModel(T5PreTrainedModel):\n authorized_missing_keys = [\n r\"encoder.embed_tokens.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.forward#L2026-L2066","kind":"function","name":"forward","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":2026,"end_line":2066,"context_start_line":2006,"context_end_line":2066,"code":"\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:\n r\"\"\"\n Returns:\n\n Example:\n\n ```python\n >>> from transformers import T5Tokenizer, T5EncoderModel\n\n >>> tokenizer = T5Tokenizer.from_pretrained(\"t5-small\")\n >>> model = T5EncoderModel.from_pretrained(\"t5-small\")\n >>> input_ids = tokenizer(\n ... \"Studies have been shown that owning a dog is good for you\", return_tensors=\"pt\"\n ... ).input_ids # Batch size 1\n >>> outputs = model(input_ids=input_ids)\n >>> last_hidden_states = outputs.last_hidden_state\n ```\"\"\"\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n encoder_outputs = self.encoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n inputs_embeds=inputs_embeds,\n head_mask=head_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n return encoder_outputs","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.prune_heads","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.prune_heads#L376-L390","kind":"function","name":"prune_heads","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":376,"end_line":390,"context_start_line":356,"context_end_line":410,"code":" self.relative_attention_max_distance = config.relative_attention_max_distance\n self.d_model = config.d_model\n self.key_value_proj_dim = config.d_kv\n self.n_heads = config.num_heads\n self.dropout = config.dropout_rate\n self.inner_dim = self.n_heads * self.key_value_proj_dim\n\n # Mesh TensorFlow initialization to avoid scaling before softmax\n self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)\n self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)\n\n if self.has_relative_attention_bias:\n self.relative_attention_bias = nn.Embedding(\n self.relative_attention_num_buckets, self.n_heads\n )\n self.pruned_heads = set()\n self.gradient_checkpointing = False\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads\n )\n # Prune linear layers\n self.q = prune_linear_layer(self.q, index)\n self.k = prune_linear_layer(self.k, index)\n self.v = prune_linear_layer(self.v, index)\n self.o = prune_linear_layer(self.o, index, dim=1)\n # Update hyper params\n self.n_heads = self.n_heads - len(heads)\n self.inner_dim = self.key_value_proj_dim * self.n_heads\n self.pruned_heads = self.pruned_heads.union(heads)\n\n @staticmethod\n def _relative_position_bucket(\n relative_position, bidirectional=True, num_buckets=32, max_distance=128\n ):\n \"\"\"\n Adapted from Mesh Tensorflow:\n https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593\n\n Translate relative position to a bucket number for relative attention. The relative position is defined as\n memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to\n position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for\n small absolute relative_position and larger buckets for larger absolute relative_positions. All relative\n positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.\n This should allow for more graceful generalization to longer sequences than the model has been trained on\n\n Args:\n relative_position: an int32 Tensor\n bidirectional: a boolean - whether the attention is bidirectional\n num_buckets: an integer","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._relative_position_bucket","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._relative_position_bucket#L393-L445","kind":"function","name":"_relative_position_bucket","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":393,"end_line":445,"context_start_line":373,"context_end_line":465,"code":" self.pruned_heads = set()\n self.gradient_checkpointing = False\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads\n )\n # Prune linear layers\n self.q = prune_linear_layer(self.q, index)\n self.k = prune_linear_layer(self.k, index)\n self.v = prune_linear_layer(self.v, index)\n self.o = prune_linear_layer(self.o, index, dim=1)\n # Update hyper params\n self.n_heads = self.n_heads - len(heads)\n self.inner_dim = self.key_value_proj_dim * self.n_heads\n self.pruned_heads = self.pruned_heads.union(heads)\n\n @staticmethod\n def _relative_position_bucket(\n relative_position, bidirectional=True, num_buckets=32, max_distance=128\n ):\n \"\"\"\n Adapted from Mesh Tensorflow:\n https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593\n\n Translate relative position to a bucket number for relative attention. The relative position is defined as\n memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to\n position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for\n small absolute relative_position and larger buckets for larger absolute relative_positions. All relative\n positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.\n This should allow for more graceful generalization to longer sequences than the model has been trained on\n\n Args:\n relative_position: an int32 Tensor\n bidirectional: a boolean - whether the attention is bidirectional\n num_buckets: an integer\n max_distance: an integer\n\n Returns:\n a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)\n \"\"\"\n relative_buckets = 0\n if bidirectional:\n num_buckets //= 2\n relative_buckets += (relative_position > 0).to(torch.long) * num_buckets\n relative_position = torch.abs(relative_position)\n else:\n relative_position = -torch.min(\n relative_position, torch.zeros_like(relative_position)\n )\n # now relative_position is in the range [0, inf)\n\n # half of the buckets are for exact increments in positions\n max_exact = num_buckets // 2\n is_small = relative_position < max_exact\n\n # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance\n relative_position_if_large = max_exact + (\n torch.log(relative_position.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_position_if_large = torch.min(\n relative_position_if_large,\n torch.full_like(relative_position_if_large, num_buckets - 1),\n )\n\n relative_buckets += torch.where(\n is_small, relative_position, relative_position_if_large\n )\n return relative_buckets\n\n def compute_bias(self, query_length, key_length, device=None):\n \"\"\"Compute binned relative position bias\"\"\"\n if device is None:\n device = self.relative_attention_bias.weight.device\n context_position = torch.arange(query_length, dtype=torch.long, device=device)[\n :, None\n ]\n memory_position = torch.arange(key_length, dtype=torch.long, device=device)[\n None, :\n ]\n relative_position = (\n memory_position - context_position\n ) # shape (query_length, key_length)\n relative_position_bucket = self._relative_position_bucket(\n relative_position, # shape (query_length, key_length)\n bidirectional=(not self.is_decoder),\n num_buckets=self.relative_attention_num_buckets,\n max_distance=self.relative_attention_max_distance,\n )","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.compute_bias","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.compute_bias#L447-L472","kind":"function","name":"compute_bias","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":447,"end_line":472,"context_start_line":427,"context_end_line":492,"code":" # half of the buckets are for exact increments in positions\n max_exact = num_buckets // 2\n is_small = relative_position < max_exact\n\n # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance\n relative_position_if_large = max_exact + (\n torch.log(relative_position.float() / max_exact)\n / math.log(max_distance / max_exact)\n * (num_buckets - max_exact)\n ).to(torch.long)\n relative_position_if_large = torch.min(\n relative_position_if_large,\n torch.full_like(relative_position_if_large, num_buckets - 1),\n )\n\n relative_buckets += torch.where(\n is_small, relative_position, relative_position_if_large\n )\n return relative_buckets\n\n def compute_bias(self, query_length, key_length, device=None):\n \"\"\"Compute binned relative position bias\"\"\"\n if device is None:\n device = self.relative_attention_bias.weight.device\n context_position = torch.arange(query_length, dtype=torch.long, device=device)[\n :, None\n ]\n memory_position = torch.arange(key_length, dtype=torch.long, device=device)[\n None, :\n ]\n relative_position = (\n memory_position - context_position\n ) # shape (query_length, key_length)\n relative_position_bucket = self._relative_position_bucket(\n relative_position, # shape (query_length, key_length)\n bidirectional=(not self.is_decoder),\n num_buckets=self.relative_attention_num_buckets,\n max_distance=self.relative_attention_max_distance,\n )\n values = self.relative_attention_bias(\n relative_position_bucket\n ) # shape (query_length, key_length, num_heads)\n values = values.permute([2, 0, 1]).unsqueeze(\n 0\n ) # shape (1, num_heads, query_length, key_length)\n return values\n\n def forward(\n self,\n hidden_states,\n mask=None,\n key_value_states=None,\n position_bias=None,\n past_key_value=None,\n layer_head_mask=None,\n query_length=None,\n use_cache=False,\n output_attentions=False,\n ):\n \"\"\"\n Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).\n \"\"\"\n # Input is (batch_size, seq_length, dim)\n # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)\n # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)\n batch_size, seq_length = hidden_states.shape[:2]","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.dummy_inputs","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.dummy_inputs#L843-L851","kind":"function","name":"dummy_inputs","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":843,"end_line":851,"context_start_line":823,"context_end_line":871,"code":" else:\n outputs = outputs + attention_outputs\n\n return outputs # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)\n\n\nclass T5PreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = T5Config\n load_tf_weights = load_tf_weights_in_t5\n base_model_prefix = \"transformer\"\n is_parallelizable = True\n supports_gradient_checkpointing = True\n _no_split_modules = [\"T5Block\"]\n\n @property\n def dummy_inputs(self):\n input_ids = torch.tensor(DUMMY_INPUTS)\n input_mask = torch.tensor(DUMMY_MASK)\n dummy_inputs = {\n \"decoder_input_ids\": input_ids,\n \"input_ids\": input_ids,\n \"decoder_attention_mask\": input_mask,\n }\n return dummy_inputs\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n factor = (\n self.config.initializer_factor\n ) # Used for testing weights initialization\n if isinstance(module, T5LayerNorm):\n module.weight.data.fill_(factor * 1.0)\n elif isinstance(module, (T5Model, T5ForConditionalGeneration, T5EncoderModel)):\n # Mesh TensorFlow embeddings initialization\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624\n module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)\n if hasattr(module, \"lm_head\") and not self.config.tie_word_embeddings:\n module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)\n elif isinstance(module, T5DenseActDense):\n # Mesh TensorFlow FF initialization\n # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56\n # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89\n module.wi.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._init_weights","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._init_weights#L853-L913","kind":"function","name":"_init_weights","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":853,"end_line":913,"context_start_line":833,"context_end_line":933,"code":" \"\"\"\n\n config_class = T5Config\n load_tf_weights = load_tf_weights_in_t5\n base_model_prefix = \"transformer\"\n is_parallelizable = True\n supports_gradient_checkpointing = True\n _no_split_modules = [\"T5Block\"]\n\n @property\n def dummy_inputs(self):\n input_ids = torch.tensor(DUMMY_INPUTS)\n input_mask = torch.tensor(DUMMY_MASK)\n dummy_inputs = {\n \"decoder_input_ids\": input_ids,\n \"input_ids\": input_ids,\n \"decoder_attention_mask\": input_mask,\n }\n return dummy_inputs\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n factor = (\n self.config.initializer_factor\n ) # Used for testing weights initialization\n if isinstance(module, T5LayerNorm):\n module.weight.data.fill_(factor * 1.0)\n elif isinstance(module, (T5Model, T5ForConditionalGeneration, T5EncoderModel)):\n # Mesh TensorFlow embeddings initialization\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624\n module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)\n if hasattr(module, \"lm_head\") and not self.config.tie_word_embeddings:\n module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)\n elif isinstance(module, T5DenseActDense):\n # Mesh TensorFlow FF initialization\n # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56\n # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89\n module.wi.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi, \"bias\") and module.wi.bias is not None:\n module.wi.bias.data.zero_()\n module.wo.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_ff) ** -0.5)\n )\n if hasattr(module.wo, \"bias\") and module.wo.bias is not None:\n module.wo.bias.data.zero_()\n elif isinstance(module, T5DenseGatedActDense):\n module.wi_0.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi_0, \"bias\") and module.wi_0.bias is not None:\n module.wi_0.bias.data.zero_()\n module.wi_1.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_model) ** -0.5)\n )\n if hasattr(module.wi_1, \"bias\") and module.wi_1.bias is not None:\n module.wi_1.bias.data.zero_()\n module.wo.weight.data.normal_(\n mean=0.0, std=factor * ((self.config.d_ff) ** -0.5)\n )\n if hasattr(module.wo, \"bias\") and module.wo.bias is not None:\n module.wo.bias.data.zero_()\n elif isinstance(module, T5Attention):\n # Mesh TensorFlow attention initialization to avoid scaling before softmax\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136\n d_model = self.config.d_model\n key_value_proj_dim = self.config.d_kv\n n_heads = self.config.num_heads\n module.q.weight.data.normal_(\n mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5)\n )\n module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.o.weight.data.normal_(\n mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5)\n )\n if module.has_relative_attention_bias:\n module.relative_attention_bias.weight.data.normal_(\n mean=0.0, std=factor * ((d_model) ** -0.5)\n )\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (T5Attention, T5Stack)):\n module.gradient_checkpointing = value\n\n def _shift_right(self, input_ids):\n decoder_start_token_id = self.config.decoder_start_token_id\n pad_token_id = self.config.pad_token_id\n\n assert decoder_start_token_id is not None, (\n \"self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id.\"\n \" See T5 docs for more information\"\n )\n\n # shift inputs to the right\n if is_torch_fx_proxy(input_ids):\n # Item assignment is not supported natively for proxies.\n shifted_input_ids = torch.full(\n input_ids.shape[:-1] + (1,), decoder_start_token_id\n )","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._set_gradient_checkpointing","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._set_gradient_checkpointing#L915-L917","kind":"function","name":"_set_gradient_checkpointing","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":915,"end_line":917,"context_start_line":895,"context_end_line":937,"code":" module.wo.bias.data.zero_()\n elif isinstance(module, T5Attention):\n # Mesh TensorFlow attention initialization to avoid scaling before softmax\n # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136\n d_model = self.config.d_model\n key_value_proj_dim = self.config.d_kv\n n_heads = self.config.num_heads\n module.q.weight.data.normal_(\n mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5)\n )\n module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.o.weight.data.normal_(\n mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5)\n )\n if module.has_relative_attention_bias:\n module.relative_attention_bias.weight.data.normal_(\n mean=0.0, std=factor * ((d_model) ** -0.5)\n )\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (T5Attention, T5Stack)):\n module.gradient_checkpointing = value\n\n def _shift_right(self, input_ids):\n decoder_start_token_id = self.config.decoder_start_token_id\n pad_token_id = self.config.pad_token_id\n\n assert decoder_start_token_id is not None, (\n \"self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id.\"\n \" See T5 docs for more information\"\n )\n\n # shift inputs to the right\n if is_torch_fx_proxy(input_ids):\n # Item assignment is not supported natively for proxies.\n shifted_input_ids = torch.full(\n input_ids.shape[:-1] + (1,), decoder_start_token_id\n )\n shifted_input_ids = torch.cat(\n [shifted_input_ids, input_ids[..., :-1]], dim=-1\n )\n else:","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._shift_right","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._shift_right#L919-L948","kind":"function","name":"_shift_right","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":919,"end_line":948,"context_start_line":899,"context_end_line":968,"code":" d_model = self.config.d_model\n key_value_proj_dim = self.config.d_kv\n n_heads = self.config.num_heads\n module.q.weight.data.normal_(\n mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5)\n )\n module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))\n module.o.weight.data.normal_(\n mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5)\n )\n if module.has_relative_attention_bias:\n module.relative_attention_bias.weight.data.normal_(\n mean=0.0, std=factor * ((d_model) ** -0.5)\n )\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (T5Attention, T5Stack)):\n module.gradient_checkpointing = value\n\n def _shift_right(self, input_ids):\n decoder_start_token_id = self.config.decoder_start_token_id\n pad_token_id = self.config.pad_token_id\n\n assert decoder_start_token_id is not None, (\n \"self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id.\"\n \" See T5 docs for more information\"\n )\n\n # shift inputs to the right\n if is_torch_fx_proxy(input_ids):\n # Item assignment is not supported natively for proxies.\n shifted_input_ids = torch.full(\n input_ids.shape[:-1] + (1,), decoder_start_token_id\n )\n shifted_input_ids = torch.cat(\n [shifted_input_ids, input_ids[..., :-1]], dim=-1\n )\n else:\n shifted_input_ids = input_ids.new_zeros(input_ids.shape)\n shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()\n shifted_input_ids[..., 0] = decoder_start_token_id\n\n assert (\n pad_token_id is not None\n ), \"self.model.config.pad_token_id has to be defined.\"\n # replace possible -100 values in labels by `pad_token_id`\n shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)\n\n return shifted_input_ids\n\n\nclass T5Stack(T5PreTrainedModel):\n def __init__(self, config, embed_tokens=None):\n super().__init__(config)\n\n self.embed_tokens = embed_tokens\n self.is_decoder = config.is_decoder\n\n self.block = nn.ModuleList(\n [\n T5Block(config, has_relative_attention_bias=bool(i == 0))\n for i in range(config.num_layers)\n ]\n )\n self.final_layer_norm = T5LayerNorm(\n config.d_model, eps=config.layer_norm_epsilon\n )\n self.dropout = nn.Dropout(config.dropout_rate)\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.parallelize","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.parallelize#L1986-L1994","kind":"function","name":"parallelize","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1986,"end_line":1994,"context_start_line":1966,"context_end_line":2014,"code":" r\"encoder.embed_tokens.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.deparallelize","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.deparallelize#L1997-L2002","kind":"function","name":"deparallelize","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1997,"end_line":2002,"context_start_line":1977,"context_end_line":2022,"code":"\n # Initialize weights and apply final processing\n self.post_init()\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.get_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.get_input_embeddings#L2004-L2005","kind":"function","name":"get_input_embeddings","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":2004,"end_line":2005,"context_start_line":1984,"context_end_line":2025,"code":"\n @add_start_docstrings(PARALLELIZE_DOCSTRING)\n def parallelize(self, device_map=None):\n self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.set_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.set_input_embeddings#L2007-L2009","kind":"function","name":"set_input_embeddings","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":2007,"end_line":2009,"context_start_line":1987,"context_end_line":2029,"code":" self.device_map = (\n get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))\n if device_map is None\n else device_map\n )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.get_encoder","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.get_encoder#L2011-L2012","kind":"function","name":"get_encoder","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":2011,"end_line":2012,"context_start_line":1991,"context_end_line":2032,"code":" )\n assert_device_map(self.device_map, len(self.encoder.block))\n self.encoder.parallelize(self.device_map)\n self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n output_attentions: Optional[bool] = None,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.get_decoder","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.get_decoder#L1727-L1728","kind":"function","name":"get_decoder","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1727,"end_line":1728,"context_start_line":1707,"context_end_line":1748,"code":" self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n self.decoder.set_input_embeddings(new_embeddings)\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def get_encoder(self):\n return self.encoder\n\n def get_decoder(self):\n return self.decoder\n\n @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n decoder_input_ids: Optional[torch.LongTensor] = None,\n decoder_attention_mask: Optional[torch.BoolTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n decoder_head_mask: Optional[torch.FloatTensor] = None,\n cross_attn_head_mask: Optional[torch.Tensor] = None,\n encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,\n past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n decoder_inputs_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._prune_heads","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._prune_heads#L2014-L2020","kind":"function","name":"_prune_heads","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":2014,"end_line":2020,"context_start_line":1994,"context_end_line":2040,"code":" self.model_parallel = True\n\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n\n def get_encoder(self):\n return self.encoder\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.block[layer].layer[0].SelfAttention.prune_heads(heads)\n\n @add_start_docstrings_to_model_forward(T5_ENCODER_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:\n r\"\"\"\n Returns:\n\n Example:\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.set_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.set_output_embeddings#L1718-L1719","kind":"function","name":"set_output_embeddings","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1718,"end_line":1719,"context_start_line":1698,"context_end_line":1739,"code":"\n @add_start_docstrings(DEPARALLELIZE_DOCSTRING)\n def deparallelize(self):\n self.encoder.deparallelize()\n self.decoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.decoder = self.decoder.to(\"cpu\")\n self.lm_head = self.lm_head.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n self.decoder.set_input_embeddings(new_embeddings)\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def get_encoder(self):\n return self.encoder\n\n def get_decoder(self):\n return self.decoder\n\n @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n decoder_input_ids: Optional[torch.LongTensor] = None,\n decoder_attention_mask: Optional[torch.BoolTensor] = None,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.get_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.get_output_embeddings#L1721-L1722","kind":"function","name":"get_output_embeddings","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1721,"end_line":1722,"context_start_line":1701,"context_end_line":1742,"code":" self.encoder.deparallelize()\n self.decoder.deparallelize()\n self.encoder = self.encoder.to(\"cpu\")\n self.decoder = self.decoder.to(\"cpu\")\n self.lm_head = self.lm_head.to(\"cpu\")\n self.model_parallel = False\n self.device_map = None\n torch.cuda.empty_cache()\n\n def get_input_embeddings(self):\n return self.shared\n\n def set_input_embeddings(self, new_embeddings):\n self.shared = new_embeddings\n self.encoder.set_input_embeddings(new_embeddings)\n self.decoder.set_input_embeddings(new_embeddings)\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def get_encoder(self):\n return self.encoder\n\n def get_decoder(self):\n return self.decoder\n\n @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)\n @replace_return_docstrings(\n output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: Optional[torch.LongTensor] = None,\n attention_mask: Optional[torch.FloatTensor] = None,\n decoder_input_ids: Optional[torch.LongTensor] = None,\n decoder_attention_mask: Optional[torch.BoolTensor] = None,\n head_mask: Optional[torch.FloatTensor] = None,\n decoder_head_mask: Optional[torch.FloatTensor] = None,\n cross_attn_head_mask: Optional[torch.Tensor] = None,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.prepare_inputs_for_generation","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.prepare_inputs_for_generation#L1898-L1924","kind":"function","name":"prepare_inputs_for_generation","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1898,"end_line":1924,"context_start_line":1878,"context_end_line":1944,"code":" loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))\n if reduction == \"none\":\n loss = loss.view(lm_logits.size(0), -1).sum(1)\n\n if not return_dict:\n output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs\n return ((loss,) + output) if loss is not None else output\n\n return Seq2SeqLMOutput(\n loss=loss,\n logits=lm_logits,\n past_key_values=decoder_outputs.past_key_values,\n decoder_hidden_states=decoder_outputs.hidden_states,\n decoder_attentions=decoder_outputs.attentions,\n cross_attentions=decoder_outputs.cross_attentions,\n encoder_last_hidden_state=encoder_outputs.last_hidden_state,\n encoder_hidden_states=encoder_outputs.hidden_states,\n encoder_attentions=encoder_outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self,\n input_ids,\n past=None,\n attention_mask=None,\n head_mask=None,\n decoder_head_mask=None,\n cross_attn_head_mask=None,\n use_cache=None,\n encoder_outputs=None,\n **kwargs,\n ):\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"decoder_input_ids\": input_ids,\n \"past_key_values\": past,\n \"encoder_outputs\": encoder_outputs,\n \"attention_mask\": attention_mask,\n \"head_mask\": head_mask,\n \"decoder_head_mask\": decoder_head_mask,\n \"cross_attn_head_mask\": cross_attn_head_mask,\n \"use_cache\": use_cache,\n }\n\n def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):\n return self._shift_right(labels)\n\n def _reorder_cache(self, past, beam_idx):\n # if decoder past is not included in output\n # speedy decoding is disabled and no need to reorder\n if past is None:\n logger.warning(\n \"You might want to consider setting `use_cache=True` to speed up decoding\"\n )\n return past\n\n reordered_decoder_past = ()\n for layer_past_states in past:\n # get the correct batch idx from layer past batch dim\n # batch dim of `past` is at 2nd position\n reordered_layer_past_states = ()\n for layer_past_state in layer_past_states:\n # need to set correct `past` for each of the four key / value states","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.prepare_decoder_input_ids_from_labels","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.prepare_decoder_input_ids_from_labels#L1926-L1927","kind":"function","name":"prepare_decoder_input_ids_from_labels","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1926,"end_line":1927,"context_start_line":1906,"context_end_line":1947,"code":" use_cache=None,\n encoder_outputs=None,\n **kwargs,\n ):\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"decoder_input_ids\": input_ids,\n \"past_key_values\": past,\n \"encoder_outputs\": encoder_outputs,\n \"attention_mask\": attention_mask,\n \"head_mask\": head_mask,\n \"decoder_head_mask\": decoder_head_mask,\n \"cross_attn_head_mask\": cross_attn_head_mask,\n \"use_cache\": use_cache,\n }\n\n def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):\n return self._shift_right(labels)\n\n def _reorder_cache(self, past, beam_idx):\n # if decoder past is not included in output\n # speedy decoding is disabled and no need to reorder\n if past is None:\n logger.warning(\n \"You might want to consider setting `use_cache=True` to speed up decoding\"\n )\n return past\n\n reordered_decoder_past = ()\n for layer_past_states in past:\n # get the correct batch idx from layer past batch dim\n # batch dim of `past` is at 2nd position\n reordered_layer_past_states = ()\n for layer_past_state in layer_past_states:\n # need to set correct `past` for each of the four key / value states\n reordered_layer_past_states = reordered_layer_past_states + (\n layer_past_state.index_select(\n 0, beam_idx.to(layer_past_state.device)","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5._reorder_cache","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5._reorder_cache#L1929-L1957","kind":"function","name":"_reorder_cache","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1929,"end_line":1957,"context_start_line":1909,"context_end_line":1977,"code":" ):\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"decoder_input_ids\": input_ids,\n \"past_key_values\": past,\n \"encoder_outputs\": encoder_outputs,\n \"attention_mask\": attention_mask,\n \"head_mask\": head_mask,\n \"decoder_head_mask\": decoder_head_mask,\n \"cross_attn_head_mask\": cross_attn_head_mask,\n \"use_cache\": use_cache,\n }\n\n def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):\n return self._shift_right(labels)\n\n def _reorder_cache(self, past, beam_idx):\n # if decoder past is not included in output\n # speedy decoding is disabled and no need to reorder\n if past is None:\n logger.warning(\n \"You might want to consider setting `use_cache=True` to speed up decoding\"\n )\n return past\n\n reordered_decoder_past = ()\n for layer_past_states in past:\n # get the correct batch idx from layer past batch dim\n # batch dim of `past` is at 2nd position\n reordered_layer_past_states = ()\n for layer_past_state in layer_past_states:\n # need to set correct `past` for each of the four key / value states\n reordered_layer_past_states = reordered_layer_past_states + (\n layer_past_state.index_select(\n 0, beam_idx.to(layer_past_state.device)\n ),\n )\n\n assert reordered_layer_past_states[0].shape == layer_past_states[0].shape\n assert len(reordered_layer_past_states) == len(layer_past_states)\n\n reordered_decoder_past = reordered_decoder_past + (\n reordered_layer_past_states,\n )\n return reordered_decoder_past\n\n\n@add_start_docstrings(\n \"The bare T5 Model transformer outputting encoder's raw hidden-states without any specific head on top.\",\n T5_START_DOCSTRING,\n)\nclass T5EncoderModel(T5PreTrainedModel):\n authorized_missing_keys = [\n r\"encoder.embed_tokens.weight\",\n ]\n\n def __init__(self, config: T5Config):\n super().__init__(config)\n self.shared = nn.Embedding(config.vocab_size, config.d_model)\n\n encoder_config = copy.deepcopy(config)\n encoder_config.use_cache = False\n encoder_config.is_encoder_decoder = False\n self.encoder = T5Stack(encoder_config, self.shared)\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.shape","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.shape#L508-L512","kind":"function","name":"shape","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":508,"end_line":512,"context_start_line":488,"context_end_line":532,"code":" \"\"\"\n # Input is (batch_size, seq_length, dim)\n # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)\n # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)\n batch_size, seq_length = hidden_states.shape[:2]\n\n real_seq_length = seq_length\n\n if past_key_value is not None:\n assert (\n len(past_key_value) == 2\n ), f\"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states\"\n real_seq_length += (\n past_key_value[0].shape[2] if query_length is None else query_length\n )\n\n key_length = (\n real_seq_length if key_value_states is None else key_value_states.shape[1]\n )\n\n def shape(states):\n \"\"\"projection\"\"\"\n return states.view(\n batch_size, -1, self.n_heads, self.key_value_proj_dim\n ).transpose(1, 2)\n\n def unshape(states):\n \"\"\"reshape\"\"\"\n return (\n states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)\n )\n\n def project(hidden_states, proj_layer, key_value_states, past_key_value):\n \"\"\"projects hidden states correctly to key/query states\"\"\"\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(hidden_states))\n elif past_key_value is None:\n # cross-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(key_value_states))\n\n if past_key_value is not None:\n if key_value_states is None:","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.unshape","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.unshape#L514-L518","kind":"function","name":"unshape","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":514,"end_line":518,"context_start_line":494,"context_end_line":538,"code":" real_seq_length = seq_length\n\n if past_key_value is not None:\n assert (\n len(past_key_value) == 2\n ), f\"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states\"\n real_seq_length += (\n past_key_value[0].shape[2] if query_length is None else query_length\n )\n\n key_length = (\n real_seq_length if key_value_states is None else key_value_states.shape[1]\n )\n\n def shape(states):\n \"\"\"projection\"\"\"\n return states.view(\n batch_size, -1, self.n_heads, self.key_value_proj_dim\n ).transpose(1, 2)\n\n def unshape(states):\n \"\"\"reshape\"\"\"\n return (\n states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)\n )\n\n def project(hidden_states, proj_layer, key_value_states, past_key_value):\n \"\"\"projects hidden states correctly to key/query states\"\"\"\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(hidden_states))\n elif past_key_value is None:\n # cross-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(key_value_states))\n\n if past_key_value is not None:\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, key_length, dim_per_head)\n hidden_states = torch.cat([past_key_value, hidden_states], dim=2)\n else:\n # cross-attn\n hidden_states = past_key_value","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.project","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.project#L520-L539","kind":"function","name":"project","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":520,"end_line":539,"context_start_line":500,"context_end_line":559,"code":" real_seq_length += (\n past_key_value[0].shape[2] if query_length is None else query_length\n )\n\n key_length = (\n real_seq_length if key_value_states is None else key_value_states.shape[1]\n )\n\n def shape(states):\n \"\"\"projection\"\"\"\n return states.view(\n batch_size, -1, self.n_heads, self.key_value_proj_dim\n ).transpose(1, 2)\n\n def unshape(states):\n \"\"\"reshape\"\"\"\n return (\n states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)\n )\n\n def project(hidden_states, proj_layer, key_value_states, past_key_value):\n \"\"\"projects hidden states correctly to key/query states\"\"\"\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(hidden_states))\n elif past_key_value is None:\n # cross-attn\n # (batch_size, n_heads, seq_length, dim_per_head)\n hidden_states = shape(proj_layer(key_value_states))\n\n if past_key_value is not None:\n if key_value_states is None:\n # self-attn\n # (batch_size, n_heads, key_length, dim_per_head)\n hidden_states = torch.cat([past_key_value, hidden_states], dim=2)\n else:\n # cross-attn\n hidden_states = past_key_value\n return hidden_states\n\n # get query states\n query_states = shape(\n self.q(hidden_states)\n ) # (batch_size, n_heads, seq_length, dim_per_head)\n\n # get key/value states\n key_states = project(\n hidden_states,\n self.k,\n key_value_states,\n past_key_value[0] if past_key_value is not None else None,\n )\n value_states = project(\n hidden_states,\n self.v,\n key_value_states,\n past_key_value[1] if past_key_value is not None else None,\n )\n","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.create_custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.create_custom_forward#L1192-L1196","kind":"function","name":"create_custom_forward","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1192,"end_line":1196,"context_start_line":1172,"context_end_line":1216,"code":" if encoder_decoder_position_bias is not None:\n encoder_decoder_position_bias = encoder_decoder_position_bias.to(\n hidden_states.device\n )\n if layer_head_mask is not None:\n layer_head_mask = layer_head_mask.to(hidden_states.device)\n if cross_attn_layer_head_mask is not None:\n cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(\n hidden_states.device\n )\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if self.gradient_checkpointing and self.training:\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return tuple(module(*inputs, use_cache, output_attentions))\n\n return custom_forward\n\n layer_outputs = checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n extended_attention_mask,\n position_bias,\n encoder_hidden_states,\n encoder_extended_attention_mask,\n encoder_decoder_position_bias,\n layer_head_mask,\n cross_attn_layer_head_mask,\n None, # past_key_value is always None with gradient checkpointing\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask=extended_attention_mask,\n position_bias=position_bias,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_t5.custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_t5.custom_forward#L1193-L1194","kind":"function","name":"custom_forward","path":"lavis/models/blip2_models/modeling_t5.py","language":"python","start_line":1193,"end_line":1194,"context_start_line":1173,"context_end_line":1214,"code":" encoder_decoder_position_bias = encoder_decoder_position_bias.to(\n hidden_states.device\n )\n if layer_head_mask is not None:\n layer_head_mask = layer_head_mask.to(hidden_states.device)\n if cross_attn_layer_head_mask is not None:\n cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(\n hidden_states.device\n )\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if self.gradient_checkpointing and self.training:\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return tuple(module(*inputs, use_cache, output_attentions))\n\n return custom_forward\n\n layer_outputs = checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n extended_attention_mask,\n position_bias,\n encoder_hidden_states,\n encoder_extended_attention_mask,\n encoder_decoder_position_bias,\n layer_head_mask,\n cross_attn_layer_head_mask,\n None, # past_key_value is always None with gradient checkpointing\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask=extended_attention_mask,\n position_bias=position_bias,","source_hash":"f92aca700799d11cd1c99a20e8ec93b2cffe693485acc11e5124b06454247516","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt","uri":"program://CREMA/module/lavis.models.blip2_models.modeling_opt#L1-L1113","kind":"module","name":"lavis.models.blip2_models.modeling_opt","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":1,"end_line":1113,"context_start_line":1,"context_end_line":1113,"code":"# coding=utf-8\n# Copyright 2022 The Fairseq Authors and The HuggingFace Inc. team. All rights reserved.\n#\n# Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# http://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.\n\"\"\" PyTorch OPT model.\"\"\"\nimport random\nfrom typing import List, Optional, Tuple, Union\n\nimport torch\nimport torch.utils.checkpoint\nfrom torch import nn\nfrom torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss\n\nfrom transformers.activations import ACT2FN\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPast,\n CausalLMOutputWithPast,\n)\nfrom transformers.modeling_utils import PreTrainedModel\nfrom transformers.utils import (\n add_code_sample_docstrings,\n add_start_docstrings,\n add_start_docstrings_to_model_forward,\n logging,\n replace_return_docstrings,\n)\nfrom transformers.models.opt.configuration_opt import OPTConfig\n\n\nlogger = logging.get_logger(__name__)\n\n_CHECKPOINT_FOR_DOC = \"facebook/opt-350m\"\n_CONFIG_FOR_DOC = \"OPTConfig\"\n_TOKENIZER_FOR_DOC = \"GPT2Tokenizer\"\n\n# Base model docstring\n_EXPECTED_OUTPUT_SHAPE = [1, 8, 1024]\n\n# SequenceClassification docstring\n_CHECKPOINT_FOR_SEQUENCE_CLASSIFICATION = \"ArthurZ/opt-350m-dummy-sc\"\n_SEQ_CLASS_EXPECTED_LOSS = 1.71\n_SEQ_CLASS_EXPECTED_OUTPUT = \"'LABEL_0'\"\n\n# QuestionAnswering docstring\n_QA_EXPECTED_OUTPUT = \"'a nice puppet'\"\n_QA_EXPECTED_LOSS = 7.41\n_QA_TARGET_START_INDEX = 14\n_QA_TARGET_END_INDEX = 15\n\nOPT_PRETRAINED_MODEL_ARCHIVE_LIST = [\n \"facebook/opt-125m\",\n \"facebook/opt-350m\",\n \"facebook/opt-1.3b\",\n \"facebook/opt-2.7b\",\n \"facebook/opt-6.7b\",\n \"facebook/opt-13b\",\n \"facebook/opt-30b\",\n # See all OPT models at https://huggingface.co/models?filter=opt\n]\n\n\ndef _make_causal_mask(\n input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0\n):\n \"\"\"\n Make causal mask used for bi-directional self-attention.\n \"\"\"\n bsz, tgt_len = input_ids_shape\n mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min))\n mask_cond = torch.arange(mask.size(-1))\n mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)\n mask = mask.to(dtype)\n\n if past_key_values_length > 0:\n mask = torch.cat(\n [torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1\n )\n return mask[None, None, :, :].expand(\n bsz, 1, tgt_len, tgt_len + past_key_values_length\n )\n\n\ndef _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):\n \"\"\"\n Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.\n \"\"\"\n bsz, src_len = mask.size()\n tgt_len = tgt_len if tgt_len is not None else src_len\n\n expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)\n\n inverted_mask = 1.0 - expanded_mask\n\n return inverted_mask.masked_fill(\n inverted_mask.to(torch.bool), torch.finfo(dtype).min\n )\n\n\nclass OPTLearnedPositionalEmbedding(nn.Embedding):\n \"\"\"\n This module learns positional embeddings up to a fixed maximum size.\n \"\"\"\n\n def __init__(self, num_embeddings: int, embedding_dim: int):\n # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2\n # and adjust num_embeddings appropriately. Other models don't have this hack\n self.offset = 2\n super().__init__(num_embeddings + self.offset, embedding_dim)\n\n def forward(\n self, attention_mask: torch.LongTensor, past_key_values_length: int = 0\n ):\n \"\"\"`input_ids_shape` is expected to be [bsz x seqlen].\"\"\"\n attention_mask = attention_mask.long()\n\n # create positions depending on attention_mask\n positions = (\n torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask\n ).long() - 1\n\n # cut positions if `past_key_values_length` is > 0\n positions = positions[:, past_key_values_length:]\n\n return super().forward(positions + self.offset)\n\n\nclass OPTAttention(nn.Module):\n \"\"\"Multi-headed attention from 'Attention Is All You Need' paper\"\"\"\n\n def __init__(\n self,\n embed_dim: int,\n num_heads: int,\n dropout: float = 0.0,\n is_decoder: bool = False,\n bias: bool = True,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.num_heads = num_heads\n self.dropout = dropout\n self.head_dim = embed_dim // num_heads\n\n if (self.head_dim * num_heads) != self.embed_dim:\n raise ValueError(\n f\"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}\"\n f\" and `num_heads`: {num_heads}).\"\n )\n self.scaling = self.head_dim**-0.5\n self.is_decoder = is_decoder\n\n self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n\n def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):\n return (\n tensor.view(bsz, seq_len, self.num_heads, self.head_dim)\n .transpose(1, 2)\n .contiguous()\n )\n\n def forward(\n self,\n hidden_states: torch.Tensor,\n key_value_states: Optional[torch.Tensor] = None,\n past_key_value: Optional[Tuple[torch.Tensor]] = None,\n attention_mask: Optional[torch.Tensor] = None,\n layer_head_mask: Optional[torch.Tensor] = None,\n output_attentions: bool = False,\n ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:\n \"\"\"Input shape: Batch x Time x Channel\"\"\"\n\n # if key_value_states are provided this layer is used as a cross-attention layer\n # for the decoder\n is_cross_attention = key_value_states is not None\n\n bsz, tgt_len, _ = hidden_states.size()\n\n # get query proj\n query_states = self.q_proj(hidden_states) * self.scaling\n # get key, value proj\n if is_cross_attention and past_key_value is not None:\n # reuse k,v, cross_attentions\n key_states = past_key_value[0]\n value_states = past_key_value[1]\n elif is_cross_attention:\n # cross_attentions\n key_states = self._shape(self.k_proj(key_value_states), -1, bsz)\n value_states = self._shape(self.v_proj(key_value_states), -1, bsz)\n elif past_key_value is not None:\n # reuse k, v, self_attention\n key_states = self._shape(self.k_proj(hidden_states), -1, bsz)\n value_states = self._shape(self.v_proj(hidden_states), -1, bsz)\n key_states = torch.cat([past_key_value[0], key_states], dim=2)\n value_states = torch.cat([past_key_value[1], value_states], dim=2)\n else:\n # self_attention\n key_states = self._shape(self.k_proj(hidden_states), -1, bsz)\n value_states = self._shape(self.v_proj(hidden_states), -1, bsz)\n\n if self.is_decoder:\n # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.\n # Further calls to cross_attention layer can then reuse all cross-attention\n # key/value_states (first \"if\" case)\n # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of\n # all previous decoder key/value_states. Further calls to uni-directional self-attention\n # can concat previous decoder key/value_states to current projected key/value_states (third \"elif\" case)\n # if encoder bi-directional self-attention `past_key_value` is always `None`\n past_key_value = (key_states, value_states)\n\n proj_shape = (bsz * self.num_heads, -1, self.head_dim)\n query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)\n key_states = key_states.view(*proj_shape)\n value_states = value_states.view(*proj_shape)\n\n src_len = key_states.size(1)\n attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))\n\n if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):\n raise ValueError(\n f\"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is\"\n f\" {attn_weights.size()}\"\n )\n\n if attention_mask is not None:\n if attention_mask.size() != (bsz, 1, tgt_len, src_len):\n raise ValueError(\n f\"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}\"\n )\n attn_weights = (\n attn_weights.view(bsz, self.num_heads, tgt_len, src_len)\n + attention_mask\n )\n attn_weights = torch.max(\n attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)\n )\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n # upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437\n if attn_weights.dtype == torch.float16:\n attn_weights = nn.functional.softmax(\n attn_weights, dim=-1, dtype=torch.float32\n ).to(torch.float16)\n else:\n attn_weights = nn.functional.softmax(attn_weights, dim=-1)\n\n if layer_head_mask is not None:\n if layer_head_mask.size() != (self.num_heads,):\n raise ValueError(\n f\"Head mask for a single layer should be of size {(self.num_heads,)}, but is\"\n f\" {layer_head_mask.size()}\"\n )\n attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(\n bsz, self.num_heads, tgt_len, src_len\n )\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n if output_attentions:\n # this operation is a bit awkward, but it's required to\n # make sure that attn_weights keeps its gradient.\n # In order to do so, attn_weights have to be reshaped\n # twice and have to be reused in the following\n attn_weights_reshaped = attn_weights.view(\n bsz, self.num_heads, tgt_len, src_len\n )\n attn_weights = attn_weights_reshaped.view(\n bsz * self.num_heads, tgt_len, src_len\n )\n else:\n attn_weights_reshaped = None\n\n attn_probs = nn.functional.dropout(\n attn_weights, p=self.dropout, training=self.training\n )\n\n attn_output = torch.bmm(attn_probs, value_states)\n\n if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):\n raise ValueError(\n f\"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is\"\n f\" {attn_output.size()}\"\n )\n\n attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)\n attn_output = attn_output.transpose(1, 2)\n\n # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be\n # partitioned aross GPUs when using tensor-parallelism.\n attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)\n\n attn_output = self.out_proj(attn_output)\n\n return attn_output, attn_weights_reshaped, past_key_value\n\n\nclass OPTDecoderLayer(nn.Module):\n def __init__(self, config: OPTConfig):\n super().__init__()\n self.embed_dim = config.hidden_size\n self.self_attn = OPTAttention(\n embed_dim=self.embed_dim,\n num_heads=config.num_attention_heads,\n dropout=config.attention_dropout,\n is_decoder=True,\n )\n self.do_layer_norm_before = config.do_layer_norm_before\n self.dropout = config.dropout\n self.activation_fn = ACT2FN[config.activation_function]\n\n self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)\n self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim)\n self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim)\n self.final_layer_norm = nn.LayerNorm(self.embed_dim)\n\n def forward(\n self,\n hidden_states: torch.Tensor,\n attention_mask: Optional[torch.Tensor] = None,\n layer_head_mask: Optional[torch.Tensor] = None,\n output_attentions: Optional[bool] = False,\n use_cache: Optional[bool] = False,\n past_key_value: Optional[Tuple[torch.Tensor]] = None,\n ) -> Tuple[\n torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]\n ]:\n \"\"\"\n Args:\n hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`\n attention_mask (`torch.FloatTensor`, *optional*): attention mask of size\n `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.\n layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size\n `(encoder_attention_heads,)`.\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under\n returned tensors for more detail.\n use_cache (`bool`, *optional*):\n If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding\n (see `past_key_values`).\n past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states\n \"\"\"\n\n residual = hidden_states\n\n # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention\n if self.do_layer_norm_before:\n hidden_states = self.self_attn_layer_norm(hidden_states)\n\n # Self Attention\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\n hidden_states=hidden_states,\n past_key_value=past_key_value,\n attention_mask=attention_mask,\n layer_head_mask=layer_head_mask,\n output_attentions=output_attentions,\n )\n hidden_states = nn.functional.dropout(\n hidden_states, p=self.dropout, training=self.training\n )\n hidden_states = residual + hidden_states\n\n # 350m applies layer norm AFTER attention\n if not self.do_layer_norm_before:\n hidden_states = self.self_attn_layer_norm(hidden_states)\n\n # Fully Connected\n hidden_states_shape = hidden_states.shape\n hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))\n residual = hidden_states\n\n # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention\n if self.do_layer_norm_before:\n hidden_states = self.final_layer_norm(hidden_states)\n\n hidden_states = self.fc1(hidden_states)\n hidden_states = self.activation_fn(hidden_states)\n\n hidden_states = self.fc2(hidden_states)\n hidden_states = nn.functional.dropout(\n hidden_states, p=self.dropout, training=self.training\n )\n\n hidden_states = (residual + hidden_states).view(hidden_states_shape)\n\n # 350m applies layer norm AFTER attention\n if not self.do_layer_norm_before:\n hidden_states = self.final_layer_norm(hidden_states)\n\n outputs = (hidden_states,)\n\n if output_attentions:\n outputs += (self_attn_weights,)\n\n if use_cache:\n outputs += (present_key_value,)\n\n return outputs\n\n\nOPT_START_DOCSTRING = r\"\"\"\n This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the\n library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads\n etc.)\n\n This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.\n Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage\n and behavior.\n\n Parameters:\n config ([`OPTConfig`]):\n Model configuration class with all the parameters of the model. Initializing with a config file does not\n load the weights associated with the model, only the configuration. Check out the\n [`~PreTrainedModel.from_pretrained`] method to load the model weights.\n\"\"\"\n\n\n@add_start_docstrings(\n \"The bare OPT Model outputting raw hidden-states without any specific head on top.\",\n OPT_START_DOCSTRING,\n)\nclass OPTPreTrainedModel(PreTrainedModel):\n\n config_class = OPTConfig\n base_model_prefix = \"model\"\n supports_gradient_checkpointing = True\n _no_split_modules = [\"OPTDecoderLayer\"]\n _keys_to_ignore_on_load_unexpected = [r\"decoder\\.version\"]\n\n def _init_weights(self, module):\n std = self.config.init_std\n if isinstance(module, nn.Linear):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.bias is not None:\n module.bias.data.zero_()\n elif isinstance(module, nn.Embedding):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (OPTDecoder)):\n module.gradient_checkpointing = value\n\n\nOPT_INPUTS_DOCSTRING = r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide\n it.\n\n Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n\n Indices can be obtained using [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see\n `past_key_values`).\n\n If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]\n and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more\n information on the default strategy.\n head_mask (`torch.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):\n Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:\n\n - 1 indicates the head is **not masked**,\n - 0 indicates the head is **masked**.\n\n past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cac\n# ... truncated ...","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._make_causal_mask","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._make_causal_mask#L72-L90","kind":"function","name":"_make_causal_mask","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":72,"end_line":90,"context_start_line":52,"context_end_line":110,"code":"_SEQ_CLASS_EXPECTED_OUTPUT = \"'LABEL_0'\"\n\n# QuestionAnswering docstring\n_QA_EXPECTED_OUTPUT = \"'a nice puppet'\"\n_QA_EXPECTED_LOSS = 7.41\n_QA_TARGET_START_INDEX = 14\n_QA_TARGET_END_INDEX = 15\n\nOPT_PRETRAINED_MODEL_ARCHIVE_LIST = [\n \"facebook/opt-125m\",\n \"facebook/opt-350m\",\n \"facebook/opt-1.3b\",\n \"facebook/opt-2.7b\",\n \"facebook/opt-6.7b\",\n \"facebook/opt-13b\",\n \"facebook/opt-30b\",\n # See all OPT models at https://huggingface.co/models?filter=opt\n]\n\n\ndef _make_causal_mask(\n input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0\n):\n \"\"\"\n Make causal mask used for bi-directional self-attention.\n \"\"\"\n bsz, tgt_len = input_ids_shape\n mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min))\n mask_cond = torch.arange(mask.size(-1))\n mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)\n mask = mask.to(dtype)\n\n if past_key_values_length > 0:\n mask = torch.cat(\n [torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1\n )\n return mask[None, None, :, :].expand(\n bsz, 1, tgt_len, tgt_len + past_key_values_length\n )\n\n\ndef _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):\n \"\"\"\n Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.\n \"\"\"\n bsz, src_len = mask.size()\n tgt_len = tgt_len if tgt_len is not None else src_len\n\n expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)\n\n inverted_mask = 1.0 - expanded_mask\n\n return inverted_mask.masked_fill(\n inverted_mask.to(torch.bool), torch.finfo(dtype).min\n )\n\n\nclass OPTLearnedPositionalEmbedding(nn.Embedding):\n \"\"\"","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._expand_mask","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._expand_mask#L93-L106","kind":"function","name":"_expand_mask","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":93,"end_line":106,"context_start_line":73,"context_end_line":126,"code":" input_ids_shape: torch.Size, dtype: torch.dtype, past_key_values_length: int = 0\n):\n \"\"\"\n Make causal mask used for bi-directional self-attention.\n \"\"\"\n bsz, tgt_len = input_ids_shape\n mask = torch.full((tgt_len, tgt_len), torch.tensor(torch.finfo(dtype).min))\n mask_cond = torch.arange(mask.size(-1))\n mask.masked_fill_(mask_cond < (mask_cond + 1).view(mask.size(-1), 1), 0)\n mask = mask.to(dtype)\n\n if past_key_values_length > 0:\n mask = torch.cat(\n [torch.zeros(tgt_len, past_key_values_length, dtype=dtype), mask], dim=-1\n )\n return mask[None, None, :, :].expand(\n bsz, 1, tgt_len, tgt_len + past_key_values_length\n )\n\n\ndef _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):\n \"\"\"\n Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.\n \"\"\"\n bsz, src_len = mask.size()\n tgt_len = tgt_len if tgt_len is not None else src_len\n\n expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)\n\n inverted_mask = 1.0 - expanded_mask\n\n return inverted_mask.masked_fill(\n inverted_mask.to(torch.bool), torch.finfo(dtype).min\n )\n\n\nclass OPTLearnedPositionalEmbedding(nn.Embedding):\n \"\"\"\n This module learns positional embeddings up to a fixed maximum size.\n \"\"\"\n\n def __init__(self, num_embeddings: int, embedding_dim: int):\n # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2\n # and adjust num_embeddings appropriately. Other models don't have this hack\n self.offset = 2\n super().__init__(num_embeddings + self.offset, embedding_dim)\n\n def forward(\n self, attention_mask: torch.LongTensor, past_key_values_length: int = 0\n ):\n \"\"\"`input_ids_shape` is expected to be [bsz x seqlen].\"\"\"\n attention_mask = attention_mask.long()\n\n # create positions depending on attention_mask","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTLearnedPositionalEmbedding","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTLearnedPositionalEmbedding#L109-L134","kind":"class","name":"OPTLearnedPositionalEmbedding","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":109,"end_line":134,"context_start_line":89,"context_end_line":154,"code":" bsz, 1, tgt_len, tgt_len + past_key_values_length\n )\n\n\ndef _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):\n \"\"\"\n Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.\n \"\"\"\n bsz, src_len = mask.size()\n tgt_len = tgt_len if tgt_len is not None else src_len\n\n expanded_mask = mask[:, None, None, :].expand(bsz, 1, tgt_len, src_len).to(dtype)\n\n inverted_mask = 1.0 - expanded_mask\n\n return inverted_mask.masked_fill(\n inverted_mask.to(torch.bool), torch.finfo(dtype).min\n )\n\n\nclass OPTLearnedPositionalEmbedding(nn.Embedding):\n \"\"\"\n This module learns positional embeddings up to a fixed maximum size.\n \"\"\"\n\n def __init__(self, num_embeddings: int, embedding_dim: int):\n # OPT is set up so that if padding_idx is specified then offset the embedding ids by 2\n # and adjust num_embeddings appropriately. Other models don't have this hack\n self.offset = 2\n super().__init__(num_embeddings + self.offset, embedding_dim)\n\n def forward(\n self, attention_mask: torch.LongTensor, past_key_values_length: int = 0\n ):\n \"\"\"`input_ids_shape` is expected to be [bsz x seqlen].\"\"\"\n attention_mask = attention_mask.long()\n\n # create positions depending on attention_mask\n positions = (\n torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask\n ).long() - 1\n\n # cut positions if `past_key_values_length` is > 0\n positions = positions[:, past_key_values_length:]\n\n return super().forward(positions + self.offset)\n\n\nclass OPTAttention(nn.Module):\n \"\"\"Multi-headed attention from 'Attention Is All You Need' paper\"\"\"\n\n def __init__(\n self,\n embed_dim: int,\n num_heads: int,\n dropout: float = 0.0,\n is_decoder: bool = False,\n bias: bool = True,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.num_heads = num_heads\n self.dropout = dropout\n self.head_dim = embed_dim // num_heads\n\n if (self.head_dim * num_heads) != self.embed_dim:","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTAttention","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTAttention#L137-L305","kind":"class","name":"OPTAttention","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":137,"end_line":305,"context_start_line":117,"context_end_line":325,"code":" self.offset = 2\n super().__init__(num_embeddings + self.offset, embedding_dim)\n\n def forward(\n self, attention_mask: torch.LongTensor, past_key_values_length: int = 0\n ):\n \"\"\"`input_ids_shape` is expected to be [bsz x seqlen].\"\"\"\n attention_mask = attention_mask.long()\n\n # create positions depending on attention_mask\n positions = (\n torch.cumsum(attention_mask, dim=1).type_as(attention_mask) * attention_mask\n ).long() - 1\n\n # cut positions if `past_key_values_length` is > 0\n positions = positions[:, past_key_values_length:]\n\n return super().forward(positions + self.offset)\n\n\nclass OPTAttention(nn.Module):\n \"\"\"Multi-headed attention from 'Attention Is All You Need' paper\"\"\"\n\n def __init__(\n self,\n embed_dim: int,\n num_heads: int,\n dropout: float = 0.0,\n is_decoder: bool = False,\n bias: bool = True,\n ):\n super().__init__()\n self.embed_dim = embed_dim\n self.num_heads = num_heads\n self.dropout = dropout\n self.head_dim = embed_dim // num_heads\n\n if (self.head_dim * num_heads) != self.embed_dim:\n raise ValueError(\n f\"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}\"\n f\" and `num_heads`: {num_heads}).\"\n )\n self.scaling = self.head_dim**-0.5\n self.is_decoder = is_decoder\n\n self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n\n def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):\n return (\n tensor.view(bsz, seq_len, self.num_heads, self.head_dim)\n .transpose(1, 2)\n .contiguous()\n )\n\n def forward(\n self,\n hidden_states: torch.Tensor,\n key_value_states: Optional[torch.Tensor] = None,\n past_key_value: Optional[Tuple[torch.Tensor]] = None,\n attention_mask: Optional[torch.Tensor] = None,\n layer_head_mask: Optional[torch.Tensor] = None,\n output_attentions: bool = False,\n ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:\n \"\"\"Input shape: Batch x Time x Channel\"\"\"\n\n # if key_value_states are provided this layer is used as a cross-attention layer\n # for the decoder\n is_cross_attention = key_value_states is not None\n\n bsz, tgt_len, _ = hidden_states.size()\n\n # get query proj\n query_states = self.q_proj(hidden_states) * self.scaling\n # get key, value proj\n if is_cross_attention and past_key_value is not None:\n # reuse k,v, cross_attentions\n key_states = past_key_value[0]\n value_states = past_key_value[1]\n elif is_cross_attention:\n # cross_attentions\n key_states = self._shape(self.k_proj(key_value_states), -1, bsz)\n value_states = self._shape(self.v_proj(key_value_states), -1, bsz)\n elif past_key_value is not None:\n # reuse k, v, self_attention\n key_states = self._shape(self.k_proj(hidden_states), -1, bsz)\n value_states = self._shape(self.v_proj(hidden_states), -1, bsz)\n key_states = torch.cat([past_key_value[0], key_states], dim=2)\n value_states = torch.cat([past_key_value[1], value_states], dim=2)\n else:\n # self_attention\n key_states = self._shape(self.k_proj(hidden_states), -1, bsz)\n value_states = self._shape(self.v_proj(hidden_states), -1, bsz)\n\n if self.is_decoder:\n # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.\n # Further calls to cross_attention layer can then reuse all cross-attention\n # key/value_states (first \"if\" case)\n # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of\n # all previous decoder key/value_states. Further calls to uni-directional self-attention\n # can concat previous decoder key/value_states to current projected key/value_states (third \"elif\" case)\n # if encoder bi-directional self-attention `past_key_value` is always `None`\n past_key_value = (key_states, value_states)\n\n proj_shape = (bsz * self.num_heads, -1, self.head_dim)\n query_states = self._shape(query_states, tgt_len, bsz).view(*proj_shape)\n key_states = key_states.view(*proj_shape)\n value_states = value_states.view(*proj_shape)\n\n src_len = key_states.size(1)\n attn_weights = torch.bmm(query_states, key_states.transpose(1, 2))\n\n if attn_weights.size() != (bsz * self.num_heads, tgt_len, src_len):\n raise ValueError(\n f\"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is\"\n f\" {attn_weights.size()}\"\n )\n\n if attention_mask is not None:\n if attention_mask.size() != (bsz, 1, tgt_len, src_len):\n raise ValueError(\n f\"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is {attention_mask.size()}\"\n )\n attn_weights = (\n attn_weights.view(bsz, self.num_heads, tgt_len, src_len)\n + attention_mask\n )\n attn_weights = torch.max(\n attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)\n )\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n # upcast to fp32 if the weights are in fp16. Please see https://github.com/huggingface/transformers/pull/17437\n if attn_weights.dtype == torch.float16:\n attn_weights = nn.functional.softmax(\n attn_weights, dim=-1, dtype=torch.float32\n ).to(torch.float16)\n else:\n attn_weights = nn.functional.softmax(attn_weights, dim=-1)\n\n if layer_head_mask is not None:\n if layer_head_mask.size() != (self.num_heads,):\n raise ValueError(\n f\"Head mask for a single layer should be of size {(self.num_heads,)}, but is\"\n f\" {layer_head_mask.size()}\"\n )\n attn_weights = layer_head_mask.view(1, -1, 1, 1) * attn_weights.view(\n bsz, self.num_heads, tgt_len, src_len\n )\n attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)\n\n if output_attentions:\n # this operation is a bit awkward, but it's required to\n # make sure that attn_weights keeps its gradient.\n # In order to do so, attn_weights have to be reshaped\n # twice and have to be reused in the following\n attn_weights_reshaped = attn_weights.view(\n bsz, self.num_heads, tgt_len, src_len\n )\n attn_weights = attn_weights_reshaped.view(\n bsz * self.num_heads, tgt_len, src_len\n )\n else:\n attn_weights_reshaped = None\n\n attn_probs = nn.functional.dropout(\n attn_weights, p=self.dropout, training=self.training\n )\n\n attn_output = torch.bmm(attn_probs, value_states)\n\n if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):\n raise ValueError(\n f\"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is\"\n f\" {attn_output.size()}\"\n )\n\n attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)\n attn_output = attn_output.transpose(1, 2)\n\n # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be\n # partitioned aross GPUs when using tensor-parallelism.\n attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)\n\n attn_output = self.out_proj(attn_output)\n\n return attn_output, attn_weights_reshaped, past_key_value\n\n\nclass OPTDecoderLayer(nn.Module):\n def __init__(self, config: OPTConfig):\n super().__init__()\n self.embed_dim = config.hidden_size\n self.self_attn = OPTAttention(\n embed_dim=self.embed_dim,\n num_heads=config.num_attention_heads,\n dropout=config.attention_dropout,\n is_decoder=True,\n )\n self.do_layer_norm_before = config.do_layer_norm_before\n self.dropout = config.dropout\n self.activation_fn = ACT2FN[config.activation_function]\n\n self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)\n self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim)\n self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim)\n self.final_layer_norm = nn.LayerNorm(self.embed_dim)","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTDecoderLayer","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTDecoderLayer#L308-L408","kind":"class","name":"OPTDecoderLayer","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":308,"end_line":408,"context_start_line":288,"context_end_line":428,"code":" attn_output = torch.bmm(attn_probs, value_states)\n\n if attn_output.size() != (bsz * self.num_heads, tgt_len, self.head_dim):\n raise ValueError(\n f\"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is\"\n f\" {attn_output.size()}\"\n )\n\n attn_output = attn_output.view(bsz, self.num_heads, tgt_len, self.head_dim)\n attn_output = attn_output.transpose(1, 2)\n\n # Use the `embed_dim` from the config (stored in the class) rather than `hidden_state` because `attn_output` can be\n # partitioned aross GPUs when using tensor-parallelism.\n attn_output = attn_output.reshape(bsz, tgt_len, self.embed_dim)\n\n attn_output = self.out_proj(attn_output)\n\n return attn_output, attn_weights_reshaped, past_key_value\n\n\nclass OPTDecoderLayer(nn.Module):\n def __init__(self, config: OPTConfig):\n super().__init__()\n self.embed_dim = config.hidden_size\n self.self_attn = OPTAttention(\n embed_dim=self.embed_dim,\n num_heads=config.num_attention_heads,\n dropout=config.attention_dropout,\n is_decoder=True,\n )\n self.do_layer_norm_before = config.do_layer_norm_before\n self.dropout = config.dropout\n self.activation_fn = ACT2FN[config.activation_function]\n\n self.self_attn_layer_norm = nn.LayerNorm(self.embed_dim)\n self.fc1 = nn.Linear(self.embed_dim, config.ffn_dim)\n self.fc2 = nn.Linear(config.ffn_dim, self.embed_dim)\n self.final_layer_norm = nn.LayerNorm(self.embed_dim)\n\n def forward(\n self,\n hidden_states: torch.Tensor,\n attention_mask: Optional[torch.Tensor] = None,\n layer_head_mask: Optional[torch.Tensor] = None,\n output_attentions: Optional[bool] = False,\n use_cache: Optional[bool] = False,\n past_key_value: Optional[Tuple[torch.Tensor]] = None,\n ) -> Tuple[\n torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]\n ]:\n \"\"\"\n Args:\n hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`\n attention_mask (`torch.FloatTensor`, *optional*): attention mask of size\n `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.\n layer_head_mask (`torch.FloatTensor`, *optional*): mask for attention heads in a given layer of size\n `(encoder_attention_heads,)`.\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under\n returned tensors for more detail.\n use_cache (`bool`, *optional*):\n If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding\n (see `past_key_values`).\n past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states\n \"\"\"\n\n residual = hidden_states\n\n # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention\n if self.do_layer_norm_before:\n hidden_states = self.self_attn_layer_norm(hidden_states)\n\n # Self Attention\n hidden_states, self_attn_weights, present_key_value = self.self_attn(\n hidden_states=hidden_states,\n past_key_value=past_key_value,\n attention_mask=attention_mask,\n layer_head_mask=layer_head_mask,\n output_attentions=output_attentions,\n )\n hidden_states = nn.functional.dropout(\n hidden_states, p=self.dropout, training=self.training\n )\n hidden_states = residual + hidden_states\n\n # 350m applies layer norm AFTER attention\n if not self.do_layer_norm_before:\n hidden_states = self.self_attn_layer_norm(hidden_states)\n\n # Fully Connected\n hidden_states_shape = hidden_states.shape\n hidden_states = hidden_states.reshape(-1, hidden_states.size(-1))\n residual = hidden_states\n\n # 125m, 1.7B, ..., 175B applies layer norm BEFORE attention\n if self.do_layer_norm_before:\n hidden_states = self.final_layer_norm(hidden_states)\n\n hidden_states = self.fc1(hidden_states)\n hidden_states = self.activation_fn(hidden_states)\n\n hidden_states = self.fc2(hidden_states)\n hidden_states = nn.functional.dropout(\n hidden_states, p=self.dropout, training=self.training\n )\n\n hidden_states = (residual + hidden_states).view(hidden_states_shape)\n\n # 350m applies layer norm AFTER attention\n if not self.do_layer_norm_before:\n hidden_states = self.final_layer_norm(hidden_states)\n\n outputs = (hidden_states,)\n\n if output_attentions:\n outputs += (self_attn_weights,)\n\n if use_cache:\n outputs += (present_key_value,)\n\n return outputs\n\n\nOPT_START_DOCSTRING = r\"\"\"\n This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the\n library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads\n etc.)\n\n This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.\n Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage\n and behavior.\n\n Parameters:\n config ([`OPTConfig`]):\n Model configuration class with all the parameters of the model. Initializing with a config file does not\n load the weights associated with the model, only the configuration. Check out the\n [`~PreTrainedModel.from_pretrained`] method to load the model weights.\n\"\"\"\n\n\n@add_start_docstrings(","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTPreTrainedModel","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTPreTrainedModel#L432-L453","kind":"class","name":"OPTPreTrainedModel","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":432,"end_line":453,"context_start_line":412,"context_end_line":473,"code":" This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the\n library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads\n etc.)\n\n This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.\n Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage\n and behavior.\n\n Parameters:\n config ([`OPTConfig`]):\n Model configuration class with all the parameters of the model. Initializing with a config file does not\n load the weights associated with the model, only the configuration. Check out the\n [`~PreTrainedModel.from_pretrained`] method to load the model weights.\n\"\"\"\n\n\n@add_start_docstrings(\n \"The bare OPT Model outputting raw hidden-states without any specific head on top.\",\n OPT_START_DOCSTRING,\n)\nclass OPTPreTrainedModel(PreTrainedModel):\n\n config_class = OPTConfig\n base_model_prefix = \"model\"\n supports_gradient_checkpointing = True\n _no_split_modules = [\"OPTDecoderLayer\"]\n _keys_to_ignore_on_load_unexpected = [r\"decoder\\.version\"]\n\n def _init_weights(self, module):\n std = self.config.init_std\n if isinstance(module, nn.Linear):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.bias is not None:\n module.bias.data.zero_()\n elif isinstance(module, nn.Embedding):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (OPTDecoder)):\n module.gradient_checkpointing = value\n\n\nOPT_INPUTS_DOCSTRING = r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide\n it.\n\n Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTDecoder","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTDecoder#L518-L812","kind":"class","name":"OPTDecoder","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":518,"end_line":812,"context_start_line":498,"context_end_line":832,"code":" don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all\n `decoder_input_ids` of shape `(batch_size, sequence_length)`.\n inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):\n Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This\n is useful if you want more control over how to convert `input_ids` indices into associated vectors than the\n model's internal embedding lookup matrix.\n use_cache (`bool`, *optional*):\n If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see\n `past_key_values`).\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned\n tensors for more detail.\n output_hidden_states (`bool`, *optional*):\n Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for\n more detail.\n return_dict (`bool`, *optional*):\n Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.\n\"\"\"\n\n\nclass OPTDecoder(OPTPreTrainedModel):\n \"\"\"\n Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`OPTDecoderLayer`]\n\n Args:\n config: OPTConfig\n \"\"\"\n\n def __init__(self, config: OPTConfig):\n super().__init__(config)\n self.dropout = config.dropout\n self.layerdrop = config.layerdrop\n self.padding_idx = config.pad_token_id\n self.max_target_positions = config.max_position_embeddings\n self.vocab_size = config.vocab_size\n\n self.embed_tokens = nn.Embedding(\n config.vocab_size, config.word_embed_proj_dim, self.padding_idx\n )\n self.embed_positions = OPTLearnedPositionalEmbedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n if config.word_embed_proj_dim != config.hidden_size:\n self.project_out = nn.Linear(\n config.hidden_size, config.word_embed_proj_dim, bias=False\n )\n else:\n self.project_out = None\n\n if config.word_embed_proj_dim != config.hidden_size:\n self.project_in = nn.Linear(\n config.word_embed_proj_dim, config.hidden_size, bias=False\n )\n else:\n self.project_in = None\n\n # Note that the only purpose of `config._remove_final_layer_norm` is to keep backward compatibility\n # with checkpoints that have been fine-tuned before transformers v4.20.1\n # see https://github.com/facebookresearch/metaseq/pull/164\n if config.do_layer_norm_before and not config._remove_final_layer_norm:\n self.final_layer_norm = nn.LayerNorm(config.hidden_size)\n else:\n self.final_layer_norm = None\n\n self.layers = nn.ModuleList(\n [OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)]\n )\n\n self.gradient_checkpointing = False\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.embed_tokens\n\n def set_input_embeddings(self, value):\n self.embed_tokens = value\n\n # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask\n def _prepare_decoder_attention_mask(\n self, attention_mask, input_shape, inputs_embeds, past_key_values_length\n ):\n # create causal mask\n # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]\n combined_attention_mask = None\n if input_shape[-1] > 1:\n combined_attention_mask = _make_causal_mask(\n input_shape,\n inputs_embeds.dtype,\n past_key_values_length=past_key_values_length,\n ).to(inputs_embeds.device)\n\n if attention_mask is not None:\n # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]\n expanded_attn_mask = _expand_mask(\n attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]\n ).to(inputs_embeds.device)\n combined_attention_mask = (\n expanded_attn_mask\n if combined_attention_mask is None\n else expanded_attn_mask + combined_attention_mask\n )\n\n return combined_attention_mask\n\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple, BaseModelOutputWithPast]:\n r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you\n provide it.\n\n Indices can be obtained using [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):\n Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:\n\n - 1 indicates the head is **not masked**,\n - 0 indicates the head is **masked**.\n\n past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):\n Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of\n shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of\n\n Contains pre-computed hidden-states (key and values in the self-attention blocks and in the\n cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.\n\n If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those\n that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of\n all `decoder_input_ids` of shape `(batch_size, sequence_length)`.\n\n inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):\n Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.\n This is useful if you want more control over how to convert `input_ids` indices into associated vectors\n than the model's internal embedding lookup matrix.\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under\n returned tensors for more detail.\n output_hidden_states (`bool`, *optional*):\n Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors\n for more detail.\n return_dict (`bool`, *optional*):\n Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.\n \"\"\"\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # retrieve input_ids and inputs_embeds\n if input_ids is not None and inputs_embeds is not None:\n raise ValueError(\n \"You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time\"\n )\n elif input_ids is not None:\n input_shape = input_ids.size()\n input_ids = input_ids.view(-1, input_shape[-1])\n elif inputs_embeds is not None:\n input_shape = inputs_embeds.size()[:-1]\n else:\n raise ValueError(\n \"You have to specify either decoder_input_ids or decoder_inputs_embeds\"\n )\n\n past_key_values_length = (\n past_key_values[0][0].shape[2] if past_key_values is not None else 0\n )\n\n if inputs_embeds is None:\n inputs_embeds = self.embed_tokens(input_ids)\n\n if query_embeds is not None:\n inputs_embeds = torch.cat([query_embeds, inputs_embeds], dim=1)\n input_shape = inputs_embeds.size()[:-1]\n\n # embed positions\n if attention_mask is None:\n attention_mask = torch.ones(\n inputs_embeds.shape[:2], dtype=torch.bool, device=inputs_embeds.device\n )\n pos_embeds = self.embed_positions(attention_mask, past_key_values_length)\n\n attention_mask = self._prepare_decoder_attention_mask(\n attention_mask, input_shape, inputs_embeds, past_key_values_length\n )\n\n if self.project_in is not None:\n inputs_embeds = self.project_in(inputs_embeds)\n\n hidden_states = inputs_embeds + pos_embeds\n\n # decoder layers\n all_hidden_states = () if output_hidden_states else None\n all_self_attns = () if output_attentions else None\n next_decoder_cache = () if use_cache else None\n\n # check if head_mask has a correct number of layers specified if desired\n for attn_mask, mask_name in zip([head_mask], [\"head_mask\"]):\n if attn_mask is not None:\n if attn_mask.size()[0] != (len(self.layers)):\n raise ValueError(\n f\"The `{mask_name}` should be specified for {len(self.layers)} layers, but it is for\"\n f\" {head_mask.size()[0]}.\"\n )\n\n for idx, decoder_layer in enumerate(self.layers):\n # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)\n if output_hidden_states:\n all_hidden_states += (hidden_states,)\n\n dropout_probability = random.uniform(0, 1)\n if self.training and (dropout_probability < self.layerdrop):\n continue\n\n past_key_value = (\n past_key_values[idx] if past_key_values is not None else None\n )\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n # None for past_key_value\n return module(*inputs, output_attentions, None)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(decoder_layer),\n hidden_states,\n attention_mask,\n head_mask[idx] if head_mask is not None else None,\n None,\n )\n else:\n\n layer_outputs = decoder_layer(\n hidden_states,\n attention_mask=attention_mask,\n layer_head_mask=(head_mask[idx] if head_mask is not None else None),\n past_key_value=past_key_value,\n output_attentions=output_attentions,\n use_cache=use_cache,\n )\n\n hidden_states = layer_outputs[0]\n\n if use_cache:\n next_decoder_cache += (layer_outputs[2 if output_attentions else 1],)\n\n if output_attentions:\n all_self_attns += (layer_outputs[1],)\n\n if self.final_layer_norm is not None:\n hidden_states = self.final_layer_norm(hidden_states)\n\n if self.project_out is not None:\n hidden_states = self.project_out(hidden_states)\n\n # add hidden states from the last decoder layer\n if output_hidden_states:\n all_hidden_states += (hidden_states,)\n\n next_cache = next_decoder_cache if use_cache else None\n if not return_dict:\n return tuple(\n v\n for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]\n if v is not None\n )\n return BaseModelOutputWithPast(\n last_hidden_state=hidden_states,\n past_key_values=next_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attns,\n )\n\n\n@add_start_docstrings(\n \"The bare OPT Model outputting raw hidden-states without any specific head on top.\",\n OPT_START_DOCSTRING,\n)\nclass OPTModel(OPTPreTrainedModel):\n def __init__(self, config: OPTConfig):\n super().__init__(config)\n self.decoder = OPTDecoder(config)\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.decoder.embed_tokens = value\n\n def get_decoder(self):","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTModel","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTModel#L819-L894","kind":"class","name":"OPTModel","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":819,"end_line":894,"context_start_line":799,"context_end_line":914,"code":"\n next_cache = next_decoder_cache if use_cache else None\n if not return_dict:\n return tuple(\n v\n for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]\n if v is not None\n )\n return BaseModelOutputWithPast(\n last_hidden_state=hidden_states,\n past_key_values=next_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attns,\n )\n\n\n@add_start_docstrings(\n \"The bare OPT Model outputting raw hidden-states without any specific head on top.\",\n OPT_START_DOCSTRING,\n)\nclass OPTModel(OPTPreTrainedModel):\n def __init__(self, config: OPTConfig):\n super().__init__(config)\n self.decoder = OPTDecoder(config)\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.decoder.embed_tokens = value\n\n def get_decoder(self):\n return self.decoder\n\n @add_start_docstrings_to_model_forward(OPT_INPUTS_DOCSTRING)\n @add_code_sample_docstrings(\n processor_class=_TOKENIZER_FOR_DOC,\n checkpoint=_CHECKPOINT_FOR_DOC,\n output_type=BaseModelOutputWithPast,\n config_class=_CONFIG_FOR_DOC,\n expected_output=_EXPECTED_OUTPUT_SHAPE,\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple, BaseModelOutputWithPast]:\n\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)\n decoder_outputs = self.decoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n head_mask=head_mask,\n past_key_values=past_key_values,\n inputs_embeds=inputs_embeds,\n query_embeds=query_embeds,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n if not return_dict:\n return decoder_outputs\n\n return BaseModelOutputWithPast(\n last_hidden_state=decoder_outputs.last_hidden_state,\n past_key_values=decoder_outputs.past_key_values,\n hidden_states=decoder_outputs.hidden_states,\n attentions=decoder_outputs.attentions,\n )\n\n\nclass OPTForCausalLM(OPTPreTrainedModel):\n _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.OPTForCausalLM","uri":"program://CREMA/class/lavis.models.blip2_models.modeling_opt.OPTForCausalLM#L897-L1113","kind":"class","name":"OPTForCausalLM","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":897,"end_line":1113,"context_start_line":877,"context_end_line":1113,"code":" past_key_values=past_key_values,\n inputs_embeds=inputs_embeds,\n query_embeds=query_embeds,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n if not return_dict:\n return decoder_outputs\n\n return BaseModelOutputWithPast(\n last_hidden_state=decoder_outputs.last_hidden_state,\n past_key_values=decoder_outputs.past_key_values,\n hidden_states=decoder_outputs.hidden_states,\n attentions=decoder_outputs.attentions,\n )\n\n\nclass OPTForCausalLM(OPTPreTrainedModel):\n _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n reduction: Optional[str] = \"mean\",\n ) -> Union[Tuple, CausalLMOutputWithPast]:\n r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you\n provide it.\n\n Indices can be obtained using [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):\n Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:\n\n - 1 indicates the head is **not masked**,\n - 0 indicates the head is **masked**.\n\n past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):\n Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of\n shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of\n shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional\n tensors are only required when the model is used as a decoder in a Sequence to Sequence model.\n\n Contains pre-computed hidden-states (key and values in the self-attention blocks and in the\n cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.\n\n If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those\n that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of\n all `decoder_input_ids` of shape `(batch_size, sequence_length)`.\n inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):\n Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.\n This is useful if you want more control over how to convert `input_ids` indices into associated vectors\n than the model's internal embedding lookup matrix.\n labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):\n Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,\n config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored\n (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.\n use_cache (`bool`, *optional*):\n If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding\n (see `past_key_values`).\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under\n returned tensors for more detail.\n output_hidden_states (`bool`, *optional*):\n Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors\n for more detail.\n return_dict (`bool`, *optional*):\n Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.\n\n Returns:\n\n Example:\n\n ```python\n >>> from transformers import GPT2Tokenizer, OPTForCausalLM\n\n >>> model = OPTForCausalLM.from_pretrained(\"facebook/opt-350m\")\n >>> tokenizer = GPT2Tokenizer.from_pretrained(\"facebook/opt-350m\")\n\n >>> prompt = \"Hey, are you consciours? Can you talk to me?\"\n >>> inputs = tokenizer(prompt, return_tensors=\"pt\")\n\n >>> # Generate\n >>> generate_ids = model.generate(inputs.input_ids, max_length=30)\n >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]\n \"Hey, are you consciours? Can you talk to me?\\nI'm not consciours, but I can talk to you.\"\n ```\"\"\"\n\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)\n outputs = self.model.decoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n head_mask=head_mask,\n past_key_values=past_key_values,\n inputs_embeds=inputs_embeds,\n query_embeds=query_embeds,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n logits = self.lm_head(outputs[0]).contiguous()\n\n loss = None\n if labels is not None:\n logits = logits[:, -labels.size(1) :, :]\n\n # Shift so that tokens < n predict n\n shift_logits = logits[..., :-1, :].contiguous()\n shift_labels = labels[..., 1:].contiguous()\n # Flatten the tokens\n loss_fct = CrossEntropyLoss(reduction=reduction)\n loss = loss_fct(\n shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1)\n )\n if reduction == \"none\":\n loss = loss.view(shift_logits.size(0), -1).sum(1)\n\n if not return_dict:\n output = (logits,) + outputs[1:]\n return (loss,) + output if loss is not None else output\n\n return CausalLMOutputWithPast(\n loss=loss,\n logits=logits,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self,\n input_ids=None,\n query_embeds=None,\n past=None,\n attention_mask=None,\n use_cache=None,\n **kwargs,\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n if input_ids is not None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n if past:\n input_ids = input_ids[:, -1:]\n query_embeds = None\n # first step, decoder_cached_states are empty\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"use_cache\": use_cache,\n }\n\n @staticmethod\n def _reorder_cache(past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.__init__#L900-L910","kind":"function","name":"__init__","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":900,"end_line":910,"context_start_line":880,"context_end_line":930,"code":" use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n if not return_dict:\n return decoder_outputs\n\n return BaseModelOutputWithPast(\n last_hidden_state=decoder_outputs.last_hidden_state,\n past_key_values=decoder_outputs.past_key_values,\n hidden_states=decoder_outputs.hidden_states,\n attentions=decoder_outputs.attentions,\n )\n\n\nclass OPTForCausalLM(OPTPreTrainedModel):\n _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.forward#L933-L1077","kind":"function","name":"forward","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":933,"end_line":1077,"context_start_line":913,"context_end_line":1097,"code":" return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n reduction: Optional[str] = \"mean\",\n ) -> Union[Tuple, CausalLMOutputWithPast]:\n r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you\n provide it.\n\n Indices can be obtained using [`OPTTokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n head_mask (`torch.Tensor` of shape `(num_hidden_layers, num_attention_heads)`, *optional*):\n Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:\n\n - 1 indicates the head is **not masked**,\n - 0 indicates the head is **masked**.\n\n past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):\n Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of\n shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of\n shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. The two additional\n tensors are only required when the model is used as a decoder in a Sequence to Sequence model.\n\n Contains pre-computed hidden-states (key and values in the self-attention blocks and in the\n cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.\n\n If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those\n that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of\n all `decoder_input_ids` of shape `(batch_size, sequence_length)`.\n inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):\n Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.\n This is useful if you want more control over how to convert `input_ids` indices into associated vectors\n than the model's internal embedding lookup matrix.\n labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):\n Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,\n config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored\n (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.\n use_cache (`bool`, *optional*):\n If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding\n (see `past_key_values`).\n output_attentions (`bool`, *optional*):\n Whether or not to return the attentions tensors of all attention layers. See `attentions` under\n returned tensors for more detail.\n output_hidden_states (`bool`, *optional*):\n Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors\n for more detail.\n return_dict (`bool`, *optional*):\n Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.\n\n Returns:\n\n Example:\n\n ```python\n >>> from transformers import GPT2Tokenizer, OPTForCausalLM\n\n >>> model = OPTForCausalLM.from_pretrained(\"facebook/opt-350m\")\n >>> tokenizer = GPT2Tokenizer.from_pretrained(\"facebook/opt-350m\")\n\n >>> prompt = \"Hey, are you consciours? Can you talk to me?\"\n >>> inputs = tokenizer(prompt, return_tensors=\"pt\")\n\n >>> # Generate\n >>> generate_ids = model.generate(inputs.input_ids, max_length=30)\n >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]\n \"Hey, are you consciours? Can you talk to me?\\nI'm not consciours, but I can talk to you.\"\n ```\"\"\"\n\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)\n outputs = self.model.decoder(\n input_ids=input_ids,\n attention_mask=attention_mask,\n head_mask=head_mask,\n past_key_values=past_key_values,\n inputs_embeds=inputs_embeds,\n query_embeds=query_embeds,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n\n logits = self.lm_head(outputs[0]).contiguous()\n\n loss = None\n if labels is not None:\n logits = logits[:, -labels.size(1) :, :]\n\n # Shift so that tokens < n predict n\n shift_logits = logits[..., :-1, :].contiguous()\n shift_labels = labels[..., 1:].contiguous()\n # Flatten the tokens\n loss_fct = CrossEntropyLoss(reduction=reduction)\n loss = loss_fct(\n shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1)\n )\n if reduction == \"none\":\n loss = loss.view(shift_logits.size(0), -1).sum(1)\n\n if not return_dict:\n output = (logits,) + outputs[1:]\n return (loss,) + output if loss is not None else output\n\n return CausalLMOutputWithPast(\n loss=loss,\n logits=logits,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self,\n input_ids=None,\n query_embeds=None,\n past=None,\n attention_mask=None,\n use_cache=None,\n **kwargs,\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n if input_ids is not None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n if past:\n input_ids = input_ids[:, -1:]\n query_embeds = None\n # first step, decoder_cached_states are empty\n return {\n \"input_ids\": input_ids,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._shape","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._shape#L167-L172","kind":"function","name":"_shape","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":167,"end_line":172,"context_start_line":147,"context_end_line":192,"code":" ):\n super().__init__()\n self.embed_dim = embed_dim\n self.num_heads = num_heads\n self.dropout = dropout\n self.head_dim = embed_dim // num_heads\n\n if (self.head_dim * num_heads) != self.embed_dim:\n raise ValueError(\n f\"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}\"\n f\" and `num_heads`: {num_heads}).\"\n )\n self.scaling = self.head_dim**-0.5\n self.is_decoder = is_decoder\n\n self.k_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.v_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.q_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias)\n\n def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):\n return (\n tensor.view(bsz, seq_len, self.num_heads, self.head_dim)\n .transpose(1, 2)\n .contiguous()\n )\n\n def forward(\n self,\n hidden_states: torch.Tensor,\n key_value_states: Optional[torch.Tensor] = None,\n past_key_value: Optional[Tuple[torch.Tensor]] = None,\n attention_mask: Optional[torch.Tensor] = None,\n layer_head_mask: Optional[torch.Tensor] = None,\n output_attentions: bool = False,\n ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:\n \"\"\"Input shape: Batch x Time x Channel\"\"\"\n\n # if key_value_states are provided this layer is used as a cross-attention layer\n # for the decoder\n is_cross_attention = key_value_states is not None\n\n bsz, tgt_len, _ = hidden_states.size()\n\n # get query proj\n query_states = self.q_proj(hidden_states) * self.scaling","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._init_weights","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._init_weights#L440-L449","kind":"function","name":"_init_weights","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":440,"end_line":449,"context_start_line":420,"context_end_line":469,"code":" Parameters:\n config ([`OPTConfig`]):\n Model configuration class with all the parameters of the model. Initializing with a config file does not\n load the weights associated with the model, only the configuration. Check out the\n [`~PreTrainedModel.from_pretrained`] method to load the model weights.\n\"\"\"\n\n\n@add_start_docstrings(\n \"The bare OPT Model outputting raw hidden-states without any specific head on top.\",\n OPT_START_DOCSTRING,\n)\nclass OPTPreTrainedModel(PreTrainedModel):\n\n config_class = OPTConfig\n base_model_prefix = \"model\"\n supports_gradient_checkpointing = True\n _no_split_modules = [\"OPTDecoderLayer\"]\n _keys_to_ignore_on_load_unexpected = [r\"decoder\\.version\"]\n\n def _init_weights(self, module):\n std = self.config.init_std\n if isinstance(module, nn.Linear):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.bias is not None:\n module.bias.data.zero_()\n elif isinstance(module, nn.Embedding):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (OPTDecoder)):\n module.gradient_checkpointing = value\n\n\nOPT_INPUTS_DOCSTRING = r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide\n it.\n\n Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._set_gradient_checkpointing","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._set_gradient_checkpointing#L451-L453","kind":"function","name":"_set_gradient_checkpointing","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":451,"end_line":453,"context_start_line":431,"context_end_line":473,"code":")\nclass OPTPreTrainedModel(PreTrainedModel):\n\n config_class = OPTConfig\n base_model_prefix = \"model\"\n supports_gradient_checkpointing = True\n _no_split_modules = [\"OPTDecoderLayer\"]\n _keys_to_ignore_on_load_unexpected = [r\"decoder\\.version\"]\n\n def _init_weights(self, module):\n std = self.config.init_std\n if isinstance(module, nn.Linear):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.bias is not None:\n module.bias.data.zero_()\n elif isinstance(module, nn.Embedding):\n module.weight.data.normal_(mean=0.0, std=std)\n if module.padding_idx is not None:\n module.weight.data[module.padding_idx].zero_()\n\n def _set_gradient_checkpointing(self, module, value=False):\n if isinstance(module, (OPTDecoder)):\n module.gradient_checkpointing = value\n\n\nOPT_INPUTS_DOCSTRING = r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide\n it.\n\n Indices can be obtained using [`GPT2Tokenizer`]. See [`PreTrainedTokenizer.encode`] and\n [`PreTrainedTokenizer.__call__`] for details.\n\n [What are input IDs?](../glossary#input-ids)\n attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):\n Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:\n\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n\n [What are attention masks?](../glossary#attention-mask)\n","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.get_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.get_input_embeddings#L912-L913","kind":"function","name":"get_input_embeddings","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":912,"end_line":913,"context_start_line":892,"context_end_line":933,"code":" hidden_states=decoder_outputs.hidden_states,\n attentions=decoder_outputs.attentions,\n )\n\n\nclass OPTForCausalLM(OPTPreTrainedModel):\n _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.set_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.set_input_embeddings#L915-L916","kind":"function","name":"set_input_embeddings","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":915,"end_line":916,"context_start_line":895,"context_end_line":936,"code":"\n\nclass OPTForCausalLM(OPTPreTrainedModel):\n _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._prepare_decoder_attention_mask","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._prepare_decoder_attention_mask#L578-L602","kind":"function","name":"_prepare_decoder_attention_mask","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":578,"end_line":602,"context_start_line":558,"context_end_line":622,"code":" if config.do_layer_norm_before and not config._remove_final_layer_norm:\n self.final_layer_norm = nn.LayerNorm(config.hidden_size)\n else:\n self.final_layer_norm = None\n\n self.layers = nn.ModuleList(\n [OPTDecoderLayer(config) for _ in range(config.num_hidden_layers)]\n )\n\n self.gradient_checkpointing = False\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.embed_tokens\n\n def set_input_embeddings(self, value):\n self.embed_tokens = value\n\n # Copied from transformers.models.bart.modeling_bart.BartDecoder._prepare_decoder_attention_mask\n def _prepare_decoder_attention_mask(\n self, attention_mask, input_shape, inputs_embeds, past_key_values_length\n ):\n # create causal mask\n # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]\n combined_attention_mask = None\n if input_shape[-1] > 1:\n combined_attention_mask = _make_causal_mask(\n input_shape,\n inputs_embeds.dtype,\n past_key_values_length=past_key_values_length,\n ).to(inputs_embeds.device)\n\n if attention_mask is not None:\n # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]\n expanded_attn_mask = _expand_mask(\n attention_mask, inputs_embeds.dtype, tgt_len=input_shape[-1]\n ).to(inputs_embeds.device)\n combined_attention_mask = (\n expanded_attn_mask\n if combined_attention_mask is None\n else expanded_attn_mask + combined_attention_mask\n )\n\n return combined_attention_mask\n\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n ) -> Union[Tuple, BaseModelOutputWithPast]:\n r\"\"\"\n Args:\n input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):\n Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you\n provide it.\n","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.get_decoder","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.get_decoder#L927-L928","kind":"function","name":"get_decoder","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":927,"end_line":928,"context_start_line":907,"context_end_line":948,"code":" )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,\n reduction: Optional[str] = \"mean\",\n ) -> Union[Tuple, CausalLMOutputWithPast]:\n r\"\"\"","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.get_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.get_output_embeddings#L918-L919","kind":"function","name":"get_output_embeddings","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":918,"end_line":919,"context_start_line":898,"context_end_line":939,"code":" _keys_to_ignore_on_load_missing = [r\"lm_head.weight\"]\n\n def __init__(self, config):\n super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.set_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.set_output_embeddings#L921-L922","kind":"function","name":"set_output_embeddings","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":921,"end_line":922,"context_start_line":901,"context_end_line":942,"code":" super().__init__(config)\n self.model = OPTModel(config)\n\n # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.set_decoder","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.set_decoder#L924-L925","kind":"function","name":"set_decoder","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":924,"end_line":925,"context_start_line":904,"context_end_line":945,"code":" # the lm_head weight is automatically tied to the embed tokens weight\n self.lm_head = nn.Linear(\n config.word_embed_proj_dim, config.vocab_size, bias=False\n )\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def get_input_embeddings(self):\n return self.model.decoder.embed_tokens\n\n def set_input_embeddings(self, value):\n self.model.decoder.embed_tokens = value\n\n def get_output_embeddings(self):\n return self.lm_head\n\n def set_output_embeddings(self, new_embeddings):\n self.lm_head = new_embeddings\n\n def set_decoder(self, decoder):\n self.model.decoder = decoder\n\n def get_decoder(self):\n return self.model.decoder\n\n @replace_return_docstrings(\n output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC\n )\n def forward(\n self,\n input_ids: torch.LongTensor = None,\n attention_mask: Optional[torch.Tensor] = None,\n head_mask: Optional[torch.Tensor] = None,\n past_key_values: Optional[List[torch.FloatTensor]] = None,\n inputs_embeds: Optional[torch.FloatTensor] = None,\n query_embeds: Optional[torch.FloatTensor] = None,\n labels: Optional[torch.LongTensor] = None,\n use_cache: Optional[bool] = None,\n output_attentions: Optional[bool] = None,\n output_hidden_states: Optional[bool] = None,\n return_dict: Optional[bool] = None,","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.prepare_inputs_for_generation","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.prepare_inputs_for_generation#L1079-L1102","kind":"function","name":"prepare_inputs_for_generation","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":1079,"end_line":1102,"context_start_line":1059,"context_end_line":1113,"code":" # Flatten the tokens\n loss_fct = CrossEntropyLoss(reduction=reduction)\n loss = loss_fct(\n shift_logits.view(-1, self.config.vocab_size), shift_labels.view(-1)\n )\n if reduction == \"none\":\n loss = loss.view(shift_logits.size(0), -1).sum(1)\n\n if not return_dict:\n output = (logits,) + outputs[1:]\n return (loss,) + output if loss is not None else output\n\n return CausalLMOutputWithPast(\n loss=loss,\n logits=logits,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )\n\n def prepare_inputs_for_generation(\n self,\n input_ids=None,\n query_embeds=None,\n past=None,\n attention_mask=None,\n use_cache=None,\n **kwargs,\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n if input_ids is not None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n if past:\n input_ids = input_ids[:, -1:]\n query_embeds = None\n # first step, decoder_cached_states are empty\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"use_cache\": use_cache,\n }\n\n @staticmethod\n def _reorder_cache(past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt._reorder_cache","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt._reorder_cache#L1105-L1113","kind":"function","name":"_reorder_cache","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":1105,"end_line":1113,"context_start_line":1085,"context_end_line":1113,"code":" use_cache=None,\n **kwargs,\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n if input_ids is not None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n if past:\n input_ids = input_ids[:, -1:]\n query_embeds = None\n # first step, decoder_cached_states are empty\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"use_cache\": use_cache,\n }\n\n @staticmethod\n def _reorder_cache(past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.create_custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.create_custom_forward#L757-L762","kind":"function","name":"create_custom_forward","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":757,"end_line":762,"context_start_line":737,"context_end_line":782,"code":" # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)\n if output_hidden_states:\n all_hidden_states += (hidden_states,)\n\n dropout_probability = random.uniform(0, 1)\n if self.training and (dropout_probability < self.layerdrop):\n continue\n\n past_key_value = (\n past_key_values[idx] if past_key_values is not None else None\n )\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n # None for past_key_value\n return module(*inputs, output_attentions, None)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(decoder_layer),\n hidden_states,\n attention_mask,\n head_mask[idx] if head_mask is not None else None,\n None,\n )\n else:\n\n layer_outputs = decoder_layer(\n hidden_states,\n attention_mask=attention_mask,\n layer_head_mask=(head_mask[idx] if head_mask is not None else None),\n past_key_value=past_key_value,\n output_attentions=output_attentions,\n use_cache=use_cache,\n )\n\n hidden_states = layer_outputs[0]","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.modeling_opt.custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.modeling_opt.custom_forward#L758-L760","kind":"function","name":"custom_forward","path":"lavis/models/blip2_models/modeling_opt.py","language":"python","start_line":758,"end_line":760,"context_start_line":738,"context_end_line":780,"code":" if output_hidden_states:\n all_hidden_states += (hidden_states,)\n\n dropout_probability = random.uniform(0, 1)\n if self.training and (dropout_probability < self.layerdrop):\n continue\n\n past_key_value = (\n past_key_values[idx] if past_key_values is not None else None\n )\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warning(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n # None for past_key_value\n return module(*inputs, output_attentions, None)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(decoder_layer),\n hidden_states,\n attention_mask,\n head_mask[idx] if head_mask is not None else None,\n None,\n )\n else:\n\n layer_outputs = decoder_layer(\n hidden_states,\n attention_mask=attention_mask,\n layer_head_mask=(head_mask[idx] if head_mask is not None else None),\n past_key_value=past_key_value,\n output_attentions=output_attentions,\n use_cache=use_cache,\n )","source_hash":"324d5206148e98976196abb6f92dccfe22e0f28ea91d94445282af6eb511d723","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_image_text_matching","uri":"program://CREMA/module/lavis.models.blip2_models.blip2_image_text_matching#L1-L111","kind":"module","name":"lavis.models.blip2_models.blip2_image_text_matching","path":"lavis/models/blip2_models/blip2_image_text_matching.py","language":"python","start_line":1,"end_line":111,"context_start_line":1,"context_end_line":111,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2_qformer import Blip2Qformer\n\n\n@registry.register_model(\"blip2_image_text_matching\")\nclass Blip2ITM(Blip2Qformer):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_image_text_matching\", \"pretrained\")\n >>> model = load_model(\"blip2_image_text_matching\", \"coco\")\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n if match_head == \"itm\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n output_itm = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(itm_embeddings)\n itm_logit = itm_logit.mean(dim=1)\n\n return itm_logit\n\n elif match_head == \"itc\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sims = torch.bmm(image_feats, text_feat.unsqueeze(-1))\n sim, _ = torch.max(sims, dim=1)\n\n return sim","source_hash":"cc02551220e31d1bdae6155fced7dbc595f6948d5bbf135a9f596baea4094a26","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_image_text_matching.Blip2ITM","uri":"program://CREMA/class/lavis.models.blip2_models.blip2_image_text_matching.Blip2ITM#L15-L111","kind":"class","name":"Blip2ITM","path":"lavis/models/blip2_models/blip2_image_text_matching.py","language":"python","start_line":15,"end_line":111,"context_start_line":1,"context_end_line":111,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2_qformer import Blip2Qformer\n\n\n@registry.register_model(\"blip2_image_text_matching\")\nclass Blip2ITM(Blip2Qformer):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_image_text_matching\", \"pretrained\")\n >>> model = load_model(\"blip2_image_text_matching\", \"coco\")\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n if match_head == \"itm\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n output_itm = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(itm_embeddings)\n itm_logit = itm_logit.mean(dim=1)\n\n return itm_logit\n\n elif match_head == \"itc\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sims = torch.bmm(image_feats, text_feat.unsqueeze(-1))\n sim, _ = torch.max(sims, dim=1)\n\n return sim","source_hash":"cc02551220e31d1bdae6155fced7dbc595f6948d5bbf135a9f596baea4094a26","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_image_text_matching.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_image_text_matching.__init__#L27-L47","kind":"function","name":"__init__","path":"lavis/models/blip2_models/blip2_image_text_matching.py","language":"python","start_line":27,"end_line":47,"context_start_line":7,"context_end_line":67,"code":"\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2_qformer import Blip2Qformer\n\n\n@registry.register_model(\"blip2_image_text_matching\")\nclass Blip2ITM(Blip2Qformer):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_image_text_matching\", \"pretrained\")\n >>> model = load_model(\"blip2_image_text_matching\", \"coco\")\n \"\"\"\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n if match_head == \"itm\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)","source_hash":"cc02551220e31d1bdae6155fced7dbc595f6948d5bbf135a9f596baea4094a26","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_image_text_matching.forward","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_image_text_matching.forward#L49-L111","kind":"function","name":"forward","path":"lavis/models/blip2_models/blip2_image_text_matching.py","language":"python","start_line":49,"end_line":111,"context_start_line":29,"context_end_line":111,"code":" img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n if match_head == \"itm\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n output_itm = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(itm_embeddings)\n itm_logit = itm_logit.mean(dim=1)\n\n return itm_logit\n\n elif match_head == \"itc\":\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sims = torch.bmm(image_feats, text_feat.unsqueeze(-1))\n sim, _ = torch.max(sims, dim=1)\n\n return sim","source_hash":"cc02551220e31d1bdae6155fced7dbc595f6948d5bbf135a9f596baea4094a26","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt","uri":"program://CREMA/module/lavis.models.blip2_models.blip2_opt#L1-L287","kind":"module","name":"lavis.models.blip2_models.blip2_opt","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":1,"end_line":287,"context_start_line":1,"context_end_line":287,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport torch\nfrom torch.cuda.amp import autocast as autocast\nimport torch.nn as nn\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_opt import OPTForCausalLM, OPTConfig\nfrom transformers import AutoTokenizer\n\n\n@registry.register_model(\"blip2_opt\")\nclass Blip2OPT(Blip2Base):\n \"\"\"\n BLIP2 OPT model.\n Supported model types:\n - pretrained_opt2.7b: pretrained model with OPT2.7b\n - pretrained_opt6.7b: pretrained model with OPT6.7b\n - caption_coco_opt2.7b: fintuned image captioning model with OPT2.7b\n - caption_coco_opt6.7b: fintuned image captioning model with OPT6.7b\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_opt\", \"caption_coco_opt2.7b\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_opt2.7b\": \"configs/models/blip2/blip2_pretrain_opt2.7b.yaml\",\n \"pretrain_opt6.7b\": \"configs/models/blip2/blip2_pretrain_opt6.7b.yaml\",\n \"caption_coco_opt2.7b\": \"configs/models/blip2/blip2_caption_opt2.7b.yaml\",\n \"caption_coco_opt6.7b\": \"configs/models/blip2/blip2_caption_opt6.7b.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n opt_model=\"facebook/opt-2.7b\",\n prompt=\"\",\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.opt_tokenizer = AutoTokenizer.from_pretrained(opt_model, use_fast=False)\n self.opt_model = OPTForCausalLM.from_pretrained(\n opt_model, torch_dtype=torch.float16\n )\n for name, param in self.opt_model.named_parameters():\n param.requires_grad = False\n self.eos_token_id = self.opt_tokenizer(\n \"\\n\", add_special_tokens=False\n ).input_ids[0]\n\n self.opt_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.opt_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n prompt_tokens = self.opt_tokenizer(self.prompt, return_tensors=\"pt\")\n self.prompt_length = prompt_tokens.attention_mask.sum(1)\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n self.opt_tokenizer.padding_side = \"right\"\n\n text = [t + \"\\n\" for t in samples[\"text_input\"]]\n\n opt_tokens = self.opt_tokenizer(\n text,\n return_tensors=\"pt\",\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n ).to(image.device)\n\n targets = opt_tokens.input_ids.masked_fill(\n opt_tokens.input_ids == self.opt_tokenizer.pad_token_id, -100\n )\n if self.prompt:\n targets[:, : self.prompt_length] = -100 # do not apply loss to the prompt\n\n empty_targets = (\n torch.ones(atts_opt.size(), dtype=torch.long).to(image.device).fill_(-100)\n )\n targets = torch.cat([empty_targets, targets], dim=1)\n\n inputs_embeds = self.opt_model.model.decoder.embed_tokens(opt_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_opt, inputs_embeds], dim=1)\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n outputs = self.opt_model(\n inputs_embeds=inputs_embeds,\n attention_mask=attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(\n enabled=(self.device != torch.device(\"cpu\"))\n ): \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n prompt = [prompt] * image.size(0)\n\n opt_tokens = self.opt_tokenizer(prompt, return_tensors=\"pt\").to(image.device)\n input_ids = opt_tokens.input_ids\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n if use_nucleus_sampling:\n query_embeds = inputs_opt.repeat_interleave(num_captions, dim=0)\n num_beams = 1\n else:\n query_embeds = inputs_opt.repeat_interleave(num_beams, dim=0)\n\n outputs = self.opt_model.generate(\n input_ids=input_ids,\n query_embeds=query_embeds,\n attention_mask=attention_mask,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n eos_token_id=self.eos_token_id,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n\n prompt_length = opt_tokens.input_ids.shape[1]\n output_text = self.opt_tokenizer.batch_decode(\n outputs[:, prompt_length:], skip_special_tokens=True\n )\n output_text = [text.strip() for text in output_text]\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n @classmethod\n def from_config(cls, cfg):\n\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n opt_model = cfg.get(\"opt_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n opt_model=opt_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt.Blip2OPT","uri":"program://CREMA/class/lavis.models.blip2_models.blip2_opt.Blip2OPT#L20-L287","kind":"class","name":"Blip2OPT","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":20,"end_line":287,"context_start_line":1,"context_end_line":287,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport torch\nfrom torch.cuda.amp import autocast as autocast\nimport torch.nn as nn\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_opt import OPTForCausalLM, OPTConfig\nfrom transformers import AutoTokenizer\n\n\n@registry.register_model(\"blip2_opt\")\nclass Blip2OPT(Blip2Base):\n \"\"\"\n BLIP2 OPT model.\n Supported model types:\n - pretrained_opt2.7b: pretrained model with OPT2.7b\n - pretrained_opt6.7b: pretrained model with OPT6.7b\n - caption_coco_opt2.7b: fintuned image captioning model with OPT2.7b\n - caption_coco_opt6.7b: fintuned image captioning model with OPT6.7b\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_opt\", \"caption_coco_opt2.7b\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_opt2.7b\": \"configs/models/blip2/blip2_pretrain_opt2.7b.yaml\",\n \"pretrain_opt6.7b\": \"configs/models/blip2/blip2_pretrain_opt6.7b.yaml\",\n \"caption_coco_opt2.7b\": \"configs/models/blip2/blip2_caption_opt2.7b.yaml\",\n \"caption_coco_opt6.7b\": \"configs/models/blip2/blip2_caption_opt6.7b.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n opt_model=\"facebook/opt-2.7b\",\n prompt=\"\",\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.opt_tokenizer = AutoTokenizer.from_pretrained(opt_model, use_fast=False)\n self.opt_model = OPTForCausalLM.from_pretrained(\n opt_model, torch_dtype=torch.float16\n )\n for name, param in self.opt_model.named_parameters():\n param.requires_grad = False\n self.eos_token_id = self.opt_tokenizer(\n \"\\n\", add_special_tokens=False\n ).input_ids[0]\n\n self.opt_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.opt_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n prompt_tokens = self.opt_tokenizer(self.prompt, return_tensors=\"pt\")\n self.prompt_length = prompt_tokens.attention_mask.sum(1)\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n self.opt_tokenizer.padding_side = \"right\"\n\n text = [t + \"\\n\" for t in samples[\"text_input\"]]\n\n opt_tokens = self.opt_tokenizer(\n text,\n return_tensors=\"pt\",\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n ).to(image.device)\n\n targets = opt_tokens.input_ids.masked_fill(\n opt_tokens.input_ids == self.opt_tokenizer.pad_token_id, -100\n )\n if self.prompt:\n targets[:, : self.prompt_length] = -100 # do not apply loss to the prompt\n\n empty_targets = (\n torch.ones(atts_opt.size(), dtype=torch.long).to(image.device).fill_(-100)\n )\n targets = torch.cat([empty_targets, targets], dim=1)\n\n inputs_embeds = self.opt_model.model.decoder.embed_tokens(opt_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_opt, inputs_embeds], dim=1)\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n outputs = self.opt_model(\n inputs_embeds=inputs_embeds,\n attention_mask=attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(\n enabled=(self.device != torch.device(\"cpu\"))\n ): \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n prompt = [prompt] * image.size(0)\n\n opt_tokens = self.opt_tokenizer(prompt, return_tensors=\"pt\").to(image.device)\n input_ids = opt_tokens.input_ids\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n if use_nucleus_sampling:\n query_embeds = inputs_opt.repeat_interleave(num_captions, dim=0)\n num_beams = 1\n else:\n query_embeds = inputs_opt.repeat_interleave(num_beams, dim=0)\n\n outputs = self.opt_model.generate(\n input_ids=input_ids,\n query_embeds=query_embeds,\n attention_mask=attention_mask,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n eos_token_id=self.eos_token_id,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n\n prompt_length = opt_tokens.input_ids.shape[1]\n output_text = self.opt_tokenizer.batch_decode(\n outputs[:, prompt_length:], skip_special_tokens=True\n )\n output_text = [text.strip() for text in output_text]\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n @classmethod\n def from_config(cls, cfg):\n\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n opt_model = cfg.get(\"opt_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n opt_model=opt_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_opt.__init__#L40-L93","kind":"function","name":"__init__","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":40,"end_line":93,"context_start_line":20,"context_end_line":113,"code":"class Blip2OPT(Blip2Base):\n \"\"\"\n BLIP2 OPT model.\n Supported model types:\n - pretrained_opt2.7b: pretrained model with OPT2.7b\n - pretrained_opt6.7b: pretrained model with OPT6.7b\n - caption_coco_opt2.7b: fintuned image captioning model with OPT2.7b\n - caption_coco_opt6.7b: fintuned image captioning model with OPT6.7b\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_opt\", \"caption_coco_opt2.7b\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_opt2.7b\": \"configs/models/blip2/blip2_pretrain_opt2.7b.yaml\",\n \"pretrain_opt6.7b\": \"configs/models/blip2/blip2_pretrain_opt6.7b.yaml\",\n \"caption_coco_opt2.7b\": \"configs/models/blip2/blip2_caption_opt2.7b.yaml\",\n \"caption_coco_opt6.7b\": \"configs/models/blip2/blip2_caption_opt6.7b.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n opt_model=\"facebook/opt-2.7b\",\n prompt=\"\",\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.opt_tokenizer = AutoTokenizer.from_pretrained(opt_model, use_fast=False)\n self.opt_model = OPTForCausalLM.from_pretrained(\n opt_model, torch_dtype=torch.float16\n )\n for name, param in self.opt_model.named_parameters():\n param.requires_grad = False\n self.eos_token_id = self.opt_tokenizer(\n \"\\n\", add_special_tokens=False\n ).input_ids[0]\n\n self.opt_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.opt_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n prompt_tokens = self.opt_tokenizer(self.prompt, return_tensors=\"pt\")\n self.prompt_length = prompt_tokens.attention_mask.sum(1)\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n self.opt_tokenizer.padding_side = \"right\"","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt.forward","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_opt.forward#L95-L148","kind":"function","name":"forward","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":95,"end_line":148,"context_start_line":75,"context_end_line":168,"code":"\n self.opt_tokenizer = AutoTokenizer.from_pretrained(opt_model, use_fast=False)\n self.opt_model = OPTForCausalLM.from_pretrained(\n opt_model, torch_dtype=torch.float16\n )\n for name, param in self.opt_model.named_parameters():\n param.requires_grad = False\n self.eos_token_id = self.opt_tokenizer(\n \"\\n\", add_special_tokens=False\n ).input_ids[0]\n\n self.opt_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.opt_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n prompt_tokens = self.opt_tokenizer(self.prompt, return_tensors=\"pt\")\n self.prompt_length = prompt_tokens.attention_mask.sum(1)\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n self.opt_tokenizer.padding_side = \"right\"\n\n text = [t + \"\\n\" for t in samples[\"text_input\"]]\n\n opt_tokens = self.opt_tokenizer(\n text,\n return_tensors=\"pt\",\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n ).to(image.device)\n\n targets = opt_tokens.input_ids.masked_fill(\n opt_tokens.input_ids == self.opt_tokenizer.pad_token_id, -100\n )\n if self.prompt:\n targets[:, : self.prompt_length] = -100 # do not apply loss to the prompt\n\n empty_targets = (\n torch.ones(atts_opt.size(), dtype=torch.long).to(image.device).fill_(-100)\n )\n targets = torch.cat([empty_targets, targets], dim=1)\n\n inputs_embeds = self.opt_model.model.decoder.embed_tokens(opt_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_opt, inputs_embeds], dim=1)\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n outputs = self.opt_model(\n inputs_embeds=inputs_embeds,\n attention_mask=attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt.generate","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_opt.generate#L151-L257","kind":"function","name":"generate","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":151,"end_line":257,"context_start_line":131,"context_end_line":277,"code":" empty_targets = (\n torch.ones(atts_opt.size(), dtype=torch.long).to(image.device).fill_(-100)\n )\n targets = torch.cat([empty_targets, targets], dim=1)\n\n inputs_embeds = self.opt_model.model.decoder.embed_tokens(opt_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_opt, inputs_embeds], dim=1)\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n outputs = self.opt_model(\n inputs_embeds=inputs_embeds,\n attention_mask=attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(\n enabled=(self.device != torch.device(\"cpu\"))\n ): \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_opt = self.opt_proj(query_output.last_hidden_state)\n atts_opt = torch.ones(inputs_opt.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n prompt = [prompt] * image.size(0)\n\n opt_tokens = self.opt_tokenizer(prompt, return_tensors=\"pt\").to(image.device)\n input_ids = opt_tokens.input_ids\n attention_mask = torch.cat([atts_opt, opt_tokens.attention_mask], dim=1)\n\n if use_nucleus_sampling:\n query_embeds = inputs_opt.repeat_interleave(num_captions, dim=0)\n num_beams = 1\n else:\n query_embeds = inputs_opt.repeat_interleave(num_beams, dim=0)\n\n outputs = self.opt_model.generate(\n input_ids=input_ids,\n query_embeds=query_embeds,\n attention_mask=attention_mask,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n eos_token_id=self.eos_token_id,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n\n prompt_length = opt_tokens.input_ids.shape[1]\n output_text = self.opt_tokenizer.batch_decode(\n outputs[:, prompt_length:], skip_special_tokens=True\n )\n output_text = [text.strip() for text in output_text]\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n @classmethod\n def from_config(cls, cfg):\n\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n opt_model = cfg.get(\"opt_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_opt.from_config","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_opt.from_config#L260-L287","kind":"function","name":"from_config","path":"lavis/models/blip2_models/blip2_opt.py","language":"python","start_line":260,"end_line":287,"context_start_line":240,"context_end_line":287,"code":" )\n\n prompt_length = opt_tokens.input_ids.shape[1]\n output_text = self.opt_tokenizer.batch_decode(\n outputs[:, prompt_length:], skip_special_tokens=True\n )\n output_text = [text.strip() for text in output_text]\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n @classmethod\n def from_config(cls, cfg):\n\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n opt_model = cfg.get(\"opt_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n opt_model=opt_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"989c4803e6a195552ceb41a7f5f0f6ee4e3e6a061f255cd25d214bacd0961c39","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2","uri":"program://CREMA/module/lavis.models.blip2_models.blip2#L1-L554","kind":"module","name":"lavis.models.blip2_models.blip2","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":1,"end_line":554,"context_start_line":1,"context_end_line":554,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\nimport os\nimport time\nimport datetime\n\nimport torch\nimport torch.nn as nn\nimport torch.distributed as dist\nimport torch.nn.functional as F\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.common.logger import MetricLogger\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.blip2_models.Qformer import BertConfig, BertLMHeadModel\nfrom lavis.models.eva_vit import create_eva_vit_g\nfrom transformers import BertTokenizer\n\nimport math\nfrom torch import Tensor\nfrom einops import rearrange\nfrom safetensors import safe_open\nfrom safetensors.torch import save_file\nfrom torch.nn.parameter import Parameter\n\nclass _LoRALayer_multimodal(nn.Module):\n def __init__(self, w, w_a, w_b, dropout, modulars):\n super().__init__()\n self.w = w\n\n # multimodal lora\n if 'rgb' in modulars:\n self.w_a_rgb = w_a['rgb']\n self.w_b_rgb = w_b['rgb']\n self.dropout_rgb = dropout['rgb']\n if 'depth' in modulars:\n self.w_a_depth = w_a['depth']\n self.w_b_depth = w_b['depth']\n self.dropout_depth = dropout['depth']\n if 'flow' in modulars:\n self.w_a_flow = w_a['flow']\n self.w_b_flow = w_b['flow']\n self.dropout_flow = dropout['flow']\n if 'norm' in modulars:\n self.w_a_norm = w_a['norm']\n self.w_b_norm = w_b['norm']\n self.dropout_norm = dropout['norm']\n if 'audio' in modulars:\n self.w_a_audio = w_a['audio']\n self.w_b_audio = w_b['audio']\n self.dropout_audio = dropout['audio']\n if 'pc' in modulars:\n self.w_a_pc = w_a['pc']\n self.w_b_pc = w_b['pc']\n self.dropout_pc = dropout['pc']\n\n def forward(self, x, modular='rgb'):\n w_b = getattr(self, f'w_b_{modular}')\n w_a = getattr(self, f'w_a_{modular}')\n dropout = getattr(self, f'dropout_{modular}')\n \n if 'skip' in modular:\n x = self.w(x)\n else:\n x = self.w(x) + w_b(w_a(dropout(x)))\n\n return x\n\n# multitask Qformer\nclass LoRA_Multimodal_QFormer(nn.Module):\n \"\"\"Applies low-rank adaptation to a vision transformer.\n\n Args:\n vit_model: a vision transformer model, see base_vit.py\n r: rank of LoRA\n num_classes: how many classes the model output, default to the vit model\n lora_layer: which layer we apply LoRA.\n\n Examples::\n >>> model = ViT('B_16_imagenet1k')\n >>> lora_model = LoRA_ViT(model, r=4)\n >>> preds = lora_model(img)\n >>> print(preds.shape)\n torch.Size([1, 1000])\n \"\"\"\n\n def __init__(self, qformer, modulars,\n r: int, \n lora_layer=None,\n cross_attention_freq=2,\n lora_dropout=0.1):\n super(LoRA_Multimodal_QFormer, self).__init__()\n\n \n assert r > 0\n base_vit_dim = qformer.bert.encoder.layer[0].attention.self.query.out_features\n dim = base_vit_dim\n\n self.modulars = modulars\n \n if lora_layer:\n self.lora_layer = lora_layer\n else:\n self.lora_layer = list(range(len(qformer.bert.encoder.layer)))\n \n # create for storage, then we can init them or load weights\n self.w_As = [] # These are linear layers\n self.w_Bs = []\n # self.w_As_cross = [] \n # self.w_Bs_cross = []\n\n # lets freeze first / yui: nothing changes here for test\n for param in qformer.parameters():\n param.requires_grad = False\n\n # Here, we do the surgery\n # yui: yeah, let's do the second surgery for the multimodal adapter here ;-)\n \n for t_layer_i, blk in enumerate(qformer.bert.encoder.layer):\n # If we only want few lora layer instead of all\n if t_layer_i not in self.lora_layer:\n continue\n\n # replace self-attention\n w_q_linear = blk.attention.self.query\n w_v_linear = blk.attention.self.value\n w_a_linear_q, w_a_linear_v = {}, {}\n w_b_linear_q, w_b_linear_v = {}, {}\n dropout_q, dropout_v = {}, {}\n \n for m in self.modulars:\n # if m == 'rgb':\n # continue\n w_a_linear_q[m] = nn.Linear(dim, r, bias=False)\n w_a_linear_v[m] = nn.Linear(dim, r, bias=False)\n w_b_linear_q[m] = nn.Linear(r, dim, bias=False)\n w_b_linear_v[m] = nn.Linear(r, dim, bias=False)\n if lora_dropout > 0.0:\n dropout_q[m] = nn.Dropout(p=lora_dropout)\n dropout_v[m] = nn.Dropout(p=lora_dropout)\n else:\n dropout_q[m] = nn.Identity()\n dropout_v[m] = nn.Identity()\n \n self.w_As.append(w_a_linear_q[m])\n self.w_Bs.append(w_b_linear_q[m])\n self.w_As.append(w_a_linear_v[m])\n self.w_Bs.append(w_b_linear_v[m])\n \n blk.attention.self.query = _LoRALayer_multimodal(w_q_linear, w_a_linear_q, w_b_linear_q, dropout_q, modulars)\n blk.attention.self.value = _LoRALayer_multimodal(w_v_linear, w_a_linear_v, w_b_linear_v, dropout_v, modulars)\n\n self.reset_parameters()\n self.lora_qformer = qformer\n\n def save_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.fc.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n\n def load_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n \n @classmethod\n def init_Multimodal_Qformer(cls, num_query_token, vision_width, \n modulars, r=64, lora_layer=None, lora_dropout=0.1):\n\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n\n LoRA_Multimodal_QFormer(Qformer, modulars, \n r=r,\n lora_layer=lora_layer,\n cross_attention_freq=encoder_config.cross_attention_freq,\n lora_dropout=lora_dropout\n )\n \n return Qformer, encoder_config\n \n @classmethod\n def init_ln(cls, num_features, load_ln_path=False, load_ln_type=\"\"):\n ln = LayerNorm(num_features)\n if load_ln_path and load_ln_type:\n url_or_filename=load_ln_path\n logging.info(f\"Loading pretrained layer norm weights from {url_or_filename} of type {load_ln_type}\")\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n \n if load_ln_type:\n load_ln_type = f\"{load_ln_type}_ln\" if \"vision\" not in load_ln_type else \"ln_vision\"\n loaded_state_dict = {}\n if 'model' in checkpoint:\n checkpoint = checkpoint['model'] \n for k in checkpoint.keys():\n if load_ln_type in k:\n loaded_state_dict['.'.join(k.split('.')[1:])] = checkpoint[k]\n ln.load_state_dict(loaded_state_dict, strict=False)\n \n return ln\n\n @classmethod\n def init_audio_encoder(self, \n model_name, cached_audio, load_ln_path=False, load_ln_type=\"\"):\n assert model_name in [\n 'beats'\n ], \"audio model must be in [beats]\"\n\n # load_ln_path = kwargs['load_ln_path']\n # del kwargs['load_ln_path']\n # load_ln_type=kwargs['load_ln_type']\n # del kwargs['load_ln_type']\n kwargs = {}\n if \"beats\" in model_name:\n from lavis.models.beats_encoder import BeatsEncoder\n if cached_audio:\n audio_encoder = lambda x: x\n ln_audio = self.init_ln(768, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n else:\n audio_encoder = BeatsEncoder(**kwargs)\n\n if not cached_audio:\n ln_audio = self.init_ln(audio_encoder.num_features, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n self.audio_enc_name = model_name\n\n return audio_encoder, ln_audio\n \n @classmethod\n def init_TemporalQFormer(cls, num_of_frame):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.query_length = num_of_frame\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = model.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=35,\n return_tensors=\"pt\",\n ).to(model.device)\n text_feat = model.forward_text(text_input)\n text_embed = F.normalize(model.text_proj(text_feat))\n text_embeds.append(text_embed)\n text_ids.append(text_input.input_ids)\n text_atts.append(text_input.attention_mask)\n\n text_embeds = torch.cat(text_embeds, dim=0)\n text_ids = torch.cat(text_ids, dim=0)\n text_atts = torch.cat(text_atts, dim=0)\n\n vit_feats = []\n image_embeds = []\n for samples in data_loader:\n image = samples[\"image\"]\n\n image = image.to(model.device)\n image_feat, vit_feat = model.forward_image(image)\n image_embed = model.vision_proj(image_feat)\n image_embed = F.normalize(image_embed, dim=-1)\n\n vit_feats.append(vit_feat.cpu())\n image_embeds.append(image_embed)\n\n vit_feats = torch.cat(vit_feats, dim=0)\n image_embeds = torch.cat(image_embeds, dim=0)\n\n sims_matrix = []\n for image_embed in image_embeds:\n sim_q2t = image_embed @ text_embeds.t()\n sim_i2t, _ = sim_q2t.max(0)\n sims_matrix.append(sim_i2t)\n sims_matrix = torch.stack(sims_matrix, dim=0)\n\n score_matrix_i2t = torch.full(\n (len(data_loader.dataset.image), len(texts)), -100.0\n ).to(model.device)\n\n num_tasks = dist_utils.get_world_size()\n rank = dist_utils.get_rank()\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n image_inputs = vit_feats[start + i].repeat(k_test, 1, 1).to(model.device)\n score = model.compute_itm(\n image_inputs=image_inputs,\n text_ids=text_ids[topk_idx],\n text_atts=text_atts[topk_idx],\n ).float()\n score_matrix_i2t[start + i, topk_idx] = score + topk_sim\n\n sims_matrix = sims_matrix.t()\n score_matrix_t2i = torch.full(\n (len(texts), len(data_loader.dataset.image)), -100.0\n ).to(model.device)\n\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n image_inputs = vit_feats[topk_idx.cpu()].to(model.device)\n score = model.compute_itm(\n image_inputs=image_inputs,\n text_ids=text_ids[start + i].repeat(k_test, 1),\n text_atts=text_atts[start + i].repeat(k_test, 1),\n ).float()\n score_matrix_t2i[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.\n# ... truncated ...","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2._LoRALayer_multimodal","uri":"program://CREMA/class/lavis.models.blip2_models.blip2._LoRALayer_multimodal#L33-L74","kind":"class","name":"_LoRALayer_multimodal","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":33,"end_line":74,"context_start_line":13,"context_end_line":94,"code":"import torch.nn as nn\nimport torch.distributed as dist\nimport torch.nn.functional as F\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.common.logger import MetricLogger\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.blip2_models.Qformer import BertConfig, BertLMHeadModel\nfrom lavis.models.eva_vit import create_eva_vit_g\nfrom transformers import BertTokenizer\n\nimport math\nfrom torch import Tensor\nfrom einops import rearrange\nfrom safetensors import safe_open\nfrom safetensors.torch import save_file\nfrom torch.nn.parameter import Parameter\n\nclass _LoRALayer_multimodal(nn.Module):\n def __init__(self, w, w_a, w_b, dropout, modulars):\n super().__init__()\n self.w = w\n\n # multimodal lora\n if 'rgb' in modulars:\n self.w_a_rgb = w_a['rgb']\n self.w_b_rgb = w_b['rgb']\n self.dropout_rgb = dropout['rgb']\n if 'depth' in modulars:\n self.w_a_depth = w_a['depth']\n self.w_b_depth = w_b['depth']\n self.dropout_depth = dropout['depth']\n if 'flow' in modulars:\n self.w_a_flow = w_a['flow']\n self.w_b_flow = w_b['flow']\n self.dropout_flow = dropout['flow']\n if 'norm' in modulars:\n self.w_a_norm = w_a['norm']\n self.w_b_norm = w_b['norm']\n self.dropout_norm = dropout['norm']\n if 'audio' in modulars:\n self.w_a_audio = w_a['audio']\n self.w_b_audio = w_b['audio']\n self.dropout_audio = dropout['audio']\n if 'pc' in modulars:\n self.w_a_pc = w_a['pc']\n self.w_b_pc = w_b['pc']\n self.dropout_pc = dropout['pc']\n\n def forward(self, x, modular='rgb'):\n w_b = getattr(self, f'w_b_{modular}')\n w_a = getattr(self, f'w_a_{modular}')\n dropout = getattr(self, f'dropout_{modular}')\n \n if 'skip' in modular:\n x = self.w(x)\n else:\n x = self.w(x) + w_b(w_a(dropout(x)))\n\n return x\n\n# multitask Qformer\nclass LoRA_Multimodal_QFormer(nn.Module):\n \"\"\"Applies low-rank adaptation to a vision transformer.\n\n Args:\n vit_model: a vision transformer model, see base_vit.py\n r: rank of LoRA\n num_classes: how many classes the model output, default to the vit model\n lora_layer: which layer we apply LoRA.\n\n Examples::\n >>> model = ViT('B_16_imagenet1k')\n >>> lora_model = LoRA_ViT(model, r=4)\n >>> preds = lora_model(img)\n >>> print(preds.shape)\n torch.Size([1, 1000])\n \"\"\"\n\n def __init__(self, qformer, modulars,","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.LoRA_Multimodal_QFormer","uri":"program://CREMA/class/lavis.models.blip2_models.blip2.LoRA_Multimodal_QFormer#L77-L250","kind":"class","name":"LoRA_Multimodal_QFormer","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":77,"end_line":250,"context_start_line":57,"context_end_line":270,"code":" self.w_b_audio = w_b['audio']\n self.dropout_audio = dropout['audio']\n if 'pc' in modulars:\n self.w_a_pc = w_a['pc']\n self.w_b_pc = w_b['pc']\n self.dropout_pc = dropout['pc']\n\n def forward(self, x, modular='rgb'):\n w_b = getattr(self, f'w_b_{modular}')\n w_a = getattr(self, f'w_a_{modular}')\n dropout = getattr(self, f'dropout_{modular}')\n \n if 'skip' in modular:\n x = self.w(x)\n else:\n x = self.w(x) + w_b(w_a(dropout(x)))\n\n return x\n\n# multitask Qformer\nclass LoRA_Multimodal_QFormer(nn.Module):\n \"\"\"Applies low-rank adaptation to a vision transformer.\n\n Args:\n vit_model: a vision transformer model, see base_vit.py\n r: rank of LoRA\n num_classes: how many classes the model output, default to the vit model\n lora_layer: which layer we apply LoRA.\n\n Examples::\n >>> model = ViT('B_16_imagenet1k')\n >>> lora_model = LoRA_ViT(model, r=4)\n >>> preds = lora_model(img)\n >>> print(preds.shape)\n torch.Size([1, 1000])\n \"\"\"\n\n def __init__(self, qformer, modulars,\n r: int, \n lora_layer=None,\n cross_attention_freq=2,\n lora_dropout=0.1):\n super(LoRA_Multimodal_QFormer, self).__init__()\n\n \n assert r > 0\n base_vit_dim = qformer.bert.encoder.layer[0].attention.self.query.out_features\n dim = base_vit_dim\n\n self.modulars = modulars\n \n if lora_layer:\n self.lora_layer = lora_layer\n else:\n self.lora_layer = list(range(len(qformer.bert.encoder.layer)))\n \n # create for storage, then we can init them or load weights\n self.w_As = [] # These are linear layers\n self.w_Bs = []\n # self.w_As_cross = [] \n # self.w_Bs_cross = []\n\n # lets freeze first / yui: nothing changes here for test\n for param in qformer.parameters():\n param.requires_grad = False\n\n # Here, we do the surgery\n # yui: yeah, let's do the second surgery for the multimodal adapter here ;-)\n \n for t_layer_i, blk in enumerate(qformer.bert.encoder.layer):\n # If we only want few lora layer instead of all\n if t_layer_i not in self.lora_layer:\n continue\n\n # replace self-attention\n w_q_linear = blk.attention.self.query\n w_v_linear = blk.attention.self.value\n w_a_linear_q, w_a_linear_v = {}, {}\n w_b_linear_q, w_b_linear_v = {}, {}\n dropout_q, dropout_v = {}, {}\n \n for m in self.modulars:\n # if m == 'rgb':\n # continue\n w_a_linear_q[m] = nn.Linear(dim, r, bias=False)\n w_a_linear_v[m] = nn.Linear(dim, r, bias=False)\n w_b_linear_q[m] = nn.Linear(r, dim, bias=False)\n w_b_linear_v[m] = nn.Linear(r, dim, bias=False)\n if lora_dropout > 0.0:\n dropout_q[m] = nn.Dropout(p=lora_dropout)\n dropout_v[m] = nn.Dropout(p=lora_dropout)\n else:\n dropout_q[m] = nn.Identity()\n dropout_v[m] = nn.Identity()\n \n self.w_As.append(w_a_linear_q[m])\n self.w_Bs.append(w_b_linear_q[m])\n self.w_As.append(w_a_linear_v[m])\n self.w_Bs.append(w_b_linear_v[m])\n \n blk.attention.self.query = _LoRALayer_multimodal(w_q_linear, w_a_linear_q, w_b_linear_q, dropout_q, modulars)\n blk.attention.self.value = _LoRALayer_multimodal(w_v_linear, w_a_linear_v, w_b_linear_v, dropout_v, modulars)\n\n self.reset_parameters()\n self.lora_qformer = qformer\n\n def save_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.fc.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n\n def load_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.Blip2Base","uri":"program://CREMA/class/lavis.models.blip2_models.blip2.Blip2Base#L252-L423","kind":"class","name":"Blip2Base","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":252,"end_line":423,"context_start_line":232,"context_end_line":443,"code":" try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n \n @classmethod\n def init_Multimodal_Qformer(cls, num_query_token, vision_width, \n modulars, r=64, lora_layer=None, lora_dropout=0.1):\n\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n\n LoRA_Multimodal_QFormer(Qformer, modulars, \n r=r,\n lora_layer=lora_layer,\n cross_attention_freq=encoder_config.cross_attention_freq,\n lora_dropout=lora_dropout\n )\n \n return Qformer, encoder_config\n \n @classmethod\n def init_ln(cls, num_features, load_ln_path=False, load_ln_type=\"\"):\n ln = LayerNorm(num_features)\n if load_ln_path and load_ln_type:\n url_or_filename=load_ln_path\n logging.info(f\"Loading pretrained layer norm weights from {url_or_filename} of type {load_ln_type}\")\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n \n if load_ln_type:\n load_ln_type = f\"{load_ln_type}_ln\" if \"vision\" not in load_ln_type else \"ln_vision\"\n loaded_state_dict = {}\n if 'model' in checkpoint:\n checkpoint = checkpoint['model'] \n for k in checkpoint.keys():\n if load_ln_type in k:\n loaded_state_dict['.'.join(k.split('.')[1:])] = checkpoint[k]\n ln.load_state_dict(loaded_state_dict, strict=False)\n \n return ln\n\n @classmethod\n def init_audio_encoder(self, \n model_name, cached_audio, load_ln_path=False, load_ln_type=\"\"):\n assert model_name in [\n 'beats'\n ], \"audio model must be in [beats]\"\n\n # load_ln_path = kwargs['load_ln_path']\n # del kwargs['load_ln_path']\n # load_ln_type=kwargs['load_ln_type']\n # del kwargs['load_ln_type']\n kwargs = {}\n if \"beats\" in model_name:\n from lavis.models.beats_encoder import BeatsEncoder\n if cached_audio:\n audio_encoder = lambda x: x\n ln_audio = self.init_ln(768, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n else:\n audio_encoder = BeatsEncoder(**kwargs)\n\n if not cached_audio:\n ln_audio = self.init_ln(audio_encoder.num_features, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n self.audio_enc_name = model_name\n\n return audio_encoder, ln_audio\n \n @classmethod\n def init_TemporalQFormer(cls, num_of_frame):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.query_length = num_of_frame\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.disabled_train","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.disabled_train#L426-L429","kind":"function","name":"disabled_train","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":426,"end_line":429,"context_start_line":406,"context_end_line":449,"code":"\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.LayerNorm","uri":"program://CREMA/class/lavis.models.blip2_models.blip2.LayerNorm#L432-L438","kind":"class","name":"LayerNorm","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":432,"end_line":438,"context_start_line":412,"context_end_line":458,"code":" logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = model.tokenizer(","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.compute_sim_matrix#L441-L554","kind":"function","name":"compute_sim_matrix","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":441,"end_line":554,"context_start_line":421,"context_end_line":554,"code":" state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = model.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=35,\n return_tensors=\"pt\",\n ).to(model.device)\n text_feat = model.forward_text(text_input)\n text_embed = F.normalize(model.text_proj(text_feat))\n text_embeds.append(text_embed)\n text_ids.append(text_input.input_ids)\n text_atts.append(text_input.attention_mask)\n\n text_embeds = torch.cat(text_embeds, dim=0)\n text_ids = torch.cat(text_ids, dim=0)\n text_atts = torch.cat(text_atts, dim=0)\n\n vit_feats = []\n image_embeds = []\n for samples in data_loader:\n image = samples[\"image\"]\n\n image = image.to(model.device)\n image_feat, vit_feat = model.forward_image(image)\n image_embed = model.vision_proj(image_feat)\n image_embed = F.normalize(image_embed, dim=-1)\n\n vit_feats.append(vit_feat.cpu())\n image_embeds.append(image_embed)\n\n vit_feats = torch.cat(vit_feats, dim=0)\n image_embeds = torch.cat(image_embeds, dim=0)\n\n sims_matrix = []\n for image_embed in image_embeds:\n sim_q2t = image_embed @ text_embeds.t()\n sim_i2t, _ = sim_q2t.max(0)\n sims_matrix.append(sim_i2t)\n sims_matrix = torch.stack(sims_matrix, dim=0)\n\n score_matrix_i2t = torch.full(\n (len(data_loader.dataset.image), len(texts)), -100.0\n ).to(model.device)\n\n num_tasks = dist_utils.get_world_size()\n rank = dist_utils.get_rank()\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n image_inputs = vit_feats[start + i].repeat(k_test, 1, 1).to(model.device)\n score = model.compute_itm(\n image_inputs=image_inputs,\n text_ids=text_ids[topk_idx],\n text_atts=text_atts[topk_idx],\n ).float()\n score_matrix_i2t[start + i, topk_idx] = score + topk_sim\n\n sims_matrix = sims_matrix.t()\n score_matrix_t2i = torch.full(\n (len(texts), len(data_loader.dataset.image)), -100.0\n ).to(model.device)\n\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n image_inputs = vit_feats[topk_idx.cpu()].to(model.device)\n score = model.compute_itm(\n image_inputs=image_inputs,\n text_ids=text_ids[start + i].repeat(k_test, 1),\n text_atts=text_atts[start + i].repeat(k_test, 1),\n ).float()\n score_matrix_t2i[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.barrier()\n torch.distributed.all_reduce(\n score_matrix_i2t, op=torch.distributed.ReduceOp.SUM\n )\n torch.distributed.all_reduce(\n score_matrix_t2i, op=torch.distributed.ReduceOp.SUM\n )\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return score_matrix_i2t.cpu().numpy(), score_matrix_t2i.cpu().numpy()","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.__init__#L94-L161","kind":"function","name":"__init__","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":94,"end_line":161,"context_start_line":74,"context_end_line":181,"code":" return x\n\n# multitask Qformer\nclass LoRA_Multimodal_QFormer(nn.Module):\n \"\"\"Applies low-rank adaptation to a vision transformer.\n\n Args:\n vit_model: a vision transformer model, see base_vit.py\n r: rank of LoRA\n num_classes: how many classes the model output, default to the vit model\n lora_layer: which layer we apply LoRA.\n\n Examples::\n >>> model = ViT('B_16_imagenet1k')\n >>> lora_model = LoRA_ViT(model, r=4)\n >>> preds = lora_model(img)\n >>> print(preds.shape)\n torch.Size([1, 1000])\n \"\"\"\n\n def __init__(self, qformer, modulars,\n r: int, \n lora_layer=None,\n cross_attention_freq=2,\n lora_dropout=0.1):\n super(LoRA_Multimodal_QFormer, self).__init__()\n\n \n assert r > 0\n base_vit_dim = qformer.bert.encoder.layer[0].attention.self.query.out_features\n dim = base_vit_dim\n\n self.modulars = modulars\n \n if lora_layer:\n self.lora_layer = lora_layer\n else:\n self.lora_layer = list(range(len(qformer.bert.encoder.layer)))\n \n # create for storage, then we can init them or load weights\n self.w_As = [] # These are linear layers\n self.w_Bs = []\n # self.w_As_cross = [] \n # self.w_Bs_cross = []\n\n # lets freeze first / yui: nothing changes here for test\n for param in qformer.parameters():\n param.requires_grad = False\n\n # Here, we do the surgery\n # yui: yeah, let's do the second surgery for the multimodal adapter here ;-)\n \n for t_layer_i, blk in enumerate(qformer.bert.encoder.layer):\n # If we only want few lora layer instead of all\n if t_layer_i not in self.lora_layer:\n continue\n\n # replace self-attention\n w_q_linear = blk.attention.self.query\n w_v_linear = blk.attention.self.value\n w_a_linear_q, w_a_linear_v = {}, {}\n w_b_linear_q, w_b_linear_v = {}, {}\n dropout_q, dropout_v = {}, {}\n \n for m in self.modulars:\n # if m == 'rgb':\n # continue\n w_a_linear_q[m] = nn.Linear(dim, r, bias=False)\n w_a_linear_v[m] = nn.Linear(dim, r, bias=False)\n w_b_linear_q[m] = nn.Linear(r, dim, bias=False)\n w_b_linear_v[m] = nn.Linear(r, dim, bias=False)\n if lora_dropout > 0.0:\n dropout_q[m] = nn.Dropout(p=lora_dropout)\n dropout_v[m] = nn.Dropout(p=lora_dropout)\n else:\n dropout_q[m] = nn.Identity()\n dropout_v[m] = nn.Identity()\n \n self.w_As.append(w_a_linear_q[m])\n self.w_Bs.append(w_b_linear_q[m])\n self.w_As.append(w_a_linear_v[m])\n self.w_Bs.append(w_b_linear_v[m])\n \n blk.attention.self.query = _LoRALayer_multimodal(w_q_linear, w_a_linear_q, w_b_linear_q, dropout_q, modulars)\n blk.attention.self.value = _LoRALayer_multimodal(w_v_linear, w_a_linear_v, w_b_linear_v, dropout_v, modulars)\n\n self.reset_parameters()\n self.lora_qformer = qformer\n\n def save_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.forward","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.forward#L435-L438","kind":"function","name":"forward","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":435,"end_line":438,"context_start_line":415,"context_end_line":458,"code":" \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = model.tokenizer(","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.save_fc_parameters","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.save_fc_parameters#L163-L172","kind":"function","name":"save_fc_parameters","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":163,"end_line":172,"context_start_line":143,"context_end_line":192,"code":" w_b_linear_q[m] = nn.Linear(r, dim, bias=False)\n w_b_linear_v[m] = nn.Linear(r, dim, bias=False)\n if lora_dropout > 0.0:\n dropout_q[m] = nn.Dropout(p=lora_dropout)\n dropout_v[m] = nn.Dropout(p=lora_dropout)\n else:\n dropout_q[m] = nn.Identity()\n dropout_v[m] = nn.Identity()\n \n self.w_As.append(w_a_linear_q[m])\n self.w_Bs.append(w_b_linear_q[m])\n self.w_As.append(w_a_linear_v[m])\n self.w_Bs.append(w_b_linear_v[m])\n \n blk.attention.self.query = _LoRALayer_multimodal(w_q_linear, w_a_linear_q, w_b_linear_q, dropout_q, modulars)\n blk.attention.self.value = _LoRALayer_multimodal(w_v_linear, w_a_linear_v, w_b_linear_v, dropout_v, modulars)\n\n self.reset_parameters()\n self.lora_qformer = qformer\n\n def save_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.fc.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.load_fc_parameters","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.load_fc_parameters#L174-L189","kind":"function","name":"load_fc_parameters","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":174,"end_line":189,"context_start_line":154,"context_end_line":209,"code":" self.w_As.append(w_a_linear_v[m])\n self.w_Bs.append(w_b_linear_v[m])\n \n blk.attention.self.query = _LoRALayer_multimodal(w_q_linear, w_a_linear_q, w_b_linear_q, dropout_q, modulars)\n blk.attention.self.value = _LoRALayer_multimodal(w_v_linear, w_a_linear_v, w_b_linear_v, dropout_v, modulars)\n\n self.reset_parameters()\n self.lora_qformer = qformer\n\n def save_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.fc.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.save_lora_parameters","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.save_lora_parameters#L191-L208","kind":"function","name":"save_lora_parameters","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":191,"end_line":208,"context_start_line":171,"context_end_line":228,"code":" fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.fc.weight}\n save_file(fc_tensors, filename)\n\n def load_fc_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n _in = self.lora_vit.fc.in_features\n _out = self.lora_vit.fc.out_features\n with safe_open(filename, framework=\"pt\") as f:\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.fc.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n\n def load_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n ","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.load_lora_parameters","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.load_lora_parameters#L210-L236","kind":"function","name":"load_lora_parameters","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":210,"end_line":236,"context_start_line":190,"context_end_line":256,"code":"\n def save_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n num_layer = len(self.w_As) # actually, it is half\n a_tensors = {f\"w_a_{i:03d}\": self.w_As[i].weight for i in range(num_layer)}\n b_tensors = {f\"w_b_{i:03d}\": self.w_Bs[i].weight for i in range(num_layer)}\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n fc_tensors = {f\"fc_{_in}in_{_out}out\": self.lora_vit.head.weight}\n \n merged_dict = {**a_tensors, **b_tensors, **fc_tensors}\n save_file(merged_dict, filename)\n\n def load_lora_parameters(self, filename: str) -> None:\n r\"\"\"Only safetensors is supported now.\n\n pip install safetensor if you do not have one installed yet.\n \"\"\"\n\n assert filename.endswith(\".safetensors\")\n\n with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.reset_parameters","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.reset_parameters#L238-L242","kind":"function","name":"reset_parameters","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":238,"end_line":242,"context_start_line":218,"context_end_line":262,"code":" with safe_open(filename, framework=\"pt\") as f:\n for i, w_A_linear in enumerate(self.w_As):\n saved_key = f\"w_a_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_A_linear.weight = Parameter(saved_tensor)\n\n for i, w_B_linear in enumerate(self.w_Bs):\n saved_key = f\"w_b_{i:03d}\"\n saved_tensor = f.get_tensor(saved_key)\n w_B_linear.weight = Parameter(saved_tensor)\n \n _in = self.lora_vit.head.in_features\n _out = self.lora_vit.head.out_features\n saved_key = f\"fc_{_in}in_{_out}out\"\n try:\n saved_tensor = f.get_tensor(saved_key)\n self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_tokenizer","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_tokenizer#L254-L257","kind":"function","name":"init_tokenizer","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":254,"end_line":257,"context_start_line":234,"context_end_line":277,"code":" self.lora_vit.head.weight = Parameter(saved_tensor)\n except ValueError:\n print(\"this fc weight is not for this model\")\n\n def reset_parameters(self) -> None:\n for w_A in self.w_As:\n nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n \n @classmethod\n def init_Multimodal_Qformer(cls, num_query_token, vision_width, ","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_Qformer","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_Qformer#L260-L274","kind":"function","name":"init_Qformer","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":260,"end_line":274,"context_start_line":240,"context_end_line":294,"code":" nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n for w_B in self.w_Bs:\n nn.init.zeros_(w_B.weight)\n\n # for w_A in self.w_As_cross:\n # nn.init.kaiming_uniform_(w_A.weight, a=math.sqrt(5))\n # for w_B in self.w_Bs_cross:\n # nn.init.zeros_(w_B.weight)\n\n def forward(self, x: Tensor) -> Tensor:\n return self.lora_qformer(x)\n\nclass Blip2Base(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n \n @classmethod\n def init_Multimodal_Qformer(cls, num_query_token, vision_width, \n modulars, r=64, lora_layer=None, lora_dropout=0.1):\n\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n\n LoRA_Multimodal_QFormer(Qformer, modulars, \n r=r,\n lora_layer=lora_layer,\n cross_attention_freq=encoder_config.cross_attention_freq,\n lora_dropout=lora_dropout","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_Multimodal_Qformer","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_Multimodal_Qformer#L277-L297","kind":"function","name":"init_Multimodal_Qformer","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":277,"end_line":297,"context_start_line":257,"context_end_line":317,"code":" return tokenizer\n\n @classmethod\n def init_Qformer(cls, num_query_token, vision_width):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n \n @classmethod\n def init_Multimodal_Qformer(cls, num_query_token, vision_width, \n modulars, r=64, lora_layer=None, lora_dropout=0.1):\n\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n\n LoRA_Multimodal_QFormer(Qformer, modulars, \n r=r,\n lora_layer=lora_layer,\n cross_attention_freq=encoder_config.cross_attention_freq,\n lora_dropout=lora_dropout\n )\n \n return Qformer, encoder_config\n \n @classmethod\n def init_ln(cls, num_features, load_ln_path=False, load_ln_type=\"\"):\n ln = LayerNorm(num_features)\n if load_ln_path and load_ln_type:\n url_or_filename=load_ln_path\n logging.info(f\"Loading pretrained layer norm weights from {url_or_filename} of type {load_ln_type}\")\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n \n if load_ln_type:\n load_ln_type = f\"{load_ln_type}_ln\" if \"vision\" not in load_ln_type else \"ln_vision\"\n loaded_state_dict = {}","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_ln","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_ln#L300-L325","kind":"function","name":"init_ln","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":300,"end_line":325,"context_start_line":280,"context_end_line":345,"code":" encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.encoder_width = vision_width\n # insert cross-attention layer every other block\n encoder_config.add_cross_attention = True\n encoder_config.cross_attention_freq = 2\n encoder_config.query_length = num_query_token\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n\n LoRA_Multimodal_QFormer(Qformer, modulars, \n r=r,\n lora_layer=lora_layer,\n cross_attention_freq=encoder_config.cross_attention_freq,\n lora_dropout=lora_dropout\n )\n \n return Qformer, encoder_config\n \n @classmethod\n def init_ln(cls, num_features, load_ln_path=False, load_ln_type=\"\"):\n ln = LayerNorm(num_features)\n if load_ln_path and load_ln_type:\n url_or_filename=load_ln_path\n logging.info(f\"Loading pretrained layer norm weights from {url_or_filename} of type {load_ln_type}\")\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n \n if load_ln_type:\n load_ln_type = f\"{load_ln_type}_ln\" if \"vision\" not in load_ln_type else \"ln_vision\"\n loaded_state_dict = {}\n if 'model' in checkpoint:\n checkpoint = checkpoint['model'] \n for k in checkpoint.keys():\n if load_ln_type in k:\n loaded_state_dict['.'.join(k.split('.')[1:])] = checkpoint[k]\n ln.load_state_dict(loaded_state_dict, strict=False)\n \n return ln\n\n @classmethod\n def init_audio_encoder(self, \n model_name, cached_audio, load_ln_path=False, load_ln_type=\"\"):\n assert model_name in [\n 'beats'\n ], \"audio model must be in [beats]\"\n\n # load_ln_path = kwargs['load_ln_path']\n # del kwargs['load_ln_path']\n # load_ln_type=kwargs['load_ln_type']\n # del kwargs['load_ln_type']\n kwargs = {}\n if \"beats\" in model_name:\n from lavis.models.beats_encoder import BeatsEncoder\n if cached_audio:\n audio_encoder = lambda x: x\n ln_audio = self.init_ln(768, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n else:\n audio_encoder = BeatsEncoder(**kwargs)","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_audio_encoder","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_audio_encoder#L328-L351","kind":"function","name":"init_audio_encoder","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":328,"end_line":351,"context_start_line":308,"context_end_line":371,"code":" )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n \n if load_ln_type:\n load_ln_type = f\"{load_ln_type}_ln\" if \"vision\" not in load_ln_type else \"ln_vision\"\n loaded_state_dict = {}\n if 'model' in checkpoint:\n checkpoint = checkpoint['model'] \n for k in checkpoint.keys():\n if load_ln_type in k:\n loaded_state_dict['.'.join(k.split('.')[1:])] = checkpoint[k]\n ln.load_state_dict(loaded_state_dict, strict=False)\n \n return ln\n\n @classmethod\n def init_audio_encoder(self, \n model_name, cached_audio, load_ln_path=False, load_ln_type=\"\"):\n assert model_name in [\n 'beats'\n ], \"audio model must be in [beats]\"\n\n # load_ln_path = kwargs['load_ln_path']\n # del kwargs['load_ln_path']\n # load_ln_type=kwargs['load_ln_type']\n # del kwargs['load_ln_type']\n kwargs = {}\n if \"beats\" in model_name:\n from lavis.models.beats_encoder import BeatsEncoder\n if cached_audio:\n audio_encoder = lambda x: x\n ln_audio = self.init_ln(768, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n else:\n audio_encoder = BeatsEncoder(**kwargs)\n\n if not cached_audio:\n ln_audio = self.init_ln(audio_encoder.num_features, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n self.audio_enc_name = model_name\n\n return audio_encoder, ln_audio\n \n @classmethod\n def init_TemporalQFormer(cls, num_of_frame):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.query_length = num_of_frame\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_TemporalQFormer","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_TemporalQFormer#L354-L364","kind":"function","name":"init_TemporalQFormer","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":354,"end_line":364,"context_start_line":334,"context_end_line":384,"code":" # load_ln_path = kwargs['load_ln_path']\n # del kwargs['load_ln_path']\n # load_ln_type=kwargs['load_ln_type']\n # del kwargs['load_ln_type']\n kwargs = {}\n if \"beats\" in model_name:\n from lavis.models.beats_encoder import BeatsEncoder\n if cached_audio:\n audio_encoder = lambda x: x\n ln_audio = self.init_ln(768, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n else:\n audio_encoder = BeatsEncoder(**kwargs)\n\n if not cached_audio:\n ln_audio = self.init_ln(audio_encoder.num_features, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n self.audio_enc_name = model_name\n\n return audio_encoder, ln_audio\n \n @classmethod\n def init_TemporalQFormer(cls, num_of_frame):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.query_length = num_of_frame\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) ","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_vision_encoder","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_vision_encoder#L367-L374","kind":"function","name":"init_vision_encoder","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":367,"end_line":374,"context_start_line":347,"context_end_line":394,"code":" if not cached_audio:\n ln_audio = self.init_ln(audio_encoder.num_features, load_ln_path=load_ln_path, load_ln_type=load_ln_type)\n self.audio_enc_name = model_name\n\n return audio_encoder, ln_audio\n \n @classmethod\n def init_TemporalQFormer(cls, num_of_frame):\n encoder_config = BertConfig.from_pretrained(\"bert-base-uncased\")\n encoder_config.query_length = num_of_frame\n Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_vision_encoder_sevila","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_vision_encoder_sevila#L377-L385","kind":"function","name":"init_vision_encoder_sevila","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":377,"end_line":385,"context_start_line":357,"context_end_line":405,"code":" Qformer = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=encoder_config\n ) \n query_tokens = nn.Parameter(\n torch.zeros(1, num_of_frame, 1, encoder_config.hidden_size)\n )\n query_tokens.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n return Qformer, query_tokens\n\n @classmethod\n def init_vision_encoder(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.init_vision_encoder_only","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.init_vision_encoder_only#L388-L394","kind":"function","name":"init_vision_encoder_only","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":388,"end_line":394,"context_start_line":368,"context_end_line":414,"code":" cls, img_size, drop_path_rate, use_grad_checkpoint, precision\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n return visual_encoder, ln_vision\n\n @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.load_from_pretrained","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.load_from_pretrained#L396-L414","kind":"function","name":"load_from_pretrained","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":396,"end_line":414,"context_start_line":376,"context_end_line":434,"code":" @classmethod\n def init_vision_encoder_sevila(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n ln_vision = LayerNorm(visual_encoder.num_features)\n ln_vision2 = LayerNorm(visual_encoder.num_features) \n return visual_encoder, ln_vision, ln_vision2\n\n @classmethod\n def init_vision_encoder_only(\n cls, img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans=3\n ):\n visual_encoder = create_eva_vit_g(\n img_size, drop_path_rate, use_grad_checkpoint, precision, in_chans\n )\n return visual_encoder\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2.load_lora","uri":"program://CREMA/function/lavis.models.blip2_models.blip2.load_lora#L417-L423","kind":"function","name":"load_lora","path":"lavis/models/blip2_models/blip2.py","language":"python","start_line":417,"end_line":423,"context_start_line":397,"context_end_line":443,"code":" if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n #print('state_dict',state_dict.keys())\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg\n \n \n def load_lora(self, lora_ckpt):\n if lora_ckpt=='':\n return\n checkpoint = torch.load(lora_ckpt, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % lora_ckpt)\n \n\ndef disabled_train(self, mode=True):\n \"\"\"Overwrite model.train with this function to make sure train/eval mode\n does not change anymore.\"\"\"\n return self\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\ndef compute_sim_matrix(model, data_loader, **kwargs):\n k_test = kwargs.pop(\"k_test\")\n","source_hash":"5c0814717288af186eef99b3250fb487bbcd5a2044e39aa48b0731fc51d6a90d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer","uri":"program://CREMA/module/lavis.models.blip2_models.blip2_qformer#L1-L507","kind":"module","name":"lavis.models.blip2_models.blip2_qformer","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":1,"end_line":507,"context_start_line":1,"context_end_line":507,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport torch\nimport torch.distributed as dist\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom torch.nn import functional as F\n\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import all_gather_with_grad, concat_all_gather\nfrom lavis.models.blip2_models.blip2 import (\n Blip2Base,\n compute_sim_matrix,\n disabled_train,\n)\nfrom lavis.models.blip_models.blip_outputs import BlipOutput, BlipOutputFeatures\n\n\n\n@registry.register_model(\"blip2\")\n@registry.register_model(\"blip2_feature_extractor\")\nclass Blip2Qformer(Blip2Base):\n \"\"\"\n BLIP2 first-stage model with Q-former and ViT.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2\", \"pretrain\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain\": \"configs/models/blip2/blip2_pretrain.yaml\",\n \"coco\": \"configs/models/blip2/blip2_coco.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train \n logging.info(\"freeze vision encoder\")\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.resize_token_embeddings(len(self.tokenizer))\n state_dict = self.Qformer.state_dict()\n for name, param in self.Qformer.named_parameters():\n if \"_query\" in name:\n key_orig = name.replace(\"_query\", \"\")\n param.data.copy_(state_dict[key_orig])\n\n self.vision_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n self.text_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n\n self.itm_head = nn.Linear(self.Qformer.config.hidden_size, 2)\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n image = samples[\"image\"]\n text = samples[\"text_input\"]\n \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n use_cache=True,\n return_dict=True,\n )\n\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_tokens = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n ###============== Image-text Contrastive ===================###\n image_feats_all = concat_all_gather(\n image_feats\n ) # [batch_size*num_gpu, num_query_tokens, embed_dim]\n text_feat_all = concat_all_gather(text_feat) # [batch_size*num_gpu, embed_dim]\n\n sim_q2t = torch.matmul(\n image_feats.unsqueeze(1), text_feat_all.unsqueeze(-1)\n ).squeeze()\n # [batch_size, batch_size*num_gpu, num_query_tokens]\n\n # image-text similarity: aggregate across all query tokens\n sim_i2t, _ = sim_q2t.max(-1)\n sim_i2t = sim_i2t / self.temp\n\n # text-query similarity: [batch_size, batch_size*num_gpu, num_query_tokens]\n sim_t2q = torch.matmul(\n text_feat.unsqueeze(1).unsqueeze(1), image_feats_all.permute(0, 2, 1)\n ).squeeze()\n\n # text-image similarity: aggregate across all query tokens\n sim_t2i, _ = sim_t2q.max(-1)\n sim_t2i = sim_t2i / self.temp # [batch_size, batch_size*num_gpu]\n\n rank = dist.get_rank()\n bs = image.size(0)\n targets = torch.linspace(rank * bs, rank * bs + bs - 1, bs, dtype=int).to(\n image.device\n )\n\n loss_itc = (\n F.cross_entropy(sim_i2t, targets, label_smoothing=0.1)\n + F.cross_entropy(sim_t2i, targets, label_smoothing=0.1)\n ) / 2\n\n ###============== Image-text Matching ===================###\n text_input_ids_world = concat_all_gather(text_tokens.input_ids)\n text_attention_mask_world = concat_all_gather(text_tokens.attention_mask)\n image_embeds_world = all_gather_with_grad(image_embeds)\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i, dim=1) + 1e-4\n weights_t2i[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t, dim=1) + 1e-4\n weights_i2t[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(text_input_ids_world[neg_idx])\n text_atts_neg.append(text_attention_mask_world[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat(\n [text_tokens.input_ids, text_tokens.input_ids, text_ids_neg], dim=0\n ) # pos, pos, neg\n text_atts_all = torch.cat(\n [text_tokens.attention_mask, text_tokens.attention_mask, text_atts_neg],\n dim=0,\n )\n\n query_tokens_itm = self.query_tokens.expand(text_ids_all.shape[0], -1, -1)\n query_atts_itm = torch.ones(query_tokens_itm.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask_all = torch.cat([query_atts_itm, text_atts_all], dim=1)\n\n image_embeds_all = torch.cat(\n [image_embeds, image_embeds_neg, image_embeds], dim=0\n ) # pos, neg, pos\n image_atts_all = torch.ones(image_embeds_all.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n output_itm = self.Qformer.bert(\n text_ids_all,\n query_embeds=query_tokens_itm,\n attention_mask=attention_mask_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens_itm.size(1), :]\n vl_output = self.itm_head(vl_embeddings)\n logits = vl_output.mean(dim=1)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(logits, itm_labels)\n\n ##================= Image Captioning ========================##\n decoder_input_ids = text_tokens.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n labels = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text_tokens.attention_mask], dim=1)\n lm_output = self.Qformer(\n decoder_input_ids,\n attention_mask=attention_mask,\n past_key_values=query_output.past_key_values,\n return_dict=True,\n labels=labels,\n )\n \n loss_lm = lm_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n )\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n\n if not use_nucleus_sampling:\n image_embeds = image_embeds.repeat_interleave(num_beams, dim=0)\n else:\n num_beams = 1\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n input_ids = (\n torch.LongTensor(image.size(0), 1)\n .fill_(self.tokenizer.bos_token_id)\n .to(image.device)\n )\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n outputs = self.Qformer.generate(\n input_ids=input_ids,\n query_embeds=query_tokens,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)\n return captions\n\n def forward_image(self, image):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n return text_output.last_hidden_state[:, 0, :]\n\n def compute_itm(self, image_inputs, text_ids, text_atts):\n image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_inputs,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(vl_embeddings)\n itm_logit = itm_logit[:, :, 1].mean(dim=1)\n return itm_logit\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_embeds = query_output.last_hidden_state\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n # return text features\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_embeds = text_output.last_hidden_state\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n\n output = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state[:, : query_tokens.size(1), :]\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.Blip2Qformer","uri":"program://CREMA/class/lavis.models.blip2_models.blip2_qformer.Blip2Qformer#L28-L507","kind":"class","name":"Blip2Qformer","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":28,"end_line":507,"context_start_line":8,"context_end_line":507,"code":"\nimport torch\nimport torch.distributed as dist\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom torch.nn import functional as F\n\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import all_gather_with_grad, concat_all_gather\nfrom lavis.models.blip2_models.blip2 import (\n Blip2Base,\n compute_sim_matrix,\n disabled_train,\n)\nfrom lavis.models.blip_models.blip_outputs import BlipOutput, BlipOutputFeatures\n\n\n\n@registry.register_model(\"blip2\")\n@registry.register_model(\"blip2_feature_extractor\")\nclass Blip2Qformer(Blip2Base):\n \"\"\"\n BLIP2 first-stage model with Q-former and ViT.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2\", \"pretrain\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain\": \"configs/models/blip2/blip2_pretrain.yaml\",\n \"coco\": \"configs/models/blip2/blip2_coco.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train \n logging.info(\"freeze vision encoder\")\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.resize_token_embeddings(len(self.tokenizer))\n state_dict = self.Qformer.state_dict()\n for name, param in self.Qformer.named_parameters():\n if \"_query\" in name:\n key_orig = name.replace(\"_query\", \"\")\n param.data.copy_(state_dict[key_orig])\n\n self.vision_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n self.text_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n\n self.itm_head = nn.Linear(self.Qformer.config.hidden_size, 2)\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n image = samples[\"image\"]\n text = samples[\"text_input\"]\n \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n use_cache=True,\n return_dict=True,\n )\n\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_tokens = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n ###============== Image-text Contrastive ===================###\n image_feats_all = concat_all_gather(\n image_feats\n ) # [batch_size*num_gpu, num_query_tokens, embed_dim]\n text_feat_all = concat_all_gather(text_feat) # [batch_size*num_gpu, embed_dim]\n\n sim_q2t = torch.matmul(\n image_feats.unsqueeze(1), text_feat_all.unsqueeze(-1)\n ).squeeze()\n # [batch_size, batch_size*num_gpu, num_query_tokens]\n\n # image-text similarity: aggregate across all query tokens\n sim_i2t, _ = sim_q2t.max(-1)\n sim_i2t = sim_i2t / self.temp\n\n # text-query similarity: [batch_size, batch_size*num_gpu, num_query_tokens]\n sim_t2q = torch.matmul(\n text_feat.unsqueeze(1).unsqueeze(1), image_feats_all.permute(0, 2, 1)\n ).squeeze()\n\n # text-image similarity: aggregate across all query tokens\n sim_t2i, _ = sim_t2q.max(-1)\n sim_t2i = sim_t2i / self.temp # [batch_size, batch_size*num_gpu]\n\n rank = dist.get_rank()\n bs = image.size(0)\n targets = torch.linspace(rank * bs, rank * bs + bs - 1, bs, dtype=int).to(\n image.device\n )\n\n loss_itc = (\n F.cross_entropy(sim_i2t, targets, label_smoothing=0.1)\n + F.cross_entropy(sim_t2i, targets, label_smoothing=0.1)\n ) / 2\n\n ###============== Image-text Matching ===================###\n text_input_ids_world = concat_all_gather(text_tokens.input_ids)\n text_attention_mask_world = concat_all_gather(text_tokens.attention_mask)\n image_embeds_world = all_gather_with_grad(image_embeds)\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i, dim=1) + 1e-4\n weights_t2i[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t, dim=1) + 1e-4\n weights_i2t[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(text_input_ids_world[neg_idx])\n text_atts_neg.append(text_attention_mask_world[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat(\n [text_tokens.input_ids, text_tokens.input_ids, text_ids_neg], dim=0\n ) # pos, pos, neg\n text_atts_all = torch.cat(\n [text_tokens.attention_mask, text_tokens.attention_mask, text_atts_neg],\n dim=0,\n )\n\n query_tokens_itm = self.query_tokens.expand(text_ids_all.shape[0], -1, -1)\n query_atts_itm = torch.ones(query_tokens_itm.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask_all = torch.cat([query_atts_itm, text_atts_all], dim=1)\n\n image_embeds_all = torch.cat(\n [image_embeds, image_embeds_neg, image_embeds], dim=0\n ) # pos, neg, pos\n image_atts_all = torch.ones(image_embeds_all.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n output_itm = self.Qformer.bert(\n text_ids_all,\n query_embeds=query_tokens_itm,\n attention_mask=attention_mask_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens_itm.size(1), :]\n vl_output = self.itm_head(vl_embeddings)\n logits = vl_output.mean(dim=1)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(logits, itm_labels)\n\n ##================= Image Captioning ========================##\n decoder_input_ids = text_tokens.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n labels = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text_tokens.attention_mask], dim=1)\n lm_output = self.Qformer(\n decoder_input_ids,\n attention_mask=attention_mask,\n past_key_values=query_output.past_key_values,\n return_dict=True,\n labels=labels,\n )\n \n loss_lm = lm_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n )\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n\n if not use_nucleus_sampling:\n image_embeds = image_embeds.repeat_interleave(num_beams, dim=0)\n else:\n num_beams = 1\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n input_ids = (\n torch.LongTensor(image.size(0), 1)\n .fill_(self.tokenizer.bos_token_id)\n .to(image.device)\n )\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n outputs = self.Qformer.generate(\n input_ids=input_ids,\n query_embeds=query_tokens,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)\n return captions\n\n def forward_image(self, image):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n return text_output.last_hidden_state[:, 0, :]\n\n def compute_itm(self, image_inputs, text_ids, text_atts):\n image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_inputs,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(vl_embeddings)\n itm_logit = itm_logit[:, :, 1].mean(dim=1)\n return itm_logit\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_embeds = query_output.last_hidden_state\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n # return text features\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_embeds = text_output.last_hidden_state\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n\n output = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state[:, : query_tokens.size(1), :]\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.__init__#L44-L85","kind":"function","name":"__init__","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":44,"end_line":85,"context_start_line":24,"context_end_line":105,"code":"\n\n@registry.register_model(\"blip2\")\n@registry.register_model(\"blip2_feature_extractor\")\nclass Blip2Qformer(Blip2Base):\n \"\"\"\n BLIP2 first-stage model with Q-former and ViT.\n Supported model types:\n - pretrained: pretrained model\n - coco: fintuned model on coco\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2\", \"pretrain\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain\": \"configs/models/blip2/blip2_pretrain.yaml\",\n \"coco\": \"configs/models/blip2/blip2_coco.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n embed_dim=256,\n max_txt_len=32,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train \n logging.info(\"freeze vision encoder\")\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.resize_token_embeddings(len(self.tokenizer))\n state_dict = self.Qformer.state_dict()\n for name, param in self.Qformer.named_parameters():\n if \"_query\" in name:\n key_orig = name.replace(\"_query\", \"\")\n param.data.copy_(state_dict[key_orig])\n\n self.vision_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n self.text_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n\n self.itm_head = nn.Linear(self.Qformer.config.hidden_size, 2)\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n image = samples[\"image\"]\n text = samples[\"text_input\"]\n \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n use_cache=True,\n return_dict=True,\n )\n","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.forward","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.forward#L87-L255","kind":"function","name":"forward","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":87,"end_line":255,"context_start_line":67,"context_end_line":275,"code":" logging.info(\"freeze vision encoder\")\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.resize_token_embeddings(len(self.tokenizer))\n state_dict = self.Qformer.state_dict()\n for name, param in self.Qformer.named_parameters():\n if \"_query\" in name:\n key_orig = name.replace(\"_query\", \"\")\n param.data.copy_(state_dict[key_orig])\n\n self.vision_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n self.text_proj = nn.Linear(self.Qformer.config.hidden_size, embed_dim)\n\n self.itm_head = nn.Linear(self.Qformer.config.hidden_size, 2)\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n image = samples[\"image\"]\n text = samples[\"text_input\"]\n \n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n use_cache=True,\n return_dict=True,\n )\n\n image_feats = F.normalize(\n self.vision_proj(query_output.last_hidden_state), dim=-1\n )\n\n text_tokens = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n ###============== Image-text Contrastive ===================###\n image_feats_all = concat_all_gather(\n image_feats\n ) # [batch_size*num_gpu, num_query_tokens, embed_dim]\n text_feat_all = concat_all_gather(text_feat) # [batch_size*num_gpu, embed_dim]\n\n sim_q2t = torch.matmul(\n image_feats.unsqueeze(1), text_feat_all.unsqueeze(-1)\n ).squeeze()\n # [batch_size, batch_size*num_gpu, num_query_tokens]\n\n # image-text similarity: aggregate across all query tokens\n sim_i2t, _ = sim_q2t.max(-1)\n sim_i2t = sim_i2t / self.temp\n\n # text-query similarity: [batch_size, batch_size*num_gpu, num_query_tokens]\n sim_t2q = torch.matmul(\n text_feat.unsqueeze(1).unsqueeze(1), image_feats_all.permute(0, 2, 1)\n ).squeeze()\n\n # text-image similarity: aggregate across all query tokens\n sim_t2i, _ = sim_t2q.max(-1)\n sim_t2i = sim_t2i / self.temp # [batch_size, batch_size*num_gpu]\n\n rank = dist.get_rank()\n bs = image.size(0)\n targets = torch.linspace(rank * bs, rank * bs + bs - 1, bs, dtype=int).to(\n image.device\n )\n\n loss_itc = (\n F.cross_entropy(sim_i2t, targets, label_smoothing=0.1)\n + F.cross_entropy(sim_t2i, targets, label_smoothing=0.1)\n ) / 2\n\n ###============== Image-text Matching ===================###\n text_input_ids_world = concat_all_gather(text_tokens.input_ids)\n text_attention_mask_world = concat_all_gather(text_tokens.attention_mask)\n image_embeds_world = all_gather_with_grad(image_embeds)\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i, dim=1) + 1e-4\n weights_t2i[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t, dim=1) + 1e-4\n weights_i2t[:, rank * bs : rank * bs + bs].fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(text_input_ids_world[neg_idx])\n text_atts_neg.append(text_attention_mask_world[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat(\n [text_tokens.input_ids, text_tokens.input_ids, text_ids_neg], dim=0\n ) # pos, pos, neg\n text_atts_all = torch.cat(\n [text_tokens.attention_mask, text_tokens.attention_mask, text_atts_neg],\n dim=0,\n )\n\n query_tokens_itm = self.query_tokens.expand(text_ids_all.shape[0], -1, -1)\n query_atts_itm = torch.ones(query_tokens_itm.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask_all = torch.cat([query_atts_itm, text_atts_all], dim=1)\n\n image_embeds_all = torch.cat(\n [image_embeds, image_embeds_neg, image_embeds], dim=0\n ) # pos, neg, pos\n image_atts_all = torch.ones(image_embeds_all.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n output_itm = self.Qformer.bert(\n text_ids_all,\n query_embeds=query_tokens_itm,\n attention_mask=attention_mask_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens_itm.size(1), :]\n vl_output = self.itm_head(vl_embeddings)\n logits = vl_output.mean(dim=1)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(logits, itm_labels)\n\n ##================= Image Captioning ========================##\n decoder_input_ids = text_tokens.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n labels = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image.device\n )\n attention_mask = torch.cat([query_atts, text_tokens.attention_mask], dim=1)\n lm_output = self.Qformer(\n decoder_input_ids,\n attention_mask=attention_mask,\n past_key_values=query_output.past_key_values,\n return_dict=True,\n labels=labels,\n )\n \n loss_lm = lm_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n )\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.generate","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.generate#L258-L318","kind":"function","name":"generate","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":258,"end_line":318,"context_start_line":238,"context_end_line":338,"code":" )\n attention_mask = torch.cat([query_atts, text_tokens.attention_mask], dim=1)\n lm_output = self.Qformer(\n decoder_input_ids,\n attention_mask=attention_mask,\n past_key_values=query_output.past_key_values,\n return_dict=True,\n labels=labels,\n )\n \n loss_lm = lm_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n )\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n\n if not use_nucleus_sampling:\n image_embeds = image_embeds.repeat_interleave(num_beams, dim=0)\n else:\n num_beams = 1\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n input_ids = (\n torch.LongTensor(image.size(0), 1)\n .fill_(self.tokenizer.bos_token_id)\n .to(image.device)\n )\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n outputs = self.Qformer.generate(\n input_ids=input_ids,\n query_embeds=query_tokens,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)\n return captions\n\n def forward_image(self, image):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.forward_image","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.forward_image#L320-L334","kind":"function","name":"forward_image","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":320,"end_line":334,"context_start_line":300,"context_end_line":354,"code":" .fill_(self.tokenizer.bos_token_id)\n .to(image.device)\n )\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n outputs = self.Qformer.generate(\n input_ids=input_ids,\n query_embeds=query_tokens,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)\n return captions\n\n def forward_image(self, image):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n return text_output.last_hidden_state[:, 0, :]\n\n def compute_itm(self, image_inputs, text_ids, text_atts):\n image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.forward_text","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.forward_text#L336-L342","kind":"function","name":"forward_text","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":336,"end_line":342,"context_start_line":316,"context_end_line":362,"code":" )\n captions = self.tokenizer.batch_decode(outputs, skip_special_tokens=True)\n return captions\n\n def forward_image(self, image):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n return text_output.last_hidden_state[:, 0, :]\n\n def compute_itm(self, image_inputs, text_ids, text_atts):\n image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_inputs,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(vl_embeddings)","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.compute_itm","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.compute_itm#L344-L364","kind":"function","name":"compute_itm","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":344,"end_line":364,"context_start_line":324,"context_end_line":384,"code":" )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n return query_output.last_hidden_state, image_embeds\n\n def forward_text(self, text_tokens):\n text_output = self.Qformer.bert(\n text_tokens.input_ids,\n attention_mask=text_tokens.attention_mask,\n return_dict=True,\n )\n return text_output.last_hidden_state[:, 0, :]\n\n def compute_itm(self, image_inputs, text_ids, text_atts):\n image_atts = torch.ones(image_inputs.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_inputs,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(vl_embeddings)\n itm_logit = itm_logit[:, :, 1].mean(dim=1)\n return itm_logit\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n \"\"\"\n image = samples.get(\"image\")","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.extract_features","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.extract_features#L367-L474","kind":"function","name":"extract_features","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":367,"end_line":474,"context_start_line":347,"context_end_line":494,"code":" )\n query_tokens = self.query_tokens.expand(image_inputs.shape[0], -1, -1)\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n image_inputs.device\n )\n attention_mask = torch.cat([query_atts, text_atts], dim=1)\n output_itm = self.Qformer.bert(\n text_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_inputs,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n vl_embeddings = output_itm.last_hidden_state[:, : query_tokens.size(1), :]\n itm_logit = self.itm_head(vl_embeddings)\n itm_logit = itm_logit[:, :, 1].mean(dim=1)\n return itm_logit\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n image_embeds = query_output.last_hidden_state\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n # return text features\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n text_output = self.Qformer.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n )\n text_embeds = text_output.last_hidden_state\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel query features\n image_embeds_frozen = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(\n image_embeds_frozen.size()[:-1], dtype=torch.long\n ).to(self.device)\n query_tokens = self.query_tokens.expand(\n image_embeds_frozen.shape[0], -1, -1\n )\n query_atts = torch.ones(query_tokens.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n attention_mask = torch.cat([query_atts, text.attention_mask], dim=1)\n\n output = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state[:, : query_tokens.size(1), :]\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.from_config","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.from_config#L477-L499","kind":"function","name":"from_config","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":477,"end_line":499,"context_start_line":457,"context_end_line":507,"code":" output = self.Qformer.bert(\n text.input_ids,\n query_embeds=query_tokens,\n attention_mask=attention_mask,\n encoder_hidden_states=image_embeds_frozen,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state[:, : query_tokens.size(1), :]\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_qformer.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_qformer.compute_sim_matrix#L501-L507","kind":"function","name":"compute_sim_matrix","path":"lavis/models/blip2_models/blip2_qformer.py","language":"python","start_line":501,"end_line":507,"context_start_line":481,"context_end_line":507,"code":" drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n max_txt_len=max_txt_len,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"6abad4e82ddf2f7e7b7c856e6398201edbd20a3d0e3e430d5cb4f4a8e067dd42","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer","uri":"program://CREMA/module/lavis.models.blip2_models.Qformer#L1-L1251","kind":"module","name":"lavis.models.blip2_models.Qformer","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1,"end_line":1251,"context_start_line":1,"context_end_line":1251,"code":"\"\"\"\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see LICENSE.txt file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Junnan Li\n * Based on huggingface code base\n * https://github.com/huggingface/transformers/blob/v4.15.0/src/transformers/models/bert\n\"\"\"\n\nimport math\nimport os\nimport warnings\nfrom dataclasses import dataclass\nfrom typing import Optional, Tuple, Dict, Any\n\nimport torch\nfrom torch import Tensor, device, dtype, nn\nimport torch.utils.checkpoint\nfrom torch import nn\nfrom torch.nn import CrossEntropyLoss\nimport torch.nn.functional as F\n\nfrom transformers.activations import ACT2FN\nfrom transformers.file_utils import (\n ModelOutput,\n)\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPastAndCrossAttentions,\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n MaskedLMOutput,\n MultipleChoiceModelOutput,\n NextSentencePredictorOutput,\n QuestionAnsweringModelOutput,\n SequenceClassifierOutput,\n TokenClassifierOutput,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.utils import logging\nfrom transformers.models.bert.configuration_bert import BertConfig\n\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n position_ids=None,\n query_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n seq_length = input_ids.size()[1]\n else:\n seq_length = 0\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ].clone()\n\n if input_ids is not None:\n embeddings = self.word_embeddings(input_ids)\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings = embeddings + position_embeddings\n\n if query_embeds is not None:\n embeddings = torch.cat((query_embeds, embeddings), dim=1)\n else:\n embeddings = query_embeds\n\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n # adding for multimodal lora\n # if module is not None:\n # key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n # value_layer = self.transpose_for_scores(self.value(encoder_hidden_states, module))\n # attention_mask = encoder_attention_mask\n # else:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n\n elif past_key_value is not None:\n\n if modular is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states, modular))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n\n else:\n if modular is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states, modular))\n else: \n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n if modular is not None and not is_cross_attention:\n mixed_query_layer = self.query(hidden_states, modular)\n else:\n mixed_query_layer = self.query(hidden_states)\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n modular,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if (\n self.config.add_cross_attention\n and layer_num % self.config.cross_attention_freq == 0\n ):\n self.crossattention = BertAttention(\n config, is_cross_attention=self.config.add_cross_attention\n )\n self.has_cross_attention = True\n else:\n self.has_cross_attention = False\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n self.intermediate_query = BertIntermediate(config)\n self.output_query = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n query_length=0,\n modular=None\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n modular=modular\n )\n attention_output = self_attention_outputs[0]\n outputs = self_attention_outputs[1:-1]\n\n present_key_value = self_attention_outputs[-1]\n\n if query_length > 0:\n query_attention_output = attention_output[:, :query_length, :]\n\n if self.has_cross_attention:\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n cross_attention_outputs = self.crossattention(\n query_attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n modular=modular,\n )\n query_attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk_query,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n query_attention_output,\n )\n if attention_output.shape[1] > query_length:\n layer_output_text = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output[:, query_length:, :],\n )\n layer_output = torch.cat([layer_output, layer_output_text], dim=1)\n else:\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n def feed_forward_chunk_query(self, attention_output):\n intermediate_output = self.intermediate_query(attention_output)\n layer_output = self.output_query(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value\n# ... truncated ...","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertEmbeddings","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertEmbeddings#L51-L108","kind":"class","name":"BertEmbeddings","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":51,"end_line":108,"context_start_line":31,"context_end_line":128,"code":" CausalLMOutputWithCrossAttentions,\n MaskedLMOutput,\n MultipleChoiceModelOutput,\n NextSentencePredictorOutput,\n QuestionAnsweringModelOutput,\n SequenceClassifierOutput,\n TokenClassifierOutput,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.utils import logging\nfrom transformers.models.bert.configuration_bert import BertConfig\n\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n position_ids=None,\n query_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n seq_length = input_ids.size()[1]\n else:\n seq_length = 0\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ].clone()\n\n if input_ids is not None:\n embeddings = self.word_embeddings(input_ids)\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings = embeddings + position_embeddings\n\n if query_embeds is not None:\n embeddings = torch.cat((query_embeds, embeddings), dim=1)\n else:\n embeddings = query_embeds\n\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertSelfAttention","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertSelfAttention#L111-L298","kind":"class","name":"BertSelfAttention","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":111,"end_line":298,"context_start_line":91,"context_end_line":318,"code":" position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ].clone()\n\n if input_ids is not None:\n embeddings = self.word_embeddings(input_ids)\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings = embeddings + position_embeddings\n\n if query_embeds is not None:\n embeddings = torch.cat((query_embeds, embeddings), dim=1)\n else:\n embeddings = query_embeds\n\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n # adding for multimodal lora\n # if module is not None:\n # key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n # value_layer = self.transpose_for_scores(self.value(encoder_hidden_states, module))\n # attention_mask = encoder_attention_mask\n # else:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n\n elif past_key_value is not None:\n\n if modular is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states, modular))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n\n else:\n if modular is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states, modular))\n else: \n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n if modular is not None and not is_cross_attention:\n mixed_query_layer = self.query(hidden_states, modular)\n else:\n mixed_query_layer = self.query(hidden_states)\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertSelfOutput","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertSelfOutput#L301-L312","kind":"class","name":"BertSelfOutput","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":301,"end_line":312,"context_start_line":281,"context_end_line":332,"code":" attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertAttention","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertAttention#L315-L371","kind":"class","name":"BertAttention","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":315,"end_line":371,"context_start_line":295,"context_end_line":391,"code":" )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n modular,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertIntermediate","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertIntermediate#L374-L386","kind":"class","name":"BertIntermediate","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":374,"end_line":386,"context_start_line":354,"context_end_line":406,"code":" modular=None\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n modular,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertOutput","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertOutput#L389-L400","kind":"class","name":"BertOutput","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":389,"end_line":400,"context_start_line":369,"context_end_line":420,"code":" 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if (\n self.config.add_cross_attention\n and layer_num % self.config.cross_attention_freq == 0\n ):\n self.crossattention = BertAttention(\n config, is_cross_attention=self.config.add_cross_attention\n )\n self.has_cross_attention = True\n else:\n self.has_cross_attention = False","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertLayer","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertLayer#L403-L512","kind":"class","name":"BertLayer","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":403,"end_line":512,"context_start_line":383,"context_end_line":532,"code":" def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if (\n self.config.add_cross_attention\n and layer_num % self.config.cross_attention_freq == 0\n ):\n self.crossattention = BertAttention(\n config, is_cross_attention=self.config.add_cross_attention\n )\n self.has_cross_attention = True\n else:\n self.has_cross_attention = False\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n self.intermediate_query = BertIntermediate(config)\n self.output_query = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n query_length=0,\n modular=None\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n modular=modular\n )\n attention_output = self_attention_outputs[0]\n outputs = self_attention_outputs[1:-1]\n\n present_key_value = self_attention_outputs[-1]\n\n if query_length > 0:\n query_attention_output = attention_output[:, :query_length, :]\n\n if self.has_cross_attention:\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n cross_attention_outputs = self.crossattention(\n query_attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n modular=modular,\n )\n query_attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk_query,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n query_attention_output,\n )\n if attention_output.shape[1] > query_length:\n layer_output_text = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output[:, query_length:, :],\n )\n layer_output = torch.cat([layer_output, layer_output_text], dim=1)\n else:\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n def feed_forward_chunk_query(self, attention_output):\n intermediate_output = self.intermediate_query(attention_output)\n layer_output = self.output_query(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertEncoder","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertEncoder#L515-L620","kind":"class","name":"BertEncoder","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":515,"end_line":620,"context_start_line":495,"context_end_line":640,"code":" self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n def feed_forward_chunk_query(self, attention_output):\n intermediate_output = self.intermediate_query(attention_output)\n layer_output = self.output_query(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,\n output_hidden_states=False,\n return_dict=True,\n query_length=0,\n modular=None,\n ):\n all_hidden_states = () if output_hidden_states else None\n all_self_attentions = () if output_attentions else None\n all_cross_attentions = (\n () if output_attentions and self.config.add_cross_attention else None\n )\n\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if getattr(self.config, \"gradient_checkpointing\", False) and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(\n *inputs, past_key_value, output_attentions, query_length\n )\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n modular\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n query_length,\n modular,\n )\n\n hidden_states = layer_outputs[0]\n if use_cache:\n next_decoder_cache += (layer_outputs[-1],)\n if output_attentions:\n all_self_attentions = all_self_attentions + (layer_outputs[1],)\n all_cross_attentions = all_cross_attentions + (layer_outputs[2],)\n\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if not return_dict:\n return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertPooler","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertPooler#L623-L635","kind":"class","name":"BertPooler","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":623,"end_line":635,"context_start_line":603,"context_end_line":655,"code":" return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertPredictionHeadTransform","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertPredictionHeadTransform#L638-L652","kind":"class","name":"BertPredictionHeadTransform","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":638,"end_line":652,"context_start_line":618,"context_end_line":672,"code":" attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertLMPredictionHead","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertLMPredictionHead#L655-L672","kind":"class","name":"BertLMPredictionHead","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":655,"end_line":672,"context_start_line":635,"context_end_line":692,"code":" return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertOnlyMLMHead","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertOnlyMLMHead#L675-L682","kind":"class","name":"BertOnlyMLMHead","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":675,"end_line":682,"context_start_line":655,"context_end_line":702,"code":"class BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertPreTrainedModel","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertPreTrainedModel#L685-L705","kind":"class","name":"BertPreTrainedModel","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":685,"end_line":705,"context_start_line":665,"context_end_line":725,"code":"\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertModel","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertModel#L708-L998","kind":"class","name":"BertModel","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":708,"end_line":998,"context_start_line":688,"context_end_line":1018,"code":" models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n has_query: bool = False,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n\n # add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n if has_query: # UniLM style attention mask\n causal_mask = torch.cat(\n [\n torch.zeros(\n (batch_size, prefix_seq_len, seq_length),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=1,\n )\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, causal_mask.shape[1], prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n modular=None\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n \"\"\"\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n # use_cache = use_cache if use_cache is not None else self.config.use_cache\n\n if input_ids is None:\n assert (\n query_embeds is not None\n ), \"You have to specify query_embeds when input_ids is None\"\n\n # past_key_values_length\n past_key_values_length = (\n past_key_values[0][0].shape[2] - self.config.query_length\n if past_key_values is not None\n else 0\n )\n\n query_length = query_embeds.shape[1] if query_embeds is not None else 0\n\n embedding_output = self.embeddings(\n input_ids=input_ids,\n position_ids=position_ids,\n query_embeds=query_embeds,\n past_key_values_length=past_key_values_length,\n )\n\n input_shape = embedding_output.size()[:-1]\n batch_size, seq_length = input_shape\n device = embedding_output.device\n\n if attention_mask is None:\n attention_mask = torch.ones(\n ((batch_size, seq_length + past_key_values_length)), device=device\n )\n\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if is_decoder:\n extended_attention_mask = self.get_extended_attention_mask(\n attention_mask,\n input_ids.shape,\n device,\n is_decoder,\n has_query=(query_embeds is not None),\n )\n else:\n extended_attention_mask = self.get_extended_attention_mask(\n attention_mask, input_shape, device, is_decoder\n )\n\n # If a 2D or 3D attention mask is provided for the cross-attention\n # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if encoder_hidden_states is not None:\n if type(encoder_hidden_states) == list:\n encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[\n 0\n ].size()\n else:\n (\n encoder_batch_size,\n encoder_sequence_length,\n _,\n ) = encoder_hidden_states.size()\n encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)\n\n if type(encoder_attention_mask) == list:\n encoder_extended_attention_mask = [\n self.invert_attention_mask(mask) for mask in encoder_attention_mask\n ]\n elif encoder_attention_mask is None:\n encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = None\n\n # Prepare head mask if needed\n # 1.0 in head_mask indicate we keep the head\n # attention_probs has shape bsz x n_heads x N x N\n # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]\n # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]\n head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)\n\n encoder_outputs = self.encoder(\n embedding_output,\n attention_mask=extended_attention_mask,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n query_length=query_length,\n modular=modular,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertLMHeadModel","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertLMHeadModel#L1001-L1163","kind":"class","name":"BertLMHeadModel","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1001,"end_line":1163,"context_start_line":981,"context_end_line":1183,"code":" modular=modular,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )\n\n\nclass BertLMHeadModel(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n past_key_values=None,\n use_cache=True,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=True,\n reduction=\"mean\",\n modular=None, # for multimodal lora\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in\n ``[-100, 0, ..., config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are\n ignored (masked), the loss is only computed for the tokens with labels n ``[0, ..., config.vocab_size]``\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n Returns:\n Example::\n >>> from transformers import BertTokenizer, BertLMHeadModel, BertConfig\n >>> import torch\n >>> tokenizer = BertTokenizer.from_pretrained('bert-base-cased')\n >>> config = BertConfig.from_pretrained(\"bert-base-cased\")\n >>> model = BertLMHeadModel.from_pretrained('bert-base-cased', config=config)\n >>> inputs = tokenizer(\"Hello, my dog is cute\", return_tensors=\"pt\")\n >>> outputs = model(**inputs)\n >>> prediction_logits = outputs.logits\n \"\"\"\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n if labels is not None:\n use_cache = False\n if past_key_values is not None:\n query_embeds = None\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n position_ids=position_ids,\n head_mask=head_mask,\n query_embeds=query_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n modular=modular,\n )\n\n sequence_output = outputs[0]\n if query_embeds is not None:\n sequence_output = outputs[0][:, query_embeds.shape[1] :, :]\n\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores[:, :-1, :].contiguous()\n\n lm_loss = None\n if labels is not None:\n # we are doing next-token prediction; shift prediction scores and input ids by one\n shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()\n labels = labels[:, 1:].contiguous()\n loss_fct = CrossEntropyLoss(reduction=reduction, label_smoothing=0.1)\n lm_loss = loss_fct(\n shifted_prediction_scores.view(-1, self.config.vocab_size),\n labels.view(-1),\n )\n if reduction == \"none\":\n lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return ((lm_loss,) + output) if lm_loss is not None else output\n\n return CausalLMOutputWithCrossAttentions(\n loss=lm_loss,\n logits=prediction_scores,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n cross_attentions=outputs.cross_attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, query_embeds, past=None, attention_mask=None, **model_kwargs\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n query_mask = input_ids.new_ones(query_embeds.shape[:-1])\n attention_mask = torch.cat([query_mask, attention_mask], dim=-1)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.BertForMaskedLM","uri":"program://CREMA/class/lavis.models.blip2_models.Qformer.BertForMaskedLM#L1166-L1251","kind":"class","name":"BertForMaskedLM","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1166,"end_line":1251,"context_start_line":1146,"context_end_line":1251,"code":" \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=False,\n ):\n r\"\"\"\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,\n config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored\n (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``\n \"\"\"\n\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n position_ids=position_ids,\n head_mask=head_mask,\n query_embeds=query_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n )\n\n if query_embeds is not None:\n sequence_output = outputs[0][:, query_embeds.shape[1] :, :]\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores\n\n masked_lm_loss = None\n if labels is not None:\n loss_fct = CrossEntropyLoss() # -100 index = padding token\n masked_lm_loss = loss_fct(\n prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)\n )\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return (\n ((masked_lm_loss,) + output) if masked_lm_loss is not None else output\n )\n\n return MaskedLMOutput(\n loss=masked_lm_loss,\n logits=prediction_scores,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.__init__#L1171-L1177","kind":"function","name":"__init__","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1171,"end_line":1177,"context_start_line":1151,"context_end_line":1197,"code":" \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.forward","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.forward#L1185-L1251","kind":"function","name":"forward","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1185,"end_line":1251,"context_start_line":1165,"context_end_line":1251,"code":"\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=False,\n ):\n r\"\"\"\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,\n config.vocab_size]`` (see ``input_ids`` docstring) Tokens with indices set to ``-100`` are ignored\n (masked), the loss is only computed for the tokens with labels in ``[0, ..., config.vocab_size]``\n \"\"\"\n\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n outputs = self.bert(\n input_ids,\n attention_mask=attention_mask,\n position_ids=position_ids,\n head_mask=head_mask,\n query_embeds=query_embeds,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n is_decoder=is_decoder,\n )\n\n if query_embeds is not None:\n sequence_output = outputs[0][:, query_embeds.shape[1] :, :]\n prediction_scores = self.cls(sequence_output)\n\n if return_logits:\n return prediction_scores\n\n masked_lm_loss = None\n if labels is not None:\n loss_fct = CrossEntropyLoss() # -100 index = padding token\n masked_lm_loss = loss_fct(\n prediction_scores.view(-1, self.config.vocab_size), labels.view(-1)\n )\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return (\n ((masked_lm_loss,) + output) if masked_lm_loss is not None else output\n )\n\n return MaskedLMOutput(\n loss=masked_lm_loss,\n logits=prediction_scores,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n )","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.save_attn_gradients","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.save_attn_gradients#L149-L150","kind":"function","name":"save_attn_gradients","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":149,"end_line":150,"context_start_line":129,"context_end_line":170,"code":" self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.get_attn_gradients","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.get_attn_gradients#L152-L153","kind":"function","name":"get_attn_gradients","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":152,"end_line":153,"context_start_line":132,"context_end_line":173,"code":" self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.save_attention_map","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.save_attention_map#L155-L156","kind":"function","name":"save_attention_map","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":155,"end_line":156,"context_start_line":135,"context_end_line":176,"code":" self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.get_attention_map","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.get_attention_map#L158-L159","kind":"function","name":"get_attention_map","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":158,"end_line":159,"context_start_line":138,"context_end_line":179,"code":" )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.transpose_for_scores","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.transpose_for_scores#L161-L167","kind":"function","name":"transpose_for_scores","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":161,"end_line":167,"context_start_line":141,"context_end_line":187,"code":" or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n # adding for multimodal lora","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.prune_heads","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.prune_heads#L322-L343","kind":"function","name":"prune_heads","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":322,"end_line":343,"context_start_line":302,"context_end_line":363,"code":" def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False):\n super().__init__()\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(config)\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n modular=None\n ):\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.feed_forward_chunk","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.feed_forward_chunk#L504-L507","kind":"function","name":"feed_forward_chunk","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":504,"end_line":507,"context_start_line":484,"context_end_line":527,"code":" layer_output_text = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output[:, query_length:, :],\n )\n layer_output = torch.cat([layer_output, layer_output_text], dim=1)\n else:\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n def feed_forward_chunk_query(self, attention_output):\n intermediate_output = self.intermediate_query(attention_output)\n layer_output = self.output_query(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.feed_forward_chunk_query","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.feed_forward_chunk_query#L509-L512","kind":"function","name":"feed_forward_chunk_query","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":509,"end_line":512,"context_start_line":489,"context_end_line":532,"code":" )\n layer_output = torch.cat([layer_output, layer_output_text], dim=1)\n else:\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n def feed_forward_chunk_query(self, attention_output):\n intermediate_output = self.intermediate_query(attention_output)\n layer_output = self.output_query(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer._init_weights","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer._init_weights#L695-L705","kind":"function","name":"_init_weights","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":695,"end_line":705,"context_start_line":675,"context_end_line":725,"code":"class BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.get_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.get_input_embeddings#L730-L731","kind":"function","name":"get_input_embeddings","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":730,"end_line":731,"context_start_line":710,"context_end_line":751,"code":" The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n has_query: bool = False,\n ) -> Tensor:","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.set_input_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.set_input_embeddings#L733-L734","kind":"function","name":"set_input_embeddings","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":733,"end_line":734,"context_start_line":713,"context_end_line":754,"code":" Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n has_query: bool = False,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer._prune_heads","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer._prune_heads#L736-L742","kind":"function","name":"_prune_heads","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":736,"end_line":742,"context_start_line":716,"context_end_line":762,"code":" \"\"\"\n\n def __init__(self, config, add_pooling_layer=False):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n has_query: bool = False,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.get_extended_attention_mask","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.get_extended_attention_mask#L744-L833","kind":"function","name":"get_extended_attention_mask","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":744,"end_line":833,"context_start_line":724,"context_end_line":853,"code":" self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n has_query: bool = False,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n\n # add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n if has_query: # UniLM style attention mask\n causal_mask = torch.cat(\n [\n torch.zeros(\n (batch_size, prefix_seq_len, seq_length),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=1,\n )\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, causal_mask.shape[1], prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n modular=None\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.get_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.get_output_embeddings#L1179-L1180","kind":"function","name":"get_output_embeddings","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1179,"end_line":1180,"context_start_line":1159,"context_end_line":1200,"code":" tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=False,\n ):","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.set_output_embeddings","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.set_output_embeddings#L1182-L1183","kind":"function","name":"set_output_embeddings","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1182,"end_line":1183,"context_start_line":1162,"context_end_line":1203,"code":" )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n query_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n labels=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n return_logits=False,\n is_decoder=False,\n ):\n r\"\"\"\n labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Labels for computing the masked language modeling loss. Indices should be in ``[-100, 0, ...,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.prepare_inputs_for_generation","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.prepare_inputs_for_generation#L1132-L1153","kind":"function","name":"prepare_inputs_for_generation","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1132,"end_line":1153,"context_start_line":1112,"context_end_line":1173,"code":" lm_loss = loss_fct(\n shifted_prediction_scores.view(-1, self.config.vocab_size),\n labels.view(-1),\n )\n if reduction == \"none\":\n lm_loss = lm_loss.view(prediction_scores.size(0), -1).sum(1)\n\n if not return_dict:\n output = (prediction_scores,) + outputs[2:]\n return ((lm_loss,) + output) if lm_loss is not None else output\n\n return CausalLMOutputWithCrossAttentions(\n loss=lm_loss,\n logits=prediction_scores,\n past_key_values=outputs.past_key_values,\n hidden_states=outputs.hidden_states,\n attentions=outputs.attentions,\n cross_attentions=outputs.cross_attentions,\n )\n\n def prepare_inputs_for_generation(\n self, input_ids, query_embeds, past=None, attention_mask=None, **model_kwargs\n ):\n # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n query_mask = input_ids.new_ones(query_embeds.shape[:-1])\n attention_mask = torch.cat([query_mask, attention_mask], dim=-1)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer._reorder_cache","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer._reorder_cache#L1155-L1163","kind":"function","name":"_reorder_cache","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":1155,"end_line":1163,"context_start_line":1135,"context_end_line":1183,"code":" # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly\n if attention_mask is None:\n attention_mask = input_ids.new_ones(input_ids.shape)\n query_mask = input_ids.new_ones(query_embeds.shape[:-1])\n attention_mask = torch.cat([query_mask, attention_mask], dim=-1)\n\n # cut decoder_input_ids if past is used\n if past is not None:\n input_ids = input_ids[:, -1:]\n\n return {\n \"input_ids\": input_ids,\n \"query_embeds\": query_embeds,\n \"attention_mask\": attention_mask,\n \"past_key_values\": past,\n \"encoder_hidden_states\": model_kwargs.get(\"encoder_hidden_states\", None),\n \"encoder_attention_mask\": model_kwargs.get(\"encoder_attention_mask\", None),\n \"is_decoder\": True,\n }\n\n def _reorder_cache(self, past, beam_idx):\n reordered_past = ()\n for layer_past in past:\n reordered_past += (\n tuple(\n past_state.index_select(0, beam_idx) for past_state in layer_past\n ),\n )\n return reordered_past\n\n\nclass BertForMaskedLM(BertPreTrainedModel):\n\n _keys_to_ignore_on_load_unexpected = [r\"pooler\"]\n _keys_to_ignore_on_load_missing = [r\"position_ids\", r\"predictions.decoder.bias\"]\n\n def __init__(self, config):\n super().__init__(config)\n\n self.bert = BertModel(config, add_pooling_layer=False)\n self.cls = BertOnlyMLMHead(config)\n\n self.init_weights()\n\n def get_output_embeddings(self):\n return self.cls.predictions.decoder\n\n def set_output_embeddings(self, new_embeddings):\n self.cls.predictions.decoder = new_embeddings","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.create_custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.create_custom_forward#L562-L568","kind":"function","name":"create_custom_forward","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":562,"end_line":568,"context_start_line":542,"context_end_line":588,"code":" )\n\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if getattr(self.config, \"gradient_checkpointing\", False) and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(\n *inputs, past_key_value, output_attentions, query_length\n )\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n modular\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n query_length,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.Qformer.custom_forward","uri":"program://CREMA/function/lavis.models.blip2_models.Qformer.custom_forward#L563-L566","kind":"function","name":"custom_forward","path":"lavis/models/blip2_models/Qformer.py","language":"python","start_line":563,"end_line":566,"context_start_line":543,"context_end_line":586,"code":"\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if getattr(self.config, \"gradient_checkpointing\", False) and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(\n *inputs, past_key_value, output_attentions, query_length\n )\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n modular\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,","source_hash":"277480176fd8d7fbc28e447f3b29b7305bfb727ffb7455a63bdbf9661daa789d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5","uri":"program://CREMA/module/lavis.models.blip2_models.blip2_t5#L1-L396","kind":"module","name":"lavis.models.blip2_models.blip2_t5","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":1,"end_line":396,"context_start_line":1,"context_end_line":396,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport torch\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom transformers import T5TokenizerFast\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration\n\n\n@registry.register_model(\"blip2_t5\")\nclass Blip2T5(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n t5_model=\"google/flan-t5-xl\",\n prompt=\"\",\n max_txt_len=32,\n apply_lemmatizer=False,\n ):\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(\n t5_model, config=t5_config\n )\n\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16()\n\n self.t5_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n\n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n input_tokens = self.t5_tokenizer(\n samples[\"text_input\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n output_tokens = self.t5_tokenizer(\n samples[\"text_output\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n targets = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100\n )\n\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens.attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n #print('prompt', self.prompt)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n if isinstance(prompt, str):\n prompt = [prompt] * image.size(0)\n else:\n assert len(prompt) == image.size(\n 0\n ), \"The number of prompts must be equal to the batch size.\"\n\n input_tokens = self.t5_tokenizer(\n prompt, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n image = samples[\"image\"]\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n if prompt:\n text_input = [prompt.format(question) for question in samples[\"text_input\"]]\n else:\n text_input = samples[\"text_input\"]\n\n input_tokens = self.t5_tokenizer(\n text_input, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_len,\n min_length=min_len,\n length_penalty=length_penalty,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n\n if self._apply_lemmatizer:\n output_text = self._lemmatize(output_text)\n\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n t5_model=t5_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n apply_lemmatizer=apply_lemmatizer,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.Blip2T5","uri":"program://CREMA/class/lavis.models.blip2_models.blip2_t5.Blip2T5#L20-L396","kind":"class","name":"Blip2T5","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":20,"end_line":396,"context_start_line":1,"context_end_line":396,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport torch\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom transformers import T5TokenizerFast\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration\n\n\n@registry.register_model(\"blip2_t5\")\nclass Blip2T5(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n t5_model=\"google/flan-t5-xl\",\n prompt=\"\",\n max_txt_len=32,\n apply_lemmatizer=False,\n ):\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(\n t5_model, config=t5_config\n )\n\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16()\n\n self.t5_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n\n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n input_tokens = self.t5_tokenizer(\n samples[\"text_input\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n output_tokens = self.t5_tokenizer(\n samples[\"text_output\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n targets = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100\n )\n\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens.attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n #print('prompt', self.prompt)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n if isinstance(prompt, str):\n prompt = [prompt] * image.size(0)\n else:\n assert len(prompt) == image.size(\n 0\n ), \"The number of prompts must be equal to the batch size.\"\n\n input_tokens = self.t5_tokenizer(\n prompt, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n image = samples[\"image\"]\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n if prompt:\n text_input = [prompt.format(question) for question in samples[\"text_input\"]]\n else:\n text_input = samples[\"text_input\"]\n\n input_tokens = self.t5_tokenizer(\n text_input, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_len,\n min_length=min_len,\n length_penalty=length_penalty,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n\n if self._apply_lemmatizer:\n output_text = self._lemmatize(output_text)\n\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n t5_model=t5_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n apply_lemmatizer=apply_lemmatizer,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.__init__","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.__init__#L38-L97","kind":"function","name":"__init__","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":38,"end_line":97,"context_start_line":18,"context_end_line":117,"code":"\n@registry.register_model(\"blip2_t5\")\nclass Blip2T5(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__(\n self,\n img_size=224,\n drop_path_rate=0,\n use_grad_checkpoint=False,\n vit_precision=\"fp16\",\n freeze_vit=True,\n num_query_token=32,\n t5_model=\"google/flan-t5-xl\",\n prompt=\"\",\n max_txt_len=32,\n apply_lemmatizer=False,\n ):\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder, self.ln_vision = self.init_vision_encoder(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision\n )\n if freeze_vit:\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n\n self.Qformer, self.query_tokens = self.init_Qformer(\n num_query_token, self.visual_encoder.num_features\n )\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(\n t5_model, config=t5_config\n )\n\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16()\n\n self.t5_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n\n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.forward","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.forward#L99-L151","kind":"function","name":"forward","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":99,"end_line":151,"context_start_line":79,"context_end_line":171,"code":" t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(\n t5_model, config=t5_config\n )\n\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16()\n\n self.t5_proj = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size\n )\n\n self.max_txt_len = max_txt_len\n self.prompt = prompt\n\n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n\n def forward(self, samples):\n image = samples[\"image\"]\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n input_tokens = self.t5_tokenizer(\n samples[\"text_input\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n output_tokens = self.t5_tokenizer(\n samples[\"text_output\"],\n padding=\"longest\",\n truncation=True,\n max_length=self.max_text_length,\n return_tensors=\"pt\",\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n targets = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100\n )\n\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens.attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.generate","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.generate#L154-L260","kind":"function","name":"generate","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":154,"end_line":260,"context_start_line":134,"context_end_line":280,"code":"\n targets = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100\n )\n\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens.attention_mask,\n return_dict=True,\n labels=targets,\n )\n loss = outputs.loss\n\n return {\"loss\": loss}\n\n @torch.no_grad()\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5,\n max_length=30,\n min_length=1,\n top_p=0.9,\n repetition_penalty=1.0,\n length_penalty=1.0,\n num_captions=1,\n temperature=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n if 'video' in samples:\n image = samples[\"video\"]\n vid = samples['vid']\n fids = samples['fids']\n out = {}\n #print('vid', vid)\n #print('fids', fids)\n b, t, c, w, h = image.shape\n image = image.reshape(-1, c, w, h)\n #print('prompt', self.prompt)\n else:\n image = samples[\"image\"]\n \n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if \"prompt\" in samples.keys():\n prompt = samples[\"prompt\"]\n else:\n prompt = self.prompt\n\n if isinstance(prompt, str):\n prompt = [prompt] * image.size(0)\n else:\n assert len(prompt) == image.size(\n 0\n ), \"The number of prompts must be equal to the batch size.\"\n\n input_tokens = self.t5_tokenizer(\n prompt, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=use_nucleus_sampling,\n top_p=top_p,\n temperature=temperature,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n image = samples[\"image\"]\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.predict_answers","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.predict_answers#L262-L327","kind":"function","name":"predict_answers","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":262,"end_line":327,"context_start_line":242,"context_end_line":347,"code":" min_length=min_length,\n repetition_penalty=repetition_penalty,\n length_penalty=length_penalty,\n num_return_sequences=num_captions,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n \n if 'video' in samples:\n out['vid'] = vid\n out['fids'] = fids\n caption_by_batch = []\n for i in range(b):\n caption_by_batch.append(output_text[i*t : (i+1)*t])\n out['output_text'] = caption_by_batch\n return out\n else:\n return output_text\n\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n image = samples[\"image\"]\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = self.ln_vision(self.visual_encoder(image))\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)\n query_output = self.Qformer.bert(\n query_embeds=query_tokens,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n inputs_t5 = self.t5_proj(query_output.last_hidden_state)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(image.device)\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n if prompt:\n text_input = [prompt.format(question) for question in samples[\"text_input\"]]\n else:\n text_input = samples[\"text_input\"]\n\n input_tokens = self.t5_tokenizer(\n text_input, padding=\"longest\", return_tensors=\"pt\"\n ).to(image.device)\n\n encoder_atts = torch.cat([atts_t5, input_tokens.attention_mask], dim=1)\n\n device_type = \"cuda\" if \"cuda\" in str(self.device) else \"cpu\"\n with torch.amp.autocast(device_type=device_type, dtype=torch.bfloat16):\n inputs_embeds = self.t5_model.encoder.embed_tokens(input_tokens.input_ids)\n inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_len,\n min_length=min_len,\n length_penalty=length_penalty,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n\n if self._apply_lemmatizer:\n output_text = self._lemmatize(output_text)\n\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5._lemmatize","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5._lemmatize#L329-L343","kind":"function","name":"_lemmatize","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":329,"end_line":343,"context_start_line":309,"context_end_line":363,"code":" inputs_embeds = torch.cat([inputs_t5, inputs_embeds], dim=1)\n\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_len,\n min_length=min_len,\n length_penalty=length_penalty,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n\n if self._apply_lemmatizer:\n output_text = self._lemmatize(output_text)\n\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.lemmatizer","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.lemmatizer#L346-L364","kind":"function","name":"lemmatizer","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":346,"end_line":364,"context_start_line":326,"context_end_line":384,"code":"\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.from_config","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.from_config#L367-L396","kind":"function","name":"from_config","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":367,"end_line":396,"context_start_line":347,"context_end_line":396,"code":" if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n\n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n t5_model=t5_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n apply_lemmatizer=apply_lemmatizer,\n )\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip2_models.blip2_t5.apply","uri":"program://CREMA/function/lavis.models.blip2_models.blip2_t5.apply#L330-L341","kind":"function","name":"apply","path":"lavis/models/blip2_models/blip2_t5.py","language":"python","start_line":330,"end_line":341,"context_start_line":310,"context_end_line":361,"code":"\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_len,\n min_length=min_len,\n length_penalty=length_penalty,\n )\n output_text = self.t5_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n\n if self._apply_lemmatizer:\n output_text = self._lemmatize(output_text)\n\n return output_text\n\n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )","source_hash":"571dc1c71c07320b9eacdd433b6034a568cdd4b33d5776c3b3e6febd231b53be","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs","uri":"program://CREMA/module/lavis.models.albef_models.albef_outputs#L1-L97","kind":"module","name":"lavis.models.albef_models.albef_outputs","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":1,"end_line":97,"context_start_line":1,"context_end_line":97,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlbefSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefIntermediateOutput(ModelOutput):\n # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_m: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlbefOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlbefSimilarity] = None\n\n intermediate_output: AlbefIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefOutputWithLogits(AlbefOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass AlbefOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, num_patches+1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, num_patches+1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, feature_dim)`, `optional`\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n multimodal_embeds: Optional[torch.FloatTensor] = None","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs.AlbefSimilarity","uri":"program://CREMA/class/lavis.models.albef_models.albef_outputs.AlbefSimilarity#L20-L28","kind":"class","name":"AlbefSimilarity","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":20,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlbefSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefIntermediateOutput(ModelOutput):\n # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_m: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs.AlbefIntermediateOutput","uri":"program://CREMA/class/lavis.models.albef_models.albef_outputs.AlbefIntermediateOutput#L32-L50","kind":"class","name":"AlbefIntermediateOutput","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":32,"end_line":50,"context_start_line":12,"context_end_line":70,"code":"from transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlbefSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefIntermediateOutput(ModelOutput):\n # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_m: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlbefOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlbefSimilarity] = None\n\n intermediate_output: AlbefIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefOutputWithLogits(AlbefOutput):","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs.AlbefOutput","uri":"program://CREMA/class/lavis.models.albef_models.albef_outputs.AlbefOutput#L54-L66","kind":"class","name":"AlbefOutput","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":54,"end_line":66,"context_start_line":34,"context_end_line":86,"code":" image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_m: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlbefOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlbefSimilarity] = None\n\n intermediate_output: AlbefIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefOutputWithLogits(AlbefOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass AlbefOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, num_patches+1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, num_patches+1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, feature_dim)`, `optional`\n\n The first embedding or feature is for the [CLS] token.","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs.AlbefOutputWithLogits","uri":"program://CREMA/class/lavis.models.albef_models.albef_outputs.AlbefOutputWithLogits#L70-L72","kind":"class","name":"AlbefOutputWithLogits","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":70,"end_line":72,"context_start_line":50,"context_end_line":92,"code":" decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlbefOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlbefSimilarity] = None\n\n intermediate_output: AlbefIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefOutputWithLogits(AlbefOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass AlbefOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, num_patches+1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, num_patches+1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, feature_dim)`, `optional`\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_outputs.AlbefOutputFeatures","uri":"program://CREMA/class/lavis.models.albef_models.albef_outputs.AlbefOutputFeatures#L76-L97","kind":"class","name":"AlbefOutputFeatures","path":"lavis/models/albef_models/albef_outputs.py","language":"python","start_line":76,"end_line":97,"context_start_line":56,"context_end_line":97,"code":" sims: Optional[AlbefSimilarity] = None\n\n intermediate_output: AlbefIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlbefOutputWithLogits(AlbefOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass AlbefOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, num_patches+1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, num_patches+1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, sequence_length+1, feature_dim)`, `optional`\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n multimodal_embeds: Optional[torch.FloatTensor] = None","source_hash":"bc71d46fa7e2c648ff8aaee44501282c1a2dda7b52e2714dc3708957becbc71c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_feature_extractor","uri":"program://CREMA/module/lavis.models.albef_models.albef_feature_extractor#L1-L204","kind":"module","name":"lavis.models.albef_models.albef_feature_extractor","path":"lavis/models/albef_models/albef_feature_extractor.py","language":"python","start_line":1,"end_line":204,"context_start_line":1,"context_end_line":204,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefOutputFeatures\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_feature_extractor\")\nclass AlbefFeatureExtractor(AlbefBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_feature_extractor.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=30):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n An AlbefOutputFeatures object, see lavis/models/albef_models/albef_outputs.py for details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n if isinstance(mode, str):\n mode = [mode]\n\n for m in mode:\n assert m in [\n \"multimodal\",\n \"image\",\n \"text\",\n ], \"mode must be one of [multimodal, image, text], but got {}\".format(m)\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if \"image\" in mode or \"multimodal\" in mode:\n assert (\n image is not None\n ), \"image must be provided if mode is 'image' or 'multimodal'\"\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n if \"text\" in mode or \"multimodal\" in mode:\n assert (\n caption is not None\n ), \"text must be provided if mode is 'text' or 'multimodal'\"\n\n text = self.tokenizer(\n caption,\n padding=True,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_features = F.normalize(self.text_proj(text_embeds), dim=-1)\n\n if \"multimodal\" in mode:\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward the positve image-text pair\n output = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return AlbefOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(\n url_or_filename=pretrain_path, rename_text_keys=False\n )\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"27372a918b9416b3e86ae146b75637b4332828902b65e1e5c6144fef0bdfdf8a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_feature_extractor.AlbefFeatureExtractor","uri":"program://CREMA/class/lavis.models.albef_models.albef_feature_extractor.AlbefFeatureExtractor#L23-L204","kind":"class","name":"AlbefFeatureExtractor","path":"lavis/models/albef_models/albef_feature_extractor.py","language":"python","start_line":23,"end_line":204,"context_start_line":3,"context_end_line":204,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefOutputFeatures\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_feature_extractor\")\nclass AlbefFeatureExtractor(AlbefBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_feature_extractor.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=30):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n An AlbefOutputFeatures object, see lavis/models/albef_models/albef_outputs.py for details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n if isinstance(mode, str):\n mode = [mode]\n\n for m in mode:\n assert m in [\n \"multimodal\",\n \"image\",\n \"text\",\n ], \"mode must be one of [multimodal, image, text], but got {}\".format(m)\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if \"image\" in mode or \"multimodal\" in mode:\n assert (\n image is not None\n ), \"image must be provided if mode is 'image' or 'multimodal'\"\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n if \"text\" in mode or \"multimodal\" in mode:\n assert (\n caption is not None\n ), \"text must be provided if mode is 'text' or 'multimodal'\"\n\n text = self.tokenizer(\n caption,\n padding=True,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_features = F.normalize(self.text_proj(text_embeds), dim=-1)\n\n if \"multimodal\" in mode:\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward the positve image-text pair\n output = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return AlbefOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(\n url_or_filename=pretrain_path, rename_text_keys=False\n )\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"27372a918b9416b3e86ae146b75637b4332828902b65e1e5c6144fef0bdfdf8a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_feature_extractor.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_feature_extractor.__init__#L28-L46","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_feature_extractor.py","language":"python","start_line":28,"end_line":46,"context_start_line":8,"context_end_line":66,"code":"import warnings\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefOutputFeatures\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_feature_extractor\")\nclass AlbefFeatureExtractor(AlbefBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_feature_extractor.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=30):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n An AlbefOutputFeatures object, see lavis/models/albef_models/albef_outputs.py for details.\n","source_hash":"27372a918b9416b3e86ae146b75637b4332828902b65e1e5c6144fef0bdfdf8a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_feature_extractor.extract_features","uri":"program://CREMA/function/lavis.models.albef_models.albef_feature_extractor.extract_features#L49-L172","kind":"function","name":"extract_features","path":"lavis/models/albef_models/albef_feature_extractor.py","language":"python","start_line":49,"end_line":172,"context_start_line":29,"context_end_line":192,"code":" super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n An AlbefOutputFeatures object, see lavis/models/albef_models/albef_outputs.py for details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n if isinstance(mode, str):\n mode = [mode]\n\n for m in mode:\n assert m in [\n \"multimodal\",\n \"image\",\n \"text\",\n ], \"mode must be one of [multimodal, image, text], but got {}\".format(m)\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if \"image\" in mode or \"multimodal\" in mode:\n assert (\n image is not None\n ), \"image must be provided if mode is 'image' or 'multimodal'\"\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_features = F.normalize(self.vision_proj(image_embeds), dim=-1)\n\n if \"text\" in mode or \"multimodal\" in mode:\n assert (\n caption is not None\n ), \"text must be provided if mode is 'text' or 'multimodal'\"\n\n text = self.tokenizer(\n caption,\n padding=True,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_features = F.normalize(self.text_proj(text_embeds), dim=-1)\n\n if \"multimodal\" in mode:\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward the positve image-text pair\n output = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return AlbefOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,","source_hash":"27372a918b9416b3e86ae146b75637b4332828902b65e1e5c6144fef0bdfdf8a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_feature_extractor.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_feature_extractor.from_config#L175-L204","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_feature_extractor.py","language":"python","start_line":175,"end_line":204,"context_start_line":155,"context_end_line":204,"code":" output = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return AlbefOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(\n url_or_filename=pretrain_path, rename_text_keys=False\n )\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"27372a918b9416b3e86ae146b75637b4332828902b65e1e5c6144fef0bdfdf8a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval","uri":"program://CREMA/module/lavis.models.albef_models.albef_retrieval#L1-L344","kind":"module","name":"lavis.models.albef_models.albef_retrieval","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":1,"end_line":344,"context_start_line":1,"context_end_line":344,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import AlbefBase, compute_sim_matrix\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutput,\n AlbefSimilarity,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"albef_retrieval\")\nclass AlbefRetrieval(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF retrieval model.\n\n Supported model types:\n - coco: fine-tuned ALBEF base model on COCO dataset (Karparthy split).\n - flickr: fine-tuned ALBEF base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> model = load_model(\"albef_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/albef_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/albef_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n temp=0.07,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n self.use_distill = use_distill\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(text)\n\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n if self.use_distill:\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n if self.use_distill:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n else:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n encoder_output_pos = self.text_encoder(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n with torch.no_grad():\n bs = image.size(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs] + 1e-4, dim=1)\n weights_t2i = F.softmax(sim_t2i[:, :bs] + 1e-4, dim=1)\n\n mask = torch.eq(idx, idx.T)\n weights_i2t.masked_fill_(mask, 0)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return AlbefOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=False)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 0)\n use_distill = cfg.get(\"use_distill\", True)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n use_distill=use_distill,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval.AlbefRetrieval","uri":"program://CREMA/class/lavis.models.albef_models.albef_retrieval.AlbefRetrieval#L26-L344","kind":"class","name":"AlbefRetrieval","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":26,"end_line":344,"context_start_line":6,"context_end_line":344,"code":"\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import AlbefBase, compute_sim_matrix\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutput,\n AlbefSimilarity,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"albef_retrieval\")\nclass AlbefRetrieval(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF retrieval model.\n\n Supported model types:\n - coco: fine-tuned ALBEF base model on COCO dataset (Karparthy split).\n - flickr: fine-tuned ALBEF base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> model = load_model(\"albef_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/albef_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/albef_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n temp=0.07,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n self.use_distill = use_distill\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(text)\n\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n if self.use_distill:\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n if self.use_distill:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n else:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n encoder_output_pos = self.text_encoder(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n with torch.no_grad():\n bs = image.size(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs] + 1e-4, dim=1)\n weights_t2i = F.softmax(sim_t2i[:, :bs] + 1e-4, dim=1)\n\n mask = torch.eq(idx, idx.T)\n weights_i2t.masked_fill_(mask, 0)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return AlbefOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=False)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 0)\n use_distill = cfg.get(\"use_distill\", True)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n use_distill=use_distill,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_retrieval.__init__#L45-L102","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":45,"end_line":102,"context_start_line":25,"context_end_line":122,"code":"@registry.register_model(\"albef_retrieval\")\nclass AlbefRetrieval(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF retrieval model.\n\n Supported model types:\n - coco: fine-tuned ALBEF base model on COCO dataset (Karparthy split).\n - flickr: fine-tuned ALBEF base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> model = load_model(\"albef_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/albef_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/albef_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n temp=0.07,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n self.use_distill = use_distill\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval._rampup_factor","uri":"program://CREMA/function/lavis.models.albef_models.albef_retrieval._rampup_factor#L104-L105","kind":"function","name":"_rampup_factor","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":104,"end_line":105,"context_start_line":84,"context_end_line":125,"code":" ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n self.use_distill = use_distill\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval.forward","uri":"program://CREMA/function/lavis.models.albef_models.albef_retrieval.forward#L107-L307","kind":"function","name":"forward","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":107,"end_line":307,"context_start_line":87,"context_end_line":327,"code":" # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n self.use_distill = use_distill\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(text)\n\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n if self.use_distill:\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n if self.use_distill:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n else:\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n encoder_output_pos = self.text_encoder(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n\n with torch.no_grad():\n bs = image.size(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs] + 1e-4, dim=1)\n weights_t2i = F.softmax(sim_t2i[:, :bs] + 1e-4, dim=1)\n\n mask = torch.eq(idx, idx.T)\n weights_i2t.masked_fill_(mask, 0)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return AlbefOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=False)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 0)\n use_distill = cfg.get(\"use_distill\", True)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n temp=temp,","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_retrieval.from_config#L310-L336","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":310,"end_line":336,"context_start_line":290,"context_end_line":344,"code":" sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=False)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 0)\n use_distill = cfg.get(\"use_distill\", True)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n use_distill=use_distill,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_retrieval.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.albef_models.albef_retrieval.compute_sim_matrix#L338-L344","kind":"function","name":"compute_sim_matrix","path":"lavis/models/albef_models/albef_retrieval.py","language":"python","start_line":338,"end_line":344,"context_start_line":318,"context_end_line":344,"code":" max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 0)\n use_distill = cfg.get(\"use_distill\", True)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n use_distill=use_distill,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"724cfcd7b15c8ce2f4026bcc6681cb3ee1a269f01a41980fd32e0bf6c4e5643b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr","uri":"program://CREMA/module/lavis.models.albef_models.albef_nlvr#L1-L260","kind":"module","name":"lavis.models.albef_models.albef_nlvr","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":1,"end_line":260,"context_start_line":1,"context_end_line":260,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefIntermediateOutput, AlbefOutput\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import BertModel\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_nlvr\")\nclass AlbefNLVR(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/albef_nlvr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n self.share_cross_attention(self.text_encoder.encoder)\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.share_cross_attention(self.text_encoder_m.encoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(\n text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(images)\n image0_embeds_m, image1_embeds_m = torch.split(\n image_embeds_m, targets.size(0)\n )\n encoder_output_m = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds_m, image1_embeds_m],\n encoder_attention_mask=[\n image_atts[: image0_embeds_m.size(0)],\n image_atts[image0_embeds_m.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n encoder_output_m = None\n image0_embeds_m, image1_embeds_m = None, None\n\n # return {\"loss\": loss}\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n image_embeds_m=torch.stack(\n [image0_embeds_m, image1_embeds_m], dim=0\n ),\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n bert_config.num_hidden_layers = 18\n\n text_encoder = BertModel.from_pretrained(\n \"bert-base-uncased\", config=bert_config, add_pooling_layer=False\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.AlbefNLVR","uri":"program://CREMA/class/lavis.models.albef_models.albef_nlvr.AlbefNLVR#L24-L260","kind":"class","name":"AlbefNLVR","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":24,"end_line":260,"context_start_line":4,"context_end_line":260,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefIntermediateOutput, AlbefOutput\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import BertModel\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_nlvr\")\nclass AlbefNLVR(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/albef_nlvr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n self.share_cross_attention(self.text_encoder.encoder)\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.share_cross_attention(self.text_encoder_m.encoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(\n text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(images)\n image0_embeds_m, image1_embeds_m = torch.split(\n image_embeds_m, targets.size(0)\n )\n encoder_output_m = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds_m, image1_embeds_m],\n encoder_attention_mask=[\n image_atts[: image0_embeds_m.size(0)],\n image_atts[image0_embeds_m.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n encoder_output_m = None\n image0_embeds_m, image1_embeds_m = None, None\n\n # return {\"loss\": loss}\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n image_embeds_m=torch.stack(\n [image0_embeds_m, image1_embeds_m], dim=0\n ),\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n bert_config.num_hidden_layers = 18\n\n text_encoder = BertModel.from_pretrained(\n \"bert-base-uncased\", config=bert_config, add_pooling_layer=False\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.__init__#L29-L74","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":29,"end_line":74,"context_start_line":9,"context_end_line":94,"code":"\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefIntermediateOutput, AlbefOutput\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import BertModel\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_nlvr\")\nclass AlbefNLVR(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/albef_nlvr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n self.share_cross_attention(self.text_encoder.encoder)\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.share_cross_attention(self.text_encoder_m.encoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr._rampup_factor","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr._rampup_factor#L76-L77","kind":"function","name":"_rampup_factor","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":76,"end_line":77,"context_start_line":56,"context_end_line":97,"code":" self.share_cross_attention(self.text_encoder.encoder)\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.share_cross_attention(self.text_encoder_m.encoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_nlvr\")\n >>> samples = {","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.forward","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.forward#L79-L196","kind":"function","name":"forward","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":79,"end_line":196,"context_start_line":59,"context_end_line":216,"code":" self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.share_cross_attention(self.text_encoder_m.encoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(\n text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(images)\n image0_embeds_m, image1_embeds_m = torch.split(\n image_embeds_m, targets.size(0)\n )\n encoder_output_m = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds_m, image1_embeds_m],\n encoder_attention_mask=[\n image_atts[: image0_embeds_m.size(0)],\n image_atts[image0_embeds_m.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n encoder_output_m = None\n image0_embeds_m, image1_embeds_m = None, None\n\n # return {\"loss\": loss}\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n image_embeds_m=torch.stack(\n [image0_embeds_m, image1_embeds_m], dim=0\n ),\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.share_cross_attention","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.share_cross_attention#L198-L211","kind":"function","name":"share_cross_attention","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":198,"end_line":211,"context_start_line":178,"context_end_line":231,"code":" loss = F.cross_entropy(prediction, targets)\n\n encoder_output_m = None\n image0_embeds_m, image1_embeds_m = None, None\n\n # return {\"loss\": loss}\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n image_embeds_m=torch.stack(\n [image0_embeds_m, image1_embeds_m], dim=0\n ),\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.predict","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.predict#L213-L215","kind":"function","name":"predict","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":213,"end_line":215,"context_start_line":193,"context_end_line":235,"code":" ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n bert_config.num_hidden_layers = 18\n\n text_encoder = BertModel.from_pretrained(\n \"bert-base-uncased\", config=bert_config, add_pooling_layer=False","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.load_from_pretrained","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.load_from_pretrained#L217-L224","kind":"function","name":"load_from_pretrained","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":217,"end_line":224,"context_start_line":197,"context_end_line":244,"code":"\n def share_cross_attention(self, model):\n for i in range(6):\n layer_num = 6 + i * 2\n modules_0 = model.layer[layer_num].crossattention.self._modules\n modules_1 = model.layer[layer_num + 1].crossattention.self._modules\n\n for name in modules_0.keys():\n if \"key\" in name or \"value\" in name:\n module_0 = modules_0[name]\n module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n bert_config.num_hidden_layers = 18\n\n text_encoder = BertModel.from_pretrained(\n \"bert-base-uncased\", config=bert_config, add_pooling_layer=False\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_nlvr.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_nlvr.from_config#L227-L260","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_nlvr.py","language":"python","start_line":227,"end_line":260,"context_start_line":207,"context_end_line":260,"code":" module_1 = modules_1[name]\n if hasattr(module_0, \"weight\"):\n module_0.weight = module_1.weight\n if hasattr(module_0, \"bias\"):\n module_0.bias = module_1.bias\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n def load_from_pretrained(self, url_or_filename, use_distill=True):\n _, msg = super().load_from_pretrained(url_or_filename)\n\n if use_distill and any([\"_m\" in k for k in msg.missing_keys]):\n # this is required when initializing the model from TA pre-trained weights\n self.copy_params()\n\n return msg\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n bert_config.num_hidden_layers = 18\n\n text_encoder = BertModel.from_pretrained(\n \"bert-base-uncased\", config=bert_config, add_pooling_layer=False\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"d2d74a66e4171c27b9a40a62b96436de1a184ba6ab6cd66808d57aaf902b8f40","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain","uri":"program://CREMA/module/lavis.models.albef_models.albef_pretrain#L1-L416","kind":"module","name":"lavis.models.albef_models.albef_pretrain","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":1,"end_line":416,"context_start_line":1,"context_end_line":416,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutput,\n AlbefSimilarity,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_pretrain\")\nclass AlbefPretrain(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF pretrain model.\n\n Supported model types:\n - base: ALBEF base model used for pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_pretrain_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n mlm_mask_prob=0.15,\n temp=0.07,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.mlm_probability = mlm_mask_prob\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_pretrain\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_mlm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # forward the positve image-text pair\n encoder_output_pos = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n with torch.no_grad():\n bs = image.size(0)\n\n weights_i2t = sim_i2t[:, :bs].clone()\n weights_t2i = sim_t2i[:, :bs].clone()\n\n weights_i2t.fill_diagonal_(-np.Inf)\n weights_t2i.fill_diagonal_(-np.Inf)\n\n weights_i2t = F.softmax(weights_i2t, dim=1)\n weights_t2i = F.softmax(weights_t2i, dim=1)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder.bert(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # MLM\n input_ids = text.input_ids.clone()\n labels = input_ids.clone()\n\n probability_matrix = torch.full(labels.shape, self.mlm_probability)\n input_ids, labels = self.mask(\n input_ids,\n self.text_encoder.config.vocab_size,\n self.device,\n targets=labels,\n probability_matrix=probability_matrix,\n )\n\n with torch.no_grad():\n logits_m = self.text_encoder_m(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds_m,\n encoder_attention_mask=image_atts,\n return_dict=True,\n return_logits=True,\n )\n mlm_output = self.text_encoder(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n labels=labels,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n )\n loss_mlm = mlm_output.loss\n\n return AlbefOutput(\n loss=loss_itc + loss_itm + loss_mlm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_mlm=loss_mlm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def mask(\n self,\n input_ids,\n vocab_size,\n device,\n targets=None,\n masked_indices=None,\n probability_matrix=None,\n ):\n \"\"\"\n Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.\n \"\"\"\n if masked_indices is None:\n masked_indices = torch.bernoulli(probability_matrix).bool()\n\n masked_indices[input_ids == self.tokenizer.pad_token_id] = False\n masked_indices[input_ids == self.tokenizer.cls_token_id] = False\n\n if targets is not None:\n targets[~masked_indices] = -100 # We only compute loss on masked tokens\n\n # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])\n indices_replaced = (\n torch.bernoulli(torch.full(input_ids.shape, 0.8)).bool() & masked_indices\n )\n input_ids[indices_replaced] = self.tokenizer.mask_token_id\n\n # 10% of the time, we replace masked input tokens with random word\n indices_random = (\n torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool()\n & masked_indices\n & ~indices_replaced\n )\n random_words = torch.randint(vocab_size, input_ids.shape, dtype=torch.long).to(\n device\n )\n input_ids[indices_random] = random_words[indices_random]\n # The rest of the time (10% of the time) we keep the masked input tokens unchanged\n\n if targets is not None:\n return input_ids, targets\n else:\n return input_ids\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n mlm_mask_prob = cfg.get(\"mlm_mask_prob\", 0.15)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 65536)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n mlm_mask_prob=mlm_mask_prob,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n return model","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain.AlbefPretrain","uri":"program://CREMA/class/lavis.models.albef_models.albef_pretrain.AlbefPretrain#L29-L416","kind":"class","name":"AlbefPretrain","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":29,"end_line":416,"context_start_line":9,"context_end_line":416,"code":"\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutput,\n AlbefSimilarity,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_pretrain\")\nclass AlbefPretrain(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF pretrain model.\n\n Supported model types:\n - base: ALBEF base model used for pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_pretrain_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n mlm_mask_prob=0.15,\n temp=0.07,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.mlm_probability = mlm_mask_prob\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_pretrain\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_mlm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # forward the positve image-text pair\n encoder_output_pos = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n with torch.no_grad():\n bs = image.size(0)\n\n weights_i2t = sim_i2t[:, :bs].clone()\n weights_t2i = sim_t2i[:, :bs].clone()\n\n weights_i2t.fill_diagonal_(-np.Inf)\n weights_t2i.fill_diagonal_(-np.Inf)\n\n weights_i2t = F.softmax(weights_i2t, dim=1)\n weights_t2i = F.softmax(weights_t2i, dim=1)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder.bert(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # MLM\n input_ids = text.input_ids.clone()\n labels = input_ids.clone()\n\n probability_matrix = torch.full(labels.shape, self.mlm_probability)\n input_ids, labels = self.mask(\n input_ids,\n self.text_encoder.config.vocab_size,\n self.device,\n targets=labels,\n probability_matrix=probability_matrix,\n )\n\n with torch.no_grad():\n logits_m = self.text_encoder_m(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds_m,\n encoder_attention_mask=image_atts,\n return_dict=True,\n return_logits=True,\n )\n mlm_output = self.text_encoder(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n labels=labels,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n )\n loss_mlm = mlm_output.loss\n\n return AlbefOutput(\n loss=loss_itc + loss_itm + loss_mlm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_mlm=loss_mlm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def mask(\n self,\n input_ids,\n vocab_size,\n device,\n targets=None,\n masked_indices=None,\n probability_matrix=None,\n ):\n \"\"\"\n Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.\n \"\"\"\n if masked_indices is None:\n masked_indices = torch.bernoulli(probability_matrix).bool()\n\n masked_indices[input_ids == self.tokenizer.pad_token_id] = False\n masked_indices[input_ids == self.tokenizer.cls_token_id] = False\n\n if targets is not None:\n targets[~masked_indices] = -100 # We only compute loss on masked tokens\n\n # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])\n indices_replaced = (\n torch.bernoulli(torch.full(input_ids.shape, 0.8)).bool() & masked_indices\n )\n input_ids[indices_replaced] = self.tokenizer.mask_token_id\n\n # 10% of the time, we replace masked input tokens with random word\n indices_random = (\n torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool()\n & masked_indices\n & ~indices_replaced\n )\n random_words = torch.randint(vocab_size, input_ids.shape, dtype=torch.long).to(\n device\n )\n input_ids[indices_random] = random_words[indices_random]\n # The rest of the time (10% of the time) we keep the masked input tokens unchanged\n\n if targets is not None:\n return input_ids, targets\n else:\n return input_ids\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n mlm_mask_prob = cfg.get(\"mlm_mask_prob\", 0.15)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 65536)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n mlm_mask_prob=mlm_mask_prob,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n return model","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_pretrain.__init__#L41-L100","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":41,"end_line":100,"context_start_line":21,"context_end_line":120,"code":"from lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.med import BertForMaskedLM\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"albef_pretrain\")\nclass AlbefPretrain(AlbefBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n ALBEF pretrain model.\n\n Supported model types:\n - base: ALBEF base model used for pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/albef_pretrain_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n embed_dim=256,\n mlm_mask_prob=0.15,\n temp=0.07,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.embed_dim = embed_dim\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.mlm_probability = mlm_mask_prob\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain._rampup_factor","uri":"program://CREMA/function/lavis.models.albef_models.albef_pretrain._rampup_factor#L102-L103","kind":"function","name":"_rampup_factor","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":102,"end_line":103,"context_start_line":82,"context_end_line":123,"code":" ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.mlm_probability = mlm_mask_prob\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_pretrain\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain.forward","uri":"program://CREMA/function/lavis.models.albef_models.albef_pretrain.forward#L105-L339","kind":"function","name":"forward","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":105,"end_line":339,"context_start_line":85,"context_end_line":359,"code":" # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(temp * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.mlm_probability = mlm_mask_prob\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_pretrain\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_mlm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text_output = self.text_encoder.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n text_output_m = self.text_encoder_m.bert(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # forward the positve image-text pair\n encoder_output_pos = self.text_encoder.bert(\n encoder_embeds=text_embeds,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n mode=\"fusion\",\n )\n with torch.no_grad():\n bs = image.size(0)\n\n weights_i2t = sim_i2t[:, :bs].clone()\n weights_t2i = sim_t2i[:, :bs].clone()\n\n weights_i2t.fill_diagonal_(-np.Inf)\n weights_t2i.fill_diagonal_(-np.Inf)\n\n weights_i2t = F.softmax(weights_i2t, dim=1)\n weights_t2i = F.softmax(weights_t2i, dim=1)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n encoder_output_neg = self.text_encoder.bert(\n encoder_embeds=text_embeds_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_output_pos.last_hidden_state[:, 0, :],\n encoder_output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # MLM\n input_ids = text.input_ids.clone()\n labels = input_ids.clone()\n\n probability_matrix = torch.full(labels.shape, self.mlm_probability)\n input_ids, labels = self.mask(\n input_ids,\n self.text_encoder.config.vocab_size,\n self.device,\n targets=labels,\n probability_matrix=probability_matrix,\n )\n\n with torch.no_grad():\n logits_m = self.text_encoder_m(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds_m,\n encoder_attention_mask=image_atts,\n return_dict=True,\n return_logits=True,\n )\n mlm_output = self.text_encoder(\n input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n labels=labels,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n )\n loss_mlm = mlm_output.loss\n\n return AlbefOutput(\n loss=loss_itc + loss_itm + loss_mlm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_mlm=loss_mlm,\n sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def mask(\n self,\n input_ids,\n vocab_size,\n device,\n targets=None,\n masked_indices=None,\n probability_matrix=None,\n ):\n \"\"\"\n Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.\n \"\"\"\n if masked_indices is None:\n masked_indices = torch.bernoulli(probability_matrix).bool()\n\n masked_indices[input_ids == self.tokenizer.pad_token_id] = False\n masked_indices[input_ids == self.tokenizer.cls_token_id] = False\n\n if targets is not None:","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain.mask","uri":"program://CREMA/function/lavis.models.albef_models.albef_pretrain.mask#L341-L383","kind":"function","name":"mask","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":341,"end_line":383,"context_start_line":321,"context_end_line":403,"code":" sims=AlbefSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=encoder_output_pos,\n encoder_output_neg=encoder_output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def mask(\n self,\n input_ids,\n vocab_size,\n device,\n targets=None,\n masked_indices=None,\n probability_matrix=None,\n ):\n \"\"\"\n Prepare masked tokens inputs/labels for masked language modeling: 80% MASK, 10% random, 10% original.\n \"\"\"\n if masked_indices is None:\n masked_indices = torch.bernoulli(probability_matrix).bool()\n\n masked_indices[input_ids == self.tokenizer.pad_token_id] = False\n masked_indices[input_ids == self.tokenizer.cls_token_id] = False\n\n if targets is not None:\n targets[~masked_indices] = -100 # We only compute loss on masked tokens\n\n # 80% of the time, we replace masked input tokens with tokenizer.mask_token ([MASK])\n indices_replaced = (\n torch.bernoulli(torch.full(input_ids.shape, 0.8)).bool() & masked_indices\n )\n input_ids[indices_replaced] = self.tokenizer.mask_token_id\n\n # 10% of the time, we replace masked input tokens with random word\n indices_random = (\n torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool()\n & masked_indices\n & ~indices_replaced\n )\n random_words = torch.randint(vocab_size, input_ids.shape, dtype=torch.long).to(\n device\n )\n input_ids[indices_random] = random_words[indices_random]\n # The rest of the time (10% of the time) we keep the masked input tokens unchanged\n\n if targets is not None:\n return input_ids, targets\n else:\n return input_ids\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n mlm_mask_prob = cfg.get(\"mlm_mask_prob\", 0.15)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 65536)\n","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_pretrain.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_pretrain.from_config#L386-L416","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_pretrain.py","language":"python","start_line":386,"end_line":416,"context_start_line":366,"context_end_line":416,"code":" input_ids[indices_replaced] = self.tokenizer.mask_token_id\n\n # 10% of the time, we replace masked input tokens with random word\n indices_random = (\n torch.bernoulli(torch.full(input_ids.shape, 0.5)).bool()\n & masked_indices\n & ~indices_replaced\n )\n random_words = torch.randint(vocab_size, input_ids.shape, dtype=torch.long).to(\n device\n )\n input_ids[indices_random] = random_words[indices_random]\n # The rest of the time (10% of the time) we keep the masked input tokens unchanged\n\n if targets is not None:\n return input_ids, targets\n else:\n return input_ids\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n config_text_encoder = BertConfig.from_json_file(\n get_abs_path(cfg[\"med_config_path\"])\n )\n config_text_encoder.fusion_layer = 6\n text_encoder = BertForMaskedLM.from_pretrained(\n \"bert-base-uncased\", config=config_text_encoder\n )\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n mlm_mask_prob = cfg.get(\"mlm_mask_prob\", 0.15)\n temp = cfg.get(\"temp\", 0.07)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 65536)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n embed_dim=embed_dim,\n mlm_mask_prob=mlm_mask_prob,\n temp=temp,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n return model","source_hash":"808ae15a9a98c2634db28a05f7228948ae0d20b5542d73576a8550d5a8788e09","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa","uri":"program://CREMA/module/lavis.models.albef_models.albef_vqa#L1-L442","kind":"module","name":"lavis.models.albef_models.albef_vqa","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":1,"end_line":442,"context_start_line":1,"context_end_line":442,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path, is_url\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefIntermediateOutput, AlbefOutput\nfrom lavis.models.base_model import MomentumDistilationMixin, tile\nfrom lavis.models.med import BertConfig, BertLMHeadModel, XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder, interpolate_pos_embed\nfrom lavis.common.dist_utils import download_cached_file\n\n\n@registry.register_model(\"albef_vqa\")\nclass AlbefVQA(AlbefBase, MomentumDistilationMixin):\n \"\"\"\n ALBEF VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained ALBEF base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned ALBEF base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\", \"vqav2\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/albef_vqav2.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=35,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.text_decoder_m = deepcopy(self.text_decoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.text_decoder, self.text_decoder_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n An AlbefOutput object containing loss and intermediate outputs;\n see lavis/models/albef_models/albef_outputs.py for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 384, 384),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 1000,\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n (\n encoder_output,\n encoder_output_m,\n image_embeds,\n image_embeds_m,\n ) = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples, encoder_out=(encoder_output, encoder_output_m)\n )\n\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n if self.use_distill:\n self._momentum_update()\n with torch.no_grad():\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n tokenized_text=samples[\"tokenized_text\"],\n visual_embeds=image_embeds_m,\n )\n else:\n encoder_output_m = None\n image_embeds_m = None\n\n return encoder_output, encoder_output_m, image_embeds, image_embeds_m\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output, question_output_m = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n if self.use_distill:\n with torch.no_grad():\n question_states_m = []\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states_m += [question_output_m.last_hidden_state[b]] * n\n question_states_m = torch.stack(question_states_m, 0)\n\n logits_m = self.text_decoder_m(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states_m,\n encoder_attention_mask=question_atts,\n return_logits=True,\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(self, samples, answer_list, num_ans_candidates=128, **kwargs):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n \"\"\"\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self.rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n # answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _, _, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n text_encoder = XBertEncoder.from_config(cfg)\n\n config_decoder = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n config_decoder.fusion_layer = 0\n config_decoder.num_hidden_layers = 6\n text_decoder = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=config_decoder\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n max_txt_len = cfg.get(\"max_txt_len\", 25)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n use_distill=use_distill,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint:\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n # reshape positional embedding to accomodate for image resolution change\n pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n state_dict[\"visual_encoder.pos_embed\"] = pos_embed_reshaped\n\n m_pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n state_dict[\"visual_encoder_m.pos_embed\"] = m_pos_embed_reshaped\n\n for key in list(state_dict.keys()):\n if \"bert\" in key:\n encoder_key = key.replace(\"bert.\", \"\")\n state_dict[encoder_key] = state_dict[key]\n\n # intialize text decoder as multimodal encoder (last 6 layers of model.text_encoder)\n if \"text_encoder\" in key:\n if \"layer\" in key:\n encoder_keys = key.split(\".\")\n layer_num = int(encoder_keys[4])\n\n if layer_num < 6:\n del state_dict[key]\n continue\n else:\n decoder_layer_num = layer_num - 6\n encoder_keys[4] = str(decoder_layer_num)\n encoder_key = \".\".join(encoder_keys)\n else:\n encoder_key = key\n decoder_key = encoder_key.replace(\"text_encoder\", \"text_decoder\")\n state_dict[decoder_key] = state_dict[key]\n\n del state_dict[key]\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n logging.info(f\"missing keys: {msg.missing_keys}\")\n\n return msg","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.AlbefVQA","uri":"program://CREMA/class/lavis.models.albef_models.albef_vqa.AlbefVQA#L25-L442","kind":"class","name":"AlbefVQA","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":25,"end_line":442,"context_start_line":5,"context_end_line":442,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path, is_url\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import AlbefIntermediateOutput, AlbefOutput\nfrom lavis.models.base_model import MomentumDistilationMixin, tile\nfrom lavis.models.med import BertConfig, BertLMHeadModel, XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder, interpolate_pos_embed\nfrom lavis.common.dist_utils import download_cached_file\n\n\n@registry.register_model(\"albef_vqa\")\nclass AlbefVQA(AlbefBase, MomentumDistilationMixin):\n \"\"\"\n ALBEF VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained ALBEF base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned ALBEF base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\", \"vqav2\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/albef_vqav2.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=35,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.text_decoder_m = deepcopy(self.text_decoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.text_decoder, self.text_decoder_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n An AlbefOutput object containing loss and intermediate outputs;\n see lavis/models/albef_models/albef_outputs.py for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 384, 384),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 1000,\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n (\n encoder_output,\n encoder_output_m,\n image_embeds,\n image_embeds_m,\n ) = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples, encoder_out=(encoder_output, encoder_output_m)\n )\n\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n if self.use_distill:\n self._momentum_update()\n with torch.no_grad():\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n tokenized_text=samples[\"tokenized_text\"],\n visual_embeds=image_embeds_m,\n )\n else:\n encoder_output_m = None\n image_embeds_m = None\n\n return encoder_output, encoder_output_m, image_embeds, image_embeds_m\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output, question_output_m = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n if self.use_distill:\n with torch.no_grad():\n question_states_m = []\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states_m += [question_output_m.last_hidden_state[b]] * n\n question_states_m = torch.stack(question_states_m, 0)\n\n logits_m = self.text_decoder_m(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states_m,\n encoder_attention_mask=question_atts,\n return_logits=True,\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(self, samples, answer_list, num_ans_candidates=128, **kwargs):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n \"\"\"\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self.rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n # answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _, _, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n text_encoder = XBertEncoder.from_config(cfg)\n\n config_decoder = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n config_decoder.fusion_layer = 0\n config_decoder.num_hidden_layers = 6\n text_decoder = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=config_decoder\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n max_txt_len = cfg.get(\"max_txt_len\", 25)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n use_distill=use_distill,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint:\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n # reshape positional embedding to accomodate for image resolution change\n pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n state_dict[\"visual_encoder.pos_embed\"] = pos_embed_reshaped\n\n m_pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n state_dict[\"visual_encoder_m.pos_embed\"] = m_pos_embed_reshaped\n\n for key in list(state_dict.keys()):\n if \"bert\" in key:\n encoder_key = key.replace(\"bert.\", \"\")\n state_dict[encoder_key] = state_dict[key]\n\n # intialize text decoder as multimodal encoder (last 6 layers of model.text_encoder)\n if \"text_encoder\" in key:\n if \"layer\" in key:\n encoder_keys = key.split(\".\")\n layer_num = int(encoder_keys[4])\n\n if layer_num < 6:\n del state_dict[key]\n continue\n else:\n decoder_layer_num = layer_num - 6\n encoder_keys[4] = str(decoder_layer_num)\n encoder_key = \".\".join(encoder_keys)\n else:\n encoder_key = key\n decoder_key = encoder_key.replace(\"text_encoder\", \"text_decoder\")\n state_dict[decoder_key] = state_dict[key]\n\n del state_dict[key]\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n logging.info(f\"missing keys: {msg.missing_keys}\")\n\n return msg","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.__init__#L42-L78","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":42,"end_line":78,"context_start_line":22,"context_end_line":98,"code":"\n\n@registry.register_model(\"albef_vqa\")\nclass AlbefVQA(AlbefBase, MomentumDistilationMixin):\n \"\"\"\n ALBEF VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained ALBEF base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned ALBEF base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\", \"vqav2\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/albef_vqav2.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n use_distill=True,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=35,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.text_decoder_m = deepcopy(self.text_decoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.text_decoder, self.text_decoder_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n An AlbefOutput object containing loss and intermediate outputs;\n see lavis/models/albef_models/albef_outputs.py for more details.\n","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa._rampup_factor","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa._rampup_factor#L80-L81","kind":"function","name":"_rampup_factor","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":80,"end_line":81,"context_start_line":60,"context_end_line":101,"code":"\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.text_decoder_m = deepcopy(self.text_decoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.text_decoder, self.text_decoder_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n An AlbefOutput object containing loss and intermediate outputs;\n see lavis/models/albef_models/albef_outputs.py for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.forward","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.forward#L83-L135","kind":"function","name":"forward","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":83,"end_line":135,"context_start_line":63,"context_end_line":155,"code":"\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.text_decoder_m = deepcopy(self.text_decoder)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.text_decoder, self.text_decoder_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n An AlbefOutput object containing loss and intermediate outputs;\n see lavis/models/albef_models/albef_outputs.py for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"albef_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 384, 384),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 1000,\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n (\n encoder_output,\n encoder_output_m,\n image_embeds,\n image_embeds_m,\n ) = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples, encoder_out=(encoder_output, encoder_output_m)\n )\n\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n if self.use_distill:\n self._momentum_update()\n with torch.no_grad():","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.forward_encoder","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.forward_encoder#L137-L165","kind":"function","name":"forward_encoder","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":137,"end_line":165,"context_start_line":117,"context_end_line":185,"code":" encoder_output_m,\n image_embeds,\n image_embeds_m,\n ) = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples, encoder_out=(encoder_output, encoder_output_m)\n )\n\n return AlbefOutput(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n if self.use_distill:\n self._momentum_update()\n with torch.no_grad():\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n tokenized_text=samples[\"tokenized_text\"],\n visual_embeds=image_embeds_m,\n )\n else:\n encoder_output_m = None\n image_embeds_m = None\n\n return encoder_output, encoder_output_m, image_embeds, image_embeds_m\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output, question_output_m = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.forward_decoder","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.forward_decoder#L167-L226","kind":"function","name":"forward_decoder","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":167,"end_line":226,"context_start_line":147,"context_end_line":246,"code":"\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n if self.use_distill:\n self._momentum_update()\n with torch.no_grad():\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n tokenized_text=samples[\"tokenized_text\"],\n visual_embeds=image_embeds_m,\n )\n else:\n encoder_output_m = None\n image_embeds_m = None\n\n return encoder_output, encoder_output_m, image_embeds, image_embeds_m\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output, question_output_m = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n if self.use_distill:\n with torch.no_grad():\n question_states_m = []\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states_m += [question_output_m.last_hidden_state[b]] * n\n question_states_m = torch.stack(question_states_m, 0)\n\n logits_m = self.text_decoder_m(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states_m,\n encoder_attention_mask=question_atts,\n return_logits=True,\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(self, samples, answer_list, num_ans_candidates=128, **kwargs):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.predict_answers","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.predict_answers#L228-L267","kind":"function","name":"predict_answers","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":228,"end_line":267,"context_start_line":208,"context_end_line":287,"code":"\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n soft_labels=F.softmax(logits_m, dim=-1),\n alpha=alpha,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(self, samples, answer_list, num_ans_candidates=128, **kwargs):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"albef_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n \"\"\"\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self.rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n # answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _, _, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.rank_answers","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.rank_answers#L269-L346","kind":"function","name":"rank_answers","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":269,"end_line":346,"context_start_line":249,"context_end_line":366,"code":" >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n \"\"\"\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self.rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n # answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _, _, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n text_encoder = XBertEncoder.from_config(cfg)\n\n config_decoder = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n config_decoder.fusion_layer = 0\n config_decoder.num_hidden_layers = 6\n text_decoder = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=config_decoder\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n max_txt_len = cfg.get(\"max_txt_len\", 25)\n\n model = cls(","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.from_config#L349-L379","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":349,"end_line":379,"context_start_line":329,"context_end_line":399,"code":" input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n text_encoder = XBertEncoder.from_config(cfg)\n\n config_decoder = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n config_decoder.fusion_layer = 0\n config_decoder.num_hidden_layers = 6\n text_decoder = BertLMHeadModel.from_pretrained(\n \"bert-base-uncased\", config=config_decoder\n )\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n max_txt_len = cfg.get(\"max_txt_len\", 25)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n use_distill=use_distill,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint:\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n # reshape positional embedding to accomodate for image resolution change\n pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_vqa.load_from_pretrained","uri":"program://CREMA/function/lavis.models.albef_models.albef_vqa.load_from_pretrained#L381-L442","kind":"function","name":"load_from_pretrained","path":"lavis/models/albef_models/albef_vqa.py","language":"python","start_line":381,"end_line":442,"context_start_line":361,"context_end_line":442,"code":" alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n max_txt_len = cfg.get(\"max_txt_len\", 25)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n use_distill=use_distill,\n momentum=momentum,\n alpha=alpha,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n if \"model\" in checkpoint:\n state_dict = checkpoint[\"model\"]\n else:\n state_dict = checkpoint\n\n # reshape positional embedding to accomodate for image resolution change\n pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n state_dict[\"visual_encoder.pos_embed\"] = pos_embed_reshaped\n\n m_pos_embed_reshaped = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n state_dict[\"visual_encoder_m.pos_embed\"] = m_pos_embed_reshaped\n\n for key in list(state_dict.keys()):\n if \"bert\" in key:\n encoder_key = key.replace(\"bert.\", \"\")\n state_dict[encoder_key] = state_dict[key]\n\n # intialize text decoder as multimodal encoder (last 6 layers of model.text_encoder)\n if \"text_encoder\" in key:\n if \"layer\" in key:\n encoder_keys = key.split(\".\")\n layer_num = int(encoder_keys[4])\n\n if layer_num < 6:\n del state_dict[key]\n continue\n else:\n decoder_layer_num = layer_num - 6\n encoder_keys[4] = str(decoder_layer_num)\n encoder_key = \".\".join(encoder_keys)\n else:\n encoder_key = key\n decoder_key = encoder_key.replace(\"text_encoder\", \"text_decoder\")\n state_dict[decoder_key] = state_dict[key]\n\n del state_dict[key]\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n logging.info(f\"missing keys: {msg.missing_keys}\")\n\n return msg","source_hash":"2c4d23f39e3a2d5868fbcee6931cc776007ccdbbbcc7539c1c4d7993320892ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification","uri":"program://CREMA/module/lavis.models.albef_models.albef_classification#L1-L182","kind":"module","name":"lavis.models.albef_models.albef_classification","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":1,"end_line":182,"context_start_line":1,"context_end_line":182,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutputWithLogits,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"albef_classification\")\nclass AlbefClassification(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ve\": \"configs/models/albef_classification_ve.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n\n if num_classes > 0:\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n else:\n warnings.warn(\n f\"Found num_classes=0, initializing {type(self)} without classifier.\"\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n image_embeds_m, encoder_output_m, prediction_m = None, None, None\n\n # return {\"loss\": loss}\n return AlbefOutputWithLogits(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification.AlbefClassification","uri":"program://CREMA/class/lavis.models.albef_models.albef_classification.AlbefClassification#L26-L182","kind":"class","name":"AlbefClassification","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":26,"end_line":182,"context_start_line":6,"context_end_line":182,"code":"\"\"\"\n\nimport warnings\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutputWithLogits,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"albef_classification\")\nclass AlbefClassification(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ve\": \"configs/models/albef_classification_ve.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n\n if num_classes > 0:\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n else:\n warnings.warn(\n f\"Found num_classes=0, initializing {type(self)} without classifier.\"\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n image_embeds_m, encoder_output_m, prediction_m = None, None, None\n\n # return {\"loss\": loss}\n return AlbefOutputWithLogits(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification.__init__","uri":"program://CREMA/function/lavis.models.albef_models.albef_classification.__init__#L31-L78","kind":"function","name":"__init__","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":31,"end_line":78,"context_start_line":11,"context_end_line":98,"code":"import torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import AlbefBase\nfrom lavis.models.albef_models.albef_outputs import (\n AlbefIntermediateOutput,\n AlbefOutputWithLogits,\n)\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"albef_classification\")\nclass AlbefClassification(AlbefBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ve\": \"configs/models/albef_classification_ve.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n use_distill=True,\n max_txt_len=40,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.max_txt_len = max_txt_len\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n\n if num_classes > 0:\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n else:\n warnings.warn(\n f\"Found num_classes=0, initializing {type(self)} without classifier.\"\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification._rampup_factor","uri":"program://CREMA/function/lavis.models.albef_models.albef_classification._rampup_factor#L80-L81","kind":"function","name":"_rampup_factor","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":80,"end_line":81,"context_start_line":60,"context_end_line":101,"code":" warnings.warn(\n f\"Found num_classes=0, initializing {type(self)} without classifier.\"\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification.forward","uri":"program://CREMA/function/lavis.models.albef_models.albef_classification.forward#L83-L147","kind":"function","name":"forward","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":83,"end_line":147,"context_start_line":63,"context_end_line":167,"code":"\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n image_embeds_m, encoder_output_m, prediction_m = None, None, None\n\n # return {\"loss\": loss}\n return AlbefOutputWithLogits(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification.predict","uri":"program://CREMA/function/lavis.models.albef_models.albef_classification.predict#L149-L151","kind":"function","name":"predict","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":149,"end_line":151,"context_start_line":129,"context_end_line":171,"code":" else:\n loss = F.cross_entropy(prediction, targets)\n\n image_embeds_m, encoder_output_m, prediction_m = None, None, None\n\n # return {\"loss\": loss}\n return AlbefOutputWithLogits(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.albef_models.albef_classification.from_config","uri":"program://CREMA/function/lavis.models.albef_models.albef_classification.from_config#L154-L182","kind":"function","name":"from_config","path":"lavis/models/albef_models/albef_classification.py","language":"python","start_line":154,"end_line":182,"context_start_line":134,"context_end_line":182,"code":" # return {\"loss\": loss}\n return AlbefOutputWithLogits(\n loss=loss,\n intermediate_output=AlbefIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n alpha = cfg.get(\"alpha\", 0.4)\n momentum = cfg.get(\"momentum\", 0.995)\n use_distill = cfg.get(\"use_distill\", True)\n num_classes = cfg.get(\"num_classes\", -1)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"9d7f82ac2ad6f25b61c3c12e71b33082feaf46704b0de7247a1c965f27449a32","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.losses","uri":"program://CREMA/module/lavis.models.ulip_models.losses#L1-L62","kind":"module","name":"lavis.models.ulip_models.losses","path":"lavis/models/ulip_models/losses.py","language":"python","start_line":1,"end_line":62,"context_start_line":1,"context_end_line":62,"code":"'''\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom lavis.models.ulip_models.utils import utils\n\nclass ULIPWithImageLoss(nn.Module):\n def __init__(self):\n super().__init__()\n self.labels = None\n self.last_local_batch_size = None\n\n def forward(self, outputs):\n pc_embed = outputs['pc_embed']\n text_embed = outputs['text_embed']\n image_embed = outputs['image_embed']\n logit_scale = outputs['logit_scale']\n local_batch_size = pc_embed.size(0)\n\n if local_batch_size != self.last_local_batch_size:\n self.labels = local_batch_size * utils.get_rank() + torch.arange(\n local_batch_size, device=pc_embed.device\n )\n self.last_local_batch_size = local_batch_size\n\n # normalized features\n pc_embed = F.normalize(pc_embed, dim=-1, p=2)\n text_embed = F.normalize(text_embed, dim=-1, p=2)\n image_embed = F.normalize(image_embed, dim=-1, p=2)\n\n # gather features from all GPUs\n pc_embed_all, text_embed_all, image_embed_all = \\\n utils.all_gather_batch([pc_embed, text_embed, image_embed])\n\n # cosine similarity as logits\n logits_per_pc_text = logit_scale * pc_embed @ text_embed_all.t()\n logits_per_text_pc = logit_scale * text_embed @ pc_embed_all.t()\n logits_per_pc_image = logit_scale * pc_embed @ image_embed_all.t()\n logits_per_image_pc = logit_scale * image_embed @ pc_embed_all.t()\n\n loss = (F.cross_entropy(logits_per_pc_text, self.labels) + \\\n F.cross_entropy(logits_per_text_pc, self.labels)) / 2 + \\\n (F.cross_entropy(logits_per_pc_image, self.labels) + F.cross_entropy(logits_per_image_pc, self.labels)) / 2\n\n # compute accuracy\n with torch.no_grad():\n pred = torch.argmax(logits_per_pc_text, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_text_acc = 100 * correct / local_batch_size\n\n pred = torch.argmax(logits_per_pc_image, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_image_acc = 100 * correct / local_batch_size\n\n return {'loss': loss, 'ulip_loss': loss, 'ulip_pc_image_acc': pc_image_acc, 'ulip_pc_text_acc': pc_text_acc}","source_hash":"22c319742b38b9a8ce11950311fcc04f2af7d9dc6f999b578e3ecb77a746a93c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.losses.ULIPWithImageLoss","uri":"program://CREMA/class/lavis.models.ulip_models.losses.ULIPWithImageLoss#L14-L62","kind":"class","name":"ULIPWithImageLoss","path":"lavis/models/ulip_models/losses.py","language":"python","start_line":14,"end_line":62,"context_start_line":1,"context_end_line":62,"code":"'''\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom lavis.models.ulip_models.utils import utils\n\nclass ULIPWithImageLoss(nn.Module):\n def __init__(self):\n super().__init__()\n self.labels = None\n self.last_local_batch_size = None\n\n def forward(self, outputs):\n pc_embed = outputs['pc_embed']\n text_embed = outputs['text_embed']\n image_embed = outputs['image_embed']\n logit_scale = outputs['logit_scale']\n local_batch_size = pc_embed.size(0)\n\n if local_batch_size != self.last_local_batch_size:\n self.labels = local_batch_size * utils.get_rank() + torch.arange(\n local_batch_size, device=pc_embed.device\n )\n self.last_local_batch_size = local_batch_size\n\n # normalized features\n pc_embed = F.normalize(pc_embed, dim=-1, p=2)\n text_embed = F.normalize(text_embed, dim=-1, p=2)\n image_embed = F.normalize(image_embed, dim=-1, p=2)\n\n # gather features from all GPUs\n pc_embed_all, text_embed_all, image_embed_all = \\\n utils.all_gather_batch([pc_embed, text_embed, image_embed])\n\n # cosine similarity as logits\n logits_per_pc_text = logit_scale * pc_embed @ text_embed_all.t()\n logits_per_text_pc = logit_scale * text_embed @ pc_embed_all.t()\n logits_per_pc_image = logit_scale * pc_embed @ image_embed_all.t()\n logits_per_image_pc = logit_scale * image_embed @ pc_embed_all.t()\n\n loss = (F.cross_entropy(logits_per_pc_text, self.labels) + \\\n F.cross_entropy(logits_per_text_pc, self.labels)) / 2 + \\\n (F.cross_entropy(logits_per_pc_image, self.labels) + F.cross_entropy(logits_per_image_pc, self.labels)) / 2\n\n # compute accuracy\n with torch.no_grad():\n pred = torch.argmax(logits_per_pc_text, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_text_acc = 100 * correct / local_batch_size\n\n pred = torch.argmax(logits_per_pc_image, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_image_acc = 100 * correct / local_batch_size\n\n return {'loss': loss, 'ulip_loss': loss, 'ulip_pc_image_acc': pc_image_acc, 'ulip_pc_text_acc': pc_text_acc}","source_hash":"22c319742b38b9a8ce11950311fcc04f2af7d9dc6f999b578e3ecb77a746a93c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.losses.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.losses.__init__#L15-L18","kind":"function","name":"__init__","path":"lavis/models/ulip_models/losses.py","language":"python","start_line":15,"end_line":18,"context_start_line":1,"context_end_line":38,"code":"'''\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom lavis.models.ulip_models.utils import utils\n\nclass ULIPWithImageLoss(nn.Module):\n def __init__(self):\n super().__init__()\n self.labels = None\n self.last_local_batch_size = None\n\n def forward(self, outputs):\n pc_embed = outputs['pc_embed']\n text_embed = outputs['text_embed']\n image_embed = outputs['image_embed']\n logit_scale = outputs['logit_scale']\n local_batch_size = pc_embed.size(0)\n\n if local_batch_size != self.last_local_batch_size:\n self.labels = local_batch_size * utils.get_rank() + torch.arange(\n local_batch_size, device=pc_embed.device\n )\n self.last_local_batch_size = local_batch_size\n\n # normalized features\n pc_embed = F.normalize(pc_embed, dim=-1, p=2)\n text_embed = F.normalize(text_embed, dim=-1, p=2)\n image_embed = F.normalize(image_embed, dim=-1, p=2)\n\n # gather features from all GPUs","source_hash":"22c319742b38b9a8ce11950311fcc04f2af7d9dc6f999b578e3ecb77a746a93c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.losses.forward","uri":"program://CREMA/function/lavis.models.ulip_models.losses.forward#L20-L62","kind":"function","name":"forward","path":"lavis/models/ulip_models/losses.py","language":"python","start_line":20,"end_line":62,"context_start_line":1,"context_end_line":62,"code":"'''\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nfrom lavis.models.ulip_models.utils import utils\n\nclass ULIPWithImageLoss(nn.Module):\n def __init__(self):\n super().__init__()\n self.labels = None\n self.last_local_batch_size = None\n\n def forward(self, outputs):\n pc_embed = outputs['pc_embed']\n text_embed = outputs['text_embed']\n image_embed = outputs['image_embed']\n logit_scale = outputs['logit_scale']\n local_batch_size = pc_embed.size(0)\n\n if local_batch_size != self.last_local_batch_size:\n self.labels = local_batch_size * utils.get_rank() + torch.arange(\n local_batch_size, device=pc_embed.device\n )\n self.last_local_batch_size = local_batch_size\n\n # normalized features\n pc_embed = F.normalize(pc_embed, dim=-1, p=2)\n text_embed = F.normalize(text_embed, dim=-1, p=2)\n image_embed = F.normalize(image_embed, dim=-1, p=2)\n\n # gather features from all GPUs\n pc_embed_all, text_embed_all, image_embed_all = \\\n utils.all_gather_batch([pc_embed, text_embed, image_embed])\n\n # cosine similarity as logits\n logits_per_pc_text = logit_scale * pc_embed @ text_embed_all.t()\n logits_per_text_pc = logit_scale * text_embed @ pc_embed_all.t()\n logits_per_pc_image = logit_scale * pc_embed @ image_embed_all.t()\n logits_per_image_pc = logit_scale * image_embed @ pc_embed_all.t()\n\n loss = (F.cross_entropy(logits_per_pc_text, self.labels) + \\\n F.cross_entropy(logits_per_text_pc, self.labels)) / 2 + \\\n (F.cross_entropy(logits_per_pc_image, self.labels) + F.cross_entropy(logits_per_image_pc, self.labels)) / 2\n\n # compute accuracy\n with torch.no_grad():\n pred = torch.argmax(logits_per_pc_text, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_text_acc = 100 * correct / local_batch_size\n\n pred = torch.argmax(logits_per_pc_image, dim=-1)\n correct = pred.eq(self.labels).sum()\n pc_image_acc = 100 * correct / local_batch_size\n\n return {'loss': loss, 'ulip_loss': loss, 'ulip_pc_image_acc': pc_image_acc, 'ulip_pc_text_acc': pc_text_acc}","source_hash":"22c319742b38b9a8ce11950311fcc04f2af7d9dc6f999b578e3ecb77a746a93c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models","uri":"program://CREMA/module/lavis.models.ulip_models.ULIP_models#L1-L243","kind":"module","name":"lavis.models.ulip_models.ULIP_models","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":1,"end_line":243,"context_start_line":1,"context_end_line":243,"code":"'''\n * Copyright (c) 2023, salesforce.com, inc.\n * All rights reserved.\n * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\n## FROM: https://github.com/salesforce/ULIP\n## TODO: Convert to LAVIS format. Currently only supports functionality for XInstructBLIP\n\n# Modified from github.com/openai/CLIP\nfrom collections import OrderedDict\n\nimport timm\nfrom torch import nn\nfrom lavis.models.ulip_models import losses\nfrom torch.nn.parameter import Parameter\nfrom easydict import EasyDict\nimport torch\nimport numpy as np\nfrom lavis.common.dist_utils import download_cached_file\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\nclass QuickGELU(nn.Module):\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask\n\n def attention(self, x: torch.Tensor):\n self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):\n # super().__init__(ssl_mlp_dim, ssl_emb_dim, **kwargs)\n super().__init__()\n kwargs = EasyDict(kwargs)\n self.context_length = kwargs.context_length\n self.vision_width = kwargs.vision_width\n self.visual = kwargs.vision_model\n self.num_features = kwargs.embed_dim\n\n self.transformer = Transformer(\n width=kwargs.transformer_width,\n layers=kwargs.transformer_layers,\n heads=kwargs.transformer_heads,\n attn_mask=self.build_attention_mask(),\n )\n\n self.vocab_size = kwargs.vocab_size\n self.token_embedding = nn.Embedding(kwargs.vocab_size, kwargs.transformer_width)\n self.positional_embedding = nn.Parameter(torch.empty(self.context_length, kwargs.transformer_width))\n self.ln_final = LayerNorm(kwargs.transformer_width)\n\n self.image_projection = nn.Parameter(torch.empty(kwargs.vision_width, kwargs.embed_dim))\n self.text_projection = nn.Parameter(torch.empty(kwargs.transformer_width, kwargs.embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n\n self.initialize_parameters()\n\n self.point_encoder = point_encoder\n\n self.pc_projection = nn.Parameter(torch.empty(kwargs.pc_feat_dims, kwargs.embed_dim ))\n nn.init.normal_(self.pc_projection, std= kwargs.embed_dim ** -0.5)\n\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection\n return pc_embed\n\n def forward(self, pc, text=None, image=None):\n\n if text is not None:\n text_embed_all = []\n for i in range(text.shape[0]):\n text_for_one_sample = text[i]\n text_embed = self.encode_text(text_for_one_sample)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed = text_embed.mean(dim=0)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: \n text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(\n url, check_hash=False, progress=True\n )\n config = cfg_from_yaml_file(config_addr)\n pc_feat_dims = 768 \n if ulip_v == \"ulip2_scaledup\":\n config.model.depth = 18\n transformer_layers = 18\n embed_dim=1280\n else:\n embed_dim=512\n\n transformer_layers = 12\n point_encoder = PointTransformer(config.model)\n # =====================================================================\n model = ULIP_WITH_IMAGE(embed_dim=embed_dim, vision_width=pc_feat_dims, point_encoder=point_encoder, vision_model=vision_model,\n context_length=77, vocab_size=49408,\n transformer_width=512, transformer_heads=8, transformer_layers=transformer_layers, pc_feat_dims=pc_feat_dims)\n \n ## TODO: setup config\n if ulip_v == 2:\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_5/ULIP-2_pointbert_last.pt'\n elif ulip_v == 1:\n cached_file = '/export/share/lxue/shared_models/ULIP-1/objaverse/ULIP-1_pointbert_last.pt'\n elif ulip_v == 'shapenet':\n cached_file = '/export/share/lxue/shared_models/ULIP-1/objaverse_shapenet/checkpoint_last.pt'\n elif ulip_v == 'objaverse_k_1':\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_k_1/checkpoint_last.pt'\n elif ulip_v == 'objaverse_shapenet_k_1':\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_1/checkpoint_last.pt'\n elif ulip_v == \"ulip2_scaledup\":\n cached_file = \"/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_1_scaled_up/checkpoint_last.pt\"\n # url = \"https://storage.cloud.google.com/sfr-ulip-code-release-research/pretrained_models/ckpt_zero-sho_classification/checkpoint_pointbert.pt\"\n # cached_file = download_cached_file(\n # url, check_hash=False, progress=True\n # )\n ckpt = torch.load(cached_file, map_location='cpu')\n state_dict = OrderedDict()\n for k, v in ckpt['state_dict'].items():\n state_dict[k.replace('module.', '')] = v\n # model.cuda()\n model.load_state_dict(state_dict, strict=False)\n return model","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.LayerNorm","uri":"program://CREMA/class/lavis.models.ulip_models.ULIP_models.LayerNorm#L24-L30","kind":"class","name":"LayerNorm","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":24,"end_line":30,"context_start_line":4,"context_end_line":50,"code":" * SPDX-License-Identifier: BSD-3-Clause\n * For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n * By Le Xue\n'''\n## FROM: https://github.com/salesforce/ULIP\n## TODO: Convert to LAVIS format. Currently only supports functionality for XInstructBLIP\n\n# Modified from github.com/openai/CLIP\nfrom collections import OrderedDict\n\nimport timm\nfrom torch import nn\nfrom lavis.models.ulip_models import losses\nfrom torch.nn.parameter import Parameter\nfrom easydict import EasyDict\nimport torch\nimport numpy as np\nfrom lavis.common.dist_utils import download_cached_file\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\nclass QuickGELU(nn.Module):\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.QuickGELU","uri":"program://CREMA/class/lavis.models.ulip_models.ULIP_models.QuickGELU#L33-L35","kind":"class","name":"QuickGELU","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":33,"end_line":35,"context_start_line":13,"context_end_line":55,"code":"\nimport timm\nfrom torch import nn\nfrom lavis.models.ulip_models import losses\nfrom torch.nn.parameter import Parameter\nfrom easydict import EasyDict\nimport torch\nimport numpy as np\nfrom lavis.common.dist_utils import download_cached_file\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\nclass QuickGELU(nn.Module):\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask\n\n def attention(self, x: torch.Tensor):\n self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.ResidualAttentionBlock","uri":"program://CREMA/class/lavis.models.ulip_models.ULIP_models.ResidualAttentionBlock#L38-L59","kind":"class","name":"ResidualAttentionBlock","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":38,"end_line":59,"context_start_line":18,"context_end_line":79,"code":"from easydict import EasyDict\nimport torch\nimport numpy as np\nfrom lavis.common.dist_utils import download_cached_file\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n ret = super().forward(x.type(torch.float32))\n return ret.type(orig_type)\n\n\nclass QuickGELU(nn.Module):\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask\n\n def attention(self, x: torch.Tensor):\n self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):\n # super().__init__(ssl_mlp_dim, ssl_emb_dim, **kwargs)\n super().__init__()\n kwargs = EasyDict(kwargs)\n self.context_length = kwargs.context_length\n self.vision_width = kwargs.vision_width","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.Transformer","uri":"program://CREMA/class/lavis.models.ulip_models.ULIP_models.Transformer#L62-L70","kind":"class","name":"Transformer","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":62,"end_line":70,"context_start_line":42,"context_end_line":90,"code":" self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask\n\n def attention(self, x: torch.Tensor):\n self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):\n # super().__init__(ssl_mlp_dim, ssl_emb_dim, **kwargs)\n super().__init__()\n kwargs = EasyDict(kwargs)\n self.context_length = kwargs.context_length\n self.vision_width = kwargs.vision_width\n self.visual = kwargs.vision_model\n self.num_features = kwargs.embed_dim\n\n self.transformer = Transformer(\n width=kwargs.transformer_width,\n layers=kwargs.transformer_layers,\n heads=kwargs.transformer_heads,\n attn_mask=self.build_attention_mask(),\n )\n\n self.vocab_size = kwargs.vocab_size","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.ULIP_WITH_IMAGE","uri":"program://CREMA/class/lavis.models.ulip_models.ULIP_models.ULIP_WITH_IMAGE#L73-L181","kind":"class","name":"ULIP_WITH_IMAGE","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":73,"end_line":181,"context_start_line":53,"context_end_line":201,"code":" self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):\n # super().__init__(ssl_mlp_dim, ssl_emb_dim, **kwargs)\n super().__init__()\n kwargs = EasyDict(kwargs)\n self.context_length = kwargs.context_length\n self.vision_width = kwargs.vision_width\n self.visual = kwargs.vision_model\n self.num_features = kwargs.embed_dim\n\n self.transformer = Transformer(\n width=kwargs.transformer_width,\n layers=kwargs.transformer_layers,\n heads=kwargs.transformer_heads,\n attn_mask=self.build_attention_mask(),\n )\n\n self.vocab_size = kwargs.vocab_size\n self.token_embedding = nn.Embedding(kwargs.vocab_size, kwargs.transformer_width)\n self.positional_embedding = nn.Parameter(torch.empty(self.context_length, kwargs.transformer_width))\n self.ln_final = LayerNorm(kwargs.transformer_width)\n\n self.image_projection = nn.Parameter(torch.empty(kwargs.vision_width, kwargs.embed_dim))\n self.text_projection = nn.Parameter(torch.empty(kwargs.transformer_width, kwargs.embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n\n self.initialize_parameters()\n\n self.point_encoder = point_encoder\n\n self.pc_projection = nn.Parameter(torch.empty(kwargs.pc_feat_dims, kwargs.embed_dim ))\n nn.init.normal_(self.pc_projection, std= kwargs.embed_dim ** -0.5)\n\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection\n return pc_embed\n\n def forward(self, pc, text=None, image=None):\n\n if text is not None:\n text_embed_all = []\n for i in range(text.shape[0]):\n text_for_one_sample = text[i]\n text_embed = self.encode_text(text_for_one_sample)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed = text_embed.mean(dim=0)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: \n text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.get_loss","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.get_loss#L184-L185","kind":"function","name":"get_loss","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":184,"end_line":185,"context_start_line":164,"context_end_line":205,"code":" text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: \n text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(\n url, check_hash=False, progress=True\n )\n config = cfg_from_yaml_file(config_addr)\n pc_feat_dims = 768 ","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.get_metric_names","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.get_metric_names#L188-L189","kind":"function","name":"get_metric_names","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":188,"end_line":189,"context_start_line":168,"context_end_line":209,"code":" text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(\n url, check_hash=False, progress=True\n )\n config = cfg_from_yaml_file(config_addr)\n pc_feat_dims = 768 \n if ulip_v == \"ulip2_scaledup\":\n config.model.depth = 18\n transformer_layers = 18\n embed_dim=1280","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.ULIP_PointBERT","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.ULIP_PointBERT#L191-L243","kind":"function","name":"ULIP_PointBERT","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":191,"end_line":243,"context_start_line":171,"context_end_line":243,"code":" if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(\n url, check_hash=False, progress=True\n )\n config = cfg_from_yaml_file(config_addr)\n pc_feat_dims = 768 \n if ulip_v == \"ulip2_scaledup\":\n config.model.depth = 18\n transformer_layers = 18\n embed_dim=1280\n else:\n embed_dim=512\n\n transformer_layers = 12\n point_encoder = PointTransformer(config.model)\n # =====================================================================\n model = ULIP_WITH_IMAGE(embed_dim=embed_dim, vision_width=pc_feat_dims, point_encoder=point_encoder, vision_model=vision_model,\n context_length=77, vocab_size=49408,\n transformer_width=512, transformer_heads=8, transformer_layers=transformer_layers, pc_feat_dims=pc_feat_dims)\n \n ## TODO: setup config\n if ulip_v == 2:\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_5/ULIP-2_pointbert_last.pt'\n elif ulip_v == 1:\n cached_file = '/export/share/lxue/shared_models/ULIP-1/objaverse/ULIP-1_pointbert_last.pt'\n elif ulip_v == 'shapenet':\n cached_file = '/export/share/lxue/shared_models/ULIP-1/objaverse_shapenet/checkpoint_last.pt'\n elif ulip_v == 'objaverse_k_1':\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_k_1/checkpoint_last.pt'\n elif ulip_v == 'objaverse_shapenet_k_1':\n cached_file = '/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_1/checkpoint_last.pt'\n elif ulip_v == \"ulip2_scaledup\":\n cached_file = \"/export/share/lxue/shared_models/ULIP-2/objaverse_shapenet_k_1_scaled_up/checkpoint_last.pt\"\n # url = \"https://storage.cloud.google.com/sfr-ulip-code-release-research/pretrained_models/ckpt_zero-sho_classification/checkpoint_pointbert.pt\"\n # cached_file = download_cached_file(\n # url, check_hash=False, progress=True\n # )\n ckpt = torch.load(cached_file, map_location='cpu')\n state_dict = OrderedDict()\n for k, v in ckpt['state_dict'].items():\n state_dict[k.replace('module.', '')] = v\n # model.cuda()\n model.load_state_dict(state_dict, strict=False)\n return model","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.forward","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.forward#L154-L181","kind":"function","name":"forward","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":154,"end_line":181,"context_start_line":134,"context_end_line":201,"code":" nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection\n return pc_embed\n\n def forward(self, pc, text=None, image=None):\n\n if text is not None:\n text_embed_all = []\n for i in range(text.shape[0]):\n text_for_one_sample = text[i]\n text_embed = self.encode_text(text_for_one_sample)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed = text_embed.mean(dim=0)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: \n text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)\n else:\n image_embed = None\n \n res = {'text_embed': text_embed_all,\n 'pc_embed': pc_embed,\n 'image_embed': image_embed,\n 'logit_scale': self.logit_scale.exp()\n }\n return pc_embed\n\n\ndef get_loss(args):\n return losses.ULIPWithImageLoss()\n\n\ndef get_metric_names(model):\n return ['loss', 'ulip_loss', 'ulip_pc_image_acc', 'ulip_pc_text_acc']\n\ndef ULIP_PointBERT(ulip_v=2):\n vision_model = timm.create_model('vit_base_patch16_224', num_classes=0)\n\n # =====================================================================\n # import the 3D backbone and specify the output point cloud feature dimension\n from lavis.models.ulip_models.pointbert.point_encoder import PointTransformer\n from lavis.models.ulip_models.utils.config import cfg_from_yaml_file\n ## TODO: parse as config\n # config_addr = '/export/home/LAVIS/lavis/models/ulip_models/pointbert/PointTransformer_8192point.yaml'\n url = \"https://raw.githubusercontent.com/salesforce/ULIP/48d8d00b1cdb2aee79005817a202816f1c521911/models/pointbert/PointTransformer_8192point.yaml\"\n config_addr = download_cached_file(","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.__init__#L74-L104","kind":"function","name":"__init__","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":74,"end_line":104,"context_start_line":54,"context_end_line":124,"code":" return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):\n # super().__init__(ssl_mlp_dim, ssl_emb_dim, **kwargs)\n super().__init__()\n kwargs = EasyDict(kwargs)\n self.context_length = kwargs.context_length\n self.vision_width = kwargs.vision_width\n self.visual = kwargs.vision_model\n self.num_features = kwargs.embed_dim\n\n self.transformer = Transformer(\n width=kwargs.transformer_width,\n layers=kwargs.transformer_layers,\n heads=kwargs.transformer_heads,\n attn_mask=self.build_attention_mask(),\n )\n\n self.vocab_size = kwargs.vocab_size\n self.token_embedding = nn.Embedding(kwargs.vocab_size, kwargs.transformer_width)\n self.positional_embedding = nn.Parameter(torch.empty(self.context_length, kwargs.transformer_width))\n self.ln_final = LayerNorm(kwargs.transformer_width)\n\n self.image_projection = nn.Parameter(torch.empty(kwargs.vision_width, kwargs.embed_dim))\n self.text_projection = nn.Parameter(torch.empty(kwargs.transformer_width, kwargs.embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n\n self.initialize_parameters()\n\n self.point_encoder = point_encoder\n\n self.pc_projection = nn.Parameter(torch.empty(kwargs.pc_feat_dims, kwargs.embed_dim ))\n nn.init.normal_(self.pc_projection, std= kwargs.embed_dim ** -0.5)\n\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.attention","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.attention#L52-L54","kind":"function","name":"attention","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":52,"end_line":54,"context_start_line":32,"context_end_line":74,"code":"\nclass QuickGELU(nn.Module):\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(OrderedDict([\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", QuickGELU()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model))\n ]))\n self.ln_2 = LayerNorm(d_model)\n self.attn_mask = attn_mask\n\n def attention(self, x: torch.Tensor):\n self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None\n return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]\n\n def forward(self, x: torch.Tensor):\n x = x + self.attention(self.ln_1(x))\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])\n\n def forward(self, x: torch.Tensor):\n return self.resblocks(x)\n\n\nclass ULIP_WITH_IMAGE(nn.Module):\n def __init__(self, point_encoder, **kwargs):","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.encode_image","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.encode_image#L106-L110","kind":"function","name":"encode_image","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":106,"end_line":110,"context_start_line":86,"context_end_line":130,"code":" heads=kwargs.transformer_heads,\n attn_mask=self.build_attention_mask(),\n )\n\n self.vocab_size = kwargs.vocab_size\n self.token_embedding = nn.Embedding(kwargs.vocab_size, kwargs.transformer_width)\n self.positional_embedding = nn.Parameter(torch.empty(self.context_length, kwargs.transformer_width))\n self.ln_final = LayerNorm(kwargs.transformer_width)\n\n self.image_projection = nn.Parameter(torch.empty(kwargs.vision_width, kwargs.embed_dim))\n self.text_projection = nn.Parameter(torch.empty(kwargs.transformer_width, kwargs.embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n\n self.initialize_parameters()\n\n self.point_encoder = point_encoder\n\n self.pc_projection = nn.Parameter(torch.empty(kwargs.pc_feat_dims, kwargs.embed_dim ))\n nn.init.normal_(self.pc_projection, std= kwargs.embed_dim ** -0.5)\n\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.encode_text","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.encode_text#L112-L123","kind":"function","name":"encode_text","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":112,"end_line":123,"context_start_line":92,"context_end_line":143,"code":" self.positional_embedding = nn.Parameter(torch.empty(self.context_length, kwargs.transformer_width))\n self.ln_final = LayerNorm(kwargs.transformer_width)\n\n self.image_projection = nn.Parameter(torch.empty(kwargs.vision_width, kwargs.embed_dim))\n self.text_projection = nn.Parameter(torch.empty(kwargs.transformer_width, kwargs.embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n\n self.initialize_parameters()\n\n self.point_encoder = point_encoder\n\n self.pc_projection = nn.Parameter(torch.empty(kwargs.pc_feat_dims, kwargs.embed_dim ))\n nn.init.normal_(self.pc_projection, std= kwargs.embed_dim ** -0.5)\n\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.build_attention_mask","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.build_attention_mask#L125-L131","kind":"function","name":"build_attention_mask","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":125,"end_line":131,"context_start_line":105,"context_end_line":151,"code":"\n def encode_image(self, image):\n x = self.visual(image)\n x = x @ self.image_projection\n\n return x\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.initialize_parameters","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.initialize_parameters#L133-L147","kind":"function","name":"initialize_parameters","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":133,"end_line":147,"context_start_line":113,"context_end_line":167,"code":" x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection\n return pc_embed\n\n def forward(self, pc, text=None, image=None):\n\n if text is not None:\n text_embed_all = []\n for i in range(text.shape[0]):\n text_for_one_sample = text[i]\n text_embed = self.encode_text(text_for_one_sample)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed = text_embed.mean(dim=0)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: ","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.ULIP_models.encode_pc","uri":"program://CREMA/function/lavis.models.ulip_models.ULIP_models.encode_pc#L149-L152","kind":"function","name":"encode_pc","path":"lavis/models/ulip_models/ULIP_models.py","language":"python","start_line":149,"end_line":152,"context_start_line":129,"context_end_line":172,"code":" mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def initialize_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n\n proj_std = (self.transformer.width ** -0.5) * ((2 * self.transformer.layers) ** -0.5)\n attn_std = self.transformer.width ** -0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n nn.init.normal_(self.image_projection, std=self.vision_width ** -0.5)\n nn.init.normal_(self.text_projection, std=self.transformer.width ** -0.5)\n\n def encode_pc(self, pc):\n pc_feat = self.point_encoder(pc)\n pc_embed = pc_feat @ self.pc_projection\n return pc_embed\n\n def forward(self, pc, text=None, image=None):\n\n if text is not None:\n text_embed_all = []\n for i in range(text.shape[0]):\n text_for_one_sample = text[i]\n text_embed = self.encode_text(text_for_one_sample)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed = text_embed.mean(dim=0)\n text_embed = text_embed / text_embed.norm(dim=-1, keepdim=True)\n text_embed_all.append(text_embed)\n\n text_embed_all = torch.stack(text_embed_all)\n else: \n text_embed_all = None\n\n pc_embed = self.encode_pc(pc)\n if image is not None:\n image_embed = self.encode_image(image)","source_hash":"4af44caf0ea29b46d9020aa95cc362bd7f237a7d0105dfe5e7a7b11fe0432072","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config","uri":"program://CREMA/module/lavis.models.ulip_models.utils.config#L1-L63","kind":"module","name":"lavis.models.ulip_models.utils.config","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":1,"end_line":63,"context_start_line":1,"context_end_line":63,"code":"import yaml\nfrom easydict import EasyDict\nimport os\nfrom .logger import print_log\n\ndef log_args_to_file(args, pre='args', logger=None):\n for key, val in args.__dict__.items():\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef log_config_to_file(cfg, pre='cfg', logger=None):\n for key, val in cfg.items():\n if isinstance(cfg[key], EasyDict):\n print_log(f'{pre}.{key} = edict()', logger = logger)\n log_config_to_file(cfg[key], pre=pre + '.' + key, logger=logger)\n continue\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\n\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n except:\n new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_config(args, logger=None):\n if args.resume:\n cfg_path = os.path.join(args.experiment_path, 'config.yaml')\n if not os.path.exists(cfg_path):\n print_log(\"Failed to resume\", logger = logger)\n raise FileNotFoundError()\n print_log(f'Resume yaml from {cfg_path}', logger = logger)\n args.config = cfg_path\n config = cfg_from_yaml_file(args.config)\n if not args.resume and args.local_rank == 0:\n save_experiment_config(args, config, logger)\n return config\n\ndef save_experiment_config(args, config, logger = None):\n config_path = os.path.join(args.experiment_path, 'config.yaml')\n os.system('cp %s %s' % (args.config, config_path))\n print_log(f'Copy the Config file from {args.config} to {config_path}',logger = logger )","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.log_args_to_file","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.log_args_to_file#L6-L8","kind":"function","name":"log_args_to_file","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":6,"end_line":8,"context_start_line":1,"context_end_line":28,"code":"import yaml\nfrom easydict import EasyDict\nimport os\nfrom .logger import print_log\n\ndef log_args_to_file(args, pre='args', logger=None):\n for key, val in args.__dict__.items():\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef log_config_to_file(cfg, pre='cfg', logger=None):\n for key, val in cfg.items():\n if isinstance(cfg[key], EasyDict):\n print_log(f'{pre}.{key} = edict()', logger = logger)\n log_config_to_file(cfg[key], pre=pre + '.' + key, logger=logger)\n continue\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.log_config_to_file","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.log_config_to_file#L10-L16","kind":"function","name":"log_config_to_file","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":10,"end_line":16,"context_start_line":1,"context_end_line":36,"code":"import yaml\nfrom easydict import EasyDict\nimport os\nfrom .logger import print_log\n\ndef log_args_to_file(args, pre='args', logger=None):\n for key, val in args.__dict__.items():\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef log_config_to_file(cfg, pre='cfg', logger=None):\n for key, val in cfg.items():\n if isinstance(cfg[key], EasyDict):\n print_log(f'{pre}.{key} = edict()', logger = logger)\n log_config_to_file(cfg[key], pre=pre + '.' + key, logger=logger)\n continue\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\n","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.merge_new_config","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.merge_new_config#L18-L35","kind":"function","name":"merge_new_config","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":18,"end_line":35,"context_start_line":1,"context_end_line":55,"code":"import yaml\nfrom easydict import EasyDict\nimport os\nfrom .logger import print_log\n\ndef log_args_to_file(args, pre='args', logger=None):\n for key, val in args.__dict__.items():\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef log_config_to_file(cfg, pre='cfg', logger=None):\n for key, val in cfg.items():\n if isinstance(cfg[key], EasyDict):\n print_log(f'{pre}.{key} = edict()', logger = logger)\n log_config_to_file(cfg[key], pre=pre + '.' + key, logger=logger)\n continue\n print_log(f'{pre}.{key} : {val}', logger = logger)\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\n\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n except:\n new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_config(args, logger=None):\n if args.resume:\n cfg_path = os.path.join(args.experiment_path, 'config.yaml')\n if not os.path.exists(cfg_path):\n print_log(\"Failed to resume\", logger = logger)\n raise FileNotFoundError()\n print_log(f'Resume yaml from {cfg_path}', logger = logger)\n args.config = cfg_path\n config = cfg_from_yaml_file(args.config)","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.cfg_from_yaml_file","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.cfg_from_yaml_file#L37-L45","kind":"function","name":"cfg_from_yaml_file","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":37,"end_line":45,"context_start_line":17,"context_end_line":63,"code":"\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\n\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n except:\n new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_config(args, logger=None):\n if args.resume:\n cfg_path = os.path.join(args.experiment_path, 'config.yaml')\n if not os.path.exists(cfg_path):\n print_log(\"Failed to resume\", logger = logger)\n raise FileNotFoundError()\n print_log(f'Resume yaml from {cfg_path}', logger = logger)\n args.config = cfg_path\n config = cfg_from_yaml_file(args.config)\n if not args.resume and args.local_rank == 0:\n save_experiment_config(args, config, logger)\n return config\n\ndef save_experiment_config(args, config, logger = None):\n config_path = os.path.join(args.experiment_path, 'config.yaml')\n os.system('cp %s %s' % (args.config, config_path))\n print_log(f'Copy the Config file from {args.config} to {config_path}',logger = logger )","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.get_config","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.get_config#L47-L58","kind":"function","name":"get_config","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":47,"end_line":58,"context_start_line":27,"context_end_line":63,"code":" config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\n\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n except:\n new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_config(args, logger=None):\n if args.resume:\n cfg_path = os.path.join(args.experiment_path, 'config.yaml')\n if not os.path.exists(cfg_path):\n print_log(\"Failed to resume\", logger = logger)\n raise FileNotFoundError()\n print_log(f'Resume yaml from {cfg_path}', logger = logger)\n args.config = cfg_path\n config = cfg_from_yaml_file(args.config)\n if not args.resume and args.local_rank == 0:\n save_experiment_config(args, config, logger)\n return config\n\ndef save_experiment_config(args, config, logger = None):\n config_path = os.path.join(args.experiment_path, 'config.yaml')\n os.system('cp %s %s' % (args.config, config_path))\n print_log(f'Copy the Config file from {args.config} to {config_path}',logger = logger )","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.config.save_experiment_config","uri":"program://CREMA/function/lavis.models.ulip_models.utils.config.save_experiment_config#L60-L63","kind":"function","name":"save_experiment_config","path":"lavis/models/ulip_models/utils/config.py","language":"python","start_line":60,"end_line":63,"context_start_line":40,"context_end_line":63,"code":" try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n except:\n new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_config(args, logger=None):\n if args.resume:\n cfg_path = os.path.join(args.experiment_path, 'config.yaml')\n if not os.path.exists(cfg_path):\n print_log(\"Failed to resume\", logger = logger)\n raise FileNotFoundError()\n print_log(f'Resume yaml from {cfg_path}', logger = logger)\n args.config = cfg_path\n config = cfg_from_yaml_file(args.config)\n if not args.resume and args.local_rank == 0:\n save_experiment_config(args, config, logger)\n return config\n\ndef save_experiment_config(args, config, logger = None):\n config_path = os.path.join(args.experiment_path, 'config.yaml')\n os.system('cp %s %s' % (args.config, config_path))\n print_log(f'Copy the Config file from {args.config} to {config_path}',logger = logger )","source_hash":"a718798663c866aa817aed958566f5320d56a733f38d6fff65057be3f408b927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io","uri":"program://CREMA/module/lavis.models.ulip_models.utils.io#L1-L42","kind":"module","name":"lavis.models.ulip_models.utils.io","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":1,"end_line":42,"context_start_line":1,"context_end_line":42,"code":"import h5py\nimport numpy as np\nimport open3d\nimport os\n\nclass IO:\n @classmethod\n def get(cls, file_path):\n _, file_extension = os.path.splitext(file_path)\n\n if file_extension in ['.npy']:\n return cls._read_npy(file_path)\n elif file_extension in ['.pcd']:\n return cls._read_pcd(file_path)\n elif file_extension in ['.h5']:\n return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io.IO","uri":"program://CREMA/class/lavis.models.ulip_models.utils.io.IO#L6-L42","kind":"class","name":"IO","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":6,"end_line":42,"context_start_line":1,"context_end_line":42,"code":"import h5py\nimport numpy as np\nimport open3d\nimport os\n\nclass IO:\n @classmethod\n def get(cls, file_path):\n _, file_extension = os.path.splitext(file_path)\n\n if file_extension in ['.npy']:\n return cls._read_npy(file_path)\n elif file_extension in ['.pcd']:\n return cls._read_pcd(file_path)\n elif file_extension in ['.h5']:\n return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io.get","uri":"program://CREMA/function/lavis.models.ulip_models.utils.io.get#L8-L20","kind":"function","name":"get","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":8,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"import h5py\nimport numpy as np\nimport open3d\nimport os\n\nclass IO:\n @classmethod\n def get(cls, file_path):\n _, file_extension = os.path.splitext(file_path)\n\n if file_extension in ['.npy']:\n return cls._read_npy(file_path)\n elif file_extension in ['.pcd']:\n return cls._read_pcd(file_path)\n elif file_extension in ['.h5']:\n return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io._read_npy","uri":"program://CREMA/function/lavis.models.ulip_models.utils.io._read_npy#L24-L25","kind":"function","name":"_read_npy","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":24,"end_line":25,"context_start_line":4,"context_end_line":42,"code":"import os\n\nclass IO:\n @classmethod\n def get(cls, file_path):\n _, file_extension = os.path.splitext(file_path)\n\n if file_extension in ['.npy']:\n return cls._read_npy(file_path)\n elif file_extension in ['.pcd']:\n return cls._read_pcd(file_path)\n elif file_extension in ['.h5']:\n return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io._read_pcd","uri":"program://CREMA/function/lavis.models.ulip_models.utils.io._read_pcd#L30-L33","kind":"function","name":"_read_pcd","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":30,"end_line":33,"context_start_line":10,"context_end_line":42,"code":"\n if file_extension in ['.npy']:\n return cls._read_npy(file_path)\n elif file_extension in ['.pcd']:\n return cls._read_pcd(file_path)\n elif file_extension in ['.h5']:\n return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io._read_txt","uri":"program://CREMA/function/lavis.models.ulip_models.utils.io._read_txt#L36-L37","kind":"function","name":"_read_txt","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":36,"end_line":37,"context_start_line":16,"context_end_line":42,"code":" return cls._read_h5(file_path)\n elif file_extension in ['.txt']:\n return cls._read_txt(file_path)\n else:\n raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.io._read_h5","uri":"program://CREMA/function/lavis.models.ulip_models.utils.io._read_h5#L40-L42","kind":"function","name":"_read_h5","path":"lavis/models/ulip_models/utils/io.py","language":"python","start_line":40,"end_line":42,"context_start_line":20,"context_end_line":42,"code":" raise Exception('Unsupported file extension: %s' % file_extension)\n\n # References: https://github.com/numpy/numpy/blob/master/numpy/lib/format.py\n @classmethod\n def _read_npy(cls, file_path):\n return np.load(file_path)\n \n # References: https://github.com/dimatura/pypcd/blob/master/pypcd/pypcd.py#L275\n # Support PCD files without compression ONLY!\n @classmethod\n def _read_pcd(cls, file_path):\n pc = open3d.io.read_point_cloud(file_path)\n ptcloud = np.array(pc.points)\n return ptcloud\n\n @classmethod\n def _read_txt(cls, file_path):\n return np.loadtxt(file_path)\n\n @classmethod\n def _read_h5(cls, file_path):\n f = h5py.File(file_path, 'r')\n return f['data'][()]","source_hash":"b15d29f319b4c29bd7b3d4a973e151f2303c918235e756296c389d414f05b491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry","uri":"program://CREMA/module/lavis.models.ulip_models.utils.registry#L1-L288","kind":"module","name":"lavis.models.ulip_models.utils.registry","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":1,"end_line":288,"context_start_line":1,"context_end_line":288,"code":"import inspect\nimport warnings\nfrom functools import partial\nfrom . import config\n\nclass Registry:\n \"\"\"A registry to map strings to classes.\n Registered object could be built from registry.\n Example:\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = MODELS.build(dict(NAME='ResNet'))\n Please refer to https://mmcv.readthedocs.io/en/latest/registry.html for\n advanced useage.\n Args:\n name (str): Registry name.\n build_func(func, optional): Build function to construct instance from\n Registry, func:`build_from_cfg` is used if neither ``parent`` or\n ``build_func`` is specified. If ``parent`` is specified and\n ``build_func`` is not given, ``build_func`` will be inherited\n from ``parent``. Default: None.\n parent (Registry, optional): Parent registry. The class registered in\n children registry could be built from parent. Default: None.\n scope (str, optional): The scope of registry. It is the key to search\n for children registry. If not specified, scope will be the name of\n the package where class is defined, e.g. mmdet, mmcls, mmseg.\n Default: None.\n \"\"\"\n\n def __init__(self, name, build_func=None, parent=None, scope=None):\n self._name = name\n self._module_dict = dict()\n self._children = dict()\n self._scope = self.infer_scope() if scope is None else scope\n\n # self.build_func will be set with the following priority:\n # 1. build_func\n # 2. parent.build_func\n # 3. build_from_cfg\n if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.\n \"\"\"\n # inspect.stack() trace where this function is called, the index-2\n # indicates the frame where `infer_scope()` is called\n filename = inspect.getmodule(inspect.stack()[2][0]).__name__\n split_filename = filename.split('.')\n return split_filename[0]\n\n @staticmethod\n def split_scope_key(key):\n \"\"\"Split scope and key.\n The first scope will be split from key.\n Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent\n while parent.parent is not None:\n parent = parent.parent\n return parent.get(key)\n\n def build(self, *args, **kwargs):\n return self.build_func(*args, **kwargs, registry=self)\n\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n\n def _register_module(self, module_class, module_name=None, force=False):\n if not inspect.isclass(module_class):\n raise TypeError('module must be a class, '\n f'but got {type(module_class)}')\n\n if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '\n 'register_module(name=None, force=False, module=None) instead.')\n if cls is None:\n return partial(self.deprecated_register_module, force=force)\n self._register_module(cls, force=force)\n return cls\n\n def register_module(self, name=None, force=False, module=None):\n \"\"\"Register a module.\n A record will be added to `self._module_dict`, whose key is the class\n name or the specified name, and value is the class itself.\n It can be used as a decorator or a normal function.\n Example:\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module()\n >>> class ResNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module(name='mnet')\n >>> class MobileNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> class ResNet:\n >>> pass\n >>> backbones.register_module(ResNet)\n Args:\n name (str | None): The module name to be registered. If not\n specified, the class name will be used.\n force (bool, optional): Whether to override an existing class with\n the same name. Default: False.\n module (type): Module class to be registered.\n \"\"\"\n if not isinstance(force, bool):\n raise TypeError(f'force must be a boolean, but got {type(force)}')\n # NOTE: This is a walkaround to be compatible with the old api,\n # while it may introduce unexpected bugs.\n if isinstance(name, type):\n return self.deprecated_register_module(name, force=force)\n\n # raise the error ahead of time\n if not (name is None or isinstance(name, str) or misc.is_seq_of(name, str)):\n raise TypeError(\n 'name must be either of None, an instance of str or a sequence'\n f' of str, but got {type(name)}')\n\n # use it as a normal method: x.register_module(module=SomeClass)\n if module is not None:\n self._register_module(\n module_class=module, module_name=name, force=force)\n return module\n\n # use it as a decorator: @x.register_module()\n def _register(cls):\n self._register_module(\n module_class=cls, module_name=name, force=force)\n return cls\n\n return _register\n\n\ndef build_from_cfg(cfg, registry, default_args=None):\n \"\"\"Build a module from config dict.\n Args:\n cfg (edict): Config dict. It should at least contain the key \"NAME\".\n registry (:obj:`Registry`): The registry to search the type from.\n Returns:\n object: The constructed object.\n \"\"\"\n if not isinstance(cfg, dict):\n raise TypeError(f'cfg must be a dict, but got {type(cfg)}')\n if 'NAME' not in cfg:\n if default_args is None or 'NAME' not in default_args:\n raise KeyError(\n '`cfg` or `default_args` must contain the key \"NAME\", '\n f'but got {cfg}\\n{default_args}')\n if not isinstance(registry, Registry):\n raise TypeError('registry must be an mmcv.Registry object, '\n f'but got {type(registry)}')\n\n if not (isinstance(default_args, dict) or default_args is None):\n raise TypeError('default_args must be a dict or None, '\n f'but got {type(default_args)}')\n\n if default_args is not None:\n cfg = config.merge_new_config(cfg, default_args)\n\n obj_type = cfg.get('NAME')\n\n if isinstance(obj_type, str):\n obj_cls = registry.get(obj_type)\n if obj_cls is None:\n raise KeyError(\n f'{obj_type} is not in the {registry.name} registry')\n elif inspect.isclass(obj_type):\n obj_cls = obj_type\n else:\n raise TypeError(\n f'type must be a str or valid type, but got {type(obj_type)}')\n try:\n return obj_cls(cfg)\n except Exception as e:\n # Normal TypeError does not print class name.\n raise type(e)(f'{obj_cls.__name__}: {e}')","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.Registry","uri":"program://CREMA/class/lavis.models.ulip_models.utils.registry.Registry#L6-L243","kind":"class","name":"Registry","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":6,"end_line":243,"context_start_line":1,"context_end_line":263,"code":"import inspect\nimport warnings\nfrom functools import partial\nfrom . import config\n\nclass Registry:\n \"\"\"A registry to map strings to classes.\n Registered object could be built from registry.\n Example:\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = MODELS.build(dict(NAME='ResNet'))\n Please refer to https://mmcv.readthedocs.io/en/latest/registry.html for\n advanced useage.\n Args:\n name (str): Registry name.\n build_func(func, optional): Build function to construct instance from\n Registry, func:`build_from_cfg` is used if neither ``parent`` or\n ``build_func`` is specified. If ``parent`` is specified and\n ``build_func`` is not given, ``build_func`` will be inherited\n from ``parent``. Default: None.\n parent (Registry, optional): Parent registry. The class registered in\n children registry could be built from parent. Default: None.\n scope (str, optional): The scope of registry. It is the key to search\n for children registry. If not specified, scope will be the name of\n the package where class is defined, e.g. mmdet, mmcls, mmseg.\n Default: None.\n \"\"\"\n\n def __init__(self, name, build_func=None, parent=None, scope=None):\n self._name = name\n self._module_dict = dict()\n self._children = dict()\n self._scope = self.infer_scope() if scope is None else scope\n\n # self.build_func will be set with the following priority:\n # 1. build_func\n # 2. parent.build_func\n # 3. build_from_cfg\n if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.\n \"\"\"\n # inspect.stack() trace where this function is called, the index-2\n # indicates the frame where `infer_scope()` is called\n filename = inspect.getmodule(inspect.stack()[2][0]).__name__\n split_filename = filename.split('.')\n return split_filename[0]\n\n @staticmethod\n def split_scope_key(key):\n \"\"\"Split scope and key.\n The first scope will be split from key.\n Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent\n while parent.parent is not None:\n parent = parent.parent\n return parent.get(key)\n\n def build(self, *args, **kwargs):\n return self.build_func(*args, **kwargs, registry=self)\n\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n\n def _register_module(self, module_class, module_name=None, force=False):\n if not inspect.isclass(module_class):\n raise TypeError('module must be a class, '\n f'but got {type(module_class)}')\n\n if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '\n 'register_module(name=None, force=False, module=None) instead.')\n if cls is None:\n return partial(self.deprecated_register_module, force=force)\n self._register_module(cls, force=force)\n return cls\n\n def register_module(self, name=None, force=False, module=None):\n \"\"\"Register a module.\n A record will be added to `self._module_dict`, whose key is the class\n name or the specified name, and value is the class itself.\n It can be used as a decorator or a normal function.\n Example:\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module()\n >>> class ResNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module(name='mnet')\n >>> class MobileNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> class ResNet:\n >>> pass\n >>> backbones.register_module(ResNet)\n Args:\n name (str | None): The module name to be registered. If not\n specified, the class name will be used.\n force (bool, optional): Whether to override an existing class with\n the same name. Default: False.\n module (type): Module class to be registered.\n \"\"\"\n if not isinstance(force, bool):\n raise TypeError(f'force must be a boolean, but got {type(force)}')\n # NOTE: This is a walkaround to be compatible with the old api,\n # while it may introduce unexpected bugs.\n if isinstance(name, type):\n return self.deprecated_register_module(name, force=force)\n\n # raise the error ahead of time\n if not (name is None or isinstance(name, str) or misc.is_seq_of(name, str)):\n raise TypeError(\n 'name must be either of None, an instance of str or a sequence'\n f' of str, but got {type(name)}')\n\n # use it as a normal method: x.register_module(module=SomeClass)\n if module is not None:\n self._register_module(\n module_class=module, module_name=name, force=force)\n return module\n\n # use it as a decorator: @x.register_module()\n def _register(cls):\n self._register_module(\n module_class=cls, module_name=name, force=force)\n return cls\n\n return _register\n\n\ndef build_from_cfg(cfg, registry, default_args=None):\n \"\"\"Build a module from config dict.\n Args:\n cfg (edict): Config dict. It should at least contain the key \"NAME\".\n registry (:obj:`Registry`): The registry to search the type from.\n Returns:\n object: The constructed object.\n \"\"\"\n if not isinstance(cfg, dict):\n raise TypeError(f'cfg must be a dict, but got {type(cfg)}')\n if 'NAME' not in cfg:\n if default_args is None or 'NAME' not in default_args:\n raise KeyError(\n '`cfg` or `default_args` must contain the key \"NAME\", '\n f'but got {cfg}\\n{default_args}')\n if not isinstance(registry, Registry):\n raise TypeError('registry must be an mmcv.Registry object, '\n f'but got {type(registry)}')","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.build_from_cfg","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.build_from_cfg#L246-L288","kind":"function","name":"build_from_cfg","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":246,"end_line":288,"context_start_line":226,"context_end_line":288,"code":" if not (name is None or isinstance(name, str) or misc.is_seq_of(name, str)):\n raise TypeError(\n 'name must be either of None, an instance of str or a sequence'\n f' of str, but got {type(name)}')\n\n # use it as a normal method: x.register_module(module=SomeClass)\n if module is not None:\n self._register_module(\n module_class=module, module_name=name, force=force)\n return module\n\n # use it as a decorator: @x.register_module()\n def _register(cls):\n self._register_module(\n module_class=cls, module_name=name, force=force)\n return cls\n\n return _register\n\n\ndef build_from_cfg(cfg, registry, default_args=None):\n \"\"\"Build a module from config dict.\n Args:\n cfg (edict): Config dict. It should at least contain the key \"NAME\".\n registry (:obj:`Registry`): The registry to search the type from.\n Returns:\n object: The constructed object.\n \"\"\"\n if not isinstance(cfg, dict):\n raise TypeError(f'cfg must be a dict, but got {type(cfg)}')\n if 'NAME' not in cfg:\n if default_args is None or 'NAME' not in default_args:\n raise KeyError(\n '`cfg` or `default_args` must contain the key \"NAME\", '\n f'but got {cfg}\\n{default_args}')\n if not isinstance(registry, Registry):\n raise TypeError('registry must be an mmcv.Registry object, '\n f'but got {type(registry)}')\n\n if not (isinstance(default_args, dict) or default_args is None):\n raise TypeError('default_args must be a dict or None, '\n f'but got {type(default_args)}')\n\n if default_args is not None:\n cfg = config.merge_new_config(cfg, default_args)\n\n obj_type = cfg.get('NAME')\n\n if isinstance(obj_type, str):\n obj_cls = registry.get(obj_type)\n if obj_cls is None:\n raise KeyError(\n f'{obj_type} is not in the {registry.name} registry')\n elif inspect.isclass(obj_type):\n obj_cls = obj_type\n else:\n raise TypeError(\n f'type must be a str or valid type, but got {type(obj_type)}')\n try:\n return obj_cls(cfg)\n except Exception as e:\n # Normal TypeError does not print class name.\n raise type(e)(f'{obj_cls.__name__}: {e}')","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.__init__#L32-L54","kind":"function","name":"__init__","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":32,"end_line":54,"context_start_line":12,"context_end_line":74,"code":" >>> class ResNet:\n >>> pass\n >>> resnet = MODELS.build(dict(NAME='ResNet'))\n Please refer to https://mmcv.readthedocs.io/en/latest/registry.html for\n advanced useage.\n Args:\n name (str): Registry name.\n build_func(func, optional): Build function to construct instance from\n Registry, func:`build_from_cfg` is used if neither ``parent`` or\n ``build_func`` is specified. If ``parent`` is specified and\n ``build_func`` is not given, ``build_func`` will be inherited\n from ``parent``. Default: None.\n parent (Registry, optional): Parent registry. The class registered in\n children registry could be built from parent. Default: None.\n scope (str, optional): The scope of registry. It is the key to search\n for children registry. If not specified, scope will be the name of\n the package where class is defined, e.g. mmdet, mmcls, mmseg.\n Default: None.\n \"\"\"\n\n def __init__(self, name, build_func=None, parent=None, scope=None):\n self._name = name\n self._module_dict = dict()\n self._children = dict()\n self._scope = self.infer_scope() if scope is None else scope\n\n # self.build_func will be set with the following priority:\n # 1. build_func\n # 2. parent.build_func\n # 3. build_from_cfg\n if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.__len__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.__len__#L56-L57","kind":"function","name":"__len__","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":56,"end_line":57,"context_start_line":36,"context_end_line":77,"code":" self._scope = self.infer_scope() if scope is None else scope\n\n # self.build_func will be set with the following priority:\n # 1. build_func\n # 2. parent.build_func\n # 3. build_from_cfg\n if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.__contains__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.__contains__#L59-L60","kind":"function","name":"__contains__","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":59,"end_line":60,"context_start_line":39,"context_end_line":80,"code":" # 1. build_func\n # 2. parent.build_func\n # 3. build_from_cfg\n if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.__repr__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.__repr__#L62-L66","kind":"function","name":"__repr__","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":62,"end_line":66,"context_start_line":42,"context_end_line":86,"code":" if build_func is None:\n if parent is not None:\n self.build_func = parent.build_func\n else:\n self.build_func = build_from_cfg\n else:\n self.build_func = build_func\n if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.\n \"\"\"\n # inspect.stack() trace where this function is called, the index-2\n # indicates the frame where `infer_scope()` is called\n filename = inspect.getmodule(inspect.stack()[2][0]).__name__\n split_filename = filename.split('.')\n return split_filename[0]","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.infer_scope","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.infer_scope#L69-L86","kind":"function","name":"infer_scope","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":69,"end_line":86,"context_start_line":49,"context_end_line":106,"code":" if parent is not None:\n assert isinstance(parent, Registry)\n parent._add_children(self)\n self.parent = parent\n else:\n self.parent = None\n\n def __len__(self):\n return len(self._module_dict)\n\n def __contains__(self, key):\n return self.get(key) is not None\n\n def __repr__(self):\n format_str = self.__class__.__name__ + \\\n f'(name={self._name}, ' \\\n f'items={self._module_dict})'\n return format_str\n\n @staticmethod\n def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.\n \"\"\"\n # inspect.stack() trace where this function is called, the index-2\n # indicates the frame where `infer_scope()` is called\n filename = inspect.getmodule(inspect.stack()[2][0]).__name__\n split_filename = filename.split('.')\n return split_filename[0]\n\n @staticmethod\n def split_scope_key(key):\n \"\"\"Split scope and key.\n The first scope will be split from key.\n Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.split_scope_key","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.split_scope_key#L89-L105","kind":"function","name":"split_scope_key","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":89,"end_line":105,"context_start_line":69,"context_end_line":125,"code":" def infer_scope():\n \"\"\"Infer the scope of registry.\n The name of the package where registry is defined will be returned.\n Example:\n # in mmdet/models/backbone/resnet.py\n >>> MODELS = Registry('models')\n >>> @MODELS.register_module()\n >>> class ResNet:\n >>> pass\n The scope of ``ResNet`` will be ``mmdet``.\n Returns:\n scope (str): The inferred scope name.\n \"\"\"\n # inspect.stack() trace where this function is called, the index-2\n # indicates the frame where `infer_scope()` is called\n filename = inspect.getmodule(inspect.stack()[2][0]).__name__\n split_filename = filename.split('.')\n return split_filename[0]\n\n @staticmethod\n def split_scope_key(key):\n \"\"\"Split scope and key.\n The first scope will be split from key.\n Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.name","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.name#L108-L109","kind":"function","name":"name","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":108,"end_line":109,"context_start_line":88,"context_end_line":129,"code":" @staticmethod\n def split_scope_key(key):\n \"\"\"Split scope and key.\n The first scope will be split from key.\n Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.scope","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.scope#L112-L113","kind":"function","name":"scope","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":112,"end_line":113,"context_start_line":92,"context_end_line":133,"code":" Examples:\n >>> Registry.split_scope_key('mmdet.ResNet')\n 'mmdet', 'ResNet'\n >>> Registry.split_scope_key('ResNet')\n None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.module_dict","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.module_dict#L116-L117","kind":"function","name":"module_dict","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":116,"end_line":117,"context_start_line":96,"context_end_line":137,"code":" None, 'ResNet'\n Return:\n scope (str, None): The first scope.\n key (str): The remaining key.\n \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.children","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.children#L120-L121","kind":"function","name":"children","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":120,"end_line":121,"context_start_line":100,"context_end_line":141,"code":" \"\"\"\n split_index = key.find('.')\n if split_index != -1:\n return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.get","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.get#L123-L144","kind":"function","name":"get","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":123,"end_line":144,"context_start_line":103,"context_end_line":164,"code":" return key[:split_index], key[split_index + 1:]\n else:\n return None, key\n\n @property\n def name(self):\n return self._name\n\n @property\n def scope(self):\n return self._scope\n\n @property\n def module_dict(self):\n return self._module_dict\n\n @property\n def children(self):\n return self._children\n\n def get(self, key):\n \"\"\"Get the registry record.\n Args:\n key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent\n while parent.parent is not None:\n parent = parent.parent\n return parent.get(key)\n\n def build(self, *args, **kwargs):\n return self.build_func(*args, **kwargs, registry=self)\n\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.build","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.build#L146-L147","kind":"function","name":"build","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":146,"end_line":147,"context_start_line":126,"context_end_line":167,"code":" key (str): The class name in string format.\n Returns:\n class: The corresponding class.\n \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent\n while parent.parent is not None:\n parent = parent.parent\n return parent.get(key)\n\n def build(self, *args, **kwargs):\n return self.build_func(*args, **kwargs, registry=self)\n\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry._add_children","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry._add_children#L149-L166","kind":"function","name":"_add_children","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":149,"end_line":166,"context_start_line":129,"context_end_line":186,"code":" \"\"\"\n scope, real_key = self.split_scope_key(key)\n if scope is None or scope == self._scope:\n # get from self\n if real_key in self._module_dict:\n return self._module_dict[real_key]\n else:\n # get from self._children\n if scope in self._children:\n return self._children[scope].get(real_key)\n else:\n # goto root\n parent = self.parent\n while parent.parent is not None:\n parent = parent.parent\n return parent.get(key)\n\n def build(self, *args, **kwargs):\n return self.build_func(*args, **kwargs, registry=self)\n\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n\n def _register_module(self, module_class, module_name=None, force=False):\n if not inspect.isclass(module_class):\n raise TypeError('module must be a class, '\n f'but got {type(module_class)}')\n\n if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry._register_module","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry._register_module#L168-L181","kind":"function","name":"_register_module","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":168,"end_line":181,"context_start_line":148,"context_end_line":201,"code":"\n def _add_children(self, registry):\n \"\"\"Add children for a registry.\n The ``registry`` will be added as children based on its scope.\n The parent registry could build objects from children registry.\n Example:\n >>> models = Registry('models')\n >>> mmdet_models = Registry('models', parent=models)\n >>> @mmdet_models.register_module()\n >>> class ResNet:\n >>> pass\n >>> resnet = models.build(dict(NAME='mmdet.ResNet'))\n \"\"\"\n\n assert isinstance(registry, Registry)\n assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n\n def _register_module(self, module_class, module_name=None, force=False):\n if not inspect.isclass(module_class):\n raise TypeError('module must be a class, '\n f'but got {type(module_class)}')\n\n if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '\n 'register_module(name=None, force=False, module=None) instead.')\n if cls is None:\n return partial(self.deprecated_register_module, force=force)\n self._register_module(cls, force=force)\n return cls\n\n def register_module(self, name=None, force=False, module=None):\n \"\"\"Register a module.\n A record will be added to `self._module_dict`, whose key is the class\n name or the specified name, and value is the class itself.\n It can be used as a decorator or a normal function.\n Example:\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module()\n >>> class ResNet:","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.deprecated_register_module","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.deprecated_register_module#L183-L191","kind":"function","name":"deprecated_register_module","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":183,"end_line":191,"context_start_line":163,"context_end_line":211,"code":" assert registry.scope is not None\n assert registry.scope not in self.children, \\\n f'scope {registry.scope} exists in {self.name} registry'\n self.children[registry.scope] = registry\n\n def _register_module(self, module_class, module_name=None, force=False):\n if not inspect.isclass(module_class):\n raise TypeError('module must be a class, '\n f'but got {type(module_class)}')\n\n if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '\n 'register_module(name=None, force=False, module=None) instead.')\n if cls is None:\n return partial(self.deprecated_register_module, force=force)\n self._register_module(cls, force=force)\n return cls\n\n def register_module(self, name=None, force=False, module=None):\n \"\"\"Register a module.\n A record will be added to `self._module_dict`, whose key is the class\n name or the specified name, and value is the class itself.\n It can be used as a decorator or a normal function.\n Example:\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module()\n >>> class ResNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module(name='mnet')\n >>> class MobileNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> class ResNet:\n >>> pass\n >>> backbones.register_module(ResNet)\n Args:","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry.register_module","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry.register_module#L193-L243","kind":"function","name":"register_module","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":193,"end_line":243,"context_start_line":173,"context_end_line":263,"code":" if module_name is None:\n module_name = module_class.__name__\n if isinstance(module_name, str):\n module_name = [module_name]\n for name in module_name:\n if not force and name in self._module_dict:\n raise KeyError(f'{name} is already registered '\n f'in {self.name}')\n self._module_dict[name] = module_class\n\n def deprecated_register_module(self, cls=None, force=False):\n warnings.warn(\n 'The old API of register_module(module, force=False) '\n 'is deprecated and will be removed, please use the new API '\n 'register_module(name=None, force=False, module=None) instead.')\n if cls is None:\n return partial(self.deprecated_register_module, force=force)\n self._register_module(cls, force=force)\n return cls\n\n def register_module(self, name=None, force=False, module=None):\n \"\"\"Register a module.\n A record will be added to `self._module_dict`, whose key is the class\n name or the specified name, and value is the class itself.\n It can be used as a decorator or a normal function.\n Example:\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module()\n >>> class ResNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> @backbones.register_module(name='mnet')\n >>> class MobileNet:\n >>> pass\n >>> backbones = Registry('backbone')\n >>> class ResNet:\n >>> pass\n >>> backbones.register_module(ResNet)\n Args:\n name (str | None): The module name to be registered. If not\n specified, the class name will be used.\n force (bool, optional): Whether to override an existing class with\n the same name. Default: False.\n module (type): Module class to be registered.\n \"\"\"\n if not isinstance(force, bool):\n raise TypeError(f'force must be a boolean, but got {type(force)}')\n # NOTE: This is a walkaround to be compatible with the old api,\n # while it may introduce unexpected bugs.\n if isinstance(name, type):\n return self.deprecated_register_module(name, force=force)\n\n # raise the error ahead of time\n if not (name is None or isinstance(name, str) or misc.is_seq_of(name, str)):\n raise TypeError(\n 'name must be either of None, an instance of str or a sequence'\n f' of str, but got {type(name)}')\n\n # use it as a normal method: x.register_module(module=SomeClass)\n if module is not None:\n self._register_module(\n module_class=module, module_name=name, force=force)\n return module\n\n # use it as a decorator: @x.register_module()\n def _register(cls):\n self._register_module(\n module_class=cls, module_name=name, force=force)\n return cls\n\n return _register\n\n\ndef build_from_cfg(cfg, registry, default_args=None):\n \"\"\"Build a module from config dict.\n Args:\n cfg (edict): Config dict. It should at least contain the key \"NAME\".\n registry (:obj:`Registry`): The registry to search the type from.\n Returns:\n object: The constructed object.\n \"\"\"\n if not isinstance(cfg, dict):\n raise TypeError(f'cfg must be a dict, but got {type(cfg)}')\n if 'NAME' not in cfg:\n if default_args is None or 'NAME' not in default_args:\n raise KeyError(\n '`cfg` or `default_args` must contain the key \"NAME\", '\n f'but got {cfg}\\n{default_args}')\n if not isinstance(registry, Registry):\n raise TypeError('registry must be an mmcv.Registry object, '\n f'but got {type(registry)}')","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.registry._register","uri":"program://CREMA/function/lavis.models.ulip_models.utils.registry._register#L238-L241","kind":"function","name":"_register","path":"lavis/models/ulip_models/utils/registry.py","language":"python","start_line":238,"end_line":241,"context_start_line":218,"context_end_line":261,"code":" if not isinstance(force, bool):\n raise TypeError(f'force must be a boolean, but got {type(force)}')\n # NOTE: This is a walkaround to be compatible with the old api,\n # while it may introduce unexpected bugs.\n if isinstance(name, type):\n return self.deprecated_register_module(name, force=force)\n\n # raise the error ahead of time\n if not (name is None or isinstance(name, str) or misc.is_seq_of(name, str)):\n raise TypeError(\n 'name must be either of None, an instance of str or a sequence'\n f' of str, but got {type(name)}')\n\n # use it as a normal method: x.register_module(module=SomeClass)\n if module is not None:\n self._register_module(\n module_class=module, module_name=name, force=force)\n return module\n\n # use it as a decorator: @x.register_module()\n def _register(cls):\n self._register_module(\n module_class=cls, module_name=name, force=force)\n return cls\n\n return _register\n\n\ndef build_from_cfg(cfg, registry, default_args=None):\n \"\"\"Build a module from config dict.\n Args:\n cfg (edict): Config dict. It should at least contain the key \"NAME\".\n registry (:obj:`Registry`): The registry to search the type from.\n Returns:\n object: The constructed object.\n \"\"\"\n if not isinstance(cfg, dict):\n raise TypeError(f'cfg must be a dict, but got {type(cfg)}')\n if 'NAME' not in cfg:\n if default_args is None or 'NAME' not in default_args:\n raise KeyError(\n '`cfg` or `default_args` must contain the key \"NAME\", '\n f'but got {cfg}\\n{default_args}')\n if not isinstance(registry, Registry):","source_hash":"27573c4851a6db807956d371bd8d1f01f0fee5fd0b92c53496355db1480d285b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils","uri":"program://CREMA/module/lavis.models.ulip_models.utils.utils#L1-L242","kind":"module","name":"lavis.models.ulip_models.utils.utils","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":1,"end_line":242,"context_start_line":1,"context_end_line":242,"code":"import numpy as np\nimport os\nimport random\nimport shutil\nimport torch\nimport torch.distributed as dist\nimport torch.autograd as autograd\n\nfrom PIL import ImageFilter\nfrom easydict import EasyDict\nimport yaml\n# from data.dataset_3d import Dataset_3D\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n # try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n\n\ndef init_distributed_mode(args):\n if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:\n args.rank = int(os.environ[\"RANK\"])\n args.world_size = int(os.environ['WORLD_SIZE'])\n args.gpu = int(os.environ['LOCAL_RANK'])\n elif 'SLURM_PROCID' in os.environ:\n args.rank = int(os.environ['SLURM_PROCID'])\n args.gpu = args.rank % torch.cuda.device_count()\n else:\n print('Not using distributed mode')\n args.distributed = False\n return\n\n args.distributed = True\n\n torch.cuda.set_device(args.gpu)\n args.dist_backend = 'nccl'\n print('| distributed init (rank {}): {}'.format(\n args.rank, args.dist_url), flush=True)\n torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url,\n world_size=args.world_size, rank=args.rank)\n torch.distributed.barrier()\n setup_for_distributed(args.rank == 0)\n\n\ndef scaled_all_reduce(tensors, is_scale=True):\n \"\"\"Performs the scaled all_reduce operation on the provided tensors.\n The input tensors are modified in-place. Currently supports only the sum\n reduction operator. The reduced values are scaled by the inverse size of the\n world size.\n \"\"\"\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n # Queue the reductions\n reductions = []\n for tensor in tensors:\n reduction = dist.all_reduce(tensor, async_op=True)\n reductions.append(reduction)\n # Wait for reductions to finish\n for reduction in reductions:\n reduction.wait()\n # Scale the results\n if is_scale:\n for tensor in tensors:\n tensor.mul_(1.0 / world_size)\n return tensors\n\n\ndef all_gather_batch(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n for tensor in tensors:\n tensor_all = [torch.ones_like(tensor) for _ in range(world_size)]\n dist.all_gather(\n tensor_all,\n tensor,\n async_op=False # performance opt\n )\n\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n\n\ndef all_gather_batch_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n\n for tensor in tensors:\n tensor_all = GatherLayer.apply(tensor)\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\ndef cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):\n self.sigma = sigma\n\n def __call__(self, x):\n sigma = random.uniform(self.sigma[0], self.sigma[1])\n x = x.filter(ImageFilter.GaussianBlur(radius=sigma))\n return x\n\n# def get_dataset(train_transform, tokenizer, args, dataset_name=None, files_list=None):\n# dataset_3d = Dataset_3D(args, tokenizer, dataset_name, train_transform, files_list=files_list)\n# return dataset_3d.dataset","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.merge_new_config","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.merge_new_config#L14-L31","kind":"function","name":"merge_new_config","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":14,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"import numpy as np\nimport os\nimport random\nimport shutil\nimport torch\nimport torch.distributed as dist\nimport torch.autograd as autograd\n\nfrom PIL import ImageFilter\nfrom easydict import EasyDict\nimport yaml\n# from data.dataset_3d import Dataset_3D\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n # try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.cfg_from_yaml_file","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.cfg_from_yaml_file#L32-L40","kind":"function","name":"cfg_from_yaml_file","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":32,"end_line":40,"context_start_line":12,"context_end_line":60,"code":"# from data.dataset_3d import Dataset_3D\n\ndef merge_new_config(config, new_config):\n for key, val in new_config.items():\n if not isinstance(val, dict):\n if key == '_base_':\n with open(new_config['_base_'], 'r') as f:\n try:\n val = yaml.load(f, Loader=yaml.FullLoader)\n except:\n val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n # try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.get_model","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.get_model#L42-L47","kind":"function","name":"get_model","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":42,"end_line":47,"context_start_line":22,"context_end_line":67,"code":" val = yaml.load(f)\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n else:\n config[key] = val\n continue\n if key not in config:\n config[key] = EasyDict()\n merge_new_config(config[key], val)\n return config\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n # try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.setup_for_distributed","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.setup_for_distributed#L50-L62","kind":"function","name":"setup_for_distributed","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":50,"end_line":62,"context_start_line":30,"context_end_line":82,"code":" merge_new_config(config[key], val)\n return config\ndef cfg_from_yaml_file(cfg_file):\n config = EasyDict()\n with open(cfg_file, 'r') as f:\n # try:\n new_config = yaml.load(f, Loader=yaml.FullLoader)\n # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.is_dist_avail_and_initialized","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.is_dist_avail_and_initialized#L65-L70","kind":"function","name":"is_dist_avail_and_initialized","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":65,"end_line":70,"context_start_line":45,"context_end_line":90,"code":" return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.get_world_size","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.get_world_size#L73-L76","kind":"function","name":"get_world_size","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":73,"end_line":76,"context_start_line":53,"context_end_line":96,"code":" \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.get_rank","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.get_rank#L79-L82","kind":"function","name":"get_rank","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":79,"end_line":82,"context_start_line":59,"context_end_line":102,"code":" if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n\n\ndef init_distributed_mode(args):\n if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:\n args.rank = int(os.environ[\"RANK\"])\n args.world_size = int(os.environ['WORLD_SIZE'])\n args.gpu = int(os.environ['LOCAL_RANK'])","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.is_main_process","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.is_main_process#L85-L86","kind":"function","name":"is_main_process","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":85,"end_line":86,"context_start_line":65,"context_end_line":106,"code":"def is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n\n\ndef init_distributed_mode(args):\n if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:\n args.rank = int(os.environ[\"RANK\"])\n args.world_size = int(os.environ['WORLD_SIZE'])\n args.gpu = int(os.environ['LOCAL_RANK'])\n elif 'SLURM_PROCID' in os.environ:\n args.rank = int(os.environ['SLURM_PROCID'])\n args.gpu = args.rank % torch.cuda.device_count()\n else:","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.save_on_master","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.save_on_master#L89-L95","kind":"function","name":"save_on_master","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":89,"end_line":95,"context_start_line":69,"context_end_line":115,"code":" return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n\n\ndef init_distributed_mode(args):\n if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:\n args.rank = int(os.environ[\"RANK\"])\n args.world_size = int(os.environ['WORLD_SIZE'])\n args.gpu = int(os.environ['LOCAL_RANK'])\n elif 'SLURM_PROCID' in os.environ:\n args.rank = int(os.environ['SLURM_PROCID'])\n args.gpu = args.rank % torch.cuda.device_count()\n else:\n print('Not using distributed mode')\n args.distributed = False\n return\n\n args.distributed = True\n\n torch.cuda.set_device(args.gpu)\n args.dist_backend = 'nccl'\n print('| distributed init (rank {}): {}'.format(","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.init_distributed_mode","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.init_distributed_mode#L98-L120","kind":"function","name":"init_distributed_mode","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":98,"end_line":120,"context_start_line":78,"context_end_line":140,"code":"\ndef get_rank():\n if not is_dist_avail_and_initialized():\n return 0\n return dist.get_rank()\n\n\ndef is_main_process():\n return get_rank() == 0\n\n\ndef save_on_master(state, is_best, output_dir):\n if is_main_process():\n ckpt_path = '{}/checkpoint_{}.pt'.format(output_dir, state['epoch'])\n best_path = f'{output_dir}/checkpoint_best.pt'\n torch.save(state, ckpt_path)\n if is_best:\n shutil.copyfile(ckpt_path, best_path)\n\n\ndef init_distributed_mode(args):\n if 'RANK' in os.environ and 'WORLD_SIZE' in os.environ:\n args.rank = int(os.environ[\"RANK\"])\n args.world_size = int(os.environ['WORLD_SIZE'])\n args.gpu = int(os.environ['LOCAL_RANK'])\n elif 'SLURM_PROCID' in os.environ:\n args.rank = int(os.environ['SLURM_PROCID'])\n args.gpu = args.rank % torch.cuda.device_count()\n else:\n print('Not using distributed mode')\n args.distributed = False\n return\n\n args.distributed = True\n\n torch.cuda.set_device(args.gpu)\n args.dist_backend = 'nccl'\n print('| distributed init (rank {}): {}'.format(\n args.rank, args.dist_url), flush=True)\n torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url,\n world_size=args.world_size, rank=args.rank)\n torch.distributed.barrier()\n setup_for_distributed(args.rank == 0)\n\n\ndef scaled_all_reduce(tensors, is_scale=True):\n \"\"\"Performs the scaled all_reduce operation on the provided tensors.\n The input tensors are modified in-place. Currently supports only the sum\n reduction operator. The reduced values are scaled by the inverse size of the\n world size.\n \"\"\"\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n # Queue the reductions\n reductions = []\n for tensor in tensors:\n reduction = dist.all_reduce(tensor, async_op=True)\n reductions.append(reduction)\n # Wait for reductions to finish\n for reduction in reductions:\n reduction.wait()","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.scaled_all_reduce","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.scaled_all_reduce#L123-L145","kind":"function","name":"scaled_all_reduce","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":123,"end_line":145,"context_start_line":103,"context_end_line":165,"code":" elif 'SLURM_PROCID' in os.environ:\n args.rank = int(os.environ['SLURM_PROCID'])\n args.gpu = args.rank % torch.cuda.device_count()\n else:\n print('Not using distributed mode')\n args.distributed = False\n return\n\n args.distributed = True\n\n torch.cuda.set_device(args.gpu)\n args.dist_backend = 'nccl'\n print('| distributed init (rank {}): {}'.format(\n args.rank, args.dist_url), flush=True)\n torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url,\n world_size=args.world_size, rank=args.rank)\n torch.distributed.barrier()\n setup_for_distributed(args.rank == 0)\n\n\ndef scaled_all_reduce(tensors, is_scale=True):\n \"\"\"Performs the scaled all_reduce operation on the provided tensors.\n The input tensors are modified in-place. Currently supports only the sum\n reduction operator. The reduced values are scaled by the inverse size of the\n world size.\n \"\"\"\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n # Queue the reductions\n reductions = []\n for tensor in tensors:\n reduction = dist.all_reduce(tensor, async_op=True)\n reductions.append(reduction)\n # Wait for reductions to finish\n for reduction in reductions:\n reduction.wait()\n # Scale the results\n if is_scale:\n for tensor in tensors:\n tensor.mul_(1.0 / world_size)\n return tensors\n\n\ndef all_gather_batch(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n for tensor in tensors:\n tensor_all = [torch.ones_like(tensor) for _ in range(world_size)]\n dist.all_gather(\n tensor_all,\n tensor,\n async_op=False # performance opt\n )","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.all_gather_batch","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.all_gather_batch#L148-L171","kind":"function","name":"all_gather_batch","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":148,"end_line":171,"context_start_line":128,"context_end_line":191,"code":" \"\"\"\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n # Queue the reductions\n reductions = []\n for tensor in tensors:\n reduction = dist.all_reduce(tensor, async_op=True)\n reductions.append(reduction)\n # Wait for reductions to finish\n for reduction in reductions:\n reduction.wait()\n # Scale the results\n if is_scale:\n for tensor in tensors:\n tensor.mul_(1.0 / world_size)\n return tensors\n\n\ndef all_gather_batch(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n for tensor in tensors:\n tensor_all = [torch.ones_like(tensor) for _ in range(world_size)]\n dist.all_gather(\n tensor_all,\n tensor,\n async_op=False # performance opt\n )\n\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.GatherLayer","uri":"program://CREMA/class/lavis.models.ulip_models.utils.utils.GatherLayer#L174-L190","kind":"class","name":"GatherLayer","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":174,"end_line":190,"context_start_line":154,"context_end_line":210,"code":" # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n for tensor in tensors:\n tensor_all = [torch.ones_like(tensor) for _ in range(world_size)]\n dist.all_gather(\n tensor_all,\n tensor,\n async_op=False # performance opt\n )\n\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n\n\ndef all_gather_batch_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n\n for tensor in tensors:\n tensor_all = GatherLayer.apply(tensor)\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.all_gather_batch_with_grad","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.all_gather_batch_with_grad#L193-L212","kind":"function","name":"all_gather_batch_with_grad","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":193,"end_line":212,"context_start_line":173,"context_end_line":232,"code":"\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n\n\ndef all_gather_batch_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n\n for tensor in tensors:\n tensor_all = GatherLayer.apply(tensor)\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\ndef cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.cosine_scheduler","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.cosine_scheduler#L215-L226","kind":"function","name":"cosine_scheduler","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":215,"end_line":226,"context_start_line":195,"context_end_line":242,"code":" Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n\n for tensor in tensors:\n tensor_all = GatherLayer.apply(tensor)\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\ndef cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):\n self.sigma = sigma\n\n def __call__(self, x):\n sigma = random.uniform(self.sigma[0], self.sigma[1])\n x = x.filter(ImageFilter.GaussianBlur(radius=sigma))\n return x\n\n# def get_dataset(train_transform, tokenizer, args, dataset_name=None, files_list=None):\n# dataset_3d = Dataset_3D(args, tokenizer, dataset_name, train_transform, files_list=files_list)\n# return dataset_3d.dataset","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.GaussianBlur","uri":"program://CREMA/class/lavis.models.ulip_models.utils.utils.GaussianBlur#L229-L238","kind":"class","name":"GaussianBlur","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":229,"end_line":238,"context_start_line":209,"context_end_line":242,"code":"\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\ndef cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):\n self.sigma = sigma\n\n def __call__(self, x):\n sigma = random.uniform(self.sigma[0], self.sigma[1])\n x = x.filter(ImageFilter.GaussianBlur(radius=sigma))\n return x\n\n# def get_dataset(train_transform, tokenizer, args, dataset_name=None, files_list=None):\n# dataset_3d = Dataset_3D(args, tokenizer, dataset_name, train_transform, files_list=files_list)\n# return dataset_3d.dataset","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.print","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.print#L57-L60","kind":"function","name":"print","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":57,"end_line":60,"context_start_line":37,"context_end_line":80,"code":" # except:\n # new_config = yaml.load(f)\n merge_new_config(config=config, new_config=new_config)\n return config\n\ndef get_model(model):\n if isinstance(model, torch.nn.DataParallel) \\\n or isinstance(model, torch.nn.parallel.DistributedDataParallel):\n return model.module\n else:\n return model\n\n\ndef setup_for_distributed(is_master):\n \"\"\"\n This function disables printing when not in master process\n \"\"\"\n import builtins as __builtin__\n builtin_print = __builtin__.print\n\n def print(*args, **kwargs):\n force = kwargs.pop('force', False)\n if is_master or force:\n builtin_print(*args, **kwargs)\n\n __builtin__.print = print\n\n\ndef is_dist_avail_and_initialized():\n if not dist.is_available():\n return False\n if not dist.is_initialized():\n return False\n return True\n\n\ndef get_world_size():\n if not is_dist_avail_and_initialized():\n return 1\n return dist.get_world_size()\n\n\ndef get_rank():\n if not is_dist_avail_and_initialized():","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.forward","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.forward#L181-L184","kind":"function","name":"forward","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":181,"end_line":184,"context_start_line":161,"context_end_line":204,"code":" dist.all_gather(\n tensor_all,\n tensor,\n async_op=False # performance opt\n )\n\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n\n\ndef all_gather_batch_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.backward","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.backward#L187-L190","kind":"function","name":"backward","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":187,"end_line":190,"context_start_line":167,"context_end_line":210,"code":" tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:\n output_tensor.append(torch.cat(tensor_all, dim=0))\n return output_tensor\n\n\nclass GatherLayer(autograd.Function):\n \"\"\"\n Gather tensors from all workers with support for backward propagation:\n This implementation does not cut the gradients as torch.distributed.all_gather does.\n \"\"\"\n\n @staticmethod\n def forward(ctx, x):\n output = [torch.zeros_like(x) for _ in range(dist.get_world_size())]\n dist.all_gather(output, x)\n return tuple(output)\n\n @staticmethod\n def backward(ctx, *grads):\n all_gradients = torch.stack(grads)\n dist.all_reduce(all_gradients)\n return all_gradients[dist.get_rank()]\n\n\ndef all_gather_batch_with_grad(tensors):\n \"\"\"\n Performs all_gather operation on the provided tensors.\n Graph remains connected for backward grad computation.\n \"\"\"\n # Queue the gathered tensors\n world_size = get_world_size()\n # There is no need for reduction in the single-proc case\n if world_size == 1:\n return tensors\n tensor_list = []\n output_tensor = []\n\n for tensor in tensors:\n tensor_all = GatherLayer.apply(tensor)\n tensor_list.append(tensor_all)\n\n for tensor_all in tensor_list:","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.__init__#L232-L233","kind":"function","name":"__init__","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":232,"end_line":233,"context_start_line":212,"context_end_line":242,"code":" return output_tensor\n\n\ndef cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):\n self.sigma = sigma\n\n def __call__(self, x):\n sigma = random.uniform(self.sigma[0], self.sigma[1])\n x = x.filter(ImageFilter.GaussianBlur(radius=sigma))\n return x\n\n# def get_dataset(train_transform, tokenizer, args, dataset_name=None, files_list=None):\n# dataset_3d = Dataset_3D(args, tokenizer, dataset_name, train_transform, files_list=files_list)\n# return dataset_3d.dataset","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.utils.__call__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.utils.__call__#L235-L238","kind":"function","name":"__call__","path":"lavis/models/ulip_models/utils/utils.py","language":"python","start_line":235,"end_line":238,"context_start_line":215,"context_end_line":242,"code":"def cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, start_warmup_value=0):\n warmup_schedule = np.array([])\n warmup_iters = warmup_epochs * niter_per_ep\n if warmup_epochs > 0:\n warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters)\n\n iters = np.arange(epochs * niter_per_ep - warmup_iters)\n schedule = final_value + 0.5 * (base_value - final_value) * (1 + np.cos(np.pi * iters / len(iters)))\n\n schedule = np.concatenate((warmup_schedule, schedule))\n assert len(schedule) == epochs * niter_per_ep\n return schedule\n\n\nclass GaussianBlur(object):\n \"\"\"Gaussian blur augmentation in SimCLR https://arxiv.org/abs/2002.05709\"\"\"\n\n def __init__(self, sigma=[.1, 2.]):\n self.sigma = sigma\n\n def __call__(self, x):\n sigma = random.uniform(self.sigma[0], self.sigma[1])\n x = x.filter(ImageFilter.GaussianBlur(radius=sigma))\n return x\n\n# def get_dataset(train_transform, tokenizer, args, dataset_name=None, files_list=None):\n# dataset_3d = Dataset_3D(args, tokenizer, dataset_name, train_transform, files_list=files_list)\n# return dataset_3d.dataset","source_hash":"cff06979f73a295b19d5f0702a0cdfa6ff2d65816f6e4773c82940fa4f2691e3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.logger","uri":"program://CREMA/module/lavis.models.ulip_models.utils.logger#L1-L127","kind":"module","name":"lavis.models.ulip_models.utils.logger","path":"lavis/models/ulip_models/utils/logger.py","language":"python","start_line":1,"end_line":127,"context_start_line":1,"context_end_line":127,"code":"import logging\nimport torch.distributed as dist\n\nlogger_initialized = {}\n\ndef get_root_logger(log_file=None, log_level=logging.INFO, name='main'):\n \"\"\"Get root logger and add a keyword filter to it.\n The logger will be initialized if it has not been initialized. By default a\n StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.\n \"\"\"\n logger = logging.getLogger(name)\n if name in logger_initialized:\n return logger\n # handle hierarchical names\n # e.g., logger \"a\" is initialized, then logger \"a.b\" will skip the\n # initialization since it is a child of \"a\".\n for logger_name in logger_initialized:\n if name.startswith(logger_name):\n return logger\n\n # handle duplicate logs to the console\n # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler (NOTSET)\n # to the root logger. As logger.propagate is True by default, this root\n # level handler causes logging messages from rank>0 processes to\n # unexpectedly show up on the console, creating much unwanted clutter.\n # To fix this issue, we set the root logger's StreamHandler, if any, to log\n # at the ERROR level.\n for handler in logger.root.handlers:\n if type(handler) is logging.StreamHandler:\n handler.setLevel(logging.ERROR)\n\n stream_handler = logging.StreamHandler()\n handlers = [stream_handler]\n\n if dist.is_available() and dist.is_initialized():\n rank = dist.get_rank()\n else:\n rank = 0\n\n # only rank 0 will add a FileHandler\n if rank == 0 and log_file is not None:\n # Here, the default behaviour of the official logger is 'a'. Thus, we\n # provide an interface to change the file mode to the default\n # behaviour.\n file_handler = logging.FileHandler(log_file, file_mode)\n handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass\n elif isinstance(logger, str):\n _logger = get_logger(logger)\n _logger.log(level, msg)\n else:\n raise TypeError(\n 'logger should be either a logging.Logger object, str, '\n f'\"silent\" or None, but got {type(logger)}')","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.logger.get_root_logger","uri":"program://CREMA/function/lavis.models.ulip_models.utils.logger.get_root_logger#L6-L26","kind":"function","name":"get_root_logger","path":"lavis/models/ulip_models/utils/logger.py","language":"python","start_line":6,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"import logging\nimport torch.distributed as dist\n\nlogger_initialized = {}\n\ndef get_root_logger(log_file=None, log_level=logging.INFO, name='main'):\n \"\"\"Get root logger and add a keyword filter to it.\n The logger will be initialized if it has not been initialized. By default a\n StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.logger.get_logger","uri":"program://CREMA/function/lavis.models.ulip_models.utils.logger.get_logger#L29-L100","kind":"function","name":"get_logger","path":"lavis/models/ulip_models/utils/logger.py","language":"python","start_line":29,"end_line":100,"context_start_line":9,"context_end_line":120,"code":" StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.\n \"\"\"\n logger = logging.getLogger(name)\n if name in logger_initialized:\n return logger\n # handle hierarchical names\n # e.g., logger \"a\" is initialized, then logger \"a.b\" will skip the\n # initialization since it is a child of \"a\".\n for logger_name in logger_initialized:\n if name.startswith(logger_name):\n return logger\n\n # handle duplicate logs to the console\n # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler (NOTSET)\n # to the root logger. As logger.propagate is True by default, this root\n # level handler causes logging messages from rank>0 processes to\n # unexpectedly show up on the console, creating much unwanted clutter.\n # To fix this issue, we set the root logger's StreamHandler, if any, to log\n # at the ERROR level.\n for handler in logger.root.handlers:\n if type(handler) is logging.StreamHandler:\n handler.setLevel(logging.ERROR)\n\n stream_handler = logging.StreamHandler()\n handlers = [stream_handler]\n\n if dist.is_available() and dist.is_initialized():\n rank = dist.get_rank()\n else:\n rank = 0\n\n # only rank 0 will add a FileHandler\n if rank == 0 and log_file is not None:\n # Here, the default behaviour of the official logger is 'a'. Thus, we\n # provide an interface to change the file mode to the default\n # behaviour.\n file_handler = logging.FileHandler(log_file, file_mode)\n handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.logger.print_log","uri":"program://CREMA/function/lavis.models.ulip_models.utils.logger.print_log#L103-L127","kind":"function","name":"print_log","path":"lavis/models/ulip_models/utils/logger.py","language":"python","start_line":103,"end_line":127,"context_start_line":83,"context_end_line":127,"code":" handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass\n elif isinstance(logger, str):\n _logger = get_logger(logger)\n _logger.log(level, msg)\n else:\n raise TypeError(\n 'logger should be either a logging.Logger object, str, '\n f'\"silent\" or None, but got {type(logger)}')","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer","uri":"program://CREMA/module/lavis.models.ulip_models.utils.tokenizer#L1-L151","kind":"module","name":"lavis.models.ulip_models.utils.tokenizer","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":1,"end_line":151,"context_start_line":1,"context_end_line":151,"code":"# Modified from github.com/openai/CLIP\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\")\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = list(range(ord(\"!\"), ord(\"~\")+1))+list(range(ord(\"¡\"), ord(\"¬\")+1))+list(range(ord(\"®\"), ord(\"ÿ\")+1))\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8+n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}\n self.pat = re.compile(r\"\"\"<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\", re.IGNORECASE)\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + ( token[-1] + '',)\n pairs = get_pairs(word)\n\n if not pairs:\n return token+''\n\n while True:\n bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word)-1 and word[i+1] == second:\n new_word.append(first+second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]\n all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n tokens = tokens[:context_length]\n result[i, :len(tokens)] = torch.tensor(tokens)\n\n if len(result) == 1:\n return result[0]\n return result","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.default_bpe","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.default_bpe#L13-L14","kind":"function","name":"default_bpe","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":13,"end_line":14,"context_start_line":1,"context_end_line":34,"code":"# Modified from github.com/openai/CLIP\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\")\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = list(range(ord(\"!\"), ord(\"~\")+1))+list(range(ord(\"¡\"), ord(\"¬\")+1))+list(range(ord(\"®\"), ord(\"ÿ\")+1))\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8+n)","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.bytes_to_unicode","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.bytes_to_unicode#L18-L37","kind":"function","name":"bytes_to_unicode","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":18,"end_line":37,"context_start_line":1,"context_end_line":57,"code":"# Modified from github.com/openai/CLIP\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\")\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = list(range(ord(\"!\"), ord(\"~\")+1))+list(range(ord(\"¡\"), ord(\"¬\")+1))+list(range(ord(\"®\"), ord(\"ÿ\")+1))\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8+n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.get_pairs","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.get_pairs#L40-L49","kind":"function","name":"get_pairs","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":40,"end_line":49,"context_start_line":20,"context_end_line":69,"code":" Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = list(range(ord(\"!\"), ord(\"~\")+1))+list(range(ord(\"¡\"), ord(\"¬\")+1))+list(range(ord(\"®\"), ord(\"ÿ\")+1))\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8+n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.basic_clean","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.basic_clean#L52-L55","kind":"function","name":"basic_clean","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":52,"end_line":55,"context_start_line":32,"context_end_line":75,"code":" if b not in bs:\n bs.append(b)\n cs.append(2**8+n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.whitespace_clean","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.whitespace_clean#L58-L61","kind":"function","name":"whitespace_clean","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":58,"end_line":61,"context_start_line":38,"context_end_line":81,"code":"\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}\n self.pat = re.compile(r\"\"\"<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\", re.IGNORECASE)\n","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.SimpleTokenizer","uri":"program://CREMA/class/lavis.models.ulip_models.utils.tokenizer.SimpleTokenizer#L64-L151","kind":"class","name":"SimpleTokenizer","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":64,"end_line":151,"context_start_line":44,"context_end_line":151,"code":" pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}\n self.pat = re.compile(r\"\"\"<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\", re.IGNORECASE)\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + ( token[-1] + '',)\n pairs = get_pairs(word)\n\n if not pairs:\n return token+''\n\n while True:\n bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word)-1 and word[i+1] == second:\n new_word.append(first+second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]\n all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n tokens = tokens[:context_length]\n result[i, :len(tokens)] = torch.tensor(tokens)\n\n if len(result) == 1:\n return result[0]\n return result","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.__init__#L65-L80","kind":"function","name":"__init__","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":65,"end_line":80,"context_start_line":45,"context_end_line":100,"code":" prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r'\\s+', ' ', text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}\n self.pat = re.compile(r\"\"\"<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\", re.IGNORECASE)\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + ( token[-1] + '',)\n pairs = get_pairs(word)\n\n if not pairs:\n return token+''\n\n while True:\n bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.bpe","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.bpe#L82-L121","kind":"function","name":"bpe","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":82,"end_line":121,"context_start_line":62,"context_end_line":141,"code":"\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe()):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split('\\n')\n merges = merges[1:49152-256-2+1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v+'' for v in vocab]\n for merge in merges:\n vocab.append(''.join(merge))\n vocab.extend(['<|startoftext|>', '<|endoftext|>'])\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}\n self.pat = re.compile(r\"\"\"<\\|startoftext\\|>|<\\|endoftext\\|>|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\", re.IGNORECASE)\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + ( token[-1] + '',)\n pairs = get_pairs(word)\n\n if not pairs:\n return token+''\n\n while True:\n bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word)-1 and word[i+1] == second:\n new_word.append(first+second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.encode","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.encode#L123-L129","kind":"function","name":"encode","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":123,"end_line":129,"context_start_line":103,"context_end_line":149,"code":" except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word)-1 and word[i+1] == second:\n new_word.append(first+second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]\n all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n tokens = tokens[:context_length]\n result[i, :len(tokens)] = torch.tensor(tokens)\n\n if len(result) == 1:","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.decode","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.decode#L131-L134","kind":"function","name":"decode","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":131,"end_line":134,"context_start_line":111,"context_end_line":151,"code":" new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]\n all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n tokens = tokens[:context_length]\n result[i, :len(tokens)] = torch.tensor(tokens)\n\n if len(result) == 1:\n return result[0]\n return result","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.tokenizer.__call__","uri":"program://CREMA/function/lavis.models.ulip_models.utils.tokenizer.__call__#L136-L151","kind":"function","name":"__call__","path":"lavis/models/ulip_models/utils/tokenizer.py","language":"python","start_line":136,"end_line":151,"context_start_line":116,"context_end_line":151,"code":" break\n else:\n pairs = get_pairs(word)\n word = ' '.join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))\n bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))\n return bpe_tokens\n\n def decode(self, tokens):\n text = ''.join([self.decoder[token] for token in tokens])\n text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors=\"replace\").replace('', ' ')\n return text\n\n def __call__(self, texts, context_length=77):\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = self.encoder[\"<|startoftext|>\"]\n eot_token = self.encoder[\"<|endoftext|>\"]\n all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n tokens = tokens[:context_length]\n result[i, :len(tokens)] = torch.tensor(tokens)\n\n if len(result) == 1:\n return result[0]\n return result","source_hash":"db10f2a1f8e6f5f639dfb3b4d090e13507bec3d3cb695e25fdbe044b7cbaf230","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.build","uri":"program://CREMA/module/lavis.models.ulip_models.utils.build#L1-L17","kind":"module","name":"lavis.models.ulip_models.utils.build","path":"lavis/models/ulip_models/utils/build.py","language":"python","start_line":1,"end_line":17,"context_start_line":1,"context_end_line":17,"code":"from utils import registry\n\n\nDATASETS = registry.Registry('dataset')\n\n\ndef build_dataset_from_cfg(cfg, default_args = None):\n \"\"\"\n Build a dataset, defined by `dataset_name`.\n Args:\n cfg (eDICT): \n Returns:\n Dataset: a constructed dataset specified by dataset_name.\n \"\"\"\n return DATASETS.build(cfg, default_args = default_args)\n\n","source_hash":"8d1705e603f752f90013a8b5244f7a705c75cb1926040168cbf7f78f053aaf9b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.utils.build.build_dataset_from_cfg","uri":"program://CREMA/function/lavis.models.ulip_models.utils.build.build_dataset_from_cfg#L7-L15","kind":"function","name":"build_dataset_from_cfg","path":"lavis/models/ulip_models/utils/build.py","language":"python","start_line":7,"end_line":15,"context_start_line":1,"context_end_line":17,"code":"from utils import registry\n\n\nDATASETS = registry.Registry('dataset')\n\n\ndef build_dataset_from_cfg(cfg, default_args = None):\n \"\"\"\n Build a dataset, defined by `dataset_name`.\n Args:\n cfg (eDICT): \n Returns:\n Dataset: a constructed dataset specified by dataset_name.\n \"\"\"\n return DATASETS.build(cfg, default_args = default_args)\n\n","source_hash":"8d1705e603f752f90013a8b5244f7a705c75cb1926040168cbf7f78f053aaf9b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc","uri":"program://CREMA/module/lavis.models.ulip_models.pointbert.misc#L1-L287","kind":"module","name":"lavis.models.ulip_models.pointbert.misc","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":1,"end_line":287,"context_start_line":1,"context_end_line":287,"code":"import numpy as np\nimport matplotlib.pyplot as plt\nfrom mpl_toolkits.mplot3d import Axes3D\nimport random\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport os\nfrom collections import abc\n# from pointnet2_ops import pointnet2_utils\n\n\n# def fps(data, number):\n# '''\n# data B N 3\n# number int\n# '''\n# fps_idx = pointnet2_utils.furthest_point_sample(data, number)\n# fps_data = pointnet2_utils.gather_operation(data.transpose(1, 2).contiguous(), fps_idx).transpose(1,2).contiguous()\n# return fps_data\n\ndef index_points(points, idx):\n \"\"\"\n Input:\n points: input points data, [B, N, C]\n idx: sample index data, [B, S]\n Return:\n new_points:, indexed points data, [B, S, C]\n \"\"\"\n device = points.device\n B = points.shape[0]\n view_shape = list(idx.shape)\n view_shape[1:] = [1] * (len(view_shape) - 1)\n repeat_shape = list(idx.shape)\n repeat_shape[0] = 1\n batch_indices = torch.arange(B, dtype=torch.long).to(device).view(view_shape).repeat(repeat_shape)\n new_points = points[batch_indices, idx, :]\n return new_points\n\ndef fps(xyz, npoint):\n \"\"\"\n Input:\n xyz: pointcloud data, [B, N, 3]\n npoint: number of samples\n Return:\n centroids: sampled pointcloud index, [B, npoint]\n \"\"\"\n device = xyz.device\n B, N, C = xyz.shape\n centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)\n distance = torch.ones(B, N).to(device) * 1e10\n farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)\n batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)\n farthest = torch.max(distance, -1)[1]\n return index_points(xyz, centroids)\n\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n \ndef set_random_seed(seed, deterministic=False):\n \"\"\"Set random seed.\n Args:\n seed (int): Seed to be used.\n deterministic (bool): Whether to set the deterministic option for\n CUDNN backend, i.e., set `torch.backends.cudnn.deterministic`\n to True and `torch.backends.cudnn.benchmark` to False.\n Default: False.\n\n # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html\n if cuda_deterministic: # slower, more reproducible\n cudnn.deterministic = True\n cudnn.benchmark = False\n else: # faster, less reproducible\n cudnn.deterministic = False\n cudnn.benchmark = True\n\n \"\"\"\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n if deterministic:\n torch.backends.cudnn.deterministic = True\n torch.backends.cudnn.benchmark = False\n\n\ndef is_seq_of(seq, expected_type, seq_type=None):\n \"\"\"Check whether it is a sequence of some type.\n Args:\n seq (Sequence): The sequence to be checked.\n expected_type (type): Expected type of sequence items.\n seq_type (type, optional): Expected sequence type.\n Returns:\n bool: Whether the sequence is valid.\n \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape\n\n assert n == num_points\n assert c == 3\n if crop == num_points:\n return xyz, None\n \n INPUT = []\n CROP = []\n for points in xyz:\n if isinstance(crop,list):\n num_crop = random.randint(crop[0],crop[1])\n else:\n num_crop = crop\n\n points = points.unsqueeze(0)\n\n if fixed_points is None: \n center = F.normalize(torch.randn(1,1,3),p=2,dim=-1).cuda()\n else:\n if isinstance(fixed_points,list):\n fixed_point = random.sample(fixed_points,1)[0]\n else:\n fixed_point = fixed_points\n center = fixed_point.reshape(1,1,3).cuda()\n\n distance_matrix = torch.norm(center.unsqueeze(2) - points.unsqueeze(1), p =2 ,dim = -1) # 1 1 2048\n\n idx = torch.argsort(distance_matrix,dim=-1, descending=False)[0,0] # 2048\n\n if padding_zeros:\n input_data = points.clone()\n input_data[0, idx[:num_crop]] = input_data[0,idx[:num_crop]] * 0\n\n else:\n input_data = points.clone()[0, idx[num_crop:]].unsqueeze(0) # 1 N 3\n\n crop_data = points.clone()[0, idx[:num_crop]].unsqueeze(0)\n\n if isinstance(crop,list):\n INPUT.append(fps(input_data,2048))\n CROP.append(fps(crop_data,2048))\n else:\n INPUT.append(input_data)\n CROP.append(crop_data)\n\n input_data = torch.cat(INPUT,dim=0)# B N 3\n crop_data = torch.cat(CROP,dim=0)# B M 3\n\n return input_data.contiguous(), crop_data.contiguous()\n\ndef get_ptcloud_img(ptcloud):\n fig = plt.figure(figsize=(8, 8))\n\n x, z, y = ptcloud.transpose(1, 0)\n ax = fig.gca(projection=Axes3D.name, adjustable='box')\n ax.axis('off')\n # ax.axis('scaled')\n ax.view_init(30, 45)\n max, min = np.max(ptcloud), np.min(ptcloud)\n ax.set_xbound(min, max)\n ax.set_ybound(min, max)\n ax.set_zbound(min, max)\n ax.scatter(x, y, z, zdir='z', c=x, cmap='jet')\n\n fig.canvas.draw()\n img = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\n img = img.reshape(fig.canvas.get_width_height()[::-1] + (3, ))\n return img\n\n\n\ndef visualize_KITTI(path, data_list, titles = ['input','pred'], cmap=['bwr','autumn'], zdir='y', \n xlim=(-1, 1), ylim=(-1, 1), zlim=(-1, 1) ):\n fig = plt.figure(figsize=(6*len(data_list),6))\n cmax = data_list[-1][:,0].max()\n\n for i in range(len(data_list)):\n data = data_list[i][:-2048] if i == 1 else data_list[i]\n color = data[:,0] /cmax\n ax = fig.add_subplot(1, len(data_list) , i + 1, projection='3d')\n ax.view_init(30, -120)\n b = ax.scatter(data[:, 0], data[:, 1], data[:, 2], zdir=zdir, c=color,vmin=-1,vmax=1 ,cmap = cmap[0],s=4,linewidth=0.05, edgecolors = 'black')\n ax.set_title(titles[i])\n\n ax.set_axis_off()\n ax.set_xlim(xlim)\n ax.set_ylim(ylim)\n ax.set_zlim(zlim)\n plt.subplots_adjust(left=0, right=1, bottom=0, top=1, wspace=0.2, hspace=0)\n if not os.path.exists(path):\n os.makedirs(path)\n\n pic_path = path + '.png'\n fig.savefig(pic_path)\n\n np.save(os.path.join(path, 'input.npy'), data_list[0].numpy())\n np.save(os.path.join(path, 'pred.npy'), data_list[1].numpy())\n plt.close(fig)\n\n\ndef random_dropping(pc, e):\n up_num = max(64, 768 // (e//50 + 1))\n pc = pc\n random_num = torch.randint(1, up_num, (1,1))[0,0]\n pc = fps(pc, random_num)\n padding = torch.zeros(pc.size(0), 2048 - pc.size(1), 3).to(pc.device)\n pc = torch.cat([pc, padding], dim = 1)\n return pc\n \n\ndef random_scale(partial, scale_range=[0.8, 1.2]):\n scale = torch.rand(1).cuda() * (scale_range[1] - scale_range[0]) + scale_range[0]\n return partial * scale","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.index_points","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.index_points#L22-L38","kind":"function","name":"index_points","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":22,"end_line":38,"context_start_line":2,"context_end_line":58,"code":"import matplotlib.pyplot as plt\nfrom mpl_toolkits.mplot3d import Axes3D\nimport random\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport os\nfrom collections import abc\n# from pointnet2_ops import pointnet2_utils\n\n\n# def fps(data, number):\n# '''\n# data B N 3\n# number int\n# '''\n# fps_idx = pointnet2_utils.furthest_point_sample(data, number)\n# fps_data = pointnet2_utils.gather_operation(data.transpose(1, 2).contiguous(), fps_idx).transpose(1,2).contiguous()\n# return fps_data\n\ndef index_points(points, idx):\n \"\"\"\n Input:\n points: input points data, [B, N, C]\n idx: sample index data, [B, S]\n Return:\n new_points:, indexed points data, [B, S, C]\n \"\"\"\n device = points.device\n B = points.shape[0]\n view_shape = list(idx.shape)\n view_shape[1:] = [1] * (len(view_shape) - 1)\n repeat_shape = list(idx.shape)\n repeat_shape[0] = 1\n batch_indices = torch.arange(B, dtype=torch.long).to(device).view(view_shape).repeat(repeat_shape)\n new_points = points[batch_indices, idx, :]\n return new_points\n\ndef fps(xyz, npoint):\n \"\"\"\n Input:\n xyz: pointcloud data, [B, N, 3]\n npoint: number of samples\n Return:\n centroids: sampled pointcloud index, [B, npoint]\n \"\"\"\n device = xyz.device\n B, N, C = xyz.shape\n centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)\n distance = torch.ones(B, N).to(device) * 1e10\n farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)\n batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.fps","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.fps#L40-L60","kind":"function","name":"fps","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":40,"end_line":60,"context_start_line":20,"context_end_line":80,"code":"# return fps_data\n\ndef index_points(points, idx):\n \"\"\"\n Input:\n points: input points data, [B, N, C]\n idx: sample index data, [B, S]\n Return:\n new_points:, indexed points data, [B, S, C]\n \"\"\"\n device = points.device\n B = points.shape[0]\n view_shape = list(idx.shape)\n view_shape[1:] = [1] * (len(view_shape) - 1)\n repeat_shape = list(idx.shape)\n repeat_shape[0] = 1\n batch_indices = torch.arange(B, dtype=torch.long).to(device).view(view_shape).repeat(repeat_shape)\n new_points = points[batch_indices, idx, :]\n return new_points\n\ndef fps(xyz, npoint):\n \"\"\"\n Input:\n xyz: pointcloud data, [B, N, 3]\n npoint: number of samples\n Return:\n centroids: sampled pointcloud index, [B, npoint]\n \"\"\"\n device = xyz.device\n B, N, C = xyz.shape\n centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)\n distance = torch.ones(B, N).to(device) * 1e10\n farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)\n batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)\n farthest = torch.max(distance, -1)[1]\n return index_points(xyz, centroids)\n\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n ","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.worker_init_fn","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.worker_init_fn#L62-L63","kind":"function","name":"worker_init_fn","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":62,"end_line":63,"context_start_line":42,"context_end_line":83,"code":" Input:\n xyz: pointcloud data, [B, N, 3]\n npoint: number of samples\n Return:\n centroids: sampled pointcloud index, [B, npoint]\n \"\"\"\n device = xyz.device\n B, N, C = xyz.shape\n centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)\n distance = torch.ones(B, N).to(device) * 1e10\n farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)\n batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)\n farthest = torch.max(distance, -1)[1]\n return index_points(xyz, centroids)\n\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n \ndef set_random_seed(seed, deterministic=False):\n \"\"\"Set random seed.\n Args:","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.build_lambda_sche","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.build_lambda_sche#L65-L71","kind":"function","name":"build_lambda_sche","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":65,"end_line":71,"context_start_line":45,"context_end_line":91,"code":" Return:\n centroids: sampled pointcloud index, [B, npoint]\n \"\"\"\n device = xyz.device\n B, N, C = xyz.shape\n centroids = torch.zeros(B, npoint, dtype=torch.long).to(device)\n distance = torch.ones(B, N).to(device) * 1e10\n farthest = torch.randint(0, N, (B,), dtype=torch.long).to(device)\n batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)\n farthest = torch.max(distance, -1)[1]\n return index_points(xyz, centroids)\n\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n \ndef set_random_seed(seed, deterministic=False):\n \"\"\"Set random seed.\n Args:\n seed (int): Seed to be used.\n deterministic (bool): Whether to set the deterministic option for\n CUDNN backend, i.e., set `torch.backends.cudnn.deterministic`\n to True and `torch.backends.cudnn.benchmark` to False.\n Default: False.\n\n # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html\n if cuda_deterministic: # slower, more reproducible","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.build_lambda_bnsche","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.build_lambda_bnsche#L73-L79","kind":"function","name":"build_lambda_bnsche","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":73,"end_line":79,"context_start_line":53,"context_end_line":99,"code":" batch_indices = torch.arange(B, dtype=torch.long).to(device)\n for i in range(npoint):\n centroids[:, i] = farthest\n centroid = xyz[batch_indices, farthest, :].view(B, 1, 3)\n dist = torch.sum((xyz - centroid) ** 2, -1)\n distance = torch.min(distance, dist)\n farthest = torch.max(distance, -1)[1]\n return index_points(xyz, centroids)\n\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n \ndef set_random_seed(seed, deterministic=False):\n \"\"\"Set random seed.\n Args:\n seed (int): Seed to be used.\n deterministic (bool): Whether to set the deterministic option for\n CUDNN backend, i.e., set `torch.backends.cudnn.deterministic`\n to True and `torch.backends.cudnn.benchmark` to False.\n Default: False.\n\n # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html\n if cuda_deterministic: # slower, more reproducible\n cudnn.deterministic = True\n cudnn.benchmark = False\n else: # faster, less reproducible\n cudnn.deterministic = False\n cudnn.benchmark = True\n\n \"\"\"\n random.seed(seed)","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.set_random_seed","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.set_random_seed#L81-L105","kind":"function","name":"set_random_seed","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":81,"end_line":105,"context_start_line":61,"context_end_line":125,"code":"\ndef worker_init_fn(worker_id):\n np.random.seed(np.random.get_state()[1][0] + worker_id)\n\ndef build_lambda_sche(opti, config):\n if config.get('decay_step') is not None:\n lr_lbmd = lambda e: max(config.lr_decay ** (e / config.decay_step), config.lowest_decay)\n scheduler = torch.optim.lr_scheduler.LambdaLR(opti, lr_lbmd)\n else:\n raise NotImplementedError()\n return scheduler\n\ndef build_lambda_bnsche(model, config):\n if config.get('decay_step') is not None:\n bnm_lmbd = lambda e: max(config.bn_momentum * config.bn_decay ** (e / config.decay_step), config.lowest_decay)\n bnm_scheduler = BNMomentumScheduler(model, bnm_lmbd)\n else:\n raise NotImplementedError()\n return bnm_scheduler\n \ndef set_random_seed(seed, deterministic=False):\n \"\"\"Set random seed.\n Args:\n seed (int): Seed to be used.\n deterministic (bool): Whether to set the deterministic option for\n CUDNN backend, i.e., set `torch.backends.cudnn.deterministic`\n to True and `torch.backends.cudnn.benchmark` to False.\n Default: False.\n\n # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html\n if cuda_deterministic: # slower, more reproducible\n cudnn.deterministic = True\n cudnn.benchmark = False\n else: # faster, less reproducible\n cudnn.deterministic = False\n cudnn.benchmark = True\n\n \"\"\"\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n if deterministic:\n torch.backends.cudnn.deterministic = True\n torch.backends.cudnn.benchmark = False\n\n\ndef is_seq_of(seq, expected_type, seq_type=None):\n \"\"\"Check whether it is a sequence of some type.\n Args:\n seq (Sequence): The sequence to be checked.\n expected_type (type): Expected type of sequence items.\n seq_type (type, optional): Expected sequence type.\n Returns:\n bool: Whether the sequence is valid.\n \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.is_seq_of","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.is_seq_of#L108-L127","kind":"function","name":"is_seq_of","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":108,"end_line":127,"context_start_line":88,"context_end_line":147,"code":" Default: False.\n\n # Speed-reproducibility tradeoff https://pytorch.org/docs/stable/notes/randomness.html\n if cuda_deterministic: # slower, more reproducible\n cudnn.deterministic = True\n cudnn.benchmark = False\n else: # faster, less reproducible\n cudnn.deterministic = False\n cudnn.benchmark = True\n\n \"\"\"\n random.seed(seed)\n np.random.seed(seed)\n torch.manual_seed(seed)\n torch.cuda.manual_seed_all(seed)\n if deterministic:\n torch.backends.cudnn.deterministic = True\n torch.backends.cudnn.benchmark = False\n\n\ndef is_seq_of(seq, expected_type, seq_type=None):\n \"\"\"Check whether it is a sequence of some type.\n Args:\n seq (Sequence): The sequence to be checked.\n expected_type (type): Expected type of sequence items.\n seq_type (type, optional): Expected sequence type.\n Returns:\n bool: Whether the sequence is valid.\n \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.set_bn_momentum_default","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.set_bn_momentum_default#L130-L134","kind":"function","name":"set_bn_momentum_default","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":130,"end_line":134,"context_start_line":110,"context_end_line":154,"code":" Args:\n seq (Sequence): The sequence to be checked.\n expected_type (type): Expected type of sequence items.\n seq_type (type, optional): Expected sequence type.\n Returns:\n bool: Whether the sequence is valid.\n \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.BNMomentumScheduler","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.misc.BNMomentumScheduler#L136-L166","kind":"class","name":"BNMomentumScheduler","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":136,"end_line":166,"context_start_line":116,"context_end_line":186,"code":" \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape\n\n assert n == num_points\n assert c == 3\n if crop == num_points:\n return xyz, None\n \n INPUT = []\n CROP = []\n for points in xyz:\n if isinstance(crop,list):\n num_crop = random.randint(crop[0],crop[1])\n else:","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.seprate_point_cloud","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.seprate_point_cloud#L170-L223","kind":"function","name":"seprate_point_cloud","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":170,"end_line":223,"context_start_line":150,"context_end_line":243,"code":" self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape\n\n assert n == num_points\n assert c == 3\n if crop == num_points:\n return xyz, None\n \n INPUT = []\n CROP = []\n for points in xyz:\n if isinstance(crop,list):\n num_crop = random.randint(crop[0],crop[1])\n else:\n num_crop = crop\n\n points = points.unsqueeze(0)\n\n if fixed_points is None: \n center = F.normalize(torch.randn(1,1,3),p=2,dim=-1).cuda()\n else:\n if isinstance(fixed_points,list):\n fixed_point = random.sample(fixed_points,1)[0]\n else:\n fixed_point = fixed_points\n center = fixed_point.reshape(1,1,3).cuda()\n\n distance_matrix = torch.norm(center.unsqueeze(2) - points.unsqueeze(1), p =2 ,dim = -1) # 1 1 2048\n\n idx = torch.argsort(distance_matrix,dim=-1, descending=False)[0,0] # 2048\n\n if padding_zeros:\n input_data = points.clone()\n input_data[0, idx[:num_crop]] = input_data[0,idx[:num_crop]] * 0\n\n else:\n input_data = points.clone()[0, idx[num_crop:]].unsqueeze(0) # 1 N 3\n\n crop_data = points.clone()[0, idx[:num_crop]].unsqueeze(0)\n\n if isinstance(crop,list):\n INPUT.append(fps(input_data,2048))\n CROP.append(fps(crop_data,2048))\n else:\n INPUT.append(input_data)\n CROP.append(crop_data)\n\n input_data = torch.cat(INPUT,dim=0)# B N 3\n crop_data = torch.cat(CROP,dim=0)# B M 3\n\n return input_data.contiguous(), crop_data.contiguous()\n\ndef get_ptcloud_img(ptcloud):\n fig = plt.figure(figsize=(8, 8))\n\n x, z, y = ptcloud.transpose(1, 0)\n ax = fig.gca(projection=Axes3D.name, adjustable='box')\n ax.axis('off')\n # ax.axis('scaled')\n ax.view_init(30, 45)\n max, min = np.max(ptcloud), np.min(ptcloud)\n ax.set_xbound(min, max)\n ax.set_ybound(min, max)\n ax.set_zbound(min, max)\n ax.scatter(x, y, z, zdir='z', c=x, cmap='jet')\n\n fig.canvas.draw()\n img = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\n img = img.reshape(fig.canvas.get_width_height()[::-1] + (3, ))\n return img\n","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.get_ptcloud_img","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.get_ptcloud_img#L225-L242","kind":"function","name":"get_ptcloud_img","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":225,"end_line":242,"context_start_line":205,"context_end_line":262,"code":" input_data = points.clone()\n input_data[0, idx[:num_crop]] = input_data[0,idx[:num_crop]] * 0\n\n else:\n input_data = points.clone()[0, idx[num_crop:]].unsqueeze(0) # 1 N 3\n\n crop_data = points.clone()[0, idx[:num_crop]].unsqueeze(0)\n\n if isinstance(crop,list):\n INPUT.append(fps(input_data,2048))\n CROP.append(fps(crop_data,2048))\n else:\n INPUT.append(input_data)\n CROP.append(crop_data)\n\n input_data = torch.cat(INPUT,dim=0)# B N 3\n crop_data = torch.cat(CROP,dim=0)# B M 3\n\n return input_data.contiguous(), crop_data.contiguous()\n\ndef get_ptcloud_img(ptcloud):\n fig = plt.figure(figsize=(8, 8))\n\n x, z, y = ptcloud.transpose(1, 0)\n ax = fig.gca(projection=Axes3D.name, adjustable='box')\n ax.axis('off')\n # ax.axis('scaled')\n ax.view_init(30, 45)\n max, min = np.max(ptcloud), np.min(ptcloud)\n ax.set_xbound(min, max)\n ax.set_ybound(min, max)\n ax.set_zbound(min, max)\n ax.scatter(x, y, z, zdir='z', c=x, cmap='jet')\n\n fig.canvas.draw()\n img = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\n img = img.reshape(fig.canvas.get_width_height()[::-1] + (3, ))\n return img\n\n\n\ndef visualize_KITTI(path, data_list, titles = ['input','pred'], cmap=['bwr','autumn'], zdir='y', \n xlim=(-1, 1), ylim=(-1, 1), zlim=(-1, 1) ):\n fig = plt.figure(figsize=(6*len(data_list),6))\n cmax = data_list[-1][:,0].max()\n\n for i in range(len(data_list)):\n data = data_list[i][:-2048] if i == 1 else data_list[i]\n color = data[:,0] /cmax\n ax = fig.add_subplot(1, len(data_list) , i + 1, projection='3d')\n ax.view_init(30, -120)\n b = ax.scatter(data[:, 0], data[:, 1], data[:, 2], zdir=zdir, c=color,vmin=-1,vmax=1 ,cmap = cmap[0],s=4,linewidth=0.05, edgecolors = 'black')\n ax.set_title(titles[i])\n\n ax.set_axis_off()\n ax.set_xlim(xlim)\n ax.set_ylim(ylim)\n ax.set_zlim(zlim)","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.visualize_KITTI","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.visualize_KITTI#L246-L272","kind":"function","name":"visualize_KITTI","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":246,"end_line":272,"context_start_line":226,"context_end_line":287,"code":" fig = plt.figure(figsize=(8, 8))\n\n x, z, y = ptcloud.transpose(1, 0)\n ax = fig.gca(projection=Axes3D.name, adjustable='box')\n ax.axis('off')\n # ax.axis('scaled')\n ax.view_init(30, 45)\n max, min = np.max(ptcloud), np.min(ptcloud)\n ax.set_xbound(min, max)\n ax.set_ybound(min, max)\n ax.set_zbound(min, max)\n ax.scatter(x, y, z, zdir='z', c=x, cmap='jet')\n\n fig.canvas.draw()\n img = np.fromstring(fig.canvas.tostring_rgb(), dtype=np.uint8, sep='')\n img = img.reshape(fig.canvas.get_width_height()[::-1] + (3, ))\n return img\n\n\n\ndef visualize_KITTI(path, data_list, titles = ['input','pred'], cmap=['bwr','autumn'], zdir='y', \n xlim=(-1, 1), ylim=(-1, 1), zlim=(-1, 1) ):\n fig = plt.figure(figsize=(6*len(data_list),6))\n cmax = data_list[-1][:,0].max()\n\n for i in range(len(data_list)):\n data = data_list[i][:-2048] if i == 1 else data_list[i]\n color = data[:,0] /cmax\n ax = fig.add_subplot(1, len(data_list) , i + 1, projection='3d')\n ax.view_init(30, -120)\n b = ax.scatter(data[:, 0], data[:, 1], data[:, 2], zdir=zdir, c=color,vmin=-1,vmax=1 ,cmap = cmap[0],s=4,linewidth=0.05, edgecolors = 'black')\n ax.set_title(titles[i])\n\n ax.set_axis_off()\n ax.set_xlim(xlim)\n ax.set_ylim(ylim)\n ax.set_zlim(zlim)\n plt.subplots_adjust(left=0, right=1, bottom=0, top=1, wspace=0.2, hspace=0)\n if not os.path.exists(path):\n os.makedirs(path)\n\n pic_path = path + '.png'\n fig.savefig(pic_path)\n\n np.save(os.path.join(path, 'input.npy'), data_list[0].numpy())\n np.save(os.path.join(path, 'pred.npy'), data_list[1].numpy())\n plt.close(fig)\n\n\ndef random_dropping(pc, e):\n up_num = max(64, 768 // (e//50 + 1))\n pc = pc\n random_num = torch.randint(1, up_num, (1,1))[0,0]\n pc = fps(pc, random_num)\n padding = torch.zeros(pc.size(0), 2048 - pc.size(1), 3).to(pc.device)\n pc = torch.cat([pc, padding], dim = 1)\n return pc\n \n\ndef random_scale(partial, scale_range=[0.8, 1.2]):\n scale = torch.rand(1).cuda() * (scale_range[1] - scale_range[0]) + scale_range[0]\n return partial * scale","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.random_dropping","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.random_dropping#L275-L282","kind":"function","name":"random_dropping","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":275,"end_line":282,"context_start_line":255,"context_end_line":287,"code":" ax.view_init(30, -120)\n b = ax.scatter(data[:, 0], data[:, 1], data[:, 2], zdir=zdir, c=color,vmin=-1,vmax=1 ,cmap = cmap[0],s=4,linewidth=0.05, edgecolors = 'black')\n ax.set_title(titles[i])\n\n ax.set_axis_off()\n ax.set_xlim(xlim)\n ax.set_ylim(ylim)\n ax.set_zlim(zlim)\n plt.subplots_adjust(left=0, right=1, bottom=0, top=1, wspace=0.2, hspace=0)\n if not os.path.exists(path):\n os.makedirs(path)\n\n pic_path = path + '.png'\n fig.savefig(pic_path)\n\n np.save(os.path.join(path, 'input.npy'), data_list[0].numpy())\n np.save(os.path.join(path, 'pred.npy'), data_list[1].numpy())\n plt.close(fig)\n\n\ndef random_dropping(pc, e):\n up_num = max(64, 768 // (e//50 + 1))\n pc = pc\n random_num = torch.randint(1, up_num, (1,1))[0,0]\n pc = fps(pc, random_num)\n padding = torch.zeros(pc.size(0), 2048 - pc.size(1), 3).to(pc.device)\n pc = torch.cat([pc, padding], dim = 1)\n return pc\n \n\ndef random_scale(partial, scale_range=[0.8, 1.2]):\n scale = torch.rand(1).cuda() * (scale_range[1] - scale_range[0]) + scale_range[0]\n return partial * scale","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.random_scale","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.random_scale#L285-L287","kind":"function","name":"random_scale","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":285,"end_line":287,"context_start_line":265,"context_end_line":287,"code":" os.makedirs(path)\n\n pic_path = path + '.png'\n fig.savefig(pic_path)\n\n np.save(os.path.join(path, 'input.npy'), data_list[0].numpy())\n np.save(os.path.join(path, 'pred.npy'), data_list[1].numpy())\n plt.close(fig)\n\n\ndef random_dropping(pc, e):\n up_num = max(64, 768 // (e//50 + 1))\n pc = pc\n random_num = torch.randint(1, up_num, (1,1))[0,0]\n pc = fps(pc, random_num)\n padding = torch.zeros(pc.size(0), 2048 - pc.size(1), 3).to(pc.device)\n pc = torch.cat([pc, padding], dim = 1)\n return pc\n \n\ndef random_scale(partial, scale_range=[0.8, 1.2]):\n scale = torch.rand(1).cuda() * (scale_range[1] - scale_range[0]) + scale_range[0]\n return partial * scale","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.fn","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.fn#L131-L133","kind":"function","name":"fn","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":131,"end_line":133,"context_start_line":111,"context_end_line":153,"code":" seq (Sequence): The sequence to be checked.\n expected_type (type): Expected type of sequence items.\n seq_type (type, optional): Expected sequence type.\n Returns:\n bool: Whether the sequence is valid.\n \"\"\"\n if seq_type is None:\n exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.__init__#L138-L154","kind":"function","name":"__init__","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":138,"end_line":154,"context_start_line":118,"context_end_line":174,"code":" exp_seq_type = abc.Sequence\n else:\n assert isinstance(seq_type, type)\n exp_seq_type = seq_type\n if not isinstance(seq, exp_seq_type):\n return False\n for item in seq:\n if not isinstance(item, expected_type):\n return False\n return True\n\n\ndef set_bn_momentum_default(bn_momentum):\n def fn(m):\n if isinstance(m, (nn.BatchNorm1d, nn.BatchNorm2d, nn.BatchNorm3d)):\n m.momentum = bn_momentum\n return fn\n\nclass BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.step","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.step#L156-L161","kind":"function","name":"step","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":156,"end_line":161,"context_start_line":136,"context_end_line":181,"code":"class BNMomentumScheduler(object):\n\n def __init__(\n self, model, bn_lambda, last_epoch=-1,\n setter=set_bn_momentum_default\n ):\n if not isinstance(model, nn.Module):\n raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape\n\n assert n == num_points\n assert c == 3\n if crop == num_points:\n return xyz, None\n \n INPUT = []","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.misc.get_momentum","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.misc.get_momentum#L163-L166","kind":"function","name":"get_momentum","path":"lavis/models/ulip_models/pointbert/misc.py","language":"python","start_line":163,"end_line":166,"context_start_line":143,"context_end_line":186,"code":" raise RuntimeError(\n \"Class '{}' is not a PyTorch nn Module\".format(\n type(model).__name__\n )\n )\n\n self.model = model\n self.setter = setter\n self.lmbd = bn_lambda\n\n self.step(last_epoch + 1)\n self.last_epoch = last_epoch\n\n def step(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n\n self.last_epoch = epoch\n self.model.apply(self.setter(self.lmbd(epoch)))\n\n def get_momentum(self, epoch=None):\n if epoch is None:\n epoch = self.last_epoch + 1\n return self.lmbd(epoch)\n\n\n\ndef seprate_point_cloud(xyz, num_points, crop, fixed_points = None, padding_zeros = False):\n '''\n seprate point cloud: usage : using to generate the incomplete point cloud with a setted number.\n '''\n _,n,c = xyz.shape\n\n assert n == num_points\n assert c == 3\n if crop == num_points:\n return xyz, None\n \n INPUT = []\n CROP = []\n for points in xyz:\n if isinstance(crop,list):\n num_crop = random.randint(crop[0],crop[1])\n else:","source_hash":"fc840f9ad9970d47a0eb7b38089e342ed4174ff94ff2b980487183032d888927","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint","uri":"program://CREMA/module/lavis.models.ulip_models.pointbert.checkpoint#L1-L126","kind":"module","name":"lavis.models.ulip_models.pointbert.checkpoint","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":1,"end_line":126,"context_start_line":1,"context_end_line":126,"code":"from collections import defaultdict\nimport torch.nn as nn\n\nfrom typing import Any\nfrom typing import Optional, List, Dict, NamedTuple, Tuple, Iterable\n\nfrom termcolor import colored\n\ndef get_missing_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the model but not found in a checkpoint.\n Args:\n keys (list[str]): List of keys that were not found in the checkpoint.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"Some model parameters or buffers are not found in the checkpoint:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"blue\") for k, v in groups.items()\n )\n return msg\n\n\ndef get_unexpected_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the checkpoint but not found in the model.\n Args:\n keys (list[str]): List of keys that were not found in the model.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"The checkpoint state_dict contains keys that are not used by the model:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"magenta\") for k, v in groups.items()\n )\n return msg\n\n\ndef _strip_prefix_if_present(state_dict: Dict[str, Any], prefix: str) -> None:\n \"\"\"\n Strip the prefix in metadata, if any.\n Args:\n state_dict (OrderedDict): a state-dict to be loaded to the model.\n prefix (str): prefix.\n \"\"\"\n keys = sorted(state_dict.keys())\n if not all(len(key) == 0 or key.startswith(prefix) for key in keys):\n return\n\n for key in keys:\n newkey = key[len(prefix):]\n state_dict[newkey] = state_dict.pop(key)\n\n # also strip the prefix in metadata, if any..\n try:\n metadata = state_dict._metadata # pyre-ignore\n except AttributeError:\n pass\n else:\n for key in list(metadata.keys()):\n # for the metadata dict, the key can be:\n # '': for the DDP module, which we want to remove.\n # 'module': for the actual model.\n # 'module.xx.xx': for the rest.\n\n if len(key) == 0:\n continue\n newkey = key[len(prefix):]\n metadata[newkey] = metadata.pop(key)\n\n\ndef _group_checkpoint_keys(keys: List[str]) -> Dict[str, List[str]]:\n \"\"\"\n Group keys based on common prefixes. A prefix is the string up to the final\n \".\" in each key.\n Args:\n keys (list[str]): list of parameter names, i.e. keys in the model\n checkpoint dict.\n Returns:\n dict[list]: keys with common prefixes are grouped into lists.\n \"\"\"\n groups = defaultdict(list)\n for key in keys:\n pos = key.rfind(\".\")\n if pos >= 0:\n head, tail = key[:pos], [key[pos + 1:]]\n else:\n head, tail = key, []\n groups[head].extend(tail)\n return groups\n\n\ndef _group_to_str(group: List[str]) -> str:\n \"\"\"\n Format a group of parameter name suffixes into a loggable string.\n Args:\n group (list[str]): list of parameter name suffixes.\n Returns:\n str: formated string.\n \"\"\"\n if len(group) == 0:\n return \"\"\n\n if len(group) == 1:\n return \".\" + group[0]\n\n return \".{\" + \", \".join(group) + \"}\"\n\n\ndef _named_modules_with_dup(\n model: nn.Module, prefix: str = \"\"\n) -> Iterable[Tuple[str, nn.Module]]:\n \"\"\"\n The same as `model.named_modules()`, except that it includes\n duplicated modules that have more than one name.\n \"\"\"\n yield prefix, model\n for name, module in model._modules.items(): # pyre-ignore\n if module is None:\n continue\n submodule_prefix = prefix + (\".\" if prefix else \"\") + name\n yield from _named_modules_with_dup(module, submodule_prefix)","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint.get_missing_parameters_message","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint.get_missing_parameters_message#L9-L23","kind":"function","name":"get_missing_parameters_message","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":9,"end_line":23,"context_start_line":1,"context_end_line":43,"code":"from collections import defaultdict\nimport torch.nn as nn\n\nfrom typing import Any\nfrom typing import Optional, List, Dict, NamedTuple, Tuple, Iterable\n\nfrom termcolor import colored\n\ndef get_missing_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the model but not found in a checkpoint.\n Args:\n keys (list[str]): List of keys that were not found in the checkpoint.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"Some model parameters or buffers are not found in the checkpoint:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"blue\") for k, v in groups.items()\n )\n return msg\n\n\ndef get_unexpected_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the checkpoint but not found in the model.\n Args:\n keys (list[str]): List of keys that were not found in the model.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"The checkpoint state_dict contains keys that are not used by the model:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"magenta\") for k, v in groups.items()\n )\n return msg\n\n\ndef _strip_prefix_if_present(state_dict: Dict[str, Any], prefix: str) -> None:","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint.get_unexpected_parameters_message","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint.get_unexpected_parameters_message#L26-L40","kind":"function","name":"get_unexpected_parameters_message","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":26,"end_line":40,"context_start_line":6,"context_end_line":60,"code":"\nfrom termcolor import colored\n\ndef get_missing_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the model but not found in a checkpoint.\n Args:\n keys (list[str]): List of keys that were not found in the checkpoint.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"Some model parameters or buffers are not found in the checkpoint:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"blue\") for k, v in groups.items()\n )\n return msg\n\n\ndef get_unexpected_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the checkpoint but not found in the model.\n Args:\n keys (list[str]): List of keys that were not found in the model.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"The checkpoint state_dict contains keys that are not used by the model:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"magenta\") for k, v in groups.items()\n )\n return msg\n\n\ndef _strip_prefix_if_present(state_dict: Dict[str, Any], prefix: str) -> None:\n \"\"\"\n Strip the prefix in metadata, if any.\n Args:\n state_dict (OrderedDict): a state-dict to be loaded to the model.\n prefix (str): prefix.\n \"\"\"\n keys = sorted(state_dict.keys())\n if not all(len(key) == 0 or key.startswith(prefix) for key in keys):\n return\n\n for key in keys:\n newkey = key[len(prefix):]\n state_dict[newkey] = state_dict.pop(key)\n\n # also strip the prefix in metadata, if any..\n try:\n metadata = state_dict._metadata # pyre-ignore","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint._strip_prefix_if_present","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint._strip_prefix_if_present#L43-L73","kind":"function","name":"_strip_prefix_if_present","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":43,"end_line":73,"context_start_line":23,"context_end_line":93,"code":" return msg\n\n\ndef get_unexpected_parameters_message(keys: List[str]) -> str:\n \"\"\"\n Get a logging-friendly message to report parameter names (keys) that are in\n the checkpoint but not found in the model.\n Args:\n keys (list[str]): List of keys that were not found in the model.\n Returns:\n str: message.\n \"\"\"\n groups = _group_checkpoint_keys(keys)\n msg = \"The checkpoint state_dict contains keys that are not used by the model:\\n\"\n msg += \"\\n\".join(\n \" \" + colored(k + _group_to_str(v), \"magenta\") for k, v in groups.items()\n )\n return msg\n\n\ndef _strip_prefix_if_present(state_dict: Dict[str, Any], prefix: str) -> None:\n \"\"\"\n Strip the prefix in metadata, if any.\n Args:\n state_dict (OrderedDict): a state-dict to be loaded to the model.\n prefix (str): prefix.\n \"\"\"\n keys = sorted(state_dict.keys())\n if not all(len(key) == 0 or key.startswith(prefix) for key in keys):\n return\n\n for key in keys:\n newkey = key[len(prefix):]\n state_dict[newkey] = state_dict.pop(key)\n\n # also strip the prefix in metadata, if any..\n try:\n metadata = state_dict._metadata # pyre-ignore\n except AttributeError:\n pass\n else:\n for key in list(metadata.keys()):\n # for the metadata dict, the key can be:\n # '': for the DDP module, which we want to remove.\n # 'module': for the actual model.\n # 'module.xx.xx': for the rest.\n\n if len(key) == 0:\n continue\n newkey = key[len(prefix):]\n metadata[newkey] = metadata.pop(key)\n\n\ndef _group_checkpoint_keys(keys: List[str]) -> Dict[str, List[str]]:\n \"\"\"\n Group keys based on common prefixes. A prefix is the string up to the final\n \".\" in each key.\n Args:\n keys (list[str]): list of parameter names, i.e. keys in the model\n checkpoint dict.\n Returns:\n dict[list]: keys with common prefixes are grouped into lists.\n \"\"\"\n groups = defaultdict(list)\n for key in keys:\n pos = key.rfind(\".\")\n if pos >= 0:\n head, tail = key[:pos], [key[pos + 1:]]\n else:\n head, tail = key, []\n groups[head].extend(tail)","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint._group_checkpoint_keys","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint._group_checkpoint_keys#L76-L94","kind":"function","name":"_group_checkpoint_keys","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":76,"end_line":94,"context_start_line":56,"context_end_line":114,"code":" state_dict[newkey] = state_dict.pop(key)\n\n # also strip the prefix in metadata, if any..\n try:\n metadata = state_dict._metadata # pyre-ignore\n except AttributeError:\n pass\n else:\n for key in list(metadata.keys()):\n # for the metadata dict, the key can be:\n # '': for the DDP module, which we want to remove.\n # 'module': for the actual model.\n # 'module.xx.xx': for the rest.\n\n if len(key) == 0:\n continue\n newkey = key[len(prefix):]\n metadata[newkey] = metadata.pop(key)\n\n\ndef _group_checkpoint_keys(keys: List[str]) -> Dict[str, List[str]]:\n \"\"\"\n Group keys based on common prefixes. A prefix is the string up to the final\n \".\" in each key.\n Args:\n keys (list[str]): list of parameter names, i.e. keys in the model\n checkpoint dict.\n Returns:\n dict[list]: keys with common prefixes are grouped into lists.\n \"\"\"\n groups = defaultdict(list)\n for key in keys:\n pos = key.rfind(\".\")\n if pos >= 0:\n head, tail = key[:pos], [key[pos + 1:]]\n else:\n head, tail = key, []\n groups[head].extend(tail)\n return groups\n\n\ndef _group_to_str(group: List[str]) -> str:\n \"\"\"\n Format a group of parameter name suffixes into a loggable string.\n Args:\n group (list[str]): list of parameter name suffixes.\n Returns:\n str: formated string.\n \"\"\"\n if len(group) == 0:\n return \"\"\n\n if len(group) == 1:\n return \".\" + group[0]\n\n return \".{\" + \", \".join(group) + \"}\"\n\n\ndef _named_modules_with_dup(","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint._group_to_str","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint._group_to_str#L97-L111","kind":"function","name":"_group_to_str","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":97,"end_line":111,"context_start_line":77,"context_end_line":126,"code":" \"\"\"\n Group keys based on common prefixes. A prefix is the string up to the final\n \".\" in each key.\n Args:\n keys (list[str]): list of parameter names, i.e. keys in the model\n checkpoint dict.\n Returns:\n dict[list]: keys with common prefixes are grouped into lists.\n \"\"\"\n groups = defaultdict(list)\n for key in keys:\n pos = key.rfind(\".\")\n if pos >= 0:\n head, tail = key[:pos], [key[pos + 1:]]\n else:\n head, tail = key, []\n groups[head].extend(tail)\n return groups\n\n\ndef _group_to_str(group: List[str]) -> str:\n \"\"\"\n Format a group of parameter name suffixes into a loggable string.\n Args:\n group (list[str]): list of parameter name suffixes.\n Returns:\n str: formated string.\n \"\"\"\n if len(group) == 0:\n return \"\"\n\n if len(group) == 1:\n return \".\" + group[0]\n\n return \".{\" + \", \".join(group) + \"}\"\n\n\ndef _named_modules_with_dup(\n model: nn.Module, prefix: str = \"\"\n) -> Iterable[Tuple[str, nn.Module]]:\n \"\"\"\n The same as `model.named_modules()`, except that it includes\n duplicated modules that have more than one name.\n \"\"\"\n yield prefix, model\n for name, module in model._modules.items(): # pyre-ignore\n if module is None:\n continue\n submodule_prefix = prefix + (\".\" if prefix else \"\") + name\n yield from _named_modules_with_dup(module, submodule_prefix)","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.checkpoint._named_modules_with_dup","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.checkpoint._named_modules_with_dup#L114-L126","kind":"function","name":"_named_modules_with_dup","path":"lavis/models/ulip_models/pointbert/checkpoint.py","language":"python","start_line":114,"end_line":126,"context_start_line":94,"context_end_line":126,"code":" return groups\n\n\ndef _group_to_str(group: List[str]) -> str:\n \"\"\"\n Format a group of parameter name suffixes into a loggable string.\n Args:\n group (list[str]): list of parameter name suffixes.\n Returns:\n str: formated string.\n \"\"\"\n if len(group) == 0:\n return \"\"\n\n if len(group) == 1:\n return \".\" + group[0]\n\n return \".{\" + \", \".join(group) + \"}\"\n\n\ndef _named_modules_with_dup(\n model: nn.Module, prefix: str = \"\"\n) -> Iterable[Tuple[str, nn.Module]]:\n \"\"\"\n The same as `model.named_modules()`, except that it includes\n duplicated modules that have more than one name.\n \"\"\"\n yield prefix, model\n for name, module in model._modules.items(): # pyre-ignore\n if module is None:\n continue\n submodule_prefix = prefix + (\".\" if prefix else \"\") + name\n yield from _named_modules_with_dup(module, submodule_prefix)","source_hash":"bd95d3c00499acc67b69555c7d9022731b3f53633dda6789380f32e41a084946","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder","uri":"program://CREMA/module/lavis.models.ulip_models.pointbert.point_encoder#L1-L225","kind":"module","name":"lavis.models.ulip_models.pointbert.point_encoder","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":1,"end_line":225,"context_start_line":1,"context_end_line":225,"code":"import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom timm.models.layers import DropPath\nfrom lavis.models.ulip_models.pointbert.dvae import Group\nfrom lavis.models.ulip_models.pointbert.dvae import Encoder\nfrom lavis.models.ulip_models.pointbert.logger import print_log\n\nfrom lavis.models.ulip_models.pointbert.checkpoint import get_missing_parameters_message, get_unexpected_parameters_message\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):\n super().__init__()\n self.norm1 = norm_layer(dim)\n\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)\n\n def forward(self, x):\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass TransformerEncoder(nn.Module):\n \"\"\" Transformer Encoder without hierarchical structure\n \"\"\"\n\n def __init__(self, embed_dim=768, depth=4, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None,\n drop_rate=0., attn_drop_rate=0., drop_path_rate=0.):\n super().__init__()\n\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate,\n drop_path=drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate\n )\n for i in range(depth)])\n\n def forward(self, x, pos):\n for _, block in enumerate(self.blocks):\n x = block(x + pos)\n return x\n\n\nclass PointTransformer(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.config = config\n # self.args = kwargs[\"args\"]\n self.num_features = 512\n self.trans_dim = config.trans_dim\n self.depth = config.depth\n self.drop_path_rate = config.drop_path_rate\n self.cls_dim = config.cls_dim\n self.num_heads = config.num_heads\n\n self.group_size = config.group_size\n self.num_group = config.num_group\n # grouper\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n # define the encoder\n self.encoder_dims = config.encoder_dims\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n # bridge encoder and transformer\n self.reduce_dim = nn.Linear(self.encoder_dims, self.trans_dim)\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, self.trans_dim))\n self.cls_pos = nn.Parameter(torch.randn(1, 1, self.trans_dim))\n\n self.pos_embed = nn.Sequential(\n nn.Linear(3, 128),\n nn.GELU(),\n nn.Linear(128, self.trans_dim)\n )\n\n dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, self.depth)]\n self.blocks = TransformerEncoder(\n embed_dim=self.trans_dim,\n depth=self.depth,\n drop_path_rate=dpr,\n num_heads=self.num_heads\n )\n\n self.norm = nn.LayerNorm(self.trans_dim)\n # self.load_model_from_ckpt('/export/home/repos/SLIP/pretrained_models/point_transformer_8192.pt')\n # if not self.args.evaluate_3d:\n ## TODO: pass as config\n # self.load_model_from_ckpt('/export/home/ULIP/data/initialize_models/ULIP-2_pointbert_last.pt')\n\n # self.cls_head_finetune = nn.Sequential(\n # nn.Linear(self.trans_dim * 2, 256),\n # nn.ReLU(inplace=True),\n # nn.Dropout(0.5),\n # nn.Linear(256, self.cls_dim)\n # )\n\n # self.build_loss_func()\n\n def build_loss_func(self):\n self.loss_ce = nn.CrossEntropyLoss()\n\n def get_loss_acc(self, pred, gt, smoothing=True):\n # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()\n\n if smoothing:\n eps = 0.2\n n_class = pred.size(1)\n\n one_hot = torch.zeros_like(pred).scatter(1, gt.view(-1, 1), 1)\n one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)\n log_prb = F.log_softmax(pred, dim=1)\n\n loss = -(one_hot * log_prb).sum(dim=1).mean()\n else:\n loss = self.loss_ce(pred, gt.long())\n\n pred = pred.argmax(-1)\n acc = (pred == gt).sum() / float(gt.size(0))\n\n return loss, acc * 100\n\n def load_model_from_ckpt(self, bert_ckpt_path):\n ckpt = torch.load(bert_ckpt_path, map_location='cpu')\n base_ckpt = {k.replace(\"module.\", \"\"): v for k, v in ckpt['state_dict'].items()}\n for k in list(base_ckpt.keys()):\n if k.startswith('transformer_q') and not k.startswith('transformer_q.cls_head'):\n base_ckpt[k[len('transformer_q.'):]] = base_ckpt[k]\n elif k.startswith('base_model'):\n base_ckpt[k[len('base_model.'):]] = base_ckpt[k]\n del base_ckpt[k]\n\n incompatible = self.load_state_dict(base_ckpt, strict=True)\n\n if incompatible.missing_keys:\n print_log('missing_keys', logger='Transformer')\n print_log(\n get_missing_parameters_message(incompatible.missing_keys),\n logger='Transformer'\n )\n if incompatible.unexpected_keys:\n print_log('unexpected_keys', logger='Transformer')\n print_log(\n get_unexpected_parameters_message(incompatible.unexpected_keys),\n logger='Transformer'\n )\n\n print_log(f'[Transformer] Successful Loading the ckpt from {bert_ckpt_path}', logger='Transformer')\n\n def forward(self, pts):\n # divide the point cloud in the same form. This is important\n neighborhood, center = self.group_divider(pts)\n # encoder the input cloud blocks\n group_input_tokens = self.encoder(neighborhood) # B G N\n group_input_tokens = self.reduce_dim(group_input_tokens)\n # prepare cls\n cls_tokens = self.cls_token.expand(group_input_tokens.size(0), -1, -1)\n cls_pos = self.cls_pos.expand(group_input_tokens.size(0), -1, -1)\n # add pos embedding\n pos = self.pos_embed(center)\n # final input\n x = torch.cat((cls_tokens, group_input_tokens), dim=1)\n pos = torch.cat((cls_pos, pos), dim=1)\n # transformer\n x = self.blocks(x, pos)\n x = self.norm(x)\n concat_f = torch.cat([x[:, 0], x[:, 1:].max(1)[0]], dim=-1)\n # ret = self.cls_head_finetune(concat_f)\n return concat_f","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.Mlp","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.point_encoder.Mlp#L11-L27","kind":"class","name":"Mlp","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":11,"end_line":27,"context_start_line":1,"context_end_line":47,"code":"import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom timm.models.layers import DropPath\nfrom lavis.models.ulip_models.pointbert.dvae import Group\nfrom lavis.models.ulip_models.pointbert.dvae import Encoder\nfrom lavis.models.ulip_models.pointbert.logger import print_log\n\nfrom lavis.models.ulip_models.pointbert.checkpoint import get_missing_parameters_message, get_unexpected_parameters_message\n\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.Attention","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.point_encoder.Attention#L30-L55","kind":"class","name":"Attention","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":30,"end_line":55,"context_start_line":10,"context_end_line":75,"code":"\nclass Mlp(nn.Module):\n def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights\n self.scale = qk_scale or head_dim ** -0.5\n\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):\n super().__init__()\n self.norm1 = norm_layer(dim)\n\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)\n\n def forward(self, x):\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.Block","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.point_encoder.Block#L58-L76","kind":"class","name":"Block","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":58,"end_line":76,"context_start_line":38,"context_end_line":96,"code":" self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.attn_drop = nn.Dropout(attn_drop)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)\n q, k, v = qkv[0], qkv[1], qkv[2] # make torchscript happy (cannot use tensor as tuple)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):\n super().__init__()\n self.norm1 = norm_layer(dim)\n\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)\n\n def forward(self, x):\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass TransformerEncoder(nn.Module):\n \"\"\" Transformer Encoder without hierarchical structure\n \"\"\"\n\n def __init__(self, embed_dim=768, depth=4, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None,\n drop_rate=0., attn_drop_rate=0., drop_path_rate=0.):\n super().__init__()\n\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate,\n drop_path=drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate\n )\n for i in range(depth)])\n\n def forward(self, x, pos):\n for _, block in enumerate(self.blocks):","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.TransformerEncoder","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.point_encoder.TransformerEncoder#L79-L98","kind":"class","name":"TransformerEncoder","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":79,"end_line":98,"context_start_line":59,"context_end_line":118,"code":" def __init__(self, dim, num_heads, mlp_ratio=4., qkv_bias=False, qk_scale=None, drop=0., attn_drop=0.,\n drop_path=0., act_layer=nn.GELU, norm_layer=nn.LayerNorm):\n super().__init__()\n self.norm1 = norm_layer(dim)\n\n # NOTE: drop path for stochastic depth, we shall see if this is better than dropout here\n self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)\n\n self.attn = Attention(\n dim, num_heads=num_heads, qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)\n\n def forward(self, x):\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n\n\nclass TransformerEncoder(nn.Module):\n \"\"\" Transformer Encoder without hierarchical structure\n \"\"\"\n\n def __init__(self, embed_dim=768, depth=4, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None,\n drop_rate=0., attn_drop_rate=0., drop_path_rate=0.):\n super().__init__()\n\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate,\n drop_path=drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate\n )\n for i in range(depth)])\n\n def forward(self, x, pos):\n for _, block in enumerate(self.blocks):\n x = block(x + pos)\n return x\n\n\nclass PointTransformer(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.config = config\n # self.args = kwargs[\"args\"]\n self.num_features = 512\n self.trans_dim = config.trans_dim\n self.depth = config.depth\n self.drop_path_rate = config.drop_path_rate\n self.cls_dim = config.cls_dim\n self.num_heads = config.num_heads\n\n self.group_size = config.group_size\n self.num_group = config.num_group\n # grouper\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n # define the encoder\n self.encoder_dims = config.encoder_dims","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.PointTransformer","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.point_encoder.PointTransformer#L101-L225","kind":"class","name":"PointTransformer","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":101,"end_line":225,"context_start_line":81,"context_end_line":225,"code":" \"\"\"\n\n def __init__(self, embed_dim=768, depth=4, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None,\n drop_rate=0., attn_drop_rate=0., drop_path_rate=0.):\n super().__init__()\n\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate,\n drop_path=drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate\n )\n for i in range(depth)])\n\n def forward(self, x, pos):\n for _, block in enumerate(self.blocks):\n x = block(x + pos)\n return x\n\n\nclass PointTransformer(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.config = config\n # self.args = kwargs[\"args\"]\n self.num_features = 512\n self.trans_dim = config.trans_dim\n self.depth = config.depth\n self.drop_path_rate = config.drop_path_rate\n self.cls_dim = config.cls_dim\n self.num_heads = config.num_heads\n\n self.group_size = config.group_size\n self.num_group = config.num_group\n # grouper\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n # define the encoder\n self.encoder_dims = config.encoder_dims\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n # bridge encoder and transformer\n self.reduce_dim = nn.Linear(self.encoder_dims, self.trans_dim)\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, self.trans_dim))\n self.cls_pos = nn.Parameter(torch.randn(1, 1, self.trans_dim))\n\n self.pos_embed = nn.Sequential(\n nn.Linear(3, 128),\n nn.GELU(),\n nn.Linear(128, self.trans_dim)\n )\n\n dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, self.depth)]\n self.blocks = TransformerEncoder(\n embed_dim=self.trans_dim,\n depth=self.depth,\n drop_path_rate=dpr,\n num_heads=self.num_heads\n )\n\n self.norm = nn.LayerNorm(self.trans_dim)\n # self.load_model_from_ckpt('/export/home/repos/SLIP/pretrained_models/point_transformer_8192.pt')\n # if not self.args.evaluate_3d:\n ## TODO: pass as config\n # self.load_model_from_ckpt('/export/home/ULIP/data/initialize_models/ULIP-2_pointbert_last.pt')\n\n # self.cls_head_finetune = nn.Sequential(\n # nn.Linear(self.trans_dim * 2, 256),\n # nn.ReLU(inplace=True),\n # nn.Dropout(0.5),\n # nn.Linear(256, self.cls_dim)\n # )\n\n # self.build_loss_func()\n\n def build_loss_func(self):\n self.loss_ce = nn.CrossEntropyLoss()\n\n def get_loss_acc(self, pred, gt, smoothing=True):\n # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()\n\n if smoothing:\n eps = 0.2\n n_class = pred.size(1)\n\n one_hot = torch.zeros_like(pred).scatter(1, gt.view(-1, 1), 1)\n one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)\n log_prb = F.log_softmax(pred, dim=1)\n\n loss = -(one_hot * log_prb).sum(dim=1).mean()\n else:\n loss = self.loss_ce(pred, gt.long())\n\n pred = pred.argmax(-1)\n acc = (pred == gt).sum() / float(gt.size(0))\n\n return loss, acc * 100\n\n def load_model_from_ckpt(self, bert_ckpt_path):\n ckpt = torch.load(bert_ckpt_path, map_location='cpu')\n base_ckpt = {k.replace(\"module.\", \"\"): v for k, v in ckpt['state_dict'].items()}\n for k in list(base_ckpt.keys()):\n if k.startswith('transformer_q') and not k.startswith('transformer_q.cls_head'):\n base_ckpt[k[len('transformer_q.'):]] = base_ckpt[k]\n elif k.startswith('base_model'):\n base_ckpt[k[len('base_model.'):]] = base_ckpt[k]\n del base_ckpt[k]\n\n incompatible = self.load_state_dict(base_ckpt, strict=True)\n\n if incompatible.missing_keys:\n print_log('missing_keys', logger='Transformer')\n print_log(\n get_missing_parameters_message(incompatible.missing_keys),\n logger='Transformer'\n )\n if incompatible.unexpected_keys:\n print_log('unexpected_keys', logger='Transformer')\n print_log(\n get_unexpected_parameters_message(incompatible.unexpected_keys),\n logger='Transformer'\n )\n\n print_log(f'[Transformer] Successful Loading the ckpt from {bert_ckpt_path}', logger='Transformer')\n\n def forward(self, pts):\n # divide the point cloud in the same form. This is important\n neighborhood, center = self.group_divider(pts)\n # encoder the input cloud blocks\n group_input_tokens = self.encoder(neighborhood) # B G N\n group_input_tokens = self.reduce_dim(group_input_tokens)\n # prepare cls\n cls_tokens = self.cls_token.expand(group_input_tokens.size(0), -1, -1)\n cls_pos = self.cls_pos.expand(group_input_tokens.size(0), -1, -1)\n # add pos embedding\n pos = self.pos_embed(center)\n # final input\n x = torch.cat((cls_tokens, group_input_tokens), dim=1)\n pos = torch.cat((cls_pos, pos), dim=1)\n # transformer\n x = self.blocks(x, pos)\n x = self.norm(x)\n concat_f = torch.cat([x[:, 0], x[:, 1:].max(1)[0]], dim=-1)\n # ret = self.cls_head_finetune(concat_f)\n return concat_f","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.point_encoder.__init__#L102-L140","kind":"function","name":"__init__","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":102,"end_line":140,"context_start_line":82,"context_end_line":160,"code":"\n def __init__(self, embed_dim=768, depth=4, num_heads=12, mlp_ratio=4., qkv_bias=False, qk_scale=None,\n drop_rate=0., attn_drop_rate=0., drop_path_rate=0.):\n super().__init__()\n\n self.blocks = nn.ModuleList([\n Block(\n dim=embed_dim, num_heads=num_heads, mlp_ratio=mlp_ratio, qkv_bias=qkv_bias, qk_scale=qk_scale,\n drop=drop_rate, attn_drop=attn_drop_rate,\n drop_path=drop_path_rate[i] if isinstance(drop_path_rate, list) else drop_path_rate\n )\n for i in range(depth)])\n\n def forward(self, x, pos):\n for _, block in enumerate(self.blocks):\n x = block(x + pos)\n return x\n\n\nclass PointTransformer(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.config = config\n # self.args = kwargs[\"args\"]\n self.num_features = 512\n self.trans_dim = config.trans_dim\n self.depth = config.depth\n self.drop_path_rate = config.drop_path_rate\n self.cls_dim = config.cls_dim\n self.num_heads = config.num_heads\n\n self.group_size = config.group_size\n self.num_group = config.num_group\n # grouper\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n # define the encoder\n self.encoder_dims = config.encoder_dims\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n # bridge encoder and transformer\n self.reduce_dim = nn.Linear(self.encoder_dims, self.trans_dim)\n\n self.cls_token = nn.Parameter(torch.zeros(1, 1, self.trans_dim))\n self.cls_pos = nn.Parameter(torch.randn(1, 1, self.trans_dim))\n\n self.pos_embed = nn.Sequential(\n nn.Linear(3, 128),\n nn.GELU(),\n nn.Linear(128, self.trans_dim)\n )\n\n dpr = [x.item() for x in torch.linspace(0, self.drop_path_rate, self.depth)]\n self.blocks = TransformerEncoder(\n embed_dim=self.trans_dim,\n depth=self.depth,\n drop_path_rate=dpr,\n num_heads=self.num_heads\n )\n\n self.norm = nn.LayerNorm(self.trans_dim)\n # self.load_model_from_ckpt('/export/home/repos/SLIP/pretrained_models/point_transformer_8192.pt')\n # if not self.args.evaluate_3d:\n ## TODO: pass as config\n # self.load_model_from_ckpt('/export/home/ULIP/data/initialize_models/ULIP-2_pointbert_last.pt')\n\n # self.cls_head_finetune = nn.Sequential(\n # nn.Linear(self.trans_dim * 2, 256),\n # nn.ReLU(inplace=True),\n # nn.Dropout(0.5),\n # nn.Linear(256, self.cls_dim)\n # )\n\n # self.build_loss_func()\n\n def build_loss_func(self):\n self.loss_ce = nn.CrossEntropyLoss()\n\n def get_loss_acc(self, pred, gt, smoothing=True):\n # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.forward","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.point_encoder.forward#L206-L225","kind":"function","name":"forward","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":206,"end_line":225,"context_start_line":186,"context_end_line":225,"code":" base_ckpt[k[len('base_model.'):]] = base_ckpt[k]\n del base_ckpt[k]\n\n incompatible = self.load_state_dict(base_ckpt, strict=True)\n\n if incompatible.missing_keys:\n print_log('missing_keys', logger='Transformer')\n print_log(\n get_missing_parameters_message(incompatible.missing_keys),\n logger='Transformer'\n )\n if incompatible.unexpected_keys:\n print_log('unexpected_keys', logger='Transformer')\n print_log(\n get_unexpected_parameters_message(incompatible.unexpected_keys),\n logger='Transformer'\n )\n\n print_log(f'[Transformer] Successful Loading the ckpt from {bert_ckpt_path}', logger='Transformer')\n\n def forward(self, pts):\n # divide the point cloud in the same form. This is important\n neighborhood, center = self.group_divider(pts)\n # encoder the input cloud blocks\n group_input_tokens = self.encoder(neighborhood) # B G N\n group_input_tokens = self.reduce_dim(group_input_tokens)\n # prepare cls\n cls_tokens = self.cls_token.expand(group_input_tokens.size(0), -1, -1)\n cls_pos = self.cls_pos.expand(group_input_tokens.size(0), -1, -1)\n # add pos embedding\n pos = self.pos_embed(center)\n # final input\n x = torch.cat((cls_tokens, group_input_tokens), dim=1)\n pos = torch.cat((cls_pos, pos), dim=1)\n # transformer\n x = self.blocks(x, pos)\n x = self.norm(x)\n concat_f = torch.cat([x[:, 0], x[:, 1:].max(1)[0]], dim=-1)\n # ret = self.cls_head_finetune(concat_f)\n return concat_f","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.build_loss_func","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.point_encoder.build_loss_func#L155-L156","kind":"function","name":"build_loss_func","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":155,"end_line":156,"context_start_line":135,"context_end_line":176,"code":" depth=self.depth,\n drop_path_rate=dpr,\n num_heads=self.num_heads\n )\n\n self.norm = nn.LayerNorm(self.trans_dim)\n # self.load_model_from_ckpt('/export/home/repos/SLIP/pretrained_models/point_transformer_8192.pt')\n # if not self.args.evaluate_3d:\n ## TODO: pass as config\n # self.load_model_from_ckpt('/export/home/ULIP/data/initialize_models/ULIP-2_pointbert_last.pt')\n\n # self.cls_head_finetune = nn.Sequential(\n # nn.Linear(self.trans_dim * 2, 256),\n # nn.ReLU(inplace=True),\n # nn.Dropout(0.5),\n # nn.Linear(256, self.cls_dim)\n # )\n\n # self.build_loss_func()\n\n def build_loss_func(self):\n self.loss_ce = nn.CrossEntropyLoss()\n\n def get_loss_acc(self, pred, gt, smoothing=True):\n # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()\n\n if smoothing:\n eps = 0.2\n n_class = pred.size(1)\n\n one_hot = torch.zeros_like(pred).scatter(1, gt.view(-1, 1), 1)\n one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)\n log_prb = F.log_softmax(pred, dim=1)\n\n loss = -(one_hot * log_prb).sum(dim=1).mean()\n else:\n loss = self.loss_ce(pred, gt.long())\n\n pred = pred.argmax(-1)\n acc = (pred == gt).sum() / float(gt.size(0))\n","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.get_loss_acc","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.point_encoder.get_loss_acc#L158-L177","kind":"function","name":"get_loss_acc","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":158,"end_line":177,"context_start_line":138,"context_end_line":197,"code":" )\n\n self.norm = nn.LayerNorm(self.trans_dim)\n # self.load_model_from_ckpt('/export/home/repos/SLIP/pretrained_models/point_transformer_8192.pt')\n # if not self.args.evaluate_3d:\n ## TODO: pass as config\n # self.load_model_from_ckpt('/export/home/ULIP/data/initialize_models/ULIP-2_pointbert_last.pt')\n\n # self.cls_head_finetune = nn.Sequential(\n # nn.Linear(self.trans_dim * 2, 256),\n # nn.ReLU(inplace=True),\n # nn.Dropout(0.5),\n # nn.Linear(256, self.cls_dim)\n # )\n\n # self.build_loss_func()\n\n def build_loss_func(self):\n self.loss_ce = nn.CrossEntropyLoss()\n\n def get_loss_acc(self, pred, gt, smoothing=True):\n # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()\n\n if smoothing:\n eps = 0.2\n n_class = pred.size(1)\n\n one_hot = torch.zeros_like(pred).scatter(1, gt.view(-1, 1), 1)\n one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)\n log_prb = F.log_softmax(pred, dim=1)\n\n loss = -(one_hot * log_prb).sum(dim=1).mean()\n else:\n loss = self.loss_ce(pred, gt.long())\n\n pred = pred.argmax(-1)\n acc = (pred == gt).sum() / float(gt.size(0))\n\n return loss, acc * 100\n\n def load_model_from_ckpt(self, bert_ckpt_path):\n ckpt = torch.load(bert_ckpt_path, map_location='cpu')\n base_ckpt = {k.replace(\"module.\", \"\"): v for k, v in ckpt['state_dict'].items()}\n for k in list(base_ckpt.keys()):\n if k.startswith('transformer_q') and not k.startswith('transformer_q.cls_head'):\n base_ckpt[k[len('transformer_q.'):]] = base_ckpt[k]\n elif k.startswith('base_model'):\n base_ckpt[k[len('base_model.'):]] = base_ckpt[k]\n del base_ckpt[k]\n\n incompatible = self.load_state_dict(base_ckpt, strict=True)\n\n if incompatible.missing_keys:\n print_log('missing_keys', logger='Transformer')\n print_log(\n get_missing_parameters_message(incompatible.missing_keys),\n logger='Transformer'\n )\n if incompatible.unexpected_keys:","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.point_encoder.load_model_from_ckpt","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.point_encoder.load_model_from_ckpt#L179-L204","kind":"function","name":"load_model_from_ckpt","path":"lavis/models/ulip_models/pointbert/point_encoder.py","language":"python","start_line":179,"end_line":204,"context_start_line":159,"context_end_line":224,"code":" # import pdb; pdb.set_trace()\n gt = gt.contiguous().view(-1).long()\n\n if smoothing:\n eps = 0.2\n n_class = pred.size(1)\n\n one_hot = torch.zeros_like(pred).scatter(1, gt.view(-1, 1), 1)\n one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)\n log_prb = F.log_softmax(pred, dim=1)\n\n loss = -(one_hot * log_prb).sum(dim=1).mean()\n else:\n loss = self.loss_ce(pred, gt.long())\n\n pred = pred.argmax(-1)\n acc = (pred == gt).sum() / float(gt.size(0))\n\n return loss, acc * 100\n\n def load_model_from_ckpt(self, bert_ckpt_path):\n ckpt = torch.load(bert_ckpt_path, map_location='cpu')\n base_ckpt = {k.replace(\"module.\", \"\"): v for k, v in ckpt['state_dict'].items()}\n for k in list(base_ckpt.keys()):\n if k.startswith('transformer_q') and not k.startswith('transformer_q.cls_head'):\n base_ckpt[k[len('transformer_q.'):]] = base_ckpt[k]\n elif k.startswith('base_model'):\n base_ckpt[k[len('base_model.'):]] = base_ckpt[k]\n del base_ckpt[k]\n\n incompatible = self.load_state_dict(base_ckpt, strict=True)\n\n if incompatible.missing_keys:\n print_log('missing_keys', logger='Transformer')\n print_log(\n get_missing_parameters_message(incompatible.missing_keys),\n logger='Transformer'\n )\n if incompatible.unexpected_keys:\n print_log('unexpected_keys', logger='Transformer')\n print_log(\n get_unexpected_parameters_message(incompatible.unexpected_keys),\n logger='Transformer'\n )\n\n print_log(f'[Transformer] Successful Loading the ckpt from {bert_ckpt_path}', logger='Transformer')\n\n def forward(self, pts):\n # divide the point cloud in the same form. This is important\n neighborhood, center = self.group_divider(pts)\n # encoder the input cloud blocks\n group_input_tokens = self.encoder(neighborhood) # B G N\n group_input_tokens = self.reduce_dim(group_input_tokens)\n # prepare cls\n cls_tokens = self.cls_token.expand(group_input_tokens.size(0), -1, -1)\n cls_pos = self.cls_pos.expand(group_input_tokens.size(0), -1, -1)\n # add pos embedding\n pos = self.pos_embed(center)\n # final input\n x = torch.cat((cls_tokens, group_input_tokens), dim=1)\n pos = torch.cat((cls_pos, pos), dim=1)\n # transformer\n x = self.blocks(x, pos)\n x = self.norm(x)\n concat_f = torch.cat([x[:, 0], x[:, 1:].max(1)[0]], dim=-1)\n # ret = self.cls_head_finetune(concat_f)","source_hash":"f7111e67d83b55dd7a53fbfb02665eb9a125a5408a3dc6342e79bb0064dac2fe","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.logger","uri":"program://CREMA/module/lavis.models.ulip_models.pointbert.logger#L1-L127","kind":"module","name":"lavis.models.ulip_models.pointbert.logger","path":"lavis/models/ulip_models/pointbert/logger.py","language":"python","start_line":1,"end_line":127,"context_start_line":1,"context_end_line":127,"code":"import logging\nimport torch.distributed as dist\n\nlogger_initialized = {}\n\ndef get_root_logger(log_file=None, log_level=logging.INFO, name='main'):\n \"\"\"Get root logger and add a keyword filter to it.\n The logger will be initialized if it has not been initialized. By default a\n StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.\n \"\"\"\n logger = logging.getLogger(name)\n if name in logger_initialized:\n return logger\n # handle hierarchical names\n # e.g., logger \"a\" is initialized, then logger \"a.b\" will skip the\n # initialization since it is a child of \"a\".\n for logger_name in logger_initialized:\n if name.startswith(logger_name):\n return logger\n\n # handle duplicate logs to the console\n # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler (NOTSET)\n # to the root logger. As logger.propagate is True by default, this root\n # level handler causes logging messages from rank>0 processes to\n # unexpectedly show up on the console, creating much unwanted clutter.\n # To fix this issue, we set the root logger's StreamHandler, if any, to log\n # at the ERROR level.\n for handler in logger.root.handlers:\n if type(handler) is logging.StreamHandler:\n handler.setLevel(logging.ERROR)\n\n stream_handler = logging.StreamHandler()\n handlers = [stream_handler]\n\n if dist.is_available() and dist.is_initialized():\n rank = dist.get_rank()\n else:\n rank = 0\n\n # only rank 0 will add a FileHandler\n if rank == 0 and log_file is not None:\n # Here, the default behaviour of the official logger is 'a'. Thus, we\n # provide an interface to change the file mode to the default\n # behaviour.\n file_handler = logging.FileHandler(log_file, file_mode)\n handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass\n elif isinstance(logger, str):\n _logger = get_logger(logger)\n _logger.log(level, msg)\n else:\n raise TypeError(\n 'logger should be either a logging.Logger object, str, '\n f'\"silent\" or None, but got {type(logger)}')","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.logger.get_root_logger","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.logger.get_root_logger#L6-L26","kind":"function","name":"get_root_logger","path":"lavis/models/ulip_models/pointbert/logger.py","language":"python","start_line":6,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"import logging\nimport torch.distributed as dist\n\nlogger_initialized = {}\n\ndef get_root_logger(log_file=None, log_level=logging.INFO, name='main'):\n \"\"\"Get root logger and add a keyword filter to it.\n The logger will be initialized if it has not been initialized. By default a\n StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.logger.get_logger","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.logger.get_logger#L29-L100","kind":"function","name":"get_logger","path":"lavis/models/ulip_models/pointbert/logger.py","language":"python","start_line":29,"end_line":100,"context_start_line":9,"context_end_line":120,"code":" StreamHandler will be added. If `log_file` is specified, a FileHandler will\n also be added. The name of the root logger is the top-level package name,\n e.g., \"mmdet3d\".\n Args:\n log_file (str, optional): File path of log. Defaults to None.\n log_level (int, optional): The level of logger.\n Defaults to logging.INFO.\n name (str, optional): The name of the root logger, also used as a\n filter keyword. Defaults to 'mmdet3d'.\n Returns:\n :obj:`logging.Logger`: The obtained logger\n \"\"\"\n logger = get_logger(name=name, log_file=log_file, log_level=log_level)\n # add a logging filter\n logging_filter = logging.Filter(name)\n logging_filter.filter = lambda record: record.find(name) != -1\n\n return logger\n\n\ndef get_logger(name, log_file=None, log_level=logging.INFO, file_mode='w'):\n \"\"\"Initialize and get a logger by name.\n If the logger has not been initialized, this method will initialize the\n logger by adding one or two handlers, otherwise the initialized logger will\n be directly returned. During initialization, a StreamHandler will always be\n added. If `log_file` is specified and the process rank is 0, a FileHandler\n will also be added.\n Args:\n name (str): Logger name.\n log_file (str | None): The log filename. If specified, a FileHandler\n will be added to the logger.\n log_level (int): The logger level. Note that only the process of\n rank 0 is affected, and other processes will set the level to\n \"Error\" thus be silent most of the time.\n file_mode (str): The file mode used in opening log file.\n Defaults to 'w'.\n Returns:\n logging.Logger: The expected logger.\n \"\"\"\n logger = logging.getLogger(name)\n if name in logger_initialized:\n return logger\n # handle hierarchical names\n # e.g., logger \"a\" is initialized, then logger \"a.b\" will skip the\n # initialization since it is a child of \"a\".\n for logger_name in logger_initialized:\n if name.startswith(logger_name):\n return logger\n\n # handle duplicate logs to the console\n # Starting in 1.8.0, PyTorch DDP attaches a StreamHandler (NOTSET)\n # to the root logger. As logger.propagate is True by default, this root\n # level handler causes logging messages from rank>0 processes to\n # unexpectedly show up on the console, creating much unwanted clutter.\n # To fix this issue, we set the root logger's StreamHandler, if any, to log\n # at the ERROR level.\n for handler in logger.root.handlers:\n if type(handler) is logging.StreamHandler:\n handler.setLevel(logging.ERROR)\n\n stream_handler = logging.StreamHandler()\n handlers = [stream_handler]\n\n if dist.is_available() and dist.is_initialized():\n rank = dist.get_rank()\n else:\n rank = 0\n\n # only rank 0 will add a FileHandler\n if rank == 0 and log_file is not None:\n # Here, the default behaviour of the official logger is 'a'. Thus, we\n # provide an interface to change the file mode to the default\n # behaviour.\n file_handler = logging.FileHandler(log_file, file_mode)\n handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.logger.print_log","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.logger.print_log#L103-L127","kind":"function","name":"print_log","path":"lavis/models/ulip_models/pointbert/logger.py","language":"python","start_line":103,"end_line":127,"context_start_line":83,"context_end_line":127,"code":" handlers.append(file_handler)\n\n formatter = logging.Formatter(\n '%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n for handler in handlers:\n handler.setFormatter(formatter)\n handler.setLevel(log_level)\n logger.addHandler(handler)\n\n if rank == 0:\n logger.setLevel(log_level)\n else:\n logger.setLevel(logging.ERROR)\n\n logger_initialized[name] = True\n\n\n return logger\n\n\ndef print_log(msg, logger=None, level=logging.INFO):\n \"\"\"Print a log message.\n Args:\n msg (str): The message to be logged.\n logger (logging.Logger | str | None): The logger to be used.\n Some special loggers are:\n - \"silent\": no message will be printed.\n - other str: the logger obtained with `get_root_logger(logger)`.\n - None: The `print()` method will be used to print log messages.\n level (int): Logging level. Only available when `logger` is a Logger\n object or \"root\".\n \"\"\"\n if logger is None:\n print(msg)\n elif isinstance(logger, logging.Logger):\n logger.log(level, msg)\n elif logger == 'silent':\n pass\n elif isinstance(logger, str):\n _logger = get_logger(logger)\n _logger.log(level, msg)\n else:\n raise TypeError(\n 'logger should be either a logging.Logger object, str, '\n f'\"silent\" or None, but got {type(logger)}')","source_hash":"d5ce96284512f59bfe386c857c9e3587fbd9afd605d4140f3383bfe7d0a03f69","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae","uri":"program://CREMA/module/lavis.models.ulip_models.pointbert.dvae#L1-L342","kind":"module","name":"lavis.models.ulip_models.pointbert.dvae","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":1,"end_line":342,"context_start_line":1,"context_end_line":342,"code":"import torch.nn as nn\nimport torch\nimport torch.nn.functional as F\nfrom lavis.models.ulip_models.pointbert import misc\n\nif torch.cuda.is_available():\n from knn_cuda import KNN\n\n knn = KNN(k=4, transpose_mode=False)\nelse:\n knn=None\n\nclass DGCNN(nn.Module):\n def __init__(self, encoder_channel, output_channel):\n super().__init__()\n '''\n K has to be 16\n '''\n self.input_trans = nn.Conv1d(encoder_channel, 128, 1)\n\n self.layer1 = nn.Sequential(nn.Conv2d(256, 256, kernel_size=1, bias=False),\n nn.GroupNorm(4, 256),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer2 = nn.Sequential(nn.Conv2d(512, 512, kernel_size=1, bias=False),\n nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer3 = nn.Sequential(nn.Conv2d(1024, 512, kernel_size=1, bias=False),\n nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer4 = nn.Sequential(nn.Conv2d(1024, 1024, kernel_size=1, bias=False),\n nn.GroupNorm(4, 1024),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer5 = nn.Sequential(nn.Conv1d(2304, output_channel, kernel_size=1, bias=False),\n nn.GroupNorm(4, output_channel),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n @staticmethod\n def get_graph_feature(coor_q, x_q, coor_k, x_k):\n # coor: bs, 3, np, x: bs, c, np\n\n k = 4\n batch_size = x_k.size(0)\n num_points_k = x_k.size(2)\n num_points_q = x_q.size(2)\n\n with torch.no_grad():\n _, idx = knn(coor_k, coor_q) # bs k np\n assert idx.shape[1] == k\n idx_base = torch.arange(0, batch_size, device=x_q.device).view(-1, 1, 1) * num_points_k\n idx = idx + idx_base\n idx = idx.view(-1)\n num_dims = x_k.size(1)\n x_k = x_k.transpose(2, 1).contiguous()\n feature = x_k.view(batch_size * num_points_k, -1)[idx, :]\n feature = feature.view(batch_size, k, num_points_q, num_dims).permute(0, 3, 2, 1).contiguous()\n x_q = x_q.view(batch_size, num_dims, num_points_q, 1).expand(-1, -1, -1, k)\n feature = torch.cat((feature - x_q, x_q), dim=1)\n return feature\n\n def forward(self, f, coor):\n # f: B G C\n # coor: B G 3\n\n # bs 3 N bs C N\n feature_list = []\n coor = coor.transpose(1, 2).contiguous() # B 3 N\n f = f.transpose(1, 2).contiguous() # B C N\n f = self.input_trans(f) # B 128 N\n\n f = self.get_graph_feature(coor, f, coor, f) # B 256 N k\n f = self.layer1(f) # B 256 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 256 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 512 N k\n f = self.layer2(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer3(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer4(f) # B 1024 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 1024 N\n feature_list.append(f)\n\n f = torch.cat(feature_list, dim=1) # B 2304 N\n\n f = self.layer5(f) # B C' N\n\n f = f.transpose(-1, -2)\n\n return f\n\n\n### ref https://github.com/Strawberry-Eat-Mango/PCT_Pytorch/blob/main/util.py ###\ndef knn_point(nsample, xyz, new_xyz):\n \"\"\"\n Input:\n nsample: max sample number in local region\n xyz: all points, [B, N, C]\n new_xyz: query points, [B, S, C]\n Return:\n group_idx: grouped points index, [B, S, nsample]\n \"\"\"\n sqrdists = square_distance(new_xyz, xyz)\n _, group_idx = torch.topk(sqrdists, nsample, dim=-1, largest=False, sorted=False)\n return group_idx\n\n\ndef square_distance(src, dst):\n \"\"\"\n Calculate Euclid distance between each two points.\n src^T * dst = xn * xm + yn * ym + zn * zm；\n sum(src^2, dim=-1) = xn*xn + yn*yn + zn*zn;\n sum(dst^2, dim=-1) = xm*xm + ym*ym + zm*zm;\n dist = (xn - xm)^2 + (yn - ym)^2 + (zn - zm)^2\n = sum(src**2,dim=-1)+sum(dst**2,dim=-1)-2*src^T*dst\n Input:\n src: source points, [B, N, C]\n dst: target points, [B, M, C]\n Output:\n dist: per-point square distance, [B, N, M]\n \"\"\"\n B, N, _ = src.shape\n _, M, _ = dst.shape\n dist = -2 * torch.matmul(src, dst.permute(0, 2, 1))\n dist += torch.sum(src ** 2, -1).view(B, N, 1)\n dist += torch.sum(dst ** 2, -1).view(B, 1, M)\n return dist\n\n\nclass Group(nn.Module):\n def __init__(self, num_group, group_size):\n super().__init__()\n self.num_group = num_group\n self.group_size = group_size\n self.knn = KNN(k=self.group_size, transpose_mode=True)\n\n def forward(self, xyz):\n '''\n input: B N 3\n ---------------------------\n output: B G M 3\n center : B G 3\n '''\n batch_size, num_points, _ = xyz.shape\n # fps the centers out\n center = misc.fps(xyz, self.num_group) # B G 3\n # knn to get the neighborhood\n # _, idx = self.knn(xyz, center) # B G M\n idx = knn_point(self.group_size, xyz, center) # B G M\n assert idx.size(1) == self.num_group\n assert idx.size(2) == self.group_size\n idx_base = torch.arange(0, batch_size, device=xyz.device).view(-1, 1, 1) * num_points\n idx = idx + idx_base\n idx = idx.view(-1)\n neighborhood = xyz.view(batch_size * num_points, -1)[idx, :]\n neighborhood = neighborhood.view(batch_size, self.num_group, self.group_size, 3).contiguous()\n # normalize\n neighborhood = neighborhood - center.unsqueeze(2)\n return neighborhood, center\n\n\nclass Encoder(nn.Module):\n def __init__(self, encoder_channel):\n super().__init__()\n self.encoder_channel = encoder_channel\n self.first_conv = nn.Sequential(\n nn.Conv1d(3, 128, 1),\n nn.BatchNorm1d(128),\n nn.ReLU(inplace=True),\n nn.Conv1d(128, 256, 1)\n )\n self.second_conv = nn.Sequential(\n nn.Conv1d(512, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, self.encoder_channel, 1)\n )\n\n def forward(self, point_groups):\n '''\n point_groups : B G N 3\n -----------------\n feature_global : B G C\n '''\n bs, g, n, _ = point_groups.shape\n point_groups = point_groups.reshape(bs * g, n, 3)\n # encoder\n feature = self.first_conv(point_groups.transpose(2, 1)) # BG 256 n\n feature_global = torch.max(feature, dim=2, keepdim=True)[0] # BG 256 1\n feature = torch.cat([feature_global.expand(-1, -1, n), feature], dim=1) # BG 512 n\n feature = self.second_conv(feature) # BG 1024 n\n feature_global = torch.max(feature, dim=2, keepdim=False)[0] # BG 1024\n return feature_global.reshape(bs, g, self.encoder_channel)\n\n\nclass Decoder(nn.Module):\n def __init__(self, encoder_channel, num_fine):\n super().__init__()\n self.num_fine = num_fine\n self.grid_size = 2\n self.num_coarse = self.num_fine // 4\n assert num_fine % 4 == 0\n\n self.mlp = nn.Sequential(\n nn.Linear(encoder_channel, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 3 * self.num_coarse)\n )\n self.final_conv = nn.Sequential(\n nn.Conv1d(encoder_channel + 3 + 2, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, 3, 1)\n )\n a = torch.linspace(-0.05, 0.05, steps=self.grid_size, dtype=torch.float).view(1, self.grid_size).expand(\n self.grid_size, self.grid_size).reshape(1, -1)\n b = torch.linspace(-0.05, 0.05, steps=self.grid_size, dtype=torch.float).view(self.grid_size, 1).expand(\n self.grid_size, self.grid_size).reshape(1, -1)\n self.folding_seed = torch.cat([a, b], dim=0).view(1, 2, self.grid_size ** 2) # 1 2 S\n\n def forward(self, feature_global):\n '''\n feature_global : B G C\n -------\n coarse : B G M 3\n fine : B G N 3\n\n '''\n bs, g, c = feature_global.shape\n feature_global = feature_global.reshape(bs * g, c)\n\n coarse = self.mlp(feature_global).reshape(bs * g, self.num_coarse, 3) # BG M 3\n\n point_feat = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n point_feat = point_feat.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n seed = self.folding_seed.unsqueeze(2).expand(bs * g, -1, self.num_coarse, -1) # BG 2 M (S)\n seed = seed.reshape(bs * g, -1, self.num_fine).to(feature_global.device) # BG 2 N\n\n feature_global = feature_global.unsqueeze(2).expand(-1, -1, self.num_fine) # BG 1024 N\n feat = torch.cat([feature_global, seed, point_feat], dim=1) # BG C N\n\n center = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n center = center.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n fine = self.final_conv(feat) + center # BG 3 N\n fine = fine.reshape(bs, g, 3, self.num_fine).transpose(-1, -2)\n coarse = coarse.reshape(bs, g, self.num_coarse, 3)\n return coarse, fine\n\n\nclass DiscreteVAE(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.group_size = config.group_size\n self.num_group = config.num_group\n self.encoder_dims = config.encoder_dims\n self.tokens_dims = config.tokens_dims\n\n self.decoder_dims = config.decoder_dims\n self.num_tokens = config.num_tokens\n\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n self.dgcnn_1 = DGCNN(encoder_channel=self.encoder_dims, output_channel=self.num_tokens)\n self.codebook = nn.Parameter(torch.randn(self.num_tokens, self.tokens_dims))\n\n self.dgcnn_2 = DGCNN(encoder_channel=self.tokens_dims, output_channel=self.decoder_dims)\n self.decoder = Decoder(encoder_channel=self.decoder_dims, num_fine=self.group_size)\n # self.build_loss_func()\n\n # def build_loss_func(self):\n # self.loss_func_cdl1 = ChamferDistanceL1().cuda()\n # self.loss_func_cdl2 = ChamferDistanceL2().cuda()\n # self.loss_func_emd = emd().cuda()\n\n def recon_loss(self, ret, gt):\n whole_coarse, whole_fine, coarse, fine, group_gt, _ = ret\n\n bs, g, _, _ = coarse.shape\n\n coarse = coarse.reshape(bs * g, -1, 3).contiguous()\n fine = fine.reshape(bs * g, -1, 3).contiguous()\n group_gt = group_gt.reshape(bs * g, -1, 3).contiguous()\n\n loss_coarse_block = self.loss_func_cdl1(coarse, group_gt)\n loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block\n\n return loss_recon\n\n def get_loss(self, ret, gt):\n # reconstruction loss\n loss_recon = self.recon_loss(ret, gt)\n # kl divergence\n logits = ret[-1] # B G N\n softmax = F.softmax(logits, dim=-1)\n mean_softmax = softmax.mean(dim=1)\n log_qy = torch.log(mean_softmax)\n log_uniform = torch.log(torch.tensor([1. / self.num_tokens], device=gt.device))\n loss_klv = F.kl_div(log_qy, log_uniform.expand(log_qy.size(0), log_qy.size(1)), None, None, 'batchmean',\n log_target=True)\n\n return loss_recon, loss_klv\n\n def forward(self, inp, temperature=1., hard=False, **kwargs):\n neighborhood, center = self.group_divider(inp)\n logits = self.encoder(neighborhood) # B G C\n logits = self.dgcnn_1(logits, center) # B G N\n soft_one_hot = F.gumbel_softmax(logits, tau=temperature, dim=2, hard=hard) # B G N\n sampled = torch.einsum('b g n, n c -> b g c', soft_one_hot, self.codebook) # B G C\n feature = self.dgcnn_2(sampled, center)\n coarse, fine = self.decoder(feature)\n\n with torch.no_grad():\n whole_fine = (fine + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n whole_coarse = (coarse + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n\n assert fine.size(2) == self.group_size\n ret = (whole_coarse, whole_fine, coarse, fine, neighborhood, logits)\n return ret","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.DGCNN","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.dvae.DGCNN#L13-L105","kind":"class","name":"DGCNN","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":13,"end_line":105,"context_start_line":1,"context_end_line":125,"code":"import torch.nn as nn\nimport torch\nimport torch.nn.functional as F\nfrom lavis.models.ulip_models.pointbert import misc\n\nif torch.cuda.is_available():\n from knn_cuda import KNN\n\n knn = KNN(k=4, transpose_mode=False)\nelse:\n knn=None\n\nclass DGCNN(nn.Module):\n def __init__(self, encoder_channel, output_channel):\n super().__init__()\n '''\n K has to be 16\n '''\n self.input_trans = nn.Conv1d(encoder_channel, 128, 1)\n\n self.layer1 = nn.Sequential(nn.Conv2d(256, 256, kernel_size=1, bias=False),\n nn.GroupNorm(4, 256),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer2 = nn.Sequential(nn.Conv2d(512, 512, kernel_size=1, bias=False),\n nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer3 = nn.Sequential(nn.Conv2d(1024, 512, kernel_size=1, bias=False),\n nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer4 = nn.Sequential(nn.Conv2d(1024, 1024, kernel_size=1, bias=False),\n nn.GroupNorm(4, 1024),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer5 = nn.Sequential(nn.Conv1d(2304, output_channel, kernel_size=1, bias=False),\n nn.GroupNorm(4, output_channel),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n @staticmethod\n def get_graph_feature(coor_q, x_q, coor_k, x_k):\n # coor: bs, 3, np, x: bs, c, np\n\n k = 4\n batch_size = x_k.size(0)\n num_points_k = x_k.size(2)\n num_points_q = x_q.size(2)\n\n with torch.no_grad():\n _, idx = knn(coor_k, coor_q) # bs k np\n assert idx.shape[1] == k\n idx_base = torch.arange(0, batch_size, device=x_q.device).view(-1, 1, 1) * num_points_k\n idx = idx + idx_base\n idx = idx.view(-1)\n num_dims = x_k.size(1)\n x_k = x_k.transpose(2, 1).contiguous()\n feature = x_k.view(batch_size * num_points_k, -1)[idx, :]\n feature = feature.view(batch_size, k, num_points_q, num_dims).permute(0, 3, 2, 1).contiguous()\n x_q = x_q.view(batch_size, num_dims, num_points_q, 1).expand(-1, -1, -1, k)\n feature = torch.cat((feature - x_q, x_q), dim=1)\n return feature\n\n def forward(self, f, coor):\n # f: B G C\n # coor: B G 3\n\n # bs 3 N bs C N\n feature_list = []\n coor = coor.transpose(1, 2).contiguous() # B 3 N\n f = f.transpose(1, 2).contiguous() # B C N\n f = self.input_trans(f) # B 128 N\n\n f = self.get_graph_feature(coor, f, coor, f) # B 256 N k\n f = self.layer1(f) # B 256 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 256 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 512 N k\n f = self.layer2(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer3(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer4(f) # B 1024 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 1024 N\n feature_list.append(f)\n\n f = torch.cat(feature_list, dim=1) # B 2304 N\n\n f = self.layer5(f) # B C' N\n\n f = f.transpose(-1, -2)\n\n return f\n\n\n### ref https://github.com/Strawberry-Eat-Mango/PCT_Pytorch/blob/main/util.py ###\ndef knn_point(nsample, xyz, new_xyz):\n \"\"\"\n Input:\n nsample: max sample number in local region\n xyz: all points, [B, N, C]\n new_xyz: query points, [B, S, C]\n Return:\n group_idx: grouped points index, [B, S, nsample]\n \"\"\"\n sqrdists = square_distance(new_xyz, xyz)\n _, group_idx = torch.topk(sqrdists, nsample, dim=-1, largest=False, sorted=False)\n return group_idx\n\n\ndef square_distance(src, dst):\n \"\"\"\n Calculate Euclid distance between each two points.","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.knn_point","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.knn_point#L109-L120","kind":"function","name":"knn_point","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":109,"end_line":120,"context_start_line":89,"context_end_line":140,"code":" f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer3(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 1024 N k\n f = self.layer4(f) # B 1024 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 1024 N\n feature_list.append(f)\n\n f = torch.cat(feature_list, dim=1) # B 2304 N\n\n f = self.layer5(f) # B C' N\n\n f = f.transpose(-1, -2)\n\n return f\n\n\n### ref https://github.com/Strawberry-Eat-Mango/PCT_Pytorch/blob/main/util.py ###\ndef knn_point(nsample, xyz, new_xyz):\n \"\"\"\n Input:\n nsample: max sample number in local region\n xyz: all points, [B, N, C]\n new_xyz: query points, [B, S, C]\n Return:\n group_idx: grouped points index, [B, S, nsample]\n \"\"\"\n sqrdists = square_distance(new_xyz, xyz)\n _, group_idx = torch.topk(sqrdists, nsample, dim=-1, largest=False, sorted=False)\n return group_idx\n\n\ndef square_distance(src, dst):\n \"\"\"\n Calculate Euclid distance between each two points.\n src^T * dst = xn * xm + yn * ym + zn * zm；\n sum(src^2, dim=-1) = xn*xn + yn*yn + zn*zn;\n sum(dst^2, dim=-1) = xm*xm + ym*ym + zm*zm;\n dist = (xn - xm)^2 + (yn - ym)^2 + (zn - zm)^2\n = sum(src**2,dim=-1)+sum(dst**2,dim=-1)-2*src^T*dst\n Input:\n src: source points, [B, N, C]\n dst: target points, [B, M, C]\n Output:\n dist: per-point square distance, [B, N, M]\n \"\"\"\n B, N, _ = src.shape\n _, M, _ = dst.shape\n dist = -2 * torch.matmul(src, dst.permute(0, 2, 1))\n dist += torch.sum(src ** 2, -1).view(B, N, 1)","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.square_distance","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.square_distance#L123-L142","kind":"function","name":"square_distance","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":123,"end_line":142,"context_start_line":103,"context_end_line":162,"code":" f = f.transpose(-1, -2)\n\n return f\n\n\n### ref https://github.com/Strawberry-Eat-Mango/PCT_Pytorch/blob/main/util.py ###\ndef knn_point(nsample, xyz, new_xyz):\n \"\"\"\n Input:\n nsample: max sample number in local region\n xyz: all points, [B, N, C]\n new_xyz: query points, [B, S, C]\n Return:\n group_idx: grouped points index, [B, S, nsample]\n \"\"\"\n sqrdists = square_distance(new_xyz, xyz)\n _, group_idx = torch.topk(sqrdists, nsample, dim=-1, largest=False, sorted=False)\n return group_idx\n\n\ndef square_distance(src, dst):\n \"\"\"\n Calculate Euclid distance between each two points.\n src^T * dst = xn * xm + yn * ym + zn * zm；\n sum(src^2, dim=-1) = xn*xn + yn*yn + zn*zn;\n sum(dst^2, dim=-1) = xm*xm + ym*ym + zm*zm;\n dist = (xn - xm)^2 + (yn - ym)^2 + (zn - zm)^2\n = sum(src**2,dim=-1)+sum(dst**2,dim=-1)-2*src^T*dst\n Input:\n src: source points, [B, N, C]\n dst: target points, [B, M, C]\n Output:\n dist: per-point square distance, [B, N, M]\n \"\"\"\n B, N, _ = src.shape\n _, M, _ = dst.shape\n dist = -2 * torch.matmul(src, dst.permute(0, 2, 1))\n dist += torch.sum(src ** 2, -1).view(B, N, 1)\n dist += torch.sum(dst ** 2, -1).view(B, 1, M)\n return dist\n\n\nclass Group(nn.Module):\n def __init__(self, num_group, group_size):\n super().__init__()\n self.num_group = num_group\n self.group_size = group_size\n self.knn = KNN(k=self.group_size, transpose_mode=True)\n\n def forward(self, xyz):\n '''\n input: B N 3\n ---------------------------\n output: B G M 3\n center : B G 3\n '''\n batch_size, num_points, _ = xyz.shape\n # fps the centers out\n center = misc.fps(xyz, self.num_group) # B G 3\n # knn to get the neighborhood","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.Group","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.dvae.Group#L145-L174","kind":"class","name":"Group","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":145,"end_line":174,"context_start_line":125,"context_end_line":194,"code":" Calculate Euclid distance between each two points.\n src^T * dst = xn * xm + yn * ym + zn * zm；\n sum(src^2, dim=-1) = xn*xn + yn*yn + zn*zn;\n sum(dst^2, dim=-1) = xm*xm + ym*ym + zm*zm;\n dist = (xn - xm)^2 + (yn - ym)^2 + (zn - zm)^2\n = sum(src**2,dim=-1)+sum(dst**2,dim=-1)-2*src^T*dst\n Input:\n src: source points, [B, N, C]\n dst: target points, [B, M, C]\n Output:\n dist: per-point square distance, [B, N, M]\n \"\"\"\n B, N, _ = src.shape\n _, M, _ = dst.shape\n dist = -2 * torch.matmul(src, dst.permute(0, 2, 1))\n dist += torch.sum(src ** 2, -1).view(B, N, 1)\n dist += torch.sum(dst ** 2, -1).view(B, 1, M)\n return dist\n\n\nclass Group(nn.Module):\n def __init__(self, num_group, group_size):\n super().__init__()\n self.num_group = num_group\n self.group_size = group_size\n self.knn = KNN(k=self.group_size, transpose_mode=True)\n\n def forward(self, xyz):\n '''\n input: B N 3\n ---------------------------\n output: B G M 3\n center : B G 3\n '''\n batch_size, num_points, _ = xyz.shape\n # fps the centers out\n center = misc.fps(xyz, self.num_group) # B G 3\n # knn to get the neighborhood\n # _, idx = self.knn(xyz, center) # B G M\n idx = knn_point(self.group_size, xyz, center) # B G M\n assert idx.size(1) == self.num_group\n assert idx.size(2) == self.group_size\n idx_base = torch.arange(0, batch_size, device=xyz.device).view(-1, 1, 1) * num_points\n idx = idx + idx_base\n idx = idx.view(-1)\n neighborhood = xyz.view(batch_size * num_points, -1)[idx, :]\n neighborhood = neighborhood.view(batch_size, self.num_group, self.group_size, 3).contiguous()\n # normalize\n neighborhood = neighborhood - center.unsqueeze(2)\n return neighborhood, center\n\n\nclass Encoder(nn.Module):\n def __init__(self, encoder_channel):\n super().__init__()\n self.encoder_channel = encoder_channel\n self.first_conv = nn.Sequential(\n nn.Conv1d(3, 128, 1),\n nn.BatchNorm1d(128),\n nn.ReLU(inplace=True),\n nn.Conv1d(128, 256, 1)\n )\n self.second_conv = nn.Sequential(\n nn.Conv1d(512, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, self.encoder_channel, 1)\n )\n\n def forward(self, point_groups):","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.Encoder","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.dvae.Encoder#L177-L208","kind":"class","name":"Encoder","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":177,"end_line":208,"context_start_line":157,"context_end_line":228,"code":" center : B G 3\n '''\n batch_size, num_points, _ = xyz.shape\n # fps the centers out\n center = misc.fps(xyz, self.num_group) # B G 3\n # knn to get the neighborhood\n # _, idx = self.knn(xyz, center) # B G M\n idx = knn_point(self.group_size, xyz, center) # B G M\n assert idx.size(1) == self.num_group\n assert idx.size(2) == self.group_size\n idx_base = torch.arange(0, batch_size, device=xyz.device).view(-1, 1, 1) * num_points\n idx = idx + idx_base\n idx = idx.view(-1)\n neighborhood = xyz.view(batch_size * num_points, -1)[idx, :]\n neighborhood = neighborhood.view(batch_size, self.num_group, self.group_size, 3).contiguous()\n # normalize\n neighborhood = neighborhood - center.unsqueeze(2)\n return neighborhood, center\n\n\nclass Encoder(nn.Module):\n def __init__(self, encoder_channel):\n super().__init__()\n self.encoder_channel = encoder_channel\n self.first_conv = nn.Sequential(\n nn.Conv1d(3, 128, 1),\n nn.BatchNorm1d(128),\n nn.ReLU(inplace=True),\n nn.Conv1d(128, 256, 1)\n )\n self.second_conv = nn.Sequential(\n nn.Conv1d(512, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, self.encoder_channel, 1)\n )\n\n def forward(self, point_groups):\n '''\n point_groups : B G N 3\n -----------------\n feature_global : B G C\n '''\n bs, g, n, _ = point_groups.shape\n point_groups = point_groups.reshape(bs * g, n, 3)\n # encoder\n feature = self.first_conv(point_groups.transpose(2, 1)) # BG 256 n\n feature_global = torch.max(feature, dim=2, keepdim=True)[0] # BG 256 1\n feature = torch.cat([feature_global.expand(-1, -1, n), feature], dim=1) # BG 512 n\n feature = self.second_conv(feature) # BG 1024 n\n feature_global = torch.max(feature, dim=2, keepdim=False)[0] # BG 1024\n return feature_global.reshape(bs, g, self.encoder_channel)\n\n\nclass Decoder(nn.Module):\n def __init__(self, encoder_channel, num_fine):\n super().__init__()\n self.num_fine = num_fine\n self.grid_size = 2\n self.num_coarse = self.num_fine // 4\n assert num_fine % 4 == 0\n\n self.mlp = nn.Sequential(\n nn.Linear(encoder_channel, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 3 * self.num_coarse)\n )\n self.final_conv = nn.Sequential(\n nn.Conv1d(encoder_channel + 3 + 2, 512, 1),\n nn.BatchNorm1d(512),","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.Decoder","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.dvae.Decoder#L211-L269","kind":"class","name":"Decoder","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":211,"end_line":269,"context_start_line":191,"context_end_line":289,"code":" nn.Conv1d(512, self.encoder_channel, 1)\n )\n\n def forward(self, point_groups):\n '''\n point_groups : B G N 3\n -----------------\n feature_global : B G C\n '''\n bs, g, n, _ = point_groups.shape\n point_groups = point_groups.reshape(bs * g, n, 3)\n # encoder\n feature = self.first_conv(point_groups.transpose(2, 1)) # BG 256 n\n feature_global = torch.max(feature, dim=2, keepdim=True)[0] # BG 256 1\n feature = torch.cat([feature_global.expand(-1, -1, n), feature], dim=1) # BG 512 n\n feature = self.second_conv(feature) # BG 1024 n\n feature_global = torch.max(feature, dim=2, keepdim=False)[0] # BG 1024\n return feature_global.reshape(bs, g, self.encoder_channel)\n\n\nclass Decoder(nn.Module):\n def __init__(self, encoder_channel, num_fine):\n super().__init__()\n self.num_fine = num_fine\n self.grid_size = 2\n self.num_coarse = self.num_fine // 4\n assert num_fine % 4 == 0\n\n self.mlp = nn.Sequential(\n nn.Linear(encoder_channel, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 1024),\n nn.ReLU(inplace=True),\n nn.Linear(1024, 3 * self.num_coarse)\n )\n self.final_conv = nn.Sequential(\n nn.Conv1d(encoder_channel + 3 + 2, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, 512, 1),\n nn.BatchNorm1d(512),\n nn.ReLU(inplace=True),\n nn.Conv1d(512, 3, 1)\n )\n a = torch.linspace(-0.05, 0.05, steps=self.grid_size, dtype=torch.float).view(1, self.grid_size).expand(\n self.grid_size, self.grid_size).reshape(1, -1)\n b = torch.linspace(-0.05, 0.05, steps=self.grid_size, dtype=torch.float).view(self.grid_size, 1).expand(\n self.grid_size, self.grid_size).reshape(1, -1)\n self.folding_seed = torch.cat([a, b], dim=0).view(1, 2, self.grid_size ** 2) # 1 2 S\n\n def forward(self, feature_global):\n '''\n feature_global : B G C\n -------\n coarse : B G M 3\n fine : B G N 3\n\n '''\n bs, g, c = feature_global.shape\n feature_global = feature_global.reshape(bs * g, c)\n\n coarse = self.mlp(feature_global).reshape(bs * g, self.num_coarse, 3) # BG M 3\n\n point_feat = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n point_feat = point_feat.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n seed = self.folding_seed.unsqueeze(2).expand(bs * g, -1, self.num_coarse, -1) # BG 2 M (S)\n seed = seed.reshape(bs * g, -1, self.num_fine).to(feature_global.device) # BG 2 N\n\n feature_global = feature_global.unsqueeze(2).expand(-1, -1, self.num_fine) # BG 1024 N\n feat = torch.cat([feature_global, seed, point_feat], dim=1) # BG C N\n\n center = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n center = center.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n fine = self.final_conv(feat) + center # BG 3 N\n fine = fine.reshape(bs, g, 3, self.num_fine).transpose(-1, -2)\n coarse = coarse.reshape(bs, g, self.num_coarse, 3)\n return coarse, fine\n\n\nclass DiscreteVAE(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.group_size = config.group_size\n self.num_group = config.num_group\n self.encoder_dims = config.encoder_dims\n self.tokens_dims = config.tokens_dims\n\n self.decoder_dims = config.decoder_dims\n self.num_tokens = config.num_tokens\n\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n self.dgcnn_1 = DGCNN(encoder_channel=self.encoder_dims, output_channel=self.num_tokens)\n self.codebook = nn.Parameter(torch.randn(self.num_tokens, self.tokens_dims))\n\n self.dgcnn_2 = DGCNN(encoder_channel=self.tokens_dims, output_channel=self.decoder_dims)\n self.decoder = Decoder(encoder_channel=self.decoder_dims, num_fine=self.group_size)","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.DiscreteVAE","uri":"program://CREMA/class/lavis.models.ulip_models.pointbert.dvae.DiscreteVAE#L272-L342","kind":"class","name":"DiscreteVAE","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":272,"end_line":342,"context_start_line":252,"context_end_line":342,"code":" coarse = self.mlp(feature_global).reshape(bs * g, self.num_coarse, 3) # BG M 3\n\n point_feat = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n point_feat = point_feat.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n seed = self.folding_seed.unsqueeze(2).expand(bs * g, -1, self.num_coarse, -1) # BG 2 M (S)\n seed = seed.reshape(bs * g, -1, self.num_fine).to(feature_global.device) # BG 2 N\n\n feature_global = feature_global.unsqueeze(2).expand(-1, -1, self.num_fine) # BG 1024 N\n feat = torch.cat([feature_global, seed, point_feat], dim=1) # BG C N\n\n center = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n center = center.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n fine = self.final_conv(feat) + center # BG 3 N\n fine = fine.reshape(bs, g, 3, self.num_fine).transpose(-1, -2)\n coarse = coarse.reshape(bs, g, self.num_coarse, 3)\n return coarse, fine\n\n\nclass DiscreteVAE(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.group_size = config.group_size\n self.num_group = config.num_group\n self.encoder_dims = config.encoder_dims\n self.tokens_dims = config.tokens_dims\n\n self.decoder_dims = config.decoder_dims\n self.num_tokens = config.num_tokens\n\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n self.dgcnn_1 = DGCNN(encoder_channel=self.encoder_dims, output_channel=self.num_tokens)\n self.codebook = nn.Parameter(torch.randn(self.num_tokens, self.tokens_dims))\n\n self.dgcnn_2 = DGCNN(encoder_channel=self.tokens_dims, output_channel=self.decoder_dims)\n self.decoder = Decoder(encoder_channel=self.decoder_dims, num_fine=self.group_size)\n # self.build_loss_func()\n\n # def build_loss_func(self):\n # self.loss_func_cdl1 = ChamferDistanceL1().cuda()\n # self.loss_func_cdl2 = ChamferDistanceL2().cuda()\n # self.loss_func_emd = emd().cuda()\n\n def recon_loss(self, ret, gt):\n whole_coarse, whole_fine, coarse, fine, group_gt, _ = ret\n\n bs, g, _, _ = coarse.shape\n\n coarse = coarse.reshape(bs * g, -1, 3).contiguous()\n fine = fine.reshape(bs * g, -1, 3).contiguous()\n group_gt = group_gt.reshape(bs * g, -1, 3).contiguous()\n\n loss_coarse_block = self.loss_func_cdl1(coarse, group_gt)\n loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block\n\n return loss_recon\n\n def get_loss(self, ret, gt):\n # reconstruction loss\n loss_recon = self.recon_loss(ret, gt)\n # kl divergence\n logits = ret[-1] # B G N\n softmax = F.softmax(logits, dim=-1)\n mean_softmax = softmax.mean(dim=1)\n log_qy = torch.log(mean_softmax)\n log_uniform = torch.log(torch.tensor([1. / self.num_tokens], device=gt.device))\n loss_klv = F.kl_div(log_qy, log_uniform.expand(log_qy.size(0), log_qy.size(1)), None, None, 'batchmean',\n log_target=True)\n\n return loss_recon, loss_klv\n\n def forward(self, inp, temperature=1., hard=False, **kwargs):\n neighborhood, center = self.group_divider(inp)\n logits = self.encoder(neighborhood) # B G C\n logits = self.dgcnn_1(logits, center) # B G N\n soft_one_hot = F.gumbel_softmax(logits, tau=temperature, dim=2, hard=hard) # B G N\n sampled = torch.einsum('b g n, n c -> b g c', soft_one_hot, self.codebook) # B G C\n feature = self.dgcnn_2(sampled, center)\n coarse, fine = self.decoder(feature)\n\n with torch.no_grad():\n whole_fine = (fine + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n whole_coarse = (coarse + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n\n assert fine.size(2) == self.group_size\n ret = (whole_coarse, whole_fine, coarse, fine, neighborhood, logits)\n return ret","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.__init__","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.__init__#L273-L289","kind":"function","name":"__init__","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":273,"end_line":289,"context_start_line":253,"context_end_line":309,"code":"\n point_feat = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n point_feat = point_feat.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n seed = self.folding_seed.unsqueeze(2).expand(bs * g, -1, self.num_coarse, -1) # BG 2 M (S)\n seed = seed.reshape(bs * g, -1, self.num_fine).to(feature_global.device) # BG 2 N\n\n feature_global = feature_global.unsqueeze(2).expand(-1, -1, self.num_fine) # BG 1024 N\n feat = torch.cat([feature_global, seed, point_feat], dim=1) # BG C N\n\n center = coarse.unsqueeze(2).expand(-1, -1, self.grid_size ** 2, -1) # BG (M) S 3\n center = center.reshape(bs * g, self.num_fine, 3).transpose(2, 1) # BG 3 N\n\n fine = self.final_conv(feat) + center # BG 3 N\n fine = fine.reshape(bs, g, 3, self.num_fine).transpose(-1, -2)\n coarse = coarse.reshape(bs, g, self.num_coarse, 3)\n return coarse, fine\n\n\nclass DiscreteVAE(nn.Module):\n def __init__(self, config, **kwargs):\n super().__init__()\n self.group_size = config.group_size\n self.num_group = config.num_group\n self.encoder_dims = config.encoder_dims\n self.tokens_dims = config.tokens_dims\n\n self.decoder_dims = config.decoder_dims\n self.num_tokens = config.num_tokens\n\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n self.dgcnn_1 = DGCNN(encoder_channel=self.encoder_dims, output_channel=self.num_tokens)\n self.codebook = nn.Parameter(torch.randn(self.num_tokens, self.tokens_dims))\n\n self.dgcnn_2 = DGCNN(encoder_channel=self.tokens_dims, output_channel=self.decoder_dims)\n self.decoder = Decoder(encoder_channel=self.decoder_dims, num_fine=self.group_size)\n # self.build_loss_func()\n\n # def build_loss_func(self):\n # self.loss_func_cdl1 = ChamferDistanceL1().cuda()\n # self.loss_func_cdl2 = ChamferDistanceL2().cuda()\n # self.loss_func_emd = emd().cuda()\n\n def recon_loss(self, ret, gt):\n whole_coarse, whole_fine, coarse, fine, group_gt, _ = ret\n\n bs, g, _, _ = coarse.shape\n\n coarse = coarse.reshape(bs * g, -1, 3).contiguous()\n fine = fine.reshape(bs * g, -1, 3).contiguous()\n group_gt = group_gt.reshape(bs * g, -1, 3).contiguous()\n\n loss_coarse_block = self.loss_func_cdl1(coarse, group_gt)\n loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.get_graph_feature","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.get_graph_feature#L47-L67","kind":"function","name":"get_graph_feature","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":47,"end_line":67,"context_start_line":27,"context_end_line":87,"code":" nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer3 = nn.Sequential(nn.Conv2d(1024, 512, kernel_size=1, bias=False),\n nn.GroupNorm(4, 512),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer4 = nn.Sequential(nn.Conv2d(1024, 1024, kernel_size=1, bias=False),\n nn.GroupNorm(4, 1024),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n self.layer5 = nn.Sequential(nn.Conv1d(2304, output_channel, kernel_size=1, bias=False),\n nn.GroupNorm(4, output_channel),\n nn.LeakyReLU(negative_slope=0.2)\n )\n\n @staticmethod\n def get_graph_feature(coor_q, x_q, coor_k, x_k):\n # coor: bs, 3, np, x: bs, c, np\n\n k = 4\n batch_size = x_k.size(0)\n num_points_k = x_k.size(2)\n num_points_q = x_q.size(2)\n\n with torch.no_grad():\n _, idx = knn(coor_k, coor_q) # bs k np\n assert idx.shape[1] == k\n idx_base = torch.arange(0, batch_size, device=x_q.device).view(-1, 1, 1) * num_points_k\n idx = idx + idx_base\n idx = idx.view(-1)\n num_dims = x_k.size(1)\n x_k = x_k.transpose(2, 1).contiguous()\n feature = x_k.view(batch_size * num_points_k, -1)[idx, :]\n feature = feature.view(batch_size, k, num_points_q, num_dims).permute(0, 3, 2, 1).contiguous()\n x_q = x_q.view(batch_size, num_dims, num_points_q, 1).expand(-1, -1, -1, k)\n feature = torch.cat((feature - x_q, x_q), dim=1)\n return feature\n\n def forward(self, f, coor):\n # f: B G C\n # coor: B G 3\n\n # bs 3 N bs C N\n feature_list = []\n coor = coor.transpose(1, 2).contiguous() # B 3 N\n f = f.transpose(1, 2).contiguous() # B C N\n f = self.input_trans(f) # B 128 N\n\n f = self.get_graph_feature(coor, f, coor, f) # B 256 N k\n f = self.layer1(f) # B 256 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 256 N\n feature_list.append(f)\n\n f = self.get_graph_feature(coor, f, coor, f) # B 512 N k\n f = self.layer2(f) # B 512 N k\n f = f.max(dim=-1, keepdim=False)[0] # B 512 N\n feature_list.append(f)","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.forward","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.forward#L327-L342","kind":"function","name":"forward","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":327,"end_line":342,"context_start_line":307,"context_end_line":342,"code":" loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block\n\n return loss_recon\n\n def get_loss(self, ret, gt):\n # reconstruction loss\n loss_recon = self.recon_loss(ret, gt)\n # kl divergence\n logits = ret[-1] # B G N\n softmax = F.softmax(logits, dim=-1)\n mean_softmax = softmax.mean(dim=1)\n log_qy = torch.log(mean_softmax)\n log_uniform = torch.log(torch.tensor([1. / self.num_tokens], device=gt.device))\n loss_klv = F.kl_div(log_qy, log_uniform.expand(log_qy.size(0), log_qy.size(1)), None, None, 'batchmean',\n log_target=True)\n\n return loss_recon, loss_klv\n\n def forward(self, inp, temperature=1., hard=False, **kwargs):\n neighborhood, center = self.group_divider(inp)\n logits = self.encoder(neighborhood) # B G C\n logits = self.dgcnn_1(logits, center) # B G N\n soft_one_hot = F.gumbel_softmax(logits, tau=temperature, dim=2, hard=hard) # B G N\n sampled = torch.einsum('b g n, n c -> b g c', soft_one_hot, self.codebook) # B G C\n feature = self.dgcnn_2(sampled, center)\n coarse, fine = self.decoder(feature)\n\n with torch.no_grad():\n whole_fine = (fine + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n whole_coarse = (coarse + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n\n assert fine.size(2) == self.group_size\n ret = (whole_coarse, whole_fine, coarse, fine, neighborhood, logits)\n return ret","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.recon_loss","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.recon_loss#L297-L311","kind":"function","name":"recon_loss","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":297,"end_line":311,"context_start_line":277,"context_end_line":331,"code":" self.encoder_dims = config.encoder_dims\n self.tokens_dims = config.tokens_dims\n\n self.decoder_dims = config.decoder_dims\n self.num_tokens = config.num_tokens\n\n self.group_divider = Group(num_group=self.num_group, group_size=self.group_size)\n self.encoder = Encoder(encoder_channel=self.encoder_dims)\n self.dgcnn_1 = DGCNN(encoder_channel=self.encoder_dims, output_channel=self.num_tokens)\n self.codebook = nn.Parameter(torch.randn(self.num_tokens, self.tokens_dims))\n\n self.dgcnn_2 = DGCNN(encoder_channel=self.tokens_dims, output_channel=self.decoder_dims)\n self.decoder = Decoder(encoder_channel=self.decoder_dims, num_fine=self.group_size)\n # self.build_loss_func()\n\n # def build_loss_func(self):\n # self.loss_func_cdl1 = ChamferDistanceL1().cuda()\n # self.loss_func_cdl2 = ChamferDistanceL2().cuda()\n # self.loss_func_emd = emd().cuda()\n\n def recon_loss(self, ret, gt):\n whole_coarse, whole_fine, coarse, fine, group_gt, _ = ret\n\n bs, g, _, _ = coarse.shape\n\n coarse = coarse.reshape(bs * g, -1, 3).contiguous()\n fine = fine.reshape(bs * g, -1, 3).contiguous()\n group_gt = group_gt.reshape(bs * g, -1, 3).contiguous()\n\n loss_coarse_block = self.loss_func_cdl1(coarse, group_gt)\n loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block\n\n return loss_recon\n\n def get_loss(self, ret, gt):\n # reconstruction loss\n loss_recon = self.recon_loss(ret, gt)\n # kl divergence\n logits = ret[-1] # B G N\n softmax = F.softmax(logits, dim=-1)\n mean_softmax = softmax.mean(dim=1)\n log_qy = torch.log(mean_softmax)\n log_uniform = torch.log(torch.tensor([1. / self.num_tokens], device=gt.device))\n loss_klv = F.kl_div(log_qy, log_uniform.expand(log_qy.size(0), log_qy.size(1)), None, None, 'batchmean',\n log_target=True)\n\n return loss_recon, loss_klv\n\n def forward(self, inp, temperature=1., hard=False, **kwargs):\n neighborhood, center = self.group_divider(inp)\n logits = self.encoder(neighborhood) # B G C\n logits = self.dgcnn_1(logits, center) # B G N\n soft_one_hot = F.gumbel_softmax(logits, tau=temperature, dim=2, hard=hard) # B G N","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.ulip_models.pointbert.dvae.get_loss","uri":"program://CREMA/function/lavis.models.ulip_models.pointbert.dvae.get_loss#L313-L325","kind":"function","name":"get_loss","path":"lavis/models/ulip_models/pointbert/dvae.py","language":"python","start_line":313,"end_line":325,"context_start_line":293,"context_end_line":342,"code":" # self.loss_func_cdl1 = ChamferDistanceL1().cuda()\n # self.loss_func_cdl2 = ChamferDistanceL2().cuda()\n # self.loss_func_emd = emd().cuda()\n\n def recon_loss(self, ret, gt):\n whole_coarse, whole_fine, coarse, fine, group_gt, _ = ret\n\n bs, g, _, _ = coarse.shape\n\n coarse = coarse.reshape(bs * g, -1, 3).contiguous()\n fine = fine.reshape(bs * g, -1, 3).contiguous()\n group_gt = group_gt.reshape(bs * g, -1, 3).contiguous()\n\n loss_coarse_block = self.loss_func_cdl1(coarse, group_gt)\n loss_fine_block = self.loss_func_cdl1(fine, group_gt)\n\n loss_recon = loss_coarse_block + loss_fine_block\n\n return loss_recon\n\n def get_loss(self, ret, gt):\n # reconstruction loss\n loss_recon = self.recon_loss(ret, gt)\n # kl divergence\n logits = ret[-1] # B G N\n softmax = F.softmax(logits, dim=-1)\n mean_softmax = softmax.mean(dim=1)\n log_qy = torch.log(mean_softmax)\n log_uniform = torch.log(torch.tensor([1. / self.num_tokens], device=gt.device))\n loss_klv = F.kl_div(log_qy, log_uniform.expand(log_qy.size(0), log_qy.size(1)), None, None, 'batchmean',\n log_target=True)\n\n return loss_recon, loss_klv\n\n def forward(self, inp, temperature=1., hard=False, **kwargs):\n neighborhood, center = self.group_divider(inp)\n logits = self.encoder(neighborhood) # B G C\n logits = self.dgcnn_1(logits, center) # B G N\n soft_one_hot = F.gumbel_softmax(logits, tau=temperature, dim=2, hard=hard) # B G N\n sampled = torch.einsum('b g n, n c -> b g c', soft_one_hot, self.codebook) # B G C\n feature = self.dgcnn_2(sampled, center)\n coarse, fine = self.decoder(feature)\n\n with torch.no_grad():\n whole_fine = (fine + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n whole_coarse = (coarse + center.unsqueeze(2)).reshape(inp.size(0), -1, 3)\n\n assert fine.size(2) == self.group_size\n ret = (whole_coarse, whole_fine, coarse, fine, neighborhood, logits)\n return ret","source_hash":"298c39908c1fff73dfa55322b6bee11393e8b44881d657e008d7012527df0b13","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa","uri":"program://CREMA/module/lavis.models.img2prompt_models.img2prompt_vqa#L1-L582","kind":"module","name":"lavis.models.img2prompt_models.img2prompt_vqa","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":1,"end_line":582,"context_start_line":1,"context_end_line":582,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport random\n\nimport spacy\nimport torch\nimport torch.nn.functional as F\nfrom transformers import T5ForConditionalGeneration, T5Tokenizer\n\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.blip_models.blip_image_text_matching import compute_gradcam\n\nopen_pos = [\"NOUN\", \"VERB\", \"ADJ\", \"ADV\", \"NUM\"]\n\n\n\n@registry.register_model(\"img2prompt_vqa\")\nclass Img2PromptVQA(BaseModel):\n \"\"\"\n Img2Prompt_VQA model consists of three submodels for zero-shot VQA:\n 1. Image-questioning matching model\n 2. Image captioning model\n 3. Large Language model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"img2prompt_vqa\", \"base\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/img2prompt-vqa/img2prompt_vqa_base.yaml\",\n }\n\n def __init__(\n self,\n image_question_matching_model,\n image_captioning_model,\n question_generation_model,\n question_generation_tokenizer,\n offload_model=False,\n ):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_generation_model = question_generation_model\n self.question_generation_tokenizer = question_generation_tokenizer\n self.offload_model = offload_model\n self.nlp = spacy.load(\"en_core_web_sm\")\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples[\"image\"]\n question = [text.strip(\"?\") for text in samples[\"text_input\"]]\n tokenized_text = self.image_question_matching_model.tokenizer(\n question, padding=\"longest\", truncation=True, return_tensors=\"pt\"\n ).to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(\n model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num,\n )\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples[\"gradcams\"] = torch.stack(gradcams).reshape(\n samples[\"image\"].size(0), -1\n )\n\n return samples\n\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head=\"itm\"):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, self.prompt_length - 1 :]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == \"itm\":\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output # , mask, token_length\n\n elif match_head == \"itc\":\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = (\n torch.multinomial(\n samples[\"gradcams\"].to(self.image_captioning_model.device),\n num_patches,\n ).reshape(encoder_out.size(0), -1)\n + 1\n )\n patch_id = (\n patch_id.sort(dim=1)\n .values.unsqueeze(-1)\n .expand(-1, -1, encoder_out.size(2))\n )\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(\n stacked, start_dim=0, end_dim=1\n ) # (bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.image_captioning_model.device\n )\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(\n prompt, return_tensors=\"pt\"\n ).to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n itm_outputs = self.image_question_matching_model.itm_rank(\n image_embeds, image_atts, encoder_input_ids=decoder_out\n ) # caption filter\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(\n decoder_out, skip_special_tokens=True\n )\n\n for counter, output in enumerate(outputs):\n ind = counter // num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt) :]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n # print(itm_outputs)\n if (\n len(overlap_caption) == 0 and itm_outputs[counter] >= 0.5\n ): # image filter\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples[\"captions\"] = captions\n\n return samples\n\n def answer_extraction(self, caption, num_question_generation=30):\n cap_use = \"\"\n # print(caption)\n caption = caption\n ans_to_cap_dict = {}\n answers = []\n for cap_idx, cap in enumerate(caption):\n # print(cap)\n cap_use += cap\n cap = cap.strip().strip(\".\")\n # print(cap)\n cap = self.nlp(cap)\n for token in cap: # Noun /Verb/Adj//NUM\n if token.pos_ in open_pos:\n if token.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[token.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[token.text.lower()]:\n ans_to_cap_dict[token.text.lower()].append(cap_idx)\n answers.append(token.text)\n for ent in cap.ents:\n\n if ent.text not in answers:\n if ent.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[ent.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[ent.text.lower()]:\n ans_to_cap_dict[ent.text.lower()].append(cap_idx)\n answers.append(ent.text)\n for chunk in cap.noun_chunks:\n if len(chunk.text.split()) < 4:\n if chunk.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[chunk.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[chunk.text.lower()]:\n ans_to_cap_dict[chunk.text.lower()].append(cap_idx)\n # print(chunk.text)\n answers.append(chunk.text)\n answers = sorted(answers, key=answers.count, reverse=True)\n real_answers = []\n for i in answers:\n i = i + \".\"\n if i not in real_answers:\n real_answers.append(i)\n\n contexts_for_question_generation = []\n answers = []\n for ans in real_answers[\n :num_question_generation\n ]: # Generate questions for 30 answers with max frequencies.\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (ans, cap_use)\n )\n answers.append(ans)\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (\"yes.\", cap_use)\n )\n answers.append(\"yes.\")\n return contexts_for_question_generation, answers, ans_to_cap_dict\n\n def forward_qa_generation(self, samples):\n caption = samples[\"captions\"][0]\n (\n contexts_for_question_generation,\n answers,\n ans_to_cap_dict,\n ) = self.answer_extraction(caption)\n inputs = self.question_generation_tokenizer(\n contexts_for_question_generation,\n padding=\"longest\",\n truncation=True,\n max_length=2048,\n return_tensors=\"pt\",\n ).to(self.device)\n question_size = inputs.input_ids.shape[0]\n cur_b = 0\n true_input_size = 10\n outputs_list = []\n while cur_b < question_size:\n outputs = self.question_generation_model.generate(\n input_ids=inputs.input_ids[cur_b : cur_b + true_input_size],\n attention_mask=inputs.attention_mask[cur_b : cur_b + true_input_size],\n num_beams=3,\n max_length=30,\n )\n questions = self.question_generation_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n outputs_list += questions\n cur_b += true_input_size\n questions = outputs_list\n samples[\"questions\"] = questions\n samples[\"answers\"] = answers\n samples[\"ans_to_cap_dict\"] = ans_to_cap_dict\n # results.append({\"question_id\": ques_id, \"question\":questions,\"answer\":answers})\n return samples\n\n def create_context_prompt(self, samples, num_caps_per_img=30):\n ans_dict_queid = samples[\"ans_to_cap_dict\"]\n # print(ans_dict_queid)\n caption = samples[\"captions\"][0]\n answers = samples[\"answers\"]\n Context_Prompt = \"\"\n mycontexts_id = []\n for idx in range(num_caps_per_img):\n cap_id_list = ans_dict_queid.get(\n answers[(len(answers) - 1 - idx) % len(answers)][:-1].lower(), [0]\n )\n for cap_id in cap_id_list:\n if cap_id not in mycontexts_id:\n Context_Prompt += caption[cap_id]\n mycontexts_id.append(cap_id)\n break # We just take one cap for each answer\n samples[\"Context_Prompt\"] = Context_Prompt\n return Context_Prompt\n\n def create_task_prompt(\n self, samples, question_type=\"neural\", num_question_per_img=30\n ):\n syn_question_queid = samples[\"questions\"]\n syn_ans_queid = samples[\"answers\"]\n Task_Prompt = \"\"\n for idx in range(num_question_per_img):\n # if config['random_question']:\n # qa_idx = random.randint(0, len(syn_question_queid) - 1)\n # else:\n qa_idx = idx\n if (\n question_type != \"rule\" and num_question_per_img > 0 and idx < 1\n ): ## yes and no questions for vqav2\n # Task_Prompt += \"Question:\"\n # Task_Prompt += syn_question_queid_next[-1]\n # Task_Prompt += '\\n'\n # Task_Prompt += \"Answer:no\\n\"\n Task_Prompt += \"Question:\"\n Task_Prompt += syn_question_queid[-1]\n Task_Prompt += \"\\n\"\n Task_Prompt += \"Answer:\"\n Task_Prompt += \"yes\\n\"\n Task_Prompt += \"Question:Is this a toilet?\\n\"\n Task_Prompt += \"Answer:no\\n\"\n if \"question_type\" == \"rule\": # Rule-Based Question Generation\n Noun_Questions = [\n \"What item is this in this picture?\",\n \"What item is that in this picture?\",\n ]\n\n Verb_Questions = [\n \"What action is being done in this picture?\",\n \"Why is this item doing in this picture?\",\n \"Which action is being taken in this picture?\",\n \"What action is item doing in this picture?\",\n \"What action is item performing in this picture?\",\n ]\n\n Adj_Questions = [\n \"How to describe one item in this picture?\",\n \"What is item's ADJ TYPE in this picture?\",\n \"What is the ADJ TYPE in this picture?\",\n ]\n\n Task_Prompt += \"Question:\"\n doc = self.nlp(syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower())\n if doc[-1].pos_ == \"NOUN\":\n Task_Prompt += Noun_Questions[\n random.randint(0, len(Noun_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"VERB\":\n Task_Prompt += Verb_Questions[\n random.randint(0, len(Verb_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"ADJ\":\n Task_Prompt += Adj_Questions[\n random.randint(0, len(Adj_Questions) - 1)\n ]\n\n Task_Prompt += \"\\n\"\n\n Task_Prompt += \"Answer:\"\n Task_Prompt += syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower()\n Task_Prompt += \"\\n\"\n samples[\"Task_Prompt\"] = Task_Prompt\n # print(Task_Prompt)\n return Task_Prompt\n\n def prompts_construction(\n self,\n samples,\n question_type=\"neural\",\n num_caps_per_img=30,\n num_question_per_img=30,\n ):\n Prompt = \"Please reason the answer of the questions according to the given contexts.\\n\"\n\n Context_Prompt = self.create_context_prompt(samples, num_caps_per_img)\n\n Task_Prompt = self.create_task_prompt(\n samples, question_type, num_question_per_img\n )\n\n Img2Prompt = (\n Prompt\n + \"Contexts:\"\n + Context_Prompt\n + \"\\n\"\n + Task_Prompt\n + \"Question:\"\n + samples[\"text_input\"][0]\n + \"\\nAnswer:\"\n )\n return Img2Prompt\n\n def prepare_LLM_input(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=20,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of capti\n# ... truncated ...","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.Img2PromptVQA","uri":"program://CREMA/class/lavis.models.img2prompt_models.img2prompt_vqa.Img2PromptVQA#L25-L582","kind":"class","name":"Img2PromptVQA","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":25,"end_line":582,"context_start_line":5,"context_end_line":582,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport random\n\nimport spacy\nimport torch\nimport torch.nn.functional as F\nfrom transformers import T5ForConditionalGeneration, T5Tokenizer\n\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.blip_models.blip_image_text_matching import compute_gradcam\n\nopen_pos = [\"NOUN\", \"VERB\", \"ADJ\", \"ADV\", \"NUM\"]\n\n\n\n@registry.register_model(\"img2prompt_vqa\")\nclass Img2PromptVQA(BaseModel):\n \"\"\"\n Img2Prompt_VQA model consists of three submodels for zero-shot VQA:\n 1. Image-questioning matching model\n 2. Image captioning model\n 3. Large Language model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"img2prompt_vqa\", \"base\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/img2prompt-vqa/img2prompt_vqa_base.yaml\",\n }\n\n def __init__(\n self,\n image_question_matching_model,\n image_captioning_model,\n question_generation_model,\n question_generation_tokenizer,\n offload_model=False,\n ):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_generation_model = question_generation_model\n self.question_generation_tokenizer = question_generation_tokenizer\n self.offload_model = offload_model\n self.nlp = spacy.load(\"en_core_web_sm\")\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples[\"image\"]\n question = [text.strip(\"?\") for text in samples[\"text_input\"]]\n tokenized_text = self.image_question_matching_model.tokenizer(\n question, padding=\"longest\", truncation=True, return_tensors=\"pt\"\n ).to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(\n model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num,\n )\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples[\"gradcams\"] = torch.stack(gradcams).reshape(\n samples[\"image\"].size(0), -1\n )\n\n return samples\n\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head=\"itm\"):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, self.prompt_length - 1 :]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == \"itm\":\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output # , mask, token_length\n\n elif match_head == \"itc\":\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = (\n torch.multinomial(\n samples[\"gradcams\"].to(self.image_captioning_model.device),\n num_patches,\n ).reshape(encoder_out.size(0), -1)\n + 1\n )\n patch_id = (\n patch_id.sort(dim=1)\n .values.unsqueeze(-1)\n .expand(-1, -1, encoder_out.size(2))\n )\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(\n stacked, start_dim=0, end_dim=1\n ) # (bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.image_captioning_model.device\n )\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(\n prompt, return_tensors=\"pt\"\n ).to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n itm_outputs = self.image_question_matching_model.itm_rank(\n image_embeds, image_atts, encoder_input_ids=decoder_out\n ) # caption filter\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(\n decoder_out, skip_special_tokens=True\n )\n\n for counter, output in enumerate(outputs):\n ind = counter // num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt) :]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n # print(itm_outputs)\n if (\n len(overlap_caption) == 0 and itm_outputs[counter] >= 0.5\n ): # image filter\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples[\"captions\"] = captions\n\n return samples\n\n def answer_extraction(self, caption, num_question_generation=30):\n cap_use = \"\"\n # print(caption)\n caption = caption\n ans_to_cap_dict = {}\n answers = []\n for cap_idx, cap in enumerate(caption):\n # print(cap)\n cap_use += cap\n cap = cap.strip().strip(\".\")\n # print(cap)\n cap = self.nlp(cap)\n for token in cap: # Noun /Verb/Adj//NUM\n if token.pos_ in open_pos:\n if token.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[token.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[token.text.lower()]:\n ans_to_cap_dict[token.text.lower()].append(cap_idx)\n answers.append(token.text)\n for ent in cap.ents:\n\n if ent.text not in answers:\n if ent.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[ent.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[ent.text.lower()]:\n ans_to_cap_dict[ent.text.lower()].append(cap_idx)\n answers.append(ent.text)\n for chunk in cap.noun_chunks:\n if len(chunk.text.split()) < 4:\n if chunk.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[chunk.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[chunk.text.lower()]:\n ans_to_cap_dict[chunk.text.lower()].append(cap_idx)\n # print(chunk.text)\n answers.append(chunk.text)\n answers = sorted(answers, key=answers.count, reverse=True)\n real_answers = []\n for i in answers:\n i = i + \".\"\n if i not in real_answers:\n real_answers.append(i)\n\n contexts_for_question_generation = []\n answers = []\n for ans in real_answers[\n :num_question_generation\n ]: # Generate questions for 30 answers with max frequencies.\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (ans, cap_use)\n )\n answers.append(ans)\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (\"yes.\", cap_use)\n )\n answers.append(\"yes.\")\n return contexts_for_question_generation, answers, ans_to_cap_dict\n\n def forward_qa_generation(self, samples):\n caption = samples[\"captions\"][0]\n (\n contexts_for_question_generation,\n answers,\n ans_to_cap_dict,\n ) = self.answer_extraction(caption)\n inputs = self.question_generation_tokenizer(\n contexts_for_question_generation,\n padding=\"longest\",\n truncation=True,\n max_length=2048,\n return_tensors=\"pt\",\n ).to(self.device)\n question_size = inputs.input_ids.shape[0]\n cur_b = 0\n true_input_size = 10\n outputs_list = []\n while cur_b < question_size:\n outputs = self.question_generation_model.generate(\n input_ids=inputs.input_ids[cur_b : cur_b + true_input_size],\n attention_mask=inputs.attention_mask[cur_b : cur_b + true_input_size],\n num_beams=3,\n max_length=30,\n )\n questions = self.question_generation_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n outputs_list += questions\n cur_b += true_input_size\n questions = outputs_list\n samples[\"questions\"] = questions\n samples[\"answers\"] = answers\n samples[\"ans_to_cap_dict\"] = ans_to_cap_dict\n # results.append({\"question_id\": ques_id, \"question\":questions,\"answer\":answers})\n return samples\n\n def create_context_prompt(self, samples, num_caps_per_img=30):\n ans_dict_queid = samples[\"ans_to_cap_dict\"]\n # print(ans_dict_queid)\n caption = samples[\"captions\"][0]\n answers = samples[\"answers\"]\n Context_Prompt = \"\"\n mycontexts_id = []\n for idx in range(num_caps_per_img):\n cap_id_list = ans_dict_queid.get(\n answers[(len(answers) - 1 - idx) % len(answers)][:-1].lower(), [0]\n )\n for cap_id in cap_id_list:\n if cap_id not in mycontexts_id:\n Context_Prompt += caption[cap_id]\n mycontexts_id.append(cap_id)\n break # We just take one cap for each answer\n samples[\"Context_Prompt\"] = Context_Prompt\n return Context_Prompt\n\n def create_task_prompt(\n self, samples, question_type=\"neural\", num_question_per_img=30\n ):\n syn_question_queid = samples[\"questions\"]\n syn_ans_queid = samples[\"answers\"]\n Task_Prompt = \"\"\n for idx in range(num_question_per_img):\n # if config['random_question']:\n # qa_idx = random.randint(0, len(syn_question_queid) - 1)\n # else:\n qa_idx = idx\n if (\n question_type != \"rule\" and num_question_per_img > 0 and idx < 1\n ): ## yes and no questions for vqav2\n # Task_Prompt += \"Question:\"\n # Task_Prompt += syn_question_queid_next[-1]\n # Task_Prompt += '\\n'\n # Task_Prompt += \"Answer:no\\n\"\n Task_Prompt += \"Question:\"\n Task_Prompt += syn_question_queid[-1]\n Task_Prompt += \"\\n\"\n Task_Prompt += \"Answer:\"\n Task_Prompt += \"yes\\n\"\n Task_Prompt += \"Question:Is this a toilet?\\n\"\n Task_Prompt += \"Answer:no\\n\"\n if \"question_type\" == \"rule\": # Rule-Based Question Generation\n Noun_Questions = [\n \"What item is this in this picture?\",\n \"What item is that in this picture?\",\n ]\n\n Verb_Questions = [\n \"What action is being done in this picture?\",\n \"Why is this item doing in this picture?\",\n \"Which action is being taken in this picture?\",\n \"What action is item doing in this picture?\",\n \"What action is item performing in this picture?\",\n ]\n\n Adj_Questions = [\n \"How to describe one item in this picture?\",\n \"What is item's ADJ TYPE in this picture?\",\n \"What is the ADJ TYPE in this picture?\",\n ]\n\n Task_Prompt += \"Question:\"\n doc = self.nlp(syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower())\n if doc[-1].pos_ == \"NOUN\":\n Task_Prompt += Noun_Questions[\n random.randint(0, len(Noun_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"VERB\":\n Task_Prompt += Verb_Questions[\n random.randint(0, len(Verb_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"ADJ\":\n Task_Prompt += Adj_Questions[\n random.randint(0, len(Adj_Questions) - 1)\n ]\n\n Task_Prompt += \"\\n\"\n\n Task_Prompt += \"Answer:\"\n Task_Prompt += syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower()\n Task_Prompt += \"\\n\"\n samples[\"Task_Prompt\"] = Task_Prompt\n # print(Task_Prompt)\n return Task_Prompt\n\n def prompts_construction(\n self,\n samples,\n question_type=\"neural\",\n num_caps_per_img=30,\n num_question_per_img=30,\n ):\n Prompt = \"Please reason the answer of the questions according to the given contexts.\\n\"\n\n Context_Prompt = self.create_context_prompt(samples, num_caps_per_img)\n\n Task_Prompt = self.create_task_prompt(\n samples, question_type, num_question_per_img\n )\n\n Img2Prompt = (\n Prompt\n + \"Contexts:\"\n + Context_Prompt\n + \"\\n\"\n + Task_Prompt\n + \"Question:\"\n + samples[\"text_input\"][0]\n + \"\\nAnswer:\"\n )\n return Img2Prompt\n\n def prepare_LLM_input(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=20,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n cap_max_length (int): The maximum length\n# ... truncated ...","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.__init__","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.__init__#L46-L61","kind":"function","name":"__init__","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":46,"end_line":61,"context_start_line":26,"context_end_line":81,"code":" \"\"\"\n Img2Prompt_VQA model consists of three submodels for zero-shot VQA:\n 1. Image-questioning matching model\n 2. Image captioning model\n 3. Large Language model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"img2prompt_vqa\", \"base\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/img2prompt-vqa/img2prompt_vqa_base.yaml\",\n }\n\n def __init__(\n self,\n image_question_matching_model,\n image_captioning_model,\n question_generation_model,\n question_generation_tokenizer,\n offload_model=False,\n ):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_generation_model = question_generation_model\n self.question_generation_tokenizer = question_generation_tokenizer\n self.offload_model = offload_model\n self.nlp = spacy.load(\"en_core_web_sm\")\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples[\"image\"]\n question = [text.strip(\"?\") for text in samples[\"text_input\"]]\n tokenized_text = self.image_question_matching_model.tokenizer(\n question, padding=\"longest\", truncation=True, return_tensors=\"pt\"\n ).to(self.image_question_matching_model.device)","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.forward_itm","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.forward_itm#L63-L96","kind":"function","name":"forward_itm","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":63,"end_line":96,"context_start_line":43,"context_end_line":116,"code":" \"base\": \"configs/models/img2prompt-vqa/img2prompt_vqa_base.yaml\",\n }\n\n def __init__(\n self,\n image_question_matching_model,\n image_captioning_model,\n question_generation_model,\n question_generation_tokenizer,\n offload_model=False,\n ):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_generation_model = question_generation_model\n self.question_generation_tokenizer = question_generation_tokenizer\n self.offload_model = offload_model\n self.nlp = spacy.load(\"en_core_web_sm\")\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples[\"image\"]\n question = [text.strip(\"?\") for text in samples[\"text_input\"]]\n tokenized_text = self.image_question_matching_model.tokenizer(\n question, padding=\"longest\", truncation=True, return_tensors=\"pt\"\n ).to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(\n model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num,\n )\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples[\"gradcams\"] = torch.stack(gradcams).reshape(\n samples[\"image\"].size(0), -1\n )\n\n return samples\n\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head=\"itm\"):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, self.prompt_length - 1 :]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == \"itm\":\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output # , mask, token_length\n","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.itm_rank","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.itm_rank#L98-L131","kind":"function","name":"itm_rank","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":98,"end_line":131,"context_start_line":78,"context_end_line":151,"code":" question = [text.strip(\"?\") for text in samples[\"text_input\"]]\n tokenized_text = self.image_question_matching_model.tokenizer(\n question, padding=\"longest\", truncation=True, return_tensors=\"pt\"\n ).to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(\n model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num,\n )\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples[\"gradcams\"] = torch.stack(gradcams).reshape(\n samples[\"image\"].size(0), -1\n )\n\n return samples\n\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head=\"itm\"):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, self.prompt_length - 1 :]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == \"itm\":\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output # , mask, token_length\n\n elif match_head == \"itc\":\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.forward_cap","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.forward_cap#L133-L245","kind":"function","name":"forward_cap","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":133,"end_line":245,"context_start_line":113,"context_end_line":265,"code":" )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output # , mask, token_length\n\n elif match_head == \"itc\":\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text_attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = (\n torch.multinomial(\n samples[\"gradcams\"].to(self.image_captioning_model.device),\n num_patches,\n ).reshape(encoder_out.size(0), -1)\n + 1\n )\n patch_id = (\n patch_id.sort(dim=1)\n .values.unsqueeze(-1)\n .expand(-1, -1, encoder_out.size(2))\n )\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(\n stacked, start_dim=0, end_dim=1\n ) # (bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.image_captioning_model.device\n )\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(\n prompt, return_tensors=\"pt\"\n ).to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs\n )\n\n itm_outputs = self.image_question_matching_model.itm_rank(\n image_embeds, image_atts, encoder_input_ids=decoder_out\n ) # caption filter\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(\n decoder_out, skip_special_tokens=True\n )\n\n for counter, output in enumerate(outputs):\n ind = counter // num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt) :]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n # print(itm_outputs)\n if (\n len(overlap_caption) == 0 and itm_outputs[counter] >= 0.5\n ): # image filter\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples[\"captions\"] = captions\n\n return samples\n\n def answer_extraction(self, caption, num_question_generation=30):\n cap_use = \"\"\n # print(caption)\n caption = caption\n ans_to_cap_dict = {}\n answers = []\n for cap_idx, cap in enumerate(caption):\n # print(cap)\n cap_use += cap\n cap = cap.strip().strip(\".\")\n # print(cap)\n cap = self.nlp(cap)\n for token in cap: # Noun /Verb/Adj//NUM\n if token.pos_ in open_pos:\n if token.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[token.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[token.text.lower()]:\n ans_to_cap_dict[token.text.lower()].append(cap_idx)","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.answer_extraction","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.answer_extraction#L247-L305","kind":"function","name":"answer_extraction","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":247,"end_line":305,"context_start_line":227,"context_end_line":325,"code":" decoder_out, skip_special_tokens=True\n )\n\n for counter, output in enumerate(outputs):\n ind = counter // num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt) :]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n # print(itm_outputs)\n if (\n len(overlap_caption) == 0 and itm_outputs[counter] >= 0.5\n ): # image filter\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples[\"captions\"] = captions\n\n return samples\n\n def answer_extraction(self, caption, num_question_generation=30):\n cap_use = \"\"\n # print(caption)\n caption = caption\n ans_to_cap_dict = {}\n answers = []\n for cap_idx, cap in enumerate(caption):\n # print(cap)\n cap_use += cap\n cap = cap.strip().strip(\".\")\n # print(cap)\n cap = self.nlp(cap)\n for token in cap: # Noun /Verb/Adj//NUM\n if token.pos_ in open_pos:\n if token.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[token.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[token.text.lower()]:\n ans_to_cap_dict[token.text.lower()].append(cap_idx)\n answers.append(token.text)\n for ent in cap.ents:\n\n if ent.text not in answers:\n if ent.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[ent.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[ent.text.lower()]:\n ans_to_cap_dict[ent.text.lower()].append(cap_idx)\n answers.append(ent.text)\n for chunk in cap.noun_chunks:\n if len(chunk.text.split()) < 4:\n if chunk.text.lower() not in ans_to_cap_dict:\n ans_to_cap_dict[chunk.text.lower()] = [cap_idx]\n else:\n if cap_idx not in ans_to_cap_dict[chunk.text.lower()]:\n ans_to_cap_dict[chunk.text.lower()].append(cap_idx)\n # print(chunk.text)\n answers.append(chunk.text)\n answers = sorted(answers, key=answers.count, reverse=True)\n real_answers = []\n for i in answers:\n i = i + \".\"\n if i not in real_answers:\n real_answers.append(i)\n\n contexts_for_question_generation = []\n answers = []\n for ans in real_answers[\n :num_question_generation\n ]: # Generate questions for 30 answers with max frequencies.\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (ans, cap_use)\n )\n answers.append(ans)\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (\"yes.\", cap_use)\n )\n answers.append(\"yes.\")\n return contexts_for_question_generation, answers, ans_to_cap_dict\n\n def forward_qa_generation(self, samples):\n caption = samples[\"captions\"][0]\n (\n contexts_for_question_generation,\n answers,\n ans_to_cap_dict,\n ) = self.answer_extraction(caption)\n inputs = self.question_generation_tokenizer(\n contexts_for_question_generation,\n padding=\"longest\",\n truncation=True,\n max_length=2048,\n return_tensors=\"pt\",\n ).to(self.device)\n question_size = inputs.input_ids.shape[0]\n cur_b = 0\n true_input_size = 10\n outputs_list = []\n while cur_b < question_size:","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.forward_qa_generation","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.forward_qa_generation#L307-L342","kind":"function","name":"forward_qa_generation","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":307,"end_line":342,"context_start_line":287,"context_end_line":362,"code":" for i in answers:\n i = i + \".\"\n if i not in real_answers:\n real_answers.append(i)\n\n contexts_for_question_generation = []\n answers = []\n for ans in real_answers[\n :num_question_generation\n ]: # Generate questions for 30 answers with max frequencies.\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (ans, cap_use)\n )\n answers.append(ans)\n contexts_for_question_generation.append(\n \"answer: %s context: %s.\" % (\"yes.\", cap_use)\n )\n answers.append(\"yes.\")\n return contexts_for_question_generation, answers, ans_to_cap_dict\n\n def forward_qa_generation(self, samples):\n caption = samples[\"captions\"][0]\n (\n contexts_for_question_generation,\n answers,\n ans_to_cap_dict,\n ) = self.answer_extraction(caption)\n inputs = self.question_generation_tokenizer(\n contexts_for_question_generation,\n padding=\"longest\",\n truncation=True,\n max_length=2048,\n return_tensors=\"pt\",\n ).to(self.device)\n question_size = inputs.input_ids.shape[0]\n cur_b = 0\n true_input_size = 10\n outputs_list = []\n while cur_b < question_size:\n outputs = self.question_generation_model.generate(\n input_ids=inputs.input_ids[cur_b : cur_b + true_input_size],\n attention_mask=inputs.attention_mask[cur_b : cur_b + true_input_size],\n num_beams=3,\n max_length=30,\n )\n questions = self.question_generation_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n outputs_list += questions\n cur_b += true_input_size\n questions = outputs_list\n samples[\"questions\"] = questions\n samples[\"answers\"] = answers\n samples[\"ans_to_cap_dict\"] = ans_to_cap_dict\n # results.append({\"question_id\": ques_id, \"question\":questions,\"answer\":answers})\n return samples\n\n def create_context_prompt(self, samples, num_caps_per_img=30):\n ans_dict_queid = samples[\"ans_to_cap_dict\"]\n # print(ans_dict_queid)\n caption = samples[\"captions\"][0]\n answers = samples[\"answers\"]\n Context_Prompt = \"\"\n mycontexts_id = []\n for idx in range(num_caps_per_img):\n cap_id_list = ans_dict_queid.get(\n answers[(len(answers) - 1 - idx) % len(answers)][:-1].lower(), [0]\n )\n for cap_id in cap_id_list:\n if cap_id not in mycontexts_id:\n Context_Prompt += caption[cap_id]\n mycontexts_id.append(cap_id)\n break # We just take one cap for each answer\n samples[\"Context_Prompt\"] = Context_Prompt\n return Context_Prompt\n","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.create_context_prompt","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.create_context_prompt#L344-L361","kind":"function","name":"create_context_prompt","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":344,"end_line":361,"context_start_line":324,"context_end_line":381,"code":" outputs_list = []\n while cur_b < question_size:\n outputs = self.question_generation_model.generate(\n input_ids=inputs.input_ids[cur_b : cur_b + true_input_size],\n attention_mask=inputs.attention_mask[cur_b : cur_b + true_input_size],\n num_beams=3,\n max_length=30,\n )\n questions = self.question_generation_tokenizer.batch_decode(\n outputs, skip_special_tokens=True\n )\n outputs_list += questions\n cur_b += true_input_size\n questions = outputs_list\n samples[\"questions\"] = questions\n samples[\"answers\"] = answers\n samples[\"ans_to_cap_dict\"] = ans_to_cap_dict\n # results.append({\"question_id\": ques_id, \"question\":questions,\"answer\":answers})\n return samples\n\n def create_context_prompt(self, samples, num_caps_per_img=30):\n ans_dict_queid = samples[\"ans_to_cap_dict\"]\n # print(ans_dict_queid)\n caption = samples[\"captions\"][0]\n answers = samples[\"answers\"]\n Context_Prompt = \"\"\n mycontexts_id = []\n for idx in range(num_caps_per_img):\n cap_id_list = ans_dict_queid.get(\n answers[(len(answers) - 1 - idx) % len(answers)][:-1].lower(), [0]\n )\n for cap_id in cap_id_list:\n if cap_id not in mycontexts_id:\n Context_Prompt += caption[cap_id]\n mycontexts_id.append(cap_id)\n break # We just take one cap for each answer\n samples[\"Context_Prompt\"] = Context_Prompt\n return Context_Prompt\n\n def create_task_prompt(\n self, samples, question_type=\"neural\", num_question_per_img=30\n ):\n syn_question_queid = samples[\"questions\"]\n syn_ans_queid = samples[\"answers\"]\n Task_Prompt = \"\"\n for idx in range(num_question_per_img):\n # if config['random_question']:\n # qa_idx = random.randint(0, len(syn_question_queid) - 1)\n # else:\n qa_idx = idx\n if (\n question_type != \"rule\" and num_question_per_img > 0 and idx < 1\n ): ## yes and no questions for vqav2\n # Task_Prompt += \"Question:\"\n # Task_Prompt += syn_question_queid_next[-1]\n # Task_Prompt += '\\n'\n # Task_Prompt += \"Answer:no\\n\"\n Task_Prompt += \"Question:\"","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.create_task_prompt","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.create_task_prompt#L363-L430","kind":"function","name":"create_task_prompt","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":363,"end_line":430,"context_start_line":343,"context_end_line":450,"code":"\n def create_context_prompt(self, samples, num_caps_per_img=30):\n ans_dict_queid = samples[\"ans_to_cap_dict\"]\n # print(ans_dict_queid)\n caption = samples[\"captions\"][0]\n answers = samples[\"answers\"]\n Context_Prompt = \"\"\n mycontexts_id = []\n for idx in range(num_caps_per_img):\n cap_id_list = ans_dict_queid.get(\n answers[(len(answers) - 1 - idx) % len(answers)][:-1].lower(), [0]\n )\n for cap_id in cap_id_list:\n if cap_id not in mycontexts_id:\n Context_Prompt += caption[cap_id]\n mycontexts_id.append(cap_id)\n break # We just take one cap for each answer\n samples[\"Context_Prompt\"] = Context_Prompt\n return Context_Prompt\n\n def create_task_prompt(\n self, samples, question_type=\"neural\", num_question_per_img=30\n ):\n syn_question_queid = samples[\"questions\"]\n syn_ans_queid = samples[\"answers\"]\n Task_Prompt = \"\"\n for idx in range(num_question_per_img):\n # if config['random_question']:\n # qa_idx = random.randint(0, len(syn_question_queid) - 1)\n # else:\n qa_idx = idx\n if (\n question_type != \"rule\" and num_question_per_img > 0 and idx < 1\n ): ## yes and no questions for vqav2\n # Task_Prompt += \"Question:\"\n # Task_Prompt += syn_question_queid_next[-1]\n # Task_Prompt += '\\n'\n # Task_Prompt += \"Answer:no\\n\"\n Task_Prompt += \"Question:\"\n Task_Prompt += syn_question_queid[-1]\n Task_Prompt += \"\\n\"\n Task_Prompt += \"Answer:\"\n Task_Prompt += \"yes\\n\"\n Task_Prompt += \"Question:Is this a toilet?\\n\"\n Task_Prompt += \"Answer:no\\n\"\n if \"question_type\" == \"rule\": # Rule-Based Question Generation\n Noun_Questions = [\n \"What item is this in this picture?\",\n \"What item is that in this picture?\",\n ]\n\n Verb_Questions = [\n \"What action is being done in this picture?\",\n \"Why is this item doing in this picture?\",\n \"Which action is being taken in this picture?\",\n \"What action is item doing in this picture?\",\n \"What action is item performing in this picture?\",\n ]\n\n Adj_Questions = [\n \"How to describe one item in this picture?\",\n \"What is item's ADJ TYPE in this picture?\",\n \"What is the ADJ TYPE in this picture?\",\n ]\n\n Task_Prompt += \"Question:\"\n doc = self.nlp(syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower())\n if doc[-1].pos_ == \"NOUN\":\n Task_Prompt += Noun_Questions[\n random.randint(0, len(Noun_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"VERB\":\n Task_Prompt += Verb_Questions[\n random.randint(0, len(Verb_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"ADJ\":\n Task_Prompt += Adj_Questions[\n random.randint(0, len(Adj_Questions) - 1)\n ]\n\n Task_Prompt += \"\\n\"\n\n Task_Prompt += \"Answer:\"\n Task_Prompt += syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower()\n Task_Prompt += \"\\n\"\n samples[\"Task_Prompt\"] = Task_Prompt\n # print(Task_Prompt)\n return Task_Prompt\n\n def prompts_construction(\n self,\n samples,\n question_type=\"neural\",\n num_caps_per_img=30,\n num_question_per_img=30,\n ):\n Prompt = \"Please reason the answer of the questions according to the given contexts.\\n\"\n\n Context_Prompt = self.create_context_prompt(samples, num_caps_per_img)\n\n Task_Prompt = self.create_task_prompt(\n samples, question_type, num_question_per_img\n )\n\n Img2Prompt = (\n Prompt\n + \"Contexts:\"\n + Context_Prompt","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.prompts_construction","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.prompts_construction#L432-L457","kind":"function","name":"prompts_construction","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":432,"end_line":457,"context_start_line":412,"context_end_line":477,"code":" random.randint(0, len(Noun_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"VERB\":\n Task_Prompt += Verb_Questions[\n random.randint(0, len(Verb_Questions) - 1)\n ]\n elif doc[-1].pos_ == \"ADJ\":\n Task_Prompt += Adj_Questions[\n random.randint(0, len(Adj_Questions) - 1)\n ]\n\n Task_Prompt += \"\\n\"\n\n Task_Prompt += \"Answer:\"\n Task_Prompt += syn_ans_queid[(qa_idx) % len(syn_ans_queid)][:-1].lower()\n Task_Prompt += \"\\n\"\n samples[\"Task_Prompt\"] = Task_Prompt\n # print(Task_Prompt)\n return Task_Prompt\n\n def prompts_construction(\n self,\n samples,\n question_type=\"neural\",\n num_caps_per_img=30,\n num_question_per_img=30,\n ):\n Prompt = \"Please reason the answer of the questions according to the given contexts.\\n\"\n\n Context_Prompt = self.create_context_prompt(samples, num_caps_per_img)\n\n Task_Prompt = self.create_task_prompt(\n samples, question_type, num_question_per_img\n )\n\n Img2Prompt = (\n Prompt\n + \"Contexts:\"\n + Context_Prompt\n + \"\\n\"\n + Task_Prompt\n + \"Question:\"\n + samples[\"text_input\"][0]\n + \"\\nAnswer:\"\n )\n return Img2Prompt\n\n def prepare_LLM_input(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=20,\n block_num=7,\n ):\n \"\"\"","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.prepare_LLM_input","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.prepare_LLM_input#L459-L547","kind":"function","name":"prepare_LLM_input","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":459,"end_line":547,"context_start_line":439,"context_end_line":567,"code":" Prompt = \"Please reason the answer of the questions according to the given contexts.\\n\"\n\n Context_Prompt = self.create_context_prompt(samples, num_caps_per_img)\n\n Task_Prompt = self.create_task_prompt(\n samples, question_type, num_question_per_img\n )\n\n Img2Prompt = (\n Prompt\n + \"Contexts:\"\n + Context_Prompt\n + \"\\n\"\n + Task_Prompt\n + \"Question:\"\n + samples[\"text_input\"][0]\n + \"\\nAnswer:\"\n )\n return Img2Prompt\n\n def prepare_LLM_input(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=20,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_patches (int): Number of patches sampled for each image.\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n List: A list of strings, each string is an answer.\n gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n assert inference_method in [\n \"generate\",\n ], \"Inference method must be 'generate', got {}.\".format(inference_method)\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n samples = self.forward_itm(samples, block_num=block_num)\n\n samples = self.forward_cap(\n samples,\n cap_max_length=cap_max_length,\n cap_min_length=cap_min_length,\n top_k=top_k,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n num_captions=num_captions,\n num_patches=num_patches,\n )\n\n if self.offload_model:\n samples[\"image\"] = samples[\"image\"].to(\"cpu\")\n self.image_question_matching_model.to(\"cpu\")\n self.image_captioning_model.to(\"cpu\")\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(\n samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid,\n )\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples[\"captions\"], samples[\"gradcams\"]\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n\n question_generation_tokenizer = T5Tokenizer.from_pretrained(\n \"google/t5-large-lm-adapt\"\n )\n question_generation_model = T5ForConditionalGeneration.from_pretrained(\n \"google/t5-large-lm-adapt\"\n )\n cached_file = download_cached_file(\n \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/projects/img2prompt/T5_large_QG.pth\",","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.img2prompt_models.img2prompt_vqa.from_config","uri":"program://CREMA/function/lavis.models.img2prompt_models.img2prompt_vqa.from_config#L550-L582","kind":"function","name":"from_config","path":"lavis/models/img2prompt_models/img2prompt_vqa.py","language":"python","start_line":550,"end_line":582,"context_start_line":530,"context_end_line":582,"code":" self.image_captioning_model.to(\"cpu\")\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(\n samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid,\n )\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples[\"captions\"], samples[\"gradcams\"]\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n\n question_generation_tokenizer = T5Tokenizer.from_pretrained(\n \"google/t5-large-lm-adapt\"\n )\n question_generation_model = T5ForConditionalGeneration.from_pretrained(\n \"google/t5-large-lm-adapt\"\n )\n cached_file = download_cached_file(\n \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/projects/img2prompt/T5_large_QG.pth\",\n check_hash=False,\n progress=True,\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n state_dict = checkpoint[\"model\"]\n question_generation_model.load_state_dict(state_dict)\n model = cls(\n image_question_matching_model=image_question_matching_model,\n image_captioning_model=image_captioning_model,\n question_generation_model=question_generation_model,\n question_generation_tokenizer=question_generation_tokenizer,\n offload_model=False,\n )\n\n return model","source_hash":"48e69c13a391601284408a035424a5da8e54486e666bd34a37ae2abe019b4b02","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema","uri":"program://CREMA/module/lavis.models.crema_models.crema#L1-L772","kind":"module","name":"lavis.models.crema_models.crema","path":"lavis/models/crema_models/crema.py","language":"python","start_line":1,"end_line":772,"context_start_line":1,"context_end_line":772,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport copy\nimport torch\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom transformers import T5TokenizerFast, BertTokenizer\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.Qformer import BertConfig, BertLMHeadModel\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration\nfrom positional_encodings.torch_encodings import PositionalEncoding1D\n\ndef print_trainable_parameters(model):\n trainable_params = 0\n all_param = 0\n for _, param in model.named_parameters():\n all_param += param.numel()\n if param.requires_grad:\n trainable_params += param.numel()\n print(\n f\"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.2f}\"\n )\n\n@registry.register_model(\"crema\")\nclass CREMA(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__( self, img_size=224, drop_path_rate=0,\n use_grad_checkpoint=False, vit_precision=\"fp16\", freeze_vit=True,\n num_query_token=32, t5_model=\"google/flan-t5-xl\", prompt=\"\",\n max_txt_len=32, frame_num=8, answer_num=5, apply_lemmatizer=False, \n task='concate',\n modalities='rgb',\n downstream_task='mcqa', # caption / oeqa / mcqa\n lora_rank=64,\n lora_layer=None,\n lora_dropout=0.1,\n fuse_with_base_modality=False):\n\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n \n self.task = task #.split('_')\n self.modalities = modalities.split('_')\n self.fuse_with_base_modality=fuse_with_base_modality\n\n print(self.modalities)\n num_features = 1408\n # ========= init vision encoder ============\n # init vision backbone for vision experts\n if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.visual_encoder = self.init_vision_encoder_only(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision)\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n if 'audio' in self.modalities:\n self.audio_encoder, self.ln_audio = self.init_audio_encoder('beats', cached_audio=False)\n for name, param in self.audio_encoder.named_parameters():\n param.requires_grad = False \n self.audio_encoder = self.audio_encoder.eval()\n self.audio_encoder.train = disabled_train\n logging.info(\"freeze audio encoder\")\n\n if 'pc' in self.modalities:\n # pre-extracted features\n pass\n \n # print('num_features', self.visual_encoder.num_features) 1408\n # ========= init LLM ============ \n # text backbone\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(t5_model, config=t5_config)\n # freeze T5\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16() \n\n # ========= init Qformer ============\n # if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.Qformer, encoder_config = self.init_Multimodal_Qformer(\n num_query_token, num_features, #self.visual_encoder.num_features,\n modulars=self.modalities, \n r=lora_rank, lora_layer=lora_layer, lora_dropout=lora_dropout)\n\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n self.num_query_token = num_query_token\n\n if 'rgb' in self.modalities:\n self.query_tokens_rgb = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_rgb.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_rgb = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_rgb = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'flow' in self.modalities:\n self.query_tokens_flow = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_flow.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_flow = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_flow = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'norm' in self.modalities:\n self.query_tokens_norm = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_norm.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_norm = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_norm = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'depth' in self.modalities:\n self.query_tokens_depth = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_depth.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_depth = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_depth = nn.LayerNorm(self.visual_encoder.num_features)\n \n if 'audio' in self.modalities:\n self.query_tokens_audio = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_audio.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_audio = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n \n self.projection_audio = nn.Linear(self.audio_encoder.num_features, num_features)\n self.ln_audio = nn.LayerNorm(num_features)\n \n if 'pc' in self.modalities:\n self.query_tokens_pc = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_pc.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_pc = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_pc = nn.LayerNorm(num_features)\n pos_model = PositionalEncoding1D(1408 // 3)\n x = torch.zeros(1, 256, 1408 // 3)\n self.pos_embedding = pos_model(x).squeeze().cuda()\n \n if 'espresso' in self.task:\n\n self.fusion = nn.Sequential(\n nn.Linear(2048*(len(self.modalities)-1), 2048),\n )\n self.sigmoid = nn.Sigmoid()\n \n self.downstream_task = downstream_task \n self.max_txt_len = 77\n answer_id = [71, 272, 205, 309, 262] # A B C D E\n self.answer_id = answer_id[:answer_num]\n \n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n self.frame_num = frame_num\n self.ANS_MAP = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n if frame_num == 1:\n self.vid_prefix = ['Frame: ']\n self.depth_prefix = ['Depth Map: ']\n self.flow_prefix = ['Optical Flow: ']\n self.norm_prefix = ['Surface Normalization: ']\n else:\n self.vid_prefix = ['Frame {}: '.format(str(i+1)) for i in range(frame_num)]\n self.depth_prefix = ['Depth Map {}: '.format(str(i+1)) for i in range(frame_num)]\n self.flow_prefix = ['Optical Flow {}: '.format(str(i+1)) for i in range(frame_num)]\n self.norm_prefix = ['Surface Normalization {}: '.format(str(i+1)) for i in range(frame_num)]\n\n self.audio_prefix = ['Audio: ']\n self.pc_prefix = ['3D Model: ']\n \n def forward(self, samples):\n\n # rgb visual embedding\n qa_text, answer = samples['qa_input'], samples['qa_output']\n b = len(qa_text)\n\n input_embed_dict, input_att_dict = {}, {}\n\n for modal in self.modalities:\n input = samples[modal]\n if modal in ['rgb', 'depth', 'norm', 'flow']:\n # fix some loading issue\n if input.shape[1] == 3:\n input = input.permute(0, 2, 1, 3, 4)\n \n # following 3D-LLM \n if modal == 'pc':\n with torch.cuda.amp.autocast(dtype=torch.float32):\n pc_embeds = samples[\"pc_feat\"]\n pc = samples[\"pc\"].long()\n all_pcs = torch.zeros((pc_embeds.shape))\n for j in range(pc.shape[0]):\n pcs = []\n for i in range(3):\n pc_i = pc[j][:, i]\n pcs.append(self.pos_embedding[pc_i])\n pcs = torch.cat(pcs, -1)\n all_pcs[j][:, :1407] = pcs\n all_pcs = all_pcs.cuda()\n pc_embeds = pc_embeds + 0.01 * all_pcs\n atts = torch.ones(pc_embeds.size()[:-1], dtype=torch.long).to(pc_embeds.device)\n input_embed_dict[modal], input_att_dict[modal] = pc_embeds, atts\n else:\n input_embed_dict[modal], input_att_dict[modal] = self.encode_input(input, modal)\n \n device = input_embed_dict[list(input_embed_dict.keys())[0]].device\n\n input_text= self.t5_tokenizer(\n qa_text, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n input_text_embeds = self.t5_model.encoder.embed_tokens(input_text.input_ids) \n\n\n fusion_modal = []\n t5_inputs, t5_atts, t5_query = {}, {}, {}\n for modal in self.modalities:\n t5_inputs[modal], t5_atts[modal], t5_query[modal] = self.get_qformer_embedding(input_embed_dict[modal], input_att_dict[modal], device, modal, b)\n\n if 'rgb' in self.modalities:\n inputs_t5_rgb = t5_inputs['rgb']\n atts_t5_rgb = t5_atts['rgb']\n vid_prefix_embed, vid_prefix_mask = self.get_prefix_embedding(self.vid_prefix, b, device)\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2) # b, t, n_word + m, c\n atts_t5 = torch.cat([vid_prefix_mask, atts_t5_rgb], dim=2) # b, t, n_word + m \n \n for modal in self.modalities:\n if modal == 'rgb':\n continue\n if modal in ['depth', 'norm', 'flow']:\n if 'espresso' in self.task:\n fusion_modal.append(t5_inputs[modal])\n else:\n inputs_t5 = torch.cat([inputs_t5, t5_inputs[modal]], dim=2)\n atts_t5 = torch.cat([atts_t5, t5_atts[modal]], dim=2)\n \n if modal in ['pc']:\n if 'espresso' in self.task:\n pc = t5_inputs[modal]\n pc = pc.unsqueeze(1)\n pc = torch.repeat_interleave(pc, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(pc)\n \n if modal in ['audio']:\n if 'espresso' in self.task:\n audio = t5_inputs[modal]\n audio = audio.mean(dim=1)\n audio = audio.unsqueeze(1)\n audio = torch.repeat_interleave(audio, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(audio)\n \n # visual only input\n if 'audio' not in self.modalities and 'pc' not in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n \n # [F1, F2, F3,..., A]\n elif 'audio' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n # seems no prefix works better for audio-video reasoning\n inputs_t5 = torch.cat([inputs_t5, t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, t5_atts['audio'].reshape(b, -1)], dim=1)\n \n elif 'pc' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.pc_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n inputs_t5 = torch.cat([inputs_t5, pc_prefix_embed.squeeze(), t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, pc_prefix_mask.squeeze(), t5_atts['pc'].reshape(b, -1)], dim=1)\n \n elif 'audio' in self.modalities: # audio \n inputs_t5 = t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])\n atts_t5 = t5_atts['audio'].reshape(b, -1)\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([audio_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([audio_prefix_mask.squeeze(), atts_t5], dim=1)\n \n elif 'pc' in self.modalities: # pc\n inputs_t5 = t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])\n atts_t5 = t5_atts['pc'].reshape(b, -1)\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([pc_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([pc_prefix_mask.squeeze(), atts_t5], dim=1)\n\n inputs_embeds = torch.cat([inputs_t5, input_text_embeds], dim=1)\n encoder_atts = torch.cat([atts_t5, input_text.attention_mask], dim=1)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n\n output_tokens = self.t5_tokenizer(\n answer, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n targets_qa = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100)\n output_tokens_mask = output_tokens.attention_mask\n \n outputs = self.t5_model(\n inputs_embeds=inputs_embeds, attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens_mask, return_dict=True, labels=targets_qa)\n loss = outputs.loss\n \n return {'loss': loss}\n \n def encode_input(self, input, modality, training=True):\n\n ln = getattr(self, f\"ln_{modality}\")\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n modality = 'visual'\n if modality in ['audio']:\n modality = 'audio'\n if modality in ['pc']:\n modality = 'pc'\n\n encoder = getattr(self, f\"{modality}_encoder\")\n\n if modality == 'visual':\n b, t, c, w, h = input.shape \n input = input.reshape(-1, c, w, h)\n if training:\n image_embeds = ln(encoder(input))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = ln(encoder(input))\n _, n, _ = image_embeds.shape\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(input.device) # bt n c\n return image_embeds, image_atts\n \n if modality == 'audio':\n embeds, atts = [], []\n for j in range(input.size(1)):\n this_frame = input[:,j,:,:]\n if training:\n embeds.append(encoder(this_frame))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n embeds.append(encoder(this_frame))\n atts.append(torch.ones(embeds[j].size()[:-1], dtype=torch.long).to(input.device))\n \n # print('here', len(embeds), embeds[0].shape) # 2, 3, 256, 768\n embeds = torch.stack(embeds\n# ... truncated ...","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.print_trainable_parameters","uri":"program://CREMA/function/lavis.models.crema_models.crema.print_trainable_parameters#L21-L30","kind":"function","name":"print_trainable_parameters","path":"lavis/models/crema_models/crema.py","language":"python","start_line":21,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2023, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\nimport logging\n\nimport copy\nimport torch\nimport torch.nn as nn\nfrom torch.cuda.amp import autocast as autocast\nfrom transformers import T5TokenizerFast, BertTokenizer\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.Qformer import BertConfig, BertLMHeadModel\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration\nfrom positional_encodings.torch_encodings import PositionalEncoding1D\n\ndef print_trainable_parameters(model):\n trainable_params = 0\n all_param = 0\n for _, param in model.named_parameters():\n all_param += param.numel()\n if param.requires_grad:\n trainable_params += param.numel()\n print(\n f\"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.2f}\"\n )\n\n@registry.register_model(\"crema\")\nclass CREMA(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__( self, img_size=224, drop_path_rate=0,","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.CREMA","uri":"program://CREMA/class/lavis.models.crema_models.crema.CREMA#L33-L772","kind":"class","name":"CREMA","path":"lavis/models/crema_models/crema.py","language":"python","start_line":33,"end_line":772,"context_start_line":13,"context_end_line":772,"code":"from transformers import T5TokenizerFast, BertTokenizer\n\nfrom lavis.common.registry import registry\nfrom lavis.models.blip2_models.Qformer import BertConfig, BertLMHeadModel\nfrom lavis.models.blip2_models.blip2 import Blip2Base, disabled_train\nfrom lavis.models.blip2_models.modeling_t5 import T5Config, T5ForConditionalGeneration\nfrom positional_encodings.torch_encodings import PositionalEncoding1D\n\ndef print_trainable_parameters(model):\n trainable_params = 0\n all_param = 0\n for _, param in model.named_parameters():\n all_param += param.numel()\n if param.requires_grad:\n trainable_params += param.numel()\n print(\n f\"trainable params: {trainable_params} || all params: {all_param} || trainable%: {100 * trainable_params / all_param:.2f}\"\n )\n\n@registry.register_model(\"crema\")\nclass CREMA(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__( self, img_size=224, drop_path_rate=0,\n use_grad_checkpoint=False, vit_precision=\"fp16\", freeze_vit=True,\n num_query_token=32, t5_model=\"google/flan-t5-xl\", prompt=\"\",\n max_txt_len=32, frame_num=8, answer_num=5, apply_lemmatizer=False, \n task='concate',\n modalities='rgb',\n downstream_task='mcqa', # caption / oeqa / mcqa\n lora_rank=64,\n lora_layer=None,\n lora_dropout=0.1,\n fuse_with_base_modality=False):\n\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n \n self.task = task #.split('_')\n self.modalities = modalities.split('_')\n self.fuse_with_base_modality=fuse_with_base_modality\n\n print(self.modalities)\n num_features = 1408\n # ========= init vision encoder ============\n # init vision backbone for vision experts\n if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.visual_encoder = self.init_vision_encoder_only(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision)\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n if 'audio' in self.modalities:\n self.audio_encoder, self.ln_audio = self.init_audio_encoder('beats', cached_audio=False)\n for name, param in self.audio_encoder.named_parameters():\n param.requires_grad = False \n self.audio_encoder = self.audio_encoder.eval()\n self.audio_encoder.train = disabled_train\n logging.info(\"freeze audio encoder\")\n\n if 'pc' in self.modalities:\n # pre-extracted features\n pass\n \n # print('num_features', self.visual_encoder.num_features) 1408\n # ========= init LLM ============ \n # text backbone\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(t5_model, config=t5_config)\n # freeze T5\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16() \n\n # ========= init Qformer ============\n # if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.Qformer, encoder_config = self.init_Multimodal_Qformer(\n num_query_token, num_features, #self.visual_encoder.num_features,\n modulars=self.modalities, \n r=lora_rank, lora_layer=lora_layer, lora_dropout=lora_dropout)\n\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n self.num_query_token = num_query_token\n\n if 'rgb' in self.modalities:\n self.query_tokens_rgb = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_rgb.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_rgb = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_rgb = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'flow' in self.modalities:\n self.query_tokens_flow = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_flow.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_flow = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_flow = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'norm' in self.modalities:\n self.query_tokens_norm = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_norm.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_norm = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_norm = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'depth' in self.modalities:\n self.query_tokens_depth = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_depth.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_depth = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_depth = nn.LayerNorm(self.visual_encoder.num_features)\n \n if 'audio' in self.modalities:\n self.query_tokens_audio = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_audio.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_audio = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n \n self.projection_audio = nn.Linear(self.audio_encoder.num_features, num_features)\n self.ln_audio = nn.LayerNorm(num_features)\n \n if 'pc' in self.modalities:\n self.query_tokens_pc = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_pc.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_pc = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_pc = nn.LayerNorm(num_features)\n pos_model = PositionalEncoding1D(1408 // 3)\n x = torch.zeros(1, 256, 1408 // 3)\n self.pos_embedding = pos_model(x).squeeze().cuda()\n \n if 'espresso' in self.task:\n\n self.fusion = nn.Sequential(\n nn.Linear(2048*(len(self.modalities)-1), 2048),\n )\n self.sigmoid = nn.Sigmoid()\n \n self.downstream_task = downstream_task \n self.max_txt_len = 77\n answer_id = [71, 272, 205, 309, 262] # A B C D E\n self.answer_id = answer_id[:answer_num]\n \n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n self.frame_num = frame_num\n self.ANS_MAP = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n if frame_num == 1:\n self.vid_prefix = ['Frame: ']\n self.depth_prefix = ['Depth Map: ']\n self.flow_prefix = ['Optical Flow: ']\n self.norm_prefix = ['Surface Normalization: ']\n else:\n self.vid_prefix = ['Frame {}: '.format(str(i+1)) for i in range(frame_num)]\n self.depth_prefix = ['Depth Map {}: '.format(str(i+1)) for i in range(frame_num)]\n self.flow_prefix = ['Optical Flow {}: '.format(str(i+1)) for i in range(frame_num)]\n self.norm_prefix = ['Surface Normalization {}: '.format(str(i+1)) for i in range(frame_num)]\n\n self.audio_prefix = ['Audio: ']\n self.pc_prefix = ['3D Model: ']\n \n def forward(self, samples):\n\n # rgb visual embedding\n qa_text, answer = samples['qa_input'], samples['qa_output']\n b = len(qa_text)\n\n input_embed_dict, input_att_dict = {}, {}\n\n for modal in self.modalities:\n input = samples[modal]\n if modal in ['rgb', 'depth', 'norm', 'flow']:\n # fix some loading issue\n if input.shape[1] == 3:\n input = input.permute(0, 2, 1, 3, 4)\n \n # following 3D-LLM \n if modal == 'pc':\n with torch.cuda.amp.autocast(dtype=torch.float32):\n pc_embeds = samples[\"pc_feat\"]\n pc = samples[\"pc\"].long()\n all_pcs = torch.zeros((pc_embeds.shape))\n for j in range(pc.shape[0]):\n pcs = []\n for i in range(3):\n pc_i = pc[j][:, i]\n pcs.append(self.pos_embedding[pc_i])\n pcs = torch.cat(pcs, -1)\n all_pcs[j][:, :1407] = pcs\n all_pcs = all_pcs.cuda()\n pc_embeds = pc_embeds + 0.01 * all_pcs\n atts = torch.ones(pc_embeds.size()[:-1], dtype=torch.long).to(pc_embeds.device)\n input_embed_dict[modal], input_att_dict[modal] = pc_embeds, atts\n else:\n input_embed_dict[modal], input_att_dict[modal] = self.encode_input(input, modal)\n \n device = input_embed_dict[list(input_embed_dict.keys())[0]].device\n\n input_text= self.t5_tokenizer(\n qa_text, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n input_text_embeds = self.t5_model.encoder.embed_tokens(input_text.input_ids) \n\n\n fusion_modal = []\n t5_inputs, t5_atts, t5_query = {}, {}, {}\n for modal in self.modalities:\n t5_inputs[modal], t5_atts[modal], t5_query[modal] = self.get_qformer_embedding(input_embed_dict[modal], input_att_dict[modal], device, modal, b)\n\n if 'rgb' in self.modalities:\n inputs_t5_rgb = t5_inputs['rgb']\n atts_t5_rgb = t5_atts['rgb']\n vid_prefix_embed, vid_prefix_mask = self.get_prefix_embedding(self.vid_prefix, b, device)\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2) # b, t, n_word + m, c\n atts_t5 = torch.cat([vid_prefix_mask, atts_t5_rgb], dim=2) # b, t, n_word + m \n \n for modal in self.modalities:\n if modal == 'rgb':\n continue\n if modal in ['depth', 'norm', 'flow']:\n if 'espresso' in self.task:\n fusion_modal.append(t5_inputs[modal])\n else:\n inputs_t5 = torch.cat([inputs_t5, t5_inputs[modal]], dim=2)\n atts_t5 = torch.cat([atts_t5, t5_atts[modal]], dim=2)\n \n if modal in ['pc']:\n if 'espresso' in self.task:\n pc = t5_inputs[modal]\n pc = pc.unsqueeze(1)\n pc = torch.repeat_interleave(pc, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(pc)\n \n if modal in ['audio']:\n if 'espresso' in self.task:\n audio = t5_inputs[modal]\n audio = audio.mean(dim=1)\n audio = audio.unsqueeze(1)\n audio = torch.repeat_interleave(audio, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(audio)\n \n # visual only input\n if 'audio' not in self.modalities and 'pc' not in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n \n # [F1, F2, F3,..., A]\n elif 'audio' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n # seems no prefix works better for audio-video reasoning\n inputs_t5 = torch.cat([inputs_t5, t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, t5_atts['audio'].reshape(b, -1)], dim=1)\n \n elif 'pc' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.pc_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n inputs_t5 = torch.cat([inputs_t5, pc_prefix_embed.squeeze(), t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, pc_prefix_mask.squeeze(), t5_atts['pc'].reshape(b, -1)], dim=1)\n \n elif 'audio' in self.modalities: # audio \n inputs_t5 = t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])\n atts_t5 = t5_atts['audio'].reshape(b, -1)\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([audio_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([audio_prefix_mask.squeeze(), atts_t5], dim=1)\n \n elif 'pc' in self.modalities: # pc\n inputs_t5 = t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])\n atts_t5 = t5_atts['pc'].reshape(b, -1)\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([pc_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([pc_prefix_mask.squeeze(), atts_t5], dim=1)\n\n inputs_embeds = torch.cat([inputs_t5, input_text_embeds], dim=1)\n encoder_atts = torch.cat([atts_t5, input_text.attention_mask], dim=1)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n\n output_tokens = self.t5_tokenizer(\n answer, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n targets_qa = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100)\n output_tokens_mask = output_tokens.attention_mask\n \n outputs = self.t5_model(\n inputs_embeds=inputs_embeds, attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens_mask, return_dict=True, labels=targets_qa)\n loss = outputs.loss\n \n return {'loss': loss}\n \n def encode_input(self, input, modality, training=True):\n\n ln = getattr(self, f\"ln_{modality}\")\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n modality = 'visual'\n if modality in ['audio']:\n modality = 'audio'\n if modality in ['pc']:\n modality = 'pc'\n\n encoder = getattr(self, f\"{modality}_encoder\")\n\n if modality == 'visual':\n b, t, c, w, h = input.shape \n input = input.reshape(-1, c, w, h)\n if training:\n image_embeds = ln(encoder(input))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = ln(encoder(input))\n _, n, _ = image_embeds.shape\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(input.device) # bt n c\n return image_embeds, image_atts\n \n if modality == 'audio':\n embeds, atts = [], []\n for j in range(input.size(1)):\n this_frame = input[:,j,:,:]\n if training:\n embeds.append(encoder(this_frame))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n embeds.append(encoder(this_frame))\n atts.append(torch.ones(embeds[j].size()[:-1], dtype=torch.long).to(input.device))\n \n # print('here', len(embeds), embeds[0].shape) # 2, 3, 256, 768\n embeds = torch.stack(embeds, dim=1)\n # print('audio_embeds 1', embeds.shape) # 3, 2, 256, 768\n atts = torch.stack(atts, dim=1)\n embeds = self.projection_audio(embeds) # 3, 2, 256, 1408\n embeds = ln(embeds.reshape(-1, embeds.shape[-2], embeds.shape[-1]))\n atts = atts.reshape(-1, atts.shape[-1])\n# ... truncated ...","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.__init__","uri":"program://CREMA/function/lavis.models.crema_models.crema.__init__#L50-L204","kind":"function","name":"__init__","path":"lavis/models/crema_models/crema.py","language":"python","start_line":50,"end_line":204,"context_start_line":30,"context_end_line":224,"code":" )\n\n@registry.register_model(\"crema\")\nclass CREMA(Blip2Base):\n \"\"\"\n BLIP2 T5 model.\n Supported model types:\n - pretrain_flant5xl: pretrained model with FlanT5-XL\n - pretrain_flant5xxl: pretrained model with FlanT5-XXL\n - caption_coco_flant5xl: fintuned image captioning model with FlanT5-XL\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip2_t5\", \"pretrain_flant5xl\")\n \"\"\"\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"pretrain_flant5xl\": \"configs/models/blip2/blip2_pretrain_flant5xl.yaml\",\n \"pretrain_flant5xxl\": \"configs/models/blip2/blip2_pretrain_flant5xxl.yaml\",\n \"caption_coco_flant5xl\": \"configs/models/blip2/blip2_caption_flant5xl.yaml\",\n }\n\n def __init__( self, img_size=224, drop_path_rate=0,\n use_grad_checkpoint=False, vit_precision=\"fp16\", freeze_vit=True,\n num_query_token=32, t5_model=\"google/flan-t5-xl\", prompt=\"\",\n max_txt_len=32, frame_num=8, answer_num=5, apply_lemmatizer=False, \n task='concate',\n modalities='rgb',\n downstream_task='mcqa', # caption / oeqa / mcqa\n lora_rank=64,\n lora_layer=None,\n lora_dropout=0.1,\n fuse_with_base_modality=False):\n\n \"\"\"\n apply_lemmatizer: when set to True, postprocess predict_answers() result with lemmas.\n \"\"\"\n super().__init__()\n \n self.task = task #.split('_')\n self.modalities = modalities.split('_')\n self.fuse_with_base_modality=fuse_with_base_modality\n\n print(self.modalities)\n num_features = 1408\n # ========= init vision encoder ============\n # init vision backbone for vision experts\n if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.visual_encoder = self.init_vision_encoder_only(\n img_size, drop_path_rate, use_grad_checkpoint, vit_precision)\n for name, param in self.visual_encoder.named_parameters():\n param.requires_grad = False \n self.visual_encoder = self.visual_encoder.eval()\n self.visual_encoder.train = disabled_train\n logging.info(\"freeze vision encoder\")\n if 'audio' in self.modalities:\n self.audio_encoder, self.ln_audio = self.init_audio_encoder('beats', cached_audio=False)\n for name, param in self.audio_encoder.named_parameters():\n param.requires_grad = False \n self.audio_encoder = self.audio_encoder.eval()\n self.audio_encoder.train = disabled_train\n logging.info(\"freeze audio encoder\")\n\n if 'pc' in self.modalities:\n # pre-extracted features\n pass\n \n # print('num_features', self.visual_encoder.num_features) 1408\n # ========= init LLM ============ \n # text backbone\n self.t5_tokenizer = T5TokenizerFast.from_pretrained(t5_model)\n t5_config = T5Config.from_pretrained(t5_model)\n t5_config.dense_act_fn = \"gelu\"\n self.t5_model = T5ForConditionalGeneration.from_pretrained(t5_model, config=t5_config)\n # freeze T5\n for name, param in self.t5_model.named_parameters():\n param.requires_grad = False\n param.data = param.data.bfloat16() \n\n # ========= init Qformer ============\n # if 'rgb' in self.modalities or 'depth' in self.modalities or 'flow' in self.modalities or 'norm' in self.modalities:\n self.Qformer, encoder_config = self.init_Multimodal_Qformer(\n num_query_token, num_features, #self.visual_encoder.num_features,\n modulars=self.modalities, \n r=lora_rank, lora_layer=lora_layer, lora_dropout=lora_dropout)\n\n self.Qformer.cls = None\n self.Qformer.bert.embeddings.word_embeddings = None\n self.Qformer.bert.embeddings.position_embeddings = None\n for layer in self.Qformer.bert.encoder.layer:\n layer.output = None\n layer.intermediate = None\n self.num_query_token = num_query_token\n\n if 'rgb' in self.modalities:\n self.query_tokens_rgb = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_rgb.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_rgb = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_rgb = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'flow' in self.modalities:\n self.query_tokens_flow = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_flow.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_flow = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_flow = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'norm' in self.modalities:\n self.query_tokens_norm = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_norm.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_norm = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_norm = nn.LayerNorm(self.visual_encoder.num_features)\n\n if 'depth' in self.modalities:\n self.query_tokens_depth = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_depth.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_depth = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_depth = nn.LayerNorm(self.visual_encoder.num_features)\n \n if 'audio' in self.modalities:\n self.query_tokens_audio = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_audio.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_audio = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n \n self.projection_audio = nn.Linear(self.audio_encoder.num_features, num_features)\n self.ln_audio = nn.LayerNorm(num_features)\n \n if 'pc' in self.modalities:\n self.query_tokens_pc = nn.Parameter(\n torch.zeros(1, num_query_token, encoder_config.hidden_size))\n self.query_tokens_pc.data.normal_(mean=0.0, std=encoder_config.initializer_range)\n self.t5_proj_pc = nn.Linear(\n self.Qformer.config.hidden_size, self.t5_model.config.hidden_size)\n self.ln_pc = nn.LayerNorm(num_features)\n pos_model = PositionalEncoding1D(1408 // 3)\n x = torch.zeros(1, 256, 1408 // 3)\n self.pos_embedding = pos_model(x).squeeze().cuda()\n \n if 'espresso' in self.task:\n\n self.fusion = nn.Sequential(\n nn.Linear(2048*(len(self.modalities)-1), 2048),\n )\n self.sigmoid = nn.Sigmoid()\n \n self.downstream_task = downstream_task \n self.max_txt_len = 77\n answer_id = [71, 272, 205, 309, 262] # A B C D E\n self.answer_id = answer_id[:answer_num]\n \n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n self.frame_num = frame_num\n self.ANS_MAP = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n if frame_num == 1:\n self.vid_prefix = ['Frame: ']\n self.depth_prefix = ['Depth Map: ']\n self.flow_prefix = ['Optical Flow: ']\n self.norm_prefix = ['Surface Normalization: ']\n else:\n self.vid_prefix = ['Frame {}: '.format(str(i+1)) for i in range(frame_num)]\n self.depth_prefix = ['Depth Map {}: '.format(str(i+1)) for i in range(frame_num)]\n self.flow_prefix = ['Optical Flow {}: '.format(str(i+1)) for i in range(frame_num)]\n self.norm_prefix = ['Surface Normalization {}: '.format(str(i+1)) for i in range(frame_num)]\n\n self.audio_prefix = ['Audio: ']\n self.pc_prefix = ['3D Model: ']\n \n def forward(self, samples):\n\n # rgb visual embedding\n qa_text, answer = samples['qa_input'], samples['qa_output']\n b = len(qa_text)\n\n input_embed_dict, input_att_dict = {}, {}\n\n for modal in self.modalities:\n input = samples[modal]\n if modal in ['rgb', 'depth', 'norm', 'flow']:\n # fix some loading issue\n if input.shape[1] == 3:\n input = input.permute(0, 2, 1, 3, 4)\n \n # following 3D-LLM \n if modal == 'pc':\n with torch.cuda.amp.autocast(dtype=torch.float32):\n pc_embeds = samples[\"pc_feat\"]","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.forward","uri":"program://CREMA/function/lavis.models.crema_models.crema.forward#L206-L365","kind":"function","name":"forward","path":"lavis/models/crema_models/crema.py","language":"python","start_line":206,"end_line":365,"context_start_line":186,"context_end_line":385,"code":" \n self._apply_lemmatizer = apply_lemmatizer\n self._lemmatizer = None\n self.frame_num = frame_num\n self.ANS_MAP = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n if frame_num == 1:\n self.vid_prefix = ['Frame: ']\n self.depth_prefix = ['Depth Map: ']\n self.flow_prefix = ['Optical Flow: ']\n self.norm_prefix = ['Surface Normalization: ']\n else:\n self.vid_prefix = ['Frame {}: '.format(str(i+1)) for i in range(frame_num)]\n self.depth_prefix = ['Depth Map {}: '.format(str(i+1)) for i in range(frame_num)]\n self.flow_prefix = ['Optical Flow {}: '.format(str(i+1)) for i in range(frame_num)]\n self.norm_prefix = ['Surface Normalization {}: '.format(str(i+1)) for i in range(frame_num)]\n\n self.audio_prefix = ['Audio: ']\n self.pc_prefix = ['3D Model: ']\n \n def forward(self, samples):\n\n # rgb visual embedding\n qa_text, answer = samples['qa_input'], samples['qa_output']\n b = len(qa_text)\n\n input_embed_dict, input_att_dict = {}, {}\n\n for modal in self.modalities:\n input = samples[modal]\n if modal in ['rgb', 'depth', 'norm', 'flow']:\n # fix some loading issue\n if input.shape[1] == 3:\n input = input.permute(0, 2, 1, 3, 4)\n \n # following 3D-LLM \n if modal == 'pc':\n with torch.cuda.amp.autocast(dtype=torch.float32):\n pc_embeds = samples[\"pc_feat\"]\n pc = samples[\"pc\"].long()\n all_pcs = torch.zeros((pc_embeds.shape))\n for j in range(pc.shape[0]):\n pcs = []\n for i in range(3):\n pc_i = pc[j][:, i]\n pcs.append(self.pos_embedding[pc_i])\n pcs = torch.cat(pcs, -1)\n all_pcs[j][:, :1407] = pcs\n all_pcs = all_pcs.cuda()\n pc_embeds = pc_embeds + 0.01 * all_pcs\n atts = torch.ones(pc_embeds.size()[:-1], dtype=torch.long).to(pc_embeds.device)\n input_embed_dict[modal], input_att_dict[modal] = pc_embeds, atts\n else:\n input_embed_dict[modal], input_att_dict[modal] = self.encode_input(input, modal)\n \n device = input_embed_dict[list(input_embed_dict.keys())[0]].device\n\n input_text= self.t5_tokenizer(\n qa_text, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n input_text_embeds = self.t5_model.encoder.embed_tokens(input_text.input_ids) \n\n\n fusion_modal = []\n t5_inputs, t5_atts, t5_query = {}, {}, {}\n for modal in self.modalities:\n t5_inputs[modal], t5_atts[modal], t5_query[modal] = self.get_qformer_embedding(input_embed_dict[modal], input_att_dict[modal], device, modal, b)\n\n if 'rgb' in self.modalities:\n inputs_t5_rgb = t5_inputs['rgb']\n atts_t5_rgb = t5_atts['rgb']\n vid_prefix_embed, vid_prefix_mask = self.get_prefix_embedding(self.vid_prefix, b, device)\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2) # b, t, n_word + m, c\n atts_t5 = torch.cat([vid_prefix_mask, atts_t5_rgb], dim=2) # b, t, n_word + m \n \n for modal in self.modalities:\n if modal == 'rgb':\n continue\n if modal in ['depth', 'norm', 'flow']:\n if 'espresso' in self.task:\n fusion_modal.append(t5_inputs[modal])\n else:\n inputs_t5 = torch.cat([inputs_t5, t5_inputs[modal]], dim=2)\n atts_t5 = torch.cat([atts_t5, t5_atts[modal]], dim=2)\n \n if modal in ['pc']:\n if 'espresso' in self.task:\n pc = t5_inputs[modal]\n pc = pc.unsqueeze(1)\n pc = torch.repeat_interleave(pc, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(pc)\n \n if modal in ['audio']:\n if 'espresso' in self.task:\n audio = t5_inputs[modal]\n audio = audio.mean(dim=1)\n audio = audio.unsqueeze(1)\n audio = torch.repeat_interleave(audio, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(audio)\n \n # visual only input\n if 'audio' not in self.modalities and 'pc' not in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n \n # [F1, F2, F3,..., A]\n elif 'audio' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n # seems no prefix works better for audio-video reasoning\n inputs_t5 = torch.cat([inputs_t5, t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, t5_atts['audio'].reshape(b, -1)], dim=1)\n \n elif 'pc' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.pc_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n inputs_t5 = torch.cat([inputs_t5, pc_prefix_embed.squeeze(), t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, pc_prefix_mask.squeeze(), t5_atts['pc'].reshape(b, -1)], dim=1)\n \n elif 'audio' in self.modalities: # audio \n inputs_t5 = t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])\n atts_t5 = t5_atts['audio'].reshape(b, -1)\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([audio_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([audio_prefix_mask.squeeze(), atts_t5], dim=1)\n \n elif 'pc' in self.modalities: # pc\n inputs_t5 = t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])\n atts_t5 = t5_atts['pc'].reshape(b, -1)\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([pc_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([pc_prefix_mask.squeeze(), atts_t5], dim=1)\n\n inputs_embeds = torch.cat([inputs_t5, input_text_embeds], dim=1)\n encoder_atts = torch.cat([atts_t5, input_text.attention_mask], dim=1)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n\n output_tokens = self.t5_tokenizer(\n answer, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n targets_qa = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100)\n output_tokens_mask = output_tokens.attention_mask\n \n outputs = self.t5_model(\n inputs_embeds=inputs_embeds, attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens_mask, return_dict=True, labels=targets_qa)\n loss = outputs.loss\n \n return {'loss': loss}\n \n def encode_input(self, input, modality, training=True):\n\n ln = getattr(self, f\"ln_{modality}\")\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n modality = 'visual'\n if modality in ['audio']:\n modality = 'audio'\n if modality in ['pc']:\n modality = 'pc'\n\n encoder = getattr(self, f\"{modality}_encoder\")\n\n if modality == 'visual':\n b, t, c, w, h = input.shape \n input = input.reshape(-1, c, w, h)\n if training:\n image_embeds = ln(encoder(input))\n else:","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.encode_input","uri":"program://CREMA/function/lavis.models.crema_models.crema.encode_input#L367-L415","kind":"function","name":"encode_input","path":"lavis/models/crema_models/crema.py","language":"python","start_line":367,"end_line":415,"context_start_line":347,"context_end_line":435,"code":"\n inputs_embeds = torch.cat([inputs_t5, input_text_embeds], dim=1)\n encoder_atts = torch.cat([atts_t5, input_text.attention_mask], dim=1)\n\n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n\n output_tokens = self.t5_tokenizer(\n answer, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n targets_qa = output_tokens.input_ids.masked_fill(\n output_tokens.input_ids == self.t5_tokenizer.pad_token_id, -100)\n output_tokens_mask = output_tokens.attention_mask\n \n outputs = self.t5_model(\n inputs_embeds=inputs_embeds, attention_mask=encoder_atts,\n decoder_attention_mask=output_tokens_mask, return_dict=True, labels=targets_qa)\n loss = outputs.loss\n \n return {'loss': loss}\n \n def encode_input(self, input, modality, training=True):\n\n ln = getattr(self, f\"ln_{modality}\")\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n modality = 'visual'\n if modality in ['audio']:\n modality = 'audio'\n if modality in ['pc']:\n modality = 'pc'\n\n encoder = getattr(self, f\"{modality}_encoder\")\n\n if modality == 'visual':\n b, t, c, w, h = input.shape \n input = input.reshape(-1, c, w, h)\n if training:\n image_embeds = ln(encoder(input))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n image_embeds = ln(encoder(input))\n _, n, _ = image_embeds.shape\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(input.device) # bt n c\n return image_embeds, image_atts\n \n if modality == 'audio':\n embeds, atts = [], []\n for j in range(input.size(1)):\n this_frame = input[:,j,:,:]\n if training:\n embeds.append(encoder(this_frame))\n else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n embeds.append(encoder(this_frame))\n atts.append(torch.ones(embeds[j].size()[:-1], dtype=torch.long).to(input.device))\n \n # print('here', len(embeds), embeds[0].shape) # 2, 3, 256, 768\n embeds = torch.stack(embeds, dim=1)\n # print('audio_embeds 1', embeds.shape) # 3, 2, 256, 768\n atts = torch.stack(atts, dim=1)\n embeds = self.projection_audio(embeds) # 3, 2, 256, 1408\n embeds = ln(embeds.reshape(-1, embeds.shape[-2], embeds.shape[-1]))\n atts = atts.reshape(-1, atts.shape[-1])\n\n return embeds, atts\n \n if modality == 'pc':\n # use pre-extracted features\n pass\n #return embeds, atts\n \n def get_qformer_embedding(self, embeds, atts, device, modality, bs):\n\n project = getattr(self, f\"t5_proj_{modality}\")\n query_tokens = getattr(self, f\"query_tokens_{modality}\")\n query_tokens = query_tokens.expand(embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens, encoder_hidden_states=embeds,\n encoder_attention_mask=atts, return_dict=True, modular=modality)\n \n query = query_output.last_hidden_state.clone()\n inputs_t5 = project(query_output.last_hidden_state)\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n inputs_t5 = inputs_t5.reshape(-1, self.frame_num, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['audio']:","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.get_qformer_embedding","uri":"program://CREMA/function/lavis.models.crema_models.crema.get_qformer_embedding#L418-L442","kind":"function","name":"get_qformer_embedding","path":"lavis/models/crema_models/crema.py","language":"python","start_line":418,"end_line":442,"context_start_line":398,"context_end_line":462,"code":" else:\n with torch.cuda.amp.autocast(enabled=(self.device != torch.device(\"cpu\"))):\n embeds.append(encoder(this_frame))\n atts.append(torch.ones(embeds[j].size()[:-1], dtype=torch.long).to(input.device))\n \n # print('here', len(embeds), embeds[0].shape) # 2, 3, 256, 768\n embeds = torch.stack(embeds, dim=1)\n # print('audio_embeds 1', embeds.shape) # 3, 2, 256, 768\n atts = torch.stack(atts, dim=1)\n embeds = self.projection_audio(embeds) # 3, 2, 256, 1408\n embeds = ln(embeds.reshape(-1, embeds.shape[-2], embeds.shape[-1]))\n atts = atts.reshape(-1, atts.shape[-1])\n\n return embeds, atts\n \n if modality == 'pc':\n # use pre-extracted features\n pass\n #return embeds, atts\n \n def get_qformer_embedding(self, embeds, atts, device, modality, bs):\n\n project = getattr(self, f\"t5_proj_{modality}\")\n query_tokens = getattr(self, f\"query_tokens_{modality}\")\n query_tokens = query_tokens.expand(embeds.shape[0], -1, -1)\n\n query_output = self.Qformer.bert(\n query_embeds=query_tokens, encoder_hidden_states=embeds,\n encoder_attention_mask=atts, return_dict=True, modular=modality)\n \n query = query_output.last_hidden_state.clone()\n inputs_t5 = project(query_output.last_hidden_state)\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n inputs_t5 = inputs_t5.reshape(-1, self.frame_num, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['audio']:\n inputs_t5 = inputs_t5.reshape(bs, -1, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['pc']:\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n\n return inputs_t5, atts_t5, query\n \n def get_prefix_embedding(self, prefix_, b, device):\n prefix = self.t5_tokenizer(\n prefix_, padding=\"longest\", add_special_tokens=False,\n truncation=True, max_length=self.max_txt_len, return_tensors=\"pt\",).to(device) # \n prefix_id = torch.repeat_interleave(prefix.input_ids.unsqueeze(0), b, 0)\n prefix_mask = torch.repeat_interleave(prefix.attention_mask.unsqueeze(0), b, 0)\n prefix_embed = self.t5_model.encoder.embed_tokens(prefix_id) # b t n_word c\n return prefix_embed, prefix_mask\n \n @torch.no_grad()\n def generate(self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5, max_length=30,\n min_length=1, top_p=0.9,\n repetition_penalty=1.0, length_penalty=1.0,\n num_captions=1, temperature=1,):\n \"\"\"\n Args:","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.get_prefix_embedding","uri":"program://CREMA/function/lavis.models.crema_models.crema.get_prefix_embedding#L444-L451","kind":"function","name":"get_prefix_embedding","path":"lavis/models/crema_models/crema.py","language":"python","start_line":444,"end_line":451,"context_start_line":424,"context_end_line":471,"code":" query_output = self.Qformer.bert(\n query_embeds=query_tokens, encoder_hidden_states=embeds,\n encoder_attention_mask=atts, return_dict=True, modular=modality)\n \n query = query_output.last_hidden_state.clone()\n inputs_t5 = project(query_output.last_hidden_state)\n\n if modality in ['rgb', 'depth', 'flow', 'norm']:\n inputs_t5 = inputs_t5.reshape(-1, self.frame_num, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['audio']:\n inputs_t5 = inputs_t5.reshape(bs, -1, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['pc']:\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n\n return inputs_t5, atts_t5, query\n \n def get_prefix_embedding(self, prefix_, b, device):\n prefix = self.t5_tokenizer(\n prefix_, padding=\"longest\", add_special_tokens=False,\n truncation=True, max_length=self.max_txt_len, return_tensors=\"pt\",).to(device) # \n prefix_id = torch.repeat_interleave(prefix.input_ids.unsqueeze(0), b, 0)\n prefix_mask = torch.repeat_interleave(prefix.attention_mask.unsqueeze(0), b, 0)\n prefix_embed = self.t5_model.encoder.embed_tokens(prefix_id) # b t n_word c\n return prefix_embed, prefix_mask\n \n @torch.no_grad()\n def generate(self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5, max_length=30,\n min_length=1, top_p=0.9,\n repetition_penalty=1.0, length_penalty=1.0,\n num_captions=1, temperature=1,):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.generate","uri":"program://CREMA/function/lavis.models.crema_models.crema.generate#L454-L650","kind":"function","name":"generate","path":"lavis/models/crema_models/crema.py","language":"python","start_line":454,"end_line":650,"context_start_line":434,"context_end_line":670,"code":" \n if modality in ['audio']:\n inputs_t5 = inputs_t5.reshape(bs, -1, inputs_t5.shape[-2], inputs_t5.shape[-1])\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n \n if modality in ['pc']:\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n\n return inputs_t5, atts_t5, query\n \n def get_prefix_embedding(self, prefix_, b, device):\n prefix = self.t5_tokenizer(\n prefix_, padding=\"longest\", add_special_tokens=False,\n truncation=True, max_length=self.max_txt_len, return_tensors=\"pt\",).to(device) # \n prefix_id = torch.repeat_interleave(prefix.input_ids.unsqueeze(0), b, 0)\n prefix_mask = torch.repeat_interleave(prefix.attention_mask.unsqueeze(0), b, 0)\n prefix_embed = self.t5_model.encoder.embed_tokens(prefix_id) # b t n_word c\n return prefix_embed, prefix_mask\n \n @torch.no_grad()\n def generate(self,\n samples,\n use_nucleus_sampling=False,\n num_beams=5, max_length=30,\n min_length=1, top_p=0.9,\n repetition_penalty=1.0, length_penalty=1.0,\n num_captions=1, temperature=1,):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n \"\"\"\n out = {}\n qid = samples['question_id']\n qa_text = samples['qa_input']\n answer = samples['qa_output']\n b = len(qa_text)\n\n input_embed_dict, input_att_dict = {}, {}\n\n for modal in self.modalities:\n input = samples[modal]\n # visual modality pre-process\n if modal in ['rgb', 'depth', 'norm', 'flow']:\n if input.shape[1] == 3:\n input = input.permute(0, 2, 1, 3, 4)\n # 3d: direct load pre-processed features\n if modal == 'pc':\n with torch.cuda.amp.autocast(dtype=torch.float32):\n pc_embeds = samples[\"pc_feat\"]\n pc = samples[\"pc\"].long()\n all_pcs = torch.zeros((pc_embeds.shape))\n for j in range(pc.shape[0]):\n pcs = []\n for i in range(3):\n pc_i = pc[j][:, i]\n pcs.append(self.pos_embedding[pc_i])\n pcs = torch.cat(pcs, -1)\n all_pcs[j][:, :1407] = pcs\n all_pcs = all_pcs.cuda()\n pc_embeds = pc_embeds + 0.01 * all_pcs\n atts = torch.ones(pc_embeds.size()[:-1], dtype=torch.long).to(pc_embeds.device)\n input_embed_dict[modal], input_att_dict[modal] = pc_embeds, atts\n else:\n input_embed_dict[modal], input_att_dict[modal] = self.encode_input(input, modal, training=False)\n \n device = input_embed_dict[list(input_embed_dict.keys())[0]].device\n fusion_modal = []\n input_text= self.t5_tokenizer(\n qa_text, padding=\"longest\", truncation=True,\n max_length=self.max_txt_len, return_tensors=\"pt\").to(device)\n input_text_embeds = self.t5_model.encoder.embed_tokens(input_text.input_ids) \n\n with torch.no_grad():\n \n t5_inputs, t5_atts, t5_query = {}, {}, {}\n for modal in self.modalities:\n t5_inputs[modal], t5_atts[modal], t5_query[modal] = self.get_qformer_embedding(input_embed_dict[modal], input_att_dict[modal], device, modal, b)\n\n # different modality combination\n if 'rgb' in self.modalities:\n inputs_t5_rgb = t5_inputs['rgb']\n atts_t5_rgb = t5_atts['rgb']\n vid_prefix_embed, vid_prefix_mask = self.get_prefix_embedding(self.vid_prefix, b, device)\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2) # b, t, n_word + m, c\n atts_t5 = torch.cat([vid_prefix_mask, atts_t5_rgb], dim=2) # b, t, n_word + m \n for modal in self.modalities:\n if modal == 'rgb':\n continue\n if modal in ['depth', 'norm', 'flow']:\n if 'espresso' in self.task:\n fusion_modal.append(t5_inputs[modal])\n else:\n inputs_t5 = torch.cat([inputs_t5, t5_inputs[modal]], dim=2)\n atts_t5 = torch.cat([atts_t5, t5_atts[modal]], dim=2) \n if modal in ['pc']:\n if 'espresso' in self.task:\n pc = t5_inputs[modal]\n pc = pc.unsqueeze(1)\n pc = torch.repeat_interleave(pc, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(pc)\n\n if modal in ['audio']:\n if 'espresso' in self.task:\n audio = t5_inputs[modal]\n audio = audio.mean(dim=1)\n audio = audio.unsqueeze(1)\n audio = torch.repeat_interleave(audio, self.frame_num, 1) # [16, 4, 32, 2048]\n fusion_modal.append(audio)\n \n # visual only input\n if 'audio' not in self.modalities and 'pc' not in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1)\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n # [F1, F2, F3,..., A]\n elif 'audio' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n inputs_t5 = torch.cat([inputs_t5, t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, t5_atts['audio'].reshape(b, -1)], dim=1)\n \n elif 'pc' in self.modalities:\n if 'espresso' in self.task:\n fusion_modal = torch.cat(fusion_modal, dim=-1) # 16, 4, 32, 8192\n inputs_t5_extra = self.fusion(fusion_modal)\n inputs_t5_rgb += self.sigmoid(inputs_t5_extra) * inputs_t5_extra\n inputs_t5 = torch.cat([vid_prefix_embed, inputs_t5_rgb], dim=2)\n atts_t5 = torch.ones(inputs_t5.size()[:-1], dtype=torch.long).to(device)\n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n else:\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.pc_prefix, b, device) \n inputs_t5 = inputs_t5.reshape(b, -1, inputs_t5.shape[-1])\n atts_t5 = atts_t5.reshape(b, -1)\n inputs_t5 = torch.cat([inputs_t5, pc_prefix_embed.squeeze(), t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])], dim=1)\n atts_t5 = torch.cat([atts_t5, pc_prefix_mask.squeeze(), t5_atts['pc'].reshape(b, -1)], dim=1)\n \n elif 'audio' in self.modalities: # audio only\n inputs_t5 = t5_inputs['audio'].reshape(b, -1, t5_inputs['audio'].shape[-1])\n atts_t5 = t5_atts['audio'].reshape(b, -1)\n audio_prefix_embed, audio_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([audio_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([audio_prefix_mask.squeeze(), atts_t5], dim=1)\n \n elif 'pc' in self.modalities: # pc only\n inputs_t5 = t5_inputs['pc'].reshape(b, -1, t5_inputs['pc'].shape[-1])\n atts_t5 = t5_atts['pc'].reshape(b, -1)\n pc_prefix_embed, pc_prefix_mask = self.get_prefix_embedding(self.audio_prefix, b, device) \n inputs_t5 = torch.cat([pc_prefix_embed.squeeze(), inputs_t5], dim=1)\n atts_t5 = torch.cat([pc_prefix_mask.squeeze(), atts_t5], dim=1)\n\n inputs_embeds = torch.cat([inputs_t5, input_text_embeds], dim=1)\n encoder_atts = torch.cat([atts_t5, input_text.attention_mask], dim=1)\n \n with torch.cuda.amp.autocast(dtype=torch.bfloat16):\n \n if self.downstream_task == 'mcqa':\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds, attention_mask=encoder_atts,\n do_sample=use_nucleus_sampling, top_p=top_p,\n temperature=temperature, num_beams=1,\n max_new_tokens=max_length, min_length=min_length,\n repetition_penalty=repetition_penalty, length_penalty=length_penalty,\n num_return_sequences=num_captions, return_dict_in_generate=True,\n output_hidden_states=True, output_scores=True)\n try:\n pred_logits = outputs.scores[1]\n except:\n pred_logits = outputs.scores[0]\n pred_logits = pred_logits[:, self.answer_id] # b, 5\n pred_ans = torch.argmax(pred_logits, dim=-1).cpu().tolist() \n\n elif self.downstream_task == 'oeqa':\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n length_penalty=length_penalty,\n )\n pred_ans = self.t5_tokenizer.batch_decode(outputs, skip_special_tokens=True)\n \n out['output_text'] = pred_ans\n out['answer'] = answer\n out['qid'] = qid\n\n return out\n\n @torch.no_grad()\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n if isinstance(samples[\"qa_input\"], str):\n samples[\"qa_input\"] = [samples[\"qa_input\"]]\n \n text_input = samples[\"qa_input\"]\n samples[\"prompt\"] = text_input","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.predict_answers","uri":"program://CREMA/function/lavis.models.crema_models.crema.predict_answers#L653-L684","kind":"function","name":"predict_answers","path":"lavis/models/crema_models/crema.py","language":"python","start_line":653,"end_line":684,"context_start_line":633,"context_end_line":704,"code":"\n elif self.downstream_task == 'oeqa':\n outputs = self.t5_model.generate(\n inputs_embeds=inputs_embeds,\n attention_mask=encoder_atts,\n do_sample=False,\n num_beams=num_beams,\n max_new_tokens=max_length,\n min_length=min_length,\n length_penalty=length_penalty,\n )\n pred_ans = self.t5_tokenizer.batch_decode(outputs, skip_special_tokens=True)\n \n out['output_text'] = pred_ans\n out['answer'] = answer\n out['qid'] = qid\n\n return out\n\n @torch.no_grad()\n def predict_answers(\n self,\n samples,\n num_beams=5,\n inference_method=\"generate\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n prompt=\"\",\n length_penalty=-1,\n **kwargs\n ):\n if isinstance(samples[\"qa_input\"], str):\n samples[\"qa_input\"] = [samples[\"qa_input\"]]\n \n text_input = samples[\"qa_input\"]\n samples[\"prompt\"] = text_input\n\n output_text = self.generate(\n samples,\n num_beams=num_beams,\n max_length=max_len,\n min_length=min_len,\n length_penalty=length_penalty\n )['output_text']\n\n if \"apply_lemmatizer\" in samples.keys() and samples[\"apply_lemmatizer\"]:\n output_text = self._lemmatize(output_text)\n \n output_text = [o if o != \"\" else \"unanswerable\" for o in output_text]\n return output_text\n \n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema._lemmatize","uri":"program://CREMA/function/lavis.models.crema_models.crema._lemmatize#L686-L700","kind":"function","name":"_lemmatize","path":"lavis/models/crema_models/crema.py","language":"python","start_line":686,"end_line":700,"context_start_line":666,"context_end_line":720,"code":" if isinstance(samples[\"qa_input\"], str):\n samples[\"qa_input\"] = [samples[\"qa_input\"]]\n \n text_input = samples[\"qa_input\"]\n samples[\"prompt\"] = text_input\n\n output_text = self.generate(\n samples,\n num_beams=num_beams,\n max_length=max_len,\n min_length=min_len,\n length_penalty=length_penalty\n )['output_text']\n\n if \"apply_lemmatizer\" in samples.keys() and samples[\"apply_lemmatizer\"]:\n output_text = self._lemmatize(output_text)\n \n output_text = [o if o != \"\" else \"unanswerable\" for o in output_text]\n return output_text\n \n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.lemmatizer","uri":"program://CREMA/function/lavis.models.crema_models.crema.lemmatizer#L703-L721","kind":"function","name":"lemmatizer","path":"lavis/models/crema_models/crema.py","language":"python","start_line":703,"end_line":721,"context_start_line":683,"context_end_line":741,"code":" output_text = [o if o != \"\" else \"unanswerable\" for o in output_text]\n return output_text\n \n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n frame_num = cfg.get(\"frame_num\", 8)\n answer_num = cfg.get(\"answer_num\", 5) \n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n task = cfg.get(\"task\", 'train')\n modalities = cfg.get(\"modalities\", 'rgb')\n downstream_task = cfg.get(\"downstream_task\", 'mcqa')","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.from_config","uri":"program://CREMA/function/lavis.models.crema_models.crema.from_config#L724-L772","kind":"function","name":"from_config","path":"lavis/models/crema_models/crema.py","language":"python","start_line":724,"end_line":772,"context_start_line":704,"context_end_line":772,"code":" if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )\n exit(1)\n\n return self._lemmatizer\n\n @classmethod\n def from_config(cls, cfg):\n img_size = cfg.get(\"image_size\")\n num_query_token = cfg.get(\"num_query_token\")\n t5_model = cfg.get(\"t5_model\")\n\n drop_path_rate = cfg.get(\"drop_path_rate\", 0)\n use_grad_checkpoint = cfg.get(\"use_grad_checkpoint\", False)\n vit_precision = cfg.get(\"vit_precision\", \"fp16\")\n freeze_vit = cfg.get(\"freeze_vit\", True)\n\n prompt = cfg.get(\"prompt\", \"\")\n max_txt_len = cfg.get(\"max_txt_len\", 32)\n frame_num = cfg.get(\"frame_num\", 8)\n answer_num = cfg.get(\"answer_num\", 5) \n apply_lemmatizer = cfg.get(\"apply_lemmatizer\", False)\n task = cfg.get(\"task\", 'train')\n modalities = cfg.get(\"modalities\", 'rgb')\n downstream_task = cfg.get(\"downstream_task\", 'mcqa')\n\n lora_rank = cfg.get(\"lora_rank\", 64)\n lora_layer = cfg.get(\"lora_layer\", None)\n lora_dropout = cfg.get(\"lora_dropout\", 0.1)\n\n model = cls(\n img_size=img_size,\n drop_path_rate=drop_path_rate,\n use_grad_checkpoint=use_grad_checkpoint,\n vit_precision=vit_precision,\n freeze_vit=freeze_vit,\n num_query_token=num_query_token,\n t5_model=t5_model,\n prompt=prompt,\n max_txt_len=max_txt_len,\n apply_lemmatizer=apply_lemmatizer,\n frame_num=frame_num,\n answer_num=answer_num,\n task=task,\n downstream_task=downstream_task,\n modalities=modalities,\n lora_rank=lora_rank,\n lora_layer=lora_layer,\n lora_dropout=lora_dropout,\n )\n \n model.load_checkpoint_from_config(cfg)\n model.load_lora('')\n print_trainable_parameters(model)\n\n return model","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.crema_models.crema.apply","uri":"program://CREMA/function/lavis.models.crema_models.crema.apply#L687-L698","kind":"function","name":"apply","path":"lavis/models/crema_models/crema.py","language":"python","start_line":687,"end_line":698,"context_start_line":667,"context_end_line":718,"code":" samples[\"qa_input\"] = [samples[\"qa_input\"]]\n \n text_input = samples[\"qa_input\"]\n samples[\"prompt\"] = text_input\n\n output_text = self.generate(\n samples,\n num_beams=num_beams,\n max_length=max_len,\n min_length=min_len,\n length_penalty=length_penalty\n )['output_text']\n\n if \"apply_lemmatizer\" in samples.keys() and samples[\"apply_lemmatizer\"]:\n output_text = self._lemmatize(output_text)\n \n output_text = [o if o != \"\" else \"unanswerable\" for o in output_text]\n return output_text\n \n def _lemmatize(self, answers):\n def apply(answer):\n doc = self.lemmatizer(answer)\n\n words = []\n for token in doc:\n if token.pos_ in [\"NOUN\", \"VERB\"]:\n words.append(token.lemma_)\n else:\n words.append(token.text)\n answer = \" \".join(words)\n\n return answer\n\n return [apply(answer) for answer in answers]\n\n @property\n def lemmatizer(self):\n if self._lemmatizer is None:\n try:\n import spacy\n\n self._lemmatizer = spacy.load(\"en_core_web_sm\")\n except ImportError:\n logging.error(\n \"\"\"\n Please install spacy and en_core_web_sm model to apply lemmatization.\n python -m spacy download en_core_web_sm\n OR\n import spacy.cli\n spacy.cli.download(\"en_core_web_sm\")\n \"\"\"\n )","source_hash":"e1d7c620f2da9e8760b699640dca1337d01b85ed61630198ed80a953dfacd74f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid","uri":"program://CREMA/module/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid#L1-L87","kind":"module","name":"lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":1,"end_line":87,"context_start_line":1,"context_end_line":87,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on facebookresearch code base\n https://github.com/facebookresearch/FiD\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom lavis.common.utils import get_abs_path\nfrom transformers import T5Config, T5Tokenizer, T5ForConditionalGeneration\n\n\n@registry.register_model(\"pnp_unifiedqav2_fid\")\nclass PNPUnifiedQAv2FiD(T5ForConditionalGeneration, BaseModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {}\n\n def __init__(self, config, model_path):\n super().__init__(config)\n \n self.tokenizer = T5Tokenizer.from_pretrained(model_path)\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n if input_ids != None:\n if input_ids.dim() == 3:\n self.encoder.num_contexts = input_ids.size(1)\n input_ids = input_ids.view(input_ids.size(0), -1)\n if attention_mask != None:\n attention_mask = attention_mask.view(attention_mask.size(0), -1)\n\n return super().forward(\n input_ids=input_ids,\n attention_mask=attention_mask,\n **kwargs\n )\n\n def generate(self, input_ids, attention_mask, num_beams=1, min_length=0, max_length=20):\n self.encoder.num_contexts = input_ids.size(1)\n\n return super().generate(\n input_ids=input_ids.view(input_ids.size(0), -1),\n attention_mask=attention_mask.view(attention_mask.size(0), -1),\n num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n\n return outputs","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.PNPUnifiedQAv2FiD","uri":"program://CREMA/class/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.PNPUnifiedQAv2FiD#L20-L66","kind":"class","name":"PNPUnifiedQAv2FiD","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":20,"end_line":66,"context_start_line":1,"context_end_line":86,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on facebookresearch code base\n https://github.com/facebookresearch/FiD\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom lavis.common.utils import get_abs_path\nfrom transformers import T5Config, T5Tokenizer, T5ForConditionalGeneration\n\n\n@registry.register_model(\"pnp_unifiedqav2_fid\")\nclass PNPUnifiedQAv2FiD(T5ForConditionalGeneration, BaseModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {}\n\n def __init__(self, config, model_path):\n super().__init__(config)\n \n self.tokenizer = T5Tokenizer.from_pretrained(model_path)\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n if input_ids != None:\n if input_ids.dim() == 3:\n self.encoder.num_contexts = input_ids.size(1)\n input_ids = input_ids.view(input_ids.size(0), -1)\n if attention_mask != None:\n attention_mask = attention_mask.view(attention_mask.size(0), -1)\n\n return super().forward(\n input_ids=input_ids,\n attention_mask=attention_mask,\n **kwargs\n )\n\n def generate(self, input_ids, attention_mask, num_beams=1, min_length=0, max_length=20):\n self.encoder.num_contexts = input_ids.size(1)\n\n return super().generate(\n input_ids=input_ids.view(input_ids.size(0), -1),\n attention_mask=attention_mask.view(attention_mask.size(0), -1),\n num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.T5EncoderWrapper","uri":"program://CREMA/class/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.T5EncoderWrapper#L69-L87","kind":"class","name":"T5EncoderWrapper","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":69,"end_line":87,"context_start_line":49,"context_end_line":87,"code":" num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n\n return outputs","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.__init__","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.__init__#L71-L77","kind":"function","name":"__init__","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":71,"end_line":77,"context_start_line":51,"context_end_line":87,"code":" max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n\n return outputs","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.forward","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.forward#L79-L87","kind":"function","name":"forward","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":79,"end_line":87,"context_start_line":59,"context_end_line":87,"code":" def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n\n return outputs","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.generate","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.generate#L43-L52","kind":"function","name":"generate","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":43,"end_line":52,"context_start_line":23,"context_end_line":72,"code":"\n def __init__(self, config, model_path):\n super().__init__(config)\n \n self.tokenizer = T5Tokenizer.from_pretrained(model_path)\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n if input_ids != None:\n if input_ids.dim() == 3:\n self.encoder.num_contexts = input_ids.size(1)\n input_ids = input_ids.view(input_ids.size(0), -1)\n if attention_mask != None:\n attention_mask = attention_mask.view(attention_mask.size(0), -1)\n\n return super().forward(\n input_ids=input_ids,\n attention_mask=attention_mask,\n **kwargs\n )\n\n def generate(self, input_ids, attention_mask, num_beams=1, min_length=0, max_length=20):\n self.encoder.num_contexts = input_ids.size(1)\n\n return super().generate(\n input_ids=input_ids.view(input_ids.size(0), -1),\n attention_mask=attention_mask.view(attention_mask.size(0), -1),\n num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.load_unifiedqa","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.load_unifiedqa#L54-L56","kind":"function","name":"load_unifiedqa","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":54,"end_line":56,"context_start_line":34,"context_end_line":76,"code":" if attention_mask != None:\n attention_mask = attention_mask.view(attention_mask.size(0), -1)\n\n return super().forward(\n input_ids=input_ids,\n attention_mask=attention_mask,\n **kwargs\n )\n\n def generate(self, input_ids, attention_mask, num_beams=1, min_length=0, max_length=20):\n self.encoder.num_contexts = input_ids.size(1)\n\n return super().generate(\n input_ids=input_ids.view(input_ids.size(0), -1),\n attention_mask=attention_mask.view(attention_mask.size(0), -1),\n num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.from_config","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_unifiedqav2_fid.from_config#L59-L66","kind":"function","name":"from_config","path":"lavis/models/pnp_vqa_models/pnp_unifiedqav2_fid.py","language":"python","start_line":59,"end_line":66,"context_start_line":39,"context_end_line":86,"code":" attention_mask=attention_mask,\n **kwargs\n )\n\n def generate(self, input_ids, attention_mask, num_beams=1, min_length=0, max_length=20):\n self.encoder.num_contexts = input_ids.size(1)\n\n return super().generate(\n input_ids=input_ids.view(input_ids.size(0), -1),\n attention_mask=attention_mask.view(attention_mask.size(0), -1),\n num_beams=num_beams,\n min_length=min_length,\n max_length=max_length\n )\n\n def load_unifiedqa(self, state_dict):\n self.load_state_dict(state_dict)\n self.encoder = T5EncoderWrapper(self.encoder)\n\n @classmethod\n def from_config(cls, cfg):\n model_path = cfg.get('pretrained')\n t5_config_path = get_abs_path(cfg.get(\"t5_config_path\"))\n t5_config = T5Config.from_json_file(t5_config_path)\n model = cls(t5_config, model_path)\n model.load_unifiedqa(T5ForConditionalGeneration.from_pretrained(model_path).state_dict())\n\n return model\n\n\nclass T5EncoderWrapper(torch.nn.Module):\n\n def __init__(self, encoder):\n super().__init__()\n\n self.encoder = encoder\n self.block = self.encoder.block\n self.parallelize = self.encoder.parallelize\n self.main_input_name = encoder.main_input_name\n\n def forward(self, input_ids=None, attention_mask=None, **kwargs):\n bsz, total_length = input_ids.shape\n context_length = total_length // self.num_contexts\n input_ids = input_ids.view(bsz*self.num_contexts, context_length)\n attention_mask = attention_mask.view(bsz*self.num_contexts, context_length)\n outputs = self.encoder(input_ids, attention_mask, **kwargs)\n outputs = (outputs[0].view(bsz, self.num_contexts*context_length, -1), ) + outputs[1:]\n","source_hash":"446ae61c8e2fbb09e560acab11819e17e83a25c2575859cca4f070c2a5e456a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa","uri":"program://CREMA/module/lavis.models.pnp_vqa_models.pnp_vqa#L1-L340","kind":"module","name":"lavis.models.pnp_vqa_models.pnp_vqa","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":1,"end_line":340,"context_start_line":1,"context_end_line":340,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom itertools import chain\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom torch.nn import CrossEntropyLoss, MSELoss\nfrom transformers import T5ForConditionalGeneration\nfrom lavis.models.pnp_vqa_models import prepare_qa_input\nfrom lavis.models.blip_models.blip_image_text_matching import compute_gradcam\nfrom transformers.modeling_outputs import CausalLMOutputWithCrossAttentions\n\n\n@registry.register_model(\"pnp_vqa\")\nclass PNPVQA(BaseModel):\n \"\"\"\n PNPVQA model consists of three submodels for zero-shot VQA:\n 1. Image-questioning matching model\n 2. Image captioning model\n 3. Question answering model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"pnp_vqa\", \"base\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"large\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"3b\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/pnp-vqa/pnp_vqa_base.yaml\",\n \"large\": \"configs/models/pnp-vqa/pnp_vqa_large.yaml\",\n \"3b\": \"configs/models/pnp-vqa/pnp_vqa_3b.yaml\",\n }\n\n def __init__(self, image_question_matching_model, image_captioning_model,\n question_answering_model, offload_model=False):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_answering_model = question_answering_model\n self.offload_model = offload_model\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples['image']\n question = [text.strip('?') for text in samples['text_input']]\n tokenized_text = self.image_question_matching_model.tokenizer(question, padding='longest', truncation=True,\n return_tensors=\"pt\").to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num)\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples['gradcams'] = torch.stack(gradcams).reshape(samples['image'].size(0), -1)\n\n return samples\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = torch.multinomial(samples['gradcams'].to(self.image_captioning_model.device),\n num_patches).reshape(encoder_out.size(0), -1) + 1\n patch_id = patch_id.sort(dim=1).values.unsqueeze(-1).expand(-1, -1, encoder_out.size(2))\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(stacked, start_dim=0, end_dim=1) #(bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(self.image_captioning_model.device)\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(prompt,\n return_tensors=\"pt\").to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs)\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n\n for counter, output in enumerate(outputs):\n ind = counter//num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt):]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n if len(overlap_caption) == 0:\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples['captions'] = captions\n\n return samples\n\n def forward_qa(\n self,\n samples,\n num_beams=1,\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=100,\n num_captions_fid=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n - question_captions (nested list): A nested list of concatenated strings of questions and captions\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n\n Returns:\n List: A list of strings, each string is an answer.\n \"\"\"\n prepare_qa_input(samples, num_captions=num_captions, num_captions_fid=num_captions_fid)\n\n pred_answers = []\n question_captions = samples['question_captions']\n question_captions_chunk = [question_captions[i:i + internal_bsz_fid]\n for i in range(0, len(question_captions), internal_bsz_fid)]\n question_captions_chunk = list(chain(*question_captions_chunk))\n\n for question_caption in question_captions_chunk:\n question_caption_input = self.question_answering_model.tokenizer(question_caption, padding='longest',\n truncation=True, return_tensors=\"pt\").to(self.question_answering_model.device)\n\n question_caption_input.input_ids = question_caption_input.input_ids.reshape(\n internal_bsz_fid, -1, question_caption_input.input_ids.size(1))\n question_caption_input.attention_mask = question_caption_input.attention_mask.reshape(\n internal_bsz_fid, -1, question_caption_input.attention_mask.size(1))\n\n outputs = self.question_answering_model.generate(input_ids=question_caption_input.input_ids,\n attention_mask=question_caption_input.attention_mask,\n num_beams=num_beams,\n min_length=min_len,\n max_length=max_len,\n )\n\n for output in outputs:\n pred_answer = self.question_answering_model.tokenizer.decode(output, skip_special_tokens=True)\n pred_answers.append(pred_answer)\n\n return pred_answers\n\n def predict_answers(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=50,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_patches (int): Number of patches sampled for each image.\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n List: A list of strings, each string is an answer.\n gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n assert inference_method in [\n \"generate\",\n ], \"Inference method must be 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n samples = self.forward_itm(samples, block_num=block_num)\n\n samples = self.forward_cap(samples,\n cap_max_length=cap_max_length,\n cap_min_length=cap_min_length,\n top_k=top_k,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n num_captions=num_captions,\n num_patches=num_patches)\n\n if self.offload_model:\n samples['image'] = samples['image'].to('cpu')\n self.image_question_matching_model.to('cpu')\n self.image_captioning_model.to('cpu')\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid)\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples['captions'], samples['gradcams']\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n qa_config = model_config.question_answering_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n qa_cls = registry.get_model_class(qa_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n question_answering_model = qa_cls.from_config(qa_config)\n\n model = cls(image_question_matching_model=image_question_matching_model,\n image_captioning_model=image_captioning_model,\n question_answering_model=question_answering_model,\n offload_model= True if model_config.model_type == '3b' else False,\n )\n\n return model","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.PNPVQA","uri":"program://CREMA/class/lavis.models.pnp_vqa_models.pnp_vqa.PNPVQA#L21-L340","kind":"class","name":"PNPVQA","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":21,"end_line":340,"context_start_line":1,"context_end_line":340,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom itertools import chain\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom torch.nn import CrossEntropyLoss, MSELoss\nfrom transformers import T5ForConditionalGeneration\nfrom lavis.models.pnp_vqa_models import prepare_qa_input\nfrom lavis.models.blip_models.blip_image_text_matching import compute_gradcam\nfrom transformers.modeling_outputs import CausalLMOutputWithCrossAttentions\n\n\n@registry.register_model(\"pnp_vqa\")\nclass PNPVQA(BaseModel):\n \"\"\"\n PNPVQA model consists of three submodels for zero-shot VQA:\n 1. Image-questioning matching model\n 2. Image captioning model\n 3. Question answering model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"pnp_vqa\", \"base\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"large\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"3b\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/pnp-vqa/pnp_vqa_base.yaml\",\n \"large\": \"configs/models/pnp-vqa/pnp_vqa_large.yaml\",\n \"3b\": \"configs/models/pnp-vqa/pnp_vqa_3b.yaml\",\n }\n\n def __init__(self, image_question_matching_model, image_captioning_model,\n question_answering_model, offload_model=False):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_answering_model = question_answering_model\n self.offload_model = offload_model\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples['image']\n question = [text.strip('?') for text in samples['text_input']]\n tokenized_text = self.image_question_matching_model.tokenizer(question, padding='longest', truncation=True,\n return_tensors=\"pt\").to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num)\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples['gradcams'] = torch.stack(gradcams).reshape(samples['image'].size(0), -1)\n\n return samples\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = torch.multinomial(samples['gradcams'].to(self.image_captioning_model.device),\n num_patches).reshape(encoder_out.size(0), -1) + 1\n patch_id = patch_id.sort(dim=1).values.unsqueeze(-1).expand(-1, -1, encoder_out.size(2))\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(stacked, start_dim=0, end_dim=1) #(bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(self.image_captioning_model.device)\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(prompt,\n return_tensors=\"pt\").to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs)\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n\n for counter, output in enumerate(outputs):\n ind = counter//num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt):]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n if len(overlap_caption) == 0:\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples['captions'] = captions\n\n return samples\n\n def forward_qa(\n self,\n samples,\n num_beams=1,\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=100,\n num_captions_fid=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n - question_captions (nested list): A nested list of concatenated strings of questions and captions\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n\n Returns:\n List: A list of strings, each string is an answer.\n \"\"\"\n prepare_qa_input(samples, num_captions=num_captions, num_captions_fid=num_captions_fid)\n\n pred_answers = []\n question_captions = samples['question_captions']\n question_captions_chunk = [question_captions[i:i + internal_bsz_fid]\n for i in range(0, len(question_captions), internal_bsz_fid)]\n question_captions_chunk = list(chain(*question_captions_chunk))\n\n for question_caption in question_captions_chunk:\n question_caption_input = self.question_answering_model.tokenizer(question_caption, padding='longest',\n truncation=True, return_tensors=\"pt\").to(self.question_answering_model.device)\n\n question_caption_input.input_ids = question_caption_input.input_ids.reshape(\n internal_bsz_fid, -1, question_caption_input.input_ids.size(1))\n question_caption_input.attention_mask = question_caption_input.attention_mask.reshape(\n internal_bsz_fid, -1, question_caption_input.attention_mask.size(1))\n\n outputs = self.question_answering_model.generate(input_ids=question_caption_input.input_ids,\n attention_mask=question_caption_input.attention_mask,\n num_beams=num_beams,\n min_length=min_len,\n max_length=max_len,\n )\n\n for output in outputs:\n pred_answer = self.question_answering_model.tokenizer.decode(output, skip_special_tokens=True)\n pred_answers.append(pred_answer)\n\n return pred_answers\n\n def predict_answers(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=50,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_patches (int): Number of patches sampled for each image.\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n List: A list of strings, each string is an answer.\n gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n assert inference_method in [\n \"generate\",\n ], \"Inference method must be 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n samples = self.forward_itm(samples, block_num=block_num)\n\n samples = self.forward_cap(samples,\n cap_max_length=cap_max_length,\n cap_min_length=cap_min_length,\n top_k=top_k,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n num_captions=num_captions,\n num_patches=num_patches)\n\n if self.offload_model:\n samples['image'] = samples['image'].to('cpu')\n self.image_question_matching_model.to('cpu')\n self.image_captioning_model.to('cpu')\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid)\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples['captions'], samples['gradcams']\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n qa_config = model_config.question_answering_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n qa_cls = registry.get_model_class(qa_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n question_answering_model = qa_cls.from_config(qa_config)\n\n model = cls(image_question_matching_model=image_question_matching_model,\n image_captioning_model=image_captioning_model,\n question_answering_model=question_answering_model,\n offload_model= True if model_config.model_type == '3b' else False,\n )\n\n return model","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.__init__","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.__init__#L45-L52","kind":"function","name":"__init__","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":45,"end_line":52,"context_start_line":25,"context_end_line":72,"code":" 2. Image captioning model\n 3. Question answering model\n\n Supported model types:\n - base: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-base)\n - large: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-large)\n - 3b: BLIPITM, BLIPCaption, PNPUnifiedQAv2FiD (t5-3b)\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"pnp_vqa\", \"base\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"large\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"3b\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/pnp-vqa/pnp_vqa_base.yaml\",\n \"large\": \"configs/models/pnp-vqa/pnp_vqa_large.yaml\",\n \"3b\": \"configs/models/pnp-vqa/pnp_vqa_3b.yaml\",\n }\n\n def __init__(self, image_question_matching_model, image_captioning_model,\n question_answering_model, offload_model=False):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_answering_model = question_answering_model\n self.offload_model = offload_model\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples['image']\n question = [text.strip('?') for text in samples['text_input']]\n tokenized_text = self.image_question_matching_model.tokenizer(question, padding='longest', truncation=True,\n return_tensors=\"pt\").to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.forward_itm","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.forward_itm#L54-L82","kind":"function","name":"forward_itm","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":54,"end_line":82,"context_start_line":34,"context_end_line":102,"code":" >>> from lavis.models import load_model\n >>> model = load_model(\"pnp_vqa\", \"base\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"large\", is_eval=True)\n >>> model = load_model(\"pnp_vqa\", \"3b\", is_eval=True)\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/pnp-vqa/pnp_vqa_base.yaml\",\n \"large\": \"configs/models/pnp-vqa/pnp_vqa_large.yaml\",\n \"3b\": \"configs/models/pnp-vqa/pnp_vqa_3b.yaml\",\n }\n\n def __init__(self, image_question_matching_model, image_captioning_model,\n question_answering_model, offload_model=False):\n super().__init__()\n\n self.image_question_matching_model = image_question_matching_model\n self.image_captioning_model = image_captioning_model\n self.question_answering_model = question_answering_model\n self.offload_model = offload_model\n\n def forward_itm(self, samples, block_num=7):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples['image']\n question = [text.strip('?') for text in samples['text_input']]\n tokenized_text = self.image_question_matching_model.tokenizer(question, padding='longest', truncation=True,\n return_tensors=\"pt\").to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num)\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples['gradcams'] = torch.stack(gradcams).reshape(samples['image'].size(0), -1)\n\n return samples\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.forward_cap","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.forward_cap#L84-L172","kind":"function","name":"forward_cap","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":84,"end_line":172,"context_start_line":64,"context_end_line":192,"code":" - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n \"\"\"\n image = samples['image']\n question = [text.strip('?') for text in samples['text_input']]\n tokenized_text = self.image_question_matching_model.tokenizer(question, padding='longest', truncation=True,\n return_tensors=\"pt\").to(self.image_question_matching_model.device)\n with torch.set_grad_enabled(True):\n gradcams, _ = compute_gradcam(model=self.image_question_matching_model,\n visual_input=image,\n text_input=question,\n tokenized_text=tokenized_text,\n block_num=block_num)\n\n gradcams = [gradcam_[1] for gradcam_ in gradcams]\n samples['gradcams'] = torch.stack(gradcams).reshape(samples['image'].size(0), -1)\n\n return samples\n\n def forward_cap(\n self,\n samples,\n cap_max_length=20,\n cap_min_length=0,\n top_p=1,\n top_k=50,\n repetition_penalty=1.0,\n num_captions=100,\n num_patches=20,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions generated for each image.\n num_patches (int): Number of patches sampled for each image.\n\n Returns:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n encoder_out = self.image_captioning_model.forward_encoder(samples)\n captions = [[] for _ in range(encoder_out.size(0))]\n\n min_num_captions = 0\n\n while min_num_captions < num_captions:\n encoder_out_samples = []\n for i in range(num_captions):\n patch_id = torch.multinomial(samples['gradcams'].to(self.image_captioning_model.device),\n num_patches).reshape(encoder_out.size(0), -1) + 1\n patch_id = patch_id.sort(dim=1).values.unsqueeze(-1).expand(-1, -1, encoder_out.size(2))\n encoder_out_sample = torch.gather(encoder_out, 1, patch_id)\n encoder_out_samples.append(encoder_out_sample)\n\n stacked = torch.stack(encoder_out_samples, dim=1)\n image_embeds = torch.flatten(stacked, start_dim=0, end_dim=1) #(bsz*num_seq, num_patch, dim)\n\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(self.image_captioning_model.device)\n model_kwargs = {\n \"encoder_hidden_states\": image_embeds,\n \"encoder_attention_mask\": image_atts,\n }\n\n prompt = [self.image_captioning_model.prompt] * image_embeds.size(0)\n prompt = self.image_captioning_model.tokenizer(prompt,\n return_tensors=\"pt\").to(self.image_captioning_model.device)\n prompt.input_ids[:, 0] = self.image_captioning_model.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n decoder_out = self.image_captioning_model.text_decoder.generate(\n input_ids=prompt.input_ids,\n max_length=cap_max_length,\n min_length=cap_min_length,\n do_sample=True,\n top_p=top_p,\n top_k=top_k,\n num_return_sequences=1,\n eos_token_id=self.image_captioning_model.tokenizer.sep_token_id,\n pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs)\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n\n for counter, output in enumerate(outputs):\n ind = counter//num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt):]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n if len(overlap_caption) == 0:\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples['captions'] = captions\n\n return samples\n\n def forward_qa(\n self,\n samples,\n num_beams=1,\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=100,\n num_captions_fid=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n - question_captions (nested list): A nested list of concatenated strings of questions and captions\n num_beams (int): Number of beams for beam search. 1 means no beam search.","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.forward_qa","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.forward_qa#L174-L230","kind":"function","name":"forward_qa","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":174,"end_line":230,"context_start_line":154,"context_end_line":250,"code":" pad_token_id=self.image_captioning_model.tokenizer.pad_token_id,\n repetition_penalty=repetition_penalty,\n **model_kwargs)\n\n outputs = self.image_captioning_model.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n\n for counter, output in enumerate(outputs):\n ind = counter//num_captions\n if len(captions[ind]) < num_captions:\n caption = output[len(self.image_captioning_model.prompt):]\n overlap_caption = [1 for caps in captions[ind] if caption in caps]\n if len(overlap_caption) == 0:\n captions[ind].append(caption)\n\n min_num_captions = min([len(i) for i in captions])\n\n samples['captions'] = captions\n\n return samples\n\n def forward_qa(\n self,\n samples,\n num_beams=1,\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=100,\n num_captions_fid=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size\n - gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n - captions (nested list): A nested list of strings of total length batch_size * num_captions\n - question_captions (nested list): A nested list of concatenated strings of questions and captions\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n\n Returns:\n List: A list of strings, each string is an answer.\n \"\"\"\n prepare_qa_input(samples, num_captions=num_captions, num_captions_fid=num_captions_fid)\n\n pred_answers = []\n question_captions = samples['question_captions']\n question_captions_chunk = [question_captions[i:i + internal_bsz_fid]\n for i in range(0, len(question_captions), internal_bsz_fid)]\n question_captions_chunk = list(chain(*question_captions_chunk))\n\n for question_caption in question_captions_chunk:\n question_caption_input = self.question_answering_model.tokenizer(question_caption, padding='longest',\n truncation=True, return_tensors=\"pt\").to(self.question_answering_model.device)\n\n question_caption_input.input_ids = question_caption_input.input_ids.reshape(\n internal_bsz_fid, -1, question_caption_input.input_ids.size(1))\n question_caption_input.attention_mask = question_caption_input.attention_mask.reshape(\n internal_bsz_fid, -1, question_caption_input.attention_mask.size(1))\n\n outputs = self.question_answering_model.generate(input_ids=question_caption_input.input_ids,\n attention_mask=question_caption_input.attention_mask,\n num_beams=num_beams,\n min_length=min_len,\n max_length=max_len,\n )\n\n for output in outputs:\n pred_answer = self.question_answering_model.tokenizer.decode(output, skip_special_tokens=True)\n pred_answers.append(pred_answer)\n\n return pred_answers\n\n def predict_answers(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=50,\n block_num=7,\n ):\n \"\"\"","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.predict_answers","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.predict_answers#L232-L318","kind":"function","name":"predict_answers","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":232,"end_line":318,"context_start_line":212,"context_end_line":338,"code":" truncation=True, return_tensors=\"pt\").to(self.question_answering_model.device)\n\n question_caption_input.input_ids = question_caption_input.input_ids.reshape(\n internal_bsz_fid, -1, question_caption_input.input_ids.size(1))\n question_caption_input.attention_mask = question_caption_input.attention_mask.reshape(\n internal_bsz_fid, -1, question_caption_input.attention_mask.size(1))\n\n outputs = self.question_answering_model.generate(input_ids=question_caption_input.input_ids,\n attention_mask=question_caption_input.attention_mask,\n num_beams=num_beams,\n min_length=min_len,\n max_length=max_len,\n )\n\n for output in outputs:\n pred_answer = self.question_answering_model.tokenizer.decode(output, skip_special_tokens=True)\n pred_answers.append(pred_answer)\n\n return pred_answers\n\n def predict_answers(\n self,\n samples,\n num_beams=1,\n inference_method=\"generate\",\n max_len=20,\n min_len=0,\n internal_bsz_fid=1,\n num_captions=50,\n num_captions_fid=1,\n cap_max_length=20,\n cap_min_length=10,\n top_k=50,\n top_p=1,\n repetition_penalty=1,\n num_patches=50,\n block_num=7,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. Must be \"generate\". The model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n internal_bsz_fid (int): Internal batch size when using FiD decoding.\n num_captions (int): Number of captions generated for each image.\n num_captions_fid (int): Number of captions concatenated with a question during FiD decoding.\n cap_max_length (int): The maximum length of the caption to be generated.\n cap_min_length (int): The minimum length of the caption to be generated.\n top_k (float): The number of the highest probability tokens for top-k sampling.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_patches (int): Number of patches sampled for each image.\n block_num (int): The index of cross-attention block for gradcam computation.\n\n Returns:\n List: A list of strings, each string is an answer.\n gradcams (torch.Tensor): A tensor of shape (batch_size, H*W)\n captions (nested list): A nested list of strings of total length batch_size * num_captions\n \"\"\"\n assert inference_method in [\n \"generate\",\n ], \"Inference method must be 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n samples = self.forward_itm(samples, block_num=block_num)\n\n samples = self.forward_cap(samples,\n cap_max_length=cap_max_length,\n cap_min_length=cap_min_length,\n top_k=top_k,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n num_captions=num_captions,\n num_patches=num_patches)\n\n if self.offload_model:\n samples['image'] = samples['image'].to('cpu')\n self.image_question_matching_model.to('cpu')\n self.image_captioning_model.to('cpu')\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid)\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples['captions'], samples['gradcams']\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n qa_config = model_config.question_answering_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n qa_cls = registry.get_model_class(qa_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n question_answering_model = qa_cls.from_config(qa_config)\n\n model = cls(image_question_matching_model=image_question_matching_model,\n image_captioning_model=image_captioning_model,\n question_answering_model=question_answering_model,\n offload_model= True if model_config.model_type == '3b' else False,\n )","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.pnp_vqa_models.pnp_vqa.from_config","uri":"program://CREMA/function/lavis.models.pnp_vqa_models.pnp_vqa.from_config#L321-L340","kind":"function","name":"from_config","path":"lavis/models/pnp_vqa_models/pnp_vqa.py","language":"python","start_line":321,"end_line":340,"context_start_line":301,"context_end_line":340,"code":" samples['image'] = samples['image'].to('cpu')\n self.image_question_matching_model.to('cpu')\n self.image_captioning_model.to('cpu')\n torch.cuda.empty_cache()\n\n pred_answers = self.forward_qa(samples,\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n internal_bsz_fid=internal_bsz_fid,\n num_captions=num_captions,\n num_captions_fid=num_captions_fid)\n\n if self.offload_model:\n self.image_question_matching_model.to(self.question_answering_model.device)\n self.image_captioning_model.to(self.question_answering_model.device)\n\n return pred_answers, samples['captions'], samples['gradcams']\n\n @classmethod\n def from_config(cls, model_config):\n itm_config = model_config.image_question_matching_model\n cap_config = model_config.image_captioning_model\n qa_config = model_config.question_answering_model\n\n itm_cls = registry.get_model_class(itm_config.arch)\n cap_cls = registry.get_model_class(cap_config.arch)\n qa_cls = registry.get_model_class(qa_config.arch)\n\n image_question_matching_model = itm_cls.from_config(itm_config)\n image_captioning_model = cap_cls.from_config(cap_config)\n question_answering_model = qa_cls.from_config(qa_config)\n\n model = cls(image_question_matching_model=image_question_matching_model,\n image_captioning_model=image_captioning_model,\n question_answering_model=question_answering_model,\n offload_model= True if model_config.model_type == '3b' else False,\n )\n\n return model","source_hash":"aa85db45f05df71d5d0eab9d0e606522c5d891896b4eec4246aa5ad8c4d5c0ea","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit","uri":"program://CREMA/module/lavis.models.timesformer.vit#L1-L634","kind":"module","name":"lavis.models.timesformer.vit","path":"lavis/models/timesformer/vit.py","language":"python","start_line":1,"end_line":634,"context_start_line":1,"context_end_line":634,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.\n# Copyright 2020 Ross Wightman\n# Modified Model definition\n\nimport logging\nfrom functools import partial\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils\nimport torch.utils.checkpoint\nfrom einops import rearrange\nfrom fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper\n\nfrom .helpers import load_pretrained, load_pretrained_imagenet, load_pretrained_kinetics\nfrom .vit_utils import (\n IMAGENET_DEFAULT_MEAN,\n IMAGENET_DEFAULT_STD,\n DropPath,\n to_2tuple,\n trunc_normal_,\n)\n\n\ndef _cfg(url=\"\", **kwargs):\n return {\n \"url\": url,\n \"num_classes\": 1000,\n \"input_size\": (3, 224, 224),\n \"pool_size\": None,\n \"crop_pct\": 0.9,\n \"interpolation\": \"bicubic\",\n \"mean\": IMAGENET_DEFAULT_MEAN,\n \"std\": IMAGENET_DEFAULT_STD,\n \"first_conv\": \"patch_embed.proj\",\n \"classifier\": \"head\",\n **kwargs,\n }\n\n\ndefault_cfgs = {\n \"vit_base_patch16_224\": _cfg(\n url=\"https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth\",\n mean=(0.5, 0.5, 0.5),\n std=(0.5, 0.5, 0.5),\n ),\n}\n\n\nclass Mlp(nn.Module):\n def __init__(\n self,\n in_features,\n hidden_features=None,\n out_features=None,\n act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n with_qkv=True,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n self.scale = qk_scale or head_dim**-0.5\n self.with_qkv = with_qkv\n if self.with_qkv:\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_drop = nn.Dropout(attn_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n if self.with_qkv:\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = qkv[0], qkv[1], qkv[2]\n else:\n qkv = x.reshape(B, N, self.num_heads, C // self.num_heads).permute(\n 0, 2, 1, 3\n )\n q, k, v = qkv, qkv, qkv\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n if self.with_qkv:\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n layer_num,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.1,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n attention_type=\"divided_space_time\",\n use_grad_checkpointing=False,\n ):\n super().__init__()\n self.attention_type = attention_type\n assert attention_type in [\n \"divided_space_time\",\n \"space_only\",\n \"joint_space_time\",\n ]\n\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n\n # Temporal Attention Parameters\n if self.attention_type == \"divided_space_time\":\n self.temporal_norm1 = norm_layer(dim)\n self.temporal_attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n self.temporal_fc = nn.Linear(dim, dim)\n\n # drop path\n self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(\n in_features=dim,\n hidden_features=mlp_hidden_dim,\n act_layer=act_layer,\n drop=drop,\n )\n\n # [dxli]\n self.layer_num = layer_num\n self.use_grad_checkpointing = use_grad_checkpointing\n\n if use_grad_checkpointing:\n self.temporal_attn = checkpoint_wrapper(self.temporal_attn)\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, B, T, W):\n num_spatial_tokens = (x.size(1) - 1) // T\n H = num_spatial_tokens // W\n\n if self.attention_type in [\"space_only\", \"joint_space_time\"]:\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n elif self.attention_type == \"divided_space_time\":\n # Temporal\n xt = x[:, 1:, :]\n xt = rearrange(xt, \"b (h w t) m -> (b h w) t m\", b=B, h=H, w=W, t=T)\n\n temporal_attn_out = self.temporal_attn(self.temporal_norm1(xt))\n\n res_temporal = self.drop_path(temporal_attn_out)\n\n res_temporal = rearrange(\n res_temporal, \"(b h w) t m -> b (h w t) m\", b=B, h=H, w=W, t=T\n )\n res_temporal = self.temporal_fc(res_temporal)\n xt = x[:, 1:, :] + res_temporal\n\n # Spatial\n init_cls_token = x[:, 0, :].unsqueeze(1)\n cls_token = init_cls_token.repeat(1, T, 1)\n cls_token = rearrange(cls_token, \"b t m -> (b t) m\", b=B, t=T).unsqueeze(1)\n xs = xt\n xs = rearrange(xs, \"b (h w t) m -> (b t) (h w) m\", b=B, h=H, w=W, t=T)\n xs = torch.cat((cls_token, xs), 1)\n\n spatial_attn_out = self.attn(self.norm1(xs))\n res_spatial = self.drop_path(spatial_attn_out)\n\n # Taking care of CLS token\n cls_token = res_spatial[:, 0, :]\n cls_token = rearrange(cls_token, \"(b t) m -> b t m\", b=B, t=T)\n # averaging for every frame\n cls_token = torch.mean(cls_token, 1, True)\n res_spatial = res_spatial[:, 1:, :]\n res_spatial = rearrange(\n res_spatial, \"(b t) (h w) m -> b (h w t) m\", b=B, h=H, w=W, t=T\n )\n res = res_spatial\n x = xt\n\n # Mlp\n x = torch.cat((init_cls_token, x), 1) + torch.cat((cls_token, res), 1)\n\n x_res = x\n\n x = self.norm2(x)\n # x = x + self.drop_path(self.mlp(self.norm2(x)))\n\n # MLP\n mlp_out = self.mlp(x)\n\n x = x_res + self.drop_path(mlp_out)\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\"Image to Patch Embedding\"\"\"\n\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(\n in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n )\n\n def forward(self, x):\n B, C, T, H, W = x.shape\n x = rearrange(x, \"b c t h w -> (b t) c h w\")\n x = self.proj(x)\n W = x.size(-1)\n x = x.flatten(2).transpose(1, 2)\n return x, T, W\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformere\"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n hybrid_backbone=None,\n norm_layer=nn.LayerNorm,\n num_frames=8,\n attention_type=\"divided_space_time\",\n dropout=0.0,\n use_grad_checkpointing=False,\n ckpt_layer=0,\n ):\n super().__init__()\n\n self.attention_type = attention_type\n self.depth = depth\n self.dropout = nn.Dropout(dropout)\n self.num_classes = num_classes\n # num_features for consistency with other models\n self.num_features = self.embed_dim = embed_dim\n self.patch_embed = PatchEmbed(\n img_size=img_size,\n patch_size=patch_size,\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n num_patches = self.patch_embed.num_patches\n\n # Positional Embeddings\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n self.pos_drop = nn.Dropout(p=drop_rate)\n if self.attention_type != \"space_only\":\n self.time_embed = nn.Parameter(torch.zeros(1, num_frames, embed_dim))\n self.time_drop = nn.Dropout(p=drop_rate)\n\n # Attention Blocks\n dpr = [\n x.item() for x in torch.linspace(0, drop_path_rate, self.depth)\n ] # stochastic depth decay rule\n self.blocks = nn.ModuleList(\n [\n Block(\n layer_num=i,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= self.depth - ckpt_layer\n ),\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n attention_type=self.attention_type,\n )\n for i in range(self.depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n # Classifier head\n self.head = (\n nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n # initialization of temporal attention weights\n if self.attention_type == \"divided_space_time\":\n i = 0\n for m in self.blocks.modules():\n m_str = str(m)\n if \"Block\" in m_str:\n if i > 0:\n nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # resizing the positional embeddings in case they don't match the input at inference\n if x.size(1) != self.pos_embed.size(1):\n pos_embed = self.pos_embed\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n P = int(other_pos_embed.size(2) ** 0.5)\n H = x.size(1) // W\n other_pos_embed = other_pos_embed.reshape(1, x.size(2), P, P)\n new_pos_embed = F.interpolate(other_pos_embed, size=(H, W), mode=\"nearest\")\n new_pos_embed = new_pos_embed.flatten(2)\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n x = x + new_pos_embed\n else:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n # Time Embeddings\n if self.attention_type != \"space_only\":\n cls_tokens = x[:B, 0, :].unsqueeze(1)\n x = x[:, 1:]\n x = rearrange(x, \"(b t) n m -> (b n) t m\", b=B, t=T)\n # Resizing time embeddings in case they don't match\n if T != self.time_embed.size(1):\n time_embed = self.time_embed.transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(T), mode=\"nearest\")\n new_time_embed = new_time_embed.transpose(1, 2)\n x = x + new_time_embed\n else:\n x = x + self.time_embed\n x = self.time_drop(x)\n x = rearrange(x, \"(b n) t m -> b (n t) m\", b=B, t=T)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # Attention blocks\n for blk in self.blocks:\n x = blk(x, B, T, W)\n\n # Predictions for space-only baseline\n if self.attention_type == \"space_only\":\n x = rearrange(x, \"(b t) n m -> b t n m\", b=B, t=T)\n x = torch.mean(x, 1) # averaging predictions for every frame\n\n x = self.norm(x)\n\n return x\n\n def forward(self, x):\n x = self.forward_features(x)\n x = self.head(x)\n return x\n\n\ndef _conv_filter(state_dict, patch_size=16):\n \"\"\"convert patch embedding weight from manual patchify + linear proj to conv\"\"\"\n out_dict = {}\n for k, v in state_dict.items():\n if \"patch_embed.proj.weight\" in k:\n if v.shape[-1] != patch_size:\n patch_size = v.shape[-1]\n v = v.reshape((v.shape[0], 3, patch_size, patch_size))\n out_dict[k] = v\n return out_dict\n\n\nclass vit_base_patch16_224(nn.Module):\n def __init__(self, cfg, **kwargs):\n super(vit_base_patch16_224, self).__init__()\n self.pretrained = True\n patch_size = 16\n self.model = VisionTransformer(\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_classes=cfg.MODEL.NUM_CLASSES,\n patch_size=patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n num_frames=cfg.DATA.NUM_FRAMES,\n attention_type=cfg.TIMESFORMER.ATTENTION_TYPE,\n **kwargs,\n )\n\n self.attention_type = cfg.TIMESFORMER.ATTENTION_TYPE\n self.model.default_cfg = default_cfgs[\"vit_base_patch16_224\"]\n self.num_patches = (cfg.DATA.TRAIN_CROP_SIZE // patch_size) * (\n cfg.DATA.TRAIN_CROP_SIZE // patch_size\n )\n pretrained_model = cfg.TIMESFORMER.PRETRAINED_MODEL\n if self.pretrained:\n load_pretrained(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=kwargs.get(\"in_chans\", 3),\n filter_fn=_conv_filter,\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_model,\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n\nclass TimeSformer(nn.Module):\n def __init__(\n self,\n image_size=224,\n patch_size=16,\n n_frms=8,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n drop_rate=0,\n use_grad_ckpt=False,\n ckpt_layer=0,\n remove_classifier=True,\n **kwargs,\n ):\n super(TimeSformer, self).__init__()\n\n self.img_size = image_size\n self.patch_size = patch_size\n self.num_frames = n_frms\n self.attn_drop_rate = attn_drop_rate\n self.drop_path_rate = drop_path_rate\n self.drop_rate = drop_rate\n self.use_grad_ckpt = use_grad_ckpt\n self.ckpt_layer = ckpt_layer\n\n self.attention_type = \"divided_space_time\"\n\n logging.info(\n f\"Initializing TimeSformer with img_size={self.img_size}, patch_size={self.patch_size}, num_frames={self.num_frames}\"\n )\n\n # will be ignored when loading official pretrained ckpt\n self.num_classes = 400\n\n self.model = VisionTransformer(\n img_size=self.img_size,\n num_classes=self.num_classes,\n patch_size=self.patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=self.drop_rate,\n attn_drop_rate=self.attn_drop_rate,\n drop_path_rate=self.drop_path_rate,\n num_frames=self.num_frames,\n attention_type=self.attention_type,\n use_grad_checkpointing=self.use_grad_ckpt,\n ckpt_layer=self.ckpt_layer,\n **kwargs,\n )\n\n if remove_classifier:\n self.model.remove_classifier()\n\n self.model.default_cfg = default_cfgs[\n \"vit_base_patch\" + str(self.patch_size) + \"_224\"\n ]\n self.num_patches = (self.img_size // self.patch_size) * (\n self.img_size // self.patch_size\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n return x\n\n def load_state_dict(self, pretrained_ckpt_path):\n logging.info(\n \"Loading TimeSformer checkpoints from {}\".format(pretrained_ckpt_path)\n )\n\n if pretrained_ckpt_path == \"vit_base_patch16_224\":\n load_ckpt_func = load_pretrained_imagenet\n else:\n load_ckpt_func = load_pretrained_kinetics\n\n load_ckpt_func(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=3\n# ... truncated ...","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit._cfg","uri":"program://CREMA/function/lavis.models.timesformer.vit._cfg#L35-L48","kind":"function","name":"_cfg","path":"lavis/models/timesformer/vit.py","language":"python","start_line":35,"end_line":48,"context_start_line":15,"context_end_line":68,"code":"from functools import partial\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.utils\nimport torch.utils.checkpoint\nfrom einops import rearrange\nfrom fairscale.nn.checkpoint.checkpoint_activations import checkpoint_wrapper\n\nfrom .helpers import load_pretrained, load_pretrained_imagenet, load_pretrained_kinetics\nfrom .vit_utils import (\n IMAGENET_DEFAULT_MEAN,\n IMAGENET_DEFAULT_STD,\n DropPath,\n to_2tuple,\n trunc_normal_,\n)\n\n\ndef _cfg(url=\"\", **kwargs):\n return {\n \"url\": url,\n \"num_classes\": 1000,\n \"input_size\": (3, 224, 224),\n \"pool_size\": None,\n \"crop_pct\": 0.9,\n \"interpolation\": \"bicubic\",\n \"mean\": IMAGENET_DEFAULT_MEAN,\n \"std\": IMAGENET_DEFAULT_STD,\n \"first_conv\": \"patch_embed.proj\",\n \"classifier\": \"head\",\n **kwargs,\n }\n\n\ndefault_cfgs = {\n \"vit_base_patch16_224\": _cfg(\n url=\"https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth\",\n mean=(0.5, 0.5, 0.5),\n std=(0.5, 0.5, 0.5),\n ),\n}\n\n\nclass Mlp(nn.Module):\n def __init__(\n self,\n in_features,\n hidden_features=None,\n out_features=None,\n act_layer=nn.GELU,\n drop=0.0,\n ):","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.Mlp","uri":"program://CREMA/class/lavis.models.timesformer.vit.Mlp#L60-L83","kind":"class","name":"Mlp","path":"lavis/models/timesformer/vit.py","language":"python","start_line":60,"end_line":83,"context_start_line":40,"context_end_line":103,"code":" \"pool_size\": None,\n \"crop_pct\": 0.9,\n \"interpolation\": \"bicubic\",\n \"mean\": IMAGENET_DEFAULT_MEAN,\n \"std\": IMAGENET_DEFAULT_STD,\n \"first_conv\": \"patch_embed.proj\",\n \"classifier\": \"head\",\n **kwargs,\n }\n\n\ndefault_cfgs = {\n \"vit_base_patch16_224\": _cfg(\n url=\"https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth\",\n mean=(0.5, 0.5, 0.5),\n std=(0.5, 0.5, 0.5),\n ),\n}\n\n\nclass Mlp(nn.Module):\n def __init__(\n self,\n in_features,\n hidden_features=None,\n out_features=None,\n act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n with_qkv=True,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n self.scale = qk_scale or head_dim**-0.5\n self.with_qkv = with_qkv\n if self.with_qkv:\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.Attention","uri":"program://CREMA/class/lavis.models.timesformer.vit.Attention#L86-L131","kind":"class","name":"Attention","path":"lavis/models/timesformer/vit.py","language":"python","start_line":86,"end_line":131,"context_start_line":66,"context_end_line":151,"code":" act_layer=nn.GELU,\n drop=0.0,\n ):\n super().__init__()\n out_features = out_features or in_features\n hidden_features = hidden_features or in_features\n self.fc1 = nn.Linear(in_features, hidden_features)\n self.act = act_layer()\n self.fc2 = nn.Linear(hidden_features, out_features)\n self.drop = nn.Dropout(drop)\n\n def forward(self, x):\n x = self.fc1(x)\n x = self.act(x)\n x = self.drop(x)\n x = self.fc2(x)\n x = self.drop(x)\n return x\n\n\nclass Attention(nn.Module):\n def __init__(\n self,\n dim,\n num_heads=8,\n qkv_bias=False,\n qk_scale=None,\n attn_drop=0.0,\n proj_drop=0.0,\n with_qkv=True,\n ):\n super().__init__()\n self.num_heads = num_heads\n head_dim = dim // num_heads\n self.scale = qk_scale or head_dim**-0.5\n self.with_qkv = with_qkv\n if self.with_qkv:\n self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(dim, dim)\n self.proj_drop = nn.Dropout(proj_drop)\n self.attn_drop = nn.Dropout(attn_drop)\n\n def forward(self, x):\n B, N, C = x.shape\n if self.with_qkv:\n qkv = (\n self.qkv(x)\n .reshape(B, N, 3, self.num_heads, C // self.num_heads)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = qkv[0], qkv[1], qkv[2]\n else:\n qkv = x.reshape(B, N, self.num_heads, C // self.num_heads).permute(\n 0, 2, 1, 3\n )\n q, k, v = qkv, qkv, qkv\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n if self.with_qkv:\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n layer_num,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.1,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n attention_type=\"divided_space_time\",\n use_grad_checkpointing=False,\n ):\n super().__init__()","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.Block","uri":"program://CREMA/class/lavis.models.timesformer.vit.Block#L134-L260","kind":"class","name":"Block","path":"lavis/models/timesformer/vit.py","language":"python","start_line":134,"end_line":260,"context_start_line":114,"context_end_line":280,"code":" .permute(2, 0, 3, 1, 4)\n )\n q, k, v = qkv[0], qkv[1], qkv[2]\n else:\n qkv = x.reshape(B, N, self.num_heads, C // self.num_heads).permute(\n 0, 2, 1, 3\n )\n q, k, v = qkv, qkv, qkv\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n attn = self.attn_drop(attn)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N, C)\n if self.with_qkv:\n x = self.proj(x)\n x = self.proj_drop(x)\n return x\n\n\nclass Block(nn.Module):\n def __init__(\n self,\n dim,\n num_heads,\n layer_num,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop=0.0,\n attn_drop=0.0,\n drop_path=0.1,\n act_layer=nn.GELU,\n norm_layer=nn.LayerNorm,\n attention_type=\"divided_space_time\",\n use_grad_checkpointing=False,\n ):\n super().__init__()\n self.attention_type = attention_type\n assert attention_type in [\n \"divided_space_time\",\n \"space_only\",\n \"joint_space_time\",\n ]\n\n self.norm1 = norm_layer(dim)\n self.attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n\n # Temporal Attention Parameters\n if self.attention_type == \"divided_space_time\":\n self.temporal_norm1 = norm_layer(dim)\n self.temporal_attn = Attention(\n dim,\n num_heads=num_heads,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n attn_drop=attn_drop,\n proj_drop=drop,\n )\n self.temporal_fc = nn.Linear(dim, dim)\n\n # drop path\n self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()\n self.norm2 = norm_layer(dim)\n mlp_hidden_dim = int(dim * mlp_ratio)\n self.mlp = Mlp(\n in_features=dim,\n hidden_features=mlp_hidden_dim,\n act_layer=act_layer,\n drop=drop,\n )\n\n # [dxli]\n self.layer_num = layer_num\n self.use_grad_checkpointing = use_grad_checkpointing\n\n if use_grad_checkpointing:\n self.temporal_attn = checkpoint_wrapper(self.temporal_attn)\n self.attn = checkpoint_wrapper(self.attn)\n self.mlp = checkpoint_wrapper(self.mlp)\n\n def forward(self, x, B, T, W):\n num_spatial_tokens = (x.size(1) - 1) // T\n H = num_spatial_tokens // W\n\n if self.attention_type in [\"space_only\", \"joint_space_time\"]:\n x = x + self.drop_path(self.attn(self.norm1(x)))\n x = x + self.drop_path(self.mlp(self.norm2(x)))\n return x\n elif self.attention_type == \"divided_space_time\":\n # Temporal\n xt = x[:, 1:, :]\n xt = rearrange(xt, \"b (h w t) m -> (b h w) t m\", b=B, h=H, w=W, t=T)\n\n temporal_attn_out = self.temporal_attn(self.temporal_norm1(xt))\n\n res_temporal = self.drop_path(temporal_attn_out)\n\n res_temporal = rearrange(\n res_temporal, \"(b h w) t m -> b (h w t) m\", b=B, h=H, w=W, t=T\n )\n res_temporal = self.temporal_fc(res_temporal)\n xt = x[:, 1:, :] + res_temporal\n\n # Spatial\n init_cls_token = x[:, 0, :].unsqueeze(1)\n cls_token = init_cls_token.repeat(1, T, 1)\n cls_token = rearrange(cls_token, \"b t m -> (b t) m\", b=B, t=T).unsqueeze(1)\n xs = xt\n xs = rearrange(xs, \"b (h w t) m -> (b t) (h w) m\", b=B, h=H, w=W, t=T)\n xs = torch.cat((cls_token, xs), 1)\n\n spatial_attn_out = self.attn(self.norm1(xs))\n res_spatial = self.drop_path(spatial_attn_out)\n\n # Taking care of CLS token\n cls_token = res_spatial[:, 0, :]\n cls_token = rearrange(cls_token, \"(b t) m -> b t m\", b=B, t=T)\n # averaging for every frame\n cls_token = torch.mean(cls_token, 1, True)\n res_spatial = res_spatial[:, 1:, :]\n res_spatial = rearrange(\n res_spatial, \"(b t) (h w) m -> b (h w t) m\", b=B, h=H, w=W, t=T\n )\n res = res_spatial\n x = xt\n\n # Mlp\n x = torch.cat((init_cls_token, x), 1) + torch.cat((cls_token, res), 1)\n\n x_res = x\n\n x = self.norm2(x)\n # x = x + self.drop_path(self.mlp(self.norm2(x)))\n\n # MLP\n mlp_out = self.mlp(x)\n\n x = x_res + self.drop_path(mlp_out)\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\"Image to Patch Embedding\"\"\"\n\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(\n in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n )\n\n def forward(self, x):\n B, C, T, H, W = x.shape","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.PatchEmbed","uri":"program://CREMA/class/lavis.models.timesformer.vit.PatchEmbed#L263-L285","kind":"class","name":"PatchEmbed","path":"lavis/models/timesformer/vit.py","language":"python","start_line":263,"end_line":285,"context_start_line":243,"context_end_line":305,"code":" res_spatial, \"(b t) (h w) m -> b (h w t) m\", b=B, h=H, w=W, t=T\n )\n res = res_spatial\n x = xt\n\n # Mlp\n x = torch.cat((init_cls_token, x), 1) + torch.cat((cls_token, res), 1)\n\n x_res = x\n\n x = self.norm2(x)\n # x = x + self.drop_path(self.mlp(self.norm2(x)))\n\n # MLP\n mlp_out = self.mlp(x)\n\n x = x_res + self.drop_path(mlp_out)\n return x\n\n\nclass PatchEmbed(nn.Module):\n \"\"\"Image to Patch Embedding\"\"\"\n\n def __init__(self, img_size=224, patch_size=16, in_chans=3, embed_dim=768):\n super().__init__()\n img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(\n in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n )\n\n def forward(self, x):\n B, C, T, H, W = x.shape\n x = rearrange(x, \"b c t h w -> (b t) c h w\")\n x = self.proj(x)\n W = x.size(-1)\n x = x.flatten(2).transpose(1, 2)\n return x, T, W\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformere\"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.VisionTransformer","uri":"program://CREMA/class/lavis.models.timesformer.vit.VisionTransformer#L288-L467","kind":"class","name":"VisionTransformer","path":"lavis/models/timesformer/vit.py","language":"python","start_line":288,"end_line":467,"context_start_line":268,"context_end_line":487,"code":" img_size = to_2tuple(img_size)\n patch_size = to_2tuple(patch_size)\n num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])\n self.img_size = img_size\n self.patch_size = patch_size\n self.num_patches = num_patches\n\n self.proj = nn.Conv2d(\n in_chans, embed_dim, kernel_size=patch_size, stride=patch_size\n )\n\n def forward(self, x):\n B, C, T, H, W = x.shape\n x = rearrange(x, \"b c t h w -> (b t) c h w\")\n x = self.proj(x)\n W = x.size(-1)\n x = x.flatten(2).transpose(1, 2)\n return x, T, W\n\n\nclass VisionTransformer(nn.Module):\n \"\"\"Vision Transformere\"\"\"\n\n def __init__(\n self,\n img_size=224,\n patch_size=16,\n in_chans=3,\n num_classes=1000,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4.0,\n qkv_bias=False,\n qk_scale=None,\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n hybrid_backbone=None,\n norm_layer=nn.LayerNorm,\n num_frames=8,\n attention_type=\"divided_space_time\",\n dropout=0.0,\n use_grad_checkpointing=False,\n ckpt_layer=0,\n ):\n super().__init__()\n\n self.attention_type = attention_type\n self.depth = depth\n self.dropout = nn.Dropout(dropout)\n self.num_classes = num_classes\n # num_features for consistency with other models\n self.num_features = self.embed_dim = embed_dim\n self.patch_embed = PatchEmbed(\n img_size=img_size,\n patch_size=patch_size,\n in_chans=in_chans,\n embed_dim=embed_dim,\n )\n num_patches = self.patch_embed.num_patches\n\n # Positional Embeddings\n self.cls_token = nn.Parameter(torch.zeros(1, 1, embed_dim))\n self.pos_embed = nn.Parameter(torch.zeros(1, num_patches + 1, embed_dim))\n self.pos_drop = nn.Dropout(p=drop_rate)\n if self.attention_type != \"space_only\":\n self.time_embed = nn.Parameter(torch.zeros(1, num_frames, embed_dim))\n self.time_drop = nn.Dropout(p=drop_rate)\n\n # Attention Blocks\n dpr = [\n x.item() for x in torch.linspace(0, drop_path_rate, self.depth)\n ] # stochastic depth decay rule\n self.blocks = nn.ModuleList(\n [\n Block(\n layer_num=i,\n use_grad_checkpointing=(\n use_grad_checkpointing and i >= self.depth - ckpt_layer\n ),\n dim=embed_dim,\n num_heads=num_heads,\n mlp_ratio=mlp_ratio,\n qkv_bias=qkv_bias,\n qk_scale=qk_scale,\n drop=drop_rate,\n attn_drop=attn_drop_rate,\n drop_path=dpr[i],\n norm_layer=norm_layer,\n attention_type=self.attention_type,\n )\n for i in range(self.depth)\n ]\n )\n self.norm = norm_layer(embed_dim)\n\n # Classifier head\n self.head = (\n nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n # initialization of temporal attention weights\n if self.attention_type == \"divided_space_time\":\n i = 0\n for m in self.blocks.modules():\n m_str = str(m)\n if \"Block\" in m_str:\n if i > 0:\n nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # resizing the positional embeddings in case they don't match the input at inference\n if x.size(1) != self.pos_embed.size(1):\n pos_embed = self.pos_embed\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n P = int(other_pos_embed.size(2) ** 0.5)\n H = x.size(1) // W\n other_pos_embed = other_pos_embed.reshape(1, x.size(2), P, P)\n new_pos_embed = F.interpolate(other_pos_embed, size=(H, W), mode=\"nearest\")\n new_pos_embed = new_pos_embed.flatten(2)\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n x = x + new_pos_embed\n else:\n x = x + self.pos_embed\n x = self.pos_drop(x)\n\n # Time Embeddings\n if self.attention_type != \"space_only\":\n cls_tokens = x[:B, 0, :].unsqueeze(1)\n x = x[:, 1:]\n x = rearrange(x, \"(b t) n m -> (b n) t m\", b=B, t=T)\n # Resizing time embeddings in case they don't match\n if T != self.time_embed.size(1):\n time_embed = self.time_embed.transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(T), mode=\"nearest\")\n new_time_embed = new_time_embed.transpose(1, 2)\n x = x + new_time_embed\n else:\n x = x + self.time_embed\n x = self.time_drop(x)\n x = rearrange(x, \"(b n) t m -> b (n t) m\", b=B, t=T)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # Attention blocks\n for blk in self.blocks:\n x = blk(x, B, T, W)\n\n # Predictions for space-only baseline\n if self.attention_type == \"space_only\":\n x = rearrange(x, \"(b t) n m -> b t n m\", b=B, t=T)\n x = torch.mean(x, 1) # averaging predictions for every frame\n\n x = self.norm(x)\n\n return x\n\n def forward(self, x):\n x = self.forward_features(x)\n x = self.head(x)\n return x\n\n\ndef _conv_filter(state_dict, patch_size=16):\n \"\"\"convert patch embedding weight from manual patchify + linear proj to conv\"\"\"\n out_dict = {}\n for k, v in state_dict.items():\n if \"patch_embed.proj.weight\" in k:\n if v.shape[-1] != patch_size:\n patch_size = v.shape[-1]\n v = v.reshape((v.shape[0], 3, patch_size, patch_size))\n out_dict[k] = v\n return out_dict\n\n\nclass vit_base_patch16_224(nn.Module):\n def __init__(self, cfg, **kwargs):\n super(vit_base_patch16_224, self).__init__()\n self.pretrained = True\n patch_size = 16\n self.model = VisionTransformer(","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit._conv_filter","uri":"program://CREMA/function/lavis.models.timesformer.vit._conv_filter#L470-L479","kind":"function","name":"_conv_filter","path":"lavis/models/timesformer/vit.py","language":"python","start_line":470,"end_line":479,"context_start_line":450,"context_end_line":499,"code":"\n # Attention blocks\n for blk in self.blocks:\n x = blk(x, B, T, W)\n\n # Predictions for space-only baseline\n if self.attention_type == \"space_only\":\n x = rearrange(x, \"(b t) n m -> b t n m\", b=B, t=T)\n x = torch.mean(x, 1) # averaging predictions for every frame\n\n x = self.norm(x)\n\n return x\n\n def forward(self, x):\n x = self.forward_features(x)\n x = self.head(x)\n return x\n\n\ndef _conv_filter(state_dict, patch_size=16):\n \"\"\"convert patch embedding weight from manual patchify + linear proj to conv\"\"\"\n out_dict = {}\n for k, v in state_dict.items():\n if \"patch_embed.proj.weight\" in k:\n if v.shape[-1] != patch_size:\n patch_size = v.shape[-1]\n v = v.reshape((v.shape[0], 3, patch_size, patch_size))\n out_dict[k] = v\n return out_dict\n\n\nclass vit_base_patch16_224(nn.Module):\n def __init__(self, cfg, **kwargs):\n super(vit_base_patch16_224, self).__init__()\n self.pretrained = True\n patch_size = 16\n self.model = VisionTransformer(\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_classes=cfg.MODEL.NUM_CLASSES,\n patch_size=patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.vit_base_patch16_224","uri":"program://CREMA/class/lavis.models.timesformer.vit.vit_base_patch16_224#L482-L525","kind":"class","name":"vit_base_patch16_224","path":"lavis/models/timesformer/vit.py","language":"python","start_line":482,"end_line":525,"context_start_line":462,"context_end_line":545,"code":" return x\n\n def forward(self, x):\n x = self.forward_features(x)\n x = self.head(x)\n return x\n\n\ndef _conv_filter(state_dict, patch_size=16):\n \"\"\"convert patch embedding weight from manual patchify + linear proj to conv\"\"\"\n out_dict = {}\n for k, v in state_dict.items():\n if \"patch_embed.proj.weight\" in k:\n if v.shape[-1] != patch_size:\n patch_size = v.shape[-1]\n v = v.reshape((v.shape[0], 3, patch_size, patch_size))\n out_dict[k] = v\n return out_dict\n\n\nclass vit_base_patch16_224(nn.Module):\n def __init__(self, cfg, **kwargs):\n super(vit_base_patch16_224, self).__init__()\n self.pretrained = True\n patch_size = 16\n self.model = VisionTransformer(\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_classes=cfg.MODEL.NUM_CLASSES,\n patch_size=patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=0.0,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n num_frames=cfg.DATA.NUM_FRAMES,\n attention_type=cfg.TIMESFORMER.ATTENTION_TYPE,\n **kwargs,\n )\n\n self.attention_type = cfg.TIMESFORMER.ATTENTION_TYPE\n self.model.default_cfg = default_cfgs[\"vit_base_patch16_224\"]\n self.num_patches = (cfg.DATA.TRAIN_CROP_SIZE // patch_size) * (\n cfg.DATA.TRAIN_CROP_SIZE // patch_size\n )\n pretrained_model = cfg.TIMESFORMER.PRETRAINED_MODEL\n if self.pretrained:\n load_pretrained(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=kwargs.get(\"in_chans\", 3),\n filter_fn=_conv_filter,\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_model,\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n\nclass TimeSformer(nn.Module):\n def __init__(\n self,\n image_size=224,\n patch_size=16,\n n_frms=8,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n drop_rate=0,\n use_grad_ckpt=False,\n ckpt_layer=0,\n remove_classifier=True,\n **kwargs,\n ):\n super(TimeSformer, self).__init__()\n\n self.img_size = image_size\n self.patch_size = patch_size","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.TimeSformer","uri":"program://CREMA/class/lavis.models.timesformer.vit.TimeSformer#L528-L634","kind":"class","name":"TimeSformer","path":"lavis/models/timesformer/vit.py","language":"python","start_line":528,"end_line":634,"context_start_line":508,"context_end_line":634,"code":" cfg.DATA.TRAIN_CROP_SIZE // patch_size\n )\n pretrained_model = cfg.TIMESFORMER.PRETRAINED_MODEL\n if self.pretrained:\n load_pretrained(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=kwargs.get(\"in_chans\", 3),\n filter_fn=_conv_filter,\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_model,\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n\nclass TimeSformer(nn.Module):\n def __init__(\n self,\n image_size=224,\n patch_size=16,\n n_frms=8,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n drop_rate=0,\n use_grad_ckpt=False,\n ckpt_layer=0,\n remove_classifier=True,\n **kwargs,\n ):\n super(TimeSformer, self).__init__()\n\n self.img_size = image_size\n self.patch_size = patch_size\n self.num_frames = n_frms\n self.attn_drop_rate = attn_drop_rate\n self.drop_path_rate = drop_path_rate\n self.drop_rate = drop_rate\n self.use_grad_ckpt = use_grad_ckpt\n self.ckpt_layer = ckpt_layer\n\n self.attention_type = \"divided_space_time\"\n\n logging.info(\n f\"Initializing TimeSformer with img_size={self.img_size}, patch_size={self.patch_size}, num_frames={self.num_frames}\"\n )\n\n # will be ignored when loading official pretrained ckpt\n self.num_classes = 400\n\n self.model = VisionTransformer(\n img_size=self.img_size,\n num_classes=self.num_classes,\n patch_size=self.patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=self.drop_rate,\n attn_drop_rate=self.attn_drop_rate,\n drop_path_rate=self.drop_path_rate,\n num_frames=self.num_frames,\n attention_type=self.attention_type,\n use_grad_checkpointing=self.use_grad_ckpt,\n ckpt_layer=self.ckpt_layer,\n **kwargs,\n )\n\n if remove_classifier:\n self.model.remove_classifier()\n\n self.model.default_cfg = default_cfgs[\n \"vit_base_patch\" + str(self.patch_size) + \"_224\"\n ]\n self.num_patches = (self.img_size // self.patch_size) * (\n self.img_size // self.patch_size\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n return x\n\n def load_state_dict(self, pretrained_ckpt_path):\n logging.info(\n \"Loading TimeSformer checkpoints from {}\".format(pretrained_ckpt_path)\n )\n\n if pretrained_ckpt_path == \"vit_base_patch16_224\":\n load_ckpt_func = load_pretrained_imagenet\n else:\n load_ckpt_func = load_pretrained_kinetics\n\n load_ckpt_func(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=3,\n filter_fn=_conv_filter,\n img_size=self.img_size,\n num_frames=self.num_frames,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_ckpt_path,\n )","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.__init__","uri":"program://CREMA/function/lavis.models.timesformer.vit.__init__#L529-L590","kind":"function","name":"__init__","path":"lavis/models/timesformer/vit.py","language":"python","start_line":529,"end_line":590,"context_start_line":509,"context_end_line":610,"code":" )\n pretrained_model = cfg.TIMESFORMER.PRETRAINED_MODEL\n if self.pretrained:\n load_pretrained(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=kwargs.get(\"in_chans\", 3),\n filter_fn=_conv_filter,\n img_size=cfg.DATA.TRAIN_CROP_SIZE,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_model,\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n\nclass TimeSformer(nn.Module):\n def __init__(\n self,\n image_size=224,\n patch_size=16,\n n_frms=8,\n attn_drop_rate=0.0,\n drop_path_rate=0.1,\n drop_rate=0,\n use_grad_ckpt=False,\n ckpt_layer=0,\n remove_classifier=True,\n **kwargs,\n ):\n super(TimeSformer, self).__init__()\n\n self.img_size = image_size\n self.patch_size = patch_size\n self.num_frames = n_frms\n self.attn_drop_rate = attn_drop_rate\n self.drop_path_rate = drop_path_rate\n self.drop_rate = drop_rate\n self.use_grad_ckpt = use_grad_ckpt\n self.ckpt_layer = ckpt_layer\n\n self.attention_type = \"divided_space_time\"\n\n logging.info(\n f\"Initializing TimeSformer with img_size={self.img_size}, patch_size={self.patch_size}, num_frames={self.num_frames}\"\n )\n\n # will be ignored when loading official pretrained ckpt\n self.num_classes = 400\n\n self.model = VisionTransformer(\n img_size=self.img_size,\n num_classes=self.num_classes,\n patch_size=self.patch_size,\n embed_dim=768,\n depth=12,\n num_heads=12,\n mlp_ratio=4,\n qkv_bias=True,\n norm_layer=partial(nn.LayerNorm, eps=1e-6),\n drop_rate=self.drop_rate,\n attn_drop_rate=self.attn_drop_rate,\n drop_path_rate=self.drop_path_rate,\n num_frames=self.num_frames,\n attention_type=self.attention_type,\n use_grad_checkpointing=self.use_grad_ckpt,\n ckpt_layer=self.ckpt_layer,\n **kwargs,\n )\n\n if remove_classifier:\n self.model.remove_classifier()\n\n self.model.default_cfg = default_cfgs[\n \"vit_base_patch\" + str(self.patch_size) + \"_224\"\n ]\n self.num_patches = (self.img_size // self.patch_size) * (\n self.img_size // self.patch_size\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.forward","uri":"program://CREMA/function/lavis.models.timesformer.vit.forward#L592-L594","kind":"function","name":"forward","path":"lavis/models/timesformer/vit.py","language":"python","start_line":592,"end_line":594,"context_start_line":572,"context_end_line":614,"code":" drop_rate=self.drop_rate,\n attn_drop_rate=self.attn_drop_rate,\n drop_path_rate=self.drop_path_rate,\n num_frames=self.num_frames,\n attention_type=self.attention_type,\n use_grad_checkpointing=self.use_grad_ckpt,\n ckpt_layer=self.ckpt_layer,\n **kwargs,\n )\n\n if remove_classifier:\n self.model.remove_classifier()\n\n self.model.default_cfg = default_cfgs[\n \"vit_base_patch\" + str(self.patch_size) + \"_224\"\n ]\n self.num_patches = (self.img_size // self.patch_size) * (\n self.img_size // self.patch_size\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n return x\n\n def load_state_dict(self, pretrained_ckpt_path):","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit._init_weights","uri":"program://CREMA/function/lavis.models.timesformer.vit._init_weights#L385-L392","kind":"function","name":"_init_weights","path":"lavis/models/timesformer/vit.py","language":"python","start_line":385,"end_line":392,"context_start_line":365,"context_end_line":412,"code":" # Classifier head\n self.head = (\n nn.Linear(embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n trunc_normal_(self.pos_embed, std=0.02)\n trunc_normal_(self.cls_token, std=0.02)\n self.apply(self._init_weights)\n\n # initialization of temporal attention weights\n if self.attention_type == \"divided_space_time\":\n i = 0\n for m in self.blocks.modules():\n m_str = str(m)\n if \"Block\" in m_str:\n if i > 0:\n nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.no_weight_decay","uri":"program://CREMA/function/lavis.models.timesformer.vit.no_weight_decay#L395-L396","kind":"function","name":"no_weight_decay","path":"lavis/models/timesformer/vit.py","language":"python","start_line":395,"end_line":396,"context_start_line":375,"context_end_line":416,"code":" if self.attention_type == \"divided_space_time\":\n i = 0\n for m in self.blocks.modules():\n m_str = str(m)\n if \"Block\" in m_str:\n if i > 0:\n nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.get_classifier","uri":"program://CREMA/function/lavis.models.timesformer.vit.get_classifier#L398-L399","kind":"function","name":"get_classifier","path":"lavis/models/timesformer/vit.py","language":"python","start_line":398,"end_line":399,"context_start_line":378,"context_end_line":419,"code":" m_str = str(m)\n if \"Block\" in m_str:\n if i > 0:\n nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # resizing the positional embeddings in case they don't match the input at inference\n if x.size(1) != self.pos_embed.size(1):\n pos_embed = self.pos_embed","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.reset_classifier","uri":"program://CREMA/function/lavis.models.timesformer.vit.reset_classifier#L401-L405","kind":"function","name":"reset_classifier","path":"lavis/models/timesformer/vit.py","language":"python","start_line":401,"end_line":405,"context_start_line":381,"context_end_line":425,"code":" nn.init.constant_(m.temporal_fc.weight, 0)\n nn.init.constant_(m.temporal_fc.bias, 0)\n i += 1\n\n def _init_weights(self, m):\n if isinstance(m, nn.Linear):\n trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # resizing the positional embeddings in case they don't match the input at inference\n if x.size(1) != self.pos_embed.size(1):\n pos_embed = self.pos_embed\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n P = int(other_pos_embed.size(2) ** 0.5)\n H = x.size(1) // W\n other_pos_embed = other_pos_embed.reshape(1, x.size(2), P, P)\n new_pos_embed = F.interpolate(other_pos_embed, size=(H, W), mode=\"nearest\")","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.remove_classifier","uri":"program://CREMA/function/lavis.models.timesformer.vit.remove_classifier#L407-L409","kind":"function","name":"remove_classifier","path":"lavis/models/timesformer/vit.py","language":"python","start_line":407,"end_line":409,"context_start_line":387,"context_end_line":429,"code":" trunc_normal_(m.weight, std=0.02)\n if isinstance(m, nn.Linear) and m.bias is not None:\n nn.init.constant_(m.bias, 0)\n elif isinstance(m, nn.LayerNorm):\n nn.init.constant_(m.bias, 0)\n nn.init.constant_(m.weight, 1.0)\n\n @torch.jit.ignore\n def no_weight_decay(self):\n return {\"pos_embed\", \"cls_token\", \"time_embed\"}\n\n def get_classifier(self):\n return self.head\n\n def reset_classifier(self, num_classes, global_pool=\"\"):\n self.num_classes = num_classes\n self.head = (\n nn.Linear(self.embed_dim, num_classes) if num_classes > 0 else nn.Identity()\n )\n\n def remove_classifier(self):\n self.num_classes = 0\n self.head = None\n\n def forward_features(self, x):\n B = x.shape[0]\n x, T, W = self.patch_embed(x)\n cls_tokens = self.cls_token.expand(x.size(0), -1, -1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n # resizing the positional embeddings in case they don't match the input at inference\n if x.size(1) != self.pos_embed.size(1):\n pos_embed = self.pos_embed\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n P = int(other_pos_embed.size(2) ** 0.5)\n H = x.size(1) // W\n other_pos_embed = other_pos_embed.reshape(1, x.size(2), P, P)\n new_pos_embed = F.interpolate(other_pos_embed, size=(H, W), mode=\"nearest\")\n new_pos_embed = new_pos_embed.flatten(2)\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n x = x + new_pos_embed","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.forward_features","uri":"program://CREMA/function/lavis.models.timesformer.vit.forward_features#L596-L612","kind":"function","name":"forward_features","path":"lavis/models/timesformer/vit.py","language":"python","start_line":596,"end_line":612,"context_start_line":576,"context_end_line":632,"code":" attention_type=self.attention_type,\n use_grad_checkpointing=self.use_grad_ckpt,\n ckpt_layer=self.ckpt_layer,\n **kwargs,\n )\n\n if remove_classifier:\n self.model.remove_classifier()\n\n self.model.default_cfg = default_cfgs[\n \"vit_base_patch\" + str(self.patch_size) + \"_224\"\n ]\n self.num_patches = (self.img_size // self.patch_size) * (\n self.img_size // self.patch_size\n )\n\n def forward(self, x):\n x = self.model(x)\n return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n return x\n\n def load_state_dict(self, pretrained_ckpt_path):\n logging.info(\n \"Loading TimeSformer checkpoints from {}\".format(pretrained_ckpt_path)\n )\n\n if pretrained_ckpt_path == \"vit_base_patch16_224\":\n load_ckpt_func = load_pretrained_imagenet\n else:\n load_ckpt_func = load_pretrained_kinetics\n\n load_ckpt_func(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=3,\n filter_fn=_conv_filter,\n img_size=self.img_size,\n num_frames=self.num_frames,\n num_patches=self.num_patches,\n attention_type=self.attention_type,","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit.load_state_dict","uri":"program://CREMA/function/lavis.models.timesformer.vit.load_state_dict#L614-L634","kind":"function","name":"load_state_dict","path":"lavis/models/timesformer/vit.py","language":"python","start_line":614,"end_line":634,"context_start_line":594,"context_end_line":634,"code":" return x\n\n def forward_features(self, x):\n # b, c, t, h, w = x.shape\n x = self.model.forward_features(x)\n\n ## apply pooling\n W = H = self.img_size // self.patch_size\n T = self.num_frames\n\n cls_tokens = x[:, 0, :].unsqueeze(1)\n other_tokens = x[:, 1:, :]\n\n x = rearrange(other_tokens, \"b (h w t) m -> b t (h w) m\", h=H, w=W, t=T)\n\n x = torch.mean(x, dim=1)\n x = torch.cat((cls_tokens, x), dim=1)\n\n return x\n\n def load_state_dict(self, pretrained_ckpt_path):\n logging.info(\n \"Loading TimeSformer checkpoints from {}\".format(pretrained_ckpt_path)\n )\n\n if pretrained_ckpt_path == \"vit_base_patch16_224\":\n load_ckpt_func = load_pretrained_imagenet\n else:\n load_ckpt_func = load_pretrained_kinetics\n\n load_ckpt_func(\n self.model,\n num_classes=self.model.num_classes,\n in_chans=3,\n filter_fn=_conv_filter,\n img_size=self.img_size,\n num_frames=self.num_frames,\n num_patches=self.num_patches,\n attention_type=self.attention_type,\n pretrained_model=pretrained_ckpt_path,\n )","source_hash":"a1fd7e795edd6eba8793f3a104895cb3b79bd4dc82064f3e581188319f93efe5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers","uri":"program://CREMA/module/lavis.models.timesformer.helpers#L1-L400","kind":"module","name":"lavis.models.timesformer.helpers","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":1,"end_line":400,"context_start_line":1,"context_end_line":400,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.\n# Copyright 2020 Ross Wightman\n# Modified model creation / weight loading / state_dict helpers\n\nimport logging, warnings\nimport os\nimport math\nfrom collections import OrderedDict\n\nimport torch\nimport torch.utils.model_zoo as model_zoo\nimport torch.nn.functional as F\n\n\ndef load_state_dict(checkpoint_path, use_ema=False):\n if checkpoint_path and os.path.isfile(checkpoint_path):\n checkpoint = torch.load(checkpoint_path, map_location=\"cpu\")\n state_dict_key = \"state_dict\"\n if isinstance(checkpoint, dict):\n if use_ema and \"state_dict_ema\" in checkpoint:\n state_dict_key = \"state_dict_ema\"\n if state_dict_key and state_dict_key in checkpoint:\n new_state_dict = OrderedDict()\n for k, v in checkpoint[state_dict_key].items():\n # strip `module.` prefix\n name = k[7:] if k.startswith(\"module\") else k\n new_state_dict[name] = v\n state_dict = new_state_dict\n elif \"model_state\" in checkpoint:\n state_dict_key = \"model_state\"\n new_state_dict = OrderedDict()\n for k, v in checkpoint[state_dict_key].items():\n # strip `model.` prefix\n name = k[6:] if k.startswith(\"model\") else k\n new_state_dict[name] = v\n state_dict = new_state_dict\n else:\n state_dict = checkpoint\n logging.info(\n \"Loaded {} from checkpoint '{}'\".format(state_dict_key, checkpoint_path)\n )\n return state_dict\n else:\n logging.error(\"No checkpoint found at '{}'\".format(checkpoint_path))\n raise FileNotFoundError()\n\n\ndef load_checkpoint(model, checkpoint_path, use_ema=False, strict=True):\n state_dict = load_state_dict(checkpoint_path, use_ema)\n model.load_state_dict(state_dict, strict=strict)\n\n\n# def resume_checkpoint(model, checkpoint_path, optimizer=None, loss_scaler=None, log_info=True):\n# resume_epoch = None\n# if os.path.isfile(checkpoint_path):\n# checkpoint = torch.load(checkpoint_path, map_location='cpu')\n# if isinstance(checkpoint, dict) and 'state_dict' in checkpoint:\n# if log_info:\n# _logger.info('Restoring model state from checkpoint...')\n# new_state_dict = OrderedDict()\n# for k, v in checkpoint['state_dict'].items():\n# name = k[7:] if k.startswith('module') else k\n# new_state_dict[name] = v\n# model.load_state_dict(new_state_dict)\n\n# if optimizer is not None and 'optimizer' in checkpoint:\n# if log_info:\n# _logger.info('Restoring optimizer state from checkpoint...')\n# optimizer.load_state_dict(checkpoint['optimizer'])\n\n# if loss_scaler is not None and loss_scaler.state_dict_key in checkpoint:\n# if log_info:\n# _logger.info('Restoring AMP loss scaler state from checkpoint...')\n# loss_scaler.load_state_dict(checkpoint[loss_scaler.state_dict_key])\n\n# if 'epoch' in checkpoint:\n# resume_epoch = checkpoint['epoch']\n# if 'version' in checkpoint and checkpoint['version'] > 1:\n# resume_epoch += 1 # start at the next epoch, old checkpoints incremented before save\n\n# if log_info:\n# _logger.info(\"Loaded checkpoint '{}' (epoch {})\".format(checkpoint_path, checkpoint['epoch']))\n# else:\n# model.load_state_dict(checkpoint)\n# if log_info:\n# _logger.info(\"Loaded checkpoint '{}'\".format(checkpoint_path))\n# return resume_epoch\n# else:\n# _logger.error(\"No checkpoint found at '{}'\".format(checkpoint_path))\n# raise FileNotFoundError()\n\n\ndef load_pretrained(\n model,\n cfg=None,\n num_classes=1000,\n in_chans=3,\n filter_fn=None,\n img_size=224,\n num_frames=8,\n num_patches=196,\n attention_type=\"divided_space_time\",\n pretrained_model=\"\",\n strict=True,\n):\n if cfg is None:\n cfg = getattr(model, \"default_cfg\")\n if cfg is None or \"url\" not in cfg or not cfg[\"url\"]:\n logging.warning(\"Pretrained model URL is invalid, using random initialization.\")\n return\n\n if len(pretrained_model) == 0:\n if cfg is None:\n logging.info(f\"loading from default config {model.default_cfg}.\")\n state_dict = model_zoo.load_url(cfg[\"url\"], progress=False, map_location=\"cpu\")\n else:\n try:\n state_dict = load_state_dict(pretrained_model)[\"model\"]\n except:\n state_dict = load_state_dict(pretrained_model)\n\n if filter_fn is not None:\n state_dict = filter_fn(state_dict)\n\n if in_chans == 1:\n conv1_name = cfg[\"first_conv\"]\n logging.info(\n \"Converting first conv (%s) pretrained weights from 3 to 1 channel\"\n % conv1_name\n )\n conv1_weight = state_dict[conv1_name + \".weight\"]\n conv1_type = conv1_weight.dtype\n conv1_weight = conv1_weight.float()\n O, I, J, K = conv1_weight.shape\n if I > 3:\n assert conv1_weight.shape[1] % 3 == 0\n # For models with space2depth stems\n conv1_weight = conv1_weight.reshape(O, I // 3, 3, J, K)\n conv1_weight = conv1_weight.sum(dim=2, keepdim=False)\n else:\n conv1_weight = conv1_weight.sum(dim=1, keepdim=True)\n conv1_weight = conv1_weight.to(conv1_type)\n state_dict[conv1_name + \".weight\"] = conv1_weight\n elif in_chans != 3:\n conv1_name = cfg[\"first_conv\"]\n conv1_weight = state_dict[conv1_name + \".weight\"]\n conv1_type = conv1_weight.dtype\n conv1_weight = conv1_weight.float()\n O, I, J, K = conv1_weight.shape\n if I != 3:\n logging.warning(\n \"Deleting first conv (%s) from pretrained weights.\" % conv1_name\n )\n del state_dict[conv1_name + \".weight\"]\n strict = False\n else:\n logging.info(\n \"Repeating first conv (%s) weights in channel dim.\" % conv1_name\n )\n repeat = int(math.ceil(in_chans / 3))\n conv1_weight = conv1_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :]\n conv1_weight *= 3 / float(in_chans)\n conv1_weight = conv1_weight.to(conv1_type)\n state_dict[conv1_name + \".weight\"] = conv1_weight\n\n classifier_name = cfg[\"classifier\"]\n if num_classes == 1000 and cfg[\"num_classes\"] == 1001:\n # special case for imagenet trained models with extra background class in pretrained weights\n classifier_weight = state_dict[classifier_name + \".weight\"]\n state_dict[classifier_name + \".weight\"] = classifier_weight[1:]\n classifier_bias = state_dict[classifier_name + \".bias\"]\n state_dict[classifier_name + \".bias\"] = classifier_bias[1:]\n elif num_classes != state_dict[classifier_name + \".weight\"].size(0):\n # print('Removing the last fully connected layer due to dimensions mismatch ('+str(num_classes)+ ' != '+str(state_dict[classifier_name + '.weight'].size(0))+').', flush=True)\n # completely discard fully connected for all other differences between pretrained and created model\n del state_dict[classifier_name + \".weight\"]\n del state_dict[classifier_name + \".bias\"]\n strict = False\n\n ## Resizing the positional embeddings in case they don't match\n logging.info(\n f\"Resizing spatial position embedding from {state_dict['pos_embed'].size(1)} to {num_patches + 1}\"\n )\n if num_patches + 1 != state_dict[\"pos_embed\"].size(1):\n pos_embed = state_dict[\"pos_embed\"]\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n new_pos_embed = F.interpolate(\n other_pos_embed, size=(num_patches), mode=\"nearest\"\n )\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n state_dict[\"pos_embed\"] = new_pos_embed\n\n ## Resizing time embeddings in case they don't match\n if \"time_embed\" in state_dict and num_frames != state_dict[\"time_embed\"].size(1):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict['time_embed'].size(1)} to {num_frames}\"\n )\n time_embed = state_dict[\"time_embed\"].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n state_dict[\"time_embed\"] = new_time_embed.transpose(1, 2)\n\n ## Initializing temporal attention\n if attention_type == \"divided_space_time\":\n new_state_dict = state_dict.copy()\n for key in state_dict:\n if \"blocks\" in key and \"attn\" in key:\n new_key = key.replace(\"attn\", \"temporal_attn\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n if \"blocks\" in key and \"norm1\" in key:\n new_key = key.replace(\"norm1\", \"temporal_norm1\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n state_dict = new_state_dict\n\n ## Loading the weights\n model.load_state_dict(state_dict, strict=False)\n\n\ndef load_pretrained_imagenet(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n import timm\n\n logging.info(f\"Loading vit_base_patch16_224 checkpoints.\")\n loaded_state_dict = timm.models.vision_transformer.vit_base_patch16_224(\n pretrained=True\n ).state_dict()\n\n del loaded_state_dict[\"head.weight\"]\n del loaded_state_dict[\"head.bias\"]\n\n ## Initializing temporal attention\n new_state_dict = loaded_state_dict.copy()\n for key in loaded_state_dict:\n if \"blocks\" in key and \"attn\" in key:\n new_key = key.replace(\"attn\", \"temporal_attn\")\n if not new_key in loaded_state_dict:\n new_state_dict[new_key] = loaded_state_dict[key]\n else:\n new_state_dict[new_key] = loaded_state_dict[new_key]\n if \"blocks\" in key and \"norm1\" in key:\n new_key = key.replace(\"norm1\", \"temporal_norm1\")\n if not new_key in loaded_state_dict:\n new_state_dict[new_key] = loaded_state_dict[key]\n else:\n new_state_dict[new_key] = loaded_state_dict[new_key]\n\n loaded_state_dict = new_state_dict\n\n loaded_keys = loaded_state_dict.keys()\n model_keys = model.state_dict().keys()\n\n load_not_in_model = [k for k in loaded_keys if k not in model_keys]\n model_not_in_load = [k for k in model_keys if k not in loaded_keys]\n\n toload = dict()\n mismatched_shape_keys = []\n for k in model_keys:\n if k in loaded_keys:\n if model.state_dict()[k].shape != loaded_state_dict[k].shape:\n mismatched_shape_keys.append(k)\n else:\n toload[k] = loaded_state_dict[k]\n\n logging.info(\"Keys in loaded but not in model:\")\n logging.info(f\"In total {len(load_not_in_model)}, {sorted(load_not_in_model)}\")\n logging.info(\"Keys in model but not in loaded:\")\n logging.info(f\"In total {len(model_not_in_load)}, {sorted(model_not_in_load)}\")\n logging.info(\"Keys in model and loaded, but shape mismatched:\")\n logging.info(\n f\"In total {len(mismatched_shape_keys)}, {sorted(mismatched_shape_keys)}\"\n )\n\n model.load_state_dict(toload, strict=False)\n\n\ndef load_pretrained_kinetics(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n if cfg is None:\n cfg = getattr(model, \"default_cfg\")\n if cfg is None or \"url\" not in cfg or not cfg[\"url\"]:\n logging.warning(\"Pretrained model URL is invalid, using random initialization.\")\n return\n\n assert (\n len(pretrained_model) > 0\n ), \"Path to pre-trained Kinetics weights not provided.\"\n\n state_dict = load_state_dict(pretrained_model)\n\n classifier_name = cfg[\"classifier\"]\n if ignore_classifier:\n\n classifier_weight_key = classifier_name + \".weight\"\n classifier_bias_key = classifier_name + \".bias\"\n\n state_dict[classifier_weight_key] = model.state_dict()[classifier_weight_key]\n state_dict[classifier_bias_key] = model.state_dict()[classifier_bias_key]\n\n else:\n raise NotImplementedError(\n \"[dxli] Not supporting loading Kinetics-pretrained ckpt with classifier.\"\n )\n\n ## Resizing the positional embeddings in case they don't match\n if num_patches + 1 != state_dict[\"pos_embed\"].size(1):\n new_pos_embed = resize_spatial_embedding(state_dict, \"pos_embed\", num_patches)\n state_dict[\"pos_embed\"] = new_pos_embed\n\n ## Resizing time embeddings in case they don't match\n if \"time_embed\" in state_dict and num_frames != state_dict[\"time_embed\"].size(1):\n state_dict[\"time_embed\"] = resize_temporal_embedding(\n state_dict, \"time_embed\", num_frames\n )\n\n ## Loading the weights\n try:\n model.load_state_dict(state_dict, strict=True)\n logging.info(\"Succeeded in loading Kinetics pre-trained weights.\")\n except:\n logging.error(\"Error in loading Kinetics pre-trained weights.\")\n\n\ndef resize_spatial_embedding(state_dict, key, num_patches):\n logging.info(\n f\"Resizing spatial position embedding from {state_dict[key].size(1)} to {num_patches + 1}\"\n )\n\n pos_embed = state_dict[key]\n\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n\n new_pos_embed = F.interpolate(other_pos_embed, size=(num_patches), mode=\"nearest\")\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n\n return new_pos_embed\n\n\ndef resize_temporal_embedding(state_dict, key, num_frames):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict[key].size(1)} to {num_frames}\"\n )\n\n time_embed = state_dict[key].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n\n return new_time_embed.transpose(1, 2)\n\n\ndef detach_variable(inputs):\n if isinstance(inputs, tuple):\n out = []\n for inp in inputs:\n x = inp.detach()\n x.requires_grad = inp.requires_grad\n out.append(x)\n return tuple(out)\n else:\n raise RuntimeError(\n \"Only tuple of tensors is supported. Got Unsupported input type: \",\n type(inputs).__name__,\n )\n\n\ndef check_backward_validity(inputs):\n if not any(inp.requires_grad for inp in inputs):\n warnings.warn(\n \"None of the inputs have requires_grad=True. Gradients will be None\"\n )","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.load_state_dict","uri":"program://CREMA/function/lavis.models.timesformer.helpers.load_state_dict#L24-L54","kind":"function","name":"load_state_dict","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":24,"end_line":54,"context_start_line":4,"context_end_line":74,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved.\n# Copyright 2020 Ross Wightman\n# Modified model creation / weight loading / state_dict helpers\n\nimport logging, warnings\nimport os\nimport math\nfrom collections import OrderedDict\n\nimport torch\nimport torch.utils.model_zoo as model_zoo\nimport torch.nn.functional as F\n\n\ndef load_state_dict(checkpoint_path, use_ema=False):\n if checkpoint_path and os.path.isfile(checkpoint_path):\n checkpoint = torch.load(checkpoint_path, map_location=\"cpu\")\n state_dict_key = \"state_dict\"\n if isinstance(checkpoint, dict):\n if use_ema and \"state_dict_ema\" in checkpoint:\n state_dict_key = \"state_dict_ema\"\n if state_dict_key and state_dict_key in checkpoint:\n new_state_dict = OrderedDict()\n for k, v in checkpoint[state_dict_key].items():\n # strip `module.` prefix\n name = k[7:] if k.startswith(\"module\") else k\n new_state_dict[name] = v\n state_dict = new_state_dict\n elif \"model_state\" in checkpoint:\n state_dict_key = \"model_state\"\n new_state_dict = OrderedDict()\n for k, v in checkpoint[state_dict_key].items():\n # strip `model.` prefix\n name = k[6:] if k.startswith(\"model\") else k\n new_state_dict[name] = v\n state_dict = new_state_dict\n else:\n state_dict = checkpoint\n logging.info(\n \"Loaded {} from checkpoint '{}'\".format(state_dict_key, checkpoint_path)\n )\n return state_dict\n else:\n logging.error(\"No checkpoint found at '{}'\".format(checkpoint_path))\n raise FileNotFoundError()\n\n\ndef load_checkpoint(model, checkpoint_path, use_ema=False, strict=True):\n state_dict = load_state_dict(checkpoint_path, use_ema)\n model.load_state_dict(state_dict, strict=strict)\n\n\n# def resume_checkpoint(model, checkpoint_path, optimizer=None, loss_scaler=None, log_info=True):\n# resume_epoch = None\n# if os.path.isfile(checkpoint_path):\n# checkpoint = torch.load(checkpoint_path, map_location='cpu')\n# if isinstance(checkpoint, dict) and 'state_dict' in checkpoint:\n# if log_info:\n# _logger.info('Restoring model state from checkpoint...')\n# new_state_dict = OrderedDict()\n# for k, v in checkpoint['state_dict'].items():\n# name = k[7:] if k.startswith('module') else k\n# new_state_dict[name] = v\n# model.load_state_dict(new_state_dict)\n","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.load_checkpoint","uri":"program://CREMA/function/lavis.models.timesformer.helpers.load_checkpoint#L57-L59","kind":"function","name":"load_checkpoint","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":57,"end_line":59,"context_start_line":37,"context_end_line":79,"code":" state_dict = new_state_dict\n elif \"model_state\" in checkpoint:\n state_dict_key = \"model_state\"\n new_state_dict = OrderedDict()\n for k, v in checkpoint[state_dict_key].items():\n # strip `model.` prefix\n name = k[6:] if k.startswith(\"model\") else k\n new_state_dict[name] = v\n state_dict = new_state_dict\n else:\n state_dict = checkpoint\n logging.info(\n \"Loaded {} from checkpoint '{}'\".format(state_dict_key, checkpoint_path)\n )\n return state_dict\n else:\n logging.error(\"No checkpoint found at '{}'\".format(checkpoint_path))\n raise FileNotFoundError()\n\n\ndef load_checkpoint(model, checkpoint_path, use_ema=False, strict=True):\n state_dict = load_state_dict(checkpoint_path, use_ema)\n model.load_state_dict(state_dict, strict=strict)\n\n\n# def resume_checkpoint(model, checkpoint_path, optimizer=None, loss_scaler=None, log_info=True):\n# resume_epoch = None\n# if os.path.isfile(checkpoint_path):\n# checkpoint = torch.load(checkpoint_path, map_location='cpu')\n# if isinstance(checkpoint, dict) and 'state_dict' in checkpoint:\n# if log_info:\n# _logger.info('Restoring model state from checkpoint...')\n# new_state_dict = OrderedDict()\n# for k, v in checkpoint['state_dict'].items():\n# name = k[7:] if k.startswith('module') else k\n# new_state_dict[name] = v\n# model.load_state_dict(new_state_dict)\n\n# if optimizer is not None and 'optimizer' in checkpoint:\n# if log_info:\n# _logger.info('Restoring optimizer state from checkpoint...')\n# optimizer.load_state_dict(checkpoint['optimizer'])\n","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.load_pretrained","uri":"program://CREMA/function/lavis.models.timesformer.helpers.load_pretrained#L102-L232","kind":"function","name":"load_pretrained","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":102,"end_line":232,"context_start_line":82,"context_end_line":252,"code":"# _logger.info('Restoring AMP loss scaler state from checkpoint...')\n# loss_scaler.load_state_dict(checkpoint[loss_scaler.state_dict_key])\n\n# if 'epoch' in checkpoint:\n# resume_epoch = checkpoint['epoch']\n# if 'version' in checkpoint and checkpoint['version'] > 1:\n# resume_epoch += 1 # start at the next epoch, old checkpoints incremented before save\n\n# if log_info:\n# _logger.info(\"Loaded checkpoint '{}' (epoch {})\".format(checkpoint_path, checkpoint['epoch']))\n# else:\n# model.load_state_dict(checkpoint)\n# if log_info:\n# _logger.info(\"Loaded checkpoint '{}'\".format(checkpoint_path))\n# return resume_epoch\n# else:\n# _logger.error(\"No checkpoint found at '{}'\".format(checkpoint_path))\n# raise FileNotFoundError()\n\n\ndef load_pretrained(\n model,\n cfg=None,\n num_classes=1000,\n in_chans=3,\n filter_fn=None,\n img_size=224,\n num_frames=8,\n num_patches=196,\n attention_type=\"divided_space_time\",\n pretrained_model=\"\",\n strict=True,\n):\n if cfg is None:\n cfg = getattr(model, \"default_cfg\")\n if cfg is None or \"url\" not in cfg or not cfg[\"url\"]:\n logging.warning(\"Pretrained model URL is invalid, using random initialization.\")\n return\n\n if len(pretrained_model) == 0:\n if cfg is None:\n logging.info(f\"loading from default config {model.default_cfg}.\")\n state_dict = model_zoo.load_url(cfg[\"url\"], progress=False, map_location=\"cpu\")\n else:\n try:\n state_dict = load_state_dict(pretrained_model)[\"model\"]\n except:\n state_dict = load_state_dict(pretrained_model)\n\n if filter_fn is not None:\n state_dict = filter_fn(state_dict)\n\n if in_chans == 1:\n conv1_name = cfg[\"first_conv\"]\n logging.info(\n \"Converting first conv (%s) pretrained weights from 3 to 1 channel\"\n % conv1_name\n )\n conv1_weight = state_dict[conv1_name + \".weight\"]\n conv1_type = conv1_weight.dtype\n conv1_weight = conv1_weight.float()\n O, I, J, K = conv1_weight.shape\n if I > 3:\n assert conv1_weight.shape[1] % 3 == 0\n # For models with space2depth stems\n conv1_weight = conv1_weight.reshape(O, I // 3, 3, J, K)\n conv1_weight = conv1_weight.sum(dim=2, keepdim=False)\n else:\n conv1_weight = conv1_weight.sum(dim=1, keepdim=True)\n conv1_weight = conv1_weight.to(conv1_type)\n state_dict[conv1_name + \".weight\"] = conv1_weight\n elif in_chans != 3:\n conv1_name = cfg[\"first_conv\"]\n conv1_weight = state_dict[conv1_name + \".weight\"]\n conv1_type = conv1_weight.dtype\n conv1_weight = conv1_weight.float()\n O, I, J, K = conv1_weight.shape\n if I != 3:\n logging.warning(\n \"Deleting first conv (%s) from pretrained weights.\" % conv1_name\n )\n del state_dict[conv1_name + \".weight\"]\n strict = False\n else:\n logging.info(\n \"Repeating first conv (%s) weights in channel dim.\" % conv1_name\n )\n repeat = int(math.ceil(in_chans / 3))\n conv1_weight = conv1_weight.repeat(1, repeat, 1, 1)[:, :in_chans, :, :]\n conv1_weight *= 3 / float(in_chans)\n conv1_weight = conv1_weight.to(conv1_type)\n state_dict[conv1_name + \".weight\"] = conv1_weight\n\n classifier_name = cfg[\"classifier\"]\n if num_classes == 1000 and cfg[\"num_classes\"] == 1001:\n # special case for imagenet trained models with extra background class in pretrained weights\n classifier_weight = state_dict[classifier_name + \".weight\"]\n state_dict[classifier_name + \".weight\"] = classifier_weight[1:]\n classifier_bias = state_dict[classifier_name + \".bias\"]\n state_dict[classifier_name + \".bias\"] = classifier_bias[1:]\n elif num_classes != state_dict[classifier_name + \".weight\"].size(0):\n # print('Removing the last fully connected layer due to dimensions mismatch ('+str(num_classes)+ ' != '+str(state_dict[classifier_name + '.weight'].size(0))+').', flush=True)\n # completely discard fully connected for all other differences between pretrained and created model\n del state_dict[classifier_name + \".weight\"]\n del state_dict[classifier_name + \".bias\"]\n strict = False\n\n ## Resizing the positional embeddings in case they don't match\n logging.info(\n f\"Resizing spatial position embedding from {state_dict['pos_embed'].size(1)} to {num_patches + 1}\"\n )\n if num_patches + 1 != state_dict[\"pos_embed\"].size(1):\n pos_embed = state_dict[\"pos_embed\"]\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n new_pos_embed = F.interpolate(\n other_pos_embed, size=(num_patches), mode=\"nearest\"\n )\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n state_dict[\"pos_embed\"] = new_pos_embed\n\n ## Resizing time embeddings in case they don't match\n if \"time_embed\" in state_dict and num_frames != state_dict[\"time_embed\"].size(1):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict['time_embed'].size(1)} to {num_frames}\"\n )\n time_embed = state_dict[\"time_embed\"].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n state_dict[\"time_embed\"] = new_time_embed.transpose(1, 2)\n\n ## Initializing temporal attention\n if attention_type == \"divided_space_time\":\n new_state_dict = state_dict.copy()\n for key in state_dict:\n if \"blocks\" in key and \"attn\" in key:\n new_key = key.replace(\"attn\", \"temporal_attn\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n if \"blocks\" in key and \"norm1\" in key:\n new_key = key.replace(\"norm1\", \"temporal_norm1\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n state_dict = new_state_dict\n\n ## Loading the weights\n model.load_state_dict(state_dict, strict=False)\n\n\ndef load_pretrained_imagenet(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n import timm\n\n logging.info(f\"Loading vit_base_patch16_224 checkpoints.\")\n loaded_state_dict = timm.models.vision_transformer.vit_base_patch16_224(\n pretrained=True\n ).state_dict()\n\n del loaded_state_dict[\"head.weight\"]\n del loaded_state_dict[\"head.bias\"]","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.load_pretrained_imagenet","uri":"program://CREMA/function/lavis.models.timesformer.helpers.load_pretrained_imagenet#L235-L296","kind":"function","name":"load_pretrained_imagenet","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":235,"end_line":296,"context_start_line":215,"context_end_line":316,"code":" new_state_dict = state_dict.copy()\n for key in state_dict:\n if \"blocks\" in key and \"attn\" in key:\n new_key = key.replace(\"attn\", \"temporal_attn\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n if \"blocks\" in key and \"norm1\" in key:\n new_key = key.replace(\"norm1\", \"temporal_norm1\")\n if not new_key in state_dict:\n new_state_dict[new_key] = state_dict[key]\n else:\n new_state_dict[new_key] = state_dict[new_key]\n state_dict = new_state_dict\n\n ## Loading the weights\n model.load_state_dict(state_dict, strict=False)\n\n\ndef load_pretrained_imagenet(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n import timm\n\n logging.info(f\"Loading vit_base_patch16_224 checkpoints.\")\n loaded_state_dict = timm.models.vision_transformer.vit_base_patch16_224(\n pretrained=True\n ).state_dict()\n\n del loaded_state_dict[\"head.weight\"]\n del loaded_state_dict[\"head.bias\"]\n\n ## Initializing temporal attention\n new_state_dict = loaded_state_dict.copy()\n for key in loaded_state_dict:\n if \"blocks\" in key and \"attn\" in key:\n new_key = key.replace(\"attn\", \"temporal_attn\")\n if not new_key in loaded_state_dict:\n new_state_dict[new_key] = loaded_state_dict[key]\n else:\n new_state_dict[new_key] = loaded_state_dict[new_key]\n if \"blocks\" in key and \"norm1\" in key:\n new_key = key.replace(\"norm1\", \"temporal_norm1\")\n if not new_key in loaded_state_dict:\n new_state_dict[new_key] = loaded_state_dict[key]\n else:\n new_state_dict[new_key] = loaded_state_dict[new_key]\n\n loaded_state_dict = new_state_dict\n\n loaded_keys = loaded_state_dict.keys()\n model_keys = model.state_dict().keys()\n\n load_not_in_model = [k for k in loaded_keys if k not in model_keys]\n model_not_in_load = [k for k in model_keys if k not in loaded_keys]\n\n toload = dict()\n mismatched_shape_keys = []\n for k in model_keys:\n if k in loaded_keys:\n if model.state_dict()[k].shape != loaded_state_dict[k].shape:\n mismatched_shape_keys.append(k)\n else:\n toload[k] = loaded_state_dict[k]\n\n logging.info(\"Keys in loaded but not in model:\")\n logging.info(f\"In total {len(load_not_in_model)}, {sorted(load_not_in_model)}\")\n logging.info(\"Keys in model but not in loaded:\")\n logging.info(f\"In total {len(model_not_in_load)}, {sorted(model_not_in_load)}\")\n logging.info(\"Keys in model and loaded, but shape mismatched:\")\n logging.info(\n f\"In total {len(mismatched_shape_keys)}, {sorted(mismatched_shape_keys)}\"\n )\n\n model.load_state_dict(toload, strict=False)\n\n\ndef load_pretrained_kinetics(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n if cfg is None:\n cfg = getattr(model, \"default_cfg\")\n if cfg is None or \"url\" not in cfg or not cfg[\"url\"]:\n logging.warning(\"Pretrained model URL is invalid, using random initialization.\")\n return\n\n assert (\n len(pretrained_model) > 0\n ), \"Path to pre-trained Kinetics weights not provided.\"","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.load_pretrained_kinetics","uri":"program://CREMA/function/lavis.models.timesformer.helpers.load_pretrained_kinetics#L299-L350","kind":"function","name":"load_pretrained_kinetics","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":299,"end_line":350,"context_start_line":279,"context_end_line":370,"code":" mismatched_shape_keys = []\n for k in model_keys:\n if k in loaded_keys:\n if model.state_dict()[k].shape != loaded_state_dict[k].shape:\n mismatched_shape_keys.append(k)\n else:\n toload[k] = loaded_state_dict[k]\n\n logging.info(\"Keys in loaded but not in model:\")\n logging.info(f\"In total {len(load_not_in_model)}, {sorted(load_not_in_model)}\")\n logging.info(\"Keys in model but not in loaded:\")\n logging.info(f\"In total {len(model_not_in_load)}, {sorted(model_not_in_load)}\")\n logging.info(\"Keys in model and loaded, but shape mismatched:\")\n logging.info(\n f\"In total {len(mismatched_shape_keys)}, {sorted(mismatched_shape_keys)}\"\n )\n\n model.load_state_dict(toload, strict=False)\n\n\ndef load_pretrained_kinetics(\n model,\n pretrained_model,\n cfg=None,\n ignore_classifier=True,\n num_frames=8,\n num_patches=196,\n **kwargs,\n):\n if cfg is None:\n cfg = getattr(model, \"default_cfg\")\n if cfg is None or \"url\" not in cfg or not cfg[\"url\"]:\n logging.warning(\"Pretrained model URL is invalid, using random initialization.\")\n return\n\n assert (\n len(pretrained_model) > 0\n ), \"Path to pre-trained Kinetics weights not provided.\"\n\n state_dict = load_state_dict(pretrained_model)\n\n classifier_name = cfg[\"classifier\"]\n if ignore_classifier:\n\n classifier_weight_key = classifier_name + \".weight\"\n classifier_bias_key = classifier_name + \".bias\"\n\n state_dict[classifier_weight_key] = model.state_dict()[classifier_weight_key]\n state_dict[classifier_bias_key] = model.state_dict()[classifier_bias_key]\n\n else:\n raise NotImplementedError(\n \"[dxli] Not supporting loading Kinetics-pretrained ckpt with classifier.\"\n )\n\n ## Resizing the positional embeddings in case they don't match\n if num_patches + 1 != state_dict[\"pos_embed\"].size(1):\n new_pos_embed = resize_spatial_embedding(state_dict, \"pos_embed\", num_patches)\n state_dict[\"pos_embed\"] = new_pos_embed\n\n ## Resizing time embeddings in case they don't match\n if \"time_embed\" in state_dict and num_frames != state_dict[\"time_embed\"].size(1):\n state_dict[\"time_embed\"] = resize_temporal_embedding(\n state_dict, \"time_embed\", num_frames\n )\n\n ## Loading the weights\n try:\n model.load_state_dict(state_dict, strict=True)\n logging.info(\"Succeeded in loading Kinetics pre-trained weights.\")\n except:\n logging.error(\"Error in loading Kinetics pre-trained weights.\")\n\n\ndef resize_spatial_embedding(state_dict, key, num_patches):\n logging.info(\n f\"Resizing spatial position embedding from {state_dict[key].size(1)} to {num_patches + 1}\"\n )\n\n pos_embed = state_dict[key]\n\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n\n new_pos_embed = F.interpolate(other_pos_embed, size=(num_patches), mode=\"nearest\")\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n\n return new_pos_embed\n\n\ndef resize_temporal_embedding(state_dict, key, num_frames):","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.resize_spatial_embedding","uri":"program://CREMA/function/lavis.models.timesformer.helpers.resize_spatial_embedding#L353-L367","kind":"function","name":"resize_spatial_embedding","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":353,"end_line":367,"context_start_line":333,"context_end_line":387,"code":"\n ## Resizing the positional embeddings in case they don't match\n if num_patches + 1 != state_dict[\"pos_embed\"].size(1):\n new_pos_embed = resize_spatial_embedding(state_dict, \"pos_embed\", num_patches)\n state_dict[\"pos_embed\"] = new_pos_embed\n\n ## Resizing time embeddings in case they don't match\n if \"time_embed\" in state_dict and num_frames != state_dict[\"time_embed\"].size(1):\n state_dict[\"time_embed\"] = resize_temporal_embedding(\n state_dict, \"time_embed\", num_frames\n )\n\n ## Loading the weights\n try:\n model.load_state_dict(state_dict, strict=True)\n logging.info(\"Succeeded in loading Kinetics pre-trained weights.\")\n except:\n logging.error(\"Error in loading Kinetics pre-trained weights.\")\n\n\ndef resize_spatial_embedding(state_dict, key, num_patches):\n logging.info(\n f\"Resizing spatial position embedding from {state_dict[key].size(1)} to {num_patches + 1}\"\n )\n\n pos_embed = state_dict[key]\n\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n\n new_pos_embed = F.interpolate(other_pos_embed, size=(num_patches), mode=\"nearest\")\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n\n return new_pos_embed\n\n\ndef resize_temporal_embedding(state_dict, key, num_frames):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict[key].size(1)} to {num_frames}\"\n )\n\n time_embed = state_dict[key].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n\n return new_time_embed.transpose(1, 2)\n\n\ndef detach_variable(inputs):\n if isinstance(inputs, tuple):\n out = []\n for inp in inputs:\n x = inp.detach()\n x.requires_grad = inp.requires_grad\n out.append(x)","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.resize_temporal_embedding","uri":"program://CREMA/function/lavis.models.timesformer.helpers.resize_temporal_embedding#L370-L378","kind":"function","name":"resize_temporal_embedding","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":370,"end_line":378,"context_start_line":350,"context_end_line":398,"code":" logging.error(\"Error in loading Kinetics pre-trained weights.\")\n\n\ndef resize_spatial_embedding(state_dict, key, num_patches):\n logging.info(\n f\"Resizing spatial position embedding from {state_dict[key].size(1)} to {num_patches + 1}\"\n )\n\n pos_embed = state_dict[key]\n\n cls_pos_embed = pos_embed[0, 0, :].unsqueeze(0).unsqueeze(1)\n other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n\n new_pos_embed = F.interpolate(other_pos_embed, size=(num_patches), mode=\"nearest\")\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n\n return new_pos_embed\n\n\ndef resize_temporal_embedding(state_dict, key, num_frames):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict[key].size(1)} to {num_frames}\"\n )\n\n time_embed = state_dict[key].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n\n return new_time_embed.transpose(1, 2)\n\n\ndef detach_variable(inputs):\n if isinstance(inputs, tuple):\n out = []\n for inp in inputs:\n x = inp.detach()\n x.requires_grad = inp.requires_grad\n out.append(x)\n return tuple(out)\n else:\n raise RuntimeError(\n \"Only tuple of tensors is supported. Got Unsupported input type: \",\n type(inputs).__name__,\n )\n\n\ndef check_backward_validity(inputs):\n if not any(inp.requires_grad for inp in inputs):\n warnings.warn(","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.detach_variable","uri":"program://CREMA/function/lavis.models.timesformer.helpers.detach_variable#L381-L393","kind":"function","name":"detach_variable","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":381,"end_line":393,"context_start_line":361,"context_end_line":400,"code":" other_pos_embed = pos_embed[0, 1:, :].unsqueeze(0).transpose(1, 2)\n\n new_pos_embed = F.interpolate(other_pos_embed, size=(num_patches), mode=\"nearest\")\n new_pos_embed = new_pos_embed.transpose(1, 2)\n new_pos_embed = torch.cat((cls_pos_embed, new_pos_embed), 1)\n\n return new_pos_embed\n\n\ndef resize_temporal_embedding(state_dict, key, num_frames):\n logging.info(\n f\"Resizing temporal position embedding from {state_dict[key].size(1)} to {num_frames}\"\n )\n\n time_embed = state_dict[key].transpose(1, 2)\n new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n\n return new_time_embed.transpose(1, 2)\n\n\ndef detach_variable(inputs):\n if isinstance(inputs, tuple):\n out = []\n for inp in inputs:\n x = inp.detach()\n x.requires_grad = inp.requires_grad\n out.append(x)\n return tuple(out)\n else:\n raise RuntimeError(\n \"Only tuple of tensors is supported. Got Unsupported input type: \",\n type(inputs).__name__,\n )\n\n\ndef check_backward_validity(inputs):\n if not any(inp.requires_grad for inp in inputs):\n warnings.warn(\n \"None of the inputs have requires_grad=True. Gradients will be None\"\n )","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.helpers.check_backward_validity","uri":"program://CREMA/function/lavis.models.timesformer.helpers.check_backward_validity#L396-L400","kind":"function","name":"check_backward_validity","path":"lavis/models/timesformer/helpers.py","language":"python","start_line":396,"end_line":400,"context_start_line":376,"context_end_line":400,"code":" new_time_embed = F.interpolate(time_embed, size=(num_frames), mode=\"nearest\")\n\n return new_time_embed.transpose(1, 2)\n\n\ndef detach_variable(inputs):\n if isinstance(inputs, tuple):\n out = []\n for inp in inputs:\n x = inp.detach()\n x.requires_grad = inp.requires_grad\n out.append(x)\n return tuple(out)\n else:\n raise RuntimeError(\n \"Only tuple of tensors is supported. Got Unsupported input type: \",\n type(inputs).__name__,\n )\n\n\ndef check_backward_validity(inputs):\n if not any(inp.requires_grad for inp in inputs):\n warnings.warn(\n \"None of the inputs have requires_grad=True. Gradients will be None\"\n )","source_hash":"53eb48a16909ffd99c3fbe2fd0a6a1327c421fd79741e091ca87f475a3ae1c4a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.linear","uri":"program://CREMA/module/lavis.models.timesformer.linear#L1-L21","kind":"module","name":"lavis.models.timesformer.linear","path":"lavis/models/timesformer/linear.py","language":"python","start_line":1,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\n\"\"\" Linear layer (alternate definition)\n\"\"\"\nimport torch\nimport torch.nn.functional as F\nfrom torch import nn as nn\n\n\nclass Linear(nn.Linear):\n def forward(self, input: torch.Tensor) -> torch.Tensor:\n if torch.jit.is_scripting():\n bias = self.bias.to(dtype=input.dtype) if self.bias is not None else None\n return F.linear(input, self.weight.to(dtype=input.dtype), bias=bias)\n else:\n return F.linear(input, self.weight, self.bias)","source_hash":"8177b9112e189da13bd659fb091cdaa6a4a9194e8a1eedbbf141100f4146b338","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.linear.Linear","uri":"program://CREMA/class/lavis.models.timesformer.linear.Linear#L15-L21","kind":"class","name":"Linear","path":"lavis/models/timesformer/linear.py","language":"python","start_line":15,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\n\"\"\" Linear layer (alternate definition)\n\"\"\"\nimport torch\nimport torch.nn.functional as F\nfrom torch import nn as nn\n\n\nclass Linear(nn.Linear):\n def forward(self, input: torch.Tensor) -> torch.Tensor:\n if torch.jit.is_scripting():\n bias = self.bias.to(dtype=input.dtype) if self.bias is not None else None\n return F.linear(input, self.weight.to(dtype=input.dtype), bias=bias)\n else:\n return F.linear(input, self.weight, self.bias)","source_hash":"8177b9112e189da13bd659fb091cdaa6a4a9194e8a1eedbbf141100f4146b338","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.linear.forward","uri":"program://CREMA/function/lavis.models.timesformer.linear.forward#L16-L21","kind":"function","name":"forward","path":"lavis/models/timesformer/linear.py","language":"python","start_line":16,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\n\"\"\" Linear layer (alternate definition)\n\"\"\"\nimport torch\nimport torch.nn.functional as F\nfrom torch import nn as nn\n\n\nclass Linear(nn.Linear):\n def forward(self, input: torch.Tensor) -> torch.Tensor:\n if torch.jit.is_scripting():\n bias = self.bias.to(dtype=input.dtype) if self.bias is not None else None\n return F.linear(input, self.weight.to(dtype=input.dtype), bias=bias)\n else:\n return F.linear(input, self.weight, self.bias)","source_hash":"8177b9112e189da13bd659fb091cdaa6a4a9194e8a1eedbbf141100f4146b338","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features","uri":"program://CREMA/module/lavis.models.timesformer.features#L1-L308","kind":"module","name":"lavis.models.timesformer.features","path":"lavis/models/timesformer/features.py","language":"python","start_line":1,"end_line":308,"context_start_line":1,"context_end_line":308,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright 2020 Ross Wightman\n\nfrom collections import OrderedDict, defaultdict\nfrom copy import deepcopy\nfrom functools import partial\nfrom typing import Dict, List, Tuple\n\nimport torch\nimport torch.nn as nn\n\n\nclass FeatureInfo:\n def __init__(self, feature_info: List[Dict], out_indices: Tuple[int]):\n prev_reduction = 1\n for fi in feature_info:\n # sanity check the mandatory fields, there may be additional fields depending on the model\n assert \"num_chs\" in fi and fi[\"num_chs\"] > 0\n assert \"reduction\" in fi and fi[\"reduction\"] >= prev_reduction\n prev_reduction = fi[\"reduction\"]\n assert \"module\" in fi\n self.out_indices = out_indices\n self.info = feature_info\n\n def from_other(self, out_indices: Tuple[int]):\n return FeatureInfo(deepcopy(self.info), out_indices)\n\n def get(self, key, idx=None):\n \"\"\"Get value by key at specified index (indices)\n if idx == None, returns value for key at each output index\n if idx is an integer, return value for that feature module index (ignoring output indices)\n if idx is a list/tupple, return value for each module index (ignoring output indices)\n \"\"\"\n if idx is None:\n return [self.info[i][key] for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [self.info[i][key] for i in idx]\n else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)\n elif hook_type == \"forward\":\n m.register_forward_hook(hook_fn)\n else:\n assert False, \"Unsupported hook type\"\n self._feature_outputs = defaultdict(OrderedDict)\n\n def _collect_output_hook(self, hook_id, *args):\n x = args[\n -1\n ] # tensor we want is last argument, output for fwd, input for fwd_pre\n if isinstance(x, tuple):\n x = x[0] # unwrap input tuple\n self._feature_outputs[x.device][hook_id] = x\n\n def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)\n elif isinstance(feature_info, (list, tuple)):\n return FeatureInfo(net.feature_info, out_indices)\n else:\n assert False, \"Provided feature_info is not valid\"\n\n\ndef _get_return_layers(feature_info, out_map):\n module_names = feature_info.module_name()\n return_layers = {}\n for i, name in enumerate(module_names):\n return_layers[name] = (\n out_map[i] if out_map is not None else feature_info.out_indices[i]\n )\n return return_layers\n\n\nclass FeatureDictNet(nn.ModuleDict):\n \"\"\"Feature extractor with OrderedDict return\n Wrap a model and extract features as specified by the out indices, the network is\n partially re-built from contained modules.\n There is a strong assumption that the modules have been registered into the model in the same\n order as they are used. There should be no reuse of the same nn.Module more than once, including\n trivial modules like `self.relu = nn.ReLU`.\n Only submodules that are directly assigned to the model class (`model.feature1`) or at most\n one Sequential container deep (`model.features.1`, with flatten_sequent=True) can be captured.\n All Sequential containers that are directly assigned to the original model will have their\n modules assigned to this module with the name `model.features.1` being changed to `model.features_1`\n Arguments:\n model (nn.Module): model from which we will extract the features\n out_indices (tuple[int]): model output indices to extract features for\n out_map (sequence): list or tuple specifying desired return id for each out index,\n otherwise str(index) is used\n feature_concat (bool): whether to concatenate intermediate features that are lists or tuples\n vs select element [0]\n flatten_sequential (bool): whether to flatten sequential modules assigned to model\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureDictNet, self).__init__()\n self.feature_info = _get_feature_info(model, out_indices)\n self.concat = feature_concat\n self.return_layers = {}\n return_layers = _get_return_layers(self.feature_info, out_map)\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = set(return_layers.keys())\n layers = OrderedDict()\n for new_name, old_name, module in modules:\n layers[new_name] = module\n if old_name in remaining:\n # return id has to be consistently str type for torchscript\n self.return_layers[new_name] = str(return_layers[old_name])\n remaining.remove(old_name)\n if not remaining:\n break\n assert not remaining and len(self.return_layers) == len(\n return_layers\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n\n def _collect(self, x) -> (Dict[str, torch.Tensor]):\n out = OrderedDict()\n for name, module in self.items():\n x = module(x)\n if name in self.return_layers:\n out_id = self.return_layers[name]\n if isinstance(x, (tuple, list)):\n # If model tap is a tuple or list, concat or select first element\n # FIXME this may need to be more generic / flexible for some nets\n out[out_id] = torch.cat(x, 1) if self.concat else x[0]\n else:\n out[out_id] = x\n return out\n\n def forward(self, x) -> Dict[str, torch.Tensor]:\n return self._collect(x)\n\n\nclass FeatureListNet(FeatureDictNet):\n \"\"\"Feature extractor with list return\n See docstring for FeatureDictNet above, this class exists only to appease Torchscript typing constraints.\n In eager Python we could have returned List[Tensor] vs Dict[id, Tensor] based on a member bool.\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureListNet, self).__init__(\n model,\n out_indices=out_indices,\n out_map=out_map,\n feature_concat=feature_concat,\n flatten_sequential=flatten_sequential,\n )\n\n def forward(self, x) -> (List[torch.Tensor]):\n return list(self._collect(x).values())\n\n\nclass FeatureHookNet(nn.ModuleDict):\n \"\"\"FeatureHookNet\n Wrap a model and extract features specified by the out indices using forward/forward-pre hooks.\n If `no_rewrite` is True, features are extracted via hooks without modifying the underlying\n network in any way.\n If `no_rewrite` is False, the model will be re-written as in the\n FeatureList/FeatureDict case by folding first to second (Sequential only) level modules into this one.\n FIXME this does not currently work with Torchscript, see FeatureHooks class\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n out_as_dict=False,\n no_rewrite=False,\n feature_concat=False,\n flatten_sequential=False,\n default_hook_type=\"forward\",\n ):\n super(FeatureHookNet, self).__init__()\n assert not torch.jit.is_scripting()\n self.feature_info = _get_feature_info(model, out_indices)\n self.out_as_dict = out_as_dict\n layers = OrderedDict()\n hooks = []\n if no_rewrite:\n assert not flatten_sequential\n if hasattr(model, \"reset_classifier\"): # make sure classifier is removed?\n model.reset_classifier(0)\n layers[\"body\"] = model\n hooks.extend(self.feature_info.get_dicts())\n else:\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = {\n f[\"module\"]: f[\"hook_type\"] if \"hook_type\" in f else default_hook_type\n for f in self.feature_info.get_dicts()\n }\n for new_name, old_name, module in modules:\n layers[new_name] = module\n for fn, fm in module.named_modules(prefix=old_name):\n if fn in remaining:\n hooks.append(dict(module=fn, hook_type=remaining[fn]))\n del remaining[fn]\n if not remaining:\n break\n assert (\n not remaining\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n self.hooks = FeatureHooks(hooks, model.named_modules(), out_map=out_map)\n\n def forward(self, x):\n for name, module in self.items():\n x = module(x)\n out = self.hooks.get_output(x.device)\n return out if self.out_as_dict else list(out.values())","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.FeatureInfo","uri":"program://CREMA/class/lavis.models.timesformer.features.FeatureInfo#L21-L82","kind":"class","name":"FeatureInfo","path":"lavis/models/timesformer/features.py","language":"python","start_line":21,"end_line":82,"context_start_line":1,"context_end_line":102,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright 2020 Ross Wightman\n\nfrom collections import OrderedDict, defaultdict\nfrom copy import deepcopy\nfrom functools import partial\nfrom typing import Dict, List, Tuple\n\nimport torch\nimport torch.nn as nn\n\n\nclass FeatureInfo:\n def __init__(self, feature_info: List[Dict], out_indices: Tuple[int]):\n prev_reduction = 1\n for fi in feature_info:\n # sanity check the mandatory fields, there may be additional fields depending on the model\n assert \"num_chs\" in fi and fi[\"num_chs\"] > 0\n assert \"reduction\" in fi and fi[\"reduction\"] >= prev_reduction\n prev_reduction = fi[\"reduction\"]\n assert \"module\" in fi\n self.out_indices = out_indices\n self.info = feature_info\n\n def from_other(self, out_indices: Tuple[int]):\n return FeatureInfo(deepcopy(self.info), out_indices)\n\n def get(self, key, idx=None):\n \"\"\"Get value by key at specified index (indices)\n if idx == None, returns value for key at each output index\n if idx is an integer, return value for that feature module index (ignoring output indices)\n if idx is a list/tupple, return value for each module index (ignoring output indices)\n \"\"\"\n if idx is None:\n return [self.info[i][key] for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [self.info[i][key] for i in idx]\n else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.FeatureHooks","uri":"program://CREMA/class/lavis.models.timesformer.features.FeatureHooks#L85-L120","kind":"class","name":"FeatureHooks","path":"lavis/models/timesformer/features.py","language":"python","start_line":85,"end_line":120,"context_start_line":65,"context_end_line":140,"code":"\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)\n elif hook_type == \"forward\":\n m.register_forward_hook(hook_fn)\n else:\n assert False, \"Unsupported hook type\"\n self._feature_outputs = defaultdict(OrderedDict)\n\n def _collect_output_hook(self, hook_id, *args):\n x = args[\n -1\n ] # tensor we want is last argument, output for fwd, input for fwd_pre\n if isinstance(x, tuple):\n x = x[0] # unwrap input tuple\n self._feature_outputs[x.device][hook_id] = x\n\n def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features._module_list","uri":"program://CREMA/function/lavis.models.timesformer.features._module_list#L123-L134","kind":"function","name":"_module_list","path":"lavis/models/timesformer/features.py","language":"python","start_line":123,"end_line":134,"context_start_line":103,"context_end_line":154,"code":" elif hook_type == \"forward\":\n m.register_forward_hook(hook_fn)\n else:\n assert False, \"Unsupported hook type\"\n self._feature_outputs = defaultdict(OrderedDict)\n\n def _collect_output_hook(self, hook_id, *args):\n x = args[\n -1\n ] # tensor we want is last argument, output for fwd, input for fwd_pre\n if isinstance(x, tuple):\n x = x[0] # unwrap input tuple\n self._feature_outputs[x.device][hook_id] = x\n\n def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)\n elif isinstance(feature_info, (list, tuple)):\n return FeatureInfo(net.feature_info, out_indices)\n else:\n assert False, \"Provided feature_info is not valid\"\n\n\ndef _get_return_layers(feature_info, out_map):\n module_names = feature_info.module_name()\n return_layers = {}\n for i, name in enumerate(module_names):\n return_layers[name] = (\n out_map[i] if out_map is not None else feature_info.out_indices[i]\n )\n return return_layers","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features._get_feature_info","uri":"program://CREMA/function/lavis.models.timesformer.features._get_feature_info#L137-L144","kind":"function","name":"_get_feature_info","path":"lavis/models/timesformer/features.py","language":"python","start_line":137,"end_line":144,"context_start_line":117,"context_end_line":164,"code":" def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)\n elif isinstance(feature_info, (list, tuple)):\n return FeatureInfo(net.feature_info, out_indices)\n else:\n assert False, \"Provided feature_info is not valid\"\n\n\ndef _get_return_layers(feature_info, out_map):\n module_names = feature_info.module_name()\n return_layers = {}\n for i, name in enumerate(module_names):\n return_layers[name] = (\n out_map[i] if out_map is not None else feature_info.out_indices[i]\n )\n return return_layers\n\n\nclass FeatureDictNet(nn.ModuleDict):\n \"\"\"Feature extractor with OrderedDict return\n Wrap a model and extract features as specified by the out indices, the network is\n partially re-built from contained modules.\n There is a strong assumption that the modules have been registered into the model in the same\n order as they are used. There should be no reuse of the same nn.Module more than once, including\n trivial modules like `self.relu = nn.ReLU`.\n Only submodules that are directly assigned to the model class (`model.feature1`) or at most","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features._get_return_layers","uri":"program://CREMA/function/lavis.models.timesformer.features._get_return_layers#L147-L154","kind":"function","name":"_get_return_layers","path":"lavis/models/timesformer/features.py","language":"python","start_line":147,"end_line":154,"context_start_line":127,"context_end_line":174,"code":" if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)\n elif isinstance(feature_info, (list, tuple)):\n return FeatureInfo(net.feature_info, out_indices)\n else:\n assert False, \"Provided feature_info is not valid\"\n\n\ndef _get_return_layers(feature_info, out_map):\n module_names = feature_info.module_name()\n return_layers = {}\n for i, name in enumerate(module_names):\n return_layers[name] = (\n out_map[i] if out_map is not None else feature_info.out_indices[i]\n )\n return return_layers\n\n\nclass FeatureDictNet(nn.ModuleDict):\n \"\"\"Feature extractor with OrderedDict return\n Wrap a model and extract features as specified by the out indices, the network is\n partially re-built from contained modules.\n There is a strong assumption that the modules have been registered into the model in the same\n order as they are used. There should be no reuse of the same nn.Module more than once, including\n trivial modules like `self.relu = nn.ReLU`.\n Only submodules that are directly assigned to the model class (`model.feature1`) or at most\n one Sequential container deep (`model.features.1`, with flatten_sequent=True) can be captured.\n All Sequential containers that are directly assigned to the original model will have their\n modules assigned to this module with the name `model.features.1` being changed to `model.features_1`\n Arguments:\n model (nn.Module): model from which we will extract the features\n out_indices (tuple[int]): model output indices to extract features for\n out_map (sequence): list or tuple specifying desired return id for each out index,\n otherwise str(index) is used\n feature_concat (bool): whether to concatenate intermediate features that are lists or tuples\n vs select element [0]","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.FeatureDictNet","uri":"program://CREMA/class/lavis.models.timesformer.features.FeatureDictNet#L157-L222","kind":"class","name":"FeatureDictNet","path":"lavis/models/timesformer/features.py","language":"python","start_line":157,"end_line":222,"context_start_line":137,"context_end_line":242,"code":"def _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)\n elif isinstance(feature_info, (list, tuple)):\n return FeatureInfo(net.feature_info, out_indices)\n else:\n assert False, \"Provided feature_info is not valid\"\n\n\ndef _get_return_layers(feature_info, out_map):\n module_names = feature_info.module_name()\n return_layers = {}\n for i, name in enumerate(module_names):\n return_layers[name] = (\n out_map[i] if out_map is not None else feature_info.out_indices[i]\n )\n return return_layers\n\n\nclass FeatureDictNet(nn.ModuleDict):\n \"\"\"Feature extractor with OrderedDict return\n Wrap a model and extract features as specified by the out indices, the network is\n partially re-built from contained modules.\n There is a strong assumption that the modules have been registered into the model in the same\n order as they are used. There should be no reuse of the same nn.Module more than once, including\n trivial modules like `self.relu = nn.ReLU`.\n Only submodules that are directly assigned to the model class (`model.feature1`) or at most\n one Sequential container deep (`model.features.1`, with flatten_sequent=True) can be captured.\n All Sequential containers that are directly assigned to the original model will have their\n modules assigned to this module with the name `model.features.1` being changed to `model.features_1`\n Arguments:\n model (nn.Module): model from which we will extract the features\n out_indices (tuple[int]): model output indices to extract features for\n out_map (sequence): list or tuple specifying desired return id for each out index,\n otherwise str(index) is used\n feature_concat (bool): whether to concatenate intermediate features that are lists or tuples\n vs select element [0]\n flatten_sequential (bool): whether to flatten sequential modules assigned to model\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureDictNet, self).__init__()\n self.feature_info = _get_feature_info(model, out_indices)\n self.concat = feature_concat\n self.return_layers = {}\n return_layers = _get_return_layers(self.feature_info, out_map)\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = set(return_layers.keys())\n layers = OrderedDict()\n for new_name, old_name, module in modules:\n layers[new_name] = module\n if old_name in remaining:\n # return id has to be consistently str type for torchscript\n self.return_layers[new_name] = str(return_layers[old_name])\n remaining.remove(old_name)\n if not remaining:\n break\n assert not remaining and len(self.return_layers) == len(\n return_layers\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n\n def _collect(self, x) -> (Dict[str, torch.Tensor]):\n out = OrderedDict()\n for name, module in self.items():\n x = module(x)\n if name in self.return_layers:\n out_id = self.return_layers[name]\n if isinstance(x, (tuple, list)):\n # If model tap is a tuple or list, concat or select first element\n # FIXME this may need to be more generic / flexible for some nets\n out[out_id] = torch.cat(x, 1) if self.concat else x[0]\n else:\n out[out_id] = x\n return out\n\n def forward(self, x) -> Dict[str, torch.Tensor]:\n return self._collect(x)\n\n\nclass FeatureListNet(FeatureDictNet):\n \"\"\"Feature extractor with list return\n See docstring for FeatureDictNet above, this class exists only to appease Torchscript typing constraints.\n In eager Python we could have returned List[Tensor] vs Dict[id, Tensor] based on a member bool.\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureListNet, self).__init__(\n model,\n out_indices=out_indices,\n out_map=out_map,","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.FeatureListNet","uri":"program://CREMA/class/lavis.models.timesformer.features.FeatureListNet#L225-L248","kind":"class","name":"FeatureListNet","path":"lavis/models/timesformer/features.py","language":"python","start_line":225,"end_line":248,"context_start_line":205,"context_end_line":268,"code":" self.update(layers)\n\n def _collect(self, x) -> (Dict[str, torch.Tensor]):\n out = OrderedDict()\n for name, module in self.items():\n x = module(x)\n if name in self.return_layers:\n out_id = self.return_layers[name]\n if isinstance(x, (tuple, list)):\n # If model tap is a tuple or list, concat or select first element\n # FIXME this may need to be more generic / flexible for some nets\n out[out_id] = torch.cat(x, 1) if self.concat else x[0]\n else:\n out[out_id] = x\n return out\n\n def forward(self, x) -> Dict[str, torch.Tensor]:\n return self._collect(x)\n\n\nclass FeatureListNet(FeatureDictNet):\n \"\"\"Feature extractor with list return\n See docstring for FeatureDictNet above, this class exists only to appease Torchscript typing constraints.\n In eager Python we could have returned List[Tensor] vs Dict[id, Tensor] based on a member bool.\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureListNet, self).__init__(\n model,\n out_indices=out_indices,\n out_map=out_map,\n feature_concat=feature_concat,\n flatten_sequential=flatten_sequential,\n )\n\n def forward(self, x) -> (List[torch.Tensor]):\n return list(self._collect(x).values())\n\n\nclass FeatureHookNet(nn.ModuleDict):\n \"\"\"FeatureHookNet\n Wrap a model and extract features specified by the out indices using forward/forward-pre hooks.\n If `no_rewrite` is True, features are extracted via hooks without modifying the underlying\n network in any way.\n If `no_rewrite` is False, the model will be re-written as in the\n FeatureList/FeatureDict case by folding first to second (Sequential only) level modules into this one.\n FIXME this does not currently work with Torchscript, see FeatureHooks class\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n out_as_dict=False,\n no_rewrite=False,\n feature_concat=False,","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.FeatureHookNet","uri":"program://CREMA/class/lavis.models.timesformer.features.FeatureHookNet#L251-L308","kind":"class","name":"FeatureHookNet","path":"lavis/models/timesformer/features.py","language":"python","start_line":251,"end_line":308,"context_start_line":231,"context_end_line":308,"code":" def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureListNet, self).__init__(\n model,\n out_indices=out_indices,\n out_map=out_map,\n feature_concat=feature_concat,\n flatten_sequential=flatten_sequential,\n )\n\n def forward(self, x) -> (List[torch.Tensor]):\n return list(self._collect(x).values())\n\n\nclass FeatureHookNet(nn.ModuleDict):\n \"\"\"FeatureHookNet\n Wrap a model and extract features specified by the out indices using forward/forward-pre hooks.\n If `no_rewrite` is True, features are extracted via hooks without modifying the underlying\n network in any way.\n If `no_rewrite` is False, the model will be re-written as in the\n FeatureList/FeatureDict case by folding first to second (Sequential only) level modules into this one.\n FIXME this does not currently work with Torchscript, see FeatureHooks class\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n out_as_dict=False,\n no_rewrite=False,\n feature_concat=False,\n flatten_sequential=False,\n default_hook_type=\"forward\",\n ):\n super(FeatureHookNet, self).__init__()\n assert not torch.jit.is_scripting()\n self.feature_info = _get_feature_info(model, out_indices)\n self.out_as_dict = out_as_dict\n layers = OrderedDict()\n hooks = []\n if no_rewrite:\n assert not flatten_sequential\n if hasattr(model, \"reset_classifier\"): # make sure classifier is removed?\n model.reset_classifier(0)\n layers[\"body\"] = model\n hooks.extend(self.feature_info.get_dicts())\n else:\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = {\n f[\"module\"]: f[\"hook_type\"] if \"hook_type\" in f else default_hook_type\n for f in self.feature_info.get_dicts()\n }\n for new_name, old_name, module in modules:\n layers[new_name] = module\n for fn, fm in module.named_modules(prefix=old_name):\n if fn in remaining:\n hooks.append(dict(module=fn, hook_type=remaining[fn]))\n del remaining[fn]\n if not remaining:\n break\n assert (\n not remaining\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n self.hooks = FeatureHooks(hooks, model.named_modules(), out_map=out_map)\n\n def forward(self, x):\n for name, module in self.items():\n x = module(x)\n out = self.hooks.get_output(x.device)\n return out if self.out_as_dict else list(out.values())","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.__init__","uri":"program://CREMA/function/lavis.models.timesformer.features.__init__#L261-L302","kind":"function","name":"__init__","path":"lavis/models/timesformer/features.py","language":"python","start_line":261,"end_line":302,"context_start_line":241,"context_end_line":308,"code":" out_indices=out_indices,\n out_map=out_map,\n feature_concat=feature_concat,\n flatten_sequential=flatten_sequential,\n )\n\n def forward(self, x) -> (List[torch.Tensor]):\n return list(self._collect(x).values())\n\n\nclass FeatureHookNet(nn.ModuleDict):\n \"\"\"FeatureHookNet\n Wrap a model and extract features specified by the out indices using forward/forward-pre hooks.\n If `no_rewrite` is True, features are extracted via hooks without modifying the underlying\n network in any way.\n If `no_rewrite` is False, the model will be re-written as in the\n FeatureList/FeatureDict case by folding first to second (Sequential only) level modules into this one.\n FIXME this does not currently work with Torchscript, see FeatureHooks class\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n out_as_dict=False,\n no_rewrite=False,\n feature_concat=False,\n flatten_sequential=False,\n default_hook_type=\"forward\",\n ):\n super(FeatureHookNet, self).__init__()\n assert not torch.jit.is_scripting()\n self.feature_info = _get_feature_info(model, out_indices)\n self.out_as_dict = out_as_dict\n layers = OrderedDict()\n hooks = []\n if no_rewrite:\n assert not flatten_sequential\n if hasattr(model, \"reset_classifier\"): # make sure classifier is removed?\n model.reset_classifier(0)\n layers[\"body\"] = model\n hooks.extend(self.feature_info.get_dicts())\n else:\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = {\n f[\"module\"]: f[\"hook_type\"] if \"hook_type\" in f else default_hook_type\n for f in self.feature_info.get_dicts()\n }\n for new_name, old_name, module in modules:\n layers[new_name] = module\n for fn, fm in module.named_modules(prefix=old_name):\n if fn in remaining:\n hooks.append(dict(module=fn, hook_type=remaining[fn]))\n del remaining[fn]\n if not remaining:\n break\n assert (\n not remaining\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n self.hooks = FeatureHooks(hooks, model.named_modules(), out_map=out_map)\n\n def forward(self, x):\n for name, module in self.items():\n x = module(x)\n out = self.hooks.get_output(x.device)\n return out if self.out_as_dict else list(out.values())","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.from_other","uri":"program://CREMA/function/lavis.models.timesformer.features.from_other#L33-L34","kind":"function","name":"from_other","path":"lavis/models/timesformer/features.py","language":"python","start_line":33,"end_line":34,"context_start_line":13,"context_end_line":54,"code":"from copy import deepcopy\nfrom functools import partial\nfrom typing import Dict, List, Tuple\n\nimport torch\nimport torch.nn as nn\n\n\nclass FeatureInfo:\n def __init__(self, feature_info: List[Dict], out_indices: Tuple[int]):\n prev_reduction = 1\n for fi in feature_info:\n # sanity check the mandatory fields, there may be additional fields depending on the model\n assert \"num_chs\" in fi and fi[\"num_chs\"] > 0\n assert \"reduction\" in fi and fi[\"reduction\"] >= prev_reduction\n prev_reduction = fi[\"reduction\"]\n assert \"module\" in fi\n self.out_indices = out_indices\n self.info = feature_info\n\n def from_other(self, out_indices: Tuple[int]):\n return FeatureInfo(deepcopy(self.info), out_indices)\n\n def get(self, key, idx=None):\n \"\"\"Get value by key at specified index (indices)\n if idx == None, returns value for key at each output index\n if idx is an integer, return value for that feature module index (ignoring output indices)\n if idx is a list/tupple, return value for each module index (ignoring output indices)\n \"\"\"\n if idx is None:\n return [self.info[i][key] for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [self.info[i][key] for i in idx]\n else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.get","uri":"program://CREMA/function/lavis.models.timesformer.features.get#L36-L47","kind":"function","name":"get","path":"lavis/models/timesformer/features.py","language":"python","start_line":36,"end_line":47,"context_start_line":16,"context_end_line":67,"code":"\nimport torch\nimport torch.nn as nn\n\n\nclass FeatureInfo:\n def __init__(self, feature_info: List[Dict], out_indices: Tuple[int]):\n prev_reduction = 1\n for fi in feature_info:\n # sanity check the mandatory fields, there may be additional fields depending on the model\n assert \"num_chs\" in fi and fi[\"num_chs\"] > 0\n assert \"reduction\" in fi and fi[\"reduction\"] >= prev_reduction\n prev_reduction = fi[\"reduction\"]\n assert \"module\" in fi\n self.out_indices = out_indices\n self.info = feature_info\n\n def from_other(self, out_indices: Tuple[int]):\n return FeatureInfo(deepcopy(self.info), out_indices)\n\n def get(self, key, idx=None):\n \"\"\"Get value by key at specified index (indices)\n if idx == None, returns value for key at each output index\n if idx is an integer, return value for that feature module index (ignoring output indices)\n if idx is a list/tupple, return value for each module index (ignoring output indices)\n \"\"\"\n if idx is None:\n return [self.info[i][key] for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [self.info[i][key] for i in idx]\n else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.get_dicts","uri":"program://CREMA/function/lavis.models.timesformer.features.get_dicts#L49-L64","kind":"function","name":"get_dicts","path":"lavis/models/timesformer/features.py","language":"python","start_line":49,"end_line":64,"context_start_line":29,"context_end_line":84,"code":" assert \"module\" in fi\n self.out_indices = out_indices\n self.info = feature_info\n\n def from_other(self, out_indices: Tuple[int]):\n return FeatureInfo(deepcopy(self.info), out_indices)\n\n def get(self, key, idx=None):\n \"\"\"Get value by key at specified index (indices)\n if idx == None, returns value for key at each output index\n if idx is an integer, return value for that feature module index (ignoring output indices)\n if idx is a list/tupple, return value for each module index (ignoring output indices)\n \"\"\"\n if idx is None:\n return [self.info[i][key] for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [self.info[i][key] for i in idx]\n else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.channels","uri":"program://CREMA/function/lavis.models.timesformer.features.channels#L66-L68","kind":"function","name":"channels","path":"lavis/models/timesformer/features.py","language":"python","start_line":66,"end_line":68,"context_start_line":46,"context_end_line":88,"code":" else:\n return self.info[idx][key]\n\n def get_dicts(self, keys=None, idx=None):\n \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.reduction","uri":"program://CREMA/function/lavis.models.timesformer.features.reduction#L70-L72","kind":"function","name":"reduction","path":"lavis/models/timesformer/features.py","language":"python","start_line":70,"end_line":72,"context_start_line":50,"context_end_line":92,"code":" \"\"\"return info dicts for specified keys (or all if None) at specified indices (or out_indices if None)\"\"\"\n if idx is None:\n if keys is None:\n return [self.info[i] for i in self.out_indices]\n else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.module_name","uri":"program://CREMA/function/lavis.models.timesformer.features.module_name#L74-L76","kind":"function","name":"module_name","path":"lavis/models/timesformer/features.py","language":"python","start_line":74,"end_line":76,"context_start_line":54,"context_end_line":96,"code":" else:\n return [{k: self.info[i][k] for k in keys} for i in self.out_indices]\n if isinstance(idx, (tuple, list)):\n return [\n self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.__getitem__","uri":"program://CREMA/function/lavis.models.timesformer.features.__getitem__#L78-L79","kind":"function","name":"__getitem__","path":"lavis/models/timesformer/features.py","language":"python","start_line":78,"end_line":79,"context_start_line":58,"context_end_line":99,"code":" self.info[i] if keys is None else {k: self.info[i][k] for k in keys}\n for i in idx\n ]\n else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.__len__","uri":"program://CREMA/function/lavis.models.timesformer.features.__len__#L81-L82","kind":"function","name":"__len__","path":"lavis/models/timesformer/features.py","language":"python","start_line":81,"end_line":82,"context_start_line":61,"context_end_line":102,"code":" else:\n return (\n self.info[idx] if keys is None else {k: self.info[idx][k] for k in keys}\n )\n\n def channels(self, idx=None):\n \"\"\"feature channels accessor\"\"\"\n return self.get(\"num_chs\", idx)\n\n def reduction(self, idx=None):\n \"\"\"feature reduction (output stride) accessor\"\"\"\n return self.get(\"reduction\", idx)\n\n def module_name(self, idx=None):\n \"\"\"feature module name accessor\"\"\"\n return self.get(\"module\", idx)\n\n def __getitem__(self, item):\n return self.info[item]\n\n def __len__(self):\n return len(self.info)\n\n\nclass FeatureHooks:\n \"\"\"Feature Hook Helper\n This module helps with the setup and extraction of hooks for extracting features from\n internal nodes in a model by node name. This works quite well in eager Python but needs\n redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features._collect_output_hook","uri":"program://CREMA/function/lavis.models.timesformer.features._collect_output_hook#L109-L115","kind":"function","name":"_collect_output_hook","path":"lavis/models/timesformer/features.py","language":"python","start_line":109,"end_line":115,"context_start_line":89,"context_end_line":135,"code":" redesign for torcscript.\n \"\"\"\n\n def __init__(self, hooks, named_modules, out_map=None, default_hook_type=\"forward\"):\n # setup feature hooks\n modules = {k: v for k, v in named_modules}\n for i, h in enumerate(hooks):\n hook_name = h[\"module\"]\n m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)\n elif hook_type == \"forward\":\n m.register_forward_hook(hook_fn)\n else:\n assert False, \"Unsupported hook type\"\n self._feature_outputs = defaultdict(OrderedDict)\n\n def _collect_output_hook(self, hook_id, *args):\n x = args[\n -1\n ] # tensor we want is last argument, output for fwd, input for fwd_pre\n if isinstance(x, tuple):\n x = x[0] # unwrap input tuple\n self._feature_outputs[x.device][hook_id] = x\n\n def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.get_output","uri":"program://CREMA/function/lavis.models.timesformer.features.get_output#L117-L120","kind":"function","name":"get_output","path":"lavis/models/timesformer/features.py","language":"python","start_line":117,"end_line":120,"context_start_line":97,"context_end_line":140,"code":" m = modules[hook_name]\n hook_id = out_map[i] if out_map else hook_name\n hook_fn = partial(self._collect_output_hook, hook_id)\n hook_type = h[\"hook_type\"] if \"hook_type\" in h else default_hook_type\n if hook_type == \"forward_pre\":\n m.register_forward_pre_hook(hook_fn)\n elif hook_type == \"forward\":\n m.register_forward_hook(hook_fn)\n else:\n assert False, \"Unsupported hook type\"\n self._feature_outputs = defaultdict(OrderedDict)\n\n def _collect_output_hook(self, hook_id, *args):\n x = args[\n -1\n ] # tensor we want is last argument, output for fwd, input for fwd_pre\n if isinstance(x, tuple):\n x = x[0] # unwrap input tuple\n self._feature_outputs[x.device][hook_id] = x\n\n def get_output(self, device) -> Dict[str, torch.tensor]:\n output = self._feature_outputs[device]\n self._feature_outputs[device] = OrderedDict() # clear after reading\n return output\n\n\ndef _module_list(module, flatten_sequential=False):\n # a yield/iter would be better for this but wouldn't be compatible with torchscript\n ml = []\n for name, module in module.named_children():\n if flatten_sequential and isinstance(module, nn.Sequential):\n # first level of Sequential containers is flattened into containing model\n for child_name, child_module in module.named_children():\n combined = [name, child_name]\n ml.append((\"_\".join(combined), \".\".join(combined), child_module))\n else:\n ml.append((name, name, module))\n return ml\n\n\ndef _get_feature_info(net, out_indices):\n feature_info = getattr(net, \"feature_info\")\n if isinstance(feature_info, FeatureInfo):\n return feature_info.from_other(out_indices)","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features._collect","uri":"program://CREMA/function/lavis.models.timesformer.features._collect#L207-L219","kind":"function","name":"_collect","path":"lavis/models/timesformer/features.py","language":"python","start_line":207,"end_line":219,"context_start_line":187,"context_end_line":239,"code":" self.feature_info = _get_feature_info(model, out_indices)\n self.concat = feature_concat\n self.return_layers = {}\n return_layers = _get_return_layers(self.feature_info, out_map)\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = set(return_layers.keys())\n layers = OrderedDict()\n for new_name, old_name, module in modules:\n layers[new_name] = module\n if old_name in remaining:\n # return id has to be consistently str type for torchscript\n self.return_layers[new_name] = str(return_layers[old_name])\n remaining.remove(old_name)\n if not remaining:\n break\n assert not remaining and len(self.return_layers) == len(\n return_layers\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n\n def _collect(self, x) -> (Dict[str, torch.Tensor]):\n out = OrderedDict()\n for name, module in self.items():\n x = module(x)\n if name in self.return_layers:\n out_id = self.return_layers[name]\n if isinstance(x, (tuple, list)):\n # If model tap is a tuple or list, concat or select first element\n # FIXME this may need to be more generic / flexible for some nets\n out[out_id] = torch.cat(x, 1) if self.concat else x[0]\n else:\n out[out_id] = x\n return out\n\n def forward(self, x) -> Dict[str, torch.Tensor]:\n return self._collect(x)\n\n\nclass FeatureListNet(FeatureDictNet):\n \"\"\"Feature extractor with list return\n See docstring for FeatureDictNet above, this class exists only to appease Torchscript typing constraints.\n In eager Python we could have returned List[Tensor] vs Dict[id, Tensor] based on a member bool.\n \"\"\"\n\n def __init__(\n self,\n model,\n out_indices=(0, 1, 2, 3, 4),\n out_map=None,\n feature_concat=False,\n flatten_sequential=False,\n ):\n super(FeatureListNet, self).__init__(","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.features.forward","uri":"program://CREMA/function/lavis.models.timesformer.features.forward#L304-L308","kind":"function","name":"forward","path":"lavis/models/timesformer/features.py","language":"python","start_line":304,"end_line":308,"context_start_line":284,"context_end_line":308,"code":" else:\n modules = _module_list(model, flatten_sequential=flatten_sequential)\n remaining = {\n f[\"module\"]: f[\"hook_type\"] if \"hook_type\" in f else default_hook_type\n for f in self.feature_info.get_dicts()\n }\n for new_name, old_name, module in modules:\n layers[new_name] = module\n for fn, fm in module.named_modules(prefix=old_name):\n if fn in remaining:\n hooks.append(dict(module=fn, hook_type=remaining[fn]))\n del remaining[fn]\n if not remaining:\n break\n assert (\n not remaining\n ), f\"Return layers ({remaining}) are not present in model\"\n self.update(layers)\n self.hooks = FeatureHooks(hooks, model.named_modules(), out_map=out_map)\n\n def forward(self, x):\n for name, module in self.items():\n x = module(x)\n out = self.hooks.get_output(x.device)\n return out if self.out_as_dict else list(out.values())","source_hash":"aaa9456a25221e9b42ccba57b47c751c7c0671e2113bb44dc1d59fdfe5e8e8c9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same","uri":"program://CREMA/module/lavis.models.timesformer.conv2d_same#L1-L116","kind":"module","name":"lavis.models.timesformer.conv2d_same","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":1,"end_line":116,"context_start_line":1,"context_end_line":116,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright 2020 Ross Wightman\n# Conv2d w/ Same Padding\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom typing import Tuple, Optional\n\nimport math\nfrom typing import List, Tuple\n\nfrom .vit_utils import is_static_pad, get_padding\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\ndef pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((math.ceil(x / s) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\ndef get_padding_value(padding, kernel_size, **kwargs) -> Tuple[Tuple, bool]:\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\ndef conv2d_same(\n x,\n weight: torch.Tensor,\n bias: Optional[torch.Tensor] = None,\n stride: Tuple[int, int] = (1, 1),\n padding: Tuple[int, int] = (0, 0),\n dilation: Tuple[int, int] = (1, 1),\n groups: int = 1,\n):\n x = pad_same(x, weight.shape[-2:], stride, dilation)\n return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)\n\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,\n in_channels,\n out_channels,\n kernel_size,\n stride=1,\n padding=0,\n dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n\n def forward(self, x):\n return conv2d_same(\n x,\n self.weight,\n self.bias,\n self.stride,\n self.padding,\n self.dilation,\n self.groups,\n )\n\n\ndef create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):\n padding = kwargs.pop(\"padding\", \"\")\n kwargs.setdefault(\"bias\", False)\n padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)\n if is_dynamic:\n return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)\n else:\n return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.pad_same","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.pad_same#L24-L35","kind":"function","name":"pad_same","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":24,"end_line":35,"context_start_line":4,"context_end_line":55,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright 2020 Ross Wightman\n# Conv2d w/ Same Padding\n\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom typing import Tuple, Optional\n\nimport math\nfrom typing import List, Tuple\n\nfrom .vit_utils import is_static_pad, get_padding\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\ndef pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((math.ceil(x / s) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\ndef get_padding_value(padding, kernel_size, **kwargs) -> Tuple[Tuple, bool]:\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.get_same_padding","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.get_same_padding#L39-L40","kind":"function","name":"get_same_padding","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":39,"end_line":40,"context_start_line":19,"context_end_line":60,"code":"from typing import List, Tuple\n\nfrom .vit_utils import is_static_pad, get_padding\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\ndef pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((math.ceil(x / s) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\ndef get_padding_value(padding, kernel_size, **kwargs) -> Tuple[Tuple, bool]:\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.get_padding_value","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.get_padding_value#L43-L63","kind":"function","name":"get_padding_value","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":43,"end_line":63,"context_start_line":23,"context_end_line":83,"code":"# Dynamically pad input x with 'SAME' padding for conv with specified args\ndef pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((math.ceil(x / s) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\ndef get_padding_value(padding, kernel_size, **kwargs) -> Tuple[Tuple, bool]:\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\ndef conv2d_same(\n x,\n weight: torch.Tensor,\n bias: Optional[torch.Tensor] = None,\n stride: Tuple[int, int] = (1, 1),\n padding: Tuple[int, int] = (0, 0),\n dilation: Tuple[int, int] = (1, 1),\n groups: int = 1,\n):\n x = pad_same(x, weight.shape[-2:], stride, dilation)\n return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)\n\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.conv2d_same","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.conv2d_same#L66-L76","kind":"function","name":"conv2d_same","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":66,"end_line":76,"context_start_line":46,"context_end_line":96,"code":" # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\ndef conv2d_same(\n x,\n weight: torch.Tensor,\n bias: Optional[torch.Tensor] = None,\n stride: Tuple[int, int] = (1, 1),\n padding: Tuple[int, int] = (0, 0),\n dilation: Tuple[int, int] = (1, 1),\n groups: int = 1,\n):\n x = pad_same(x, weight.shape[-2:], stride, dilation)\n return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)\n\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,\n in_channels,\n out_channels,\n kernel_size,\n stride=1,\n padding=0,\n dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.Conv2dSame","uri":"program://CREMA/class/lavis.models.timesformer.conv2d_same.Conv2dSame#L79-L106","kind":"class","name":"Conv2dSame","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":79,"end_line":106,"context_start_line":59,"context_end_line":116,"code":" padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\ndef conv2d_same(\n x,\n weight: torch.Tensor,\n bias: Optional[torch.Tensor] = None,\n stride: Tuple[int, int] = (1, 1),\n padding: Tuple[int, int] = (0, 0),\n dilation: Tuple[int, int] = (1, 1),\n groups: int = 1,\n):\n x = pad_same(x, weight.shape[-2:], stride, dilation)\n return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)\n\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,\n in_channels,\n out_channels,\n kernel_size,\n stride=1,\n padding=0,\n dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n\n def forward(self, x):\n return conv2d_same(\n x,\n self.weight,\n self.bias,\n self.stride,\n self.padding,\n self.dilation,\n self.groups,\n )\n\n\ndef create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):\n padding = kwargs.pop(\"padding\", \"\")\n kwargs.setdefault(\"bias\", False)\n padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)\n if is_dynamic:\n return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)\n else:\n return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.create_conv2d_pad","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.create_conv2d_pad#L109-L116","kind":"function","name":"create_conv2d_pad","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":109,"end_line":116,"context_start_line":89,"context_end_line":116,"code":" dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n\n def forward(self, x):\n return conv2d_same(\n x,\n self.weight,\n self.bias,\n self.stride,\n self.padding,\n self.dilation,\n self.groups,\n )\n\n\ndef create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):\n padding = kwargs.pop(\"padding\", \"\")\n kwargs.setdefault(\"bias\", False)\n padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)\n if is_dynamic:\n return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)\n else:\n return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.__init__","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.__init__#L82-L95","kind":"function","name":"__init__","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":82,"end_line":95,"context_start_line":62,"context_end_line":115,"code":" padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\ndef conv2d_same(\n x,\n weight: torch.Tensor,\n bias: Optional[torch.Tensor] = None,\n stride: Tuple[int, int] = (1, 1),\n padding: Tuple[int, int] = (0, 0),\n dilation: Tuple[int, int] = (1, 1),\n groups: int = 1,\n):\n x = pad_same(x, weight.shape[-2:], stride, dilation)\n return F.conv2d(x, weight, bias, stride, (0, 0), dilation, groups)\n\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,\n in_channels,\n out_channels,\n kernel_size,\n stride=1,\n padding=0,\n dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n\n def forward(self, x):\n return conv2d_same(\n x,\n self.weight,\n self.bias,\n self.stride,\n self.padding,\n self.dilation,\n self.groups,\n )\n\n\ndef create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):\n padding = kwargs.pop(\"padding\", \"\")\n kwargs.setdefault(\"bias\", False)\n padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)\n if is_dynamic:\n return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)\n else:","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.conv2d_same.forward","uri":"program://CREMA/function/lavis.models.timesformer.conv2d_same.forward#L97-L106","kind":"function","name":"forward","path":"lavis/models/timesformer/conv2d_same.py","language":"python","start_line":97,"end_line":106,"context_start_line":77,"context_end_line":116,"code":"\n\nclass Conv2dSame(nn.Conv2d):\n \"\"\"Tensorflow like 'SAME' convolution wrapper for 2D convolutions\"\"\"\n\n def __init__(\n self,\n in_channels,\n out_channels,\n kernel_size,\n stride=1,\n padding=0,\n dilation=1,\n groups=1,\n bias=True,\n ):\n super(Conv2dSame, self).__init__(\n in_channels, out_channels, kernel_size, stride, 0, dilation, groups, bias\n )\n\n def forward(self, x):\n return conv2d_same(\n x,\n self.weight,\n self.bias,\n self.stride,\n self.padding,\n self.dilation,\n self.groups,\n )\n\n\ndef create_conv2d_pad(in_chs, out_chs, kernel_size, **kwargs):\n padding = kwargs.pop(\"padding\", \"\")\n kwargs.setdefault(\"bias\", False)\n padding, is_dynamic = get_padding_value(padding, kernel_size, **kwargs)\n if is_dynamic:\n return Conv2dSame(in_chs, out_chs, kernel_size, **kwargs)\n else:\n return nn.Conv2d(in_chs, out_chs, kernel_size, padding=padding, **kwargs)","source_hash":"7772e052c20952842174a335a195192beba134057246d6993cd81d70e16ca66c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils","uri":"program://CREMA/module/lavis.models.timesformer.vit_utils#L1-L189","kind":"module","name":"lavis.models.timesformer.vit_utils","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":1,"end_line":189,"context_start_line":1,"context_end_line":189,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/facebookresearch/TimeSformer\n\"\"\"\n\n# Copyright 2020 Ross Wightman\n# Various utility functions\n\nimport torch\nimport torch.nn as nn\nimport math\nimport warnings\nimport torch.nn.functional as F\n\nfrom itertools import repeat\nimport collections.abc as container_abcs\n\nDEFAULT_CROP_PCT = 0.875\nIMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)\nIMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)\nIMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5)\nIMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5)\nIMAGENET_DPN_MEAN = (124 / 255, 117 / 255, 104 / 255)\nIMAGENET_DPN_STD = tuple([1 / (0.0167 * 255)] * 3)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)\n\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n\ndef get_padding_value(padding, kernel_size, **kwargs):\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((int(math.ceil(x // s)) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\n# Can SAME padding for given args be done statically?\ndef is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\ndef adaptive_pool_feat_mult(pool_type=\"avg\"):\n if pool_type == \"catavgmax\":\n return 2\n else:\n return 1\n\n\ndef drop_path(x, drop_prob: float = 0.0, training: bool = False):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils._no_grad_trunc_normal_","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils._no_grad_trunc_normal_#L31-L64","kind":"function","name":"_no_grad_trunc_normal_","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":31,"end_line":64,"context_start_line":11,"context_end_line":84,"code":"# Various utility functions\n\nimport torch\nimport torch.nn as nn\nimport math\nimport warnings\nimport torch.nn.functional as F\n\nfrom itertools import repeat\nimport collections.abc as container_abcs\n\nDEFAULT_CROP_PCT = 0.875\nIMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)\nIMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)\nIMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5)\nIMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5)\nIMAGENET_DPN_MEAN = (124 / 255, 117 / 255, 104 / 255)\nIMAGENET_DPN_STD = tuple([1 / (0.0167 * 255)] * 3)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.trunc_normal_","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.trunc_normal_#L67-L84","kind":"function","name":"trunc_normal_","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":67,"end_line":84,"context_start_line":47,"context_end_line":104,"code":" l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)\n\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils._ntuple","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils._ntuple#L88-L94","kind":"function","name":"_ntuple","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":88,"end_line":94,"context_start_line":68,"context_end_line":114,"code":" r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)\n\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n\ndef get_padding_value(padding, kernel_size, **kwargs):\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.get_padding","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.get_padding#L100-L102","kind":"function","name":"get_padding","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":100,"end_line":102,"context_start_line":80,"context_end_line":122,"code":" Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)\n\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n\ndef get_padding_value(padding, kernel_size, **kwargs):\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.get_padding_value","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.get_padding_value#L105-L125","kind":"function","name":"get_padding_value","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":105,"end_line":125,"context_start_line":85,"context_end_line":145,"code":"\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n\ndef get_padding_value(padding, kernel_size, **kwargs):\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((int(math.ceil(x // s)) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\n# Can SAME padding for given args be done statically?\ndef is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.get_same_padding","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.get_same_padding#L129-L130","kind":"function","name":"get_same_padding","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":129,"end_line":130,"context_start_line":109,"context_end_line":150,"code":" padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):\n # static case, no extra overhead\n padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((int(math.ceil(x // s)) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\n# Can SAME padding for given args be done statically?\ndef is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.is_static_pad","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.is_static_pad#L134-L135","kind":"function","name":"is_static_pad","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":134,"end_line":135,"context_start_line":114,"context_end_line":155,"code":" padding = get_padding(kernel_size, **kwargs)\n else:\n # dynamic 'SAME' padding, has runtime/GPU memory overhead\n padding = 0\n dynamic = True\n elif padding == \"valid\":\n # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((int(math.ceil(x // s)) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\n# Can SAME padding for given args be done statically?\ndef is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\ndef adaptive_pool_feat_mult(pool_type=\"avg\"):\n if pool_type == \"catavgmax\":","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.pad_same","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.pad_same#L140-L151","kind":"function","name":"pad_same","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":140,"end_line":151,"context_start_line":120,"context_end_line":171,"code":" # 'VALID' padding, same as padding=0\n padding = 0\n else:\n # Default to PyTorch style 'same'-ish symmetric padding\n padding = get_padding(kernel_size, **kwargs)\n return padding, dynamic\n\n\n# Calculate asymmetric TensorFlow-like 'SAME' padding for a convolution\ndef get_same_padding(x: int, k: int, s: int, d: int):\n return max((int(math.ceil(x // s)) - 1) * s + (k - 1) * d + 1 - x, 0)\n\n\n# Can SAME padding for given args be done statically?\ndef is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\ndef adaptive_pool_feat_mult(pool_type=\"avg\"):\n if pool_type == \"catavgmax\":\n return 2\n else:\n return 1\n\n\ndef drop_path(x, drop_prob: float = 0.0, training: bool = False):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.adaptive_pool_feat_mult","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.adaptive_pool_feat_mult#L154-L158","kind":"function","name":"adaptive_pool_feat_mult","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":154,"end_line":158,"context_start_line":134,"context_end_line":178,"code":"def is_static_pad(kernel_size: int, stride: int = 1, dilation: int = 1, **_):\n return stride == 1 and (dilation * (kernel_size - 1)) % 2 == 0\n\n\n# Dynamically pad input x with 'SAME' padding for conv with specified args\n# def pad_same(x, k: List[int], s: List[int], d: List[int] = (1, 1), value: float = 0):\ndef pad_same(x, k, s, d=(1, 1), value=0):\n ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\ndef adaptive_pool_feat_mult(pool_type=\"avg\"):\n if pool_type == \"catavgmax\":\n return 2\n else:\n return 1\n\n\ndef drop_path(x, drop_prob: float = 0.0, training: bool = False):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.drop_path","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.drop_path#L161-L178","kind":"function","name":"drop_path","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":161,"end_line":178,"context_start_line":141,"context_end_line":189,"code":" ih, iw = x.size()[-2:]\n pad_h, pad_w = get_same_padding(ih, k[0], s[0], d[0]), get_same_padding(\n iw, k[1], s[1], d[1]\n )\n if pad_h > 0 or pad_w > 0:\n x = F.pad(\n x,\n [pad_w // 2, pad_w - pad_w // 2, pad_h // 2, pad_h - pad_h // 2],\n value=value,\n )\n return x\n\n\ndef adaptive_pool_feat_mult(pool_type=\"avg\"):\n if pool_type == \"catavgmax\":\n return 2\n else:\n return 1\n\n\ndef drop_path(x, drop_prob: float = 0.0, training: bool = False):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.DropPath","uri":"program://CREMA/class/lavis.models.timesformer.vit_utils.DropPath#L181-L189","kind":"class","name":"DropPath","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":181,"end_line":189,"context_start_line":161,"context_end_line":189,"code":"def drop_path(x, drop_prob: float = 0.0, training: bool = False):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\n This is the same as the DropConnect impl I created for EfficientNet, etc networks, however,\n the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.norm_cdf","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.norm_cdf#L32-L34","kind":"function","name":"norm_cdf","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":32,"end_line":34,"context_start_line":12,"context_end_line":54,"code":"\nimport torch\nimport torch.nn as nn\nimport math\nimport warnings\nimport torch.nn.functional as F\n\nfrom itertools import repeat\nimport collections.abc as container_abcs\n\nDEFAULT_CROP_PCT = 0.875\nIMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)\nIMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)\nIMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5)\nIMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5)\nIMAGENET_DPN_MEAN = (124 / 255, 117 / 255, 104 / 255)\nIMAGENET_DPN_STD = tuple([1 / (0.0167 * 255)] * 3)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.parse","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.parse#L89-L92","kind":"function","name":"parse","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":89,"end_line":92,"context_start_line":69,"context_end_line":112,"code":" normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)\n\n\n# From PyTorch internals\ndef _ntuple(n):\n def parse(x):\n if isinstance(x, container_abcs.Iterable):\n return x\n return tuple(repeat(x, n))\n\n return parse\n\n\nto_2tuple = _ntuple(2)\n\n# Calculate symmetric padding for a convolution\ndef get_padding(kernel_size: int, stride: int = 1, dilation: int = 1, **_) -> int:\n padding = ((stride - 1) + dilation * (kernel_size - 1)) // 2\n return padding\n\n\ndef get_padding_value(padding, kernel_size, **kwargs):\n dynamic = False\n if isinstance(padding, str):\n # for any string padding, the padding will be calculated for you, one of three ways\n padding = padding.lower()\n if padding == \"same\":\n # TF compatible 'SAME' padding, has a performance and GPU memory allocation impact\n if is_static_pad(kernel_size, **kwargs):","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.__init__","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.__init__#L184-L186","kind":"function","name":"__init__","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":184,"end_line":186,"context_start_line":164,"context_end_line":189,"code":" the original name is misleading as 'Drop Connect' is a different form of dropout in a separate paper...\n See discussion: https://github.com/tensorflow/tpu/issues/494#issuecomment-532968956 ... I've opted for\n changing the layer and argument names to 'drop path' rather than mix DropConnect as a layer name and use\n 'survival rate' as the argument.\n \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.timesformer.vit_utils.forward","uri":"program://CREMA/function/lavis.models.timesformer.vit_utils.forward#L188-L189","kind":"function","name":"forward","path":"lavis/models/timesformer/vit_utils.py","language":"python","start_line":188,"end_line":189,"context_start_line":168,"context_end_line":189,"code":" \"\"\"\n if drop_prob == 0.0 or not training:\n return x\n keep_prob = 1 - drop_prob\n shape = (x.shape[0],) + (1,) * (\n x.ndim - 1\n ) # work with diff dim tensors, not just 2D ConvNets\n random_tensor = keep_prob + torch.rand(shape, dtype=x.dtype, device=x.device)\n random_tensor.floor_() # binarize\n output = x.div(keep_prob) * random_tensor\n return output\n\n\nclass DropPath(nn.Module):\n \"\"\"Drop paths (Stochastic Depth) per sample (when applied in main path of residual blocks).\"\"\"\n\n def __init__(self, drop_prob=None):\n super(DropPath, self).__init__()\n self.drop_prob = drop_prob\n\n def forward(self, x):\n return drop_path(x, self.drop_prob, self.training)","source_hash":"eada494467dbaf354a20f6a02f731d9d812b9ff7d7a2cb86b682e26859c6137f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval","uri":"program://CREMA/module/lavis.models.blip_models.blip_retrieval#L1-L396","kind":"module","name":"lavis.models.blip_models.blip_retrieval","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":1,"end_line":396,"context_start_line":1,"context_end_line":396,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.albef_models import compute_sim_matrix\nfrom lavis.models.base_model import (\n MomentumDistilationMixin,\n SharedQueueMixin,\n all_gather_with_grad,\n concat_all_gather,\n)\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipSimilarity,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_retrieval\")\nclass BlipRetrieval(BlipBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n BLIP retrieval model.\n\n Supported model types:\n - coco: fine-tuned BLIP base model on COCO dataset (Karpathy split).\n - flickr: fine-tuned BLIP base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> model = load_model(\"blip_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/blip_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/blip_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n negative_all_rank=False,\n max_txt_len=35,\n ):\n \"\"\" \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.negative_all_rank = negative_all_rank\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # Image-text Contrastive Learning\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_m_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_m_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_m_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_m_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_m_all / self.temp\n sim_t2i = text_feat @ image_feat_m_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n idxs = concat_all_gather(idx)\n if self.negative_all_rank:\n # compute sample similarity\n with torch.no_grad():\n mask = torch.eq(idx, idxs.t())\n\n image_feat_world = concat_all_gather(image_feat)\n text_feat_world = concat_all_gather(text_feat)\n\n sim_i2t = image_feat @ text_feat_world.t() / self.temp\n sim_t2i = text_feat @ image_feat_world.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n image_embeds_world = all_gather_with_grad(image_embeds)\n\n # select a negative image (from all ranks) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from all ranks) for each image\n input_ids_world = concat_all_gather(encoder_input_ids)\n att_mask_world = concat_all_gather(text.attention_mask)\n\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(input_ids_world[neg_idx])\n text_atts_neg.append(att_mask_world[neg_idx])\n\n else:\n with torch.no_grad():\n mask = torch.eq(idx, idx.t())\n\n sim_i2t = image_feat @ text_feat.t() / self.temp\n sim_t2i = text_feat @ image_feat.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image (from same rank) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from same rank) for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return BlipOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n negative_all_rank = cfg.get(\"negative_all_rank\", False)\n\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n alpha=alpha,\n embed_dim=embed_dim,\n momentum=momentum,\n negative_all_rank=negative_all_rank,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n model.reset_queue_ptr()\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.BlipRetrieval","uri":"program://CREMA/class/lavis.models.blip_models.blip_retrieval.BlipRetrieval#L32-L396","kind":"class","name":"BlipRetrieval","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":32,"end_line":396,"context_start_line":12,"context_end_line":396,"code":"from lavis.common.registry import registry\nfrom lavis.models.albef_models import compute_sim_matrix\nfrom lavis.models.base_model import (\n MomentumDistilationMixin,\n SharedQueueMixin,\n all_gather_with_grad,\n concat_all_gather,\n)\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipSimilarity,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_retrieval\")\nclass BlipRetrieval(BlipBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n BLIP retrieval model.\n\n Supported model types:\n - coco: fine-tuned BLIP base model on COCO dataset (Karpathy split).\n - flickr: fine-tuned BLIP base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> model = load_model(\"blip_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/blip_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/blip_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n negative_all_rank=False,\n max_txt_len=35,\n ):\n \"\"\" \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.negative_all_rank = negative_all_rank\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # Image-text Contrastive Learning\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_m_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_m_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_m_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_m_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_m_all / self.temp\n sim_t2i = text_feat @ image_feat_m_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n idxs = concat_all_gather(idx)\n if self.negative_all_rank:\n # compute sample similarity\n with torch.no_grad():\n mask = torch.eq(idx, idxs.t())\n\n image_feat_world = concat_all_gather(image_feat)\n text_feat_world = concat_all_gather(text_feat)\n\n sim_i2t = image_feat @ text_feat_world.t() / self.temp\n sim_t2i = text_feat @ image_feat_world.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n image_embeds_world = all_gather_with_grad(image_embeds)\n\n # select a negative image (from all ranks) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from all ranks) for each image\n input_ids_world = concat_all_gather(encoder_input_ids)\n att_mask_world = concat_all_gather(text.attention_mask)\n\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(input_ids_world[neg_idx])\n text_atts_neg.append(att_mask_world[neg_idx])\n\n else:\n with torch.no_grad():\n mask = torch.eq(idx, idx.t())\n\n sim_i2t = image_feat @ text_feat.t() / self.temp\n sim_t2i = text_feat @ image_feat.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image (from same rank) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from same rank) for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return BlipOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n negative_all_rank = cfg.get(\"negative_all_rank\", False)\n\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n alpha=alpha,\n embed_dim=embed_dim,\n momentum=momentum,\n negative_all_rank=negative_all_rank,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n model.reset_queue_ptr()\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval.__init__#L51-L111","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":51,"end_line":111,"context_start_line":31,"context_end_line":131,"code":"@registry.register_model(\"blip_retrieval\")\nclass BlipRetrieval(BlipBase, MomentumDistilationMixin, SharedQueueMixin):\n \"\"\"\n BLIP retrieval model.\n\n Supported model types:\n - coco: fine-tuned BLIP base model on COCO dataset (Karpathy split).\n - flickr: fine-tuned BLIP base model on Flickr30k dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> model = load_model(\"blip_retrieval\", \"flickr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"coco\": \"configs/models/blip_retrieval_coco.yaml\",\n \"flickr\": \"configs/models/blip_retrieval_flickr.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n negative_all_rank=False,\n max_txt_len=35,\n ):\n \"\"\" \"\"\"\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.negative_all_rank = negative_all_rank\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval._rampup_factor","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval._rampup_factor#L113-L114","kind":"function","name":"_rampup_factor","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":113,"end_line":114,"context_start_line":93,"context_end_line":134,"code":" self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.negative_all_rank = negative_all_rank\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval.forward#L116-L355","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":116,"end_line":355,"context_start_line":96,"context_end_line":375,"code":" self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"idx_queue\", torch.full((1, queue_size), -100))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n self.negative_all_rank = negative_all_rank\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - image_id (torch.Tensor): A tensor of shape (batch_size, ). The image ids, used to identify same images in batch.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_retrieval\", \"coco\")\n >>> images = torch.randn(4, 3, 384, 384)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> image_id = torch.tensor([1, 1, 2, 3])\n >>> samples = {\"image\": images, \"text_input\": text_input, \"image_id\": image_id, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm'])\n \"\"\"\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n idx = samples[\"image_id\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # Image-text Contrastive Learning\n idx = idx.view(-1, 1)\n idx_all = torch.cat([idx.t(), self.idx_queue.clone().detach()], dim=1)\n pos_idx = torch.eq(idx, idx_all).float()\n sim_targets = pos_idx / pos_idx.sum(1, keepdim=True)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_m_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_m_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_m_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_m_all / self.temp\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_m_all / self.temp\n sim_t2i = text_feat @ image_feat_m_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m, idx)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n idxs = concat_all_gather(idx)\n if self.negative_all_rank:\n # compute sample similarity\n with torch.no_grad():\n mask = torch.eq(idx, idxs.t())\n\n image_feat_world = concat_all_gather(image_feat)\n text_feat_world = concat_all_gather(text_feat)\n\n sim_i2t = image_feat @ text_feat_world.t() / self.temp\n sim_t2i = text_feat @ image_feat_world.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n image_embeds_world = all_gather_with_grad(image_embeds)\n\n # select a negative image (from all ranks) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds_world[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from all ranks) for each image\n input_ids_world = concat_all_gather(encoder_input_ids)\n att_mask_world = concat_all_gather(text.attention_mask)\n\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(input_ids_world[neg_idx])\n text_atts_neg.append(att_mask_world[neg_idx])\n\n else:\n with torch.no_grad():\n mask = torch.eq(idx, idx.t())\n\n sim_i2t = image_feat @ text_feat.t() / self.temp\n sim_t2i = text_feat @ image_feat.t() / self.temp\n\n weights_i2t = F.softmax(sim_i2t, dim=1)\n weights_i2t.masked_fill_(mask, 0)\n\n weights_t2i = F.softmax(sim_t2i, dim=1)\n weights_t2i.masked_fill_(mask, 0)\n\n # select a negative image (from same rank) for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text (from same rank) for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(self.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n return BlipOutput(\n loss=loss_itc + loss_itm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n negative_all_rank = cfg.get(\"negative_all_rank\", False)\n\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.reset_queue_ptr","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval.reset_queue_ptr#L357-L358","kind":"function","name":"reset_queue_ptr","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":357,"end_line":358,"context_start_line":337,"context_end_line":378,"code":" sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n negative_all_rank = cfg.get(\"negative_all_rank\", False)\n\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n alpha=alpha,","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval.from_config#L361-L388","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":361,"end_line":388,"context_start_line":341,"context_end_line":396,"code":" sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds=text_embeds,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n negative_all_rank = cfg.get(\"negative_all_rank\", False)\n\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n alpha=alpha,\n embed_dim=embed_dim,\n momentum=momentum,\n negative_all_rank=negative_all_rank,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n model.reset_queue_ptr()\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_retrieval.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.blip_models.blip_retrieval.compute_sim_matrix#L390-L396","kind":"function","name":"compute_sim_matrix","path":"lavis/models/blip_models/blip_retrieval.py","language":"python","start_line":390,"end_line":396,"context_start_line":370,"context_end_line":396,"code":"\n queue_size = cfg.get(\"queue_size\", 0)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n queue_size=queue_size,\n alpha=alpha,\n embed_dim=embed_dim,\n momentum=momentum,\n negative_all_rank=negative_all_rank,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n model.reset_queue_ptr()\n\n return model\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n \"\"\"\n Compute similarity i2t, t2i matrix for the given data loader.\n \"\"\"\n k_test = task_cfg.k_test\n\n return compute_sim_matrix(model=self, data_loader=data_loader, k_test=k_test)","source_hash":"5a3294a6f7c62348531871de4d987bccd9e699f1958ecc43fdba8a458effa03b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption","uri":"program://CREMA/module/lavis.models.blip_models.blip_caption#L1-L219","kind":"module","name":"lavis.models.blip_models.blip_caption","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":1,"end_line":219,"context_start_line":1,"context_end_line":219,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.common.registry import registry\n\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_caption\")\nclass BlipCaption(BlipBase):\n \"\"\"\n BLIP captioning model.\n\n Supported model types:\n - base_coco: fine-tuned BLIP base model on COCO caption dataset (Karparthy split).\n - large_coco: fine-tuned BLIP large model on COCO caption dataset (Karparthy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_caption\", \"base_coco\")\n >>> model = load_model(\"blip_caption\", \"large_coco\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base_coco\": \"configs/models/blip_caption_base_coco.yaml\",\n \"large_coco\": \"configs/models/blip_caption_large_coco.yaml\",\n }\n\n def __init__(self, image_encoder, text_decoder, prompt=None, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_decoder = text_decoder\n\n self.prompt = prompt\n self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1\n\n self.max_txt_len = max_txt_len\n\n def forward_encoder(self, samples):\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n return image_embeds\n\n def forward_decoder(self, samples, image_embeds):\n # prepare inputs for forwarding decoder\n raw_text = samples[\"text_input\"]\n text = self.tokenizer(\n raw_text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n # prepare targets for forwarding decoder\n decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100\n )\n decoder_targets[:, : self.prompt_length] = -100\n\n # forward decoder\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n decoder_output = self.text_decoder(\n input_ids=text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n return decoder_output, decoder_targets\n\n def forward(self, samples):\n r\"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size.\n Returns:\n output (BlipOutput): A BlipOutput object containing the following\n attributes:\n - loss (torch.Tensor): A scalar tensor containing the total loss. For BlipCaption, this is the same as the LM loss.\n - loss_lm (torch.Tensor): A scalar tensor containing the LM loss.\n - intermediate_outputs (BlipIntermediateOutput): A BlipIntermediateOutput object containing intermediate outputs.\n see :class:`lavis.models.blip_models.blip_outputs.BlipOutput` for more details.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = [\"a large statue of a person spraying water from a fountain\"]\n >>> samples = {\"image\": image, \"text_input\": text_input}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss', 'loss_lm'])\n >>> output.intermediate_output.image_embeds.shape\n torch.Size([1, 577, 768])\n >>> output.intermediate_output.decoder_labels.shape\n torch.Size([1, 13])\n ```\"\"\"\n\n image_embeds = self.forward_encoder(samples)\n decoder_output, decoder_targets = self.forward_decoder(samples, image_embeds)\n\n # return decoder_out\n return BlipOutput(\n loss=decoder_output.loss,\n loss_lm=decoder_output.loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n num_captions=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> samples = {\"image\": image}\n >>> captions = model.generate(samples)\n >>> captions\n ['a large statue of a person spraying water from a fountain']\n >>> captions = model.generate(samples, use_nucleus_sampling=True, num_captions=3)\n >>> captions # example output, results may vary due to randomness\n ['singapore showing the view of some building',\n 'the singapore harbor in twilight, as the weather is going down',\n 'the famous singapore fountain at sunset']\n \"\"\"\n # prepare inputs for decoder generation.\n encoder_out = self.forward_encoder(samples)\n image_embeds = torch.repeat_interleave(encoder_out, num_captions, 0)\n\n prompt = [self.prompt] * image_embeds.size(0)\n prompt = self.tokenizer(prompt, return_tensors=\"pt\").to(self.device)\n prompt.input_ids[:, 0] = self.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n # get decoded text\n decoder_out = self.text_decoder.generate_from_encoder(\n tokenized_prompt=prompt,\n visual_embeds=image_embeds,\n sep_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n use_nucleus_sampling=use_nucleus_sampling,\n num_beams=num_beams,\n max_length=max_length,\n min_length=min_length,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n )\n\n outputs = self.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n captions = [output[len(self.prompt) :] for output in outputs]\n\n return captions\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n # text encoder + multimodal decoder\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n prompt = cfg.get(\"prompt\", None)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n model = cls(image_encoder, text_decoder, prompt=prompt, max_txt_len=max_txt_len)\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.BlipCaption","uri":"program://CREMA/class/lavis.models.blip_models.blip_caption.BlipCaption#L21-L219","kind":"class","name":"BlipCaption","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":21,"end_line":219,"context_start_line":1,"context_end_line":219,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.common.registry import registry\n\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_caption\")\nclass BlipCaption(BlipBase):\n \"\"\"\n BLIP captioning model.\n\n Supported model types:\n - base_coco: fine-tuned BLIP base model on COCO caption dataset (Karparthy split).\n - large_coco: fine-tuned BLIP large model on COCO caption dataset (Karparthy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_caption\", \"base_coco\")\n >>> model = load_model(\"blip_caption\", \"large_coco\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base_coco\": \"configs/models/blip_caption_base_coco.yaml\",\n \"large_coco\": \"configs/models/blip_caption_large_coco.yaml\",\n }\n\n def __init__(self, image_encoder, text_decoder, prompt=None, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_decoder = text_decoder\n\n self.prompt = prompt\n self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1\n\n self.max_txt_len = max_txt_len\n\n def forward_encoder(self, samples):\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n return image_embeds\n\n def forward_decoder(self, samples, image_embeds):\n # prepare inputs for forwarding decoder\n raw_text = samples[\"text_input\"]\n text = self.tokenizer(\n raw_text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n # prepare targets for forwarding decoder\n decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100\n )\n decoder_targets[:, : self.prompt_length] = -100\n\n # forward decoder\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n decoder_output = self.text_decoder(\n input_ids=text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n return decoder_output, decoder_targets\n\n def forward(self, samples):\n r\"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size.\n Returns:\n output (BlipOutput): A BlipOutput object containing the following\n attributes:\n - loss (torch.Tensor): A scalar tensor containing the total loss. For BlipCaption, this is the same as the LM loss.\n - loss_lm (torch.Tensor): A scalar tensor containing the LM loss.\n - intermediate_outputs (BlipIntermediateOutput): A BlipIntermediateOutput object containing intermediate outputs.\n see :class:`lavis.models.blip_models.blip_outputs.BlipOutput` for more details.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = [\"a large statue of a person spraying water from a fountain\"]\n >>> samples = {\"image\": image, \"text_input\": text_input}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss', 'loss_lm'])\n >>> output.intermediate_output.image_embeds.shape\n torch.Size([1, 577, 768])\n >>> output.intermediate_output.decoder_labels.shape\n torch.Size([1, 13])\n ```\"\"\"\n\n image_embeds = self.forward_encoder(samples)\n decoder_output, decoder_targets = self.forward_decoder(samples, image_embeds)\n\n # return decoder_out\n return BlipOutput(\n loss=decoder_output.loss,\n loss_lm=decoder_output.loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n num_captions=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> samples = {\"image\": image}\n >>> captions = model.generate(samples)\n >>> captions\n ['a large statue of a person spraying water from a fountain']\n >>> captions = model.generate(samples, use_nucleus_sampling=True, num_captions=3)\n >>> captions # example output, results may vary due to randomness\n ['singapore showing the view of some building',\n 'the singapore harbor in twilight, as the weather is going down',\n 'the famous singapore fountain at sunset']\n \"\"\"\n # prepare inputs for decoder generation.\n encoder_out = self.forward_encoder(samples)\n image_embeds = torch.repeat_interleave(encoder_out, num_captions, 0)\n\n prompt = [self.prompt] * image_embeds.size(0)\n prompt = self.tokenizer(prompt, return_tensors=\"pt\").to(self.device)\n prompt.input_ids[:, 0] = self.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n # get decoded text\n decoder_out = self.text_decoder.generate_from_encoder(\n tokenized_prompt=prompt,\n visual_embeds=image_embeds,\n sep_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n use_nucleus_sampling=use_nucleus_sampling,\n num_beams=num_beams,\n max_length=max_length,\n min_length=min_length,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n )\n\n outputs = self.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n captions = [output[len(self.prompt) :] for output in outputs]\n\n return captions\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n # text encoder + multimodal decoder\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n prompt = cfg.get(\"prompt\", None)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n model = cls(image_encoder, text_decoder, prompt=prompt, max_txt_len=max_txt_len)\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.__init__#L40-L51","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":40,"end_line":51,"context_start_line":20,"context_end_line":71,"code":"@registry.register_model(\"blip_caption\")\nclass BlipCaption(BlipBase):\n \"\"\"\n BLIP captioning model.\n\n Supported model types:\n - base_coco: fine-tuned BLIP base model on COCO caption dataset (Karparthy split).\n - large_coco: fine-tuned BLIP large model on COCO caption dataset (Karparthy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_caption\", \"base_coco\")\n >>> model = load_model(\"blip_caption\", \"large_coco\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base_coco\": \"configs/models/blip_caption_base_coco.yaml\",\n \"large_coco\": \"configs/models/blip_caption_large_coco.yaml\",\n }\n\n def __init__(self, image_encoder, text_decoder, prompt=None, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_decoder = text_decoder\n\n self.prompt = prompt\n self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1\n\n self.max_txt_len = max_txt_len\n\n def forward_encoder(self, samples):\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n return image_embeds\n\n def forward_decoder(self, samples, image_embeds):\n # prepare inputs for forwarding decoder\n raw_text = samples[\"text_input\"]\n text = self.tokenizer(\n raw_text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n # prepare targets for forwarding decoder\n decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.forward_encoder","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.forward_encoder#L53-L55","kind":"function","name":"forward_encoder","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":53,"end_line":55,"context_start_line":33,"context_end_line":75,"code":" \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base_coco\": \"configs/models/blip_caption_base_coco.yaml\",\n \"large_coco\": \"configs/models/blip_caption_large_coco.yaml\",\n }\n\n def __init__(self, image_encoder, text_decoder, prompt=None, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_decoder = text_decoder\n\n self.prompt = prompt\n self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1\n\n self.max_txt_len = max_txt_len\n\n def forward_encoder(self, samples):\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n return image_embeds\n\n def forward_decoder(self, samples, image_embeds):\n # prepare inputs for forwarding decoder\n raw_text = samples[\"text_input\"]\n text = self.tokenizer(\n raw_text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n # prepare targets for forwarding decoder\n decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100\n )\n decoder_targets[:, : self.prompt_length] = -100\n\n # forward decoder","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.forward_decoder","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.forward_decoder#L57-L88","kind":"function","name":"forward_decoder","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":57,"end_line":88,"context_start_line":37,"context_end_line":108,"code":" \"large_coco\": \"configs/models/blip_caption_large_coco.yaml\",\n }\n\n def __init__(self, image_encoder, text_decoder, prompt=None, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_decoder = text_decoder\n\n self.prompt = prompt\n self.prompt_length = len(self.tokenizer(self.prompt).input_ids) - 1\n\n self.max_txt_len = max_txt_len\n\n def forward_encoder(self, samples):\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n return image_embeds\n\n def forward_decoder(self, samples, image_embeds):\n # prepare inputs for forwarding decoder\n raw_text = samples[\"text_input\"]\n text = self.tokenizer(\n raw_text,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n # prepare targets for forwarding decoder\n decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100\n )\n decoder_targets[:, : self.prompt_length] = -100\n\n # forward decoder\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n decoder_output = self.text_decoder(\n input_ids=text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n return decoder_output, decoder_targets\n\n def forward(self, samples):\n r\"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size.\n Returns:\n output (BlipOutput): A BlipOutput object containing the following\n attributes:\n - loss (torch.Tensor): A scalar tensor containing the total loss. For BlipCaption, this is the same as the LM loss.\n - loss_lm (torch.Tensor): A scalar tensor containing the LM loss.\n - intermediate_outputs (BlipIntermediateOutput): A BlipIntermediateOutput object containing intermediate outputs.\n see :class:`lavis.models.blip_models.blip_outputs.BlipOutput` for more details.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.forward#L90-L134","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":90,"end_line":134,"context_start_line":70,"context_end_line":154,"code":" decoder_targets = text.input_ids.masked_fill(\n text.input_ids == self.tokenizer.pad_token_id, -100\n )\n decoder_targets[:, : self.prompt_length] = -100\n\n # forward decoder\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n decoder_output = self.text_decoder(\n input_ids=text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n return decoder_output, decoder_targets\n\n def forward(self, samples):\n r\"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n - text_input (list): A list of strings of length batch_size.\n Returns:\n output (BlipOutput): A BlipOutput object containing the following\n attributes:\n - loss (torch.Tensor): A scalar tensor containing the total loss. For BlipCaption, this is the same as the LM loss.\n - loss_lm (torch.Tensor): A scalar tensor containing the LM loss.\n - intermediate_outputs (BlipIntermediateOutput): A BlipIntermediateOutput object containing intermediate outputs.\n see :class:`lavis.models.blip_models.blip_outputs.BlipOutput` for more details.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = [\"a large statue of a person spraying water from a fountain\"]\n >>> samples = {\"image\": image, \"text_input\": text_input}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss', 'loss_lm'])\n >>> output.intermediate_output.image_embeds.shape\n torch.Size([1, 577, 768])\n >>> output.intermediate_output.decoder_labels.shape\n torch.Size([1, 13])\n ```\"\"\"\n\n image_embeds = self.forward_encoder(samples)\n decoder_output, decoder_targets = self.forward_decoder(samples, image_embeds)\n\n # return decoder_out\n return BlipOutput(\n loss=decoder_output.loss,\n loss_lm=decoder_output.loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n num_captions=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.generate","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.generate#L136-L204","kind":"function","name":"generate","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":136,"end_line":204,"context_start_line":116,"context_end_line":219,"code":" >>> output.intermediate_output.image_embeds.shape\n torch.Size([1, 577, 768])\n >>> output.intermediate_output.decoder_labels.shape\n torch.Size([1, 13])\n ```\"\"\"\n\n image_embeds = self.forward_encoder(samples)\n decoder_output, decoder_targets = self.forward_decoder(samples, image_embeds)\n\n # return decoder_out\n return BlipOutput(\n loss=decoder_output.loss,\n loss_lm=decoder_output.loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def generate(\n self,\n samples,\n use_nucleus_sampling=False,\n num_beams=3,\n max_length=30,\n min_length=10,\n top_p=0.9,\n repetition_penalty=1.0,\n num_captions=1,\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W)\n use_nucleus_sampling (bool): Whether to use nucleus sampling. If False, use top-k sampling.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n max_length (int): The maximum length of the sequence to be generated.\n min_length (int): The minimum length of the sequence to be generated.\n top_p (float): The cumulative probability for nucleus sampling.\n repetition_penalty (float): The parameter for repetition penalty. 1.0 means no penalty.\n num_captions (int): Number of captions to be generated for each image.\n Returns:\n captions (list): A list of strings of length batch_size * num_captions.\n\n Example:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_caption\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> samples = {\"image\": image}\n >>> captions = model.generate(samples)\n >>> captions\n ['a large statue of a person spraying water from a fountain']\n >>> captions = model.generate(samples, use_nucleus_sampling=True, num_captions=3)\n >>> captions # example output, results may vary due to randomness\n ['singapore showing the view of some building',\n 'the singapore harbor in twilight, as the weather is going down',\n 'the famous singapore fountain at sunset']\n \"\"\"\n # prepare inputs for decoder generation.\n encoder_out = self.forward_encoder(samples)\n image_embeds = torch.repeat_interleave(encoder_out, num_captions, 0)\n\n prompt = [self.prompt] * image_embeds.size(0)\n prompt = self.tokenizer(prompt, return_tensors=\"pt\").to(self.device)\n prompt.input_ids[:, 0] = self.tokenizer.bos_token_id\n prompt.input_ids = prompt.input_ids[:, :-1]\n\n # get decoded text\n decoder_out = self.text_decoder.generate_from_encoder(\n tokenized_prompt=prompt,\n visual_embeds=image_embeds,\n sep_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n use_nucleus_sampling=use_nucleus_sampling,\n num_beams=num_beams,\n max_length=max_length,\n min_length=min_length,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n )\n\n outputs = self.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n captions = [output[len(self.prompt) :] for output in outputs]\n\n return captions\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n # text encoder + multimodal decoder\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n prompt = cfg.get(\"prompt\", None)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n model = cls(image_encoder, text_decoder, prompt=prompt, max_txt_len=max_txt_len)\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_caption.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_caption.from_config#L207-L219","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_caption.py","language":"python","start_line":207,"end_line":219,"context_start_line":187,"context_end_line":219,"code":" # get decoded text\n decoder_out = self.text_decoder.generate_from_encoder(\n tokenized_prompt=prompt,\n visual_embeds=image_embeds,\n sep_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n use_nucleus_sampling=use_nucleus_sampling,\n num_beams=num_beams,\n max_length=max_length,\n min_length=min_length,\n top_p=top_p,\n repetition_penalty=repetition_penalty,\n )\n\n outputs = self.tokenizer.batch_decode(decoder_out, skip_special_tokens=True)\n captions = [output[len(self.prompt) :] for output in outputs]\n\n return captions\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n # text encoder + multimodal decoder\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n prompt = cfg.get(\"prompt\", None)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n model = cls(image_encoder, text_decoder, prompt=prompt, max_txt_len=max_txt_len)\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"343d4fd79df135e07484b40e42ba0a9954ab2a9e447ffa729f1180097f1d5169","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching","uri":"program://CREMA/module/lavis.models.blip_models.blip_image_text_matching#L1-L199","kind":"module","name":"lavis.models.blip_models.blip_image_text_matching","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":1,"end_line":199,"context_start_line":1,"context_end_line":199,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip_models.blip import BlipBase\nfrom torch import nn\nfrom lavis.models.med import XBertEncoder\n\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_image_text_matching\")\nclass BlipITM(BlipBase):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n\n Supported model types:\n - base: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n - large: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_image_text_matching\", \"base\")\n >>> model = load_model(\"blip_image_text_matching\", \"large\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_itm_base.yaml\",\n \"large\": \"configs/models/blip_itm_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=35):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.text_encoder = text_encoder\n\n self.visual_encoder = image_encoder\n\n self.max_txt_len = max_txt_len\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n if match_head == \"itm\":\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id # extra code\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output\n\n elif match_head == \"itc\":\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head='itm'):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, 3:]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == 'itm':\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n # print(output.last_hidden_state.shape)\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n itm_output = F.softmax(itm_output, dim=1)[:,1]\n return itm_output #, mask, token_length\n\n elif match_head == 'itc':\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(encoder_input_ids, attention_mask=text_attention_mask,\n return_dict=True, mode='text')\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1)\n\n sim = image_feat @ text_feat.t()\n return sim\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n\ndef compute_gradcam(model, visual_input, text_input, tokenized_text, block_num=6):\n model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.save_attention = True\n\n output = model({\"image\": visual_input, \"text_input\": text_input}, match_head=\"itm\")\n loss = output[:, 1].sum()\n\n model.zero_grad()\n loss.backward()\n with torch.no_grad():\n mask = tokenized_text.attention_mask.view(\n tokenized_text.attention_mask.size(0), 1, -1, 1, 1\n ) # (bsz,1,token_len, 1,1)\n token_length = tokenized_text.attention_mask.sum(dim=-1) - 2\n token_length = token_length.cpu()\n # grads and cams [bsz, num_head, seq_len, image_patch]\n grads = model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.get_attn_gradients()\n cams = model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.get_attention_map()\n\n # assume using vit with 576 num image patch\n cams = cams[:, :, :, 1:].reshape(visual_input.size(0), 12, -1, 24, 24) * mask\n grads = (\n grads[:, :, :, 1:].clamp(0).reshape(visual_input.size(0), 12, -1, 24, 24)\n * mask\n )\n\n gradcams = cams * grads\n gradcam_list = []\n\n for ind in range(visual_input.size(0)):\n token_length_ = token_length[ind]\n gradcam = gradcams[ind].mean(0).cpu().detach()\n # [enc token gradcam, average gradcam across token, gradcam for individual token]\n gradcam = torch.cat(\n (\n gradcam[0:1, :],\n gradcam[1 : token_length_ + 1, :].sum(dim=0, keepdim=True)\n / token_length_,\n gradcam[1:, :],\n )\n )\n gradcam_list.append(gradcam)\n \n return gradcam_list, output","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.BlipITM","uri":"program://CREMA/class/lavis.models.blip_models.blip_image_text_matching.BlipITM#L19-L148","kind":"class","name":"BlipITM","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":19,"end_line":148,"context_start_line":1,"context_end_line":168,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip_models.blip import BlipBase\nfrom torch import nn\nfrom lavis.models.med import XBertEncoder\n\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_image_text_matching\")\nclass BlipITM(BlipBase):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n\n Supported model types:\n - base: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n - large: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_image_text_matching\", \"base\")\n >>> model = load_model(\"blip_image_text_matching\", \"large\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_itm_base.yaml\",\n \"large\": \"configs/models/blip_itm_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=35):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.text_encoder = text_encoder\n\n self.visual_encoder = image_encoder\n\n self.max_txt_len = max_txt_len\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n if match_head == \"itm\":\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id # extra code\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output\n\n elif match_head == \"itc\":\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head='itm'):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, 3:]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == 'itm':\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n # print(output.last_hidden_state.shape)\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n itm_output = F.softmax(itm_output, dim=1)[:,1]\n return itm_output #, mask, token_length\n\n elif match_head == 'itc':\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(encoder_input_ids, attention_mask=text_attention_mask,\n return_dict=True, mode='text')\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1)\n\n sim = image_feat @ text_feat.t()\n return sim\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n\ndef compute_gradcam(model, visual_input, text_input, tokenized_text, block_num=6):\n model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.save_attention = True\n\n output = model({\"image\": visual_input, \"text_input\": text_input}, match_head=\"itm\")\n loss = output[:, 1].sum()\n\n model.zero_grad()\n loss.backward()\n with torch.no_grad():\n mask = tokenized_text.attention_mask.view(\n tokenized_text.attention_mask.size(0), 1, -1, 1, 1\n ) # (bsz,1,token_len, 1,1)\n token_length = tokenized_text.attention_mask.sum(dim=-1) - 2\n token_length = token_length.cpu()\n # grads and cams [bsz, num_head, seq_len, image_patch]\n grads = model.text_encoder.base_model.base_model.encoder.layer[","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.compute_gradcam","uri":"program://CREMA/function/lavis.models.blip_models.blip_image_text_matching.compute_gradcam#L151-L199","kind":"function","name":"compute_gradcam","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":151,"end_line":199,"context_start_line":131,"context_end_line":199,"code":" @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n\ndef compute_gradcam(model, visual_input, text_input, tokenized_text, block_num=6):\n model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.save_attention = True\n\n output = model({\"image\": visual_input, \"text_input\": text_input}, match_head=\"itm\")\n loss = output[:, 1].sum()\n\n model.zero_grad()\n loss.backward()\n with torch.no_grad():\n mask = tokenized_text.attention_mask.view(\n tokenized_text.attention_mask.size(0), 1, -1, 1, 1\n ) # (bsz,1,token_len, 1,1)\n token_length = tokenized_text.attention_mask.sum(dim=-1) - 2\n token_length = token_length.cpu()\n # grads and cams [bsz, num_head, seq_len, image_patch]\n grads = model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.get_attn_gradients()\n cams = model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.get_attention_map()\n\n # assume using vit with 576 num image patch\n cams = cams[:, :, :, 1:].reshape(visual_input.size(0), 12, -1, 24, 24) * mask\n grads = (\n grads[:, :, :, 1:].clamp(0).reshape(visual_input.size(0), 12, -1, 24, 24)\n * mask\n )\n\n gradcams = cams * grads\n gradcam_list = []\n\n for ind in range(visual_input.size(0)):\n token_length_ = token_length[ind]\n gradcam = gradcams[ind].mean(0).cpu().detach()\n # [enc token gradcam, average gradcam across token, gradcam for individual token]\n gradcam = torch.cat(\n (\n gradcam[0:1, :],\n gradcam[1 : token_length_ + 1, :].sum(dim=0, keepdim=True)\n / token_length_,\n gradcam[1:, :],\n )\n )\n gradcam_list.append(gradcam)\n \n return gradcam_list, output","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_image_text_matching.__init__#L38-L56","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":38,"end_line":56,"context_start_line":18,"context_end_line":76,"code":"@registry.register_model(\"blip_image_text_matching\")\nclass BlipITM(BlipBase):\n \"\"\"\n BLIP Image-Text Matching (ITM) model.\n\n Supported model types:\n - base: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n - large: fine-tuned BLIP retrieval weights on COCO dataset (Karpathy split).\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_image_text_matching\", \"base\")\n >>> model = load_model(\"blip_image_text_matching\", \"large\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_itm_base.yaml\",\n \"large\": \"configs/models/blip_itm_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=35):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.text_encoder = text_encoder\n\n self.visual_encoder = image_encoder\n\n self.max_txt_len = max_txt_len\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n if match_head == \"itm\":\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id # extra code","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_image_text_matching.forward#L58-L100","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":58,"end_line":100,"context_start_line":38,"context_end_line":120,"code":" def __init__(self, image_encoder, text_encoder, embed_dim=256, max_txt_len=35):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.text_encoder = text_encoder\n\n self.visual_encoder = image_encoder\n\n self.max_txt_len = max_txt_len\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n def forward(self, samples, match_head=\"itm\"):\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n\n text = self.tokenizer(\n caption,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n if match_head == \"itm\":\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id # extra code\n output = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output\n\n elif match_head == \"itc\":\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head='itm'):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, 3:]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == 'itm':\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n # print(output.last_hidden_state.shape)\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n itm_output = F.softmax(itm_output, dim=1)[:,1]\n return itm_output #, mask, token_length\n","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.itm_rank","uri":"program://CREMA/function/lavis.models.blip_models.blip_image_text_matching.itm_rank#L101-L129","kind":"function","name":"itm_rank","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":101,"end_line":129,"context_start_line":81,"context_end_line":149,"code":" encoder_attention_mask=image_atts,\n return_dict=True,\n )\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n return itm_output\n\n elif match_head == \"itc\":\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1\n )\n\n sim = image_feat @ text_feat.t()\n return sim\n def itm_rank(self, image_embeds, image_atts, encoder_input_ids, match_head='itm'):\n # breakpoint()\n encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids = encoder_input_ids[:, 3:]\n text_attention_mask = (encoder_input_ids != self.tokenizer.pad_token_id).long()\n\n if match_head == 'itm':\n # encoder_input_ids = encoder_input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n output = self.text_encoder(encoder_input_ids,\n attention_mask=text_attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n # print(output.last_hidden_state.shape)\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n itm_output = F.softmax(itm_output, dim=1)[:,1]\n return itm_output #, mask, token_length\n\n elif match_head == 'itc':\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(encoder_input_ids, attention_mask=text_attention_mask,\n return_dict=True, mode='text')\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1)\n\n sim = image_feat @ text_feat.t()\n return sim\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_image_text_matching.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_image_text_matching.from_config#L132-L148","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_image_text_matching.py","language":"python","start_line":132,"end_line":148,"context_start_line":112,"context_end_line":168,"code":" encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n # print(output.last_hidden_state.shape)\n itm_output = self.itm_head(output.last_hidden_state[:, 0, :])\n itm_output = F.softmax(itm_output, dim=1)[:,1]\n return itm_output #, mask, token_length\n\n elif match_head == 'itc':\n encoder_input_ids[:, 0] = self.tokenizer.cls_token_id\n text_output = self.text_encoder(encoder_input_ids, attention_mask=text_attention_mask,\n return_dict=True, mode='text')\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n text_feat = F.normalize(self.text_proj(text_output.last_hidden_state[:, 0, :]), dim=-1)\n\n sim = image_feat @ text_feat.t()\n return sim\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n\ndef compute_gradcam(model, visual_input, text_input, tokenized_text, block_num=6):\n model.text_encoder.base_model.base_model.encoder.layer[\n block_num\n ].crossattention.self.save_attention = True\n\n output = model({\"image\": visual_input, \"text_input\": text_input}, match_head=\"itm\")\n loss = output[:, 1].sum()\n\n model.zero_grad()\n loss.backward()\n with torch.no_grad():\n mask = tokenized_text.attention_mask.view(\n tokenized_text.attention_mask.size(0), 1, -1, 1, 1\n ) # (bsz,1,token_len, 1,1)\n token_length = tokenized_text.attention_mask.sum(dim=-1) - 2\n token_length = token_length.cpu()\n # grads and cams [bsz, num_head, seq_len, image_patch]\n grads = model.text_encoder.base_model.base_model.encoder.layer[","source_hash":"0d5ead4ea7b846f2dbab4da9ade3d89c74f1fc86f8ff3f673b2eef08c434d73f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification","uri":"program://CREMA/module/lavis.models.blip_models.blip_classification#L1-L177","kind":"module","name":"lavis.models.blip_models.blip_classification","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":1,"end_line":177,"context_start_line":1,"context_end_line":177,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipIntermediateOutput,\n BlipOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_classification\")\nclass BlipClassification(BlipBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_classification_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=40,\n use_distill=True,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n # return {\"loss\": loss}\n return BlipOutputWithLogits(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n use_distill = cfg.get(\"use_distill\", True)\n momentum = cfg.get(\"momentum\", 0.995)\n num_classes = cfg.get(\"num_classes\", -1)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n\n return model","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification.BlipClassification","uri":"program://CREMA/class/lavis.models.blip_models.blip_classification.BlipClassification#L25-L177","kind":"class","name":"BlipClassification","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":25,"end_line":177,"context_start_line":5,"context_end_line":177,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipIntermediateOutput,\n BlipOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_classification\")\nclass BlipClassification(BlipBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_classification_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=40,\n use_distill=True,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n # return {\"loss\": loss}\n return BlipOutputWithLogits(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n use_distill = cfg.get(\"use_distill\", True)\n momentum = cfg.get(\"momentum\", 0.995)\n num_classes = cfg.get(\"num_classes\", -1)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n\n return model","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_classification.__init__#L30-L72","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":30,"end_line":72,"context_start_line":10,"context_end_line":92,"code":"import torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipIntermediateOutput,\n BlipOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_classification\")\nclass BlipClassification(BlipBase, MomentumDistilationMixin):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_classification_base.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n num_classes,\n momentum=0.995,\n alpha=0.4,\n max_txt_len=40,\n use_distill=True,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.use_distill = use_distill\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification._rampup_factor","uri":"program://CREMA/function/lavis.models.blip_models.blip_classification._rampup_factor#L74-L75","kind":"function","name":"_rampup_factor","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":74,"end_line":75,"context_start_line":54,"context_end_line":95,"code":" )\n\n if self.use_distill:\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_classification.forward#L77-L140","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":77,"end_line":140,"context_start_line":57,"context_end_line":160,"code":" self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n self.cls_head_m = deepcopy(self.cls_head)\n\n self.momentum = momentum\n self.alpha = alpha\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.cls_head, self.cls_head_m],\n ]\n\n self.copy_params()\n\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / num_iters_per_epoch)\n\n def forward(self, samples, is_train=True):\n sentences = samples[\"text_input\"]\n sentences = self.tokenizer(\n sentences,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n samples.update({\"tokenized_text\": sentences})\n\n targets = samples[\"label\"]\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n samples[\"tokenized_text\"], image_embeds\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n if self.use_distill:\n with torch.no_grad():\n self._momentum_update()\n\n image_embeds_m = self.visual_encoder_m(samples[\"image\"])\n encoder_output_m = self.text_encoder_m.forward_automask(\n samples[\"tokenized_text\"], image_embeds_m\n )\n\n prediction_m = self.cls_head_m(\n encoder_output_m.last_hidden_state[:, 0, :]\n )\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n loss = (1 - alpha) * F.cross_entropy(\n prediction, targets\n ) - alpha * torch.sum(\n F.log_softmax(prediction, dim=1) * F.softmax(prediction_m, dim=1),\n dim=1,\n ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n # return {\"loss\": loss}\n return BlipOutputWithLogits(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n use_distill = cfg.get(\"use_distill\", True)\n momentum = cfg.get(\"momentum\", 0.995)\n num_classes = cfg.get(\"num_classes\", -1)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification.predict","uri":"program://CREMA/function/lavis.models.blip_models.blip_classification.predict#L142-L144","kind":"function","name":"predict","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":142,"end_line":144,"context_start_line":122,"context_end_line":164,"code":" ).mean()\n else:\n loss = F.cross_entropy(prediction, targets)\n\n # return {\"loss\": loss}\n return BlipOutputWithLogits(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n use_distill = cfg.get(\"use_distill\", True)\n momentum = cfg.get(\"momentum\", 0.995)\n num_classes = cfg.get(\"num_classes\", -1)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_classification.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_classification.from_config#L147-L177","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_classification.py","language":"python","start_line":147,"end_line":177,"context_start_line":127,"context_end_line":177,"code":" return BlipOutputWithLogits(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n image_embeds_m=image_embeds_m,\n encoder_output=encoder_output,\n encoder_output_m=encoder_output_m,\n ),\n logits=prediction,\n logits_m=prediction_m,\n )\n\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n use_distill = cfg.get(\"use_distill\", True)\n momentum = cfg.get(\"momentum\", 0.995)\n num_classes = cfg.get(\"num_classes\", -1)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 40)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n use_distill=use_distill,\n alpha=alpha,\n num_classes=num_classes,\n momentum=momentum,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n\n return model","source_hash":"c2bf4c539d469eb5cae16af25ee633324a831e5ee3c9af0e24a5f777758052bc","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder","uri":"program://CREMA/module/lavis.models.blip_models.nlvr_encoder#L1-L960","kind":"module","name":"lavis.models.blip_models.nlvr_encoder","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":1,"end_line":960,"context_start_line":1,"context_end_line":960,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport math\nfrom typing import Tuple\n\nimport torch\nimport torch.utils.checkpoint\nfrom torch import Tensor, device, nn\nfrom transformers.activations import ACT2FN\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPastAndCrossAttentions,\n BaseModelOutputWithPoolingAndCrossAttentions,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.models.bert.configuration_bert import BertConfig\nfrom transformers.utils import logging\n\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n position_ids=None,\n inputs_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n input_shape = input_ids.size()\n else:\n input_shape = inputs_embeds.size()[:-1]\n\n seq_length = input_shape[1]\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n elif past_key_value is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config, twin=False, merge=False):\n super().__init__()\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n if twin:\n self.dense0 = nn.Linear(config.hidden_size, config.hidden_size)\n self.dense1 = nn.Linear(config.hidden_size, config.hidden_size)\n else:\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if merge:\n self.act = ACT2FN[config.hidden_act]\n self.merge_layer = nn.Linear(config.hidden_size * 2, config.hidden_size)\n self.merge = True\n else:\n self.merge = False\n\n def forward(self, hidden_states, input_tensor):\n if type(hidden_states) == list:\n hidden_states0 = self.dense0(hidden_states[0])\n hidden_states1 = self.dense1(hidden_states[1])\n if self.merge:\n # hidden_states = self.merge_layer(self.act(torch.cat([hidden_states0,hidden_states1],dim=-1)))\n hidden_states = self.merge_layer(\n torch.cat([hidden_states0, hidden_states1], dim=-1)\n )\n else:\n hidden_states = (hidden_states0 + hidden_states1) / 2\n else:\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False, layer_num=-1):\n super().__init__()\n if is_cross_attention:\n self.self0 = BertSelfAttention(config, is_cross_attention)\n self.self1 = BertSelfAttention(config, is_cross_attention)\n else:\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(\n config,\n twin=is_cross_attention,\n merge=(is_cross_attention and layer_num >= 6),\n )\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n if type(encoder_hidden_states) == list:\n self_outputs0 = self.self0(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states[0],\n encoder_attention_mask[0],\n past_key_value,\n output_attentions,\n )\n self_outputs1 = self.self1(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states[1],\n encoder_attention_mask[1],\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(\n [self_outputs0[0], self_outputs1[0]], hidden_states\n )\n\n outputs = (attention_output,) + self_outputs0[\n 1:\n ] # add attentions if we output them\n else:\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if self.config.add_cross_attention:\n self.crossattention = BertAttention(\n config,\n is_cross_attention=self.config.add_cross_attention,\n layer_num=layer_num,\n )\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n mode=None,\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n )\n attention_output = self_attention_outputs[0]\n\n outputs = self_attention_outputs[1:-1]\n present_key_value = self_attention_outputs[-1]\n\n if mode == \"multimodal\":\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,\n output_hidden_states=False,\n return_dict=True,\n mode=\"multimodal\",\n ):\n all_hidden_states = () if output_hidden_states else None\n all_self_attentions = () if output_attentions else None\n all_cross_attentions = (\n () if output_attentions and self.config.add_cross_attention else None\n )\n\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompat\n# ... truncated ...","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertEmbeddings","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertEmbeddings#L31-L87","kind":"class","name":"BertEmbeddings","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":31,"end_line":87,"context_start_line":11,"context_end_line":107,"code":"import torch\nimport torch.utils.checkpoint\nfrom torch import Tensor, device, nn\nfrom transformers.activations import ACT2FN\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPastAndCrossAttentions,\n BaseModelOutputWithPoolingAndCrossAttentions,\n)\nfrom transformers.modeling_utils import (\n PreTrainedModel,\n apply_chunking_to_forward,\n find_pruneable_heads_and_indices,\n prune_linear_layer,\n)\nfrom transformers.models.bert.configuration_bert import BertConfig\nfrom transformers.utils import logging\n\nlogger = logging.get_logger(__name__)\n\n\nclass BertEmbeddings(nn.Module):\n \"\"\"Construct the embeddings from word and position embeddings.\"\"\"\n\n def __init__(self, config):\n super().__init__()\n self.word_embeddings = nn.Embedding(\n config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id\n )\n self.position_embeddings = nn.Embedding(\n config.max_position_embeddings, config.hidden_size\n )\n\n # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load\n # any TensorFlow checkpoint file\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n # position_ids (1, len position emb) is contiguous in memory and exported when serialized\n self.register_buffer(\n \"position_ids\", torch.arange(config.max_position_embeddings).expand((1, -1))\n )\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n\n self.config = config\n\n def forward(\n self,\n input_ids=None,\n position_ids=None,\n inputs_embeds=None,\n past_key_values_length=0,\n ):\n if input_ids is not None:\n input_shape = input_ids.size()\n else:\n input_shape = inputs_embeds.size()[:-1]\n\n seq_length = input_shape[1]\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertSelfAttention","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertSelfAttention#L90-L253","kind":"class","name":"BertSelfAttention","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":90,"end_line":253,"context_start_line":70,"context_end_line":273,"code":" seq_length = input_shape[1]\n\n if position_ids is None:\n position_ids = self.position_ids[\n :, past_key_values_length : seq_length + past_key_values_length\n ]\n\n if inputs_embeds is None:\n inputs_embeds = self.word_embeddings(input_ids)\n\n embeddings = inputs_embeds\n\n if self.position_embedding_type == \"absolute\":\n position_embeddings = self.position_embeddings(position_ids)\n embeddings += position_embeddings\n embeddings = self.LayerNorm(embeddings)\n embeddings = self.dropout(embeddings)\n return embeddings\n\n\nclass BertSelfAttention(nn.Module):\n def __init__(self, config, is_cross_attention):\n super().__init__()\n self.config = config\n if config.hidden_size % config.num_attention_heads != 0 and not hasattr(\n config, \"embedding_size\"\n ):\n raise ValueError(\n \"The hidden size (%d) is not a multiple of the number of attention \"\n \"heads (%d)\" % (config.hidden_size, config.num_attention_heads)\n )\n\n self.num_attention_heads = config.num_attention_heads\n self.attention_head_size = int(config.hidden_size / config.num_attention_heads)\n self.all_head_size = self.num_attention_heads * self.attention_head_size\n\n self.query = nn.Linear(config.hidden_size, self.all_head_size)\n if is_cross_attention:\n self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))\n value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))\n attention_mask = encoder_attention_mask\n elif past_key_value is not None:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n key_layer = torch.cat([past_key_value[0], key_layer], dim=2)\n value_layer = torch.cat([past_key_value[1], value_layer], dim=2)\n else:\n key_layer = self.transpose_for_scores(self.key(hidden_states))\n value_layer = self.transpose_for_scores(self.value(hidden_states))\n\n query_layer = self.transpose_for_scores(mixed_query_layer)\n\n past_key_value = (key_layer, value_layer)\n\n # Take the dot product between \"query\" and \"key\" to get the raw attention scores.\n attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))\n\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n seq_length = hidden_states.size()[1]\n position_ids_l = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(-1, 1)\n position_ids_r = torch.arange(\n seq_length, dtype=torch.long, device=hidden_states.device\n ).view(1, -1)\n distance = position_ids_l - position_ids_r\n positional_embedding = self.distance_embedding(\n distance + self.max_position_embeddings - 1\n )\n positional_embedding = positional_embedding.to(\n dtype=query_layer.dtype\n ) # fp16 compatibility\n\n if self.position_embedding_type == \"relative_key\":\n relative_position_scores = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n attention_scores = attention_scores + relative_position_scores\n elif self.position_embedding_type == \"relative_key_query\":\n relative_position_scores_query = torch.einsum(\n \"bhld,lrd->bhlr\", query_layer, positional_embedding\n )\n relative_position_scores_key = torch.einsum(\n \"bhrd,lrd->bhlr\", key_layer, positional_embedding\n )\n attention_scores = (\n attention_scores\n + relative_position_scores_query\n + relative_position_scores_key\n )\n\n attention_scores = attention_scores / math.sqrt(self.attention_head_size)\n if attention_mask is not None:\n # Apply the attention mask is (precomputed for all layers in BertModel forward() function)\n attention_scores = attention_scores + attention_mask\n\n # Normalize the attention scores to probabilities.\n attention_probs = nn.Softmax(dim=-1)(attention_scores)\n\n if is_cross_attention and self.save_attention:\n self.save_attention_map(attention_probs)\n attention_probs.register_hook(self.save_attn_gradients)\n\n # This is actually dropping out entire tokens to attend to, which might\n # seem a bit unusual, but is taken from the original Transformer paper.\n attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config, twin=False, merge=False):\n super().__init__()\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n if twin:\n self.dense0 = nn.Linear(config.hidden_size, config.hidden_size)\n self.dense1 = nn.Linear(config.hidden_size, config.hidden_size)\n else:\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if merge:\n self.act = ACT2FN[config.hidden_act]\n self.merge_layer = nn.Linear(config.hidden_size * 2, config.hidden_size)\n self.merge = True\n else:\n self.merge = False\n\n def forward(self, hidden_states, input_tensor):","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertSelfOutput","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertSelfOutput#L256-L288","kind":"class","name":"BertSelfOutput","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":256,"end_line":288,"context_start_line":236,"context_end_line":308,"code":" attention_probs_dropped = self.dropout(attention_probs)\n\n # Mask heads if we want to\n if head_mask is not None:\n attention_probs_dropped = attention_probs_dropped * head_mask\n\n context_layer = torch.matmul(attention_probs_dropped, value_layer)\n\n context_layer = context_layer.permute(0, 2, 1, 3).contiguous()\n new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)\n context_layer = context_layer.view(*new_context_layer_shape)\n\n outputs = (\n (context_layer, attention_probs) if output_attentions else (context_layer,)\n )\n\n outputs = outputs + (past_key_value,)\n return outputs\n\n\nclass BertSelfOutput(nn.Module):\n def __init__(self, config, twin=False, merge=False):\n super().__init__()\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n if twin:\n self.dense0 = nn.Linear(config.hidden_size, config.hidden_size)\n self.dense1 = nn.Linear(config.hidden_size, config.hidden_size)\n else:\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if merge:\n self.act = ACT2FN[config.hidden_act]\n self.merge_layer = nn.Linear(config.hidden_size * 2, config.hidden_size)\n self.merge = True\n else:\n self.merge = False\n\n def forward(self, hidden_states, input_tensor):\n if type(hidden_states) == list:\n hidden_states0 = self.dense0(hidden_states[0])\n hidden_states1 = self.dense1(hidden_states[1])\n if self.merge:\n # hidden_states = self.merge_layer(self.act(torch.cat([hidden_states0,hidden_states1],dim=-1)))\n hidden_states = self.merge_layer(\n torch.cat([hidden_states0, hidden_states1], dim=-1)\n )\n else:\n hidden_states = (hidden_states0 + hidden_states1) / 2\n else:\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False, layer_num=-1):\n super().__init__()\n if is_cross_attention:\n self.self0 = BertSelfAttention(config, is_cross_attention)\n self.self1 = BertSelfAttention(config, is_cross_attention)\n else:\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(\n config,\n twin=is_cross_attention,\n merge=(is_cross_attention and layer_num >= 6),\n )\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertAttention","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertAttention#L291-L379","kind":"class","name":"BertAttention","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":291,"end_line":379,"context_start_line":271,"context_end_line":399,"code":" self.merge = False\n\n def forward(self, hidden_states, input_tensor):\n if type(hidden_states) == list:\n hidden_states0 = self.dense0(hidden_states[0])\n hidden_states1 = self.dense1(hidden_states[1])\n if self.merge:\n # hidden_states = self.merge_layer(self.act(torch.cat([hidden_states0,hidden_states1],dim=-1)))\n hidden_states = self.merge_layer(\n torch.cat([hidden_states0, hidden_states1], dim=-1)\n )\n else:\n hidden_states = (hidden_states0 + hidden_states1) / 2\n else:\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False, layer_num=-1):\n super().__init__()\n if is_cross_attention:\n self.self0 = BertSelfAttention(config, is_cross_attention)\n self.self1 = BertSelfAttention(config, is_cross_attention)\n else:\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(\n config,\n twin=is_cross_attention,\n merge=(is_cross_attention and layer_num >= 6),\n )\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n if type(encoder_hidden_states) == list:\n self_outputs0 = self.self0(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states[0],\n encoder_attention_mask[0],\n past_key_value,\n output_attentions,\n )\n self_outputs1 = self.self1(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states[1],\n encoder_attention_mask[1],\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(\n [self_outputs0[0], self_outputs1[0]], hidden_states\n )\n\n outputs = (attention_output,) + self_outputs0[\n 1:\n ] # add attentions if we output them\n else:\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertIntermediate","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertIntermediate#L382-L394","kind":"class","name":"BertIntermediate","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":382,"end_line":394,"context_start_line":362,"context_end_line":414,"code":" outputs = (attention_output,) + self_outputs0[\n 1:\n ] # add attentions if we output them\n else:\n self_outputs = self.self(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n )\n attention_output = self.output(self_outputs[0], hidden_states)\n outputs = (attention_output,) + self_outputs[\n 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertOutput","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertOutput#L397-L408","kind":"class","name":"BertOutput","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":397,"end_line":408,"context_start_line":377,"context_end_line":428,"code":" 1:\n ] # add attentions if we output them\n return outputs\n\n\nclass BertIntermediate(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.intermediate_size)\n if isinstance(config.hidden_act, str):\n self.intermediate_act_fn = ACT2FN[config.hidden_act]\n else:\n self.intermediate_act_fn = config.hidden_act\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if self.config.add_cross_attention:\n self.crossattention = BertAttention(\n config,\n is_cross_attention=self.config.add_cross_attention,\n layer_num=layer_num,\n )\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n def forward(","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertLayer","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertLayer#L411-L486","kind":"class","name":"BertLayer","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":411,"end_line":486,"context_start_line":391,"context_end_line":506,"code":" def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.intermediate_act_fn(hidden_states)\n return hidden_states\n\n\nclass BertOutput(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.intermediate_size, config.hidden_size)\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n self.dropout = nn.Dropout(config.hidden_dropout_prob)\n\n def forward(self, hidden_states, input_tensor):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertLayer(nn.Module):\n def __init__(self, config, layer_num):\n super().__init__()\n self.config = config\n self.chunk_size_feed_forward = config.chunk_size_feed_forward\n self.seq_len_dim = 1\n self.attention = BertAttention(config)\n self.layer_num = layer_num\n if self.config.add_cross_attention:\n self.crossattention = BertAttention(\n config,\n is_cross_attention=self.config.add_cross_attention,\n layer_num=layer_num,\n )\n self.intermediate = BertIntermediate(config)\n self.output = BertOutput(config)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n mode=None,\n ):\n # decoder uni-directional self-attention cached key/values tuple is at positions 1,2\n self_attn_past_key_value = (\n past_key_value[:2] if past_key_value is not None else None\n )\n self_attention_outputs = self.attention(\n hidden_states,\n attention_mask,\n head_mask,\n output_attentions=output_attentions,\n past_key_value=self_attn_past_key_value,\n )\n attention_output = self_attention_outputs[0]\n\n outputs = self_attention_outputs[1:-1]\n present_key_value = self_attention_outputs[-1]\n\n if mode == \"multimodal\":\n assert (\n encoder_hidden_states is not None\n ), \"encoder_hidden_states must be given for cross-attention layers\"\n cross_attention_outputs = self.crossattention(\n attention_output,\n attention_mask,\n head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertEncoder","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertEncoder#L489-L590","kind":"class","name":"BertEncoder","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":489,"end_line":590,"context_start_line":469,"context_end_line":610,"code":" outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=False,\n output_hidden_states=False,\n return_dict=True,\n mode=\"multimodal\",\n ):\n all_hidden_states = () if output_hidden_states else None\n all_self_attentions = () if output_attentions else None\n all_cross_attentions = (\n () if output_attentions and self.config.add_cross_attention else None\n )\n\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n mode=mode,\n )\n\n hidden_states = layer_outputs[0]\n if use_cache:\n next_decoder_cache += (layer_outputs[-1],)\n if output_attentions:\n all_self_attentions = all_self_attentions + (layer_outputs[1],)\n\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n if not return_dict:\n return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertPooler","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertPooler#L593-L605","kind":"class","name":"BertPooler","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":593,"end_line":605,"context_start_line":573,"context_end_line":625,"code":" return tuple(\n v\n for v in [\n hidden_states,\n next_decoder_cache,\n all_hidden_states,\n all_self_attentions,\n all_cross_attentions,\n ]\n if v is not None\n )\n return BaseModelOutputWithPastAndCrossAttentions(\n last_hidden_state=hidden_states,\n past_key_values=next_decoder_cache,\n hidden_states=all_hidden_states,\n attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertPredictionHeadTransform","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertPredictionHeadTransform#L608-L622","kind":"class","name":"BertPredictionHeadTransform","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":608,"end_line":622,"context_start_line":588,"context_end_line":642,"code":" attentions=all_self_attentions,\n cross_attentions=all_cross_attentions,\n )\n\n\nclass BertPooler(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n self.activation = nn.Tanh()\n\n def forward(self, hidden_states):\n # We \"pool\" the model by simply taking the hidden state corresponding\n # to the first token.\n first_token_tensor = hidden_states[:, 0]\n pooled_output = self.dense(first_token_tensor)\n pooled_output = self.activation(pooled_output)\n return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertLMPredictionHead","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertLMPredictionHead#L625-L642","kind":"class","name":"BertLMPredictionHead","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":625,"end_line":642,"context_start_line":605,"context_end_line":662,"code":" return pooled_output\n\n\nclass BertPredictionHeadTransform(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.dense = nn.Linear(config.hidden_size, config.hidden_size)\n if isinstance(config.hidden_act, str):\n self.transform_act_fn = ACT2FN[config.hidden_act]\n else:\n self.transform_act_fn = config.hidden_act\n self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)\n\n def forward(self, hidden_states):\n hidden_states = self.dense(hidden_states)\n hidden_states = self.transform_act_fn(hidden_states)\n hidden_states = self.LayerNorm(hidden_states)\n return hidden_states\n\n\nclass BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertOnlyMLMHead","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertOnlyMLMHead#L645-L652","kind":"class","name":"BertOnlyMLMHead","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":645,"end_line":652,"context_start_line":625,"context_end_line":672,"code":"class BertLMPredictionHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.transform = BertPredictionHeadTransform(config)\n\n # The output weights are the same as the input embeddings, but there is\n # an output-only bias for each token.\n self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)\n\n self.bias = nn.Parameter(torch.zeros(config.vocab_size))\n\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertPreTrainedModel","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertPreTrainedModel#L655-L675","kind":"class","name":"BertPreTrainedModel","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":655,"end_line":675,"context_start_line":635,"context_end_line":695,"code":"\n # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`\n self.decoder.bias = self.bias\n\n def forward(self, hidden_states):\n hidden_states = self.transform(hidden_states)\n hidden_states = self.decoder(hidden_states)\n return hidden_states\n\n\nclass BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.BertModel","uri":"program://CREMA/class/lavis.models.blip_models.nlvr_encoder.BertModel#L678-L960","kind":"class","name":"BertModel","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":678,"end_line":960,"context_start_line":658,"context_end_line":960,"code":" models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n # in case past_key_values are used we need to add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, seq_length, prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n \"\"\"\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n if is_decoder:\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n else:\n use_cache = False\n\n if input_ids is not None and inputs_embeds is not None:\n raise ValueError(\n \"You cannot specify both input_ids and inputs_embeds at the same time\"\n )\n elif input_ids is not None:\n input_shape = input_ids.size()\n batch_size, seq_length = input_shape\n device = input_ids.device\n elif inputs_embeds is not None:\n input_shape = inputs_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = inputs_embeds.device\n elif encoder_embeds is not None:\n input_shape = encoder_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = encoder_embeds.device\n else:\n raise ValueError(\n \"You have to specify either input_ids or inputs_embeds or encoder_embeds\"\n )\n\n # past_key_values_length\n past_key_values_length = (\n past_key_values[0][0].shape[2] if past_key_values is not None else 0\n )\n\n if attention_mask is None:\n attention_mask = torch.ones(\n ((batch_size, seq_length + past_key_values_length)), device=device\n )\n\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(\n attention_mask, input_shape, device, is_decoder\n )\n\n # If a 2D or 3D attention mask is provided for the cross-attention\n # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if encoder_hidden_states is not None:\n if type(encoder_hidden_states) == list:\n encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[\n 0\n ].size()\n else:\n (\n encoder_batch_size,\n encoder_sequence_length,\n _,\n ) = encoder_hidden_states.size()\n encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)\n\n if type(encoder_attention_mask) == list:\n encoder_extended_attention_mask = [\n self.invert_attention_mask(mask) for mask in encoder_attention_mask\n ]\n elif encoder_attention_mask is None:\n encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = None\n\n # Prepare head mask if needed\n # 1.0 in head_mask indicate we keep the head\n # attention_probs has shape bsz x n_heads x N x N\n # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]\n # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]\n head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)\n\n if encoder_embeds is None:\n embedding_output = self.embeddings(\n input_ids=input_ids,\n position_ids=position_ids,\n inputs_embeds=inputs_embeds,\n past_key_values_length=past_key_values_length,\n )\n else:\n embedding_output = encoder_embeds\n\n encoder_outputs = self.encoder(\n embedding_output,\n attention_mask=extended_attention_mask,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n mode=mode,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.__init__","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.__init__#L688-L698","kind":"function","name":"__init__","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":688,"end_line":698,"context_start_line":668,"context_end_line":718,"code":" # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.forward","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.forward#L792-L960","kind":"function","name":"forward","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":792,"end_line":960,"context_start_line":772,"context_end_line":960,"code":" else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n ):\n r\"\"\"\n encoder_hidden_states (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):\n Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if\n the model is configured as a decoder.\n encoder_attention_mask (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length)`, `optional`):\n Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in\n the cross-attention if the model is configured as a decoder. Mask values selected in ``[0, 1]``:\n - 1 for tokens that are **not masked**,\n - 0 for tokens that are **masked**.\n past_key_values (:obj:`tuple(tuple(torch.FloatTensor))` of length :obj:`config.n_layers` with each tuple having 4 tensors of shape :obj:`(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):\n Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.\n If :obj:`past_key_values` are used, the user can optionally input only the last :obj:`decoder_input_ids`\n (those that don't have their past key value states given to this model) of shape :obj:`(batch_size, 1)`\n instead of all :obj:`decoder_input_ids` of shape :obj:`(batch_size, sequence_length)`.\n use_cache (:obj:`bool`, `optional`):\n If set to :obj:`True`, :obj:`past_key_values` key value states are returned and can be used to speed up\n decoding (see :obj:`past_key_values`).\n \"\"\"\n output_attentions = (\n output_attentions\n if output_attentions is not None\n else self.config.output_attentions\n )\n output_hidden_states = (\n output_hidden_states\n if output_hidden_states is not None\n else self.config.output_hidden_states\n )\n return_dict = (\n return_dict if return_dict is not None else self.config.use_return_dict\n )\n\n if is_decoder:\n use_cache = use_cache if use_cache is not None else self.config.use_cache\n else:\n use_cache = False\n\n if input_ids is not None and inputs_embeds is not None:\n raise ValueError(\n \"You cannot specify both input_ids and inputs_embeds at the same time\"\n )\n elif input_ids is not None:\n input_shape = input_ids.size()\n batch_size, seq_length = input_shape\n device = input_ids.device\n elif inputs_embeds is not None:\n input_shape = inputs_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = inputs_embeds.device\n elif encoder_embeds is not None:\n input_shape = encoder_embeds.size()[:-1]\n batch_size, seq_length = input_shape\n device = encoder_embeds.device\n else:\n raise ValueError(\n \"You have to specify either input_ids or inputs_embeds or encoder_embeds\"\n )\n\n # past_key_values_length\n past_key_values_length = (\n past_key_values[0][0].shape[2] if past_key_values is not None else 0\n )\n\n if attention_mask is None:\n attention_mask = torch.ones(\n ((batch_size, seq_length + past_key_values_length)), device=device\n )\n\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(\n attention_mask, input_shape, device, is_decoder\n )\n\n # If a 2D or 3D attention mask is provided for the cross-attention\n # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if encoder_hidden_states is not None:\n if type(encoder_hidden_states) == list:\n encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[\n 0\n ].size()\n else:\n (\n encoder_batch_size,\n encoder_sequence_length,\n _,\n ) = encoder_hidden_states.size()\n encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)\n\n if type(encoder_attention_mask) == list:\n encoder_extended_attention_mask = [\n self.invert_attention_mask(mask) for mask in encoder_attention_mask\n ]\n elif encoder_attention_mask is None:\n encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = self.invert_attention_mask(\n encoder_attention_mask\n )\n else:\n encoder_extended_attention_mask = None\n\n # Prepare head mask if needed\n # 1.0 in head_mask indicate we keep the head\n # attention_probs has shape bsz x n_heads x N x N\n # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]\n # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]\n head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)\n\n if encoder_embeds is None:\n embedding_output = self.embeddings(\n input_ids=input_ids,\n position_ids=position_ids,\n inputs_embeds=inputs_embeds,\n past_key_values_length=past_key_values_length,\n )\n else:\n embedding_output = encoder_embeds\n\n encoder_outputs = self.encoder(\n embedding_output,\n attention_mask=extended_attention_mask,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_extended_attention_mask,\n past_key_values=past_key_values,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n mode=mode,\n )\n sequence_output = encoder_outputs[0]\n pooled_output = (\n self.pooler(sequence_output) if self.pooler is not None else None\n )\n\n if not return_dict:\n return (sequence_output, pooled_output) + encoder_outputs[1:]\n\n return BaseModelOutputWithPoolingAndCrossAttentions(\n last_hidden_state=sequence_output,\n pooler_output=pooled_output,\n past_key_values=encoder_outputs.past_key_values,\n hidden_states=encoder_outputs.hidden_states,\n attentions=encoder_outputs.attentions,\n cross_attentions=encoder_outputs.cross_attentions,\n )","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.save_attn_gradients","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.save_attn_gradients#L128-L129","kind":"function","name":"save_attn_gradients","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":128,"end_line":129,"context_start_line":108,"context_end_line":149,"code":" self.key = nn.Linear(config.encoder_width, self.all_head_size)\n self.value = nn.Linear(config.encoder_width, self.all_head_size)\n else:\n self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.get_attn_gradients","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.get_attn_gradients#L131-L132","kind":"function","name":"get_attn_gradients","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":131,"end_line":132,"context_start_line":111,"context_end_line":152,"code":" self.key = nn.Linear(config.hidden_size, self.all_head_size)\n self.value = nn.Linear(config.hidden_size, self.all_head_size)\n\n self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.save_attention_map","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.save_attention_map#L134-L135","kind":"function","name":"save_attention_map","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":134,"end_line":135,"context_start_line":114,"context_end_line":155,"code":" self.dropout = nn.Dropout(config.attention_probs_dropout_prob)\n self.position_embedding_type = getattr(\n config, \"position_embedding_type\", \"absolute\"\n )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.get_attention_map","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.get_attention_map#L137-L138","kind":"function","name":"get_attention_map","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":137,"end_line":138,"context_start_line":117,"context_end_line":158,"code":" )\n if (\n self.position_embedding_type == \"relative_key\"\n or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.transpose_for_scores","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.transpose_for_scores#L140-L146","kind":"function","name":"transpose_for_scores","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":140,"end_line":146,"context_start_line":120,"context_end_line":166,"code":" or self.position_embedding_type == \"relative_key_query\"\n ):\n self.max_position_embeddings = config.max_position_embeddings\n self.distance_embedding = nn.Embedding(\n 2 * config.max_position_embeddings - 1, self.attention_head_size\n )\n self.save_attention = False\n\n def save_attn_gradients(self, attn_gradients):\n self.attn_gradients = attn_gradients\n\n def get_attn_gradients(self):\n return self.attn_gradients\n\n def save_attention_map(self, attention_map):\n self.attention_map = attention_map\n\n def get_attention_map(self):\n return self.attention_map\n\n def transpose_for_scores(self, x):\n new_x_shape = x.size()[:-1] + (\n self.num_attention_heads,\n self.attention_head_size,\n )\n x = x.view(*new_x_shape)\n return x.permute(0, 2, 1, 3)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n mixed_query_layer = self.query(hidden_states)\n\n # If this is instantiated as a cross-attention module, the keys\n # and values come from an encoder; the attention mask needs to be\n # such that the encoder's padding tokens are not attended to.\n is_cross_attention = encoder_hidden_states is not None\n\n if is_cross_attention:\n key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.prune_heads","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.prune_heads#L306-L327","kind":"function","name":"prune_heads","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":306,"end_line":327,"context_start_line":286,"context_end_line":347,"code":" hidden_states = self.dropout(hidden_states)\n hidden_states = self.LayerNorm(hidden_states + input_tensor)\n return hidden_states\n\n\nclass BertAttention(nn.Module):\n def __init__(self, config, is_cross_attention=False, layer_num=-1):\n super().__init__()\n if is_cross_attention:\n self.self0 = BertSelfAttention(config, is_cross_attention)\n self.self1 = BertSelfAttention(config, is_cross_attention)\n else:\n self.self = BertSelfAttention(config, is_cross_attention)\n self.output = BertSelfOutput(\n config,\n twin=is_cross_attention,\n merge=(is_cross_attention and layer_num >= 6),\n )\n self.pruned_heads = set()\n\n def prune_heads(self, heads):\n if len(heads) == 0:\n return\n heads, index = find_pruneable_heads_and_indices(\n heads,\n self.self.num_attention_heads,\n self.self.attention_head_size,\n self.pruned_heads,\n )\n\n # Prune linear layers\n self.self.query = prune_linear_layer(self.self.query, index)\n self.self.key = prune_linear_layer(self.self.key, index)\n self.self.value = prune_linear_layer(self.self.value, index)\n self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)\n\n # Update hyper params and store pruned heads\n self.self.num_attention_heads = self.self.num_attention_heads - len(heads)\n self.self.all_head_size = (\n self.self.attention_head_size * self.self.num_attention_heads\n )\n self.pruned_heads = self.pruned_heads.union(heads)\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_value=None,\n output_attentions=False,\n ):\n if type(encoder_hidden_states) == list:\n self_outputs0 = self.self0(\n hidden_states,\n attention_mask,\n head_mask,\n encoder_hidden_states[0],\n encoder_attention_mask[0],\n past_key_value,\n output_attentions,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.feed_forward_chunk","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.feed_forward_chunk#L483-L486","kind":"function","name":"feed_forward_chunk","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":483,"end_line":486,"context_start_line":463,"context_end_line":506,"code":" encoder_hidden_states,\n encoder_attention_mask,\n output_attentions=output_attentions,\n )\n attention_output = cross_attention_outputs[0]\n outputs = (\n outputs + cross_attention_outputs[1:-1]\n ) # add cross attentions if we output attention weights\n layer_output = apply_chunking_to_forward(\n self.feed_forward_chunk,\n self.chunk_size_feed_forward,\n self.seq_len_dim,\n attention_output,\n )\n outputs = (layer_output,) + outputs\n\n outputs = outputs + (present_key_value,)\n\n return outputs\n\n def feed_forward_chunk(self, attention_output):\n intermediate_output = self.intermediate(attention_output)\n layer_output = self.output(intermediate_output, attention_output)\n return layer_output\n\n\nclass BertEncoder(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.config = config\n self.layer = nn.ModuleList(\n [BertLayer(config, i) for i in range(config.num_hidden_layers)]\n )\n self.gradient_checkpointing = False\n\n def forward(\n self,\n hidden_states,\n attention_mask=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder._init_weights","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder._init_weights#L665-L675","kind":"function","name":"_init_weights","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":665,"end_line":675,"context_start_line":645,"context_end_line":695,"code":"class BertOnlyMLMHead(nn.Module):\n def __init__(self, config):\n super().__init__()\n self.predictions = BertLMPredictionHead(config)\n\n def forward(self, sequence_output):\n prediction_scores = self.predictions(sequence_output)\n return prediction_scores\n\n\nclass BertPreTrainedModel(PreTrainedModel):\n \"\"\"\n An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained\n models.\n \"\"\"\n\n config_class = BertConfig\n base_model_prefix = \"bert\"\n _keys_to_ignore_on_load_missing = [r\"position_ids\"]\n\n def _init_weights(self, module):\n \"\"\"Initialize the weights\"\"\"\n if isinstance(module, (nn.Linear, nn.Embedding)):\n # Slightly different from the TF version which uses truncated_normal for initialization\n # cf https://github.com/pytorch/pytorch/pull/5617\n module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)\n elif isinstance(module, nn.LayerNorm):\n module.bias.data.zero_()\n module.weight.data.fill_(1.0)\n if isinstance(module, nn.Linear) and module.bias is not None:\n module.bias.data.zero_()\n\n\nclass BertModel(BertPreTrainedModel):\n \"\"\"\n The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.get_input_embeddings","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.get_input_embeddings#L700-L701","kind":"function","name":"get_input_embeddings","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":700,"end_line":701,"context_start_line":680,"context_end_line":721,"code":" The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of\n cross-attention is added between the self-attention layers, following the architecture described in `Attention is\n all you need `__ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit,\n Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.set_input_embeddings","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.set_input_embeddings#L703-L704","kind":"function","name":"set_input_embeddings","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":703,"end_line":704,"context_start_line":683,"context_end_line":724,"code":" Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.\n argument and :obj:`add_cross_attention` set to :obj:`True`; an :obj:`encoder_hidden_states` is then expected as an\n input to the forward pass.\n \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder._prune_heads","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder._prune_heads#L706-L712","kind":"function","name":"_prune_heads","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":706,"end_line":712,"context_start_line":686,"context_end_line":732,"code":" \"\"\"\n\n def __init__(self, config, add_pooling_layer=True):\n super().__init__(config)\n self.config = config\n\n self.embeddings = BertEmbeddings(config)\n\n self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.get_extended_attention_mask","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.get_extended_attention_mask#L714-L790","kind":"function","name":"get_extended_attention_mask","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":714,"end_line":790,"context_start_line":694,"context_end_line":810,"code":" self.encoder = BertEncoder(config)\n\n self.pooler = BertPooler(config) if add_pooling_layer else None\n\n self.init_weights()\n\n def get_input_embeddings(self):\n return self.embeddings.word_embeddings\n\n def set_input_embeddings(self, value):\n self.embeddings.word_embeddings = value\n\n def _prune_heads(self, heads_to_prune):\n \"\"\"\n Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base\n class PreTrainedModel\n \"\"\"\n for layer, heads in heads_to_prune.items():\n self.encoder.layer[layer].attention.prune_heads(heads)\n\n def get_extended_attention_mask(\n self,\n attention_mask: Tensor,\n input_shape: Tuple[int],\n device: device,\n is_decoder: bool,\n ) -> Tensor:\n \"\"\"\n Makes broadcastable attention and causal masks so that future and masked tokens are ignored.\n\n Arguments:\n attention_mask (:obj:`torch.Tensor`):\n Mask with ones indicating tokens to attend to, zeros for tokens to ignore.\n input_shape (:obj:`Tuple[int]`):\n The shape of the input to the model.\n device: (:obj:`torch.device`):\n The device of the input to the model.\n\n Returns:\n :obj:`torch.Tensor` The extended attention mask, with a the same dtype as :obj:`attention_mask.dtype`.\n \"\"\"\n # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]\n # ourselves in which case we just need to make it broadcastable to all heads.\n if attention_mask.dim() == 3:\n extended_attention_mask = attention_mask[:, None, :, :]\n elif attention_mask.dim() == 2:\n # Provided a padding mask of dimensions [batch_size, seq_length]\n # - if the model is a decoder, apply a causal mask in addition to the padding mask\n # - if the model is an encoder, make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]\n if is_decoder:\n batch_size, seq_length = input_shape\n\n seq_ids = torch.arange(seq_length, device=device)\n causal_mask = (\n seq_ids[None, None, :].repeat(batch_size, seq_length, 1)\n <= seq_ids[None, :, None]\n )\n # in case past_key_values are used we need to add a prefix ones mask to the causal mask\n # causal and attention masks must have same type with pytorch version < 1.3\n causal_mask = causal_mask.to(attention_mask.dtype)\n\n if causal_mask.shape[1] < attention_mask.shape[1]:\n prefix_seq_len = attention_mask.shape[1] - causal_mask.shape[1]\n causal_mask = torch.cat(\n [\n torch.ones(\n (batch_size, seq_length, prefix_seq_len),\n device=device,\n dtype=causal_mask.dtype,\n ),\n causal_mask,\n ],\n axis=-1,\n )\n\n extended_attention_mask = (\n causal_mask[:, None, :, :] * attention_mask[:, None, None, :]\n )\n else:\n extended_attention_mask = attention_mask[:, None, None, :]\n else:\n raise ValueError(\n \"Wrong shape for input_ids (shape {}) or attention_mask (shape {})\".format(\n input_shape, attention_mask.shape\n )\n )\n\n # Since attention_mask is 1.0 for positions we want to attend and 0.0 for\n # masked positions, this operation will create a tensor which is 0.0 for\n # positions we want to attend and -10000.0 for masked positions.\n # Since we are adding it to the raw scores before the softmax, this is\n # effectively the same as removing these entirely.\n extended_attention_mask = extended_attention_mask.to(\n dtype=self.dtype\n ) # fp16 compatibility\n extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0\n return extended_attention_mask\n\n def forward(\n self,\n input_ids=None,\n attention_mask=None,\n position_ids=None,\n head_mask=None,\n inputs_embeds=None,\n encoder_embeds=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n past_key_values=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n is_decoder=False,\n mode=\"multimodal\",\n ):\n r\"\"\"","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.create_custom_forward","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.create_custom_forward#L536-L540","kind":"function","name":"create_custom_forward","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":536,"end_line":540,"context_start_line":516,"context_end_line":560,"code":" )\n\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,\n output_attentions,\n mode=mode,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.nlvr_encoder.custom_forward","uri":"program://CREMA/function/lavis.models.blip_models.nlvr_encoder.custom_forward#L537-L538","kind":"function","name":"custom_forward","path":"lavis/models/blip_models/nlvr_encoder.py","language":"python","start_line":537,"end_line":538,"context_start_line":517,"context_end_line":558,"code":"\n next_decoder_cache = () if use_cache else None\n\n for i in range(self.config.num_hidden_layers):\n layer_module = self.layer[i]\n if output_hidden_states:\n all_hidden_states = all_hidden_states + (hidden_states,)\n\n layer_head_mask = head_mask[i] if head_mask is not None else None\n past_key_value = past_key_values[i] if past_key_values is not None else None\n\n if self.gradient_checkpointing and self.training:\n\n if use_cache:\n logger.warn(\n \"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...\"\n )\n use_cache = False\n\n def create_custom_forward(module):\n def custom_forward(*inputs):\n return module(*inputs, past_key_value, output_attentions)\n\n return custom_forward\n\n layer_outputs = torch.utils.checkpoint.checkpoint(\n create_custom_forward(layer_module),\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n mode=mode,\n )\n else:\n layer_outputs = layer_module(\n hidden_states,\n attention_mask,\n layer_head_mask,\n encoder_hidden_states,\n encoder_attention_mask,\n past_key_value,","source_hash":"137792e6ec89ac212d158b69314f5bdae897ce64c6d71db3f1265bae86db90b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa","uri":"program://CREMA/module/lavis.models.blip_models.blip_vqa#L1-L375","kind":"module","name":"lavis.models.blip_models.blip_vqa","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":1,"end_line":375,"context_start_line":1,"context_end_line":375,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import tile\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder, XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_vqa\")\nclass BlipVQA(BlipBase):\n \"\"\"\n BLIP VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained BLIP base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned BLIP base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\", \"vqav2\")\n >>> model = load_model(\"blip_vqa\", \"okvqa\")\n >>> model = load_model(\"blip_vqa\", \"aokvqa\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/blip_vqav2.yaml\",\n \"okvqa\": \"configs/models/blip_vqa_okvqa.yaml\",\n \"aokvqa\": \"configs/models/blip_vqa_aokvqa.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, text_decoder, max_txt_len=35):\n super().__init__()\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n A BlipOutput object containing loss and intermediate outputs,\n see :class:`lavis.models.blip_outputs.BlipOutput` for more details.\n\n Examples:\n ```python\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 480, 480),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'encoder_output', 'decoder_output', 'decoder_labels'])\n ```\n \"\"\"\n encoder_output, image_embeds = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples=samples, encoder_out=encoder_output\n )\n\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n encoder_output=encoder_output,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n questions.input_ids[:, 0] = self.tokenizer.enc_token_id\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n return encoder_output, image_embeds\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answers.input_ids[:, 0] = self.tokenizer.bos_token_id\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(\n self,\n samples,\n num_beams=3,\n inference_method=\"rank\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n **kwargs\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. One of \"rank\", \"generate\".\n - If \"rank\", the model will return answers with the highest probability from the answer list.\n - If \"generate\", the model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answers = model.predict_answers(samples)\n >>> answers\n ['singapore']\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n ```\n \"\"\"\n assert inference_method in [\n \"rank\",\n \"generate\",\n ], \"Inference method must be one of 'rank' or 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n if inference_method == \"generate\":\n return self._generate_answers(\n samples, num_beams=num_beams, max_length=max_len, min_length=min_len\n )\n elif inference_method == \"rank\":\n assert answer_list is not None, \"answer_list must be provided for ranking\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self._rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def _generate_answers(self, samples, num_beams=3, max_length=10, min_length=1):\n encoder_out, _ = self.forward_encoder(samples)\n\n question_output = encoder_out\n\n question_states = question_output.last_hidden_state.repeat_interleave(\n num_beams, dim=0\n )\n question_atts = torch.ones(question_states.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": question_states,\n \"encoder_attention_mask\": question_atts,\n }\n\n bsz = samples[\"image\"].size(0)\n bos_ids = torch.full(\n (bsz, 1), fill_value=self.tokenizer.bos_token_id, device=self.device\n )\n\n outputs = self.text_decoder.generate(\n input_ids=bos_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n\n # collect answers\n answers = []\n for output in outputs:\n answer = self.tokenizer.decode(output, skip_special_tokens=True)\n answers.append(answer)\n\n return answers\n\n def _rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.BlipVQA","uri":"program://CREMA/class/lavis.models.blip_models.blip_vqa.BlipVQA#L22-L375","kind":"class","name":"BlipVQA","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":22,"end_line":375,"context_start_line":2,"context_end_line":375,"code":" Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import tile\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder, XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\n\n\n@registry.register_model(\"blip_vqa\")\nclass BlipVQA(BlipBase):\n \"\"\"\n BLIP VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained BLIP base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned BLIP base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\", \"vqav2\")\n >>> model = load_model(\"blip_vqa\", \"okvqa\")\n >>> model = load_model(\"blip_vqa\", \"aokvqa\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/blip_vqav2.yaml\",\n \"okvqa\": \"configs/models/blip_vqa_okvqa.yaml\",\n \"aokvqa\": \"configs/models/blip_vqa_aokvqa.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, text_decoder, max_txt_len=35):\n super().__init__()\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n A BlipOutput object containing loss and intermediate outputs,\n see :class:`lavis.models.blip_outputs.BlipOutput` for more details.\n\n Examples:\n ```python\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 480, 480),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'encoder_output', 'decoder_output', 'decoder_labels'])\n ```\n \"\"\"\n encoder_output, image_embeds = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples=samples, encoder_out=encoder_output\n )\n\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n encoder_output=encoder_output,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n questions.input_ids[:, 0] = self.tokenizer.enc_token_id\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n return encoder_output, image_embeds\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answers.input_ids[:, 0] = self.tokenizer.bos_token_id\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(\n self,\n samples,\n num_beams=3,\n inference_method=\"rank\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n **kwargs\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. One of \"rank\", \"generate\".\n - If \"rank\", the model will return answers with the highest probability from the answer list.\n - If \"generate\", the model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answers = model.predict_answers(samples)\n >>> answers\n ['singapore']\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n ```\n \"\"\"\n assert inference_method in [\n \"rank\",\n \"generate\",\n ], \"Inference method must be one of 'rank' or 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n if inference_method == \"generate\":\n return self._generate_answers(\n samples, num_beams=num_beams, max_length=max_len, min_length=min_len\n )\n elif inference_method == \"rank\":\n assert answer_list is not None, \"answer_list must be provided for ranking\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self._rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def _generate_answers(self, samples, num_beams=3, max_length=10, min_length=1):\n encoder_out, _ = self.forward_encoder(samples)\n\n question_output = encoder_out\n\n question_states = question_output.last_hidden_state.repeat_interleave(\n num_beams, dim=0\n )\n question_atts = torch.ones(question_states.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": question_states,\n \"encoder_attention_mask\": question_atts,\n }\n\n bsz = samples[\"image\"].size(0)\n bos_ids = torch.full(\n (bsz, 1), fill_value=self.tokenizer.bos_token_id, device=self.device\n )\n\n outputs = self.text_decoder.generate(\n input_ids=bos_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n\n # collect answers\n answers = []\n for output in outputs:\n answer = self.tokenizer.decode(output, skip_special_tokens=True)\n answers.append(answer)\n\n return answers\n\n def _rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.__init__#L43-L52","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":43,"end_line":52,"context_start_line":23,"context_end_line":72,"code":" \"\"\"\n BLIP VQA models.\n\n Supported model types:\n - base: vqa model initialized with pre-trained BLIP base model on 115M image-text pairs after CapFilt; not fine-tuned.\n - vqav2: fine-tuned BLIP base model on VQA v2.0 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\", \"vqav2\")\n >>> model = load_model(\"blip_vqa\", \"okvqa\")\n >>> model = load_model(\"blip_vqa\", \"aokvqa\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/blip_vqav2.yaml\",\n \"okvqa\": \"configs/models/blip_vqa_okvqa.yaml\",\n \"aokvqa\": \"configs/models/blip_vqa_aokvqa.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, text_decoder, max_txt_len=35):\n super().__init__()\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n A BlipOutput object containing loss and intermediate outputs,\n see :class:`lavis.models.blip_outputs.BlipOutput` for more details.\n\n Examples:\n ```python\n >>> import torch","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.forward#L54-L102","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":54,"end_line":102,"context_start_line":34,"context_end_line":122,"code":" >>> model = load_model(\"blip_vqa\", \"aokvqa\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"vqav2\": \"configs/models/blip_vqav2.yaml\",\n \"okvqa\": \"configs/models/blip_vqa_okvqa.yaml\",\n \"aokvqa\": \"configs/models/blip_vqa_aokvqa.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, text_decoder, max_txt_len=35):\n super().__init__()\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (list): A list of strings, each string is a question\n - answer (list): A list of strings, each string is an answer\n - weight (torch.Tensor): A tensor used to weigh each answer in the loss computation.\n The shape of the tensor is (sum(n_answers),)\n - n_answers (torch.Tensor): A tensor shape (batch_size,) containing the number of answers\n for each question in the batch.\n\n Returns:\n A BlipOutput object containing loss and intermediate outputs,\n see :class:`lavis.models.blip_outputs.BlipOutput` for more details.\n\n Examples:\n ```python\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_vqa\")\n >>> samples = {\n ... \"image\": torch.rand(2, 3, 480, 480),\n ... \"text_input\": [\"What is this?\", \"What is that?\"],\n ... \"answer\": [\"cat\", \"cat\", \"dog\"],\n ... \"weight\": torch.tensor([1.0, 1.0, 1.0]),\n ... \"n_answers\": torch.tensor([2, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'encoder_output', 'decoder_output', 'decoder_labels'])\n ```\n \"\"\"\n encoder_output, image_embeds = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples=samples, encoder_out=encoder_output\n )\n\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n encoder_output=encoder_output,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n questions.input_ids[:, 0] = self.tokenizer.enc_token_id\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n return encoder_output, image_embeds\n","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.forward_encoder","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.forward_encoder#L104-L121","kind":"function","name":"forward_encoder","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":104,"end_line":121,"context_start_line":84,"context_end_line":141,"code":" odict_keys(['intermediate_output', 'loss'])\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'encoder_output', 'decoder_output', 'decoder_labels'])\n ```\n \"\"\"\n encoder_output, image_embeds = self.forward_encoder(samples)\n loss, decoder_output, decoder_targets = self.forward_decoder(\n samples=samples, encoder_out=encoder_output\n )\n\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n encoder_output=encoder_output,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n questions.input_ids[:, 0] = self.tokenizer.enc_token_id\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n return encoder_output, image_embeds\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answers.input_ids[:, 0] = self.tokenizer.bos_token_id\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.forward_decoder","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.forward_decoder#L123-L160","kind":"function","name":"forward_decoder","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":123,"end_line":160,"context_start_line":103,"context_end_line":180,"code":"\n def forward_encoder(self, samples):\n questions = samples[\"text_input\"]\n questions = self.tokenizer(\n questions,\n padding=\"longest\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n questions.input_ids[:, 0] = self.tokenizer.enc_token_id\n samples.update({\"tokenized_text\": questions})\n\n image_embeds = self.visual_encoder.forward_features(samples[\"image\"])\n encoder_output = self.text_encoder.forward_automask(\n tokenized_text=samples[\"tokenized_text\"], visual_embeds=image_embeds\n )\n\n return encoder_output, image_embeds\n\n def forward_decoder(self, samples, encoder_out, **kwargs):\n answers = self.tokenizer(\n samples[\"answer\"], padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answers.input_ids[:, 0] = self.tokenizer.bos_token_id\n answer_targets = answers.input_ids.masked_fill(\n answers.input_ids == self.tokenizer.pad_token_id, -100\n )\n\n question_states = []\n question_atts = []\n\n question = samples[\"tokenized_text\"]\n question_output = encoder_out\n\n for b, n in enumerate(samples[\"n_answers\"]):\n question_states += [question_output.last_hidden_state[b]] * n\n question_atts += [question.attention_mask[b]] * n\n\n question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(\n self,\n samples,\n num_beams=3,\n inference_method=\"rank\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n **kwargs\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. One of \"rank\", \"generate\".","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.predict_answers","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.predict_answers#L162-L235","kind":"function","name":"predict_answers","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":162,"end_line":235,"context_start_line":142,"context_end_line":255,"code":" question_states = torch.stack(question_states, dim=0)\n question_atts = torch.stack(question_atts, dim=0)\n\n answer_output = self.text_decoder(\n answers.input_ids,\n attention_mask=answers.attention_mask,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=answer_targets,\n return_dict=True,\n reduction=\"none\",\n )\n\n loss = samples[\"weight\"] * answer_output.loss\n bsz = samples[\"image\"].size(0)\n\n loss = loss.sum() / bsz\n\n return loss, answer_output, answer_targets\n\n def predict_answers(\n self,\n samples,\n num_beams=3,\n inference_method=\"rank\",\n max_len=10,\n min_len=1,\n num_ans_candidates=128,\n answer_list=None,\n **kwargs\n ):\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). Default H=480, W=480.\n - text_input (str or [str]): String or a list of strings, each string is a question.\n The number of questions must be equal to the batch size. If a single string, will be converted to a list of string, with length 1 first.\n num_beams (int): Number of beams for beam search. 1 means no beam search.\n inference_method (str): Inference method. One of \"rank\", \"generate\".\n - If \"rank\", the model will return answers with the highest probability from the answer list.\n - If \"generate\", the model will generate answers.\n max_len (int): Maximum length of generated answers.\n min_len (int): Minimum length of generated answers.\n num_ans_candidates (int): Number of answer candidates, used to filter out answers with low probability.\n answer_list (list): A list of strings, each string is an answer.\n\n Returns:\n List: A list of strings, each string is an answer.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_vqa\", \"vqav2\")\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> question = \"Which city is this photo taken?\"\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> question = txt_processors[\"eval\"](question)\n >>> samples = {\"image\": image, \"text_input\": [question]}\n >>> answers = model.predict_answers(samples)\n >>> answers\n ['singapore']\n >>> answer_list = [\"Singapore\", \"London\", \"Palo Alto\", \"Tokyo\"]\n >>> answers = model.predict_answers(samples, answer_list=answer_list)\n >>> answers\n ['Singapore']\n ```\n \"\"\"\n assert inference_method in [\n \"rank\",\n \"generate\",\n ], \"Inference method must be one of 'rank' or 'generate', got {}.\".format(\n inference_method\n )\n\n if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n if inference_method == \"generate\":\n return self._generate_answers(\n samples, num_beams=num_beams, max_length=max_len, min_length=min_len\n )\n elif inference_method == \"rank\":\n assert answer_list is not None, \"answer_list must be provided for ranking\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self._rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def _generate_answers(self, samples, num_beams=3, max_length=10, min_length=1):\n encoder_out, _ = self.forward_encoder(samples)\n\n question_output = encoder_out\n\n question_states = question_output.last_hidden_state.repeat_interleave(\n num_beams, dim=0\n )\n question_atts = torch.ones(question_states.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": question_states,\n \"encoder_attention_mask\": question_atts,\n }\n\n bsz = samples[\"image\"].size(0)\n bos_ids = torch.full(","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa._generate_answers","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa._generate_answers#L237-L275","kind":"function","name":"_generate_answers","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":237,"end_line":275,"context_start_line":217,"context_end_line":295,"code":" if isinstance(samples[\"text_input\"], str):\n samples[\"text_input\"] = [samples[\"text_input\"]]\n\n assert len(samples[\"text_input\"]) == samples[\"image\"].size(\n 0\n ), \"The number of questions must be equal to the batch size.\"\n\n if inference_method == \"generate\":\n return self._generate_answers(\n samples, num_beams=num_beams, max_length=max_len, min_length=min_len\n )\n elif inference_method == \"rank\":\n assert answer_list is not None, \"answer_list must be provided for ranking\"\n\n num_ans_candidates = min(num_ans_candidates, len(answer_list))\n\n return self._rank_answers(\n samples, answer_list=answer_list, num_ans_candidates=num_ans_candidates\n )\n\n def _generate_answers(self, samples, num_beams=3, max_length=10, min_length=1):\n encoder_out, _ = self.forward_encoder(samples)\n\n question_output = encoder_out\n\n question_states = question_output.last_hidden_state.repeat_interleave(\n num_beams, dim=0\n )\n question_atts = torch.ones(question_states.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n model_kwargs = {\n \"encoder_hidden_states\": question_states,\n \"encoder_attention_mask\": question_atts,\n }\n\n bsz = samples[\"image\"].size(0)\n bos_ids = torch.full(\n (bsz, 1), fill_value=self.tokenizer.bos_token_id, device=self.device\n )\n\n outputs = self.text_decoder.generate(\n input_ids=bos_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n\n # collect answers\n answers = []\n for output in outputs:\n answer = self.tokenizer.decode(output, skip_special_tokens=True)\n answers.append(answer)\n\n return answers\n\n def _rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa._rank_answers","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa._rank_answers#L277-L354","kind":"function","name":"_rank_answers","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":277,"end_line":354,"context_start_line":257,"context_end_line":374,"code":" )\n\n outputs = self.text_decoder.generate(\n input_ids=bos_ids,\n max_length=max_length,\n min_length=min_length,\n num_beams=num_beams,\n eos_token_id=self.tokenizer.sep_token_id,\n pad_token_id=self.tokenizer.pad_token_id,\n **model_kwargs\n )\n\n # collect answers\n answers = []\n for output in outputs:\n answer = self.tokenizer.decode(output, skip_special_tokens=True)\n answers.append(answer)\n\n return answers\n\n def _rank_answers(self, samples, answer_list, num_ans_candidates):\n \"\"\"\n Generate the first token of answers using decoder and select ${num_ans_candidates}\n most probable ones. Then select answers from answer list, which start with the probable tokens.\n Lastly, use the selected answers as the ground-truth labels for decoding and calculating LM loss.\n Return the answers that minimize the losses as result.\n\n \"\"\"\n answer_candidates = self.tokenizer(\n answer_list, padding=\"longest\", return_tensors=\"pt\"\n ).to(self.device)\n answer_candidates.input_ids[:, 0] = self.tokenizer.bos_token_id\n\n answer_ids = answer_candidates.input_ids\n answer_atts = answer_candidates.attention_mask\n\n question_output, _ = self.forward_encoder(samples)\n question_states = question_output.last_hidden_state\n\n tokenized_question = samples[\"tokenized_text\"]\n question_atts = tokenized_question.attention_mask\n\n num_ques = question_states.size(0)\n start_ids = answer_ids[0, 0].repeat(num_ques, 1) # bos token\n\n start_output = self.text_decoder(\n start_ids,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n return_dict=True,\n reduction=\"none\",\n )\n logits = start_output.logits[:, 0, :] # first token's logit\n\n # topk_probs: top-k probability\n # topk_ids: [num_question, k]\n answer_first_token = answer_ids[:, 1]\n prob_first_token = F.softmax(logits, dim=1).index_select(\n dim=1, index=answer_first_token\n )\n topk_probs, topk_ids = prob_first_token.topk(num_ans_candidates, dim=1)\n\n # answer input: [num_question*k, answer_len]\n input_ids = []\n input_atts = []\n for b, topk_id in enumerate(topk_ids):\n input_ids.append(answer_ids.index_select(dim=0, index=topk_id))\n input_atts.append(answer_atts.index_select(dim=0, index=topk_id))\n input_ids = torch.cat(input_ids, dim=0)\n input_atts = torch.cat(input_atts, dim=0)\n\n targets_ids = input_ids.masked_fill(\n input_ids == self.tokenizer.pad_token_id, -100\n )\n\n # repeat encoder's output for top-k answers\n question_states = tile(question_states, 0, num_ans_candidates)\n question_atts = tile(question_atts, 0, num_ans_candidates)\n\n output = self.text_decoder(\n input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_vqa.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_vqa.from_config#L357-L375","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_vqa.py","language":"python","start_line":357,"end_line":375,"context_start_line":337,"context_end_line":375,"code":" input_ids,\n attention_mask=input_atts,\n encoder_hidden_states=question_states,\n encoder_attention_mask=question_atts,\n labels=targets_ids,\n return_dict=True,\n reduction=\"none\",\n )\n\n log_probs_sum = -output.loss\n log_probs_sum = log_probs_sum.view(num_ques, num_ans_candidates)\n\n max_topk_ids = log_probs_sum.argmax(dim=1)\n max_ids = topk_ids[max_topk_ids >= 0, max_topk_ids]\n\n answers = [answer_list[max_id] for max_id in max_ids]\n\n return answers\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n text_encoder = XBertEncoder.from_config(cfg)\n text_decoder = XBertLMHeadDecoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n max_txt_len=max_txt_len,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model","source_hash":"43957021a6f8cf4f9c60ac21cf2d8a2bbc5b907756bdeaf3054b62dc5f9ea0c6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain","uri":"program://CREMA/module/lavis.models.blip_models.blip_pretrain#L1-L394","kind":"module","name":"lavis.models.blip_models.blip_pretrain","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":1,"end_line":394,"context_start_line":1,"context_end_line":394,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.blip_models import tie_encoder_decoder_weights\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipSimilarity,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder, XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_pretrain\")\nclass BlipPretrain(BlipBase, SharedQueueMixin, MomentumDistilationMixin):\n \"\"\"\n BLIP pretrain model.\n\n Supported model types:\n - base: BLIP base model before pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_pretrain_base.yaml\",\n # \"large\": \"configs/models/blip_pretrain_large.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n tie_enc_dec_weights=True,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n text_encoder.resize_token_embeddings(len(self.tokenizer))\n text_decoder.resize_token_embeddings(len(self.tokenizer))\n\n if tie_enc_dec_weights:\n tie_encoder_decoder_weights(\n encoder=text_encoder,\n decoder=text_decoder.bert,\n base_model_prefix=\"\",\n skip_key=\"/attention\",\n )\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_pretrain\", \"base\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_lm'])\n\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'text_embeds', 'image_embeds_m', 'text_embeds_m', 'encoder_output', 'encoder_output_neg', 'itm_logits', 'itm_labels', 'decoder_output', 'decoder_labels'])\n >>> output.intermediate_output.image_embeds.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.image_embeds_m.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds_m.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.itm_logits.shape\n >>> # shape: (batch_size * 3, 2)\n torch.Size([12, 2])\n >>> output.intermediate_output.itm_labels.shape\n >>> # shape: (batch_size * 3,)\n torch.Size([12])\n >>> output.intermediate_output.encoder_output.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.encoder_output_m.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.decoder_output.logits.shape\n >>> # shape: (batch_size, max_txt_len, vocab_size)\n torch.Size([4, 30, 30524])\n >>> output.intermediate_output.decoder_labels.shape\n >>> # shape: (batch_size, max_txt_len)\n torch.Size([4, 30])\n \"\"\"\n\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n # image embeddings and features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n # text embeddings and features\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i[:, :bs], dim=1) + 1e-4\n weights_t2i.fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs], dim=1) + 1e-4\n weights_i2t.fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # LM\n decoder_input_ids = text.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n decoder_targets = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n decoder_output = self.text_decoder(\n decoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n loss_lm = decoder_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n text_embeds=text_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n text_encoder = XBertEncoder.from_config(cfg, from_pretrained=True)\n text_decoder = XBertLMHeadDecoder.from_config(cfg, from_pretrained=True)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 57600)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n embed_dim=embed_dim,\n queue_size=queue_size,\n momentum=momentum,\n alpha=alpha,\n tie_enc_dec_weights=True,\n max_txt_len=max_txt_len,\n )\n\n # [IMPORTANT] to reset queue pointer to 0.\n # Otherwise when updating last batch in the queue, the batch size and remaining queue length may be un-equal.\n model.reset_queue_ptr()\n\n return model","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain.BlipPretrain","uri":"program://CREMA/class/lavis.models.blip_models.blip_pretrain.BlipPretrain#L27-L394","kind":"class","name":"BlipPretrain","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":27,"end_line":394,"context_start_line":7,"context_end_line":394,"code":"\nfrom copy import deepcopy\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import MomentumDistilationMixin, SharedQueueMixin\nfrom lavis.models.blip_models import tie_encoder_decoder_weights\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import (\n BlipOutput,\n BlipSimilarity,\n BlipIntermediateOutput,\n)\nfrom lavis.models.med import XBertEncoder, XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_pretrain\")\nclass BlipPretrain(BlipBase, SharedQueueMixin, MomentumDistilationMixin):\n \"\"\"\n BLIP pretrain model.\n\n Supported model types:\n - base: BLIP base model before pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_pretrain_base.yaml\",\n # \"large\": \"configs/models/blip_pretrain_large.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n tie_enc_dec_weights=True,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n text_encoder.resize_token_embeddings(len(self.tokenizer))\n text_decoder.resize_token_embeddings(len(self.tokenizer))\n\n if tie_enc_dec_weights:\n tie_encoder_decoder_weights(\n encoder=text_encoder,\n decoder=text_decoder.bert,\n base_model_prefix=\"\",\n skip_key=\"/attention\",\n )\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_pretrain\", \"base\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_lm'])\n\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'text_embeds', 'image_embeds_m', 'text_embeds_m', 'encoder_output', 'encoder_output_neg', 'itm_logits', 'itm_labels', 'decoder_output', 'decoder_labels'])\n >>> output.intermediate_output.image_embeds.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.image_embeds_m.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds_m.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.itm_logits.shape\n >>> # shape: (batch_size * 3, 2)\n torch.Size([12, 2])\n >>> output.intermediate_output.itm_labels.shape\n >>> # shape: (batch_size * 3,)\n torch.Size([12])\n >>> output.intermediate_output.encoder_output.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.encoder_output_m.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.decoder_output.logits.shape\n >>> # shape: (batch_size, max_txt_len, vocab_size)\n torch.Size([4, 30, 30524])\n >>> output.intermediate_output.decoder_labels.shape\n >>> # shape: (batch_size, max_txt_len)\n torch.Size([4, 30])\n \"\"\"\n\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n # image embeddings and features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n # text embeddings and features\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i[:, :bs], dim=1) + 1e-4\n weights_t2i.fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs], dim=1) + 1e-4\n weights_i2t.fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # LM\n decoder_input_ids = text.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n decoder_targets = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n decoder_output = self.text_decoder(\n decoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n loss_lm = decoder_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n text_embeds=text_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n text_encoder = XBertEncoder.from_config(cfg, from_pretrained=True)\n text_decoder = XBertLMHeadDecoder.from_config(cfg, from_pretrained=True)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 57600)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n embed_dim=embed_dim,\n queue_size=queue_size,\n momentum=momentum,\n alpha=alpha,\n tie_enc_dec_weights=True,\n max_txt_len=max_txt_len,\n )\n\n # [IMPORTANT] to reset queue pointer to 0.\n # Otherwise when updating last batch in the queue, the batch size and remaining queue length may be un-equal.\n model.reset_queue_ptr()\n\n return model","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_pretrain.__init__#L40-L109","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":40,"end_line":109,"context_start_line":20,"context_end_line":129,"code":")\nfrom lavis.models.med import XBertEncoder, XBertLMHeadDecoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_pretrain\")\nclass BlipPretrain(BlipBase, SharedQueueMixin, MomentumDistilationMixin):\n \"\"\"\n BLIP pretrain model.\n\n Supported model types:\n - base: BLIP base model before pretraining.\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_pretrain_base.yaml\",\n # \"large\": \"configs/models/blip_pretrain_large.yaml\",\n }\n\n def __init__(\n self,\n image_encoder,\n text_encoder,\n text_decoder,\n queue_size,\n alpha=0.4,\n embed_dim=256,\n momentum=0.995,\n tie_enc_dec_weights=True,\n max_txt_len=30,\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n text_encoder.resize_token_embeddings(len(self.tokenizer))\n text_decoder.resize_token_embeddings(len(self.tokenizer))\n\n if tie_enc_dec_weights:\n tie_encoder_decoder_weights(\n encoder=text_encoder,\n decoder=text_decoder.bert,\n base_model_prefix=\"\",\n skip_key=\"/attention\",\n )\n\n self.visual_encoder = image_encoder\n\n self.text_encoder = text_encoder\n self.text_decoder = text_decoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n # create the momentum encoder\n self.visual_encoder_m = deepcopy(self.visual_encoder)\n self.text_encoder_m = deepcopy(self.text_encoder)\n\n self.vision_proj_m = deepcopy(self.vision_proj)\n self.text_proj_m = deepcopy(self.text_proj)\n\n self.model_pairs = [\n [self.visual_encoder, self.visual_encoder_m],\n [self.text_encoder, self.text_encoder_m],\n [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain._rampup_factor","uri":"program://CREMA/function/lavis.models.blip_models.blip_pretrain._rampup_factor#L111-L112","kind":"function","name":"_rampup_factor","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":111,"end_line":112,"context_start_line":91,"context_end_line":132,"code":" [self.vision_proj, self.vision_proj_m],\n [self.text_proj, self.text_proj_m],\n ]\n self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_pretrain\", \"base\")\n >>> images = torch.randn(4, 3, 224, 224)","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_pretrain.forward#L114-L360","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":114,"end_line":360,"context_start_line":94,"context_end_line":380,"code":" self.copy_params()\n\n # create the queue\n self.register_buffer(\"image_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"text_queue\", torch.randn(embed_dim, queue_size))\n self.register_buffer(\"queue_ptr\", torch.zeros(1, dtype=torch.long))\n\n self.image_queue = nn.functional.normalize(self.image_queue, dim=0)\n self.text_queue = nn.functional.normalize(self.text_queue, dim=0)\n\n self.queue_size = queue_size\n self.momentum = momentum\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n self.alpha = alpha\n self.max_txt_len = max_txt_len\n\n def _rampup_factor(self, epoch, iters, num_iters_per_epoch):\n return min(1, (epoch * num_iters_per_epoch + iters) / (2 * num_iters_per_epoch))\n\n def forward(self, samples):\n\n \"\"\"\n Args:\n samples (dict): A dictionary containing the following keys:\n - image (torch.Tensor): A tensor of shape (batch_size, 3, H, W). The input images. Default: H=224, W=224.\n - text_input (list): A list of length batch_size, each element is a string of text/caption.\n - epoch (int): The current epoch.\n - iters (int): The current iteration.\n - num_iters_per_epoch (int): The number of iterations per epoch.\n\n Returns:\n BlipOutput: A BlipOutput object containing loss and intermediate output. See ``lavis.models.blip_models.blip_outputs.BlipOutput`` for more details.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_pretrain\", \"base\")\n >>> images = torch.randn(4, 3, 224, 224)\n >>> text_input = [\"caption of image 1\", \"another caption of image 1\", \"caption of image 2\", \"caption of image 3\"]\n >>> samples = {\"image\": images, \"text_input\": text_input, \"epoch\": 0, \"iters\": 0, \"num_iters_per_epoch\": 100}\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['sims', 'intermediate_output', 'loss', 'loss_itc', 'loss_itm', 'loss_lm'])\n\n >>> output.intermediate_output.keys()\n odict_keys(['image_embeds', 'text_embeds', 'image_embeds_m', 'text_embeds_m', 'encoder_output', 'encoder_output_neg', 'itm_logits', 'itm_labels', 'decoder_output', 'decoder_labels'])\n >>> output.intermediate_output.image_embeds.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.image_embeds_m.shape\n >>> # shape: (batch_size, num_patches, embed_dim)\n torch.Size([4, 197, 768])\n >>> output.intermediate_output.text_embeds_m.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.itm_logits.shape\n >>> # shape: (batch_size * 3, 2)\n torch.Size([12, 2])\n >>> output.intermediate_output.itm_labels.shape\n >>> # shape: (batch_size * 3,)\n torch.Size([12])\n >>> output.intermediate_output.encoder_output.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.encoder_output_m.last_hidden_state.shape\n >>> # shape: (batch_size, max_txt_len, embed_dim)\n torch.Size([4, 30, 768])\n >>> output.intermediate_output.decoder_output.logits.shape\n >>> # shape: (batch_size, max_txt_len, vocab_size)\n torch.Size([4, 30, 30524])\n >>> output.intermediate_output.decoder_labels.shape\n >>> # shape: (batch_size, max_txt_len)\n torch.Size([4, 30])\n \"\"\"\n\n image = samples[\"image\"]\n caption = samples[\"text_input\"]\n\n alpha = self.alpha * self._rampup_factor(\n epoch=samples[\"epoch\"],\n iters=samples[\"iters\"],\n num_iters_per_epoch=samples[\"num_iters_per_epoch\"],\n )\n\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n # image embeddings and features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n image.device\n )\n image_feat = F.normalize(self.vision_proj(image_embeds[:, 0, :]), dim=-1)\n\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(image.device)\n\n # text embeddings and features\n text_output = self.text_encoder.forward_text(text)\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # get momentum features\n with torch.no_grad():\n self._momentum_update()\n image_embeds_m = self.visual_encoder_m(image)\n image_feat_m = F.normalize(\n self.vision_proj_m(image_embeds_m[:, 0, :]), dim=-1\n )\n image_feat_all = torch.cat(\n [image_feat_m.t(), self.image_queue.clone().detach()], dim=1\n )\n\n text_output_m = self.text_encoder_m.forward_text(text)\n text_embeds_m = text_output_m.last_hidden_state\n text_feat_m = F.normalize(self.text_proj_m(text_embeds_m[:, 0, :]), dim=-1)\n text_feat_all = torch.cat(\n [text_feat_m.t(), self.text_queue.clone().detach()], dim=1\n )\n\n sim_i2t_m = image_feat_m @ text_feat_all / self.temp\n sim_t2i_m = text_feat_m @ image_feat_all / self.temp\n\n sim_targets = torch.zeros(sim_i2t_m.size()).to(image.device)\n sim_targets.fill_diagonal_(1)\n\n sim_i2t_targets = (\n alpha * F.softmax(sim_i2t_m, dim=1) + (1 - alpha) * sim_targets\n )\n sim_t2i_targets = (\n alpha * F.softmax(sim_t2i_m, dim=1) + (1 - alpha) * sim_targets\n )\n\n sim_i2t = image_feat @ text_feat_all / self.temp\n sim_t2i = text_feat @ image_feat_all / self.temp\n\n loss_i2t = -torch.sum(\n F.log_softmax(sim_i2t, dim=1) * sim_i2t_targets, dim=1\n ).mean()\n loss_t2i = -torch.sum(\n F.log_softmax(sim_t2i, dim=1) * sim_t2i_targets, dim=1\n ).mean()\n\n loss_itc = (loss_i2t + loss_t2i) / 2\n\n self._dequeue_and_enqueue(image_feat_m, text_feat_m)\n\n # Image-text Matching\n encoder_input_ids = text.input_ids.clone()\n encoder_input_ids[:, 0] = self.tokenizer.enc_token_id\n\n # forward the positve image-text pair\n bs = image.size(0)\n output_pos = self.text_encoder(\n encoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n with torch.no_grad():\n weights_t2i = F.softmax(sim_t2i[:, :bs], dim=1) + 1e-4\n weights_t2i.fill_diagonal_(0)\n weights_i2t = F.softmax(sim_i2t[:, :bs], dim=1) + 1e-4\n weights_i2t.fill_diagonal_(0)\n\n # select a negative image for each text\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2i[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_ids_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_i2t[b], 1).item()\n text_ids_neg.append(encoder_input_ids[neg_idx])\n text_atts_neg.append(text.attention_mask[neg_idx])\n\n text_ids_neg = torch.stack(text_ids_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_ids_all = torch.cat([encoder_input_ids, text_ids_neg], dim=0)\n text_atts_all = torch.cat([text.attention_mask, text_atts_neg], dim=0)\n\n image_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n image_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n output_neg = self.text_encoder(\n text_ids_all,\n attention_mask=text_atts_all,\n encoder_hidden_states=image_embeds_all,\n encoder_attention_mask=image_atts_all,\n return_dict=True,\n )\n\n vl_embeddings = torch.cat(\n [\n output_pos.last_hidden_state[:, 0, :],\n output_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n itm_logits = self.itm_head(vl_embeddings)\n\n itm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(image.device)\n loss_itm = F.cross_entropy(itm_logits, itm_labels)\n\n # LM\n decoder_input_ids = text.input_ids.clone()\n decoder_input_ids[:, 0] = self.tokenizer.bos_token_id\n decoder_targets = decoder_input_ids.masked_fill(\n decoder_input_ids == self.tokenizer.pad_token_id, -100\n )\n\n decoder_output = self.text_decoder(\n decoder_input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n labels=decoder_targets,\n return_dict=True,\n )\n\n loss_lm = decoder_output.loss\n\n return BlipOutput(\n loss=loss_itc + loss_itm + loss_lm,\n loss_itc=loss_itc,\n loss_itm=loss_itm,\n loss_lm=loss_lm,\n sims=BlipSimilarity(\n sim_i2t=sim_i2t,\n sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n text_embeds=text_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n text_encoder = XBertEncoder.from_config(cfg, from_pretrained=True)\n text_decoder = XBertLMHeadDecoder.from_config(cfg, from_pretrained=True)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 57600)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain.reset_queue_ptr","uri":"program://CREMA/function/lavis.models.blip_models.blip_pretrain.reset_queue_ptr#L362-L363","kind":"function","name":"reset_queue_ptr","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":362,"end_line":363,"context_start_line":342,"context_end_line":383,"code":" sim_t2i=sim_t2i,\n sim_i2t_m=sim_i2t_m,\n sim_t2i_m=sim_t2i_m,\n sim_i2t_targets=sim_i2t_targets,\n sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n text_embeds=text_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n text_encoder = XBertEncoder.from_config(cfg, from_pretrained=True)\n text_decoder = XBertLMHeadDecoder.from_config(cfg, from_pretrained=True)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 57600)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n embed_dim=embed_dim,\n queue_size=queue_size,","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_pretrain.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_pretrain.from_config#L366-L394","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_pretrain.py","language":"python","start_line":366,"end_line":394,"context_start_line":346,"context_end_line":394,"code":" sim_t2i_targets=sim_t2i_targets,\n ),\n intermediate_output=BlipIntermediateOutput(\n image_embeds=image_embeds,\n text_embeds=text_embeds,\n image_embeds_m=image_embeds_m,\n text_embeds_m=text_embeds_m,\n encoder_output=output_pos,\n encoder_output_neg=output_neg,\n itm_logits=itm_logits,\n itm_labels=itm_labels,\n decoder_output=decoder_output,\n decoder_labels=decoder_targets,\n ),\n )\n\n def reset_queue_ptr(self):\n self.queue_ptr = torch.zeros(1, dtype=torch.long)\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg, from_pretrained=True)\n text_encoder = XBertEncoder.from_config(cfg, from_pretrained=True)\n text_decoder = XBertLMHeadDecoder.from_config(cfg, from_pretrained=True)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n momentum = cfg.get(\"momentum\", 0.995)\n alpha = cfg.get(\"alpha\", 0.4)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n queue_size = cfg.get(\"queue_size\", 57600)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n text_decoder=text_decoder,\n embed_dim=embed_dim,\n queue_size=queue_size,\n momentum=momentum,\n alpha=alpha,\n tie_enc_dec_weights=True,\n max_txt_len=max_txt_len,\n )\n\n # [IMPORTANT] to reset queue pointer to 0.\n # Otherwise when updating last batch in the queue, the batch size and remaining queue length may be un-equal.\n model.reset_queue_ptr()\n\n return model","source_hash":"60aafbc1535a06ec8027dc209021ec71037d0a1c4700f63d78cbccd81d9b1a16","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs","uri":"program://CREMA/module/lavis.models.blip_models.blip_outputs#L1-L116","kind":"module","name":"lavis.models.blip_models.blip_outputs","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":1,"end_line":116,"context_start_line":1,"context_end_line":116,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n ModelOutput,\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n)\n\n\n@dataclass\nclass BlipSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipIntermediateOutput(ModelOutput):\n \"\"\"\n Data class for intermediate outputs of BLIP models.\n\n image_embeds (torch.FloatTensor): Image embeddings, shape (batch_size, num_patches, embed_dim).\n text_embeds (torch.FloatTensor): Text embeddings, shape (batch_size, seq_len, embed_dim).\n\n image_embeds_m (torch.FloatTensor): Image embeddings from momentum visual encoder, shape (batch_size, num_patches, embed_dim).\n text_embeds_m (torch.FloatTensor): Text embeddings from momentum text encoder, shape (batch_size, seq_len, embed_dim).\n\n encoder_output (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder.\n encoder_output_neg (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder for negative pairs.\n\n decoder_output (CausalLMOutputWithCrossAttentions): output from the image-grounded text decoder.\n decoder_labels (torch.LongTensor): labels for the captioning loss.\n\n itm_logits (torch.FloatTensor): logits for the image-text matching loss, shape (batch_size * 3, 2).\n itm_labels (torch.LongTensor): labels for the image-text matching loss, shape (batch_size * 3,)\n\n \"\"\"\n\n # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass BlipOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[BlipSimilarity] = None\n\n intermediate_output: BlipIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_lm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipOutputWithLogits(BlipOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass BlipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from BlipFeatureExtractor.\n\n Args:\n image_embeds: (torch.FloatTensor) of shape (batch_size, num_patches+1, embed_dim), optional\n image_features: (torch.FloatTensor) of shape (batch_size, num_patches+1, feature_dim), optional\n text_embeds: (torch.FloatTensor) of shape (batch_size, sequence_length+1, embed_dim), optional\n text_features: (torch.FloatTensor) of shape (batch_size, sequence_length+1, feature_dim), optional\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n multimodal_embeds: Optional[torch.FloatTensor] = None","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs.BlipSimilarity","uri":"program://CREMA/class/lavis.models.blip_models.blip_outputs.BlipSimilarity#L20-L28","kind":"class","name":"BlipSimilarity","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":20,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n ModelOutput,\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n)\n\n\n@dataclass\nclass BlipSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipIntermediateOutput(ModelOutput):\n \"\"\"\n Data class for intermediate outputs of BLIP models.\n\n image_embeds (torch.FloatTensor): Image embeddings, shape (batch_size, num_patches, embed_dim).\n text_embeds (torch.FloatTensor): Text embeddings, shape (batch_size, seq_len, embed_dim).\n\n image_embeds_m (torch.FloatTensor): Image embeddings from momentum visual encoder, shape (batch_size, num_patches, embed_dim).\n text_embeds_m (torch.FloatTensor): Text embeddings from momentum text encoder, shape (batch_size, seq_len, embed_dim).\n\n encoder_output (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder.\n encoder_output_neg (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder for negative pairs.\n\n decoder_output (CausalLMOutputWithCrossAttentions): output from the image-grounded text decoder.\n decoder_labels (torch.LongTensor): labels for the captioning loss.\n\n itm_logits (torch.FloatTensor): logits for the image-text matching loss, shape (batch_size * 3, 2).","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs.BlipIntermediateOutput","uri":"program://CREMA/class/lavis.models.blip_models.blip_outputs.BlipIntermediateOutput#L32-L69","kind":"class","name":"BlipIntermediateOutput","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":32,"end_line":69,"context_start_line":12,"context_end_line":89,"code":"from transformers.modeling_outputs import (\n ModelOutput,\n BaseModelOutputWithPoolingAndCrossAttentions,\n CausalLMOutputWithCrossAttentions,\n)\n\n\n@dataclass\nclass BlipSimilarity(ModelOutput):\n sim_i2t: torch.FloatTensor = None\n sim_t2i: torch.FloatTensor = None\n\n sim_i2t_m: Optional[torch.FloatTensor] = None\n sim_t2i_m: Optional[torch.FloatTensor] = None\n\n sim_i2t_targets: Optional[torch.FloatTensor] = None\n sim_t2i_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipIntermediateOutput(ModelOutput):\n \"\"\"\n Data class for intermediate outputs of BLIP models.\n\n image_embeds (torch.FloatTensor): Image embeddings, shape (batch_size, num_patches, embed_dim).\n text_embeds (torch.FloatTensor): Text embeddings, shape (batch_size, seq_len, embed_dim).\n\n image_embeds_m (torch.FloatTensor): Image embeddings from momentum visual encoder, shape (batch_size, num_patches, embed_dim).\n text_embeds_m (torch.FloatTensor): Text embeddings from momentum text encoder, shape (batch_size, seq_len, embed_dim).\n\n encoder_output (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder.\n encoder_output_neg (BaseModelOutputWithPoolingAndCrossAttentions): output from the image-grounded text encoder for negative pairs.\n\n decoder_output (CausalLMOutputWithCrossAttentions): output from the image-grounded text decoder.\n decoder_labels (torch.LongTensor): labels for the captioning loss.\n\n itm_logits (torch.FloatTensor): logits for the image-text matching loss, shape (batch_size * 3, 2).\n itm_labels (torch.LongTensor): labels for the image-text matching loss, shape (batch_size * 3,)\n\n \"\"\"\n\n # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass BlipOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[BlipSimilarity] = None\n\n intermediate_output: BlipIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_lm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipOutputWithLogits(BlipOutput):","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs.BlipOutput","uri":"program://CREMA/class/lavis.models.blip_models.blip_outputs.BlipOutput#L73-L85","kind":"class","name":"BlipOutput","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":73,"end_line":85,"context_start_line":53,"context_end_line":105,"code":" # uni-modal features\n image_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n image_embeds_m: Optional[torch.FloatTensor] = None\n text_embeds_m: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n itm_logits: Optional[torch.FloatTensor] = None\n itm_labels: Optional[torch.LongTensor] = None\n\n # intermediate outputs of multimodal decoder\n decoder_output: Optional[CausalLMOutputWithCrossAttentions] = None\n decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass BlipOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[BlipSimilarity] = None\n\n intermediate_output: BlipIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_lm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipOutputWithLogits(BlipOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass BlipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from BlipFeatureExtractor.\n\n Args:\n image_embeds: (torch.FloatTensor) of shape (batch_size, num_patches+1, embed_dim), optional\n image_features: (torch.FloatTensor) of shape (batch_size, num_patches+1, feature_dim), optional\n text_embeds: (torch.FloatTensor) of shape (batch_size, sequence_length+1, embed_dim), optional\n text_features: (torch.FloatTensor) of shape (batch_size, sequence_length+1, feature_dim), optional\n\n The first embedding or feature is for the [CLS] token.","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs.BlipOutputWithLogits","uri":"program://CREMA/class/lavis.models.blip_models.blip_outputs.BlipOutputWithLogits#L89-L91","kind":"class","name":"BlipOutputWithLogits","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":89,"end_line":91,"context_start_line":69,"context_end_line":111,"code":" decoder_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass BlipOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[BlipSimilarity] = None\n\n intermediate_output: BlipIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_lm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipOutputWithLogits(BlipOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass BlipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from BlipFeatureExtractor.\n\n Args:\n image_embeds: (torch.FloatTensor) of shape (batch_size, num_patches+1, embed_dim), optional\n image_features: (torch.FloatTensor) of shape (batch_size, num_patches+1, feature_dim), optional\n text_embeds: (torch.FloatTensor) of shape (batch_size, sequence_length+1, embed_dim), optional\n text_features: (torch.FloatTensor) of shape (batch_size, sequence_length+1, feature_dim), optional\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_outputs.BlipOutputFeatures","uri":"program://CREMA/class/lavis.models.blip_models.blip_outputs.BlipOutputFeatures#L95-L116","kind":"class","name":"BlipOutputFeatures","path":"lavis/models/blip_models/blip_outputs.py","language":"python","start_line":95,"end_line":116,"context_start_line":75,"context_end_line":116,"code":" sims: Optional[BlipSimilarity] = None\n\n intermediate_output: BlipIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_itc: Optional[torch.FloatTensor] = None\n\n loss_itm: Optional[torch.FloatTensor] = None\n\n loss_lm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass BlipOutputWithLogits(BlipOutput):\n logits: torch.FloatTensor = None\n logits_m: torch.FloatTensor = None\n\n\n@dataclass\nclass BlipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from BlipFeatureExtractor.\n\n Args:\n image_embeds: (torch.FloatTensor) of shape (batch_size, num_patches+1, embed_dim), optional\n image_features: (torch.FloatTensor) of shape (batch_size, num_patches+1, feature_dim), optional\n text_embeds: (torch.FloatTensor) of shape (batch_size, sequence_length+1, embed_dim), optional\n text_features: (torch.FloatTensor) of shape (batch_size, sequence_length+1, feature_dim), optional\n\n The first embedding or feature is for the [CLS] token.\n\n Features are obtained by projecting the corresponding embedding into a normalized low-dimensional space.\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n multimodal_embeds: Optional[torch.FloatTensor] = None","source_hash":"d655e46f9c61dd0d4c609b066a24e0df203053250c0c1954af69093bca750602","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_feature_extractor","uri":"program://CREMA/module/lavis.models.blip_models.blip_feature_extractor#L1-L212","kind":"module","name":"lavis.models.blip_models.blip_feature_extractor","path":"lavis/models/blip_models/blip_feature_extractor.py","language":"python","start_line":1,"end_line":212,"context_start_line":1,"context_end_line":212,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import BlipOutputFeatures\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_feature_extractor\")\nclass BlipFeatureExtractor(BlipBase):\n \"\"\"\n Class for BLIP feature extractor.\n\n Supported model types:\n - base: BLIP base model with pre-trained weights from capfilt by BLIP large model.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_feature_extractor\", \"base\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_feature_extractor_base.yaml\",\n # \"large\": \"configs/models/blip_feature_extractor_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return image features\n image_embeds = self.visual_encoder.forward_features(image)\n\n image_features = self.vision_proj(image_embeds)\n image_features = F.normalize(image_features, dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n # return text features\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"146f2ef7e0c7eee0fa2c647e136952e5360f8063a9f93f1fec6eb8d6e05fdbb0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_feature_extractor.BlipFeatureExtractor","uri":"program://CREMA/class/lavis.models.blip_models.blip_feature_extractor.BlipFeatureExtractor#L21-L212","kind":"class","name":"BlipFeatureExtractor","path":"lavis/models/blip_models/blip_feature_extractor.py","language":"python","start_line":21,"end_line":212,"context_start_line":1,"context_end_line":212,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import BlipOutputFeatures\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.vit import VisionTransformerEncoder\nfrom torch import nn\n\n\n@registry.register_model(\"blip_feature_extractor\")\nclass BlipFeatureExtractor(BlipBase):\n \"\"\"\n Class for BLIP feature extractor.\n\n Supported model types:\n - base: BLIP base model with pre-trained weights from capfilt by BLIP large model.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_feature_extractor\", \"base\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_feature_extractor_base.yaml\",\n # \"large\": \"configs/models/blip_feature_extractor_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return image features\n image_embeds = self.visual_encoder.forward_features(image)\n\n image_features = self.vision_proj(image_embeds)\n image_features = F.normalize(image_features, dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n # return text features\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"146f2ef7e0c7eee0fa2c647e136952e5360f8063a9f93f1fec6eb8d6e05fdbb0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_feature_extractor.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_feature_extractor.__init__#L38-L55","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_feature_extractor.py","language":"python","start_line":38,"end_line":55,"context_start_line":18,"context_end_line":75,"code":"\n\n@registry.register_model(\"blip_feature_extractor\")\nclass BlipFeatureExtractor(BlipBase):\n \"\"\"\n Class for BLIP feature extractor.\n\n Supported model types:\n - base: BLIP base model with pre-trained weights from capfilt by BLIP large model.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_feature_extractor\", \"base\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"base\": \"configs/models/blip_feature_extractor_base.yaml\",\n # \"large\": \"configs/models/blip_feature_extractor_large.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, embed_dim, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.","source_hash":"146f2ef7e0c7eee0fa2c647e136952e5360f8063a9f93f1fec6eb8d6e05fdbb0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_feature_extractor.extract_features","uri":"program://CREMA/function/lavis.models.blip_models.blip_feature_extractor.extract_features#L58-L187","kind":"function","name":"extract_features","path":"lavis/models/blip_models/blip_feature_extractor.py","language":"python","start_line":58,"end_line":187,"context_start_line":38,"context_end_line":207,"code":" def __init__(self, image_encoder, text_encoder, embed_dim, max_txt_len=40):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n # creating projection layers for ITC\n text_width = text_encoder.config.hidden_size\n vision_width = image_encoder.vision_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.max_txt_len = max_txt_len\n\n self.temp = nn.Parameter(0.07 * torch.ones([]))\n\n @torch.no_grad()\n def extract_features(self, samples, mode=\"multimodal\"):\n \"\"\"\n Extract features for multimodal or unimodal samples.\n\n Args:\n samples (dict): A dictionary of samples, containing the following keys:\n - image (torch.Tensor): A tensor of shape (B, C, H, W) containing the image.\n Raw images should be preprocessed before being passed to feature extractor.\n - text_input (list): A list of strings containing the text, length B.\n mode (str): The mode of feature extraction. Can be either \"multimodal\", \"text\" or \"image\".\n If \"multimodal\", return image features and multimodal features;\n if \"text\", return text features;\n if \"image\", return image features.\n Default: \"multimodal\".\n\n Returns:\n BlipOutputFeatures: A BlipOutputFeatures object containing the features.\n See lavis/models/blip_models/blip_outputs.py for more details.\n\n Examples:\n ```python\n >>> from PIL import Image\n >>> from lavis.models import load_model_and_preprocess\n >>> raw_image = Image.open(\"docs/data/merlion.png\").convert(\"RGB\")\n >>> caption = \"a large fountain spewing water into the air\"\n >>> model, vis_processors, txt_processors = load_model_and_preprocess(\"blip_feature_extractor\", is_eval=True)\n >>> image = vis_processors[\"eval\"](raw_image).unsqueeze(0)\n >>> text_input = txt_processors[\"eval\"](caption)\n\n >>> sample = {\"image\": image, \"text_input\": [text_input]}\n\n >>> features_multimodal = model.extract_features(sample)\n >>> features_multimodal.keys()\n odict_keys(['image_embeds', 'multimodal_embeds'])\n >>> features_multimodal.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_multimodal.multimodal_embeds.shape\n torch.Size([1, 12, 768])\n\n >>> features_text = model.extract_features(sample, mode=\"text\")\n >>> features_text.keys()\n odict_keys(['text_embeds', 'text_features'])\n >>> features_text.text_embeds.shape\n torch.Size([1, 12, 768])\n >>> features_text.text_features.shape\n torch.Size([1, 12, 256])\n\n >>> features_image = model.extract_features(sample, mode=\"image\")\n >>> features_image.keys()\n odict_keys(['image_embeds', 'image_features'])\n >>> features_image.image_embeds.shape\n torch.Size([1, 197, 768])\n >>> features_image.image_features.shape\n torch.Size([1, 197, 256])\n ```\n \"\"\"\n image = samples.get(\"image\")\n caption = samples.get(\"text_input\")\n\n # assert mode is one of \"image\", \"text\", \"multimodal\"\n assert mode in [\n \"image\",\n \"text\",\n \"multimodal\",\n ], \"mode must be one of 'image', 'text', 'multimodal'\"\n\n # initalize output\n image_embeds, text_embeds, multimodal_embeds = None, None, None\n image_features, text_features = None, None\n\n if mode == \"image\":\n assert (\n image is not None\n ), \"Image is not provided for mode 'image' or 'multimodal'\"\n # return image features\n image_embeds = self.visual_encoder.forward_features(image)\n\n image_features = self.vision_proj(image_embeds)\n image_features = F.normalize(image_features, dim=-1)\n\n elif mode == \"text\":\n assert (\n caption is not None\n ), \"text input is None for mode 'text' or 'multimodal'\"\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n\n # return text features\n text_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n return_dict=True,\n mode=\"text\",\n )\n text_embeds = text_output.last_hidden_state\n\n text_features = self.text_proj(text_embeds)\n text_features = F.normalize(text_features, dim=-1)\n\n elif mode == \"multimodal\":\n # return multimodel features\n image_embeds = self.visual_encoder.forward_features(image)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n text = self.tokenizer(caption, return_tensors=\"pt\", padding=True).to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:","source_hash":"146f2ef7e0c7eee0fa2c647e136952e5360f8063a9f93f1fec6eb8d6e05fdbb0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_feature_extractor.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_feature_extractor.from_config#L190-L212","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_feature_extractor.py","language":"python","start_line":190,"end_line":212,"context_start_line":170,"context_end_line":212,"code":"\n output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=image_embeds,\n encoder_attention_mask=image_atts,\n return_dict=True,\n )\n\n multimodal_embeds = output.last_hidden_state\n\n return BlipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n multimodal_embeds=multimodal_embeds,\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n # set from_pretrained=True to load weights for 'bert-base-uncased'\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n text_encoder = XBertEncoder.from_config(cfg)\n\n embed_dim = cfg.get(\"embed_dim\", 256)\n max_txt_len = cfg.get(\"max_txt_len\", 30)\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n embed_dim=embed_dim,\n max_txt_len=max_txt_len,\n )\n\n # load pre-trained weights\n pretrain_path = cfg.get(\"pretrained\", None)\n if pretrain_path is not None:\n msg = model.load_from_pretrained(url_or_filename=pretrain_path)\n else:\n warnings.warn(\"No pretrained weights are loaded.\")\n\n return model","source_hash":"146f2ef7e0c7eee0fa2c647e136952e5360f8063a9f93f1fec6eb8d6e05fdbb0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr","uri":"program://CREMA/module/lavis.models.blip_models.blip_nlvr#L1-L187","kind":"module","name":"lavis.models.blip_models.blip_nlvr","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":1,"end_line":187,"context_start_line":1,"context_end_line":187,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path, is_url\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import BlipIntermediateOutput, BlipOutput\nfrom lavis.models.blip_models.nlvr_encoder import BertModel\nfrom lavis.models.vit import VisionTransformerEncoder, interpolate_pos_embed\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"blip_nlvr\")\nclass BlipNLVR(BlipBase, MomentumDistilationMixin):\n \"\"\"\n Class for BLIP NLVR model.\n\n Supported model types:\n - base: model with pre-trained BLIP weights, used as initialization for fine-tuning.\n - nlvr: finetuned model on NLVR2 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/blip_nlvr.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, num_classes):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(text, padding=\"longest\", return_tensors=\"pt\").to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n encoder_output=encoder_output,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n num_classes=num_classes,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n\n for key in list(state_dict.keys()):\n if \"crossattention.self.\" in key:\n new_key0 = key.replace(\"self\", \"self0\")\n new_key1 = key.replace(\"self\", \"self1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n elif \"crossattention.output.dense.\" in key:\n new_key0 = key.replace(\"dense\", \"dense0\")\n new_key1 = key.replace(\"dense\", \"dense1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n print(\"load checkpoint from %s\" % url_or_filename)\n print(f\"missing keys {msg.missing_keys}\")\n return msg","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.BlipNLVR","uri":"program://CREMA/class/lavis.models.blip_models.blip_nlvr.BlipNLVR#L25-L187","kind":"class","name":"BlipNLVR","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":25,"end_line":187,"context_start_line":5,"context_end_line":187,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path, is_url\nfrom lavis.models.base_model import MomentumDistilationMixin\nfrom lavis.models.blip_models.blip import BlipBase\nfrom lavis.models.blip_models.blip_outputs import BlipIntermediateOutput, BlipOutput\nfrom lavis.models.blip_models.nlvr_encoder import BertModel\nfrom lavis.models.vit import VisionTransformerEncoder, interpolate_pos_embed\nfrom torch import nn\nfrom transformers import BertConfig\n\n\n@registry.register_model(\"blip_nlvr\")\nclass BlipNLVR(BlipBase, MomentumDistilationMixin):\n \"\"\"\n Class for BLIP NLVR model.\n\n Supported model types:\n - base: model with pre-trained BLIP weights, used as initialization for fine-tuning.\n - nlvr: finetuned model on NLVR2 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/blip_nlvr.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, num_classes):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(text, padding=\"longest\", return_tensors=\"pt\").to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n encoder_output=encoder_output,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n num_classes=num_classes,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n\n for key in list(state_dict.keys()):\n if \"crossattention.self.\" in key:\n new_key0 = key.replace(\"self\", \"self0\")\n new_key1 = key.replace(\"self\", \"self1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n elif \"crossattention.output.dense.\" in key:\n new_key0 = key.replace(\"dense\", \"dense0\")\n new_key1 = key.replace(\"dense\", \"dense1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n print(\"load checkpoint from %s\" % url_or_filename)\n print(f\"missing keys {msg.missing_keys}\")\n return msg","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.__init__","uri":"program://CREMA/function/lavis.models.blip_models.blip_nlvr.__init__#L42-L54","kind":"function","name":"__init__","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":42,"end_line":54,"context_start_line":22,"context_end_line":74,"code":"\n\n@registry.register_model(\"blip_nlvr\")\nclass BlipNLVR(BlipBase, MomentumDistilationMixin):\n \"\"\"\n Class for BLIP NLVR model.\n\n Supported model types:\n - base: model with pre-trained BLIP weights, used as initialization for fine-tuning.\n - nlvr: finetuned model on NLVR2 dataset.\n\n Usage:\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/blip_nlvr.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, num_classes):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n >>> samples = {","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.forward","uri":"program://CREMA/function/lavis.models.blip_models.blip_nlvr.forward#L56-L126","kind":"function","name":"forward","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":56,"end_line":126,"context_start_line":36,"context_end_line":146,"code":" \"\"\"\n\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"nlvr\": \"configs/models/blip_nlvr.yaml\",\n }\n\n def __init__(self, image_encoder, text_encoder, num_classes):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n self.visual_encoder = image_encoder\n self.text_encoder = text_encoder\n\n hidden_size = text_encoder.config.hidden_size\n self.cls_head = nn.Sequential(\n nn.Linear(hidden_size, hidden_size),\n nn.ReLU(),\n nn.Linear(hidden_size, num_classes),\n )\n\n def forward(self, samples, is_train=True):\n \"\"\"\n Forward function for training and evaluation.\n\n Args:\n samples (dict): a dict of input samples, which contains the following keys:\n - image0 (torch.Tensor): input image 0, shape (batch_size, 3, H, W), default H=384, W=384.\n - image1 (torch.Tensor): input image 1, shape (batch_size, 3, H, W), default H=384, W=384.\n - text_input (list): list of strings, each string is a natural language sentence.\n - label (torch.LongTensor): ground truth label with shape (batch_size,).\n is_train (bool): whether the model is in training mode.\n If True, the model will return the loss;\n If False, the model will return the prediction.\n\n Examples:\n >>> import torch\n >>> from lavis.models import load_model\n >>> model = load_model(\"blip_nlvr\", \"nlvr\")\n >>> samples = {\n ... \"image0\": torch.randn(2, 3, 384, 384),\n ... \"image1\": torch.randn(2, 3, 384, 384),\n ... \"text_input\": [\"there is a ferret in tall grass\", \"there are lips in one of the images\"],\n ... \"label\": torch.tensor([0, 1]),\n ... }\n >>> output = model(samples)\n >>> output.keys()\n odict_keys(['intermediate_output', 'loss'])\n \"\"\"\n text = samples[\"text_input\"]\n text = self.tokenizer(text, padding=\"longest\", return_tensors=\"pt\").to(\n self.device\n )\n text.input_ids[:, 0] = self.tokenizer.enc_token_id\n\n targets = samples[\"label\"]\n\n image0 = samples[\"image0\"]\n image1 = samples[\"image1\"]\n images = torch.cat([image0, image1], dim=0)\n\n image_embeds = self.visual_encoder.forward_features(images)\n image_atts = torch.ones(image_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n image0_embeds, image1_embeds = torch.split(image_embeds, targets.size(0))\n\n encoder_output = self.text_encoder(\n text.input_ids,\n attention_mask=text.attention_mask,\n encoder_hidden_states=[image0_embeds, image1_embeds],\n encoder_attention_mask=[\n image_atts[: image0_embeds.size(0)],\n image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n encoder_output=encoder_output,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.predict","uri":"program://CREMA/function/lavis.models.blip_models.blip_nlvr.predict#L128-L130","kind":"function","name":"predict","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":128,"end_line":130,"context_start_line":108,"context_end_line":150,"code":" image_atts[image0_embeds.size(0) :],\n ],\n return_dict=True,\n )\n\n prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n encoder_output=encoder_output,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n num_classes=num_classes,\n )","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.from_config","uri":"program://CREMA/function/lavis.models.blip_models.blip_nlvr.from_config#L133-L154","kind":"function","name":"from_config","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":133,"end_line":154,"context_start_line":113,"context_end_line":174,"code":" prediction = self.cls_head(encoder_output.last_hidden_state[:, 0, :])\n\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return BlipOutput(\n loss=loss,\n intermediate_output=BlipIntermediateOutput(\n image_embeds=torch.stack([image0_embeds, image1_embeds], dim=0),\n encoder_output=encoder_output,\n ),\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg=None):\n image_encoder = VisionTransformerEncoder.from_config(cfg)\n\n # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n num_classes=num_classes,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n\n for key in list(state_dict.keys()):\n if \"crossattention.self.\" in key:\n new_key0 = key.replace(\"self\", \"self0\")","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip_nlvr.load_from_pretrained","uri":"program://CREMA/function/lavis.models.blip_models.blip_nlvr.load_from_pretrained#L156-L187","kind":"function","name":"load_from_pretrained","path":"lavis/models/blip_models/blip_nlvr.py","language":"python","start_line":156,"end_line":187,"context_start_line":136,"context_end_line":187,"code":" # text encoder + multimodal encoder\n bert_config = BertConfig.from_json_file(get_abs_path(cfg[\"med_config_path\"]))\n text_encoder = BertModel(config=bert_config, add_pooling_layer=False)\n\n num_classes = cfg.get(\"num_classes\", 3)\n\n assert num_classes > 1, \"Invalid number of classes provided, found {}\".format(\n num_classes\n )\n\n model = cls(\n image_encoder=image_encoder,\n text_encoder=text_encoder,\n num_classes=num_classes,\n )\n\n model.load_checkpoint_from_config(cfg)\n\n return model\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n\n for key in list(state_dict.keys()):\n if \"crossattention.self.\" in key:\n new_key0 = key.replace(\"self\", \"self0\")\n new_key1 = key.replace(\"self\", \"self1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n elif \"crossattention.output.dense.\" in key:\n new_key0 = key.replace(\"dense\", \"dense0\")\n new_key1 = key.replace(\"dense\", \"dense1\")\n state_dict[new_key0] = state_dict[key]\n state_dict[new_key1] = state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n print(\"load checkpoint from %s\" % url_or_filename)\n print(f\"missing keys {msg.missing_keys}\")\n return msg","source_hash":"180f06feb27958bd49c29dff1d8cd565b5552fe1509416cac3cde56fdc1fcfc1","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip","uri":"program://CREMA/module/lavis.models.blip_models.blip#L1-L59","kind":"module","name":"lavis.models.blip_models.blip","path":"lavis/models/blip_models/blip.py","language":"python","start_line":1,"end_line":59,"context_start_line":1,"context_end_line":59,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.vit import interpolate_pos_embed\nfrom transformers import BertTokenizer\n\n\nclass BlipBase(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n tokenizer.add_special_tokens({\"additional_special_tokens\": [\"[ENC]\"]})\n tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]\n return tokenizer\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n if \"visual_encoder_m.pos_embed\" in self.state_dict().keys():\n state_dict[\"visual_encoder_m.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg","source_hash":"baa9da1d24ec8956a509dda74ac6f8db9bff9511dbedb515cf6976f2dd0289d6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip.BlipBase","uri":"program://CREMA/class/lavis.models.blip_models.blip.BlipBase#L19-L59","kind":"class","name":"BlipBase","path":"lavis/models/blip_models/blip.py","language":"python","start_line":19,"end_line":59,"context_start_line":1,"context_end_line":59,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.vit import interpolate_pos_embed\nfrom transformers import BertTokenizer\n\n\nclass BlipBase(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n tokenizer.add_special_tokens({\"additional_special_tokens\": [\"[ENC]\"]})\n tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]\n return tokenizer\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n if \"visual_encoder_m.pos_embed\" in self.state_dict().keys():\n state_dict[\"visual_encoder_m.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg","source_hash":"baa9da1d24ec8956a509dda74ac6f8db9bff9511dbedb515cf6976f2dd0289d6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip.init_tokenizer","uri":"program://CREMA/function/lavis.models.blip_models.blip.init_tokenizer#L21-L26","kind":"function","name":"init_tokenizer","path":"lavis/models/blip_models/blip.py","language":"python","start_line":21,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.vit import interpolate_pos_embed\nfrom transformers import BertTokenizer\n\n\nclass BlipBase(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n tokenizer.add_special_tokens({\"additional_special_tokens\": [\"[ENC]\"]})\n tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]\n return tokenizer\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n if \"visual_encoder_m.pos_embed\" in self.state_dict().keys():\n state_dict[\"visual_encoder_m.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m","source_hash":"baa9da1d24ec8956a509dda74ac6f8db9bff9511dbedb515cf6976f2dd0289d6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.blip_models.blip.load_from_pretrained","uri":"program://CREMA/function/lavis.models.blip_models.blip.load_from_pretrained#L28-L59","kind":"function","name":"load_from_pretrained","path":"lavis/models/blip_models/blip.py","language":"python","start_line":28,"end_line":59,"context_start_line":8,"context_end_line":59,"code":"import logging\nimport os\n\nimport torch\nfrom lavis.common.dist_utils import download_cached_file\nfrom lavis.common.utils import is_url\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.vit import interpolate_pos_embed\nfrom transformers import BertTokenizer\n\n\nclass BlipBase(BaseModel):\n @classmethod\n def init_tokenizer(cls):\n tokenizer = BertTokenizer.from_pretrained(\"bert-base-uncased\")\n tokenizer.add_special_tokens({\"bos_token\": \"[DEC]\"})\n tokenizer.add_special_tokens({\"additional_special_tokens\": [\"[ENC]\"]})\n tokenizer.enc_token_id = tokenizer.additional_special_tokens_ids[0]\n return tokenizer\n\n def load_from_pretrained(self, url_or_filename):\n if is_url(url_or_filename):\n cached_file = download_cached_file(\n url_or_filename, check_hash=False, progress=True\n )\n checkpoint = torch.load(cached_file, map_location=\"cpu\")\n elif os.path.isfile(url_or_filename):\n checkpoint = torch.load(url_or_filename, map_location=\"cpu\")\n else:\n raise RuntimeError(\"checkpoint url or path is invalid\")\n\n state_dict = checkpoint[\"model\"]\n\n state_dict[\"visual_encoder.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder.pos_embed\"], self.visual_encoder\n )\n if \"visual_encoder_m.pos_embed\" in self.state_dict().keys():\n state_dict[\"visual_encoder_m.pos_embed\"] = interpolate_pos_embed(\n state_dict[\"visual_encoder_m.pos_embed\"], self.visual_encoder_m\n )\n\n for key in self.state_dict().keys():\n if key in state_dict.keys():\n if state_dict[key].shape != self.state_dict()[key].shape:\n del state_dict[key]\n\n msg = self.load_state_dict(state_dict, strict=False)\n\n logging.info(\"Missing keys {}\".format(msg.missing_keys))\n logging.info(\"load checkpoint from %s\" % url_or_filename)\n\n return msg","source_hash":"baa9da1d24ec8956a509dda74ac6f8db9bff9511dbedb515cf6976f2dd0289d6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained","uri":"program://CREMA/module/lavis.models.clip_models.pretrained#L1-L182","kind":"module","name":"lavis.models.clip_models.pretrained","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":1,"end_line":182,"context_start_line":1,"context_end_line":182,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nimport hashlib\nimport os\nimport urllib\nimport warnings\n\nfrom tqdm import tqdm\n\n_RN50 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt\",\n yfcc15m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt\",\n cc12m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt\",\n)\n\n_RN50_quickgelu = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/afeb0e10f9e5a86da6080e35cf09123aca3b358a0c3e3b6c78a7b63bc04b6762/RN50.pt\",\n yfcc15m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-yfcc15m-455df137.pt\",\n cc12m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn50-quickgelu-cc12m-f000538c.pt\",\n)\n\n_RN101 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt\",\n yfcc15m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt\",\n)\n\n_RN101_quickgelu = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/8fa8567bab74a42d41c5915025a8e4538c3bdbe8804a470a72f30b0d94fab599/RN101.pt\",\n yfcc15m=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/rn101-quickgelu-yfcc15m-3e04b30e.pt\",\n)\n\n_RN50x4 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/7e526bd135e493cef0776de27d5f42653e6b4c8bf9e0f653bb11773263205fdd/RN50x4.pt\",\n)\n\n_RN50x16 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/52378b407f34354e150460fe41077663dd5b39c54cd0bfd2b27167a4a06ec9aa/RN50x16.pt\",\n)\n\n_RN50x64 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/be1cfb55d75a9666199fb2206c106743da0f6468c9d327f3e0d0a543a9919d9c/RN50x64.pt\",\n)\n\n_VITB32 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt\",\n laion400m_e31=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt\",\n laion400m_e32=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt\",\n laion400m_avg=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt\",\n)\n\n_VITB32_quickgelu = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt\",\n laion400m_e31=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e31-d867053b.pt\",\n laion400m_e32=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_e32-46683a32.pt\",\n laion400m_avg=\"https://github.com/mlfoundations/open_clip/releases/download/v0.2-weights/vit_b_32-quickgelu-laion400m_avg-8a00ab3c.pt\",\n)\n\n_VITB16 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt\",\n)\n\n_VITL14 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt\",\n)\n\n_VITL14_336 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt\"\n)\n\n_PRETRAINED = {\n \"RN50\": _RN50,\n \"RN50-quickgelu\": _RN50_quickgelu,\n \"RN101\": _RN101,\n \"RN101-quickgelu\": _RN101_quickgelu,\n \"RN50x4\": _RN50x4,\n \"RN50x16\": _RN50x16,\n \"ViT-B-32\": _VITB32,\n \"ViT-B-32-quickgelu\": _VITB32_quickgelu,\n \"ViT-B-16\": _VITB16,\n \"ViT-L-14\": _VITL14,\n \"ViT-L-14-336\": _VITL14_336,\n}\n\n\ndef list_pretrained(as_str: bool = False):\n \"\"\"returns list of pretrained models\n Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True\n \"\"\"\n return [\n \":\".join([k, t]) if as_str else (k, t)\n for k in _PRETRAINED.keys()\n for t in _PRETRAINED[k].keys()\n ]\n\n\ndef list_pretrained_tag_models(tag: str):\n \"\"\"return all models having the specified pretrain tag\"\"\"\n models = []\n for k in _PRETRAINED.keys():\n if tag in _PRETRAINED[k]:\n models.append(k)\n return models\n\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):\n if model not in _PRETRAINED:\n return \"\"\n model_pretrained = _PRETRAINED[model]\n tag = tag.lower()\n if tag not in model_pretrained:\n return \"\"\n return model_pretrained[tag]\n\n\ndef download_pretrained(url: str, root: str = os.path.expanduser(\"~/.cache/clip\")):\n os.makedirs(root, exist_ok=True)\n filename = os.path.basename(url)\n\n if \"openaipublic\" in url:\n expected_sha256 = url.split(\"/\")[-2]\n else:\n expected_sha256 = \"\"\n\n download_target = os.path.join(root, filename)\n\n if os.path.exists(download_target) and not os.path.isfile(download_target):\n raise RuntimeError(f\"{download_target} exists and is not a regular file\")\n\n if os.path.isfile(download_target):\n if expected_sha256:\n if (\n hashlib.sha256(open(download_target, \"rb\").read()).hexdigest()\n == expected_sha256\n ):\n return download_target\n else:\n warnings.warn(\n f\"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file\"\n )\n else:\n return download_target\n\n with urllib.request.urlopen(url) as source, open(download_target, \"wb\") as output:\n with tqdm(\n total=int(source.info().get(\"Content-Length\")),\n ncols=80,\n unit=\"iB\",\n unit_scale=True,\n ) as loop:\n while True:\n buffer = source.read(8192)\n if not buffer:\n break\n\n output.write(buffer)\n loop.update(len(buffer))\n\n if (\n expected_sha256\n and hashlib.sha256(open(download_target, \"rb\").read()).hexdigest()\n != expected_sha256\n ):\n raise RuntimeError(\n f\"Model has been downloaded but the SHA256 checksum does not not match\"\n )\n\n return download_target","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained.list_pretrained","uri":"program://CREMA/function/lavis.models.clip_models.pretrained.list_pretrained#L92-L100","kind":"function","name":"list_pretrained","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":92,"end_line":100,"context_start_line":72,"context_end_line":120,"code":"\n_VITL14_336 = dict(\n openai=\"https://openaipublic.azureedge.net/clip/models/3035c92b350959924f9f00213499208652fc7ea050643e8b385c2dac08641f02/ViT-L-14-336px.pt\"\n)\n\n_PRETRAINED = {\n \"RN50\": _RN50,\n \"RN50-quickgelu\": _RN50_quickgelu,\n \"RN101\": _RN101,\n \"RN101-quickgelu\": _RN101_quickgelu,\n \"RN50x4\": _RN50x4,\n \"RN50x16\": _RN50x16,\n \"ViT-B-32\": _VITB32,\n \"ViT-B-32-quickgelu\": _VITB32_quickgelu,\n \"ViT-B-16\": _VITB16,\n \"ViT-L-14\": _VITL14,\n \"ViT-L-14-336\": _VITL14_336,\n}\n\n\ndef list_pretrained(as_str: bool = False):\n \"\"\"returns list of pretrained models\n Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True\n \"\"\"\n return [\n \":\".join([k, t]) if as_str else (k, t)\n for k in _PRETRAINED.keys()\n for t in _PRETRAINED[k].keys()\n ]\n\n\ndef list_pretrained_tag_models(tag: str):\n \"\"\"return all models having the specified pretrain tag\"\"\"\n models = []\n for k in _PRETRAINED.keys():\n if tag in _PRETRAINED[k]:\n models.append(k)\n return models\n\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained.list_pretrained_tag_models","uri":"program://CREMA/function/lavis.models.clip_models.pretrained.list_pretrained_tag_models#L103-L109","kind":"function","name":"list_pretrained_tag_models","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":103,"end_line":109,"context_start_line":83,"context_end_line":129,"code":" \"RN50x16\": _RN50x16,\n \"ViT-B-32\": _VITB32,\n \"ViT-B-32-quickgelu\": _VITB32_quickgelu,\n \"ViT-B-16\": _VITB16,\n \"ViT-L-14\": _VITL14,\n \"ViT-L-14-336\": _VITL14_336,\n}\n\n\ndef list_pretrained(as_str: bool = False):\n \"\"\"returns list of pretrained models\n Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True\n \"\"\"\n return [\n \":\".join([k, t]) if as_str else (k, t)\n for k in _PRETRAINED.keys()\n for t in _PRETRAINED[k].keys()\n ]\n\n\ndef list_pretrained_tag_models(tag: str):\n \"\"\"return all models having the specified pretrain tag\"\"\"\n models = []\n for k in _PRETRAINED.keys():\n if tag in _PRETRAINED[k]:\n models.append(k)\n return models\n\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):\n if model not in _PRETRAINED:\n return \"\"\n model_pretrained = _PRETRAINED[model]\n tag = tag.lower()\n if tag not in model_pretrained:\n return \"\"\n return model_pretrained[tag]\n\n","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained.list_pretrained_model_tags","uri":"program://CREMA/function/lavis.models.clip_models.pretrained.list_pretrained_model_tags#L112-L117","kind":"function","name":"list_pretrained_model_tags","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":112,"end_line":117,"context_start_line":92,"context_end_line":137,"code":"def list_pretrained(as_str: bool = False):\n \"\"\"returns list of pretrained models\n Returns a tuple (model_name, pretrain_tag) by default or 'name:tag' if as_str == True\n \"\"\"\n return [\n \":\".join([k, t]) if as_str else (k, t)\n for k in _PRETRAINED.keys()\n for t in _PRETRAINED[k].keys()\n ]\n\n\ndef list_pretrained_tag_models(tag: str):\n \"\"\"return all models having the specified pretrain tag\"\"\"\n models = []\n for k in _PRETRAINED.keys():\n if tag in _PRETRAINED[k]:\n models.append(k)\n return models\n\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):\n if model not in _PRETRAINED:\n return \"\"\n model_pretrained = _PRETRAINED[model]\n tag = tag.lower()\n if tag not in model_pretrained:\n return \"\"\n return model_pretrained[tag]\n\n\ndef download_pretrained(url: str, root: str = os.path.expanduser(\"~/.cache/clip\")):\n os.makedirs(root, exist_ok=True)\n filename = os.path.basename(url)\n\n if \"openaipublic\" in url:\n expected_sha256 = url.split(\"/\")[-2]\n else:\n expected_sha256 = \"\"","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained.get_pretrained_url","uri":"program://CREMA/function/lavis.models.clip_models.pretrained.get_pretrained_url#L120-L127","kind":"function","name":"get_pretrained_url","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":120,"end_line":127,"context_start_line":100,"context_end_line":147,"code":" ]\n\n\ndef list_pretrained_tag_models(tag: str):\n \"\"\"return all models having the specified pretrain tag\"\"\"\n models = []\n for k in _PRETRAINED.keys():\n if tag in _PRETRAINED[k]:\n models.append(k)\n return models\n\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):\n if model not in _PRETRAINED:\n return \"\"\n model_pretrained = _PRETRAINED[model]\n tag = tag.lower()\n if tag not in model_pretrained:\n return \"\"\n return model_pretrained[tag]\n\n\ndef download_pretrained(url: str, root: str = os.path.expanduser(\"~/.cache/clip\")):\n os.makedirs(root, exist_ok=True)\n filename = os.path.basename(url)\n\n if \"openaipublic\" in url:\n expected_sha256 = url.split(\"/\")[-2]\n else:\n expected_sha256 = \"\"\n\n download_target = os.path.join(root, filename)\n\n if os.path.exists(download_target) and not os.path.isfile(download_target):\n raise RuntimeError(f\"{download_target} exists and is not a regular file\")\n\n if os.path.isfile(download_target):\n if expected_sha256:\n if (\n hashlib.sha256(open(download_target, \"rb\").read()).hexdigest()","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.pretrained.download_pretrained","uri":"program://CREMA/function/lavis.models.clip_models.pretrained.download_pretrained#L130-L182","kind":"function","name":"download_pretrained","path":"lavis/models/clip_models/pretrained.py","language":"python","start_line":130,"end_line":182,"context_start_line":110,"context_end_line":182,"code":"\n\ndef list_pretrained_model_tags(model: str):\n \"\"\"return all pretrain tags for the specified model architecture\"\"\"\n tags = []\n if model in _PRETRAINED:\n tags.extend(_PRETRAINED[model].keys())\n return tags\n\n\ndef get_pretrained_url(model: str, tag: str):\n if model not in _PRETRAINED:\n return \"\"\n model_pretrained = _PRETRAINED[model]\n tag = tag.lower()\n if tag not in model_pretrained:\n return \"\"\n return model_pretrained[tag]\n\n\ndef download_pretrained(url: str, root: str = os.path.expanduser(\"~/.cache/clip\")):\n os.makedirs(root, exist_ok=True)\n filename = os.path.basename(url)\n\n if \"openaipublic\" in url:\n expected_sha256 = url.split(\"/\")[-2]\n else:\n expected_sha256 = \"\"\n\n download_target = os.path.join(root, filename)\n\n if os.path.exists(download_target) and not os.path.isfile(download_target):\n raise RuntimeError(f\"{download_target} exists and is not a regular file\")\n\n if os.path.isfile(download_target):\n if expected_sha256:\n if (\n hashlib.sha256(open(download_target, \"rb\").read()).hexdigest()\n == expected_sha256\n ):\n return download_target\n else:\n warnings.warn(\n f\"{download_target} exists, but the SHA256 checksum does not match; re-downloading the file\"\n )\n else:\n return download_target\n\n with urllib.request.urlopen(url) as source, open(download_target, \"wb\") as output:\n with tqdm(\n total=int(source.info().get(\"Content-Length\")),\n ncols=80,\n unit=\"iB\",\n unit_scale=True,\n ) as loop:\n while True:\n buffer = source.read(8192)\n if not buffer:\n break\n\n output.write(buffer)\n loop.update(len(buffer))\n\n if (\n expected_sha256\n and hashlib.sha256(open(download_target, \"rb\").read()).hexdigest()\n != expected_sha256\n ):\n raise RuntimeError(\n f\"Model has been downloaded but the SHA256 checksum does not not match\"\n )\n\n return download_target","source_hash":"8bb8584e9490b371a382f389ed77e983120e39ac2d0591a02cb7319842a43ae3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.loss","uri":"program://CREMA/module/lavis.models.clip_models.loss#L1-L141","kind":"module","name":"lavis.models.clip_models.loss","path":"lavis/models/clip_models/loss.py","language":"python","start_line":1,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport torch\nimport torch.distributed.nn\nfrom torch import distributed as dist, nn as nn\nfrom torch.nn import functional as F\n\ntry:\n import horovod.torch as hvd\nexcept ImportError:\n hvd = None\n\n\ndef gather_features(\n image_features,\n text_features,\n local_loss=False,\n gather_with_grad=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n):\n if use_horovod:\n assert hvd is not None, \"Please install horovod\"\n if gather_with_grad:\n all_image_features = hvd.allgather(image_features)\n all_text_features = hvd.allgather(text_features)\n else:\n with torch.no_grad():\n all_image_features = hvd.allgather(image_features)\n all_text_features = hvd.allgather(text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features = list(\n all_image_features.chunk(world_size, dim=0)\n )\n gathered_text_features = list(\n all_text_features.chunk(world_size, dim=0)\n )\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n else:\n # We gather tensors from all gpus\n if gather_with_grad:\n all_image_features = torch.cat(\n torch.distributed.nn.all_gather(image_features), dim=0\n )\n all_text_features = torch.cat(\n torch.distributed.nn.all_gather(text_features), dim=0\n )\n else:\n gathered_image_features = [\n torch.zeros_like(image_features) for _ in range(world_size)\n ]\n gathered_text_features = [\n torch.zeros_like(text_features) for _ in range(world_size)\n ]\n dist.all_gather(gathered_image_features, image_features)\n dist.all_gather(gathered_text_features, text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n\n return all_image_features, all_text_features\n\n\nclass ClipLoss(nn.Module):\n def __init__(\n self,\n local_loss=False,\n gather_with_grad=False,\n cache_labels=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n ):\n super().__init__()\n self.local_loss = local_loss\n self.gather_with_grad = gather_with_grad\n self.cache_labels = cache_labels\n self.rank = rank\n self.world_size = world_size\n self.use_horovod = use_horovod\n\n # cache state\n self.prev_num_logits = 0\n self.labels = {}\n\n def forward(self, image_features, text_features, logit_scale):\n device = image_features.device\n if self.world_size > 1:\n all_image_features, all_text_features = gather_features(\n image_features,\n text_features,\n self.local_loss,\n self.gather_with_grad,\n self.rank,\n self.world_size,\n self.use_horovod,\n )\n\n if self.local_loss:\n logits_per_image = logit_scale * image_features @ all_text_features.T\n logits_per_text = logit_scale * text_features @ all_image_features.T\n else:\n logits_per_image = (\n logit_scale * all_image_features @ all_text_features.T\n )\n logits_per_text = logits_per_image.T\n else:\n logits_per_image = logit_scale * image_features @ text_features.T\n logits_per_text = logit_scale * text_features @ image_features.T\n\n # calculated ground-truth and cache if enabled\n num_logits = logits_per_image.shape[0]\n if self.prev_num_logits != num_logits or device not in self.labels:\n labels = torch.arange(num_logits, device=device, dtype=torch.long)\n if self.world_size > 1 and self.local_loss:\n labels = labels + num_logits * self.rank\n if self.cache_labels:\n self.labels[device] = labels\n self.prev_num_logits = num_logits\n else:\n labels = self.labels[device]\n\n total_loss = (\n F.cross_entropy(logits_per_image, labels)\n + F.cross_entropy(logits_per_text, labels)\n ) / 2\n return total_loss","source_hash":"df55b80c98fb24ec2f940d26f73df84a0f9518db17537ccaa4df955179212118","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.loss.gather_features","uri":"program://CREMA/function/lavis.models.clip_models.loss.gather_features#L20-L75","kind":"function","name":"gather_features","path":"lavis/models/clip_models/loss.py","language":"python","start_line":20,"end_line":75,"context_start_line":1,"context_end_line":95,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport torch\nimport torch.distributed.nn\nfrom torch import distributed as dist, nn as nn\nfrom torch.nn import functional as F\n\ntry:\n import horovod.torch as hvd\nexcept ImportError:\n hvd = None\n\n\ndef gather_features(\n image_features,\n text_features,\n local_loss=False,\n gather_with_grad=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n):\n if use_horovod:\n assert hvd is not None, \"Please install horovod\"\n if gather_with_grad:\n all_image_features = hvd.allgather(image_features)\n all_text_features = hvd.allgather(text_features)\n else:\n with torch.no_grad():\n all_image_features = hvd.allgather(image_features)\n all_text_features = hvd.allgather(text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features = list(\n all_image_features.chunk(world_size, dim=0)\n )\n gathered_text_features = list(\n all_text_features.chunk(world_size, dim=0)\n )\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n else:\n # We gather tensors from all gpus\n if gather_with_grad:\n all_image_features = torch.cat(\n torch.distributed.nn.all_gather(image_features), dim=0\n )\n all_text_features = torch.cat(\n torch.distributed.nn.all_gather(text_features), dim=0\n )\n else:\n gathered_image_features = [\n torch.zeros_like(image_features) for _ in range(world_size)\n ]\n gathered_text_features = [\n torch.zeros_like(text_features) for _ in range(world_size)\n ]\n dist.all_gather(gathered_image_features, image_features)\n dist.all_gather(gathered_text_features, text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n\n return all_image_features, all_text_features\n\n\nclass ClipLoss(nn.Module):\n def __init__(\n self,\n local_loss=False,\n gather_with_grad=False,\n cache_labels=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n ):\n super().__init__()\n self.local_loss = local_loss\n self.gather_with_grad = gather_with_grad\n self.cache_labels = cache_labels\n self.rank = rank\n self.world_size = world_size\n self.use_horovod = use_horovod\n","source_hash":"df55b80c98fb24ec2f940d26f73df84a0f9518db17537ccaa4df955179212118","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.loss.ClipLoss","uri":"program://CREMA/class/lavis.models.clip_models.loss.ClipLoss#L78-L141","kind":"class","name":"ClipLoss","path":"lavis/models/clip_models/loss.py","language":"python","start_line":78,"end_line":141,"context_start_line":58,"context_end_line":141,"code":" )\n else:\n gathered_image_features = [\n torch.zeros_like(image_features) for _ in range(world_size)\n ]\n gathered_text_features = [\n torch.zeros_like(text_features) for _ in range(world_size)\n ]\n dist.all_gather(gathered_image_features, image_features)\n dist.all_gather(gathered_text_features, text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n\n return all_image_features, all_text_features\n\n\nclass ClipLoss(nn.Module):\n def __init__(\n self,\n local_loss=False,\n gather_with_grad=False,\n cache_labels=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n ):\n super().__init__()\n self.local_loss = local_loss\n self.gather_with_grad = gather_with_grad\n self.cache_labels = cache_labels\n self.rank = rank\n self.world_size = world_size\n self.use_horovod = use_horovod\n\n # cache state\n self.prev_num_logits = 0\n self.labels = {}\n\n def forward(self, image_features, text_features, logit_scale):\n device = image_features.device\n if self.world_size > 1:\n all_image_features, all_text_features = gather_features(\n image_features,\n text_features,\n self.local_loss,\n self.gather_with_grad,\n self.rank,\n self.world_size,\n self.use_horovod,\n )\n\n if self.local_loss:\n logits_per_image = logit_scale * image_features @ all_text_features.T\n logits_per_text = logit_scale * text_features @ all_image_features.T\n else:\n logits_per_image = (\n logit_scale * all_image_features @ all_text_features.T\n )\n logits_per_text = logits_per_image.T\n else:\n logits_per_image = logit_scale * image_features @ text_features.T\n logits_per_text = logit_scale * text_features @ image_features.T\n\n # calculated ground-truth and cache if enabled\n num_logits = logits_per_image.shape[0]\n if self.prev_num_logits != num_logits or device not in self.labels:\n labels = torch.arange(num_logits, device=device, dtype=torch.long)\n if self.world_size > 1 and self.local_loss:\n labels = labels + num_logits * self.rank\n if self.cache_labels:\n self.labels[device] = labels\n self.prev_num_logits = num_logits\n else:\n labels = self.labels[device]\n\n total_loss = (\n F.cross_entropy(logits_per_image, labels)\n + F.cross_entropy(logits_per_text, labels)\n ) / 2\n return total_loss","source_hash":"df55b80c98fb24ec2f940d26f73df84a0f9518db17537ccaa4df955179212118","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.loss.__init__","uri":"program://CREMA/function/lavis.models.clip_models.loss.__init__#L79-L98","kind":"function","name":"__init__","path":"lavis/models/clip_models/loss.py","language":"python","start_line":79,"end_line":98,"context_start_line":59,"context_end_line":118,"code":" else:\n gathered_image_features = [\n torch.zeros_like(image_features) for _ in range(world_size)\n ]\n gathered_text_features = [\n torch.zeros_like(text_features) for _ in range(world_size)\n ]\n dist.all_gather(gathered_image_features, image_features)\n dist.all_gather(gathered_text_features, text_features)\n if not local_loss:\n # ensure grads for local rank when all_* features don't have a gradient\n gathered_image_features[rank] = image_features\n gathered_text_features[rank] = text_features\n all_image_features = torch.cat(gathered_image_features, dim=0)\n all_text_features = torch.cat(gathered_text_features, dim=0)\n\n return all_image_features, all_text_features\n\n\nclass ClipLoss(nn.Module):\n def __init__(\n self,\n local_loss=False,\n gather_with_grad=False,\n cache_labels=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n ):\n super().__init__()\n self.local_loss = local_loss\n self.gather_with_grad = gather_with_grad\n self.cache_labels = cache_labels\n self.rank = rank\n self.world_size = world_size\n self.use_horovod = use_horovod\n\n # cache state\n self.prev_num_logits = 0\n self.labels = {}\n\n def forward(self, image_features, text_features, logit_scale):\n device = image_features.device\n if self.world_size > 1:\n all_image_features, all_text_features = gather_features(\n image_features,\n text_features,\n self.local_loss,\n self.gather_with_grad,\n self.rank,\n self.world_size,\n self.use_horovod,\n )\n\n if self.local_loss:\n logits_per_image = logit_scale * image_features @ all_text_features.T\n logits_per_text = logit_scale * text_features @ all_image_features.T\n else:\n logits_per_image = (\n logit_scale * all_image_features @ all_text_features.T","source_hash":"df55b80c98fb24ec2f940d26f73df84a0f9518db17537ccaa4df955179212118","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.loss.forward","uri":"program://CREMA/function/lavis.models.clip_models.loss.forward#L100-L141","kind":"function","name":"forward","path":"lavis/models/clip_models/loss.py","language":"python","start_line":100,"end_line":141,"context_start_line":80,"context_end_line":141,"code":" self,\n local_loss=False,\n gather_with_grad=False,\n cache_labels=False,\n rank=0,\n world_size=1,\n use_horovod=False,\n ):\n super().__init__()\n self.local_loss = local_loss\n self.gather_with_grad = gather_with_grad\n self.cache_labels = cache_labels\n self.rank = rank\n self.world_size = world_size\n self.use_horovod = use_horovod\n\n # cache state\n self.prev_num_logits = 0\n self.labels = {}\n\n def forward(self, image_features, text_features, logit_scale):\n device = image_features.device\n if self.world_size > 1:\n all_image_features, all_text_features = gather_features(\n image_features,\n text_features,\n self.local_loss,\n self.gather_with_grad,\n self.rank,\n self.world_size,\n self.use_horovod,\n )\n\n if self.local_loss:\n logits_per_image = logit_scale * image_features @ all_text_features.T\n logits_per_text = logit_scale * text_features @ all_image_features.T\n else:\n logits_per_image = (\n logit_scale * all_image_features @ all_text_features.T\n )\n logits_per_text = logits_per_image.T\n else:\n logits_per_image = logit_scale * image_features @ text_features.T\n logits_per_text = logit_scale * text_features @ image_features.T\n\n # calculated ground-truth and cache if enabled\n num_logits = logits_per_image.shape[0]\n if self.prev_num_logits != num_logits or device not in self.labels:\n labels = torch.arange(num_logits, device=device, dtype=torch.long)\n if self.world_size > 1 and self.local_loss:\n labels = labels + num_logits * self.rank\n if self.cache_labels:\n self.labels[device] = labels\n self.prev_num_logits = num_logits\n else:\n labels = self.labels[device]\n\n total_loss = (\n F.cross_entropy(logits_per_image, labels)\n + F.cross_entropy(logits_per_text, labels)\n ) / 2\n return total_loss","source_hash":"df55b80c98fb24ec2f940d26f73df84a0f9518db17537ccaa4df955179212118","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model","uri":"program://CREMA/module/lavis.models.clip_models.model#L1-L1254","kind":"module","name":"lavis.models.clip_models.model","path":"lavis/models/clip_models/model.py","language":"python","start_line":1,"end_line":1254,"context_start_line":1,"context_end_line":1254,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\n\"\"\" CLIP Model\nAdapted from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.\n\"\"\"\n\nimport datetime\nimport json\nimport logging\nimport os\nimport re\nimport time\nimport warnings\nfrom collections import OrderedDict\nfrom copy import deepcopy\nfrom dataclasses import dataclass\nfrom pathlib import Path\nfrom typing import Callable, List, Optional, Tuple, Union\n\nimport numpy as np\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.clip_models.clip_outputs import ClipOutput, ClipOutputFeatures\nfrom lavis.models.clip_models.timm_model import TimmModel\nfrom lavis.models.clip_models.transform import image_transform\nfrom lavis.models.clip_models.utils import freeze_batch_norm_2d\nfrom lavis.tasks.multimodal_classification import MultimodalClassificationTask\nfrom torch import nn\n\nfrom .pretrained import (\n download_pretrained,\n get_pretrained_url,\n list_pretrained_tag_models,\n)\n\n_MODEL_CONFIG_PATHS = [Path(__file__).parent.parent.parent / f\"configs/models/clip/\"]\n_MODEL_CONFIGS = {} # directory (model_name: config) of model architecture configs\n\n\nclass Bottleneck(nn.Module):\n expansion = 4\n\n def __init__(self, inplanes, planes, stride=1):\n super().__init__()\n\n # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1\n self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)\n self.bn1 = nn.BatchNorm2d(planes)\n\n self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)\n self.bn2 = nn.BatchNorm2d(planes)\n\n self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()\n\n self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)\n self.bn3 = nn.BatchNorm2d(planes * self.expansion)\n\n self.relu = nn.ReLU(inplace=True)\n self.downsample = None\n self.stride = stride\n\n if stride > 1 or inplanes != planes * Bottleneck.expansion:\n # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1\n self.downsample = nn.Sequential(\n OrderedDict(\n [\n (\"-1\", nn.AvgPool2d(stride)),\n (\n \"0\",\n nn.Conv2d(\n inplanes,\n planes * self.expansion,\n 1,\n stride=1,\n bias=False,\n ),\n ),\n (\"1\", nn.BatchNorm2d(planes * self.expansion)),\n ]\n )\n )\n\n def forward(self, x: torch.Tensor):\n identity = x\n\n out = self.relu(self.bn1(self.conv1(x)))\n out = self.relu(self.bn2(self.conv2(out)))\n out = self.avgpool(out)\n out = self.bn3(self.conv3(out))\n\n if self.downsample is not None:\n identity = self.downsample(x)\n\n out += identity\n out = self.relu(out)\n return out\n\n\nclass AttentionPool2d(nn.Module):\n def __init__(\n self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None\n ):\n super().__init__()\n self.positional_embedding = nn.Parameter(\n torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5\n )\n self.k_proj = nn.Linear(embed_dim, embed_dim)\n self.q_proj = nn.Linear(embed_dim, embed_dim)\n self.v_proj = nn.Linear(embed_dim, embed_dim)\n self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)\n self.num_heads = num_heads\n\n def forward(self, x):\n x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(\n 2, 0, 1\n ) # NCHW -> (HW)NC\n x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC\n x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC\n x, _ = F.multi_head_attention_forward(\n query=x,\n key=x,\n value=x,\n embed_dim_to_check=x.shape[-1],\n num_heads=self.num_heads,\n q_proj_weight=self.q_proj.weight,\n k_proj_weight=self.k_proj.weight,\n v_proj_weight=self.v_proj.weight,\n in_proj_weight=None,\n in_proj_bias=torch.cat(\n [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]\n ),\n bias_k=None,\n bias_v=None,\n add_zero_attn=False,\n dropout_p=0,\n out_proj_weight=self.c_proj.weight,\n out_proj_bias=self.c_proj.bias,\n use_separate_proj_weight=True,\n training=self.training,\n need_weights=False,\n )\n\n return x[0]\n\n\nclass ModifiedResNet(nn.Module):\n \"\"\"\n A ResNet class that is similar to torchvision's but contains the following changes:\n - There are now 3 \"stem\" convolutions as opposed to 1, with an average pool instead of a max pool.\n - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1\n - The final pooling layer is a QKV attention instead of an average pool\n \"\"\"\n\n def __init__(self, layers, output_dim, heads, image_size=224, width=64):\n super().__init__()\n self.output_dim = output_dim\n self.image_size = image_size\n\n # the 3-layer stem\n self.conv1 = nn.Conv2d(\n 3, width // 2, kernel_size=3, stride=2, padding=1, bias=False\n )\n self.bn1 = nn.BatchNorm2d(width // 2)\n self.conv2 = nn.Conv2d(\n width // 2, width // 2, kernel_size=3, padding=1, bias=False\n )\n self.bn2 = nn.BatchNorm2d(width // 2)\n self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)\n self.bn3 = nn.BatchNorm2d(width)\n self.avgpool = nn.AvgPool2d(2)\n self.relu = nn.ReLU(inplace=True)\n\n # residual layers\n self._inplanes = width # this is a *mutable* variable used during construction\n self.layer1 = self._make_layer(width, layers[0])\n self.layer2 = self._make_layer(width * 2, layers[1], stride=2)\n self.layer3 = self._make_layer(width * 4, layers[2], stride=2)\n self.layer4 = self._make_layer(width * 8, layers[3], stride=2)\n\n embed_dim = width * 32 # the ResNet feature dimension\n self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)\n\n self.init_parameters()\n\n def _make_layer(self, planes, blocks, stride=1):\n layers = [Bottleneck(self._inplanes, planes, stride)]\n\n self._inplanes = planes * Bottleneck.expansion\n for _ in range(1, blocks):\n layers.append(Bottleneck(self._inplanes, planes))\n\n return nn.Sequential(*layers)\n\n def init_parameters(self):\n if self.attnpool is not None:\n std = self.attnpool.c_proj.in_features**-0.5\n nn.init.normal_(self.attnpool.q_proj.weight, std=std)\n nn.init.normal_(self.attnpool.k_proj.weight, std=std)\n nn.init.normal_(self.attnpool.v_proj.weight, std=std)\n nn.init.normal_(self.attnpool.c_proj.weight, std=std)\n\n for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:\n for name, param in resnet_block.named_parameters():\n if name.endswith(\"bn3.weight\"):\n nn.init.zeros_(param)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self)\n\n def stem(self, x):\n for conv, bn in [\n (self.conv1, self.bn1),\n (self.conv2, self.bn2),\n (self.conv3, self.bn3),\n ]:\n x = self.relu(bn(conv(x)))\n x = self.avgpool(x)\n return x\n\n def forward(self, x):\n x = self.stem(x)\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n\n\nclass QuickGELU(nn.Module):\n # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),\n ]\n )\n )\n self.ln_2 = LayerNorm(d_model)\n\n def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(\n self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU\n ):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.ModuleList(\n [\n ResidualAttentionBlock(width, heads, act_layer=act_layer)\n for _ in range(layers)\n ]\n )\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n for r in self.resblocks:\n x = r(x, attn_mask=attn_mask)\n return x\n\n\nclass VisualTransformer(nn.Module):\n def __init__(\n self,\n image_size: int,\n patch_size: int,\n width: int,\n layers: int,\n heads: int,\n output_dim: int,\n act_layer: Callable = nn.GELU,\n ):\n super().__init__()\n self.image_size = image_size\n self.output_dim = output_dim\n self.conv1 = nn.Conv2d(\n in_channels=3,\n out_channels=width,\n kernel_size=patch_size,\n stride=patch_size,\n bias=False,\n )\n\n scale = width**-0.5\n self.class_embedding = nn.Parameter(scale * torch.randn(width))\n self.positional_embedding = nn.Parameter(\n scale * torch.randn((image_size // patch_size) ** 2 + 1, width)\n )\n self.ln_pre = LayerNorm(width)\n\n self.transformer = Transformer(width, layers, heads, act_layer=act_layer)\n\n self.ln_post = LayerNorm(width)\n self.proj = nn.Parameter(scale * torch.randn(width, output_dim))\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n\n def forward(self, x: torch.Tensor):\n x = self.conv1(x) # shape = [*, width, grid, grid]\n x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]\n x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]\n x = torch.cat(\n [\n self.class_embedding.to(x.dtype)\n + torch.zeros(\n x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device\n ),\n x,\n ],\n dim=1,\n ) # shape = [*, grid ** 2 + 1, width]\n x = x + self.positional_embedding.to(x.dtype)\n x = self.ln_pre(x)\n\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n\n x = self.ln_post(x[:, 0, :])\n\n if self.proj is not None:\n x = x @ self.proj\n\n return x\n\n\n@dataclass\nclass CLIPVisionCfg:\n layers: Union[Tuple[int, int, int, int], int] = 12\n width: int = 768\n patch_size: int = 16\n image_size: Union[Tuple[int, int], int] = 224\n timm_model_name: str = (\n None # a valid model name overrides layers, width, patch_size\n )\n timm_model_pretrained: bool = (\n False # use (imagenet) pretrained weights for named model\n )\n timm_pool: str = (\n \"avg\" # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')\n )\n timm_proj: str = (\n \"linear\" # linear projection for timm model output ('linear', 'mlp', '')\n )\n\n\n@dataclass\nclass CLIPTextCfg:\n context_length: int\n vocab_size: int\n width: int\n heads: int\n layers: int\n\n\n@registry.register_model(\"clip\")\n@registry.register_model(\"clip_feature_extractor\")\nclass CLIP(BaseModel):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ViT-B-32\": \"configs/models/clip_vit_base32.yaml\",\n \"ViT-B-16\": \"configs/models/clip_vit_base16.yaml\",\n \"ViT-L-14\": \"configs/models/clip_vit_large14.yaml\",\n \"ViT-L-14-336\": \"configs/models/clip_vit_large14_336.yaml\",\n \"RN50\": \"configs/models/clip_resnet50.yaml\",\n }\n\n def __init__(\n self,\n embed_dim: int,\n vision_cfg: CLIPVisionCfg,\n text_cfg: CLIPTextCfg,\n quick_gelu: bool = False,\n ):\n from .tokenizer import tokenize\n\n super().__init__()\n\n self.tokenizer = tokenize\n self._loss = None\n\n if isinstance(vision_cfg, dict):\n vision_cfg = CLIPVisionCfg(**vision_cfg)\n if isinstance(text_cfg, dict):\n text_cfg = CLIPTextCfg(**text_cfg)\n\n self.context_length = text_cfg.context_length\n\n # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more\n # memory efficient in recent PyTorch releases (>= 1.10).\n # NOTE: timm models always use native GELU regardless of quick_gelu flag.\n act_layer = QuickGELU if quick_gelu else nn.GELU\n\n if vision_cfg.timm_model_name:\n self.visual = TimmModel(\n vision_cfg.timm_model_name,\n pretrained=vision_cfg.timm_model_pretrained,\n pool=vision_cfg.timm_pool,\n proj=vision_cfg.timm_proj,\n embed_dim=embed_dim,\n image_size=vision_cfg.image_size,\n )\n act_layer = (\n nn.GELU\n ) # so that text transformer doesn't use QuickGELU w/ timm models\n elif isinstance(vision_cfg.layers, (tuple, list)):\n vision_heads = vision_cfg.width * 32 // 64\n self.visual = ModifiedResNet(\n layers=vision_cfg.layers,\n output_dim=embed_dim,\n heads=vision_heads,\n image_size=vision_cfg.image_size,\n width=vision_cfg.width,\n )\n else:\n vision_heads = vision_cfg.width // 64\n self.visual = VisualTransformer(\n image_size=vision_cfg.image_size,\n patch_size=vision_cfg.patch_size,\n width=vision_cfg.width,\n layers=vision_cfg.layers,\n heads=vision_heads,\n output_dim=embed_dim,\n act_layer=act_layer,\n )\n\n self.transformer = Transformer(\n width=text_cfg.width,\n layers=text_cfg.layers,\n heads=text_cfg.heads,\n act_layer=act_layer,\n )\n\n self.vocab_size = text_cfg.vocab_size\n self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)\n self.positional_embedding = nn.Parameter(\n torch.empty(self.context_length, text_cfg.width)\n )\n self.ln_final = LayerNorm(text_cfg.width)\n\n self.text_projection = nn.Parameter(torch.empty(text_cfg.width, embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n self.register_buffer(\"attn_mask\", self.build_attention_mask(), persistent=False)\n\n self.prompt_templates = openai_imagenet_template\n self.classifier = None\n\n self.init_parameters()\n\n @property\n def loss(self):\n if self._loss is None:\n from lavis.models.clip_models.loss import ClipLoss\n from torch import distributed as dist\n\n self._loss = ClipLoss(\n world_size=dist.get_world_size(),\n rank=dist.get_rank(),\n local_loss=False,\n gather_with_grad=False,\n use_horovod=False,\n )\n\n return self._loss\n\n def init_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n nn.init.constant_(self.logit_scale, np.log(1 / 0.07))\n\n if hasattr(self.visual, \"init_parameters\"):\n self.visual.init_parameters()\n\n proj_std = (self.transformer.width**-0.5) * (\n (2 * self.transformer.layers) ** -0.5\n )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):\n def forward(self, samples):\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n# ... truncated ...","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.Bottleneck","uri":"program://CREMA/class/lavis.models.clip_models.model.Bottleneck#L50-L106","kind":"class","name":"Bottleneck","path":"lavis/models/clip_models/model.py","language":"python","start_line":50,"end_line":106,"context_start_line":30,"context_end_line":126,"code":"from lavis.common.registry import registry\nfrom lavis.common.utils import get_abs_path\nfrom lavis.models.base_model import BaseModel\nfrom lavis.models.clip_models.clip_outputs import ClipOutput, ClipOutputFeatures\nfrom lavis.models.clip_models.timm_model import TimmModel\nfrom lavis.models.clip_models.transform import image_transform\nfrom lavis.models.clip_models.utils import freeze_batch_norm_2d\nfrom lavis.tasks.multimodal_classification import MultimodalClassificationTask\nfrom torch import nn\n\nfrom .pretrained import (\n download_pretrained,\n get_pretrained_url,\n list_pretrained_tag_models,\n)\n\n_MODEL_CONFIG_PATHS = [Path(__file__).parent.parent.parent / f\"configs/models/clip/\"]\n_MODEL_CONFIGS = {} # directory (model_name: config) of model architecture configs\n\n\nclass Bottleneck(nn.Module):\n expansion = 4\n\n def __init__(self, inplanes, planes, stride=1):\n super().__init__()\n\n # all conv layers have stride 1. an avgpool is performed after the second convolution when stride > 1\n self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)\n self.bn1 = nn.BatchNorm2d(planes)\n\n self.conv2 = nn.Conv2d(planes, planes, 3, padding=1, bias=False)\n self.bn2 = nn.BatchNorm2d(planes)\n\n self.avgpool = nn.AvgPool2d(stride) if stride > 1 else nn.Identity()\n\n self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)\n self.bn3 = nn.BatchNorm2d(planes * self.expansion)\n\n self.relu = nn.ReLU(inplace=True)\n self.downsample = None\n self.stride = stride\n\n if stride > 1 or inplanes != planes * Bottleneck.expansion:\n # downsampling layer is prepended with an avgpool, and the subsequent convolution has stride 1\n self.downsample = nn.Sequential(\n OrderedDict(\n [\n (\"-1\", nn.AvgPool2d(stride)),\n (\n \"0\",\n nn.Conv2d(\n inplanes,\n planes * self.expansion,\n 1,\n stride=1,\n bias=False,\n ),\n ),\n (\"1\", nn.BatchNorm2d(planes * self.expansion)),\n ]\n )\n )\n\n def forward(self, x: torch.Tensor):\n identity = x\n\n out = self.relu(self.bn1(self.conv1(x)))\n out = self.relu(self.bn2(self.conv2(out)))\n out = self.avgpool(out)\n out = self.bn3(self.conv3(out))\n\n if self.downsample is not None:\n identity = self.downsample(x)\n\n out += identity\n out = self.relu(out)\n return out\n\n\nclass AttentionPool2d(nn.Module):\n def __init__(\n self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None\n ):\n super().__init__()\n self.positional_embedding = nn.Parameter(\n torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5\n )\n self.k_proj = nn.Linear(embed_dim, embed_dim)\n self.q_proj = nn.Linear(embed_dim, embed_dim)\n self.v_proj = nn.Linear(embed_dim, embed_dim)\n self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)\n self.num_heads = num_heads\n\n def forward(self, x):\n x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(\n 2, 0, 1\n ) # NCHW -> (HW)NC","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.AttentionPool2d","uri":"program://CREMA/class/lavis.models.clip_models.model.AttentionPool2d#L109-L153","kind":"class","name":"AttentionPool2d","path":"lavis/models/clip_models/model.py","language":"python","start_line":109,"end_line":153,"context_start_line":89,"context_end_line":173,"code":" ]\n )\n )\n\n def forward(self, x: torch.Tensor):\n identity = x\n\n out = self.relu(self.bn1(self.conv1(x)))\n out = self.relu(self.bn2(self.conv2(out)))\n out = self.avgpool(out)\n out = self.bn3(self.conv3(out))\n\n if self.downsample is not None:\n identity = self.downsample(x)\n\n out += identity\n out = self.relu(out)\n return out\n\n\nclass AttentionPool2d(nn.Module):\n def __init__(\n self, spacial_dim: int, embed_dim: int, num_heads: int, output_dim: int = None\n ):\n super().__init__()\n self.positional_embedding = nn.Parameter(\n torch.randn(spacial_dim**2 + 1, embed_dim) / embed_dim**0.5\n )\n self.k_proj = nn.Linear(embed_dim, embed_dim)\n self.q_proj = nn.Linear(embed_dim, embed_dim)\n self.v_proj = nn.Linear(embed_dim, embed_dim)\n self.c_proj = nn.Linear(embed_dim, output_dim or embed_dim)\n self.num_heads = num_heads\n\n def forward(self, x):\n x = x.reshape(x.shape[0], x.shape[1], x.shape[2] * x.shape[3]).permute(\n 2, 0, 1\n ) # NCHW -> (HW)NC\n x = torch.cat([x.mean(dim=0, keepdim=True), x], dim=0) # (HW+1)NC\n x = x + self.positional_embedding[:, None, :].to(x.dtype) # (HW+1)NC\n x, _ = F.multi_head_attention_forward(\n query=x,\n key=x,\n value=x,\n embed_dim_to_check=x.shape[-1],\n num_heads=self.num_heads,\n q_proj_weight=self.q_proj.weight,\n k_proj_weight=self.k_proj.weight,\n v_proj_weight=self.v_proj.weight,\n in_proj_weight=None,\n in_proj_bias=torch.cat(\n [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]\n ),\n bias_k=None,\n bias_v=None,\n add_zero_attn=False,\n dropout_p=0,\n out_proj_weight=self.c_proj.weight,\n out_proj_bias=self.c_proj.bias,\n use_separate_proj_weight=True,\n training=self.training,\n need_weights=False,\n )\n\n return x[0]\n\n\nclass ModifiedResNet(nn.Module):\n \"\"\"\n A ResNet class that is similar to torchvision's but contains the following changes:\n - There are now 3 \"stem\" convolutions as opposed to 1, with an average pool instead of a max pool.\n - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1\n - The final pooling layer is a QKV attention instead of an average pool\n \"\"\"\n\n def __init__(self, layers, output_dim, heads, image_size=224, width=64):\n super().__init__()\n self.output_dim = output_dim\n self.image_size = image_size\n\n # the 3-layer stem\n self.conv1 = nn.Conv2d(\n 3, width // 2, kernel_size=3, stride=2, padding=1, bias=False\n )\n self.bn1 = nn.BatchNorm2d(width // 2)","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.ModifiedResNet","uri":"program://CREMA/class/lavis.models.clip_models.model.ModifiedResNet#L156-L244","kind":"class","name":"ModifiedResNet","path":"lavis/models/clip_models/model.py","language":"python","start_line":156,"end_line":244,"context_start_line":136,"context_end_line":264,"code":" k_proj_weight=self.k_proj.weight,\n v_proj_weight=self.v_proj.weight,\n in_proj_weight=None,\n in_proj_bias=torch.cat(\n [self.q_proj.bias, self.k_proj.bias, self.v_proj.bias]\n ),\n bias_k=None,\n bias_v=None,\n add_zero_attn=False,\n dropout_p=0,\n out_proj_weight=self.c_proj.weight,\n out_proj_bias=self.c_proj.bias,\n use_separate_proj_weight=True,\n training=self.training,\n need_weights=False,\n )\n\n return x[0]\n\n\nclass ModifiedResNet(nn.Module):\n \"\"\"\n A ResNet class that is similar to torchvision's but contains the following changes:\n - There are now 3 \"stem\" convolutions as opposed to 1, with an average pool instead of a max pool.\n - Performs anti-aliasing strided convolutions, where an avgpool is prepended to convolutions with stride > 1\n - The final pooling layer is a QKV attention instead of an average pool\n \"\"\"\n\n def __init__(self, layers, output_dim, heads, image_size=224, width=64):\n super().__init__()\n self.output_dim = output_dim\n self.image_size = image_size\n\n # the 3-layer stem\n self.conv1 = nn.Conv2d(\n 3, width // 2, kernel_size=3, stride=2, padding=1, bias=False\n )\n self.bn1 = nn.BatchNorm2d(width // 2)\n self.conv2 = nn.Conv2d(\n width // 2, width // 2, kernel_size=3, padding=1, bias=False\n )\n self.bn2 = nn.BatchNorm2d(width // 2)\n self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)\n self.bn3 = nn.BatchNorm2d(width)\n self.avgpool = nn.AvgPool2d(2)\n self.relu = nn.ReLU(inplace=True)\n\n # residual layers\n self._inplanes = width # this is a *mutable* variable used during construction\n self.layer1 = self._make_layer(width, layers[0])\n self.layer2 = self._make_layer(width * 2, layers[1], stride=2)\n self.layer3 = self._make_layer(width * 4, layers[2], stride=2)\n self.layer4 = self._make_layer(width * 8, layers[3], stride=2)\n\n embed_dim = width * 32 # the ResNet feature dimension\n self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)\n\n self.init_parameters()\n\n def _make_layer(self, planes, blocks, stride=1):\n layers = [Bottleneck(self._inplanes, planes, stride)]\n\n self._inplanes = planes * Bottleneck.expansion\n for _ in range(1, blocks):\n layers.append(Bottleneck(self._inplanes, planes))\n\n return nn.Sequential(*layers)\n\n def init_parameters(self):\n if self.attnpool is not None:\n std = self.attnpool.c_proj.in_features**-0.5\n nn.init.normal_(self.attnpool.q_proj.weight, std=std)\n nn.init.normal_(self.attnpool.k_proj.weight, std=std)\n nn.init.normal_(self.attnpool.v_proj.weight, std=std)\n nn.init.normal_(self.attnpool.c_proj.weight, std=std)\n\n for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:\n for name, param in resnet_block.named_parameters():\n if name.endswith(\"bn3.weight\"):\n nn.init.zeros_(param)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self)\n\n def stem(self, x):\n for conv, bn in [\n (self.conv1, self.bn1),\n (self.conv2, self.bn2),\n (self.conv3, self.bn3),\n ]:\n x = self.relu(bn(conv(x)))\n x = self.avgpool(x)\n return x\n\n def forward(self, x):\n x = self.stem(x)\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n\n\nclass QuickGELU(nn.Module):\n # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.LayerNorm","uri":"program://CREMA/class/lavis.models.clip_models.model.LayerNorm#L247-L253","kind":"class","name":"LayerNorm","path":"lavis/models/clip_models/model.py","language":"python","start_line":247,"end_line":253,"context_start_line":227,"context_end_line":273,"code":" for conv, bn in [\n (self.conv1, self.bn1),\n (self.conv2, self.bn2),\n (self.conv3, self.bn3),\n ]:\n x = self.relu(bn(conv(x)))\n x = self.avgpool(x)\n return x\n\n def forward(self, x):\n x = self.stem(x)\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n\n\nclass QuickGELU(nn.Module):\n # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.QuickGELU","uri":"program://CREMA/class/lavis.models.clip_models.model.QuickGELU#L256-L259","kind":"class","name":"QuickGELU","path":"lavis/models/clip_models/model.py","language":"python","start_line":256,"end_line":259,"context_start_line":236,"context_end_line":279,"code":" def forward(self, x):\n x = self.stem(x)\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n\n\nclass QuickGELU(nn.Module):\n # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),\n ]\n )\n )\n self.ln_2 = LayerNorm(d_model)\n\n def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.ResidualAttentionBlock","uri":"program://CREMA/class/lavis.models.clip_models.model.ResidualAttentionBlock#L262-L285","kind":"class","name":"ResidualAttentionBlock","path":"lavis/models/clip_models/model.py","language":"python","start_line":262,"end_line":285,"context_start_line":242,"context_end_line":305,"code":" x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n\n\nclass QuickGELU(nn.Module):\n # NOTE This is slower than nn.GELU or nn.SiLU and uses more GPU memory\n def forward(self, x: torch.Tensor):\n return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),\n ]\n )\n )\n self.ln_2 = LayerNorm(d_model)\n\n def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(\n self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU\n ):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.ModuleList(\n [\n ResidualAttentionBlock(width, heads, act_layer=act_layer)\n for _ in range(layers)\n ]\n )\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n for r in self.resblocks:\n x = r(x, attn_mask=attn_mask)\n return x","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.Transformer","uri":"program://CREMA/class/lavis.models.clip_models.model.Transformer#L288-L305","kind":"class","name":"Transformer","path":"lavis/models/clip_models/model.py","language":"python","start_line":288,"end_line":305,"context_start_line":268,"context_end_line":325,"code":" self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),\n ]\n )\n )\n self.ln_2 = LayerNorm(d_model)\n\n def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(\n self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU\n ):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.ModuleList(\n [\n ResidualAttentionBlock(width, heads, act_layer=act_layer)\n for _ in range(layers)\n ]\n )\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n for r in self.resblocks:\n x = r(x, attn_mask=attn_mask)\n return x\n\n\nclass VisualTransformer(nn.Module):\n def __init__(\n self,\n image_size: int,\n patch_size: int,\n width: int,\n layers: int,\n heads: int,\n output_dim: int,\n act_layer: Callable = nn.GELU,\n ):\n super().__init__()\n self.image_size = image_size\n self.output_dim = output_dim\n self.conv1 = nn.Conv2d(\n in_channels=3,\n out_channels=width,\n kernel_size=patch_size,","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.VisualTransformer","uri":"program://CREMA/class/lavis.models.clip_models.model.VisualTransformer#L308-L375","kind":"class","name":"VisualTransformer","path":"lavis/models/clip_models/model.py","language":"python","start_line":308,"end_line":375,"context_start_line":288,"context_end_line":395,"code":"class Transformer(nn.Module):\n def __init__(\n self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU\n ):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.ModuleList(\n [\n ResidualAttentionBlock(width, heads, act_layer=act_layer)\n for _ in range(layers)\n ]\n )\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n for r in self.resblocks:\n x = r(x, attn_mask=attn_mask)\n return x\n\n\nclass VisualTransformer(nn.Module):\n def __init__(\n self,\n image_size: int,\n patch_size: int,\n width: int,\n layers: int,\n heads: int,\n output_dim: int,\n act_layer: Callable = nn.GELU,\n ):\n super().__init__()\n self.image_size = image_size\n self.output_dim = output_dim\n self.conv1 = nn.Conv2d(\n in_channels=3,\n out_channels=width,\n kernel_size=patch_size,\n stride=patch_size,\n bias=False,\n )\n\n scale = width**-0.5\n self.class_embedding = nn.Parameter(scale * torch.randn(width))\n self.positional_embedding = nn.Parameter(\n scale * torch.randn((image_size // patch_size) ** 2 + 1, width)\n )\n self.ln_pre = LayerNorm(width)\n\n self.transformer = Transformer(width, layers, heads, act_layer=act_layer)\n\n self.ln_post = LayerNorm(width)\n self.proj = nn.Parameter(scale * torch.randn(width, output_dim))\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n\n def forward(self, x: torch.Tensor):\n x = self.conv1(x) # shape = [*, width, grid, grid]\n x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]\n x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]\n x = torch.cat(\n [\n self.class_embedding.to(x.dtype)\n + torch.zeros(\n x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device\n ),\n x,\n ],\n dim=1,\n ) # shape = [*, grid ** 2 + 1, width]\n x = x + self.positional_embedding.to(x.dtype)\n x = self.ln_pre(x)\n\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n\n x = self.ln_post(x[:, 0, :])\n\n if self.proj is not None:\n x = x @ self.proj\n\n return x\n\n\n@dataclass\nclass CLIPVisionCfg:\n layers: Union[Tuple[int, int, int, int], int] = 12\n width: int = 768\n patch_size: int = 16\n image_size: Union[Tuple[int, int], int] = 224\n timm_model_name: str = (\n None # a valid model name overrides layers, width, patch_size\n )\n timm_model_pretrained: bool = (\n False # use (imagenet) pretrained weights for named model\n )\n timm_pool: str = (\n \"avg\" # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')\n )\n timm_proj: str = (\n \"linear\" # linear projection for timm model output ('linear', 'mlp', '')\n )","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.CLIPVisionCfg","uri":"program://CREMA/class/lavis.models.clip_models.model.CLIPVisionCfg#L379-L395","kind":"class","name":"CLIPVisionCfg","path":"lavis/models/clip_models/model.py","language":"python","start_line":379,"end_line":395,"context_start_line":359,"context_end_line":415,"code":" x,\n ],\n dim=1,\n ) # shape = [*, grid ** 2 + 1, width]\n x = x + self.positional_embedding.to(x.dtype)\n x = self.ln_pre(x)\n\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)\n x = x.permute(1, 0, 2) # LND -> NLD\n\n x = self.ln_post(x[:, 0, :])\n\n if self.proj is not None:\n x = x @ self.proj\n\n return x\n\n\n@dataclass\nclass CLIPVisionCfg:\n layers: Union[Tuple[int, int, int, int], int] = 12\n width: int = 768\n patch_size: int = 16\n image_size: Union[Tuple[int, int], int] = 224\n timm_model_name: str = (\n None # a valid model name overrides layers, width, patch_size\n )\n timm_model_pretrained: bool = (\n False # use (imagenet) pretrained weights for named model\n )\n timm_pool: str = (\n \"avg\" # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')\n )\n timm_proj: str = (\n \"linear\" # linear projection for timm model output ('linear', 'mlp', '')\n )\n\n\n@dataclass\nclass CLIPTextCfg:\n context_length: int\n vocab_size: int\n width: int\n heads: int\n layers: int\n\n\n@registry.register_model(\"clip\")\n@registry.register_model(\"clip_feature_extractor\")\nclass CLIP(BaseModel):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ViT-B-32\": \"configs/models/clip_vit_base32.yaml\",\n \"ViT-B-16\": \"configs/models/clip_vit_base16.yaml\",\n \"ViT-L-14\": \"configs/models/clip_vit_large14.yaml\",\n \"ViT-L-14-336\": \"configs/models/clip_vit_large14_336.yaml\",\n \"RN50\": \"configs/models/clip_resnet50.yaml\",","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.CLIPTextCfg","uri":"program://CREMA/class/lavis.models.clip_models.model.CLIPTextCfg#L399-L404","kind":"class","name":"CLIPTextCfg","path":"lavis/models/clip_models/model.py","language":"python","start_line":399,"end_line":404,"context_start_line":379,"context_end_line":424,"code":"class CLIPVisionCfg:\n layers: Union[Tuple[int, int, int, int], int] = 12\n width: int = 768\n patch_size: int = 16\n image_size: Union[Tuple[int, int], int] = 224\n timm_model_name: str = (\n None # a valid model name overrides layers, width, patch_size\n )\n timm_model_pretrained: bool = (\n False # use (imagenet) pretrained weights for named model\n )\n timm_pool: str = (\n \"avg\" # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')\n )\n timm_proj: str = (\n \"linear\" # linear projection for timm model output ('linear', 'mlp', '')\n )\n\n\n@dataclass\nclass CLIPTextCfg:\n context_length: int\n vocab_size: int\n width: int\n heads: int\n layers: int\n\n\n@registry.register_model(\"clip\")\n@registry.register_model(\"clip_feature_extractor\")\nclass CLIP(BaseModel):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ViT-B-32\": \"configs/models/clip_vit_base32.yaml\",\n \"ViT-B-16\": \"configs/models/clip_vit_base16.yaml\",\n \"ViT-L-14\": \"configs/models/clip_vit_large14.yaml\",\n \"ViT-L-14-336\": \"configs/models/clip_vit_large14_336.yaml\",\n \"RN50\": \"configs/models/clip_resnet50.yaml\",\n }\n\n def __init__(\n self,\n embed_dim: int,\n vision_cfg: CLIPVisionCfg,\n text_cfg: CLIPTextCfg,\n quick_gelu: bool = False,\n ):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.CLIP","uri":"program://CREMA/class/lavis.models.clip_models.model.CLIP#L409-L760","kind":"class","name":"CLIP","path":"lavis/models/clip_models/model.py","language":"python","start_line":409,"end_line":760,"context_start_line":389,"context_end_line":780,"code":" )\n timm_pool: str = (\n \"avg\" # feature pooling for timm model ('abs_attn', 'rot_attn', 'avg', '')\n )\n timm_proj: str = (\n \"linear\" # linear projection for timm model output ('linear', 'mlp', '')\n )\n\n\n@dataclass\nclass CLIPTextCfg:\n context_length: int\n vocab_size: int\n width: int\n heads: int\n layers: int\n\n\n@registry.register_model(\"clip\")\n@registry.register_model(\"clip_feature_extractor\")\nclass CLIP(BaseModel):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ViT-B-32\": \"configs/models/clip_vit_base32.yaml\",\n \"ViT-B-16\": \"configs/models/clip_vit_base16.yaml\",\n \"ViT-L-14\": \"configs/models/clip_vit_large14.yaml\",\n \"ViT-L-14-336\": \"configs/models/clip_vit_large14_336.yaml\",\n \"RN50\": \"configs/models/clip_resnet50.yaml\",\n }\n\n def __init__(\n self,\n embed_dim: int,\n vision_cfg: CLIPVisionCfg,\n text_cfg: CLIPTextCfg,\n quick_gelu: bool = False,\n ):\n from .tokenizer import tokenize\n\n super().__init__()\n\n self.tokenizer = tokenize\n self._loss = None\n\n if isinstance(vision_cfg, dict):\n vision_cfg = CLIPVisionCfg(**vision_cfg)\n if isinstance(text_cfg, dict):\n text_cfg = CLIPTextCfg(**text_cfg)\n\n self.context_length = text_cfg.context_length\n\n # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more\n # memory efficient in recent PyTorch releases (>= 1.10).\n # NOTE: timm models always use native GELU regardless of quick_gelu flag.\n act_layer = QuickGELU if quick_gelu else nn.GELU\n\n if vision_cfg.timm_model_name:\n self.visual = TimmModel(\n vision_cfg.timm_model_name,\n pretrained=vision_cfg.timm_model_pretrained,\n pool=vision_cfg.timm_pool,\n proj=vision_cfg.timm_proj,\n embed_dim=embed_dim,\n image_size=vision_cfg.image_size,\n )\n act_layer = (\n nn.GELU\n ) # so that text transformer doesn't use QuickGELU w/ timm models\n elif isinstance(vision_cfg.layers, (tuple, list)):\n vision_heads = vision_cfg.width * 32 // 64\n self.visual = ModifiedResNet(\n layers=vision_cfg.layers,\n output_dim=embed_dim,\n heads=vision_heads,\n image_size=vision_cfg.image_size,\n width=vision_cfg.width,\n )\n else:\n vision_heads = vision_cfg.width // 64\n self.visual = VisualTransformer(\n image_size=vision_cfg.image_size,\n patch_size=vision_cfg.patch_size,\n width=vision_cfg.width,\n layers=vision_cfg.layers,\n heads=vision_heads,\n output_dim=embed_dim,\n act_layer=act_layer,\n )\n\n self.transformer = Transformer(\n width=text_cfg.width,\n layers=text_cfg.layers,\n heads=text_cfg.heads,\n act_layer=act_layer,\n )\n\n self.vocab_size = text_cfg.vocab_size\n self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)\n self.positional_embedding = nn.Parameter(\n torch.empty(self.context_length, text_cfg.width)\n )\n self.ln_final = LayerNorm(text_cfg.width)\n\n self.text_projection = nn.Parameter(torch.empty(text_cfg.width, embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n self.register_buffer(\"attn_mask\", self.build_attention_mask(), persistent=False)\n\n self.prompt_templates = openai_imagenet_template\n self.classifier = None\n\n self.init_parameters()\n\n @property\n def loss(self):\n if self._loss is None:\n from lavis.models.clip_models.loss import ClipLoss\n from torch import distributed as dist\n\n self._loss = ClipLoss(\n world_size=dist.get_world_size(),\n rank=dist.get_rank(),\n local_loss=False,\n gather_with_grad=False,\n use_horovod=False,\n )\n\n return self._loss\n\n def init_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n nn.init.constant_(self.logit_scale, np.log(1 / 0.07))\n\n if hasattr(self.visual, \"init_parameters\"):\n self.visual.init_parameters()\n\n proj_std = (self.transformer.width**-0.5) * (\n (2 * self.transformer.layers) ** -0.5\n )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):\n def forward(self, samples):\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n loss = self.loss(image_features, text_features, self.logit_scale.exp())\n\n # return image_features, text_features, self.logit_scale.exp()\n # return {\"loss\": loss}\n return ClipOutput(\n intermediate_output=ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n ),\n loss=loss,\n logit_scale_exp=self.logit_scale.exp(),\n )\n\n def extract_features(self, samples):\n \"\"\"\n Extract features from the model for samples.\n\n Keys allowed are \"image\" and \"text_input\" in samples.\n If either key is missing, the corresponding features are not extracted.\n\n Args:\n samples: dict of samples to extract features from.\n\n Returns:\n ClipOutputFeatures object with features for the samples.\n \"\"\"\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n return ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n )\n\n def predict(self, samples):\n image = samples[\"image\"]\n targets = samples[\"label\"]\n\n image_features = self.encode_image(image)\n image_features = F.normalize(image_features, dim=-1)\n\n logits = 100.0 * image_features @ self.classifier\n\n return {\"predictions\": logits, \"targets\": targets}\n\n def before_evaluation(self, dataset, task_type, **kwargs):\n if task_type == MultimodalClassificationTask:\n self.classifier = self.zero_shot_classifier(\n classnames=dataset.classnames,\n templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in classnames:\n texts = [\n template(classname) for template in templates\n ] # format with class\n texts = self.tokenizer(texts).to(self.device) # tokenize\n\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).to(self.device)\n return zeroshot_weights\n\n @classmethod\n def default_config_path(cls, model_type=\"base\"):\n model_type = \"ViT-B-32\" if model_type == \"base\" else model_type\n\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}. \\n Available types: {}\".format(\n model_type, cls.PRETRAINED_MODEL_CONFIG_DICT.keys()\n )\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n @classmethod\n def from_config(cls, cfg=None):\n model_name = cfg.model_type\n pretrained = cfg.pretrained\n\n precision = cfg.get(\"precision\", \"fp32\")\n\n return create_model(\n model_name=model_name, pretrained=pretrained, precision=precision\n )\n\n def zero_shot_predict(self, image_path, categories):\n assert isinstance(\n categories, list\n ), f\"categories must be a list, got {type(categories)}.\"\n assert os.path.exists(image_path), f\"File {image_path} does not exist.\"\n\n from lavis.processors.clip_processors import ClipImageEvalProcessor\n from PIL import Image\n\n image_preprocess = ClipImageEvalProcessor()\n image = image_preprocess(Image.open(image_path)).unsqueeze(0)\n\n text = self.tokenizer(categories)\n\n with torch.no_grad():\n image_features = self.encode_image(image)\n text_features = self.encode_text(text)\n image_features /= image_features.norm(dim=-1, keepdim=True)\n text_features /= text_features.norm(dim=-1, keepdim=True)\n\n text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)\n\n print(\"Label probs:\", text_probs) # prints: [[1., 0., 0.]]\n\n def compute_sim_matrix(self, data_loader, **kwargs):\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_features = []\n\n for i in range(0, num_text, text_bs):\n\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(text).to(self.device)\n\n text_feat = self.encode_text(text_input)\n text_feat = F.normalize(text_feat, dim=-1)\n\n text_features.append(text_feat)\n\n text_features = torch.cat(text_features, dim=0)\n\n image_features = []\n for samples in data_loader:\n image = samples[\"image\"]\n\n image = image.to(self.device)\n image_feat = self.encode_image(image)\n image_feat = F.normalize(image_feat, dim=-1)\n\n image_features.append(image_feat)\n\n image_features = torch.cat(image_features, dim=0)\n\n sims_matrix_i2t = image_features @ text_features.t()\n sims_matrix_t2i = sims_matrix_i2t.t()\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return sims_matrix_i2t.cpu().numpy(), sims_matrix_t2i.cpu().numpy()\n\n\ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n if isinstance(l, nn.MultiheadAttention):\n for attr in [\n *[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]],\n \"in_proj_bias\",\n \"bias_k\",\n \"bias_v\",\n ]:\n tensor = getattr(l, attr)\n if tensor is not None:","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.convert_weights_to_fp16","uri":"program://CREMA/function/lavis.models.clip_models.model.convert_weights_to_fp16#L763-L789","kind":"function","name":"convert_weights_to_fp16","path":"lavis/models/clip_models/model.py","language":"python","start_line":763,"end_line":789,"context_start_line":743,"context_end_line":809,"code":" image = samples[\"image\"]\n\n image = image.to(self.device)\n image_feat = self.encode_image(image)\n image_feat = F.normalize(image_feat, dim=-1)\n\n image_features.append(image_feat)\n\n image_features = torch.cat(image_features, dim=0)\n\n sims_matrix_i2t = image_features @ text_features.t()\n sims_matrix_t2i = sims_matrix_i2t.t()\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return sims_matrix_i2t.cpu().numpy(), sims_matrix_t2i.cpu().numpy()\n\n\ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n if isinstance(l, nn.MultiheadAttention):\n for attr in [\n *[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]],\n \"in_proj_bias\",\n \"bias_k\",\n \"bias_v\",\n ]:\n tensor = getattr(l, attr)\n if tensor is not None:\n tensor.data = tensor.data.half()\n\n for name in [\"text_projection\", \"proj\"]:\n if hasattr(l, name):\n attr = getattr(l, name)\n if attr is not None:\n attr.data = attr.data.half()\n\n model.apply(_convert_weights_to_fp16)\n\n\ndef build_model_from_openai_state_dict(state_dict: dict):\n vit = \"visual.proj\" in state_dict\n\n if vit:\n vision_width = state_dict[\"visual.conv1.weight\"].shape[0]\n vision_layers = len(\n [\n k\n for k in state_dict.keys()\n if k.startswith(\"visual.\") and k.endswith(\".attn.in_proj_weight\")\n ]\n )\n vision_patch_size = state_dict[\"visual.conv1.weight\"].shape[-1]\n grid_size = round(\n (state_dict[\"visual.positional_embedding\"].shape[0] - 1) ** 0.5\n )\n image_size = vision_patch_size * grid_size\n else:","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.build_model_from_openai_state_dict","uri":"program://CREMA/function/lavis.models.clip_models.model.build_model_from_openai_state_dict#L792-L870","kind":"function","name":"build_model_from_openai_state_dict","path":"lavis/models/clip_models/model.py","language":"python","start_line":792,"end_line":870,"context_start_line":772,"context_end_line":890,"code":" if isinstance(l, nn.MultiheadAttention):\n for attr in [\n *[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]],\n \"in_proj_bias\",\n \"bias_k\",\n \"bias_v\",\n ]:\n tensor = getattr(l, attr)\n if tensor is not None:\n tensor.data = tensor.data.half()\n\n for name in [\"text_projection\", \"proj\"]:\n if hasattr(l, name):\n attr = getattr(l, name)\n if attr is not None:\n attr.data = attr.data.half()\n\n model.apply(_convert_weights_to_fp16)\n\n\ndef build_model_from_openai_state_dict(state_dict: dict):\n vit = \"visual.proj\" in state_dict\n\n if vit:\n vision_width = state_dict[\"visual.conv1.weight\"].shape[0]\n vision_layers = len(\n [\n k\n for k in state_dict.keys()\n if k.startswith(\"visual.\") and k.endswith(\".attn.in_proj_weight\")\n ]\n )\n vision_patch_size = state_dict[\"visual.conv1.weight\"].shape[-1]\n grid_size = round(\n (state_dict[\"visual.positional_embedding\"].shape[0] - 1) ** 0.5\n )\n image_size = vision_patch_size * grid_size\n else:\n counts: list = [\n len(\n set(\n k.split(\".\")[2]\n for k in state_dict\n if k.startswith(f\"visual.layer{b}\")\n )\n )\n for b in [1, 2, 3, 4]\n ]\n vision_layers = tuple(counts)\n vision_width = state_dict[\"visual.layer1.0.conv1.weight\"].shape[0]\n output_width = round(\n (state_dict[\"visual.attnpool.positional_embedding\"].shape[0] - 1) ** 0.5\n )\n vision_patch_size = None\n assert (\n output_width**2 + 1\n == state_dict[\"visual.attnpool.positional_embedding\"].shape[0]\n )\n image_size = output_width * 32\n\n embed_dim = state_dict[\"text_projection\"].shape[1]\n context_length = state_dict[\"positional_embedding\"].shape[0]\n vocab_size = state_dict[\"token_embedding.weight\"].shape[0]\n transformer_width = state_dict[\"ln_final.weight\"].shape[0]\n transformer_heads = transformer_width // 64\n transformer_layers = len(\n set(\n k.split(\".\")[2]\n for k in state_dict\n if k.startswith(f\"transformer.resblocks\")\n )\n )\n\n vision_cfg = CLIPVisionCfg(\n layers=vision_layers,\n width=vision_width,\n patch_size=vision_patch_size,\n image_size=image_size,\n )\n text_cfg = CLIPTextCfg(\n context_length=context_length,\n vocab_size=vocab_size,\n width=transformer_width,\n heads=transformer_heads,\n layers=transformer_layers,\n )\n model = CLIP(\n embed_dim,\n vision_cfg=vision_cfg,\n text_cfg=text_cfg,\n quick_gelu=True, # OpenAI models were trained with QuickGELU\n )\n\n for key in [\"input_resolution\", \"context_length\", \"vocab_size\"]:\n state_dict.pop(key, None)\n\n convert_weights_to_fp16(model)\n model.load_state_dict(state_dict)\n return model.eval()\n\n\ndef trace_model(model, batch_size=256, device=torch.device(\"cpu\")):\n model.eval()\n image_size = model.visual.image_size\n example_images = torch.ones((batch_size, 3, image_size, image_size), device=device)\n example_text = torch.zeros(\n (batch_size, model.context_length), dtype=torch.int, device=device\n )\n model = torch.jit.trace_module(\n model,\n inputs=dict(\n forward=(example_images, example_text),\n encode_text=(example_text,),\n encode_image=(example_images,),\n ),\n )\n model.visual.image_size = image_size\n return\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.trace_model","uri":"program://CREMA/function/lavis.models.clip_models.model.trace_model#L873-L889","kind":"function","name":"trace_model","path":"lavis/models/clip_models/model.py","language":"python","start_line":873,"end_line":889,"context_start_line":853,"context_end_line":909,"code":" vocab_size=vocab_size,\n width=transformer_width,\n heads=transformer_heads,\n layers=transformer_layers,\n )\n model = CLIP(\n embed_dim,\n vision_cfg=vision_cfg,\n text_cfg=text_cfg,\n quick_gelu=True, # OpenAI models were trained with QuickGELU\n )\n\n for key in [\"input_resolution\", \"context_length\", \"vocab_size\"]:\n state_dict.pop(key, None)\n\n convert_weights_to_fp16(model)\n model.load_state_dict(state_dict)\n return model.eval()\n\n\ndef trace_model(model, batch_size=256, device=torch.device(\"cpu\")):\n model.eval()\n image_size = model.visual.image_size\n example_images = torch.ones((batch_size, 3, image_size, image_size), device=device)\n example_text = torch.zeros(\n (batch_size, model.context_length), dtype=torch.int, device=device\n )\n model = torch.jit.trace_module(\n model,\n inputs=dict(\n forward=(example_images, example_text),\n encode_text=(example_text,),\n encode_image=(example_images,),\n ),\n )\n model.visual.image_size = image_size\n return\n\n\ndef _natural_key(string_):\n return [int(s) if s.isdigit() else s for s in re.split(r\"(\\d+)\", string_.lower())]\n\n\ndef _rescan_model_configs():\n global _MODEL_CONFIGS\n\n config_ext = (\".json\",)\n config_files = []\n for config_path in _MODEL_CONFIG_PATHS:\n if config_path.is_file() and config_path.suffix in config_ext:\n config_files.append(config_path)\n elif config_path.is_dir():\n for ext in config_ext:\n config_files.extend(config_path.glob(f\"*{ext}\"))\n\n for cf in config_files:\n with open(cf, \"r\") as f:","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model._natural_key","uri":"program://CREMA/function/lavis.models.clip_models.model._natural_key#L892-L893","kind":"function","name":"_natural_key","path":"lavis/models/clip_models/model.py","language":"python","start_line":892,"end_line":893,"context_start_line":872,"context_end_line":913,"code":"\ndef trace_model(model, batch_size=256, device=torch.device(\"cpu\")):\n model.eval()\n image_size = model.visual.image_size\n example_images = torch.ones((batch_size, 3, image_size, image_size), device=device)\n example_text = torch.zeros(\n (batch_size, model.context_length), dtype=torch.int, device=device\n )\n model = torch.jit.trace_module(\n model,\n inputs=dict(\n forward=(example_images, example_text),\n encode_text=(example_text,),\n encode_image=(example_images,),\n ),\n )\n model.visual.image_size = image_size\n return\n\n\ndef _natural_key(string_):\n return [int(s) if s.isdigit() else s for s in re.split(r\"(\\d+)\", string_.lower())]\n\n\ndef _rescan_model_configs():\n global _MODEL_CONFIGS\n\n config_ext = (\".json\",)\n config_files = []\n for config_path in _MODEL_CONFIG_PATHS:\n if config_path.is_file() and config_path.suffix in config_ext:\n config_files.append(config_path)\n elif config_path.is_dir():\n for ext in config_ext:\n config_files.extend(config_path.glob(f\"*{ext}\"))\n\n for cf in config_files:\n with open(cf, \"r\") as f:\n model_cfg = json.load(f)\n if all(a in model_cfg for a in (\"embed_dim\", \"vision_cfg\", \"text_cfg\")):\n _MODEL_CONFIGS[cf.stem] = model_cfg\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model._rescan_model_configs","uri":"program://CREMA/function/lavis.models.clip_models.model._rescan_model_configs#L896-L917","kind":"function","name":"_rescan_model_configs","path":"lavis/models/clip_models/model.py","language":"python","start_line":896,"end_line":917,"context_start_line":876,"context_end_line":937,"code":" example_images = torch.ones((batch_size, 3, image_size, image_size), device=device)\n example_text = torch.zeros(\n (batch_size, model.context_length), dtype=torch.int, device=device\n )\n model = torch.jit.trace_module(\n model,\n inputs=dict(\n forward=(example_images, example_text),\n encode_text=(example_text,),\n encode_image=(example_images,),\n ),\n )\n model.visual.image_size = image_size\n return\n\n\ndef _natural_key(string_):\n return [int(s) if s.isdigit() else s for s in re.split(r\"(\\d+)\", string_.lower())]\n\n\ndef _rescan_model_configs():\n global _MODEL_CONFIGS\n\n config_ext = (\".json\",)\n config_files = []\n for config_path in _MODEL_CONFIG_PATHS:\n if config_path.is_file() and config_path.suffix in config_ext:\n config_files.append(config_path)\n elif config_path.is_dir():\n for ext in config_ext:\n config_files.extend(config_path.glob(f\"*{ext}\"))\n\n for cf in config_files:\n with open(cf, \"r\") as f:\n model_cfg = json.load(f)\n if all(a in model_cfg for a in (\"embed_dim\", \"vision_cfg\", \"text_cfg\")):\n _MODEL_CONFIGS[cf.stem] = model_cfg\n\n _MODEL_CONFIGS = {\n k: v\n for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))\n }\n\n\n_rescan_model_configs() # initial populate of model config registry\n\n\ndef load_state_dict(checkpoint_path: str, map_location=\"cpu\"):\n checkpoint = torch.load(checkpoint_path, map_location=map_location)\n if isinstance(checkpoint, dict) and \"state_dict\" in checkpoint:\n state_dict = checkpoint[\"state_dict\"]\n else:\n state_dict = checkpoint\n if next(iter(state_dict.items()))[0].startswith(\"module\"):\n state_dict = {k[7:]: v for k, v in state_dict.items()}\n return state_dict\n\n\ndef create_model(\n model_name: str,\n pretrained: str = \"\",\n precision: str = \"fp32\",","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.load_state_dict","uri":"program://CREMA/function/lavis.models.clip_models.model.load_state_dict#L923-L931","kind":"function","name":"load_state_dict","path":"lavis/models/clip_models/model.py","language":"python","start_line":923,"end_line":931,"context_start_line":903,"context_end_line":951,"code":" config_files.append(config_path)\n elif config_path.is_dir():\n for ext in config_ext:\n config_files.extend(config_path.glob(f\"*{ext}\"))\n\n for cf in config_files:\n with open(cf, \"r\") as f:\n model_cfg = json.load(f)\n if all(a in model_cfg for a in (\"embed_dim\", \"vision_cfg\", \"text_cfg\")):\n _MODEL_CONFIGS[cf.stem] = model_cfg\n\n _MODEL_CONFIGS = {\n k: v\n for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))\n }\n\n\n_rescan_model_configs() # initial populate of model config registry\n\n\ndef load_state_dict(checkpoint_path: str, map_location=\"cpu\"):\n checkpoint = torch.load(checkpoint_path, map_location=map_location)\n if isinstance(checkpoint, dict) and \"state_dict\" in checkpoint:\n state_dict = checkpoint[\"state_dict\"]\n else:\n state_dict = checkpoint\n if next(iter(state_dict.items()))[0].startswith(\"module\"):\n state_dict = {k[7:]: v for k, v in state_dict.items()}\n return state_dict\n\n\ndef create_model(\n model_name: str,\n pretrained: str = \"\",\n precision: str = \"fp32\",\n device: torch.device = torch.device(\"cpu\"),\n jit: bool = False,\n force_quick_gelu: bool = False,\n pretrained_image: bool = False,\n):\n model_name = model_name.replace(\n \"/\", \"-\"\n ) # for callers using old naming with / in ViT names\n\n if pretrained.lower() == \"openai\":\n logging.info(f\"Loading pretrained {model_name} from OpenAI.\")\n model = load_openai_model(model_name, device=device, jit=jit)\n # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372\n if precision == \"amp\" or precision == \"fp32\":","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.create_model","uri":"program://CREMA/function/lavis.models.clip_models.model.create_model#L934-L1006","kind":"function","name":"create_model","path":"lavis/models/clip_models/model.py","language":"python","start_line":934,"end_line":1006,"context_start_line":914,"context_end_line":1026,"code":" _MODEL_CONFIGS = {\n k: v\n for k, v in sorted(_MODEL_CONFIGS.items(), key=lambda x: _natural_key(x[0]))\n }\n\n\n_rescan_model_configs() # initial populate of model config registry\n\n\ndef load_state_dict(checkpoint_path: str, map_location=\"cpu\"):\n checkpoint = torch.load(checkpoint_path, map_location=map_location)\n if isinstance(checkpoint, dict) and \"state_dict\" in checkpoint:\n state_dict = checkpoint[\"state_dict\"]\n else:\n state_dict = checkpoint\n if next(iter(state_dict.items()))[0].startswith(\"module\"):\n state_dict = {k[7:]: v for k, v in state_dict.items()}\n return state_dict\n\n\ndef create_model(\n model_name: str,\n pretrained: str = \"\",\n precision: str = \"fp32\",\n device: torch.device = torch.device(\"cpu\"),\n jit: bool = False,\n force_quick_gelu: bool = False,\n pretrained_image: bool = False,\n):\n model_name = model_name.replace(\n \"/\", \"-\"\n ) # for callers using old naming with / in ViT names\n\n if pretrained.lower() == \"openai\":\n logging.info(f\"Loading pretrained {model_name} from OpenAI.\")\n model = load_openai_model(model_name, device=device, jit=jit)\n # See https://discuss.pytorch.org/t/valueerror-attemting-to-unscale-fp16-gradients/81372\n if precision == \"amp\" or precision == \"fp32\":\n model = model.float()\n else:\n logging.info(f\"No pretrained weights loaded for {model_name} model.\")\n if model_name in _MODEL_CONFIGS:\n logging.info(f\"Loading {model_name} model config.\")\n model_cfg = deepcopy(_MODEL_CONFIGS[model_name])\n else:\n logging.error(\n f\"Model config for {model_name} not found; available models {list_models()}.\"\n )\n raise RuntimeError(f\"Model config for {model_name} not found.\")\n\n if force_quick_gelu:\n # override for use of QuickGELU on non-OpenAI transformer models\n model_cfg[\"quick_gelu\"] = True\n\n if pretrained_image:\n if \"timm_model_name\" in model_cfg.get(\"vision_cfg\", {}):\n # pretrained weight loading for timm models set via vision_cfg\n model_cfg[\"vision_cfg\"][\"timm_model_pretrained\"] = True\n else:\n assert (\n False\n ), \"pretrained image towers currently only supported for timm models\"\n\n model = CLIP(**model_cfg)\n\n if pretrained:\n checkpoint_path = \"\"\n url = get_pretrained_url(model_name, pretrained)\n if url:\n checkpoint_path = download_pretrained(url)\n elif os.path.exists(pretrained):\n checkpoint_path = pretrained\n\n if checkpoint_path:\n logging.info(f\"Loading pretrained {model_name} weights ({pretrained}).\")\n model.load_state_dict(load_state_dict(checkpoint_path))\n else:\n logging.warning(\n f\"Pretrained weights ({pretrained}) not found for model {model_name}.\"\n )\n raise RuntimeError(\n f\"Pretrained weights ({pretrained}) not found for model {model_name}.\"\n )\n\n model.to(device=device)\n if precision == \"fp16\":\n assert device.type != \"cpu\"\n convert_weights_to_fp16(model)\n\n if jit:\n model = torch.jit.script(model)\n\n return model\n\n\ndef create_model_and_transforms(\n model_name: str,\n pretrained: str = \"\",\n precision: str = \"fp32\",\n device: torch.device = torch.device(\"cpu\"),\n jit: bool = False,\n force_quick_gelu: bool = False,\n pretrained_image: bool = False,\n):\n model = create_model(\n model_name,\n pretrained,\n precision,\n device,\n jit,\n force_quick_gelu=force_quick_gelu,\n pretrained_image=pretrained_image,\n )","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.create_model_and_transforms","uri":"program://CREMA/function/lavis.models.clip_models.model.create_model_and_transforms#L1009-L1029","kind":"function","name":"create_model_and_transforms","path":"lavis/models/clip_models/model.py","language":"python","start_line":1009,"end_line":1029,"context_start_line":989,"context_end_line":1049,"code":" model.load_state_dict(load_state_dict(checkpoint_path))\n else:\n logging.warning(\n f\"Pretrained weights ({pretrained}) not found for model {model_name}.\"\n )\n raise RuntimeError(\n f\"Pretrained weights ({pretrained}) not found for model {model_name}.\"\n )\n\n model.to(device=device)\n if precision == \"fp16\":\n assert device.type != \"cpu\"\n convert_weights_to_fp16(model)\n\n if jit:\n model = torch.jit.script(model)\n\n return model\n\n\ndef create_model_and_transforms(\n model_name: str,\n pretrained: str = \"\",\n precision: str = \"fp32\",\n device: torch.device = torch.device(\"cpu\"),\n jit: bool = False,\n force_quick_gelu: bool = False,\n pretrained_image: bool = False,\n):\n model = create_model(\n model_name,\n pretrained,\n precision,\n device,\n jit,\n force_quick_gelu=force_quick_gelu,\n pretrained_image=pretrained_image,\n )\n preprocess_train = image_transform(model.visual.image_size, is_train=True)\n preprocess_val = image_transform(model.visual.image_size, is_train=False)\n return model, preprocess_train, preprocess_val\n\n\ndef list_models():\n \"\"\"enumerate available model architectures based on config files\"\"\"\n return list(_MODEL_CONFIGS.keys())\n\n\ndef add_model_config(path):\n \"\"\"add model config path or file and update registry\"\"\"\n if not isinstance(path, Path):\n path = Path(path)\n _MODEL_CONFIG_PATHS.append(path)\n _rescan_model_configs()\n\n\ndef list_openai_models() -> List[str]:\n \"\"\"Returns the names of available CLIP models\"\"\"\n return list_pretrained_tag_models(\"openai\")\n\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.list_models","uri":"program://CREMA/function/lavis.models.clip_models.model.list_models#L1032-L1034","kind":"function","name":"list_models","path":"lavis/models/clip_models/model.py","language":"python","start_line":1032,"end_line":1034,"context_start_line":1012,"context_end_line":1054,"code":" precision: str = \"fp32\",\n device: torch.device = torch.device(\"cpu\"),\n jit: bool = False,\n force_quick_gelu: bool = False,\n pretrained_image: bool = False,\n):\n model = create_model(\n model_name,\n pretrained,\n precision,\n device,\n jit,\n force_quick_gelu=force_quick_gelu,\n pretrained_image=pretrained_image,\n )\n preprocess_train = image_transform(model.visual.image_size, is_train=True)\n preprocess_val = image_transform(model.visual.image_size, is_train=False)\n return model, preprocess_train, preprocess_val\n\n\ndef list_models():\n \"\"\"enumerate available model architectures based on config files\"\"\"\n return list(_MODEL_CONFIGS.keys())\n\n\ndef add_model_config(path):\n \"\"\"add model config path or file and update registry\"\"\"\n if not isinstance(path, Path):\n path = Path(path)\n _MODEL_CONFIG_PATHS.append(path)\n _rescan_model_configs()\n\n\ndef list_openai_models() -> List[str]:\n \"\"\"Returns the names of available CLIP models\"\"\"\n return list_pretrained_tag_models(\"openai\")\n\n\ndef load_openai_model(\n name: str,\n device: Union[str, torch.device] = \"cuda\" if torch.cuda.is_available() else \"cpu\",\n jit=True,\n):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.add_model_config","uri":"program://CREMA/function/lavis.models.clip_models.model.add_model_config#L1037-L1042","kind":"function","name":"add_model_config","path":"lavis/models/clip_models/model.py","language":"python","start_line":1037,"end_line":1042,"context_start_line":1017,"context_end_line":1062,"code":"):\n model = create_model(\n model_name,\n pretrained,\n precision,\n device,\n jit,\n force_quick_gelu=force_quick_gelu,\n pretrained_image=pretrained_image,\n )\n preprocess_train = image_transform(model.visual.image_size, is_train=True)\n preprocess_val = image_transform(model.visual.image_size, is_train=False)\n return model, preprocess_train, preprocess_val\n\n\ndef list_models():\n \"\"\"enumerate available model architectures based on config files\"\"\"\n return list(_MODEL_CONFIGS.keys())\n\n\ndef add_model_config(path):\n \"\"\"add model config path or file and update registry\"\"\"\n if not isinstance(path, Path):\n path = Path(path)\n _MODEL_CONFIG_PATHS.append(path)\n _rescan_model_configs()\n\n\ndef list_openai_models() -> List[str]:\n \"\"\"Returns the names of available CLIP models\"\"\"\n return list_pretrained_tag_models(\"openai\")\n\n\ndef load_openai_model(\n name: str,\n device: Union[str, torch.device] = \"cuda\" if torch.cuda.is_available() else \"cpu\",\n jit=True,\n):\n \"\"\"Load a CLIP model\n Parameters\n ----------\n name : str\n A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict\n device : Union[str, torch.device]\n The device to put the loaded model\n jit : bool","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.list_openai_models","uri":"program://CREMA/function/lavis.models.clip_models.model.list_openai_models#L1045-L1047","kind":"function","name":"list_openai_models","path":"lavis/models/clip_models/model.py","language":"python","start_line":1045,"end_line":1047,"context_start_line":1025,"context_end_line":1067,"code":" pretrained_image=pretrained_image,\n )\n preprocess_train = image_transform(model.visual.image_size, is_train=True)\n preprocess_val = image_transform(model.visual.image_size, is_train=False)\n return model, preprocess_train, preprocess_val\n\n\ndef list_models():\n \"\"\"enumerate available model architectures based on config files\"\"\"\n return list(_MODEL_CONFIGS.keys())\n\n\ndef add_model_config(path):\n \"\"\"add model config path or file and update registry\"\"\"\n if not isinstance(path, Path):\n path = Path(path)\n _MODEL_CONFIG_PATHS.append(path)\n _rescan_model_configs()\n\n\ndef list_openai_models() -> List[str]:\n \"\"\"Returns the names of available CLIP models\"\"\"\n return list_pretrained_tag_models(\"openai\")\n\n\ndef load_openai_model(\n name: str,\n device: Union[str, torch.device] = \"cuda\" if torch.cuda.is_available() else \"cpu\",\n jit=True,\n):\n \"\"\"Load a CLIP model\n Parameters\n ----------\n name : str\n A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict\n device : Union[str, torch.device]\n The device to put the loaded model\n jit : bool\n Whether to load the optimized JIT model (default) or more hackable non-JIT model.\n Returns\n -------\n model : torch.nn.Module\n The CLIP model","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.load_openai_model","uri":"program://CREMA/function/lavis.models.clip_models.model.load_openai_model#L1050-L1170","kind":"function","name":"load_openai_model","path":"lavis/models/clip_models/model.py","language":"python","start_line":1050,"end_line":1170,"context_start_line":1030,"context_end_line":1190,"code":"\n\ndef list_models():\n \"\"\"enumerate available model architectures based on config files\"\"\"\n return list(_MODEL_CONFIGS.keys())\n\n\ndef add_model_config(path):\n \"\"\"add model config path or file and update registry\"\"\"\n if not isinstance(path, Path):\n path = Path(path)\n _MODEL_CONFIG_PATHS.append(path)\n _rescan_model_configs()\n\n\ndef list_openai_models() -> List[str]:\n \"\"\"Returns the names of available CLIP models\"\"\"\n return list_pretrained_tag_models(\"openai\")\n\n\ndef load_openai_model(\n name: str,\n device: Union[str, torch.device] = \"cuda\" if torch.cuda.is_available() else \"cpu\",\n jit=True,\n):\n \"\"\"Load a CLIP model\n Parameters\n ----------\n name : str\n A model name listed by `clip.available_models()`, or the path to a model checkpoint containing the state_dict\n device : Union[str, torch.device]\n The device to put the loaded model\n jit : bool\n Whether to load the optimized JIT model (default) or more hackable non-JIT model.\n Returns\n -------\n model : torch.nn.Module\n The CLIP model\n preprocess : Callable[[PIL.Image], torch.Tensor]\n A torchvision transform that converts a PIL image into a tensor that the returned model can take as its input\n \"\"\"\n if get_pretrained_url(name, \"openai\"):\n model_path = download_pretrained(get_pretrained_url(name, \"openai\"))\n elif os.path.isfile(name):\n model_path = name\n else:\n raise RuntimeError(\n f\"Model {name} not found; available models = {list_openai_models()}\"\n )\n\n try:\n # loading JIT archive\n model = torch.jit.load(model_path, map_location=device if jit else \"cpu\").eval()\n state_dict = None\n except RuntimeError:\n # loading saved state dict\n if jit:\n warnings.warn(\n f\"File {model_path} is not a JIT archive. Loading as a state dict instead\"\n )\n jit = False\n state_dict = torch.load(model_path, map_location=\"cpu\")\n\n if not jit:\n try:\n model = build_model_from_openai_state_dict(\n state_dict or model.state_dict()\n ).to(device)\n except KeyError:\n sd = {k[7:]: v for k, v in state_dict[\"state_dict\"].items()}\n model = build_model_from_openai_state_dict(sd).to(device)\n\n if str(device) == \"cpu\":\n model.float()\n return model\n\n # patch the device names\n device_holder = torch.jit.trace(\n lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]\n )\n device_node = [\n n\n for n in device_holder.graph.findAllNodes(\"prim::Constant\")\n if \"Device\" in repr(n)\n ][-1]\n\n def patch_device(module):\n try:\n graphs = [module.graph] if hasattr(module, \"graph\") else []\n except RuntimeError:\n graphs = []\n\n if hasattr(module, \"forward1\"):\n graphs.append(module.forward1.graph)\n\n for graph in graphs:\n for node in graph.findAllNodes(\"prim::Constant\"):\n if \"value\" in node.attributeNames() and str(node[\"value\"]).startswith(\n \"cuda\"\n ):\n node.copyAttributes(device_node)\n\n model.apply(patch_device)\n patch_device(model.encode_image)\n patch_device(model.encode_text)\n\n # patch dtype to float32 on CPU\n if str(device) == \"cpu\":\n float_holder = torch.jit.trace(\n lambda: torch.ones([]).float(), example_inputs=[]\n )\n float_input = list(float_holder.graph.findNode(\"aten::to\").inputs())[1]\n float_node = float_input.node()\n\n def patch_float(module):\n try:\n graphs = [module.graph] if hasattr(module, \"graph\") else []\n except RuntimeError:\n graphs = []\n\n if hasattr(module, \"forward1\"):\n graphs.append(module.forward1.graph)\n\n for graph in graphs:\n for node in graph.findAllNodes(\"aten::to\"):\n inputs = list(node.inputs())\n for i in [\n 1,\n 2,\n ]: # dtype can be the second or third argument to aten::to()\n if inputs[i].node()[\"value\"] == 5:\n inputs[i].node().copyAttributes(float_node)\n\n model.apply(patch_float)\n patch_float(model.encode_image)\n patch_float(model.encode_text)\n model.float()\n\n # ensure image_size attr available at consistent location for both jit and non-jit\n model.visual.image_size = model.input_resolution.item()\n return model\n\n\nopenai_imagenet_template = [\n lambda c: f\"a bad photo of a {c}.\",\n lambda c: f\"a photo of many {c}.\",\n lambda c: f\"a sculpture of a {c}.\",\n lambda c: f\"a photo of the hard to see {c}.\",\n lambda c: f\"a low resolution photo of the {c}.\",\n lambda c: f\"a rendering of a {c}.\",\n lambda c: f\"graffiti of a {c}.\",\n lambda c: f\"a bad photo of the {c}.\",\n lambda c: f\"a cropped photo of the {c}.\",\n lambda c: f\"a tattoo of a {c}.\",\n lambda c: f\"the embroidered {c}.\",\n lambda c: f\"a photo of a hard to see {c}.\",\n lambda c: f\"a bright photo of a {c}.\",\n lambda c: f\"a photo of a clean {c}.\",\n lambda c: f\"a photo of a dirty {c}.\",\n lambda c: f\"a dark photo of the {c}.\",\n lambda c: f\"a drawing of a {c}.\",","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.__init__","uri":"program://CREMA/function/lavis.models.clip_models.model.__init__#L418-L498","kind":"function","name":"__init__","path":"lavis/models/clip_models/model.py","language":"python","start_line":418,"end_line":498,"context_start_line":398,"context_end_line":518,"code":"@dataclass\nclass CLIPTextCfg:\n context_length: int\n vocab_size: int\n width: int\n heads: int\n layers: int\n\n\n@registry.register_model(\"clip\")\n@registry.register_model(\"clip_feature_extractor\")\nclass CLIP(BaseModel):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"ViT-B-32\": \"configs/models/clip_vit_base32.yaml\",\n \"ViT-B-16\": \"configs/models/clip_vit_base16.yaml\",\n \"ViT-L-14\": \"configs/models/clip_vit_large14.yaml\",\n \"ViT-L-14-336\": \"configs/models/clip_vit_large14_336.yaml\",\n \"RN50\": \"configs/models/clip_resnet50.yaml\",\n }\n\n def __init__(\n self,\n embed_dim: int,\n vision_cfg: CLIPVisionCfg,\n text_cfg: CLIPTextCfg,\n quick_gelu: bool = False,\n ):\n from .tokenizer import tokenize\n\n super().__init__()\n\n self.tokenizer = tokenize\n self._loss = None\n\n if isinstance(vision_cfg, dict):\n vision_cfg = CLIPVisionCfg(**vision_cfg)\n if isinstance(text_cfg, dict):\n text_cfg = CLIPTextCfg(**text_cfg)\n\n self.context_length = text_cfg.context_length\n\n # OpenAI models are pretrained w/ QuickGELU but native nn.GELU is both faster and more\n # memory efficient in recent PyTorch releases (>= 1.10).\n # NOTE: timm models always use native GELU regardless of quick_gelu flag.\n act_layer = QuickGELU if quick_gelu else nn.GELU\n\n if vision_cfg.timm_model_name:\n self.visual = TimmModel(\n vision_cfg.timm_model_name,\n pretrained=vision_cfg.timm_model_pretrained,\n pool=vision_cfg.timm_pool,\n proj=vision_cfg.timm_proj,\n embed_dim=embed_dim,\n image_size=vision_cfg.image_size,\n )\n act_layer = (\n nn.GELU\n ) # so that text transformer doesn't use QuickGELU w/ timm models\n elif isinstance(vision_cfg.layers, (tuple, list)):\n vision_heads = vision_cfg.width * 32 // 64\n self.visual = ModifiedResNet(\n layers=vision_cfg.layers,\n output_dim=embed_dim,\n heads=vision_heads,\n image_size=vision_cfg.image_size,\n width=vision_cfg.width,\n )\n else:\n vision_heads = vision_cfg.width // 64\n self.visual = VisualTransformer(\n image_size=vision_cfg.image_size,\n patch_size=vision_cfg.patch_size,\n width=vision_cfg.width,\n layers=vision_cfg.layers,\n heads=vision_heads,\n output_dim=embed_dim,\n act_layer=act_layer,\n )\n\n self.transformer = Transformer(\n width=text_cfg.width,\n layers=text_cfg.layers,\n heads=text_cfg.heads,\n act_layer=act_layer,\n )\n\n self.vocab_size = text_cfg.vocab_size\n self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)\n self.positional_embedding = nn.Parameter(\n torch.empty(self.context_length, text_cfg.width)\n )\n self.ln_final = LayerNorm(text_cfg.width)\n\n self.text_projection = nn.Parameter(torch.empty(text_cfg.width, embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n self.register_buffer(\"attn_mask\", self.build_attention_mask(), persistent=False)\n\n self.prompt_templates = openai_imagenet_template\n self.classifier = None\n\n self.init_parameters()\n\n @property\n def loss(self):\n if self._loss is None:\n from lavis.models.clip_models.loss import ClipLoss\n from torch import distributed as dist\n\n self._loss = ClipLoss(\n world_size=dist.get_world_size(),\n rank=dist.get_rank(),\n local_loss=False,\n gather_with_grad=False,\n use_horovod=False,\n )\n\n return self._loss\n\n def init_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.forward","uri":"program://CREMA/function/lavis.models.clip_models.model.forward#L571-L601","kind":"function","name":"forward","path":"lavis/models/clip_models/model.py","language":"python","start_line":571,"end_line":601,"context_start_line":551,"context_end_line":621,"code":"\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):\n def forward(self, samples):\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n loss = self.loss(image_features, text_features, self.logit_scale.exp())\n\n # return image_features, text_features, self.logit_scale.exp()\n # return {\"loss\": loss}\n return ClipOutput(\n intermediate_output=ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n ),\n loss=loss,\n logit_scale_exp=self.logit_scale.exp(),\n )\n\n def extract_features(self, samples):\n \"\"\"\n Extract features from the model for samples.\n\n Keys allowed are \"image\" and \"text_input\" in samples.\n If either key is missing, the corresponding features are not extracted.\n\n Args:\n samples: dict of samples to extract features from.\n\n Returns:\n ClipOutputFeatures object with features for the samples.\n \"\"\"\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model._make_layer","uri":"program://CREMA/function/lavis.models.clip_models.model._make_layer#L195-L202","kind":"function","name":"_make_layer","path":"lavis/models/clip_models/model.py","language":"python","start_line":195,"end_line":202,"context_start_line":175,"context_end_line":222,"code":" width // 2, width // 2, kernel_size=3, padding=1, bias=False\n )\n self.bn2 = nn.BatchNorm2d(width // 2)\n self.conv3 = nn.Conv2d(width // 2, width, kernel_size=3, padding=1, bias=False)\n self.bn3 = nn.BatchNorm2d(width)\n self.avgpool = nn.AvgPool2d(2)\n self.relu = nn.ReLU(inplace=True)\n\n # residual layers\n self._inplanes = width # this is a *mutable* variable used during construction\n self.layer1 = self._make_layer(width, layers[0])\n self.layer2 = self._make_layer(width * 2, layers[1], stride=2)\n self.layer3 = self._make_layer(width * 4, layers[2], stride=2)\n self.layer4 = self._make_layer(width * 8, layers[3], stride=2)\n\n embed_dim = width * 32 # the ResNet feature dimension\n self.attnpool = AttentionPool2d(image_size // 32, embed_dim, heads, output_dim)\n\n self.init_parameters()\n\n def _make_layer(self, planes, blocks, stride=1):\n layers = [Bottleneck(self._inplanes, planes, stride)]\n\n self._inplanes = planes * Bottleneck.expansion\n for _ in range(1, blocks):\n layers.append(Bottleneck(self._inplanes, planes))\n\n return nn.Sequential(*layers)\n\n def init_parameters(self):\n if self.attnpool is not None:\n std = self.attnpool.c_proj.in_features**-0.5\n nn.init.normal_(self.attnpool.q_proj.weight, std=std)\n nn.init.normal_(self.attnpool.k_proj.weight, std=std)\n nn.init.normal_(self.attnpool.v_proj.weight, std=std)\n nn.init.normal_(self.attnpool.c_proj.weight, std=std)\n\n for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:\n for name, param in resnet_block.named_parameters():\n if name.endswith(\"bn3.weight\"):\n nn.init.zeros_(param)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.init_parameters","uri":"program://CREMA/function/lavis.models.clip_models.model.init_parameters#L516-L536","kind":"function","name":"init_parameters","path":"lavis/models/clip_models/model.py","language":"python","start_line":516,"end_line":536,"context_start_line":496,"context_end_line":556,"code":" self.classifier = None\n\n self.init_parameters()\n\n @property\n def loss(self):\n if self._loss is None:\n from lavis.models.clip_models.loss import ClipLoss\n from torch import distributed as dist\n\n self._loss = ClipLoss(\n world_size=dist.get_world_size(),\n rank=dist.get_rank(),\n local_loss=False,\n gather_with_grad=False,\n use_horovod=False,\n )\n\n return self._loss\n\n def init_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n nn.init.constant_(self.logit_scale, np.log(1 / 0.07))\n\n if hasattr(self.visual, \"init_parameters\"):\n self.visual.init_parameters()\n\n proj_std = (self.transformer.width**-0.5) * (\n (2 * self.transformer.layers) ** -0.5\n )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.lock","uri":"program://CREMA/function/lavis.models.clip_models.model.lock#L342-L347","kind":"function","name":"lock","path":"lavis/models/clip_models/model.py","language":"python","start_line":342,"end_line":347,"context_start_line":322,"context_end_line":367,"code":" self.conv1 = nn.Conv2d(\n in_channels=3,\n out_channels=width,\n kernel_size=patch_size,\n stride=patch_size,\n bias=False,\n )\n\n scale = width**-0.5\n self.class_embedding = nn.Parameter(scale * torch.randn(width))\n self.positional_embedding = nn.Parameter(\n scale * torch.randn((image_size // patch_size) ** 2 + 1, width)\n )\n self.ln_pre = LayerNorm(width)\n\n self.transformer = Transformer(width, layers, heads, act_layer=act_layer)\n\n self.ln_post = LayerNorm(width)\n self.proj = nn.Parameter(scale * torch.randn(width, output_dim))\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n\n def forward(self, x: torch.Tensor):\n x = self.conv1(x) # shape = [*, width, grid, grid]\n x = x.reshape(x.shape[0], x.shape[1], -1) # shape = [*, width, grid ** 2]\n x = x.permute(0, 2, 1) # shape = [*, grid ** 2, width]\n x = torch.cat(\n [\n self.class_embedding.to(x.dtype)\n + torch.zeros(\n x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device\n ),\n x,\n ],\n dim=1,\n ) # shape = [*, grid ** 2 + 1, width]\n x = x + self.positional_embedding.to(x.dtype)\n x = self.ln_pre(x)\n\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x)","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.stem","uri":"program://CREMA/function/lavis.models.clip_models.model.stem#L226-L234","kind":"function","name":"stem","path":"lavis/models/clip_models/model.py","language":"python","start_line":226,"end_line":234,"context_start_line":206,"context_end_line":254,"code":" std = self.attnpool.c_proj.in_features**-0.5\n nn.init.normal_(self.attnpool.q_proj.weight, std=std)\n nn.init.normal_(self.attnpool.k_proj.weight, std=std)\n nn.init.normal_(self.attnpool.v_proj.weight, std=std)\n nn.init.normal_(self.attnpool.c_proj.weight, std=std)\n\n for resnet_block in [self.layer1, self.layer2, self.layer3, self.layer4]:\n for name, param in resnet_block.named_parameters():\n if name.endswith(\"bn3.weight\"):\n nn.init.zeros_(param)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n assert (\n unlocked_groups == 0\n ), \"partial locking not currently supported for this model\"\n for param in self.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self)\n\n def stem(self, x):\n for conv, bn in [\n (self.conv1, self.bn1),\n (self.conv2, self.bn2),\n (self.conv3, self.bn3),\n ]:\n x = self.relu(bn(conv(x)))\n x = self.avgpool(x)\n return x\n\n def forward(self, x):\n x = self.stem(x)\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n x = self.attnpool(x)\n\n return x\n\n\nclass LayerNorm(nn.LayerNorm):\n \"\"\"Subclass torch's LayerNorm to handle fp16.\"\"\"\n\n def forward(self, x: torch.Tensor):\n orig_type = x.dtype\n x = F.layer_norm(x, self.normalized_shape, self.weight, self.bias, self.eps)\n return x.to(orig_type)\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.attention","uri":"program://CREMA/function/lavis.models.clip_models.model.attention#L279-L280","kind":"function","name":"attention","path":"lavis/models/clip_models/model.py","language":"python","start_line":279,"end_line":280,"context_start_line":259,"context_end_line":300,"code":" return x * torch.sigmoid(1.702 * x)\n\n\nclass ResidualAttentionBlock(nn.Module):\n def __init__(self, d_model: int, n_head: int, act_layer: Callable = nn.GELU):\n super().__init__()\n\n self.attn = nn.MultiheadAttention(d_model, n_head)\n self.ln_1 = LayerNorm(d_model)\n self.mlp = nn.Sequential(\n OrderedDict(\n [\n (\"c_fc\", nn.Linear(d_model, d_model * 4)),\n (\"gelu\", act_layer()),\n (\"c_proj\", nn.Linear(d_model * 4, d_model)),\n ]\n )\n )\n self.ln_2 = LayerNorm(d_model)\n\n def attention(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n return self.attn(x, x, x, need_weights=False, attn_mask=attn_mask)[0]\n\n def forward(self, x: torch.Tensor, attn_mask: Optional[torch.Tensor] = None):\n x = x + self.attention(self.ln_1(x), attn_mask=attn_mask)\n x = x + self.mlp(self.ln_2(x))\n return x\n\n\nclass Transformer(nn.Module):\n def __init__(\n self, width: int, layers: int, heads: int, act_layer: Callable = nn.GELU\n ):\n super().__init__()\n self.width = width\n self.layers = layers\n self.resblocks = nn.ModuleList(\n [\n ResidualAttentionBlock(width, heads, act_layer=act_layer)\n for _ in range(layers)\n ]\n )","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.loss","uri":"program://CREMA/function/lavis.models.clip_models.model.loss#L501-L514","kind":"function","name":"loss","path":"lavis/models/clip_models/model.py","language":"python","start_line":501,"end_line":514,"context_start_line":481,"context_end_line":534,"code":" act_layer=act_layer,\n )\n\n self.vocab_size = text_cfg.vocab_size\n self.token_embedding = nn.Embedding(text_cfg.vocab_size, text_cfg.width)\n self.positional_embedding = nn.Parameter(\n torch.empty(self.context_length, text_cfg.width)\n )\n self.ln_final = LayerNorm(text_cfg.width)\n\n self.text_projection = nn.Parameter(torch.empty(text_cfg.width, embed_dim))\n self.logit_scale = nn.Parameter(torch.ones([]) * np.log(1 / 0.07))\n self.register_buffer(\"attn_mask\", self.build_attention_mask(), persistent=False)\n\n self.prompt_templates = openai_imagenet_template\n self.classifier = None\n\n self.init_parameters()\n\n @property\n def loss(self):\n if self._loss is None:\n from lavis.models.clip_models.loss import ClipLoss\n from torch import distributed as dist\n\n self._loss = ClipLoss(\n world_size=dist.get_world_size(),\n rank=dist.get_rank(),\n local_loss=False,\n gather_with_grad=False,\n use_horovod=False,\n )\n\n return self._loss\n\n def init_parameters(self):\n nn.init.normal_(self.token_embedding.weight, std=0.02)\n nn.init.normal_(self.positional_embedding, std=0.01)\n nn.init.constant_(self.logit_scale, np.log(1 / 0.07))\n\n if hasattr(self.visual, \"init_parameters\"):\n self.visual.init_parameters()\n\n proj_std = (self.transformer.width**-0.5) * (\n (2 * self.transformer.layers) ** -0.5\n )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.build_attention_mask","uri":"program://CREMA/function/lavis.models.clip_models.model.build_attention_mask#L538-L544","kind":"function","name":"build_attention_mask","path":"lavis/models/clip_models/model.py","language":"python","start_line":538,"end_line":544,"context_start_line":518,"context_end_line":564,"code":" nn.init.normal_(self.positional_embedding, std=0.01)\n nn.init.constant_(self.logit_scale, np.log(1 / 0.07))\n\n if hasattr(self.visual, \"init_parameters\"):\n self.visual.init_parameters()\n\n proj_std = (self.transformer.width**-0.5) * (\n (2 * self.transformer.layers) ** -0.5\n )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.lock_image_tower","uri":"program://CREMA/function/lavis.models.clip_models.model.lock_image_tower#L546-L550","kind":"function","name":"lock_image_tower","path":"lavis/models/clip_models/model.py","language":"python","start_line":546,"end_line":550,"context_start_line":526,"context_end_line":570,"code":" )\n attn_std = self.transformer.width**-0.5\n fc_std = (2 * self.transformer.width) ** -0.5\n for block in self.transformer.resblocks:\n nn.init.normal_(block.attn.in_proj_weight, std=attn_std)\n nn.init.normal_(block.attn.out_proj.weight, std=proj_std)\n nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.encode_image","uri":"program://CREMA/function/lavis.models.clip_models.model.encode_image#L552-L553","kind":"function","name":"encode_image","path":"lavis/models/clip_models/model.py","language":"python","start_line":552,"end_line":553,"context_start_line":532,"context_end_line":573,"code":" nn.init.normal_(block.mlp.c_fc.weight, std=fc_std)\n nn.init.normal_(block.mlp.c_proj.weight, std=proj_std)\n\n if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):\n def forward(self, samples):\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.encode_text","uri":"program://CREMA/function/lavis.models.clip_models.model.encode_text#L555-L568","kind":"function","name":"encode_text","path":"lavis/models/clip_models/model.py","language":"python","start_line":555,"end_line":568,"context_start_line":535,"context_end_line":588,"code":" if self.text_projection is not None:\n nn.init.normal_(self.text_projection, std=self.transformer.width**-0.5)\n\n def build_attention_mask(self):\n # lazily create causal attention mask, with full attention between the vision tokens\n # pytorch uses additive attention mask; fill with -inf\n mask = torch.empty(self.context_length, self.context_length)\n mask.fill_(float(\"-inf\"))\n mask.triu_(1) # zero out the lower diagonal\n return mask\n\n def lock_image_tower(self, unlocked_groups=0, freeze_bn_stats=False):\n # lock image tower as per LiT - https://arxiv.org/abs/2111.07991\n self.visual.lock(\n unlocked_groups=unlocked_groups, freeze_bn_stats=freeze_bn_stats\n )\n\n def encode_image(self, image):\n return self.visual(image)\n\n def encode_text(self, text):\n x = self.token_embedding(text) # [batch_size, n_ctx, d_model]\n\n x = x + self.positional_embedding\n x = x.permute(1, 0, 2) # NLD -> LND\n x = self.transformer(x, attn_mask=self.attn_mask)\n x = x.permute(1, 0, 2) # LND -> NLD\n x = self.ln_final(x)\n\n # x.shape = [batch_size, n_ctx, transformer.width]\n # take features from the eot embedding (eot_token is the highest number in each sequence)\n x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection\n\n return x\n\n # def forward(self, image, text):\n def forward(self, samples):\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n loss = self.loss(image_features, text_features, self.logit_scale.exp())","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.extract_features","uri":"program://CREMA/function/lavis.models.clip_models.model.extract_features#L603-L638","kind":"function","name":"extract_features","path":"lavis/models/clip_models/model.py","language":"python","start_line":603,"end_line":638,"context_start_line":583,"context_end_line":658,"code":" image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n loss = self.loss(image_features, text_features, self.logit_scale.exp())\n\n # return image_features, text_features, self.logit_scale.exp()\n # return {\"loss\": loss}\n return ClipOutput(\n intermediate_output=ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n ),\n loss=loss,\n logit_scale_exp=self.logit_scale.exp(),\n )\n\n def extract_features(self, samples):\n \"\"\"\n Extract features from the model for samples.\n\n Keys allowed are \"image\" and \"text_input\" in samples.\n If either key is missing, the corresponding features are not extracted.\n\n Args:\n samples: dict of samples to extract features from.\n\n Returns:\n ClipOutputFeatures object with features for the samples.\n \"\"\"\n image = samples.get(\"image\")\n text = samples.get(\"text_input\")\n\n if text is not None:\n text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n return ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n )\n\n def predict(self, samples):\n image = samples[\"image\"]\n targets = samples[\"label\"]\n\n image_features = self.encode_image(image)\n image_features = F.normalize(image_features, dim=-1)\n\n logits = 100.0 * image_features @ self.classifier\n\n return {\"predictions\": logits, \"targets\": targets}\n\n def before_evaluation(self, dataset, task_type, **kwargs):\n if task_type == MultimodalClassificationTask:\n self.classifier = self.zero_shot_classifier(\n classnames=dataset.classnames,\n templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.predict","uri":"program://CREMA/function/lavis.models.clip_models.model.predict#L640-L649","kind":"function","name":"predict","path":"lavis/models/clip_models/model.py","language":"python","start_line":640,"end_line":649,"context_start_line":620,"context_end_line":669,"code":" text = self.tokenizer(text).to(self.device)\n\n if image is None:\n return self.encode_text(text)\n elif text is None:\n return self.encode_image(image)\n\n image_embeds = self.encode_image(image)\n image_features = F.normalize(image_embeds, dim=-1)\n\n text_embeds = self.encode_text(text)\n text_features = F.normalize(text_embeds, dim=-1)\n\n return ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n )\n\n def predict(self, samples):\n image = samples[\"image\"]\n targets = samples[\"label\"]\n\n image_features = self.encode_image(image)\n image_features = F.normalize(image_features, dim=-1)\n\n logits = 100.0 * image_features @ self.classifier\n\n return {\"predictions\": logits, \"targets\": targets}\n\n def before_evaluation(self, dataset, task_type, **kwargs):\n if task_type == MultimodalClassificationTask:\n self.classifier = self.zero_shot_classifier(\n classnames=dataset.classnames,\n templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in classnames:\n texts = [\n template(classname) for template in templates\n ] # format with class\n texts = self.tokenizer(texts).to(self.device) # tokenize\n\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.before_evaluation","uri":"program://CREMA/function/lavis.models.clip_models.model.before_evaluation#L651-L656","kind":"function","name":"before_evaluation","path":"lavis/models/clip_models/model.py","language":"python","start_line":651,"end_line":656,"context_start_line":631,"context_end_line":676,"code":" text_features = F.normalize(text_embeds, dim=-1)\n\n return ClipOutputFeatures(\n image_embeds=image_embeds,\n image_embeds_proj=image_features,\n text_embeds=text_embeds,\n text_embeds_proj=text_features,\n )\n\n def predict(self, samples):\n image = samples[\"image\"]\n targets = samples[\"label\"]\n\n image_features = self.encode_image(image)\n image_features = F.normalize(image_features, dim=-1)\n\n logits = 100.0 * image_features @ self.classifier\n\n return {\"predictions\": logits, \"targets\": targets}\n\n def before_evaluation(self, dataset, task_type, **kwargs):\n if task_type == MultimodalClassificationTask:\n self.classifier = self.zero_shot_classifier(\n classnames=dataset.classnames,\n templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in classnames:\n texts = [\n template(classname) for template in templates\n ] # format with class\n texts = self.tokenizer(texts).to(self.device) # tokenize\n\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).to(self.device)\n return zeroshot_weights\n\n @classmethod\n def default_config_path(cls, model_type=\"base\"):\n model_type = \"ViT-B-32\" if model_type == \"base\" else model_type","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.zero_shot_classifier","uri":"program://CREMA/function/lavis.models.clip_models.model.zero_shot_classifier#L658-L672","kind":"function","name":"zero_shot_classifier","path":"lavis/models/clip_models/model.py","language":"python","start_line":658,"end_line":672,"context_start_line":638,"context_end_line":692,"code":" )\n\n def predict(self, samples):\n image = samples[\"image\"]\n targets = samples[\"label\"]\n\n image_features = self.encode_image(image)\n image_features = F.normalize(image_features, dim=-1)\n\n logits = 100.0 * image_features @ self.classifier\n\n return {\"predictions\": logits, \"targets\": targets}\n\n def before_evaluation(self, dataset, task_type, **kwargs):\n if task_type == MultimodalClassificationTask:\n self.classifier = self.zero_shot_classifier(\n classnames=dataset.classnames,\n templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in classnames:\n texts = [\n template(classname) for template in templates\n ] # format with class\n texts = self.tokenizer(texts).to(self.device) # tokenize\n\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).to(self.device)\n return zeroshot_weights\n\n @classmethod\n def default_config_path(cls, model_type=\"base\"):\n model_type = \"ViT-B-32\" if model_type == \"base\" else model_type\n\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}. \\n Available types: {}\".format(\n model_type, cls.PRETRAINED_MODEL_CONFIG_DICT.keys()\n )\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n @classmethod\n def from_config(cls, cfg=None):\n model_name = cfg.model_type\n pretrained = cfg.pretrained\n\n precision = cfg.get(\"precision\", \"fp32\")\n\n return create_model(","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.default_config_path","uri":"program://CREMA/function/lavis.models.clip_models.model.default_config_path#L675-L683","kind":"function","name":"default_config_path","path":"lavis/models/clip_models/model.py","language":"python","start_line":675,"end_line":683,"context_start_line":655,"context_end_line":703,"code":" templates=self.prompt_templates,\n )\n\n def zero_shot_classifier(self, classnames, templates):\n with torch.no_grad():\n zeroshot_weights = []\n for classname in classnames:\n texts = [\n template(classname) for template in templates\n ] # format with class\n texts = self.tokenizer(texts).to(self.device) # tokenize\n\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).to(self.device)\n return zeroshot_weights\n\n @classmethod\n def default_config_path(cls, model_type=\"base\"):\n model_type = \"ViT-B-32\" if model_type == \"base\" else model_type\n\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}. \\n Available types: {}\".format(\n model_type, cls.PRETRAINED_MODEL_CONFIG_DICT.keys()\n )\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n @classmethod\n def from_config(cls, cfg=None):\n model_name = cfg.model_type\n pretrained = cfg.pretrained\n\n precision = cfg.get(\"precision\", \"fp32\")\n\n return create_model(\n model_name=model_name, pretrained=pretrained, precision=precision\n )\n\n def zero_shot_predict(self, image_path, categories):\n assert isinstance(\n categories, list\n ), f\"categories must be a list, got {type(categories)}.\"\n assert os.path.exists(image_path), f\"File {image_path} does not exist.\"\n\n from lavis.processors.clip_processors import ClipImageEvalProcessor\n from PIL import Image","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.from_config","uri":"program://CREMA/function/lavis.models.clip_models.model.from_config#L686-L694","kind":"function","name":"from_config","path":"lavis/models/clip_models/model.py","language":"python","start_line":686,"end_line":694,"context_start_line":666,"context_end_line":714,"code":"\n class_embeddings = self.encode_text(texts)\n class_embedding = F.normalize(class_embeddings, dim=-1).mean(dim=0)\n class_embedding /= class_embedding.norm()\n zeroshot_weights.append(class_embedding)\n zeroshot_weights = torch.stack(zeroshot_weights, dim=1).to(self.device)\n return zeroshot_weights\n\n @classmethod\n def default_config_path(cls, model_type=\"base\"):\n model_type = \"ViT-B-32\" if model_type == \"base\" else model_type\n\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}. \\n Available types: {}\".format(\n model_type, cls.PRETRAINED_MODEL_CONFIG_DICT.keys()\n )\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n @classmethod\n def from_config(cls, cfg=None):\n model_name = cfg.model_type\n pretrained = cfg.pretrained\n\n precision = cfg.get(\"precision\", \"fp32\")\n\n return create_model(\n model_name=model_name, pretrained=pretrained, precision=precision\n )\n\n def zero_shot_predict(self, image_path, categories):\n assert isinstance(\n categories, list\n ), f\"categories must be a list, got {type(categories)}.\"\n assert os.path.exists(image_path), f\"File {image_path} does not exist.\"\n\n from lavis.processors.clip_processors import ClipImageEvalProcessor\n from PIL import Image\n\n image_preprocess = ClipImageEvalProcessor()\n image = image_preprocess(Image.open(image_path)).unsqueeze(0)\n\n text = self.tokenizer(categories)\n\n with torch.no_grad():\n image_features = self.encode_image(image)\n text_features = self.encode_text(text)\n image_features /= image_features.norm(dim=-1, keepdim=True)\n text_features /= text_features.norm(dim=-1, keepdim=True)","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.zero_shot_predict","uri":"program://CREMA/function/lavis.models.clip_models.model.zero_shot_predict#L696-L718","kind":"function","name":"zero_shot_predict","path":"lavis/models/clip_models/model.py","language":"python","start_line":696,"end_line":718,"context_start_line":676,"context_end_line":738,"code":" model_type = \"ViT-B-32\" if model_type == \"base\" else model_type\n\n assert (\n model_type in cls.PRETRAINED_MODEL_CONFIG_DICT\n ), \"Unknown model type {}. \\n Available types: {}\".format(\n model_type, cls.PRETRAINED_MODEL_CONFIG_DICT.keys()\n )\n return get_abs_path(cls.PRETRAINED_MODEL_CONFIG_DICT[model_type])\n\n @classmethod\n def from_config(cls, cfg=None):\n model_name = cfg.model_type\n pretrained = cfg.pretrained\n\n precision = cfg.get(\"precision\", \"fp32\")\n\n return create_model(\n model_name=model_name, pretrained=pretrained, precision=precision\n )\n\n def zero_shot_predict(self, image_path, categories):\n assert isinstance(\n categories, list\n ), f\"categories must be a list, got {type(categories)}.\"\n assert os.path.exists(image_path), f\"File {image_path} does not exist.\"\n\n from lavis.processors.clip_processors import ClipImageEvalProcessor\n from PIL import Image\n\n image_preprocess = ClipImageEvalProcessor()\n image = image_preprocess(Image.open(image_path)).unsqueeze(0)\n\n text = self.tokenizer(categories)\n\n with torch.no_grad():\n image_features = self.encode_image(image)\n text_features = self.encode_text(text)\n image_features /= image_features.norm(dim=-1, keepdim=True)\n text_features /= text_features.norm(dim=-1, keepdim=True)\n\n text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)\n\n print(\"Label probs:\", text_probs) # prints: [[1., 0., 0.]]\n\n def compute_sim_matrix(self, data_loader, **kwargs):\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_features = []\n\n for i in range(0, num_text, text_bs):\n\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(text).to(self.device)\n\n text_feat = self.encode_text(text_input)\n text_feat = F.normalize(text_feat, dim=-1)\n\n text_features.append(text_feat)\n","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.clip_models.model.compute_sim_matrix#L720-L760","kind":"function","name":"compute_sim_matrix","path":"lavis/models/clip_models/model.py","language":"python","start_line":720,"end_line":760,"context_start_line":700,"context_end_line":780,"code":" assert os.path.exists(image_path), f\"File {image_path} does not exist.\"\n\n from lavis.processors.clip_processors import ClipImageEvalProcessor\n from PIL import Image\n\n image_preprocess = ClipImageEvalProcessor()\n image = image_preprocess(Image.open(image_path)).unsqueeze(0)\n\n text = self.tokenizer(categories)\n\n with torch.no_grad():\n image_features = self.encode_image(image)\n text_features = self.encode_text(text)\n image_features /= image_features.norm(dim=-1, keepdim=True)\n text_features /= text_features.norm(dim=-1, keepdim=True)\n\n text_probs = (100.0 * image_features @ text_features.T).softmax(dim=-1)\n\n print(\"Label probs:\", text_probs) # prints: [[1., 0., 0.]]\n\n def compute_sim_matrix(self, data_loader, **kwargs):\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_features = []\n\n for i in range(0, num_text, text_bs):\n\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(text).to(self.device)\n\n text_feat = self.encode_text(text_input)\n text_feat = F.normalize(text_feat, dim=-1)\n\n text_features.append(text_feat)\n\n text_features = torch.cat(text_features, dim=0)\n\n image_features = []\n for samples in data_loader:\n image = samples[\"image\"]\n\n image = image.to(self.device)\n image_feat = self.encode_image(image)\n image_feat = F.normalize(image_feat, dim=-1)\n\n image_features.append(image_feat)\n\n image_features = torch.cat(image_features, dim=0)\n\n sims_matrix_i2t = image_features @ text_features.t()\n sims_matrix_t2i = sims_matrix_i2t.t()\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return sims_matrix_i2t.cpu().numpy(), sims_matrix_t2i.cpu().numpy()\n\n\ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n if isinstance(l, nn.MultiheadAttention):\n for attr in [\n *[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]],\n \"in_proj_bias\",\n \"bias_k\",\n \"bias_v\",\n ]:\n tensor = getattr(l, attr)\n if tensor is not None:","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model._convert_weights_to_fp16","uri":"program://CREMA/function/lavis.models.clip_models.model._convert_weights_to_fp16#L766-L787","kind":"function","name":"_convert_weights_to_fp16","path":"lavis/models/clip_models/model.py","language":"python","start_line":766,"end_line":787,"context_start_line":746,"context_end_line":807,"code":" image_feat = self.encode_image(image)\n image_feat = F.normalize(image_feat, dim=-1)\n\n image_features.append(image_feat)\n\n image_features = torch.cat(image_features, dim=0)\n\n sims_matrix_i2t = image_features @ text_features.t()\n sims_matrix_t2i = sims_matrix_i2t.t()\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return sims_matrix_i2t.cpu().numpy(), sims_matrix_t2i.cpu().numpy()\n\n\ndef convert_weights_to_fp16(model: nn.Module):\n \"\"\"Convert applicable model parameters to fp16\"\"\"\n\n def _convert_weights_to_fp16(l):\n if isinstance(l, (nn.Conv1d, nn.Conv2d, nn.Linear)):\n l.weight.data = l.weight.data.half()\n if l.bias is not None:\n l.bias.data = l.bias.data.half()\n\n if isinstance(l, nn.MultiheadAttention):\n for attr in [\n *[f\"{s}_proj_weight\" for s in [\"in\", \"q\", \"k\", \"v\"]],\n \"in_proj_bias\",\n \"bias_k\",\n \"bias_v\",\n ]:\n tensor = getattr(l, attr)\n if tensor is not None:\n tensor.data = tensor.data.half()\n\n for name in [\"text_projection\", \"proj\"]:\n if hasattr(l, name):\n attr = getattr(l, name)\n if attr is not None:\n attr.data = attr.data.half()\n\n model.apply(_convert_weights_to_fp16)\n\n\ndef build_model_from_openai_state_dict(state_dict: dict):\n vit = \"visual.proj\" in state_dict\n\n if vit:\n vision_width = state_dict[\"visual.conv1.weight\"].shape[0]\n vision_layers = len(\n [\n k\n for k in state_dict.keys()\n if k.startswith(\"visual.\") and k.endswith(\".attn.in_proj_weight\")\n ]\n )\n vision_patch_size = state_dict[\"visual.conv1.weight\"].shape[-1]\n grid_size = round(\n (state_dict[\"visual.positional_embedding\"].shape[0] - 1) ** 0.5\n )","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.patch_device","uri":"program://CREMA/function/lavis.models.clip_models.model.patch_device#L1116-L1130","kind":"function","name":"patch_device","path":"lavis/models/clip_models/model.py","language":"python","start_line":1116,"end_line":1130,"context_start_line":1096,"context_end_line":1150,"code":" state_dict or model.state_dict()\n ).to(device)\n except KeyError:\n sd = {k[7:]: v for k, v in state_dict[\"state_dict\"].items()}\n model = build_model_from_openai_state_dict(sd).to(device)\n\n if str(device) == \"cpu\":\n model.float()\n return model\n\n # patch the device names\n device_holder = torch.jit.trace(\n lambda: torch.ones([]).to(torch.device(device)), example_inputs=[]\n )\n device_node = [\n n\n for n in device_holder.graph.findAllNodes(\"prim::Constant\")\n if \"Device\" in repr(n)\n ][-1]\n\n def patch_device(module):\n try:\n graphs = [module.graph] if hasattr(module, \"graph\") else []\n except RuntimeError:\n graphs = []\n\n if hasattr(module, \"forward1\"):\n graphs.append(module.forward1.graph)\n\n for graph in graphs:\n for node in graph.findAllNodes(\"prim::Constant\"):\n if \"value\" in node.attributeNames() and str(node[\"value\"]).startswith(\n \"cuda\"\n ):\n node.copyAttributes(device_node)\n\n model.apply(patch_device)\n patch_device(model.encode_image)\n patch_device(model.encode_text)\n\n # patch dtype to float32 on CPU\n if str(device) == \"cpu\":\n float_holder = torch.jit.trace(\n lambda: torch.ones([]).float(), example_inputs=[]\n )\n float_input = list(float_holder.graph.findNode(\"aten::to\").inputs())[1]\n float_node = float_input.node()\n\n def patch_float(module):\n try:\n graphs = [module.graph] if hasattr(module, \"graph\") else []\n except RuntimeError:\n graphs = []\n\n if hasattr(module, \"forward1\"):","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.model.patch_float","uri":"program://CREMA/function/lavis.models.clip_models.model.patch_float#L1144-L1161","kind":"function","name":"patch_float","path":"lavis/models/clip_models/model.py","language":"python","start_line":1144,"end_line":1161,"context_start_line":1124,"context_end_line":1181,"code":"\n for graph in graphs:\n for node in graph.findAllNodes(\"prim::Constant\"):\n if \"value\" in node.attributeNames() and str(node[\"value\"]).startswith(\n \"cuda\"\n ):\n node.copyAttributes(device_node)\n\n model.apply(patch_device)\n patch_device(model.encode_image)\n patch_device(model.encode_text)\n\n # patch dtype to float32 on CPU\n if str(device) == \"cpu\":\n float_holder = torch.jit.trace(\n lambda: torch.ones([]).float(), example_inputs=[]\n )\n float_input = list(float_holder.graph.findNode(\"aten::to\").inputs())[1]\n float_node = float_input.node()\n\n def patch_float(module):\n try:\n graphs = [module.graph] if hasattr(module, \"graph\") else []\n except RuntimeError:\n graphs = []\n\n if hasattr(module, \"forward1\"):\n graphs.append(module.forward1.graph)\n\n for graph in graphs:\n for node in graph.findAllNodes(\"aten::to\"):\n inputs = list(node.inputs())\n for i in [\n 1,\n 2,\n ]: # dtype can be the second or third argument to aten::to()\n if inputs[i].node()[\"value\"] == 5:\n inputs[i].node().copyAttributes(float_node)\n\n model.apply(patch_float)\n patch_float(model.encode_image)\n patch_float(model.encode_text)\n model.float()\n\n # ensure image_size attr available at consistent location for both jit and non-jit\n model.visual.image_size = model.input_resolution.item()\n return model\n\n\nopenai_imagenet_template = [\n lambda c: f\"a bad photo of a {c}.\",\n lambda c: f\"a photo of many {c}.\",\n lambda c: f\"a sculpture of a {c}.\",\n lambda c: f\"a photo of the hard to see {c}.\",\n lambda c: f\"a low resolution photo of the {c}.\",\n lambda c: f\"a rendering of a {c}.\",\n lambda c: f\"graffiti of a {c}.\",\n lambda c: f\"a bad photo of the {c}.\",","source_hash":"3c051b9d9757c7f8a0914726e5f58c97ccea1078660b0cd1fd8837fa607026ab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.utils","uri":"program://CREMA/module/lavis.models.clip_models.utils#L1-L49","kind":"module","name":"lavis.models.clip_models.utils","path":"lavis/models/clip_models/utils.py","language":"python","start_line":1,"end_line":49,"context_start_line":1,"context_end_line":49,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nfrom torch import nn as nn\nfrom torchvision.ops.misc import FrozenBatchNorm2d\n\n\ndef freeze_batch_norm_2d(module, module_match={}, name=\"\"):\n \"\"\"\n Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is\n itself an instance of either `BatchNorm2d` or `SyncBatchNorm`, it is converted into `FrozenBatchNorm2d` and\n returned. Otherwise, the module is walked recursively and submodules are converted in place.\n Args:\n module (torch.nn.Module): Any PyTorch module.\n module_match (dict): Dictionary of full module names to freeze (all if empty)\n name (str): Full module name (prefix)\n Returns:\n torch.nn.Module: Resulting module\n Inspired by https://github.com/pytorch/pytorch/blob/a5895f85be0f10212791145bfedc0261d364f103/torch/nn/modules/batchnorm.py#L762\n \"\"\"\n res = module\n is_match = True\n if module_match:\n is_match = name in module_match\n if is_match and isinstance(\n module, (nn.modules.batchnorm.BatchNorm2d, nn.modules.batchnorm.SyncBatchNorm)\n ):\n res = FrozenBatchNorm2d(module.num_features)\n res.num_features = module.num_features\n res.affine = module.affine\n if module.affine:\n res.weight.data = module.weight.data.clone().detach()\n res.bias.data = module.bias.data.clone().detach()\n res.running_mean.data = module.running_mean.data\n res.running_var.data = module.running_var.data\n res.eps = module.eps\n else:\n for child_name, child in module.named_children():\n full_child_name = \".\".join([name, child_name]) if name else child_name\n new_child = freeze_batch_norm_2d(child, module_match, full_child_name)\n if new_child is not child:\n res.add_module(child_name, new_child)\n return res","source_hash":"f6d4d4fb82f3c199537e33b03bd5d4164f436615f77d0266a53615bb51e40dab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.utils.freeze_batch_norm_2d","uri":"program://CREMA/function/lavis.models.clip_models.utils.freeze_batch_norm_2d#L14-L49","kind":"function","name":"freeze_batch_norm_2d","path":"lavis/models/clip_models/utils.py","language":"python","start_line":14,"end_line":49,"context_start_line":1,"context_end_line":49,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nfrom torch import nn as nn\nfrom torchvision.ops.misc import FrozenBatchNorm2d\n\n\ndef freeze_batch_norm_2d(module, module_match={}, name=\"\"):\n \"\"\"\n Converts all `BatchNorm2d` and `SyncBatchNorm` layers of provided module into `FrozenBatchNorm2d`. If `module` is\n itself an instance of either `BatchNorm2d` or `SyncBatchNorm`, it is converted into `FrozenBatchNorm2d` and\n returned. Otherwise, the module is walked recursively and submodules are converted in place.\n Args:\n module (torch.nn.Module): Any PyTorch module.\n module_match (dict): Dictionary of full module names to freeze (all if empty)\n name (str): Full module name (prefix)\n Returns:\n torch.nn.Module: Resulting module\n Inspired by https://github.com/pytorch/pytorch/blob/a5895f85be0f10212791145bfedc0261d364f103/torch/nn/modules/batchnorm.py#L762\n \"\"\"\n res = module\n is_match = True\n if module_match:\n is_match = name in module_match\n if is_match and isinstance(\n module, (nn.modules.batchnorm.BatchNorm2d, nn.modules.batchnorm.SyncBatchNorm)\n ):\n res = FrozenBatchNorm2d(module.num_features)\n res.num_features = module.num_features\n res.affine = module.affine\n if module.affine:\n res.weight.data = module.weight.data.clone().detach()\n res.bias.data = module.bias.data.clone().detach()\n res.running_mean.data = module.running_mean.data\n res.running_var.data = module.running_var.data\n res.eps = module.eps\n else:\n for child_name, child in module.named_children():\n full_child_name = \".\".join([name, child_name]) if name else child_name\n new_child = freeze_batch_norm_2d(child, module_match, full_child_name)\n if new_child is not child:\n res.add_module(child_name, new_child)\n return res","source_hash":"f6d4d4fb82f3c199537e33b03bd5d4164f436615f77d0266a53615bb51e40dab","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform","uri":"program://CREMA/module/lavis.models.clip_models.transform#L1-L111","kind":"module","name":"lavis.models.clip_models.transform","path":"lavis/models/clip_models/transform.py","language":"python","start_line":1,"end_line":111,"context_start_line":1,"context_end_line":111,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nfrom typing import Optional, Sequence, Tuple\n\nimport torch\nimport torch.nn as nn\nimport torchvision.transforms.functional as F\n\n\nfrom torchvision.transforms import (\n Normalize,\n Compose,\n RandomResizedCrop,\n InterpolationMode,\n ToTensor,\n Resize,\n CenterCrop,\n)\n\n\nclass ResizeMaxSize(nn.Module):\n def __init__(\n self, max_size, interpolation=InterpolationMode.BICUBIC, fn=\"max\", fill=0\n ):\n super().__init__()\n if not isinstance(max_size, int):\n raise TypeError(f\"Size should be int. Got {type(max_size)}\")\n self.max_size = max_size\n self.interpolation = interpolation\n self.fn = min if fn == \"min\" else min\n self.fill = fill\n\n def forward(self, img):\n if isinstance(img, torch.Tensor):\n height, width = img.shape[:2]\n else:\n width, height = img.size\n scale = self.max_size / float(max(height, width))\n if scale != 1.0:\n new_size = tuple(round(dim * scale) for dim in (height, width))\n img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],\n fill=self.fill,\n )\n return img\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\ndef image_transform(\n image_size: int,\n is_train: bool,\n mean: Optional[Tuple[float, ...]] = None,\n std: Optional[Tuple[float, ...]] = None,\n resize_longest_max: bool = False,\n fill_color: int = 0,\n):\n mean = mean or (0.48145466, 0.4578275, 0.40821073) # OpenAI dataset mean\n std = std or (0.26862954, 0.26130258, 0.27577711) # OpenAI dataset std\n if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:\n # for square size, pass size as int so that Resize() uses aspect preserving shortest edge\n image_size = image_size[0]\n\n normalize = Normalize(mean=mean, std=std)\n if is_train:\n return Compose(\n [\n RandomResizedCrop(\n image_size,\n scale=(0.9, 1.0),\n interpolation=InterpolationMode.BICUBIC,\n ),\n _convert_to_rgb,\n ToTensor(),\n normalize,\n ]\n )\n else:\n if resize_longest_max:\n transforms = [ResizeMaxSize(image_size, fill=fill_color)]\n else:\n transforms = [\n Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n CenterCrop(image_size),\n ]\n transforms.extend(\n [\n _convert_to_rgb,\n ToTensor(),\n normalize,\n ]\n )\n return Compose(transforms)","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform.ResizeMaxSize","uri":"program://CREMA/class/lavis.models.clip_models.transform.ResizeMaxSize#L28-L61","kind":"class","name":"ResizeMaxSize","path":"lavis/models/clip_models/transform.py","language":"python","start_line":28,"end_line":61,"context_start_line":8,"context_end_line":81,"code":"\"\"\"\n\nfrom typing import Optional, Sequence, Tuple\n\nimport torch\nimport torch.nn as nn\nimport torchvision.transforms.functional as F\n\n\nfrom torchvision.transforms import (\n Normalize,\n Compose,\n RandomResizedCrop,\n InterpolationMode,\n ToTensor,\n Resize,\n CenterCrop,\n)\n\n\nclass ResizeMaxSize(nn.Module):\n def __init__(\n self, max_size, interpolation=InterpolationMode.BICUBIC, fn=\"max\", fill=0\n ):\n super().__init__()\n if not isinstance(max_size, int):\n raise TypeError(f\"Size should be int. Got {type(max_size)}\")\n self.max_size = max_size\n self.interpolation = interpolation\n self.fn = min if fn == \"min\" else min\n self.fill = fill\n\n def forward(self, img):\n if isinstance(img, torch.Tensor):\n height, width = img.shape[:2]\n else:\n width, height = img.size\n scale = self.max_size / float(max(height, width))\n if scale != 1.0:\n new_size = tuple(round(dim * scale) for dim in (height, width))\n img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],\n fill=self.fill,\n )\n return img\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\ndef image_transform(\n image_size: int,\n is_train: bool,\n mean: Optional[Tuple[float, ...]] = None,\n std: Optional[Tuple[float, ...]] = None,\n resize_longest_max: bool = False,\n fill_color: int = 0,\n):\n mean = mean or (0.48145466, 0.4578275, 0.40821073) # OpenAI dataset mean\n std = std or (0.26862954, 0.26130258, 0.27577711) # OpenAI dataset std\n if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:\n # for square size, pass size as int so that Resize() uses aspect preserving shortest edge\n image_size = image_size[0]\n","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform._convert_to_rgb","uri":"program://CREMA/function/lavis.models.clip_models.transform._convert_to_rgb#L64-L65","kind":"function","name":"_convert_to_rgb","path":"lavis/models/clip_models/transform.py","language":"python","start_line":64,"end_line":65,"context_start_line":44,"context_end_line":85,"code":" width, height = img.size\n scale = self.max_size / float(max(height, width))\n if scale != 1.0:\n new_size = tuple(round(dim * scale) for dim in (height, width))\n img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],\n fill=self.fill,\n )\n return img\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\ndef image_transform(\n image_size: int,\n is_train: bool,\n mean: Optional[Tuple[float, ...]] = None,\n std: Optional[Tuple[float, ...]] = None,\n resize_longest_max: bool = False,\n fill_color: int = 0,\n):\n mean = mean or (0.48145466, 0.4578275, 0.40821073) # OpenAI dataset mean\n std = std or (0.26862954, 0.26130258, 0.27577711) # OpenAI dataset std\n if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:\n # for square size, pass size as int so that Resize() uses aspect preserving shortest edge\n image_size = image_size[0]\n\n normalize = Normalize(mean=mean, std=std)\n if is_train:\n return Compose(\n [","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform.image_transform","uri":"program://CREMA/function/lavis.models.clip_models.transform.image_transform#L68-L111","kind":"function","name":"image_transform","path":"lavis/models/clip_models/transform.py","language":"python","start_line":68,"end_line":111,"context_start_line":48,"context_end_line":111,"code":" img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],\n fill=self.fill,\n )\n return img\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\ndef image_transform(\n image_size: int,\n is_train: bool,\n mean: Optional[Tuple[float, ...]] = None,\n std: Optional[Tuple[float, ...]] = None,\n resize_longest_max: bool = False,\n fill_color: int = 0,\n):\n mean = mean or (0.48145466, 0.4578275, 0.40821073) # OpenAI dataset mean\n std = std or (0.26862954, 0.26130258, 0.27577711) # OpenAI dataset std\n if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:\n # for square size, pass size as int so that Resize() uses aspect preserving shortest edge\n image_size = image_size[0]\n\n normalize = Normalize(mean=mean, std=std)\n if is_train:\n return Compose(\n [\n RandomResizedCrop(\n image_size,\n scale=(0.9, 1.0),\n interpolation=InterpolationMode.BICUBIC,\n ),\n _convert_to_rgb,\n ToTensor(),\n normalize,\n ]\n )\n else:\n if resize_longest_max:\n transforms = [ResizeMaxSize(image_size, fill=fill_color)]\n else:\n transforms = [\n Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n CenterCrop(image_size),\n ]\n transforms.extend(\n [\n _convert_to_rgb,\n ToTensor(),\n normalize,\n ]\n )\n return Compose(transforms)","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform.__init__","uri":"program://CREMA/function/lavis.models.clip_models.transform.__init__#L29-L38","kind":"function","name":"__init__","path":"lavis/models/clip_models/transform.py","language":"python","start_line":29,"end_line":38,"context_start_line":9,"context_end_line":58,"code":"\nfrom typing import Optional, Sequence, Tuple\n\nimport torch\nimport torch.nn as nn\nimport torchvision.transforms.functional as F\n\n\nfrom torchvision.transforms import (\n Normalize,\n Compose,\n RandomResizedCrop,\n InterpolationMode,\n ToTensor,\n Resize,\n CenterCrop,\n)\n\n\nclass ResizeMaxSize(nn.Module):\n def __init__(\n self, max_size, interpolation=InterpolationMode.BICUBIC, fn=\"max\", fill=0\n ):\n super().__init__()\n if not isinstance(max_size, int):\n raise TypeError(f\"Size should be int. Got {type(max_size)}\")\n self.max_size = max_size\n self.interpolation = interpolation\n self.fn = min if fn == \"min\" else min\n self.fill = fill\n\n def forward(self, img):\n if isinstance(img, torch.Tensor):\n height, width = img.shape[:2]\n else:\n width, height = img.size\n scale = self.max_size / float(max(height, width))\n if scale != 1.0:\n new_size = tuple(round(dim * scale) for dim in (height, width))\n img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.transform.forward","uri":"program://CREMA/function/lavis.models.clip_models.transform.forward#L40-L61","kind":"function","name":"forward","path":"lavis/models/clip_models/transform.py","language":"python","start_line":40,"end_line":61,"context_start_line":20,"context_end_line":81,"code":" RandomResizedCrop,\n InterpolationMode,\n ToTensor,\n Resize,\n CenterCrop,\n)\n\n\nclass ResizeMaxSize(nn.Module):\n def __init__(\n self, max_size, interpolation=InterpolationMode.BICUBIC, fn=\"max\", fill=0\n ):\n super().__init__()\n if not isinstance(max_size, int):\n raise TypeError(f\"Size should be int. Got {type(max_size)}\")\n self.max_size = max_size\n self.interpolation = interpolation\n self.fn = min if fn == \"min\" else min\n self.fill = fill\n\n def forward(self, img):\n if isinstance(img, torch.Tensor):\n height, width = img.shape[:2]\n else:\n width, height = img.size\n scale = self.max_size / float(max(height, width))\n if scale != 1.0:\n new_size = tuple(round(dim * scale) for dim in (height, width))\n img = F.resize(img, new_size, self.interpolation)\n pad_h = self.max_size - new_size[0]\n pad_w = self.max_size - new_size[1]\n img = F.pad(\n img,\n padding=[\n pad_w // 2,\n pad_h // 2,\n pad_w - pad_w // 2,\n pad_h - pad_h // 2,\n ],\n fill=self.fill,\n )\n return img\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\ndef image_transform(\n image_size: int,\n is_train: bool,\n mean: Optional[Tuple[float, ...]] = None,\n std: Optional[Tuple[float, ...]] = None,\n resize_longest_max: bool = False,\n fill_color: int = 0,\n):\n mean = mean or (0.48145466, 0.4578275, 0.40821073) # OpenAI dataset mean\n std = std or (0.26862954, 0.26130258, 0.27577711) # OpenAI dataset std\n if isinstance(image_size, (list, tuple)) and image_size[0] == image_size[1]:\n # for square size, pass size as int so that Resize() uses aspect preserving shortest edge\n image_size = image_size[0]\n","source_hash":"f257c3477ffc5515546454de236209ddf618dcad81e99891885dae1f8ef2ed0d","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.clip_outputs","uri":"program://CREMA/module/lavis.models.clip_models.clip_outputs#L1-L43","kind":"module","name":"lavis.models.clip_models.clip_outputs","path":"lavis/models/clip_models/clip_outputs.py","language":"python","start_line":1,"end_line":43,"context_start_line":1,"context_end_line":43,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nfrom dataclasses import dataclass\n\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import ModelOutput\n\n\n@dataclass\nclass ClipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass ClipOutput(ModelOutput):\n intermediate_output: Optional[ClipOutputFeatures] = None\n\n logit_scale_exp: Optional[torch.FloatTensor] = None\n\n loss: Optional[torch.FloatTensor] = None","source_hash":"1c631ea20d5341895fa87a417cfded4b8731365319271cc3355023d58f4c6988","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.clip_outputs.ClipOutputFeatures","uri":"program://CREMA/class/lavis.models.clip_models.clip_outputs.ClipOutputFeatures#L19-L34","kind":"class","name":"ClipOutputFeatures","path":"lavis/models/clip_models/clip_outputs.py","language":"python","start_line":19,"end_line":34,"context_start_line":1,"context_end_line":43,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\nfrom dataclasses import dataclass\n\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import ModelOutput\n\n\n@dataclass\nclass ClipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass ClipOutput(ModelOutput):\n intermediate_output: Optional[ClipOutputFeatures] = None\n\n logit_scale_exp: Optional[torch.FloatTensor] = None\n\n loss: Optional[torch.FloatTensor] = None","source_hash":"1c631ea20d5341895fa87a417cfded4b8731365319271cc3355023d58f4c6988","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.clip_outputs.ClipOutput","uri":"program://CREMA/class/lavis.models.clip_models.clip_outputs.ClipOutput#L38-L43","kind":"class","name":"ClipOutput","path":"lavis/models/clip_models/clip_outputs.py","language":"python","start_line":38,"end_line":43,"context_start_line":18,"context_end_line":43,"code":"@dataclass\nclass ClipOutputFeatures(ModelOutput):\n \"\"\"\n Data class of features from AlbefFeatureExtractor.\n\n Args:\n image_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n image_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n text_embeds: `torch.FloatTensor` of shape `(batch_size, 1, embed_dim)`, `optional`\n text_features: `torch.FloatTensor` of shape `(batch_size, 1, feature_dim)`, `optional`\n \"\"\"\n\n image_embeds: Optional[torch.FloatTensor] = None\n image_embeds_proj: Optional[torch.FloatTensor] = None\n\n text_embeds: Optional[torch.FloatTensor] = None\n text_embeds_proj: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass ClipOutput(ModelOutput):\n intermediate_output: Optional[ClipOutputFeatures] = None\n\n logit_scale_exp: Optional[torch.FloatTensor] = None\n\n loss: Optional[torch.FloatTensor] = None","source_hash":"1c631ea20d5341895fa87a417cfded4b8731365319271cc3355023d58f4c6988","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer","uri":"program://CREMA/module/lavis.models.clip_models.tokenizer#L1-L203","kind":"module","name":"lavis.models.clip_models.tokenizer","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":1,"end_line":203,"context_start_line":1,"context_end_line":203,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\n\"\"\" CLIP tokenizer\nCopied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.\n\"\"\"\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\nfrom typing import Union, List\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(\n os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\"\n )\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = (\n list(range(ord(\"!\"), ord(\"~\") + 1))\n + list(range(ord(\"¡\"), ord(\"¬\") + 1))\n + list(range(ord(\"®\"), ord(\"ÿ\") + 1))\n )\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8 + n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v + \"\" for v in vocab]\n for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:\n special_tokens = [\"\", \"\"]\n else:\n special_tokens = [\"\", \"\"] + special_tokens\n vocab.extend(special_tokens)\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {t: t for t in special_tokens}\n special = \"|\".join(special_tokens)\n self.pat = re.compile(\n special + r\"\"\"|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\",\n re.IGNORECASE,\n )\n\n self.vocab_size = len(self.encoder)\n self.all_special_ids = [self.encoder[t] for t in special_tokens]\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + (token[-1] + \"\",)\n pairs = get_pairs(word)\n\n if not pairs:\n return token + \"\"\n\n while True:\n bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float(\"inf\")))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word) - 1 and word[i + 1] == second:\n new_word.append(first + second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = \" \".join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n\n\n_tokenizer = SimpleTokenizer()\n\n\ndef tokenize(\n texts: Union[str, List[str]], context_length: int = 77\n) -> torch.LongTensor:\n \"\"\"\n Returns the tokenized representation of given input string(s)\n Parameters\n ----------\n texts : Union[str, List[str]]\n An input string or a list of input strings to tokenize\n context_length : int\n The context length to use; all CLIP models use 77 as the context length\n Returns\n -------\n A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]\n \"\"\"\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = _tokenizer.encoder[\"\"]\n eot_token = _tokenizer.encoder[\"\"]\n all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n if len(tokens) > context_length:\n tokens = tokens[:context_length] # Truncate\n result[i, : len(tokens)] = torch.tensor(tokens)\n\n return result","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.default_bpe","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.default_bpe#L25-L28","kind":"function","name":"default_bpe","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":25,"end_line":28,"context_start_line":5,"context_end_line":48,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\n\"\"\" CLIP tokenizer\nCopied from https://github.com/openai/CLIP. Originally MIT License, Copyright (c) 2021 OpenAI.\n\"\"\"\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\nfrom typing import Union, List\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(\n os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\"\n )\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = (\n list(range(ord(\"!\"), ord(\"~\") + 1))\n + list(range(ord(\"¡\"), ord(\"¬\") + 1))\n + list(range(ord(\"®\"), ord(\"ÿ\") + 1))\n )\n cs = bs[:]\n n = 0","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.bytes_to_unicode","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.bytes_to_unicode#L32-L55","kind":"function","name":"bytes_to_unicode","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":32,"end_line":55,"context_start_line":12,"context_end_line":75,"code":"\"\"\"\nimport gzip\nimport html\nimport os\nfrom functools import lru_cache\nfrom typing import Union, List\n\nimport ftfy\nimport regex as re\nimport torch\n\n\n@lru_cache()\ndef default_bpe():\n return os.path.join(\n os.path.dirname(os.path.abspath(__file__)), \"bpe_simple_vocab_16e6.txt.gz\"\n )\n\n\n@lru_cache()\ndef bytes_to_unicode():\n \"\"\"\n Returns list of utf-8 byte and a corresponding list of unicode strings.\n The reversible bpe codes work on unicode strings.\n This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.\n When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.\n This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = (\n list(range(ord(\"!\"), ord(\"~\") + 1))\n + list(range(ord(\"¡\"), ord(\"¬\") + 1))\n + list(range(ord(\"®\"), ord(\"ÿ\") + 1))\n )\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8 + n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.get_pairs","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.get_pairs#L58-L67","kind":"function","name":"get_pairs","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":58,"end_line":67,"context_start_line":38,"context_end_line":87,"code":" This is a signficant percentage of your normal, say, 32K bpe vocab.\n To avoid that, we want lookup tables between utf-8 bytes and unicode strings.\n And avoids mapping to whitespace/control characters the bpe code barfs on.\n \"\"\"\n bs = (\n list(range(ord(\"!\"), ord(\"~\") + 1))\n + list(range(ord(\"¡\"), ord(\"¬\") + 1))\n + list(range(ord(\"®\"), ord(\"ÿ\") + 1))\n )\n cs = bs[:]\n n = 0\n for b in range(2**8):\n if b not in bs:\n bs.append(b)\n cs.append(2**8 + n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.basic_clean","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.basic_clean#L70-L73","kind":"function","name":"basic_clean","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":70,"end_line":73,"context_start_line":50,"context_end_line":93,"code":" if b not in bs:\n bs.append(b)\n cs.append(2**8 + n)\n n += 1\n cs = [chr(n) for n in cs]\n return dict(zip(bs, cs))\n\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v + \"\" for v in vocab]\n for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.whitespace_clean","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.whitespace_clean#L76-L79","kind":"function","name":"whitespace_clean","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":76,"end_line":79,"context_start_line":56,"context_end_line":99,"code":"\n\ndef get_pairs(word):\n \"\"\"Return set of symbol pairs in a word.\n Word is represented as tuple of symbols (symbols being variable-length strings).\n \"\"\"\n pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v + \"\" for v in vocab]\n for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:\n special_tokens = [\"\", \"\"]\n else:\n special_tokens = [\"\", \"\"] + special_tokens\n vocab.extend(special_tokens)\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.SimpleTokenizer","uri":"program://CREMA/class/lavis.models.clip_models.tokenizer.SimpleTokenizer#L82-L169","kind":"class","name":"SimpleTokenizer","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":82,"end_line":169,"context_start_line":62,"context_end_line":189,"code":" pairs = set()\n prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v + \"\" for v in vocab]\n for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:\n special_tokens = [\"\", \"\"]\n else:\n special_tokens = [\"\", \"\"] + special_tokens\n vocab.extend(special_tokens)\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {t: t for t in special_tokens}\n special = \"|\".join(special_tokens)\n self.pat = re.compile(\n special + r\"\"\"|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\",\n re.IGNORECASE,\n )\n\n self.vocab_size = len(self.encoder)\n self.all_special_ids = [self.encoder[t] for t in special_tokens]\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + (token[-1] + \"\",)\n pairs = get_pairs(word)\n\n if not pairs:\n return token + \"\"\n\n while True:\n bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float(\"inf\")))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word) - 1 and word[i + 1] == second:\n new_word.append(first + second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = \" \".join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n\n\n_tokenizer = SimpleTokenizer()\n\n\ndef tokenize(\n texts: Union[str, List[str]], context_length: int = 77\n) -> torch.LongTensor:\n \"\"\"\n Returns the tokenized representation of given input string(s)\n Parameters\n ----------\n texts : Union[str, List[str]]\n An input string or a list of input strings to tokenize\n context_length : int\n The context length to use; all CLIP models use 77 as the context length\n Returns\n -------\n A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]\n \"\"\"","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.tokenize","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.tokenize#L175-L203","kind":"function","name":"tokenize","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":175,"end_line":203,"context_start_line":155,"context_end_line":203,"code":" for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n\n\n_tokenizer = SimpleTokenizer()\n\n\ndef tokenize(\n texts: Union[str, List[str]], context_length: int = 77\n) -> torch.LongTensor:\n \"\"\"\n Returns the tokenized representation of given input string(s)\n Parameters\n ----------\n texts : Union[str, List[str]]\n An input string or a list of input strings to tokenize\n context_length : int\n The context length to use; all CLIP models use 77 as the context length\n Returns\n -------\n A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]\n \"\"\"\n if isinstance(texts, str):\n texts = [texts]\n\n sot_token = _tokenizer.encoder[\"\"]\n eot_token = _tokenizer.encoder[\"\"]\n all_tokens = [[sot_token] + _tokenizer.encode(text) + [eot_token] for text in texts]\n result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)\n\n for i, tokens in enumerate(all_tokens):\n if len(tokens) > context_length:\n tokens = tokens[:context_length] # Truncate\n result[i, : len(tokens)] = torch.tensor(tokens)\n\n return result","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.__init__","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.__init__#L83-L109","kind":"function","name":"__init__","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":83,"end_line":109,"context_start_line":63,"context_end_line":129,"code":" prev_char = word[0]\n for char in word[1:]:\n pairs.add((prev_char, char))\n prev_char = char\n return pairs\n\n\ndef basic_clean(text):\n text = ftfy.fix_text(text)\n text = html.unescape(html.unescape(text))\n return text.strip()\n\n\ndef whitespace_clean(text):\n text = re.sub(r\"\\s+\", \" \", text)\n text = text.strip()\n return text\n\n\nclass SimpleTokenizer(object):\n def __init__(self, bpe_path: str = default_bpe(), special_tokens=None):\n self.byte_encoder = bytes_to_unicode()\n self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}\n merges = gzip.open(bpe_path).read().decode(\"utf-8\").split(\"\\n\")\n merges = merges[1 : 49152 - 256 - 2 + 1]\n merges = [tuple(merge.split()) for merge in merges]\n vocab = list(bytes_to_unicode().values())\n vocab = vocab + [v + \"\" for v in vocab]\n for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:\n special_tokens = [\"\", \"\"]\n else:\n special_tokens = [\"\", \"\"] + special_tokens\n vocab.extend(special_tokens)\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {t: t for t in special_tokens}\n special = \"|\".join(special_tokens)\n self.pat = re.compile(\n special + r\"\"\"|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\",\n re.IGNORECASE,\n )\n\n self.vocab_size = len(self.encoder)\n self.all_special_ids = [self.encoder[t] for t in special_tokens]\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + (token[-1] + \"\",)\n pairs = get_pairs(word)\n\n if not pairs:\n return token + \"\"\n\n while True:\n bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float(\"inf\")))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.bpe","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.bpe#L111-L150","kind":"function","name":"bpe","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":111,"end_line":150,"context_start_line":91,"context_end_line":170,"code":" for merge in merges:\n vocab.append(\"\".join(merge))\n if not special_tokens:\n special_tokens = [\"\", \"\"]\n else:\n special_tokens = [\"\", \"\"] + special_tokens\n vocab.extend(special_tokens)\n self.encoder = dict(zip(vocab, range(len(vocab))))\n self.decoder = {v: k for k, v in self.encoder.items()}\n self.bpe_ranks = dict(zip(merges, range(len(merges))))\n self.cache = {t: t for t in special_tokens}\n special = \"|\".join(special_tokens)\n self.pat = re.compile(\n special + r\"\"\"|'s|'t|'re|'ve|'m|'ll|'d|[\\p{L}]+|[\\p{N}]|[^\\s\\p{L}\\p{N}]+\"\"\",\n re.IGNORECASE,\n )\n\n self.vocab_size = len(self.encoder)\n self.all_special_ids = [self.encoder[t] for t in special_tokens]\n\n def bpe(self, token):\n if token in self.cache:\n return self.cache[token]\n word = tuple(token[:-1]) + (token[-1] + \"\",)\n pairs = get_pairs(word)\n\n if not pairs:\n return token + \"\"\n\n while True:\n bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float(\"inf\")))\n if bigram not in self.bpe_ranks:\n break\n first, second = bigram\n new_word = []\n i = 0\n while i < len(word):\n try:\n j = word.index(first, i)\n new_word.extend(word[i:j])\n i = j\n except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word) - 1 and word[i + 1] == second:\n new_word.append(first + second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = \" \".join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.encode","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.encode#L152-L160","kind":"function","name":"encode","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":152,"end_line":160,"context_start_line":132,"context_end_line":180,"code":" except:\n new_word.extend(word[i:])\n break\n\n if word[i] == first and i < len(word) - 1 and word[i + 1] == second:\n new_word.append(first + second)\n i += 2\n else:\n new_word.append(word[i])\n i += 1\n new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = \" \".join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n\n\n_tokenizer = SimpleTokenizer()\n\n\ndef tokenize(\n texts: Union[str, List[str]], context_length: int = 77\n) -> torch.LongTensor:\n \"\"\"\n Returns the tokenized representation of given input string(s)\n Parameters","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.tokenizer.decode","uri":"program://CREMA/function/lavis.models.clip_models.tokenizer.decode#L162-L169","kind":"function","name":"decode","path":"lavis/models/clip_models/tokenizer.py","language":"python","start_line":162,"end_line":169,"context_start_line":142,"context_end_line":189,"code":" new_word = tuple(new_word)\n word = new_word\n if len(word) == 1:\n break\n else:\n pairs = get_pairs(word)\n word = \" \".join(word)\n self.cache[token] = word\n return word\n\n def encode(self, text):\n bpe_tokens = []\n text = whitespace_clean(basic_clean(text)).lower()\n for token in re.findall(self.pat, text):\n token = \"\".join(self.byte_encoder[b] for b in token.encode(\"utf-8\"))\n bpe_tokens.extend(\n self.encoder[bpe_token] for bpe_token in self.bpe(token).split(\" \")\n )\n return bpe_tokens\n\n def decode(self, tokens):\n text = \"\".join([self.decoder[token] for token in tokens])\n text = (\n bytearray([self.byte_decoder[c] for c in text])\n .decode(\"utf-8\", errors=\"replace\")\n .replace(\"\", \" \")\n )\n return text\n\n\n_tokenizer = SimpleTokenizer()\n\n\ndef tokenize(\n texts: Union[str, List[str]], context_length: int = 77\n) -> torch.LongTensor:\n \"\"\"\n Returns the tokenized representation of given input string(s)\n Parameters\n ----------\n texts : Union[str, List[str]]\n An input string or a list of input strings to tokenize\n context_length : int\n The context length to use; all CLIP models use 77 as the context length\n Returns\n -------\n A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length]\n \"\"\"","source_hash":"7548dc1304b9e337afb9d28b4f81298b6f7ce36dfddff20fe4d69d0401a59f18","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model","uri":"program://CREMA/module/lavis.models.clip_models.timm_model#L1-L561","kind":"module","name":"lavis.models.clip_models.timm_model","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":1,"end_line":561,"context_start_line":1,"context_end_line":561,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\n Based on https://github.com/mlfoundations/open_clip\n\"\"\"\n\n\"\"\" timm model adapter\nWraps timm (https://github.com/rwightman/pytorch-image-models) models for use as a vision tower in CLIP model.\n\"\"\"\nimport math\nimport warnings\nfrom collections import OrderedDict\nfrom typing import List, Optional, Tuple, Union\n\nimport torch\nimport torch.nn as nn\nfrom torch import nn as nn\n\ntry:\n import timm\n from timm.models.layers import Mlp, to_2tuple\n\n # from timm.models.layers.attention_pool2d import RotAttentionPool2d\n # from timm.models.layers.attention_pool2d import (\n # AttentionPool2d as AbsAttentionPool2d,\n # )\n\nexcept ImportError as e:\n timm = None\n\nfrom lavis.models.clip_models.utils import freeze_batch_norm_2d\n\n\nclass TimmModel(nn.Module):\n \"\"\"timm model adapter\n # FIXME this adapter is a work in progress, may change in ways that break weight compat\n \"\"\"\n\n def __init__(\n self,\n model_name,\n embed_dim,\n image_size=224,\n pool=\"avg\",\n proj=\"linear\",\n drop=0.0,\n pretrained=False,\n ):\n super().__init__()\n if timm is None:\n raise RuntimeError(\"Please `pip install timm` to use timm models.\")\n\n self.image_size = to_2tuple(image_size)\n self.trunk = timm.create_model(model_name, pretrained=pretrained)\n feat_size = self.trunk.default_cfg.get(\"pool_size\", None)\n feature_ndim = 1 if not feat_size else 2\n if pool in (\"abs_attn\", \"rot_attn\"):\n assert feature_ndim == 2\n # if attn pooling used, remove both classifier and default pool\n self.trunk.reset_classifier(0, global_pool=\"\")\n else:\n # reset global pool if pool config set, otherwise leave as network default\n reset_kwargs = dict(global_pool=pool) if pool else {}\n self.trunk.reset_classifier(0, **reset_kwargs)\n prev_chs = self.trunk.num_features\n\n head_layers = OrderedDict()\n if pool == \"abs_attn\":\n head_layers[\"pool\"] = AttentionPool2d(\n prev_chs, feat_size=feat_size, out_features=embed_dim\n )\n prev_chs = embed_dim\n elif pool == \"rot_attn\":\n head_layers[\"pool\"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)\n prev_chs = embed_dim\n else:\n assert proj, \"projection layer needed if non-attention pooling is used.\"\n\n # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used\n if proj == \"linear\":\n head_layers[\"drop\"] = nn.Dropout(drop)\n head_layers[\"proj\"] = nn.Linear(prev_chs, embed_dim)\n elif proj == \"mlp\":\n head_layers[\"mlp\"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)\n\n self.head = nn.Sequential(head_layers)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n \"\"\"lock modules\n Args:\n unlocked_groups (int): leave last n layer groups unlocked (default: 0)\n \"\"\"\n if not unlocked_groups:\n # lock full model\n for param in self.trunk.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self.trunk)\n else:\n # NOTE: partial freeze requires latest timm (master) branch and is subject to change\n try:\n # FIXME import here until API stable and in an official release\n from timm.models.helpers import group_modules, group_parameters\n except ImportError:\n raise RuntimeError(\n \"Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`\"\n )\n matcher = self.trunk.group_matcher()\n gparams = group_parameters(self.trunk, matcher)\n max_layer_id = max(gparams.keys())\n max_layer_id = max_layer_id - unlocked_groups\n for group_idx in range(max_layer_id + 1):\n group = gparams[group_idx]\n for param in group:\n self.trunk.get_parameter(param).requires_grad = False\n if freeze_bn_stats:\n gmodules = group_modules(self.trunk, matcher, reverse=True)\n gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}\n freeze_batch_norm_2d(self.trunk, gmodules)\n\n def forward(self, x):\n x = self.trunk(x)\n x = self.head(x)\n return x\n\n\nclass RotAttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ rotary (relative) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n Adapted from the AttentionPool2d in CLIP w/ rotary embedding instead of learned embed.\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: While this impl does not require a fixed feature size, performance at differeing resolutions from\n train varies widely and falls off dramatically. I'm not sure if there is a way around this... -RW\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()\n embed_dim = embed_dim or in_features\n out_features = out_features or in_features\n self.qkv = nn.Linear(in_features, embed_dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(embed_dim, out_features)\n self.num_heads = num_heads\n assert embed_dim % num_heads == 0\n self.head_dim = embed_dim // num_heads\n self.scale = self.head_dim**-0.5\n self.pos_embed = RotaryEmbedding(self.head_dim)\n\n trunc_normal_(self.qkv.weight, std=in_features**-0.5)\n nn.init.zeros_(self.qkv.bias)\n\n def forward(self, x):\n B, _, H, W = x.shape\n N = H * W\n x = x.reshape(B, -1, N).permute(0, 2, 1)\n\n x = torch.cat([x.mean(1, keepdim=True), x], dim=1)\n\n x = (\n self.qkv(x)\n .reshape(B, N + 1, 3, self.num_heads, self.head_dim)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = x[0], x[1], x[2]\n\n qc, q = q[:, :, :1], q[:, :, 1:]\n sin_emb, cos_emb = self.pos_embed.get_embed((H, W))\n q = apply_rot_embed(q, sin_emb, cos_emb)\n q = torch.cat([qc, q], dim=2)\n\n kc, k = k[:, :, :1], k[:, :, 1:]\n k = apply_rot_embed(k, sin_emb, cos_emb)\n k = torch.cat([kc, k], dim=2)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\nclass AttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ learned (absolute) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n It was based on impl in CLIP by OpenAI\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: This requires feature size upon construction and well prevent adaptive sizing of the network.\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n feat_size: Union[int, Tuple[int, int]],\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()\n\n embed_dim = embed_dim or in_features\n out_features = out_features or in_features\n assert embed_dim % num_heads == 0\n self.feat_size = to_2tuple(feat_size)\n self.qkv = nn.Linear(in_features, embed_dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(embed_dim, out_features)\n self.num_heads = num_heads\n self.head_dim = embed_dim // num_heads\n self.scale = self.head_dim**-0.5\n\n spatial_dim = self.feat_size[0] * self.feat_size[1]\n self.pos_embed = nn.Parameter(torch.zeros(spatial_dim + 1, in_features))\n trunc_normal_(self.pos_embed, std=in_features**-0.5)\n trunc_normal_(self.qkv.weight, std=in_features**-0.5)\n nn.init.zeros_(self.qkv.bias)\n\n def forward(self, x):\n B, _, H, W = x.shape\n N = H * W\n assert self.feat_size[0] == H\n assert self.feat_size[1] == W\n x = x.reshape(B, -1, N).permute(0, 2, 1)\n x = torch.cat([x.mean(1, keepdim=True), x], dim=1)\n x = x + self.pos_embed.unsqueeze(0).to(x.dtype)\n\n x = (\n self.qkv(x)\n .reshape(B, N + 1, 3, self.num_heads, self.head_dim)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = x[0], x[1], x[2]\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\ndef pixel_freq_bands(\n num_bands: int,\n max_freq: float = 224.0,\n linear_bands: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n if linear_bands:\n bands = torch.linspace(1.0, max_freq / 2, num_bands, dtype=dtype, device=device)\n else:\n bands = 2 ** torch.linspace(\n 0, math.log(max_freq, 2) - 1, num_bands, dtype=dtype, device=device\n )\n return bands * torch.pi\n\n\ndef inv_freq_bands(\n num_bands: int,\n temperature: float = 100000.0,\n step: int = 2,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n inv_freq = 1.0 / (\n temperature\n ** (torch.arange(0, num_bands, step, dtype=dtype, device=device) / num_bands)\n )\n return inv_freq\n\n\ndef build_sincos2d_pos_embed(\n feat_shape: List[int],\n dim: int = 64,\n temperature: float = 10000.0,\n reverse_coord: bool = False,\n interleave_sin_cos: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n \"\"\"\n Args:\n feat_shape:\n dim:\n temperature:\n reverse_coord: stack grid order W, H instead of H, W\n interleave_sin_cos: sin, cos, sin, cos stack instead of sin, sin, cos, cos\n dtype:\n device:\n Returns:\n \"\"\"\n assert (\n dim % 4 == 0\n ), \"Embed dimension must be divisible by 4 for sin-cos 2D position embedding\"\n pos_dim = dim // 4\n bands = inv_freq_bands(\n pos_dim, temperature=temperature, step=1, dtype=dtype, device=device\n )\n\n if reverse_coord:\n feat_shape = feat_shape[::-1] # stack W, H instead of H, W\n grid = (\n torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n )\n )\n .flatten(1)\n .transpose(0, 1)\n )\n pos2 = grid.unsqueeze(-1) * bands.unsqueeze(0)\n # FIXME add support for unflattened spatial dim?\n\n stack_dim = (\n 2 if interleave_sin_cos else 1\n ) # stack sin, cos, sin, cos instead of sin sin cos cos\n pos_emb = torch.stack([torch.sin(pos2), torch.cos(pos2)], dim=stack_dim).flatten(1)\n return pos_emb\n\n\ndef build_fourier_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n num_bands: int = 64,\n max_res: int = 224,\n linear_bands: bool = False,\n include_grid: bool = False,\n concat_out: bool = True,\n in_pixels: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> List[torch.Tensor]:\n if bands is None:\n if in_pixels:\n bands = pixel_freq_bands(\n num_bands,\n float(max_res),\n linear_bands=linear_bands,\n dtype=dtype,\n device=device,\n )\n else:\n bands = inv_freq_bands(num_bands, step=1, dtype=dtype, device=device)\n else:\n if device is None:\n device = bands.device\n if dtype is None:\n dtype = bands.dtype\n\n if in_pixels:\n grid = torch.stack(\n torch.meshgrid(\n [\n torch.linspace(-1.0, 1.0, steps=s, device=device, dtype=dtype)\n for s in feat_shape\n ]\n ),\n dim=-1,\n )\n else:\n grid = torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n ),\n dim=-1,\n )\n grid = grid.unsqueeze(-1)\n pos = grid * bands\n\n pos_sin, pos_cos = pos.sin(), pos.cos()\n out = (grid, pos_sin, pos_cos) if include_grid else (pos_sin, pos_cos)\n # FIXME torchscript doesn't like multiple return types, probably need to always cat?\n if concat_out:\n out = torch.cat(out, dim=-1)\n return out\n\n\nclass FourierEmbed(nn.Module):\n def __init__(\n self,\n max_res: int = 224,\n num_bands: int = 64,\n concat_grid=True,\n keep_spatial=False,\n ):\n super().__init__()\n self.max_res = max_res\n self.num_bands = num_bands\n self.concat_grid = concat_grid\n self.keep_spatial = keep_spatial\n self.register_buffer(\n \"bands\", pixel_freq_bands(max_res, num_bands), persistent=False\n )\n\n def forward(self, x):\n B, C = x.shape[:2]\n feat_shape = x.shape[2:]\n emb = build_fourier_pos_embed(\n feat_shape,\n self.bands,\n include_grid=self.concat_grid,\n dtype=x.dtype,\n device=x.device,\n )\n emb = emb.transpose(-1, -2).flatten(len(feat_shape))\n batch_expand = (B,) + (-1,) * (x.ndim - 1)\n\n # FIXME support nD\n if self.keep_spatial:\n x = torch.cat(\n [x, emb.unsqueeze(0).expand(batch_expand).permute(0, 3, 1, 2)], dim=1\n )\n else:\n x = torch.cat(\n [x.permute(0, 2, 3, 1), emb.unsqueeze(0).expand(batch_expand)], dim=-1\n )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n dim: int = 64,\n max_freq: float = 224,\n linear_bands: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n \"\"\"\n NOTE: shape arg should include spatial dim only\n \"\"\"\n feat_shape = torch.Size(feat_shape)\n\n sin_emb, cos_emb = build_fourier_pos_embed(\n feat_shape,\n bands=bands,\n num_bands=dim // 4,\n max_res=max_freq,\n linear_bands=linear_bands,\n concat_out=False,\n device=device,\n dtype=dtype,\n )\n N = feat_shape.numel()\n sin_emb = sin_emb.reshape(N, -1).repeat_interleave(2, -1)\n cos_emb = cos_emb.reshape(N, -1).repeat_interleave(2, -1)\n return sin_emb, cos_emb\n\n\nclass RotaryEmbedding(nn.Module):\n \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.TimmModel","uri":"program://CREMA/class/lavis.models.clip_models.timm_model.TimmModel#L37-L127","kind":"class","name":"TimmModel","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":37,"end_line":127,"context_start_line":17,"context_end_line":147,"code":"\nimport torch\nimport torch.nn as nn\nfrom torch import nn as nn\n\ntry:\n import timm\n from timm.models.layers import Mlp, to_2tuple\n\n # from timm.models.layers.attention_pool2d import RotAttentionPool2d\n # from timm.models.layers.attention_pool2d import (\n # AttentionPool2d as AbsAttentionPool2d,\n # )\n\nexcept ImportError as e:\n timm = None\n\nfrom lavis.models.clip_models.utils import freeze_batch_norm_2d\n\n\nclass TimmModel(nn.Module):\n \"\"\"timm model adapter\n # FIXME this adapter is a work in progress, may change in ways that break weight compat\n \"\"\"\n\n def __init__(\n self,\n model_name,\n embed_dim,\n image_size=224,\n pool=\"avg\",\n proj=\"linear\",\n drop=0.0,\n pretrained=False,\n ):\n super().__init__()\n if timm is None:\n raise RuntimeError(\"Please `pip install timm` to use timm models.\")\n\n self.image_size = to_2tuple(image_size)\n self.trunk = timm.create_model(model_name, pretrained=pretrained)\n feat_size = self.trunk.default_cfg.get(\"pool_size\", None)\n feature_ndim = 1 if not feat_size else 2\n if pool in (\"abs_attn\", \"rot_attn\"):\n assert feature_ndim == 2\n # if attn pooling used, remove both classifier and default pool\n self.trunk.reset_classifier(0, global_pool=\"\")\n else:\n # reset global pool if pool config set, otherwise leave as network default\n reset_kwargs = dict(global_pool=pool) if pool else {}\n self.trunk.reset_classifier(0, **reset_kwargs)\n prev_chs = self.trunk.num_features\n\n head_layers = OrderedDict()\n if pool == \"abs_attn\":\n head_layers[\"pool\"] = AttentionPool2d(\n prev_chs, feat_size=feat_size, out_features=embed_dim\n )\n prev_chs = embed_dim\n elif pool == \"rot_attn\":\n head_layers[\"pool\"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)\n prev_chs = embed_dim\n else:\n assert proj, \"projection layer needed if non-attention pooling is used.\"\n\n # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used\n if proj == \"linear\":\n head_layers[\"drop\"] = nn.Dropout(drop)\n head_layers[\"proj\"] = nn.Linear(prev_chs, embed_dim)\n elif proj == \"mlp\":\n head_layers[\"mlp\"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)\n\n self.head = nn.Sequential(head_layers)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n \"\"\"lock modules\n Args:\n unlocked_groups (int): leave last n layer groups unlocked (default: 0)\n \"\"\"\n if not unlocked_groups:\n # lock full model\n for param in self.trunk.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self.trunk)\n else:\n # NOTE: partial freeze requires latest timm (master) branch and is subject to change\n try:\n # FIXME import here until API stable and in an official release\n from timm.models.helpers import group_modules, group_parameters\n except ImportError:\n raise RuntimeError(\n \"Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`\"\n )\n matcher = self.trunk.group_matcher()\n gparams = group_parameters(self.trunk, matcher)\n max_layer_id = max(gparams.keys())\n max_layer_id = max_layer_id - unlocked_groups\n for group_idx in range(max_layer_id + 1):\n group = gparams[group_idx]\n for param in group:\n self.trunk.get_parameter(param).requires_grad = False\n if freeze_bn_stats:\n gmodules = group_modules(self.trunk, matcher, reverse=True)\n gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}\n freeze_batch_norm_2d(self.trunk, gmodules)\n\n def forward(self, x):\n x = self.trunk(x)\n x = self.head(x)\n return x\n\n\nclass RotAttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ rotary (relative) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n Adapted from the AttentionPool2d in CLIP w/ rotary embedding instead of learned embed.\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: While this impl does not require a fixed feature size, performance at differeing resolutions from\n train varies widely and falls off dramatically. I'm not sure if there is a way around this... -RW\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.RotAttentionPool2d","uri":"program://CREMA/class/lavis.models.clip_models.timm_model.RotAttentionPool2d#L130-L189","kind":"class","name":"RotAttentionPool2d","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":130,"end_line":189,"context_start_line":110,"context_end_line":209,"code":" )\n matcher = self.trunk.group_matcher()\n gparams = group_parameters(self.trunk, matcher)\n max_layer_id = max(gparams.keys())\n max_layer_id = max_layer_id - unlocked_groups\n for group_idx in range(max_layer_id + 1):\n group = gparams[group_idx]\n for param in group:\n self.trunk.get_parameter(param).requires_grad = False\n if freeze_bn_stats:\n gmodules = group_modules(self.trunk, matcher, reverse=True)\n gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}\n freeze_batch_norm_2d(self.trunk, gmodules)\n\n def forward(self, x):\n x = self.trunk(x)\n x = self.head(x)\n return x\n\n\nclass RotAttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ rotary (relative) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n Adapted from the AttentionPool2d in CLIP w/ rotary embedding instead of learned embed.\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: While this impl does not require a fixed feature size, performance at differeing resolutions from\n train varies widely and falls off dramatically. I'm not sure if there is a way around this... -RW\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()\n embed_dim = embed_dim or in_features\n out_features = out_features or in_features\n self.qkv = nn.Linear(in_features, embed_dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(embed_dim, out_features)\n self.num_heads = num_heads\n assert embed_dim % num_heads == 0\n self.head_dim = embed_dim // num_heads\n self.scale = self.head_dim**-0.5\n self.pos_embed = RotaryEmbedding(self.head_dim)\n\n trunc_normal_(self.qkv.weight, std=in_features**-0.5)\n nn.init.zeros_(self.qkv.bias)\n\n def forward(self, x):\n B, _, H, W = x.shape\n N = H * W\n x = x.reshape(B, -1, N).permute(0, 2, 1)\n\n x = torch.cat([x.mean(1, keepdim=True), x], dim=1)\n\n x = (\n self.qkv(x)\n .reshape(B, N + 1, 3, self.num_heads, self.head_dim)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = x[0], x[1], x[2]\n\n qc, q = q[:, :, :1], q[:, :, 1:]\n sin_emb, cos_emb = self.pos_embed.get_embed((H, W))\n q = apply_rot_embed(q, sin_emb, cos_emb)\n q = torch.cat([qc, q], dim=2)\n\n kc, k = k[:, :, :1], k[:, :, 1:]\n k = apply_rot_embed(k, sin_emb, cos_emb)\n k = torch.cat([kc, k], dim=2)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\nclass AttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ learned (absolute) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n It was based on impl in CLIP by OpenAI\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: This requires feature size upon construction and well prevent adaptive sizing of the network.\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n feat_size: Union[int, Tuple[int, int]],\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.AttentionPool2d","uri":"program://CREMA/class/lavis.models.clip_models.timm_model.AttentionPool2d#L192-L247","kind":"class","name":"AttentionPool2d","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":192,"end_line":247,"context_start_line":172,"context_end_line":267,"code":" )\n q, k, v = x[0], x[1], x[2]\n\n qc, q = q[:, :, :1], q[:, :, 1:]\n sin_emb, cos_emb = self.pos_embed.get_embed((H, W))\n q = apply_rot_embed(q, sin_emb, cos_emb)\n q = torch.cat([qc, q], dim=2)\n\n kc, k = k[:, :, :1], k[:, :, 1:]\n k = apply_rot_embed(k, sin_emb, cos_emb)\n k = torch.cat([kc, k], dim=2)\n\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\nclass AttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ learned (absolute) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n It was based on impl in CLIP by OpenAI\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: This requires feature size upon construction and well prevent adaptive sizing of the network.\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n feat_size: Union[int, Tuple[int, int]],\n out_features: int = None,\n embed_dim: int = None,\n num_heads: int = 4,\n qkv_bias: bool = True,\n ):\n super().__init__()\n\n embed_dim = embed_dim or in_features\n out_features = out_features or in_features\n assert embed_dim % num_heads == 0\n self.feat_size = to_2tuple(feat_size)\n self.qkv = nn.Linear(in_features, embed_dim * 3, bias=qkv_bias)\n self.proj = nn.Linear(embed_dim, out_features)\n self.num_heads = num_heads\n self.head_dim = embed_dim // num_heads\n self.scale = self.head_dim**-0.5\n\n spatial_dim = self.feat_size[0] * self.feat_size[1]\n self.pos_embed = nn.Parameter(torch.zeros(spatial_dim + 1, in_features))\n trunc_normal_(self.pos_embed, std=in_features**-0.5)\n trunc_normal_(self.qkv.weight, std=in_features**-0.5)\n nn.init.zeros_(self.qkv.bias)\n\n def forward(self, x):\n B, _, H, W = x.shape\n N = H * W\n assert self.feat_size[0] == H\n assert self.feat_size[1] == W\n x = x.reshape(B, -1, N).permute(0, 2, 1)\n x = torch.cat([x.mean(1, keepdim=True), x], dim=1)\n x = x + self.pos_embed.unsqueeze(0).to(x.dtype)\n\n x = (\n self.qkv(x)\n .reshape(B, N + 1, 3, self.num_heads, self.head_dim)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = x[0], x[1], x[2]\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\ndef pixel_freq_bands(\n num_bands: int,\n max_freq: float = 224.0,\n linear_bands: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n if linear_bands:\n bands = torch.linspace(1.0, max_freq / 2, num_bands, dtype=dtype, device=device)\n else:\n bands = 2 ** torch.linspace(\n 0, math.log(max_freq, 2) - 1, num_bands, dtype=dtype, device=device\n )\n return bands * torch.pi\n\n\ndef inv_freq_bands(\n num_bands: int,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.pixel_freq_bands","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.pixel_freq_bands#L250-L263","kind":"function","name":"pixel_freq_bands","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":250,"end_line":263,"context_start_line":230,"context_end_line":283,"code":" assert self.feat_size[0] == H\n assert self.feat_size[1] == W\n x = x.reshape(B, -1, N).permute(0, 2, 1)\n x = torch.cat([x.mean(1, keepdim=True), x], dim=1)\n x = x + self.pos_embed.unsqueeze(0).to(x.dtype)\n\n x = (\n self.qkv(x)\n .reshape(B, N + 1, 3, self.num_heads, self.head_dim)\n .permute(2, 0, 3, 1, 4)\n )\n q, k, v = x[0], x[1], x[2]\n attn = (q @ k.transpose(-2, -1)) * self.scale\n attn = attn.softmax(dim=-1)\n\n x = (attn @ v).transpose(1, 2).reshape(B, N + 1, -1)\n x = self.proj(x)\n return x[:, 0]\n\n\ndef pixel_freq_bands(\n num_bands: int,\n max_freq: float = 224.0,\n linear_bands: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n if linear_bands:\n bands = torch.linspace(1.0, max_freq / 2, num_bands, dtype=dtype, device=device)\n else:\n bands = 2 ** torch.linspace(\n 0, math.log(max_freq, 2) - 1, num_bands, dtype=dtype, device=device\n )\n return bands * torch.pi\n\n\ndef inv_freq_bands(\n num_bands: int,\n temperature: float = 100000.0,\n step: int = 2,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n inv_freq = 1.0 / (\n temperature\n ** (torch.arange(0, num_bands, step, dtype=dtype, device=device) / num_bands)\n )\n return inv_freq\n\n\ndef build_sincos2d_pos_embed(\n feat_shape: List[int],\n dim: int = 64,\n temperature: float = 10000.0,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.inv_freq_bands","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.inv_freq_bands#L266-L277","kind":"function","name":"inv_freq_bands","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":266,"end_line":277,"context_start_line":246,"context_end_line":297,"code":" x = self.proj(x)\n return x[:, 0]\n\n\ndef pixel_freq_bands(\n num_bands: int,\n max_freq: float = 224.0,\n linear_bands: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n if linear_bands:\n bands = torch.linspace(1.0, max_freq / 2, num_bands, dtype=dtype, device=device)\n else:\n bands = 2 ** torch.linspace(\n 0, math.log(max_freq, 2) - 1, num_bands, dtype=dtype, device=device\n )\n return bands * torch.pi\n\n\ndef inv_freq_bands(\n num_bands: int,\n temperature: float = 100000.0,\n step: int = 2,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n inv_freq = 1.0 / (\n temperature\n ** (torch.arange(0, num_bands, step, dtype=dtype, device=device) / num_bands)\n )\n return inv_freq\n\n\ndef build_sincos2d_pos_embed(\n feat_shape: List[int],\n dim: int = 64,\n temperature: float = 10000.0,\n reverse_coord: bool = False,\n interleave_sin_cos: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n \"\"\"\n Args:\n feat_shape:\n dim:\n temperature:\n reverse_coord: stack grid order W, H instead of H, W\n interleave_sin_cos: sin, cos, sin, cos stack instead of sin, sin, cos, cos\n dtype:\n device:","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.build_sincos2d_pos_embed","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.build_sincos2d_pos_embed#L280-L326","kind":"function","name":"build_sincos2d_pos_embed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":280,"end_line":326,"context_start_line":260,"context_end_line":346,"code":" bands = 2 ** torch.linspace(\n 0, math.log(max_freq, 2) - 1, num_bands, dtype=dtype, device=device\n )\n return bands * torch.pi\n\n\ndef inv_freq_bands(\n num_bands: int,\n temperature: float = 100000.0,\n step: int = 2,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n inv_freq = 1.0 / (\n temperature\n ** (torch.arange(0, num_bands, step, dtype=dtype, device=device) / num_bands)\n )\n return inv_freq\n\n\ndef build_sincos2d_pos_embed(\n feat_shape: List[int],\n dim: int = 64,\n temperature: float = 10000.0,\n reverse_coord: bool = False,\n interleave_sin_cos: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> torch.Tensor:\n \"\"\"\n Args:\n feat_shape:\n dim:\n temperature:\n reverse_coord: stack grid order W, H instead of H, W\n interleave_sin_cos: sin, cos, sin, cos stack instead of sin, sin, cos, cos\n dtype:\n device:\n Returns:\n \"\"\"\n assert (\n dim % 4 == 0\n ), \"Embed dimension must be divisible by 4 for sin-cos 2D position embedding\"\n pos_dim = dim // 4\n bands = inv_freq_bands(\n pos_dim, temperature=temperature, step=1, dtype=dtype, device=device\n )\n\n if reverse_coord:\n feat_shape = feat_shape[::-1] # stack W, H instead of H, W\n grid = (\n torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n )\n )\n .flatten(1)\n .transpose(0, 1)\n )\n pos2 = grid.unsqueeze(-1) * bands.unsqueeze(0)\n # FIXME add support for unflattened spatial dim?\n\n stack_dim = (\n 2 if interleave_sin_cos else 1\n ) # stack sin, cos, sin, cos instead of sin sin cos cos\n pos_emb = torch.stack([torch.sin(pos2), torch.cos(pos2)], dim=stack_dim).flatten(1)\n return pos_emb\n\n\ndef build_fourier_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n num_bands: int = 64,\n max_res: int = 224,\n linear_bands: bool = False,\n include_grid: bool = False,\n concat_out: bool = True,\n in_pixels: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> List[torch.Tensor]:\n if bands is None:\n if in_pixels:\n bands = pixel_freq_bands(\n num_bands,\n float(max_res),\n linear_bands=linear_bands,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.build_fourier_pos_embed","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.build_fourier_pos_embed#L329-L383","kind":"function","name":"build_fourier_pos_embed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":329,"end_line":383,"context_start_line":309,"context_end_line":403,"code":" feat_shape = feat_shape[::-1] # stack W, H instead of H, W\n grid = (\n torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n )\n )\n .flatten(1)\n .transpose(0, 1)\n )\n pos2 = grid.unsqueeze(-1) * bands.unsqueeze(0)\n # FIXME add support for unflattened spatial dim?\n\n stack_dim = (\n 2 if interleave_sin_cos else 1\n ) # stack sin, cos, sin, cos instead of sin sin cos cos\n pos_emb = torch.stack([torch.sin(pos2), torch.cos(pos2)], dim=stack_dim).flatten(1)\n return pos_emb\n\n\ndef build_fourier_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n num_bands: int = 64,\n max_res: int = 224,\n linear_bands: bool = False,\n include_grid: bool = False,\n concat_out: bool = True,\n in_pixels: bool = True,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n) -> List[torch.Tensor]:\n if bands is None:\n if in_pixels:\n bands = pixel_freq_bands(\n num_bands,\n float(max_res),\n linear_bands=linear_bands,\n dtype=dtype,\n device=device,\n )\n else:\n bands = inv_freq_bands(num_bands, step=1, dtype=dtype, device=device)\n else:\n if device is None:\n device = bands.device\n if dtype is None:\n dtype = bands.dtype\n\n if in_pixels:\n grid = torch.stack(\n torch.meshgrid(\n [\n torch.linspace(-1.0, 1.0, steps=s, device=device, dtype=dtype)\n for s in feat_shape\n ]\n ),\n dim=-1,\n )\n else:\n grid = torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n ),\n dim=-1,\n )\n grid = grid.unsqueeze(-1)\n pos = grid * bands\n\n pos_sin, pos_cos = pos.sin(), pos.cos()\n out = (grid, pos_sin, pos_cos) if include_grid else (pos_sin, pos_cos)\n # FIXME torchscript doesn't like multiple return types, probably need to always cat?\n if concat_out:\n out = torch.cat(out, dim=-1)\n return out\n\n\nclass FourierEmbed(nn.Module):\n def __init__(\n self,\n max_res: int = 224,\n num_bands: int = 64,\n concat_grid=True,\n keep_spatial=False,\n ):\n super().__init__()\n self.max_res = max_res\n self.num_bands = num_bands\n self.concat_grid = concat_grid\n self.keep_spatial = keep_spatial\n self.register_buffer(\n \"bands\", pixel_freq_bands(max_res, num_bands), persistent=False\n )\n\n def forward(self, x):","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.FourierEmbed","uri":"program://CREMA/class/lavis.models.clip_models.timm_model.FourierEmbed#L386-L427","kind":"class","name":"FourierEmbed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":386,"end_line":427,"context_start_line":366,"context_end_line":447,"code":" dim=-1,\n )\n else:\n grid = torch.stack(\n torch.meshgrid(\n [torch.arange(s, device=device, dtype=dtype) for s in feat_shape]\n ),\n dim=-1,\n )\n grid = grid.unsqueeze(-1)\n pos = grid * bands\n\n pos_sin, pos_cos = pos.sin(), pos.cos()\n out = (grid, pos_sin, pos_cos) if include_grid else (pos_sin, pos_cos)\n # FIXME torchscript doesn't like multiple return types, probably need to always cat?\n if concat_out:\n out = torch.cat(out, dim=-1)\n return out\n\n\nclass FourierEmbed(nn.Module):\n def __init__(\n self,\n max_res: int = 224,\n num_bands: int = 64,\n concat_grid=True,\n keep_spatial=False,\n ):\n super().__init__()\n self.max_res = max_res\n self.num_bands = num_bands\n self.concat_grid = concat_grid\n self.keep_spatial = keep_spatial\n self.register_buffer(\n \"bands\", pixel_freq_bands(max_res, num_bands), persistent=False\n )\n\n def forward(self, x):\n B, C = x.shape[:2]\n feat_shape = x.shape[2:]\n emb = build_fourier_pos_embed(\n feat_shape,\n self.bands,\n include_grid=self.concat_grid,\n dtype=x.dtype,\n device=x.device,\n )\n emb = emb.transpose(-1, -2).flatten(len(feat_shape))\n batch_expand = (B,) + (-1,) * (x.ndim - 1)\n\n # FIXME support nD\n if self.keep_spatial:\n x = torch.cat(\n [x, emb.unsqueeze(0).expand(batch_expand).permute(0, 3, 1, 2)], dim=1\n )\n else:\n x = torch.cat(\n [x.permute(0, 2, 3, 1), emb.unsqueeze(0).expand(batch_expand)], dim=-1\n )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.rot","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.rot#L430-L431","kind":"function","name":"rot","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":430,"end_line":431,"context_start_line":410,"context_end_line":451,"code":" dtype=x.dtype,\n device=x.device,\n )\n emb = emb.transpose(-1, -2).flatten(len(feat_shape))\n batch_expand = (B,) + (-1,) * (x.ndim - 1)\n\n # FIXME support nD\n if self.keep_spatial:\n x = torch.cat(\n [x, emb.unsqueeze(0).expand(batch_expand).permute(0, 3, 1, 2)], dim=1\n )\n else:\n x = torch.cat(\n [x.permute(0, 2, 3, 1), emb.unsqueeze(0).expand(batch_expand)], dim=-1\n )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.apply_rot_embed","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.apply_rot_embed#L434-L435","kind":"function","name":"apply_rot_embed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":434,"end_line":435,"context_start_line":414,"context_end_line":455,"code":" batch_expand = (B,) + (-1,) * (x.ndim - 1)\n\n # FIXME support nD\n if self.keep_spatial:\n x = torch.cat(\n [x, emb.unsqueeze(0).expand(batch_expand).permute(0, 3, 1, 2)], dim=1\n )\n else:\n x = torch.cat(\n [x.permute(0, 2, 3, 1), emb.unsqueeze(0).expand(batch_expand)], dim=-1\n )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n dim: int = 64,\n max_freq: float = 224,\n linear_bands: bool = False,\n dtype: torch.dtype = torch.float32,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.apply_rot_embed_list","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.apply_rot_embed_list#L438-L441","kind":"function","name":"apply_rot_embed_list","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":438,"end_line":441,"context_start_line":418,"context_end_line":461,"code":" x = torch.cat(\n [x, emb.unsqueeze(0).expand(batch_expand).permute(0, 3, 1, 2)], dim=1\n )\n else:\n x = torch.cat(\n [x.permute(0, 2, 3, 1), emb.unsqueeze(0).expand(batch_expand)], dim=-1\n )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n dim: int = 64,\n max_freq: float = 224,\n linear_bands: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n \"\"\"\n NOTE: shape arg should include spatial dim only\n \"\"\"\n feat_shape = torch.Size(feat_shape)","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.apply_rot_embed_split","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.apply_rot_embed_split#L444-L446","kind":"function","name":"apply_rot_embed_split","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":444,"end_line":446,"context_start_line":424,"context_end_line":466,"code":" )\n x = x.reshape(B, feat_shape.numel(), -1)\n\n return x\n\n\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n dim: int = 64,\n max_freq: float = 224,\n linear_bands: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n \"\"\"\n NOTE: shape arg should include spatial dim only\n \"\"\"\n feat_shape = torch.Size(feat_shape)\n\n sin_emb, cos_emb = build_fourier_pos_embed(\n feat_shape,\n bands=bands,\n num_bands=dim // 4,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.build_rotary_pos_embed","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.build_rotary_pos_embed#L449-L476","kind":"function","name":"build_rotary_pos_embed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":449,"end_line":476,"context_start_line":429,"context_end_line":496,"code":"\ndef rot(x):\n return torch.stack([-x[..., 1::2], x[..., ::2]], -1).reshape(x.shape)\n\n\ndef apply_rot_embed(x: torch.Tensor, sin_emb, cos_emb):\n return x * cos_emb + rot(x) * sin_emb\n\n\ndef apply_rot_embed_list(x: List[torch.Tensor], sin_emb, cos_emb):\n if isinstance(x, torch.Tensor):\n x = [x]\n return [t * cos_emb + rot(t) * sin_emb for t in x]\n\n\ndef apply_rot_embed_split(x: torch.Tensor, emb):\n split = emb.shape[-1] // 2\n return x * emb[:, :split] + rot(x) * emb[:, split:]\n\n\ndef build_rotary_pos_embed(\n feat_shape: List[int],\n bands: Optional[torch.Tensor] = None,\n dim: int = 64,\n max_freq: float = 224,\n linear_bands: bool = False,\n dtype: torch.dtype = torch.float32,\n device: Optional[torch.device] = None,\n):\n \"\"\"\n NOTE: shape arg should include spatial dim only\n \"\"\"\n feat_shape = torch.Size(feat_shape)\n\n sin_emb, cos_emb = build_fourier_pos_embed(\n feat_shape,\n bands=bands,\n num_bands=dim // 4,\n max_res=max_freq,\n linear_bands=linear_bands,\n concat_out=False,\n device=device,\n dtype=dtype,\n )\n N = feat_shape.numel()\n sin_emb = sin_emb.reshape(N, -1).repeat_interleave(2, -1)\n cos_emb = cos_emb.reshape(N, -1).repeat_interleave(2, -1)\n return sin_emb, cos_emb\n\n\nclass RotaryEmbedding(nn.Module):\n \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.RotaryEmbedding","uri":"program://CREMA/class/lavis.models.clip_models.timm_model.RotaryEmbedding#L479-L503","kind":"class","name":"RotaryEmbedding","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":479,"end_line":503,"context_start_line":459,"context_end_line":523,"code":" NOTE: shape arg should include spatial dim only\n \"\"\"\n feat_shape = torch.Size(feat_shape)\n\n sin_emb, cos_emb = build_fourier_pos_embed(\n feat_shape,\n bands=bands,\n num_bands=dim // 4,\n max_res=max_freq,\n linear_bands=linear_bands,\n concat_out=False,\n device=device,\n dtype=dtype,\n )\n N = feat_shape.numel()\n sin_emb = sin_emb.reshape(N, -1).repeat_interleave(2, -1)\n cos_emb = cos_emb.reshape(N, -1).repeat_interleave(2, -1)\n return sin_emb, cos_emb\n\n\nclass RotaryEmbedding(nn.Module):\n \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model._no_grad_trunc_normal_","uri":"program://CREMA/function/lavis.models.clip_models.timm_model._no_grad_trunc_normal_#L506-L541","kind":"function","name":"_no_grad_trunc_normal_","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":506,"end_line":541,"context_start_line":486,"context_end_line":561,"code":" \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.trunc_normal_","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.trunc_normal_#L544-L561","kind":"function","name":"trunc_normal_","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":544,"end_line":561,"context_start_line":524,"context_end_line":561,"code":" l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated\n # standard normal\n tensor.erfinv_()\n\n # Transform to proper mean, std\n tensor.mul_(std * math.sqrt(2.0))\n tensor.add_(mean)\n\n # Clamp to ensure it's in the proper range\n tensor.clamp_(min=a, max=b)\n return tensor\n\n\ndef trunc_normal_(tensor, mean=0.0, std=1.0, a=-2.0, b=2.0):\n r\"\"\"Fills the input Tensor with values drawn from a truncated\n normal distribution. The values are effectively drawn from the\n normal distribution :math:`\\mathcal{N}(\\text{mean}, \\text{std}^2)`\n with values outside :math:`[a, b]` redrawn until they are within\n the bounds. The method used for generating the random values works\n best when :math:`a \\leq \\text{mean} \\leq b`.\n Args:\n tensor: an n-dimensional `torch.Tensor`\n mean: the mean of the normal distribution\n std: the standard deviation of the normal distribution\n a: the minimum cutoff value\n b: the maximum cutoff value\n Examples:\n >>> w = torch.empty(3, 5)\n >>> nn.init.trunc_normal_(w)\n \"\"\"\n return _no_grad_trunc_normal_(tensor, mean, std, a, b)","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.__init__","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.__init__#L488-L495","kind":"function","name":"__init__","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":488,"end_line":495,"context_start_line":468,"context_end_line":515,"code":" linear_bands=linear_bands,\n concat_out=False,\n device=device,\n dtype=dtype,\n )\n N = feat_shape.numel()\n sin_emb = sin_emb.reshape(N, -1).repeat_interleave(2, -1)\n cos_emb = cos_emb.reshape(N, -1).repeat_interleave(2, -1)\n return sin_emb, cos_emb\n\n\nclass RotaryEmbedding(nn.Module):\n \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.lock","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.lock#L91-L122","kind":"function","name":"lock","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":91,"end_line":122,"context_start_line":71,"context_end_line":142,"code":" if pool == \"abs_attn\":\n head_layers[\"pool\"] = AttentionPool2d(\n prev_chs, feat_size=feat_size, out_features=embed_dim\n )\n prev_chs = embed_dim\n elif pool == \"rot_attn\":\n head_layers[\"pool\"] = RotAttentionPool2d(prev_chs, out_features=embed_dim)\n prev_chs = embed_dim\n else:\n assert proj, \"projection layer needed if non-attention pooling is used.\"\n\n # NOTE attention pool ends with a projection layer, so proj should usually be set to '' if such pooling is used\n if proj == \"linear\":\n head_layers[\"drop\"] = nn.Dropout(drop)\n head_layers[\"proj\"] = nn.Linear(prev_chs, embed_dim)\n elif proj == \"mlp\":\n head_layers[\"mlp\"] = Mlp(prev_chs, 2 * embed_dim, embed_dim, drop=drop)\n\n self.head = nn.Sequential(head_layers)\n\n def lock(self, unlocked_groups=0, freeze_bn_stats=False):\n \"\"\"lock modules\n Args:\n unlocked_groups (int): leave last n layer groups unlocked (default: 0)\n \"\"\"\n if not unlocked_groups:\n # lock full model\n for param in self.trunk.parameters():\n param.requires_grad = False\n if freeze_bn_stats:\n freeze_batch_norm_2d(self.trunk)\n else:\n # NOTE: partial freeze requires latest timm (master) branch and is subject to change\n try:\n # FIXME import here until API stable and in an official release\n from timm.models.helpers import group_modules, group_parameters\n except ImportError:\n raise RuntimeError(\n \"Please install latest timm `pip install git+https://github.com/rwightman/pytorch-image-models`\"\n )\n matcher = self.trunk.group_matcher()\n gparams = group_parameters(self.trunk, matcher)\n max_layer_id = max(gparams.keys())\n max_layer_id = max_layer_id - unlocked_groups\n for group_idx in range(max_layer_id + 1):\n group = gparams[group_idx]\n for param in group:\n self.trunk.get_parameter(param).requires_grad = False\n if freeze_bn_stats:\n gmodules = group_modules(self.trunk, matcher, reverse=True)\n gmodules = {k for k, v in gmodules.items() if v <= max_layer_id}\n freeze_batch_norm_2d(self.trunk, gmodules)\n\n def forward(self, x):\n x = self.trunk(x)\n x = self.head(x)\n return x\n\n\nclass RotAttentionPool2d(nn.Module):\n \"\"\"Attention based 2D feature pooling w/ rotary (relative) pos embedding.\n This is a multi-head attention based replacement for (spatial) average pooling in NN architectures.\n Adapted from the AttentionPool2d in CLIP w/ rotary embedding instead of learned embed.\n https://github.com/openai/CLIP/blob/3b473b0e682c091a9e53623eebc1ca1657385717/clip/model.py\n NOTE: While this impl does not require a fixed feature size, performance at differeing resolutions from\n train varies widely and falls off dramatically. I'm not sure if there is a way around this... -RW\n \"\"\"\n\n def __init__(\n self,\n in_features: int,\n out_features: int = None,","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.forward","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.forward#L500-L503","kind":"function","name":"forward","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":500,"end_line":503,"context_start_line":480,"context_end_line":523,"code":" \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.get_embed","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.get_embed#L497-L498","kind":"function","name":"get_embed","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":497,"end_line":498,"context_start_line":477,"context_end_line":518,"code":"\n\nclass RotaryEmbedding(nn.Module):\n \"\"\"Rotary position embedding\n NOTE: This is my initial attempt at impl rotary embedding for spatial use, it has not\n been well tested, and will likely change. It will be moved to its own file.\n The following impl/resources were referenced for this impl:\n * https://github.com/lucidrains/vit-pytorch/blob/6f3a5fcf0bca1c5ec33a35ef48d97213709df4ba/vit_pytorch/rvt.py\n * https://blog.eleuther.ai/rotary-embeddings/\n \"\"\"\n\n def __init__(self, dim, max_res=224, linear_bands: bool = False):\n super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.clip_models.timm_model.norm_cdf","uri":"program://CREMA/function/lavis.models.clip_models.timm_model.norm_cdf#L509-L511","kind":"function","name":"norm_cdf","path":"lavis/models/clip_models/timm_model.py","language":"python","start_line":509,"end_line":511,"context_start_line":489,"context_end_line":531,"code":" super().__init__()\n self.dim = dim\n self.register_buffer(\n \"bands\",\n pixel_freq_bands(dim // 4, max_res, linear_bands=linear_bands),\n persistent=False,\n )\n\n def get_embed(self, shape: List[int]):\n return build_rotary_pos_embed(shape, self.bands)\n\n def forward(self, x):\n # assuming channel-first tensor where spatial dim are >= 2\n sin_emb, cos_emb = self.get_embed(x.shape[2:])\n return apply_rot_embed(x, sin_emb, cos_emb)\n\n\ndef _no_grad_trunc_normal_(tensor, mean, std, a, b):\n # Cut & paste from PyTorch official master until it's in a few official releases - RW\n # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf\n def norm_cdf(x):\n # Computes standard normal cumulative distribution function\n return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0\n\n if (mean < a - 2 * std) or (mean > b + 2 * std):\n warnings.warn(\n \"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. \"\n \"The distribution of values may be incorrect.\",\n stacklevel=2,\n )\n\n with torch.no_grad():\n # Values are generated by using a truncated uniform distribution and\n # then using the inverse CDF for the normal distribution.\n # Get upper and lower cdf values\n l = norm_cdf((a - mean) / std)\n u = norm_cdf((b - mean) / std)\n\n # Uniformly fill tensor with values from [l, u], then translate to\n # [2l-1, 2u-1].\n tensor.uniform_(2 * l - 1, 2 * u - 1)\n\n # Use inverse cdf transform for normal distribution to get truncated","source_hash":"a082e50941fc3895593324e822cb819c0e741eb1bb05eb3cc35c6076b2984180","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.gpt_models.gpt_dialogue","uri":"program://CREMA/module/lavis.models.gpt_models.gpt_dialogue#L1-L110","kind":"module","name":"lavis.models.gpt_models.gpt_dialogue","path":"lavis/models/gpt_models/gpt_dialogue.py","language":"python","start_line":1,"end_line":110,"context_start_line":1,"context_end_line":110,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom torch.nn import CrossEntropyLoss, MSELoss\nfrom transformers import GPT2LMHeadModel\nfrom transformers.modeling_outputs import CausalLMOutputWithCrossAttentions\n\n\n@registry.register_model(\"gpt_dialogue\")\nclass GPTDialogue(BaseModel, GPT2LMHeadModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/gpt_dialogue_base.yaml\"}\n\n def __init__(self, config, len_video_ft=4224):\n\n super().__init__(config)\n\n self.video_ff = nn.Linear(len_video_ft, config.n_embd)\n self.video_ff_out = nn.Linear(config.n_embd, len_video_ft)\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def forward(\n self,\n samples,\n past_key_values=None,\n position_ids=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n ):\n\n input_embs = self.transformer.wte(samples[\"input_ids\"])\n video_embs = self.video_ff(samples[\"video_fts\"])\n input_embs = torch.cat([video_embs, input_embs], dim=1)\n\n transformer_outputs = self.transformer(\n attention_mask=samples[\"attn_mask\"],\n token_type_ids=samples[\"token_type_ids\"],\n inputs_embeds=input_embs,\n position_ids=position_ids,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n hidden_states = transformer_outputs[0]\n\n lm_logits = self.lm_head(hidden_states)\n\n loss = None\n if samples[\"labels\"] is not None:\n # Shift so that tokens < n predict n\n shift_logits = lm_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"labels\"][..., 1:].contiguous()\n # Flatten the tokens\n loss_fct = CrossEntropyLoss(ignore_index=-1)\n loss = loss_fct(\n shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)\n )\n\n if samples[\"video_fts\"] is not None:\n len_video_fts = samples[\"video_fts\"].shape[1]\n video_logits = self.video_ff_out(hidden_states[:, :len_video_fts, :])\n # Shift so that tokens < n predict n\n shift_logits = video_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"video_fts\"][..., 1:, :].contiguous()\n # Flatten the tokens\n loss_fct = MSELoss(reduction=\"mean\")\n video_loss = loss_fct(shift_logits, shift_labels)\n\n if loss is not None:\n loss = loss + video_loss\n else:\n loss = video_loss\n\n return CausalLMOutputWithCrossAttentions(\n loss=loss,\n logits=lm_logits,\n past_key_values=transformer_outputs.past_key_values,\n hidden_states=transformer_outputs.hidden_states,\n attentions=transformer_outputs.attentions,\n cross_attentions=transformer_outputs.cross_attentions,\n )\n\n @classmethod\n def from_config(cls, cfg):\n model = cls.__bases__[1].from_pretrained(\"gpt2\")\n model.resize_token_embeddings(cfg[\"len_tokenizer\"])\n return model","source_hash":"3416fbd30995b50f242294c72b522ed172b5ada7bc33a7c6ee8feb6dc12be0f7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.gpt_models.gpt_dialogue.GPTDialogue","uri":"program://CREMA/class/lavis.models.gpt_models.gpt_dialogue.GPTDialogue#L18-L110","kind":"class","name":"GPTDialogue","path":"lavis/models/gpt_models/gpt_dialogue.py","language":"python","start_line":18,"end_line":110,"context_start_line":1,"context_end_line":110,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom torch.nn import CrossEntropyLoss, MSELoss\nfrom transformers import GPT2LMHeadModel\nfrom transformers.modeling_outputs import CausalLMOutputWithCrossAttentions\n\n\n@registry.register_model(\"gpt_dialogue\")\nclass GPTDialogue(BaseModel, GPT2LMHeadModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/gpt_dialogue_base.yaml\"}\n\n def __init__(self, config, len_video_ft=4224):\n\n super().__init__(config)\n\n self.video_ff = nn.Linear(len_video_ft, config.n_embd)\n self.video_ff_out = nn.Linear(config.n_embd, len_video_ft)\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def forward(\n self,\n samples,\n past_key_values=None,\n position_ids=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n ):\n\n input_embs = self.transformer.wte(samples[\"input_ids\"])\n video_embs = self.video_ff(samples[\"video_fts\"])\n input_embs = torch.cat([video_embs, input_embs], dim=1)\n\n transformer_outputs = self.transformer(\n attention_mask=samples[\"attn_mask\"],\n token_type_ids=samples[\"token_type_ids\"],\n inputs_embeds=input_embs,\n position_ids=position_ids,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n hidden_states = transformer_outputs[0]\n\n lm_logits = self.lm_head(hidden_states)\n\n loss = None\n if samples[\"labels\"] is not None:\n # Shift so that tokens < n predict n\n shift_logits = lm_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"labels\"][..., 1:].contiguous()\n # Flatten the tokens\n loss_fct = CrossEntropyLoss(ignore_index=-1)\n loss = loss_fct(\n shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)\n )\n\n if samples[\"video_fts\"] is not None:\n len_video_fts = samples[\"video_fts\"].shape[1]\n video_logits = self.video_ff_out(hidden_states[:, :len_video_fts, :])\n # Shift so that tokens < n predict n\n shift_logits = video_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"video_fts\"][..., 1:, :].contiguous()\n # Flatten the tokens\n loss_fct = MSELoss(reduction=\"mean\")\n video_loss = loss_fct(shift_logits, shift_labels)\n\n if loss is not None:\n loss = loss + video_loss\n else:\n loss = video_loss\n\n return CausalLMOutputWithCrossAttentions(\n loss=loss,\n logits=lm_logits,\n past_key_values=transformer_outputs.past_key_values,\n hidden_states=transformer_outputs.hidden_states,\n attentions=transformer_outputs.attentions,\n cross_attentions=transformer_outputs.cross_attentions,\n )\n\n @classmethod\n def from_config(cls, cfg):\n model = cls.__bases__[1].from_pretrained(\"gpt2\")\n model.resize_token_embeddings(cfg[\"len_tokenizer\"])\n return model","source_hash":"3416fbd30995b50f242294c72b522ed172b5ada7bc33a7c6ee8feb6dc12be0f7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.gpt_models.gpt_dialogue.__init__","uri":"program://CREMA/function/lavis.models.gpt_models.gpt_dialogue.__init__#L22-L34","kind":"function","name":"__init__","path":"lavis/models/gpt_models/gpt_dialogue.py","language":"python","start_line":22,"end_line":34,"context_start_line":2,"context_end_line":54,"code":" Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport torch.nn as nn\nfrom lavis.common.registry import registry\nfrom lavis.models.base_model import BaseModel\nfrom torch.nn import CrossEntropyLoss, MSELoss\nfrom transformers import GPT2LMHeadModel\nfrom transformers.modeling_outputs import CausalLMOutputWithCrossAttentions\n\n\n@registry.register_model(\"gpt_dialogue\")\nclass GPTDialogue(BaseModel, GPT2LMHeadModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/gpt_dialogue_base.yaml\"}\n\n def __init__(self, config, len_video_ft=4224):\n\n super().__init__(config)\n\n self.video_ff = nn.Linear(len_video_ft, config.n_embd)\n self.video_ff_out = nn.Linear(config.n_embd, len_video_ft)\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def forward(\n self,\n samples,\n past_key_values=None,\n position_ids=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n ):\n\n input_embs = self.transformer.wte(samples[\"input_ids\"])\n video_embs = self.video_ff(samples[\"video_fts\"])\n input_embs = torch.cat([video_embs, input_embs], dim=1)\n\n transformer_outputs = self.transformer(","source_hash":"3416fbd30995b50f242294c72b522ed172b5ada7bc33a7c6ee8feb6dc12be0f7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.gpt_models.gpt_dialogue.forward","uri":"program://CREMA/function/lavis.models.gpt_models.gpt_dialogue.forward#L36-L104","kind":"function","name":"forward","path":"lavis/models/gpt_models/gpt_dialogue.py","language":"python","start_line":36,"end_line":104,"context_start_line":16,"context_end_line":110,"code":"\n@registry.register_model(\"gpt_dialogue\")\nclass GPTDialogue(BaseModel, GPT2LMHeadModel):\n\n PRETRAINED_MODEL_CONFIG_DICT = {\"base\": \"configs/models/gpt_dialogue_base.yaml\"}\n\n def __init__(self, config, len_video_ft=4224):\n\n super().__init__(config)\n\n self.video_ff = nn.Linear(len_video_ft, config.n_embd)\n self.video_ff_out = nn.Linear(config.n_embd, len_video_ft)\n\n # Model parallel\n self.model_parallel = False\n self.device_map = None\n\n # Initialize weights and apply final processing\n self.post_init()\n\n def forward(\n self,\n samples,\n past_key_values=None,\n position_ids=None,\n head_mask=None,\n encoder_hidden_states=None,\n encoder_attention_mask=None,\n use_cache=None,\n output_attentions=None,\n output_hidden_states=None,\n return_dict=None,\n ):\n\n input_embs = self.transformer.wte(samples[\"input_ids\"])\n video_embs = self.video_ff(samples[\"video_fts\"])\n input_embs = torch.cat([video_embs, input_embs], dim=1)\n\n transformer_outputs = self.transformer(\n attention_mask=samples[\"attn_mask\"],\n token_type_ids=samples[\"token_type_ids\"],\n inputs_embeds=input_embs,\n position_ids=position_ids,\n head_mask=head_mask,\n encoder_hidden_states=encoder_hidden_states,\n encoder_attention_mask=encoder_attention_mask,\n use_cache=use_cache,\n output_attentions=output_attentions,\n output_hidden_states=output_hidden_states,\n return_dict=return_dict,\n )\n hidden_states = transformer_outputs[0]\n\n lm_logits = self.lm_head(hidden_states)\n\n loss = None\n if samples[\"labels\"] is not None:\n # Shift so that tokens < n predict n\n shift_logits = lm_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"labels\"][..., 1:].contiguous()\n # Flatten the tokens\n loss_fct = CrossEntropyLoss(ignore_index=-1)\n loss = loss_fct(\n shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)\n )\n\n if samples[\"video_fts\"] is not None:\n len_video_fts = samples[\"video_fts\"].shape[1]\n video_logits = self.video_ff_out(hidden_states[:, :len_video_fts, :])\n # Shift so that tokens < n predict n\n shift_logits = video_logits[..., :-1, :].contiguous()\n shift_labels = samples[\"video_fts\"][..., 1:, :].contiguous()\n # Flatten the tokens\n loss_fct = MSELoss(reduction=\"mean\")\n video_loss = loss_fct(shift_logits, shift_labels)\n\n if loss is not None:\n loss = loss + video_loss\n else:\n loss = video_loss\n\n return CausalLMOutputWithCrossAttentions(\n loss=loss,\n logits=lm_logits,\n past_key_values=transformer_outputs.past_key_values,\n hidden_states=transformer_outputs.hidden_states,\n attentions=transformer_outputs.attentions,\n cross_attentions=transformer_outputs.cross_attentions,\n )\n\n @classmethod\n def from_config(cls, cfg):\n model = cls.__bases__[1].from_pretrained(\"gpt2\")\n model.resize_token_embeddings(cfg[\"len_tokenizer\"])\n return model","source_hash":"3416fbd30995b50f242294c72b522ed172b5ada7bc33a7c6ee8feb6dc12be0f7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.gpt_models.gpt_dialogue.from_config","uri":"program://CREMA/function/lavis.models.gpt_models.gpt_dialogue.from_config#L107-L110","kind":"function","name":"from_config","path":"lavis/models/gpt_models/gpt_dialogue.py","language":"python","start_line":107,"end_line":110,"context_start_line":87,"context_end_line":110,"code":" shift_labels = samples[\"video_fts\"][..., 1:, :].contiguous()\n # Flatten the tokens\n loss_fct = MSELoss(reduction=\"mean\")\n video_loss = loss_fct(shift_logits, shift_labels)\n\n if loss is not None:\n loss = loss + video_loss\n else:\n loss = video_loss\n\n return CausalLMOutputWithCrossAttentions(\n loss=loss,\n logits=lm_logits,\n past_key_values=transformer_outputs.past_key_values,\n hidden_states=transformer_outputs.hidden_states,\n attentions=transformer_outputs.attentions,\n cross_attentions=transformer_outputs.cross_attentions,\n )\n\n @classmethod\n def from_config(cls, cfg):\n model = cls.__bases__[1].from_pretrained(\"gpt2\")\n model.resize_token_embeddings(cfg[\"len_tokenizer\"])\n return model","source_hash":"3416fbd30995b50f242294c72b522ed172b5ada7bc33a7c6ee8feb6dc12be0f7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval","uri":"program://CREMA/module/lavis.models.alpro_models.alpro_retrieval#L1-L422","kind":"module","name":"lavis.models.alpro_models.alpro_retrieval","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":1,"end_line":422,"context_start_line":1,"context_end_line":422,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport datetime\nimport logging\nimport time\n\nimport lavis.common.dist_utils as dist_utils\nimport numpy as np\nimport torch\nimport torch.distributed as dist\nimport torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.dist_utils import get_rank\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import AlproIntermediateOutput, AlproOutput\nfrom lavis.models.base_model import all_gather_with_grad\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_retrieval\")\nclass AlproRetrieval(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_retrieval_msrvtt.yaml\",\n \"didemo\": \"configs/models/alpro_retrieval_didemo.yaml\",\n }\n\n def __init__(\n self,\n visual_encoder,\n text_encoder,\n vision_width=768,\n text_width=768,\n embed_dim=256,\n max_txt_len=35,\n temp=0.07,\n ):\n super().__init__()\n\n self.temp = nn.Parameter(torch.ones([]) * temp)\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n self.text_encoder = text_encoder\n\n vision_width = vision_width\n text_width = text_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n visual_inputs = samples[\"video\"]\n caption = samples[\"text_input\"]\n\n b, t, c, h, w = visual_inputs.shape\n\n # forward text\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_feat = F.normalize(self.vision_proj(video_embeds[:, 0, :]), dim=-1)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # ========== (in-batch) ITC loss ==========\n gathered_video_feats = all_gather_with_grad(video_feat)\n gathered_text_feats = all_gather_with_grad(text_feat)\n\n sim_v2t = video_feat @ gathered_text_feats.t() / self.temp\n sim_t2v = text_feat @ gathered_video_feats.t() / self.temp\n\n sim_targets = torch.zeros_like(sim_v2t)\n\n local_rank = get_rank()\n b_start, b_end = b * local_rank, b * (local_rank + 1)\n sim_targets[:, b_start:b_end] = torch.eye(b)\n\n loss_v2t = -torch.sum(F.log_softmax(sim_v2t, dim=1) * sim_targets, dim=1).mean()\n loss_t2v = -torch.sum(F.log_softmax(sim_t2v, dim=1) * sim_targets, dim=1).mean()\n\n vtc_loss = (loss_v2t + loss_t2v) / 2\n\n (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_output,\n encoder_output_neg,\n ) = self.compute_vtm(\n text_embeds=text_embeds,\n text_atts=text.attention_mask,\n image_embeds=video_embeds,\n image_atts=video_atts,\n sim_i2t=sim_v2t.clone(), # for hard mining\n sim_t2i=sim_t2v.clone(), # for hard mining\n )\n\n loss = vtc_loss + vtm_loss\n\n # return {\"loss\": loss}\n return AlproOutput(\n loss=loss,\n loss_vtc=vtc_loss,\n loss_vtm=vtm_loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n encoder_output_neg=encoder_output_neg,\n vtm_logits=vtm_logits,\n vtm_labels=vtm_labels,\n ),\n )\n\n def compute_vtm(\n self, text_embeds, text_atts, image_embeds, image_atts, sim_i2t, sim_t2i\n ):\n device = self.device\n\n # ====== positive pairs =======\n attention_mask = torch.cat([text_atts, image_atts], dim=1)\n embedding_output_pos = torch.cat([text_embeds, image_embeds], dim=1)\n\n encoder_outputs_pos = self.text_encoder(\n encoder_embeds=embedding_output_pos,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n # ====== negative pairs =======\n bs = text_embeds.shape[0]\n\n local_rank = get_rank()\n b_start, b_end = bs * local_rank, bs * (local_rank + 1)\n\n with torch.no_grad():\n weights_v2t = sim_i2t[:, b_start:b_end]\n weights_t2v = sim_t2i[:, b_start:b_end]\n\n # never select self as negative\n weights_v2t.fill_diagonal_(-np.Inf)\n weights_t2v.fill_diagonal_(-np.Inf)\n\n weights_v2t = F.softmax(weights_v2t, dim=1)\n weights_t2v = F.softmax(weights_t2v, dim=1)\n\n # select a negative image for each text\n # FIXME to optimize using indexing operations\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2v[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_v2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text_atts[neg_idx])\n\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text_atts, text_atts_neg], dim=0)\n\n video_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n video_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n attention_mask_all = torch.cat([text_atts_all, video_atts_all], dim=1)\n embedding_output_all = torch.cat([text_embeds_all, video_embeds_all], dim=1)\n\n # forward negative pairs via cross encoder\n encoder_outputs_neg = self.text_encoder(\n encoder_embeds=embedding_output_all,\n attention_mask=attention_mask_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_outputs_pos.last_hidden_state[:, 0, :],\n encoder_outputs_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n vtm_logits = self.itm_head(vl_embeddings)\n\n vtm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(device)\n vtm_loss = F.cross_entropy(vtm_logits, vtm_labels)\n\n return (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_outputs_pos,\n encoder_outputs_neg,\n )\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n k_test = task_cfg.get(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_feats = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text_output = self.text_encoder.forward_text(\n text_input,\n token_type_ids=torch.zeros(\n text_input.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_feats.append(text_output.last_hidden_state.cpu())\n text_embed = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :])\n )\n text_embeds.append(text_embed)\n text_ids.append(text_input.input_ids)\n text_atts.append(text_input.attention_mask)\n\n text_embeds = torch.cat(text_embeds, dim=0)\n text_ids = torch.cat(text_ids, dim=0)\n text_atts = torch.cat(text_atts, dim=0)\n text_feats = torch.cat(text_feats, dim=0)\n\n video_feats = []\n video_embeds = []\n for samples in data_loader:\n video = samples[\"video\"]\n\n video = video.to(self.device)\n video_feat = self.visual_encoder.forward_features(video)\n video_embed = self.vision_proj(video_feat[:, 0, :])\n video_embed = F.normalize(video_embed, dim=-1)\n\n video_feats.append(video_feat.cpu())\n video_embeds.append(video_embed)\n\n video_feats = torch.cat(video_feats, dim=0)\n video_embeds = torch.cat(video_embeds, dim=0)\n\n sims_matrix = video_embeds @ text_embeds.t()\n score_matrix_v2t = torch.full(\n (len(data_loader.dataset.image), len(texts)), -100.0\n ).to(self.device)\n\n num_tasks = dist_utils.get_world_size()\n rank = dist_utils.get_rank()\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n # video-to-text\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n video_feats_repeat = (\n video_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n video_atts_repeat = torch.ones(\n video_feats_repeat.size()[:-1], dtype=torch.long\n ).to(self.device)\n\n attention_mask = torch.cat([text_atts[topk_idx], video_atts_repeat], dim=1)\n embedding_output = torch.cat(\n [text_feats[topk_idx].to(self.device), video_feats_repeat], dim=1\n )\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_v2t[start + i, topk_idx] = score + topk_sim\n\n # text-to-video\n sims_matrix = sims_matrix.t()\n score_matrix_t2v = torch.full(\n (len(texts), len(data_loader.dataset.image)), -100.0\n ).to(self.device)\n\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n text_feats_repeat = (\n text_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n text_atts_repeat = text_atts[start + i].repeat(k_test, 1).to(self.device)\n\n video_atts = torch.ones(\n video_feats[topk_idx].size()[:-1], dtype=torch.long\n ).to(self.device)\n\n embedding_output = torch.cat(\n [text_feats_repeat, video_feats[topk_idx].to(self.device)], dim=1\n )\n attention_mask = torch.cat([text_atts_repeat, video_atts], dim=1)\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_t2v[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.barrier()\n torch.distributed.all_reduce(\n score_matrix_v2t, op=torch.distributed.ReduceOp.SUM\n )\n torch.distributed.all_reduce(\n score_matrix_t2v, op=torch.distributed.ReduceOp.SUM\n )\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return score_matrix_v2t.cpu().numpy(), score_matrix_t2v.cpu().numpy()\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n max_txt_len=max_txt_len,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.AlproRetrieval","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_retrieval.AlproRetrieval#L30-L422","kind":"class","name":"AlproRetrieval","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":30,"end_line":422,"context_start_line":10,"context_end_line":422,"code":"import time\n\nimport lavis.common.dist_utils as dist_utils\nimport numpy as np\nimport torch\nimport torch.distributed as dist\nimport torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.dist_utils import get_rank\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import AlproIntermediateOutput, AlproOutput\nfrom lavis.models.base_model import all_gather_with_grad\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_retrieval\")\nclass AlproRetrieval(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_retrieval_msrvtt.yaml\",\n \"didemo\": \"configs/models/alpro_retrieval_didemo.yaml\",\n }\n\n def __init__(\n self,\n visual_encoder,\n text_encoder,\n vision_width=768,\n text_width=768,\n embed_dim=256,\n max_txt_len=35,\n temp=0.07,\n ):\n super().__init__()\n\n self.temp = nn.Parameter(torch.ones([]) * temp)\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n self.text_encoder = text_encoder\n\n vision_width = vision_width\n text_width = text_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n visual_inputs = samples[\"video\"]\n caption = samples[\"text_input\"]\n\n b, t, c, h, w = visual_inputs.shape\n\n # forward text\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_feat = F.normalize(self.vision_proj(video_embeds[:, 0, :]), dim=-1)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # ========== (in-batch) ITC loss ==========\n gathered_video_feats = all_gather_with_grad(video_feat)\n gathered_text_feats = all_gather_with_grad(text_feat)\n\n sim_v2t = video_feat @ gathered_text_feats.t() / self.temp\n sim_t2v = text_feat @ gathered_video_feats.t() / self.temp\n\n sim_targets = torch.zeros_like(sim_v2t)\n\n local_rank = get_rank()\n b_start, b_end = b * local_rank, b * (local_rank + 1)\n sim_targets[:, b_start:b_end] = torch.eye(b)\n\n loss_v2t = -torch.sum(F.log_softmax(sim_v2t, dim=1) * sim_targets, dim=1).mean()\n loss_t2v = -torch.sum(F.log_softmax(sim_t2v, dim=1) * sim_targets, dim=1).mean()\n\n vtc_loss = (loss_v2t + loss_t2v) / 2\n\n (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_output,\n encoder_output_neg,\n ) = self.compute_vtm(\n text_embeds=text_embeds,\n text_atts=text.attention_mask,\n image_embeds=video_embeds,\n image_atts=video_atts,\n sim_i2t=sim_v2t.clone(), # for hard mining\n sim_t2i=sim_t2v.clone(), # for hard mining\n )\n\n loss = vtc_loss + vtm_loss\n\n # return {\"loss\": loss}\n return AlproOutput(\n loss=loss,\n loss_vtc=vtc_loss,\n loss_vtm=vtm_loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n encoder_output_neg=encoder_output_neg,\n vtm_logits=vtm_logits,\n vtm_labels=vtm_labels,\n ),\n )\n\n def compute_vtm(\n self, text_embeds, text_atts, image_embeds, image_atts, sim_i2t, sim_t2i\n ):\n device = self.device\n\n # ====== positive pairs =======\n attention_mask = torch.cat([text_atts, image_atts], dim=1)\n embedding_output_pos = torch.cat([text_embeds, image_embeds], dim=1)\n\n encoder_outputs_pos = self.text_encoder(\n encoder_embeds=embedding_output_pos,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n # ====== negative pairs =======\n bs = text_embeds.shape[0]\n\n local_rank = get_rank()\n b_start, b_end = bs * local_rank, bs * (local_rank + 1)\n\n with torch.no_grad():\n weights_v2t = sim_i2t[:, b_start:b_end]\n weights_t2v = sim_t2i[:, b_start:b_end]\n\n # never select self as negative\n weights_v2t.fill_diagonal_(-np.Inf)\n weights_t2v.fill_diagonal_(-np.Inf)\n\n weights_v2t = F.softmax(weights_v2t, dim=1)\n weights_t2v = F.softmax(weights_t2v, dim=1)\n\n # select a negative image for each text\n # FIXME to optimize using indexing operations\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2v[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_v2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text_atts[neg_idx])\n\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text_atts, text_atts_neg], dim=0)\n\n video_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n video_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n attention_mask_all = torch.cat([text_atts_all, video_atts_all], dim=1)\n embedding_output_all = torch.cat([text_embeds_all, video_embeds_all], dim=1)\n\n # forward negative pairs via cross encoder\n encoder_outputs_neg = self.text_encoder(\n encoder_embeds=embedding_output_all,\n attention_mask=attention_mask_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_outputs_pos.last_hidden_state[:, 0, :],\n encoder_outputs_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n vtm_logits = self.itm_head(vl_embeddings)\n\n vtm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(device)\n vtm_loss = F.cross_entropy(vtm_logits, vtm_labels)\n\n return (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_outputs_pos,\n encoder_outputs_neg,\n )\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n k_test = task_cfg.get(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_feats = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text_output = self.text_encoder.forward_text(\n text_input,\n token_type_ids=torch.zeros(\n text_input.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_feats.append(text_output.last_hidden_state.cpu())\n text_embed = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :])\n )\n text_embeds.append(text_embed)\n text_ids.append(text_input.input_ids)\n text_atts.append(text_input.attention_mask)\n\n text_embeds = torch.cat(text_embeds, dim=0)\n text_ids = torch.cat(text_ids, dim=0)\n text_atts = torch.cat(text_atts, dim=0)\n text_feats = torch.cat(text_feats, dim=0)\n\n video_feats = []\n video_embeds = []\n for samples in data_loader:\n video = samples[\"video\"]\n\n video = video.to(self.device)\n video_feat = self.visual_encoder.forward_features(video)\n video_embed = self.vision_proj(video_feat[:, 0, :])\n video_embed = F.normalize(video_embed, dim=-1)\n\n video_feats.append(video_feat.cpu())\n video_embeds.append(video_embed)\n\n video_feats = torch.cat(video_feats, dim=0)\n video_embeds = torch.cat(video_embeds, dim=0)\n\n sims_matrix = video_embeds @ text_embeds.t()\n score_matrix_v2t = torch.full(\n (len(data_loader.dataset.image), len(texts)), -100.0\n ).to(self.device)\n\n num_tasks = dist_utils.get_world_size()\n rank = dist_utils.get_rank()\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n # video-to-text\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n video_feats_repeat = (\n video_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n video_atts_repeat = torch.ones(\n video_feats_repeat.size()[:-1], dtype=torch.long\n ).to(self.device)\n\n attention_mask = torch.cat([text_atts[topk_idx], video_atts_repeat], dim=1)\n embedding_output = torch.cat(\n [text_feats[topk_idx].to(self.device), video_feats_repeat], dim=1\n )\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_v2t[start + i, topk_idx] = score + topk_sim\n\n # text-to-video\n sims_matrix = sims_matrix.t()\n score_matrix_t2v = torch.full(\n (len(texts), len(data_loader.dataset.image)), -100.0\n ).to(self.device)\n\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n text_feats_repeat = (\n text_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n text_atts_repeat = text_atts[start + i].repeat(k_test, 1).to(self.device)\n\n video_atts = torch.ones(\n video_feats[topk_idx].size()[:-1], dtype=torch.long\n ).to(self.device)\n\n embedding_output = torch.cat(\n [text_feats_repeat, video_feats[topk_idx].to(self.device)], dim=1\n )\n attention_mask = torch.cat([text_atts_repeat, video_atts], dim=1)\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_t2v[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.barrier()\n torch.distributed.all_reduce(\n score_matrix_v2t, op=torch.distributed.ReduceOp.SUM\n )\n torch.distributed.all_reduce(\n score_matrix_t2v, op=torch.distributed.ReduceOp.SUM\n )\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return score_matrix_v2t.cpu().numpy(), score_matrix_t2v.cpu().numpy()\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n max_txt_len=max_txt_len,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.__init__","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_retrieval.__init__#L36-L63","kind":"function","name":"__init__","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":36,"end_line":63,"context_start_line":16,"context_end_line":83,"code":"import torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.dist_utils import get_rank\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import AlproIntermediateOutput, AlproOutput\nfrom lavis.models.base_model import all_gather_with_grad\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_retrieval\")\nclass AlproRetrieval(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_retrieval_msrvtt.yaml\",\n \"didemo\": \"configs/models/alpro_retrieval_didemo.yaml\",\n }\n\n def __init__(\n self,\n visual_encoder,\n text_encoder,\n vision_width=768,\n text_width=768,\n embed_dim=256,\n max_txt_len=35,\n temp=0.07,\n ):\n super().__init__()\n\n self.temp = nn.Parameter(torch.ones([]) * temp)\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n self.text_encoder = text_encoder\n\n vision_width = vision_width\n text_width = text_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n visual_inputs = samples[\"video\"]\n caption = samples[\"text_input\"]\n\n b, t, c, h, w = visual_inputs.shape\n\n # forward text\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.forward","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_retrieval.forward#L65-L148","kind":"function","name":"forward","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":65,"end_line":148,"context_start_line":45,"context_end_line":168,"code":" ):\n super().__init__()\n\n self.temp = nn.Parameter(torch.ones([]) * temp)\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n self.text_encoder = text_encoder\n\n vision_width = vision_width\n text_width = text_width\n\n self.vision_proj = nn.Linear(vision_width, embed_dim)\n self.text_proj = nn.Linear(text_width, embed_dim)\n\n self.itm_head = nn.Linear(text_width, 2)\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples):\n with torch.no_grad():\n self.temp.clamp_(0.001, 0.5)\n\n visual_inputs = samples[\"video\"]\n caption = samples[\"text_input\"]\n\n b, t, c, h, w = visual_inputs.shape\n\n # forward text\n text = self.tokenizer(\n caption,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n text_feat = F.normalize(self.text_proj(text_embeds[:, 0, :]), dim=-1)\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_feat = F.normalize(self.vision_proj(video_embeds[:, 0, :]), dim=-1)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # ========== (in-batch) ITC loss ==========\n gathered_video_feats = all_gather_with_grad(video_feat)\n gathered_text_feats = all_gather_with_grad(text_feat)\n\n sim_v2t = video_feat @ gathered_text_feats.t() / self.temp\n sim_t2v = text_feat @ gathered_video_feats.t() / self.temp\n\n sim_targets = torch.zeros_like(sim_v2t)\n\n local_rank = get_rank()\n b_start, b_end = b * local_rank, b * (local_rank + 1)\n sim_targets[:, b_start:b_end] = torch.eye(b)\n\n loss_v2t = -torch.sum(F.log_softmax(sim_v2t, dim=1) * sim_targets, dim=1).mean()\n loss_t2v = -torch.sum(F.log_softmax(sim_t2v, dim=1) * sim_targets, dim=1).mean()\n\n vtc_loss = (loss_v2t + loss_t2v) / 2\n\n (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_output,\n encoder_output_neg,\n ) = self.compute_vtm(\n text_embeds=text_embeds,\n text_atts=text.attention_mask,\n image_embeds=video_embeds,\n image_atts=video_atts,\n sim_i2t=sim_v2t.clone(), # for hard mining\n sim_t2i=sim_t2v.clone(), # for hard mining\n )\n\n loss = vtc_loss + vtm_loss\n\n # return {\"loss\": loss}\n return AlproOutput(\n loss=loss,\n loss_vtc=vtc_loss,\n loss_vtm=vtm_loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n encoder_output_neg=encoder_output_neg,\n vtm_logits=vtm_logits,\n vtm_labels=vtm_labels,\n ),\n )\n\n def compute_vtm(\n self, text_embeds, text_atts, image_embeds, image_atts, sim_i2t, sim_t2i\n ):\n device = self.device\n\n # ====== positive pairs =======\n attention_mask = torch.cat([text_atts, image_atts], dim=1)\n embedding_output_pos = torch.cat([text_embeds, image_embeds], dim=1)\n\n encoder_outputs_pos = self.text_encoder(\n encoder_embeds=embedding_output_pos,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n # ====== negative pairs =======\n bs = text_embeds.shape[0]\n","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.compute_vtm","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_retrieval.compute_vtm#L150-L240","kind":"function","name":"compute_vtm","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":150,"end_line":240,"context_start_line":130,"context_end_line":260,"code":" sim_t2i=sim_t2v.clone(), # for hard mining\n )\n\n loss = vtc_loss + vtm_loss\n\n # return {\"loss\": loss}\n return AlproOutput(\n loss=loss,\n loss_vtc=vtc_loss,\n loss_vtm=vtm_loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n encoder_output_neg=encoder_output_neg,\n vtm_logits=vtm_logits,\n vtm_labels=vtm_labels,\n ),\n )\n\n def compute_vtm(\n self, text_embeds, text_atts, image_embeds, image_atts, sim_i2t, sim_t2i\n ):\n device = self.device\n\n # ====== positive pairs =======\n attention_mask = torch.cat([text_atts, image_atts], dim=1)\n embedding_output_pos = torch.cat([text_embeds, image_embeds], dim=1)\n\n encoder_outputs_pos = self.text_encoder(\n encoder_embeds=embedding_output_pos,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n # ====== negative pairs =======\n bs = text_embeds.shape[0]\n\n local_rank = get_rank()\n b_start, b_end = bs * local_rank, bs * (local_rank + 1)\n\n with torch.no_grad():\n weights_v2t = sim_i2t[:, b_start:b_end]\n weights_t2v = sim_t2i[:, b_start:b_end]\n\n # never select self as negative\n weights_v2t.fill_diagonal_(-np.Inf)\n weights_t2v.fill_diagonal_(-np.Inf)\n\n weights_v2t = F.softmax(weights_v2t, dim=1)\n weights_t2v = F.softmax(weights_t2v, dim=1)\n\n # select a negative image for each text\n # FIXME to optimize using indexing operations\n image_embeds_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_t2v[b], 1).item()\n image_embeds_neg.append(image_embeds[neg_idx])\n image_embeds_neg = torch.stack(image_embeds_neg, dim=0)\n\n # select a negative text for each image\n text_embeds_neg = []\n text_atts_neg = []\n for b in range(bs):\n neg_idx = torch.multinomial(weights_v2t[b], 1).item()\n text_embeds_neg.append(text_embeds[neg_idx])\n text_atts_neg.append(text_atts[neg_idx])\n\n text_embeds_neg = torch.stack(text_embeds_neg, dim=0)\n text_atts_neg = torch.stack(text_atts_neg, dim=0)\n\n text_embeds_all = torch.cat([text_embeds, text_embeds_neg], dim=0)\n text_atts_all = torch.cat([text_atts, text_atts_neg], dim=0)\n\n video_embeds_all = torch.cat([image_embeds_neg, image_embeds], dim=0)\n video_atts_all = torch.cat([image_atts, image_atts], dim=0)\n\n attention_mask_all = torch.cat([text_atts_all, video_atts_all], dim=1)\n embedding_output_all = torch.cat([text_embeds_all, video_embeds_all], dim=1)\n\n # forward negative pairs via cross encoder\n encoder_outputs_neg = self.text_encoder(\n encoder_embeds=embedding_output_all,\n attention_mask=attention_mask_all,\n return_dict=True,\n mode=\"fusion\",\n )\n\n vl_embeddings = torch.cat(\n [\n encoder_outputs_pos.last_hidden_state[:, 0, :],\n encoder_outputs_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n vtm_logits = self.itm_head(vl_embeddings)\n\n vtm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(device)\n vtm_loss = F.cross_entropy(vtm_logits, vtm_labels)\n\n return (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_outputs_pos,\n encoder_outputs_neg,\n )\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n k_test = task_cfg.get(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_feats = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.compute_sim_matrix","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_retrieval.compute_sim_matrix#L242-L394","kind":"function","name":"compute_sim_matrix","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":242,"end_line":394,"context_start_line":222,"context_end_line":414,"code":" encoder_outputs_neg.last_hidden_state[:, 0, :],\n ],\n dim=0,\n )\n vtm_logits = self.itm_head(vl_embeddings)\n\n vtm_labels = torch.cat(\n [torch.ones(bs, dtype=torch.long), torch.zeros(2 * bs, dtype=torch.long)],\n dim=0,\n ).to(device)\n vtm_loss = F.cross_entropy(vtm_logits, vtm_labels)\n\n return (\n vtm_loss,\n vtm_logits,\n vtm_labels,\n encoder_outputs_pos,\n encoder_outputs_neg,\n )\n\n def compute_sim_matrix(self, data_loader, task_cfg):\n k_test = task_cfg.get(\"k_test\")\n\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation:\"\n\n logging.info(\"Computing features for evaluation...\")\n start_time = time.time()\n\n texts = data_loader.dataset.text\n num_text = len(texts)\n text_bs = 256\n text_ids = []\n text_embeds = []\n text_feats = []\n text_atts = []\n for i in range(0, num_text, text_bs):\n text = texts[i : min(num_text, i + text_bs)]\n text_input = self.tokenizer(\n text,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n text_output = self.text_encoder.forward_text(\n text_input,\n token_type_ids=torch.zeros(\n text_input.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_feats.append(text_output.last_hidden_state.cpu())\n text_embed = F.normalize(\n self.text_proj(text_output.last_hidden_state[:, 0, :])\n )\n text_embeds.append(text_embed)\n text_ids.append(text_input.input_ids)\n text_atts.append(text_input.attention_mask)\n\n text_embeds = torch.cat(text_embeds, dim=0)\n text_ids = torch.cat(text_ids, dim=0)\n text_atts = torch.cat(text_atts, dim=0)\n text_feats = torch.cat(text_feats, dim=0)\n\n video_feats = []\n video_embeds = []\n for samples in data_loader:\n video = samples[\"video\"]\n\n video = video.to(self.device)\n video_feat = self.visual_encoder.forward_features(video)\n video_embed = self.vision_proj(video_feat[:, 0, :])\n video_embed = F.normalize(video_embed, dim=-1)\n\n video_feats.append(video_feat.cpu())\n video_embeds.append(video_embed)\n\n video_feats = torch.cat(video_feats, dim=0)\n video_embeds = torch.cat(video_embeds, dim=0)\n\n sims_matrix = video_embeds @ text_embeds.t()\n score_matrix_v2t = torch.full(\n (len(data_loader.dataset.image), len(texts)), -100.0\n ).to(self.device)\n\n num_tasks = dist_utils.get_world_size()\n rank = dist_utils.get_rank()\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n # video-to-text\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n video_feats_repeat = (\n video_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n video_atts_repeat = torch.ones(\n video_feats_repeat.size()[:-1], dtype=torch.long\n ).to(self.device)\n\n attention_mask = torch.cat([text_atts[topk_idx], video_atts_repeat], dim=1)\n embedding_output = torch.cat(\n [text_feats[topk_idx].to(self.device), video_feats_repeat], dim=1\n )\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_v2t[start + i, topk_idx] = score + topk_sim\n\n # text-to-video\n sims_matrix = sims_matrix.t()\n score_matrix_t2v = torch.full(\n (len(texts), len(data_loader.dataset.image)), -100.0\n ).to(self.device)\n\n step = sims_matrix.size(0) // num_tasks + 1\n start = rank * step\n end = min(sims_matrix.size(0), start + step)\n\n for i, sims in enumerate(\n metric_logger.log_every(sims_matrix[start:end], 50, header)\n ):\n\n topk_sim, topk_idx = sims.topk(k=k_test, dim=0)\n\n text_feats_repeat = (\n text_feats[start + i].repeat(k_test, 1, 1).to(self.device)\n )\n text_atts_repeat = text_atts[start + i].repeat(k_test, 1).to(self.device)\n\n video_atts = torch.ones(\n video_feats[topk_idx].size()[:-1], dtype=torch.long\n ).to(self.device)\n\n embedding_output = torch.cat(\n [text_feats_repeat, video_feats[topk_idx].to(self.device)], dim=1\n )\n attention_mask = torch.cat([text_atts_repeat, video_atts], dim=1)\n\n output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_t2v[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.barrier()\n torch.distributed.all_reduce(\n score_matrix_v2t, op=torch.distributed.ReduceOp.SUM\n )\n torch.distributed.all_reduce(\n score_matrix_t2v, op=torch.distributed.ReduceOp.SUM\n )\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return score_matrix_v2t.cpu().numpy(), score_matrix_t2v.cpu().numpy()\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n max_txt_len=max_txt_len,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_retrieval.from_config","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_retrieval.from_config#L397-L422","kind":"function","name":"from_config","path":"lavis/models/alpro_models/alpro_retrieval.py","language":"python","start_line":397,"end_line":422,"context_start_line":377,"context_end_line":422,"code":"\n score = self.itm_head(output.last_hidden_state[:, 0, :])[:, 1]\n score_matrix_t2v[start + i, topk_idx] = score + topk_sim\n\n if dist_utils.is_dist_avail_and_initialized():\n dist.barrier()\n torch.distributed.all_reduce(\n score_matrix_v2t, op=torch.distributed.ReduceOp.SUM\n )\n torch.distributed.all_reduce(\n score_matrix_t2v, op=torch.distributed.ReduceOp.SUM\n )\n\n total_time = time.time() - start_time\n total_time_str = str(datetime.timedelta(seconds=int(total_time)))\n logging.info(\"Evaluation time {}\".format(total_time_str))\n\n return score_matrix_v2t.cpu().numpy(), score_matrix_t2v.cpu().numpy()\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n max_txt_len = cfg.get(\"max_txt_len\", 35)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n max_txt_len=max_txt_len,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"13db023ab0c881dfd7a1454b08f3e650f764f5d52c478a290d8179faf4a322b8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_outputs","uri":"program://CREMA/module/lavis.models.alpro_models.alpro_outputs#L1-L59","kind":"module","name":"lavis.models.alpro_models.alpro_outputs","path":"lavis/models/alpro_models/alpro_outputs.py","language":"python","start_line":1,"end_line":59,"context_start_line":1,"context_end_line":59,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlproSimilarity(ModelOutput):\n sim_v2t: torch.FloatTensor = None\n sim_t2v: torch.FloatTensor = None\n\n sim_v2t_targets: Optional[torch.FloatTensor] = None\n sim_t2v_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproIntermediateOutput(ModelOutput):\n # uni-modal features\n video_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n vtm_logits: Optional[torch.FloatTensor] = None\n vtm_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlproOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlproSimilarity] = None\n\n intermediate_output: AlproIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_vtc: Optional[torch.FloatTensor] = None\n\n loss_vtm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproOutputWithLogits(AlproOutput):\n logits: torch.FloatTensor = None","source_hash":"35cefcac85e48567e09a417eb69f5419ab55fa1a9244a56eea6cfff5c9cfba53","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_outputs.AlproSimilarity","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_outputs.AlproSimilarity#L19-L24","kind":"class","name":"AlproSimilarity","path":"lavis/models/alpro_models/alpro_outputs.py","language":"python","start_line":19,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlproSimilarity(ModelOutput):\n sim_v2t: torch.FloatTensor = None\n sim_t2v: torch.FloatTensor = None\n\n sim_v2t_targets: Optional[torch.FloatTensor] = None\n sim_t2v_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproIntermediateOutput(ModelOutput):\n # uni-modal features\n video_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n vtm_logits: Optional[torch.FloatTensor] = None\n vtm_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlproOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlproSimilarity] = None","source_hash":"35cefcac85e48567e09a417eb69f5419ab55fa1a9244a56eea6cfff5c9cfba53","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_outputs.AlproIntermediateOutput","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_outputs.AlproIntermediateOutput#L28-L38","kind":"class","name":"AlproIntermediateOutput","path":"lavis/models/alpro_models/alpro_outputs.py","language":"python","start_line":28,"end_line":38,"context_start_line":8,"context_end_line":58,"code":"from dataclasses import dataclass\nfrom typing import Optional\n\nimport torch\nfrom transformers.modeling_outputs import (\n BaseModelOutputWithPoolingAndCrossAttentions,\n ModelOutput,\n)\n\n\n@dataclass\nclass AlproSimilarity(ModelOutput):\n sim_v2t: torch.FloatTensor = None\n sim_t2v: torch.FloatTensor = None\n\n sim_v2t_targets: Optional[torch.FloatTensor] = None\n sim_t2v_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproIntermediateOutput(ModelOutput):\n # uni-modal features\n video_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n vtm_logits: Optional[torch.FloatTensor] = None\n vtm_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlproOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlproSimilarity] = None\n\n intermediate_output: AlproIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_vtc: Optional[torch.FloatTensor] = None\n\n loss_vtm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproOutputWithLogits(AlproOutput):","source_hash":"35cefcac85e48567e09a417eb69f5419ab55fa1a9244a56eea6cfff5c9cfba53","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_outputs.AlproOutput","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_outputs.AlproOutput#L42-L54","kind":"class","name":"AlproOutput","path":"lavis/models/alpro_models/alpro_outputs.py","language":"python","start_line":42,"end_line":54,"context_start_line":22,"context_end_line":59,"code":"\n sim_v2t_targets: Optional[torch.FloatTensor] = None\n sim_t2v_targets: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproIntermediateOutput(ModelOutput):\n # uni-modal features\n video_embeds: torch.FloatTensor = None\n text_embeds: Optional[torch.FloatTensor] = None\n\n # intermediate outputs of multimodal encoder\n encoder_output: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n encoder_output_neg: Optional[BaseModelOutputWithPoolingAndCrossAttentions] = None\n\n vtm_logits: Optional[torch.FloatTensor] = None\n vtm_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlproOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlproSimilarity] = None\n\n intermediate_output: AlproIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_vtc: Optional[torch.FloatTensor] = None\n\n loss_vtm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproOutputWithLogits(AlproOutput):\n logits: torch.FloatTensor = None","source_hash":"35cefcac85e48567e09a417eb69f5419ab55fa1a9244a56eea6cfff5c9cfba53","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_outputs.AlproOutputWithLogits","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_outputs.AlproOutputWithLogits#L58-L59","kind":"class","name":"AlproOutputWithLogits","path":"lavis/models/alpro_models/alpro_outputs.py","language":"python","start_line":58,"end_line":59,"context_start_line":38,"context_end_line":59,"code":" vtm_labels: Optional[torch.LongTensor] = None\n\n\n@dataclass\nclass AlproOutput(ModelOutput):\n # some finetuned models (e.g. BlipVQA) do not compute similarity, thus optional.\n sims: Optional[AlproSimilarity] = None\n\n intermediate_output: AlproIntermediateOutput = None\n\n loss: Optional[torch.FloatTensor] = None\n\n loss_vtc: Optional[torch.FloatTensor] = None\n\n loss_vtm: Optional[torch.FloatTensor] = None\n\n loss_mlm: Optional[torch.FloatTensor] = None\n\n\n@dataclass\nclass AlproOutputWithLogits(AlproOutput):\n logits: torch.FloatTensor = None","source_hash":"35cefcac85e48567e09a417eb69f5419ab55fa1a9244a56eea6cfff5c9cfba53","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa","uri":"program://CREMA/module/lavis.models.alpro_models.alpro_qa#L1-L141","kind":"module","name":"lavis.models.alpro_models.alpro_qa","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":1,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom warnings import warn\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import (\n AlproIntermediateOutput,\n AlproOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_qa\")\nclass AlproQA(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_qa_msrvtt.yaml\",\n \"msvd\": \"configs/models/alpro_qa_msvd.yaml\",\n }\n\n def __init__(\n self, visual_encoder, text_encoder, hidden_size, num_classes, max_txt_len=40\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n\n self.text_encoder = text_encoder\n\n if num_classes > 0:\n self.classifier = nn.Sequential(\n nn.Linear(hidden_size, hidden_size * 2),\n nn.ReLU(True),\n nn.Linear(hidden_size * 2, num_classes),\n )\n else:\n warn(f\"num_classes is 0. Initialized {type(self)} without classifier.\")\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples, is_train=True):\n visual_inputs = samples[\"video\"]\n question = samples[\"text_input\"]\n targets = samples[\"answers\"]\n\n # forward text\n text = self.tokenizer(\n question,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward cross-encoder\n attention_mask = torch.cat([text.attention_mask, video_atts], dim=1)\n embedding_output = torch.cat([text_embeds, video_embeds], dim=1)\n\n encoder_output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n prediction = self.classifier(encoder_output.last_hidden_state[:, 0, :])\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return AlproOutputWithLogits(\n loss=loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n ),\n logits=prediction,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n num_classes = cfg.get(\"num_classes\", -1)\n hidden_size = cfg.get(\"hidden_size\", 768)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n hidden_size=hidden_size,\n num_classes=num_classes,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa.AlproQA","uri":"program://CREMA/class/lavis.models.alpro_models.alpro_qa.AlproQA#L25-L141","kind":"class","name":"AlproQA","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":25,"end_line":141,"context_start_line":5,"context_end_line":141,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom warnings import warn\n\nimport torch\nimport torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import (\n AlproIntermediateOutput,\n AlproOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_qa\")\nclass AlproQA(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_qa_msrvtt.yaml\",\n \"msvd\": \"configs/models/alpro_qa_msvd.yaml\",\n }\n\n def __init__(\n self, visual_encoder, text_encoder, hidden_size, num_classes, max_txt_len=40\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n\n self.text_encoder = text_encoder\n\n if num_classes > 0:\n self.classifier = nn.Sequential(\n nn.Linear(hidden_size, hidden_size * 2),\n nn.ReLU(True),\n nn.Linear(hidden_size * 2, num_classes),\n )\n else:\n warn(f\"num_classes is 0. Initialized {type(self)} without classifier.\")\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples, is_train=True):\n visual_inputs = samples[\"video\"]\n question = samples[\"text_input\"]\n targets = samples[\"answers\"]\n\n # forward text\n text = self.tokenizer(\n question,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward cross-encoder\n attention_mask = torch.cat([text.attention_mask, video_atts], dim=1)\n embedding_output = torch.cat([text_embeds, video_embeds], dim=1)\n\n encoder_output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n prediction = self.classifier(encoder_output.last_hidden_state[:, 0, :])\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return AlproOutputWithLogits(\n loss=loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n ),\n logits=prediction,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n num_classes = cfg.get(\"num_classes\", -1)\n hidden_size = cfg.get(\"hidden_size\", 768)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n hidden_size=hidden_size,\n num_classes=num_classes,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa.__init__","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_qa.__init__#L31-L51","kind":"function","name":"__init__","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":31,"end_line":51,"context_start_line":11,"context_end_line":71,"code":"import torch.nn.functional as F\nfrom lavis.common.config import node_to_dict\nfrom lavis.common.registry import registry\nfrom lavis.models.alpro_models import AlproBase\nfrom lavis.models.alpro_models.alpro_outputs import (\n AlproIntermediateOutput,\n AlproOutputWithLogits,\n)\nfrom lavis.models.med import XBertEncoder\nfrom lavis.models.timesformer.vit import TimeSformer\nfrom torch import nn\n\n\n@registry.register_model(\"alpro_qa\")\nclass AlproQA(AlproBase):\n PRETRAINED_MODEL_CONFIG_DICT = {\n \"msrvtt\": \"configs/models/alpro_qa_msrvtt.yaml\",\n \"msvd\": \"configs/models/alpro_qa_msvd.yaml\",\n }\n\n def __init__(\n self, visual_encoder, text_encoder, hidden_size, num_classes, max_txt_len=40\n ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n\n self.text_encoder = text_encoder\n\n if num_classes > 0:\n self.classifier = nn.Sequential(\n nn.Linear(hidden_size, hidden_size * 2),\n nn.ReLU(True),\n nn.Linear(hidden_size * 2, num_classes),\n )\n else:\n warn(f\"num_classes is 0. Initialized {type(self)} without classifier.\")\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples, is_train=True):\n visual_inputs = samples[\"video\"]\n question = samples[\"text_input\"]\n targets = samples[\"answers\"]\n\n # forward text\n text = self.tokenizer(\n question,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa.forward","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_qa.forward#L53-L107","kind":"function","name":"forward","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":53,"end_line":107,"context_start_line":33,"context_end_line":127,"code":" ):\n super().__init__()\n\n self.tokenizer = self.init_tokenizer()\n\n self.visual_encoder = visual_encoder\n\n self.text_encoder = text_encoder\n\n if num_classes > 0:\n self.classifier = nn.Sequential(\n nn.Linear(hidden_size, hidden_size * 2),\n nn.ReLU(True),\n nn.Linear(hidden_size * 2, num_classes),\n )\n else:\n warn(f\"num_classes is 0. Initialized {type(self)} without classifier.\")\n\n self.max_txt_len = max_txt_len\n\n def forward(self, samples, is_train=True):\n visual_inputs = samples[\"video\"]\n question = samples[\"text_input\"]\n targets = samples[\"answers\"]\n\n # forward text\n text = self.tokenizer(\n question,\n padding=\"max_length\",\n truncation=True,\n max_length=self.max_txt_len,\n return_tensors=\"pt\",\n ).to(self.device)\n\n text_output = self.text_encoder.forward_text(\n text,\n token_type_ids=torch.zeros(\n text.input_ids.shape, dtype=torch.long, device=self.device\n ),\n )\n text_embeds = text_output.last_hidden_state\n\n # forward visual\n # timeSformer asks for (b, c, t, h, w) as input.\n video_embeds = self.visual_encoder.forward_features(visual_inputs)\n video_atts = torch.ones(video_embeds.size()[:-1], dtype=torch.long).to(\n self.device\n )\n\n # forward cross-encoder\n attention_mask = torch.cat([text.attention_mask, video_atts], dim=1)\n embedding_output = torch.cat([text_embeds, video_embeds], dim=1)\n\n encoder_output = self.text_encoder(\n encoder_embeds=embedding_output,\n attention_mask=attention_mask,\n return_dict=True,\n mode=\"fusion\",\n )\n\n prediction = self.classifier(encoder_output.last_hidden_state[:, 0, :])\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return AlproOutputWithLogits(\n loss=loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n ),\n logits=prediction,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n num_classes = cfg.get(\"num_classes\", -1)\n hidden_size = cfg.get(\"hidden_size\", 768)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa.predict","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_qa.predict#L109-L111","kind":"function","name":"predict","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":109,"end_line":111,"context_start_line":89,"context_end_line":131,"code":" return_dict=True,\n mode=\"fusion\",\n )\n\n prediction = self.classifier(encoder_output.last_hidden_state[:, 0, :])\n if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return AlproOutputWithLogits(\n loss=loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n ),\n logits=prediction,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n num_classes = cfg.get(\"num_classes\", -1)\n hidden_size = cfg.get(\"hidden_size\", 768)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n hidden_size=hidden_size,\n num_classes=num_classes,\n )\n","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.models.alpro_models.alpro_qa.from_config","uri":"program://CREMA/function/lavis.models.alpro_models.alpro_qa.from_config#L114-L141","kind":"function","name":"from_config","path":"lavis/models/alpro_models/alpro_qa.py","language":"python","start_line":114,"end_line":141,"context_start_line":94,"context_end_line":141,"code":" if is_train:\n loss = F.cross_entropy(prediction, targets)\n # return {\"loss\": loss}\n return AlproOutputWithLogits(\n loss=loss,\n intermediate_output=AlproIntermediateOutput(\n video_embeds=video_embeds,\n text_embeds=text_embeds,\n encoder_output=encoder_output,\n ),\n logits=prediction,\n )\n else:\n return {\"predictions\": prediction, \"targets\": targets}\n\n def predict(self, samples):\n output = self.forward(samples, is_train=False)\n return output\n\n @classmethod\n def from_config(cls, cfg):\n # vision encoder\n visual_encoder_config = node_to_dict(cfg.timesformer)\n visual_encoder = TimeSformer(**visual_encoder_config)\n\n # text encoder\n text_encoder = XBertEncoder.from_config(cfg)\n\n num_classes = cfg.get(\"num_classes\", -1)\n hidden_size = cfg.get(\"hidden_size\", 768)\n\n model = cls(\n visual_encoder=visual_encoder,\n text_encoder=text_encoder,\n hidden_size=hidden_size,\n num_classes=num_classes,\n )\n\n num_patches = (\n visual_encoder_config[\"image_size\"] // visual_encoder_config[\"patch_size\"]\n ) ** 2\n num_frames = visual_encoder_config[\"n_frms\"]\n\n model.load_checkpoint_from_config(\n cfg, num_frames=num_frames, num_patches=num_patches\n )\n\n return model","source_hash":"64229dc9745083ef2b678e2070fe731e0f8e74c12c310d7d2e1de15f6e8e74bb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils","uri":"program://CREMA/module/lavis.datasets.data_utils#L1-L510","kind":"module","name":"lavis.datasets.data_utils","path":"lavis/datasets/data_utils.py","language":"python","start_line":1,"end_line":510,"context_start_line":1,"context_end_line":510,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport gzip\nimport logging\nimport cv2\n\nimport random as rnd\nimport tarfile\nimport zipfile\n\nimport decord\nimport webdataset as wds\nimport numpy as np\nimport torch\nfrom torch.utils.data.dataset import IterableDataset, ChainDataset\nfrom decord import VideoReader\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.base_dataset import ConcatDataset\nfrom tqdm import tqdm\n\nfrom PIL import Image\n\ndecord.bridge.set_bridge(\"torch\")\nMAX_INT = registry.get(\"MAX_INT\")\n\n\ndef readFlow(filename, new_size=(256, 256)):\n \"\"\" Read optical flow from file.\n \n The function reads optical flow from a .flo file and returns it as a numpy array.\n \n Args:\n filename (str): The path to the .flo file.\n \n Returns:\n numpy.ndarray: A numpy array containing the optical flow with u and v components stacked in depth.\n \"\"\"\n # Open the file in binary mode\n with open(filename, 'rb') as f:\n # Check the magic number in the header matches the expected .flo format\n magic = np.fromfile(f, np.float32, count=1)[0]\n assert magic == 202021.25, 'Invalid .flo file'\n\n # Read the width and height of the flow field\n width = np.fromfile(f, np.int32, count=1)[0]\n height = np.fromfile(f, np.int32, count=1)[0]\n\n # Read the flow field data\n flow_data = np.fromfile(f, np.float32, count=2 * width * height)\n\n # Reshape the flow field data into a 3D numpy array with shape (height, width, 2)\n flow = np.reshape(flow_data, (height, width, 2))\n flow = np.transpose(flow, (1, 0, 2))\n\n flow_resized = cv2.resize(flow, new_size[::-1], interpolation=cv2.INTER_LINEAR)\n # min_val = flow_resized.min()\n # max_val = flow_resized.max()\n # flow_resized = 2 * (flow_resized - min_val) / (max_val - min_val) - 1\n\n return flow_resized\n\ndef load_flow(frames_dir, indices,\n height=-1, width=-1):\n\n #print(frames_dir)\n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['flo']])\n frms = []\n for idx in indices:\n # print(idx, frame_files[idx])\n flow = readFlow(frame_files[idx], (height, width))\n min_val = flow.min()\n max_val = flow.max()\n flow = 2 * (flow - min_val) / (max_val - min_val) - 1\n frms.append(np.asarray(flow))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n # print('frms',frms.shape)\n # print(frms.shape)\n frms = torch.from_numpy(frms) #.unsqueeze(1) # (T, 1, H, W)\n frms = frms.permute(0,3,1,2)\n return frms\n\ndef load_depth(frames_dir, indices,\n height=-1, width=-1, invalid_val=-99):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print(frame_files)\n # print(len(indices), indices)\n frms = []\n for idx in indices:\n depth = Image.open(frame_files[idx])\n depth = depth.resize((width, height))\n depth = np.asarray(depth)\n\n invalid_mask = depth == invalid_val\n mask = np.logical_not(invalid_mask)\n vmin = np.percentile(depth[mask],2) \n vmax = np.percentile(depth[mask],85)\n if vmin != vmax:\n depth = (depth - vmin) / (vmax - vmin) \n else:\n depth = depth * 0.\n\n frms.append(np.asarray(depth))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms).unsqueeze(1) # (1, C, T, H, W)\n\n return frms\n\ndef load_frames(frames_dir, n_frms=MAX_INT, height=-1, width=-1, \n sampling=\"uniform\", clip_proposal=None, type='rgb', indices=None):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print('frame_files', len(frame_files))\n vlen = len(frame_files) - 1 # for flow\n n_frms = n_frms #min(n_frms, vlen)\n \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*vlen), int(clip_proposal[1]*vlen)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = [(intervals[i], intervals[i + 1]) for i in range(len(intervals) - 1)]\n \n if indices is None:\n if sampling == 'random':\n indices_ = [rnd.choice(range(x[0], x[1])) if x[0] != x[1] else x[0] for x in ranges]\n elif sampling == 'uniform':\n indices_ = [(x[0] + x[1]) // 2 for x in ranges]\n # elif sampling == \"headtail\":\n # indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n # indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n # indices = indices_h + indices_t\n else:\n raise NotImplementedError\n else:\n indices_ = indices\n \n # print(len(indices_) , n_frms, indices_)\n\n if len(indices_) < n_frms:\n extra = [indices_[-1] for i in range(n_frms - len(indices_))]\n indices_.extend(extra)\n frms = []\n # debug = [frame_files[idx] for idx in indices]\n # print(debug)\n # print('frame_files',len(frame_files))\n # print('indices_', indices_)\n for idx in indices_:\n # print(frame_files[idx])\n with Image.open(frame_files[idx]) as img:\n if type == 'norm':\n img = img.convert('RGB')\n if type == 'flow':\n img = img.transpose(Image.FLIP_LEFT_RIGHT)\n img = img.transpose(2) # ROTATE_90\n # print('here')\n if type == 'depth':\n img = img.convert('RGB')\n \n if height > 0 and width > 0:\n img = img.resize((width, height))\n frms.append(np.asarray(img))\n \n frms = np.stack(frms).transpose(3, 0, 1, 2).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms)\n # print('frms', frms.shape) # 3, 4, 224, 224 for rgb, 4, 4, 224, 224 for norm, 3, 4, 224, 224 for norm.convert('RGB')\n\n return frms, indices_\n\n# add for loading video\ndef load_video(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))\n\n if sampling == 'random':\n indices = []\n for x in ranges:\n if x[0] == x[1]:\n indices.append(x[0])\n else:\n indices.append(rnd.choice(range(x[0], x[1])))\n elif sampling == 'uniform':\n \n indices = [(x[0] + x[1]) // 2 for x in ranges]\n\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n frms = vr.get_batch(indices).permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps\n\ndef load_video_demo(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))\n\n if sampling == 'random':\n indices = []\n for x in ranges:\n if x[0] == x[1]:\n indices.append(x[0])\n else:\n indices.append(rnd.choice(range(x[0], x[1])))\n elif sampling == 'uniform':\n \n indices = [(x[0] + x[1]) // 2 for x in ranges]\n\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n \n frms = vr.get_batch(indices)\n frms = frms.asnumpy()\n frms = torch.from_numpy(frms)\n frms = frms.permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"\n Organizes datasets by split.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by name.\n\n Returns:\n Dict of datasets by split {split_name: List[Datasets]}.\n \"\"\"\n # if len(datasets) == 1:\n # return datasets[list(datasets.keys())[0]]\n # else:\n reorg_datasets = dict()\n\n # reorganize by split\n for _, dataset in datasets.items():\n for split_name, dataset_split in dataset.items():\n if split_name not in reorg_datasets:\n reorg_datasets[split_name] = [dataset_split]\n else:\n reorg_datasets[split_name].append(dataset_split)\n\n return reorg_datasets\n\n\ndef concat_datasets(datasets):\n \"\"\"\n Concatenates multiple datasets into a single dataset.\n\n It supports may-style datasets and DataPipeline from WebDataset. Currently, does not support\n generic IterableDataset because it requires creating separate samplers.\n\n Now only supports conctenating training datasets and assuming validation and testing\n have only a single dataset. This is because metrics should not be computed on the concatenated\n datasets.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by split.\n\n Returns:\n Dict of concatenated datasets by split, \"train\" is the concatenation of multiple datasets,\n \"val\" and \"test\" remain the same.\n\n If the input training datasets contain both map-style and DataPipeline datasets, returns\n a tuple, where the first element is a concatenated map-style dataset and the second\n element is a chained DataPipeline dataset.\n\n \"\"\"\n # concatenate datasets in the same split\n for split_name in datasets:\n if split_name != \"train\":\n assert (\n len(datasets[split_name]) == 1\n ), \"Do not support multiple {} datasets.\".format(split_name)\n datasets[split_name] = datasets[split_name][0]\n else:\n iterable_datasets, map_datasets = [], []\n for dataset in datasets[split_name]:\n if isinstance(dataset, wds.DataPipeline):\n logging.info(\n \"Dataset {} is IterableDataset, can't be concatenated.\".format(\n dataset\n )\n )\n iterable_datasets.append(dataset)\n elif isinstance(dataset, IterableDataset):\n raise NotImplementedError(\n \"Do not support concatenation of generic IterableDataset.\"\n )\n else:\n map_datasets.append(dataset)\n\n # if len(iterable_datasets) > 0:\n # concatenate map-style datasets and iterable-style datasets separately\n chained_datasets = (\n ChainDataset(iterable_datasets) if len(iterable_datasets) > 0 else None\n )\n concat_datasets = (\n ConcatDataset(map_datasets) if len(map_datasets) > 0 else None\n )\n\n train_datasets = concat_datasets, chained_datasets\n train_datasets = tuple([x for x in train_datasets if x is not None])\n train_datasets = (\n train_datasets[0] if len(train_datasets) == 1 else train_datasets\n )\n\n datasets[split_name] = train_datasets\n\n return datasets\n\n\ndef extract_archive(from_path, to_path=None, overwrite=False):\n \"\"\"Extract archive.\n\n Args:\n from_path: the path of the archive.\n to_path: the root path of the extracted files (directory of from_path)\n overwrite: overwrite existing files (False)\n\n Returns:\n List of paths to extracted files even if not overwritten.\n\n Examples:\n >>> url = 'http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz'\n >>> from_path = './validation.tar.gz'\n >>> to_path = './'\n >>> torchtext.utils.download_from_url(url, from_path)\n >>> torchtext.utils.extract_archive(from_path, to_path)\n >>> ['.data/val.de', '.data/val.en']\n >>> torchtext.utils.download_from_url(url, from_path)\n >>> torchtext.utils.extract_archive(from_path, to_path)\n >>> ['.data/val.de', '.data/val.en']\n\n \"\"\"\n\n if to_path is None:\n to_path = os.path.dirname(from_path)\n\n if from_path.endswith((\".tar.gz\", \".tgz\")):\n logging.info(\"Opening tar file {} to {}.\".format(from_path, to_path))\n with tarfile.open(from_path, \"r\") as tar:\n files = []\n for file_ in tqdm(tar):\n file_path = os.path.join(to_path, file_.name)\n if file_.isfile():\n files.append(file_path)\n if os.path.exists(file_path):\n logging.info(\"{} already extracted.\".format(file_path))\n if not overwrite:\n continue\n tar.extract(file_, to_path)\n logging.info(\"Finished extracting tar file {}.\".format(from_path))\n return files\n\n elif from_path.endswith(\".zip\"):\n assert zipfile.is_zipfile(from_path), from_path\n logging.info(\"Opening zip file {} to {}.\".format(from_path, to_path))\n with zipfile.ZipFile(from_path, \"r\") as zfile:\n files = []\n for file_ in tqdm(zfile.namelist()):\n file_path = os.path.join(to_path, file_)\n files.append(file_path)\n if os.path.exists(file_path):\n logging.info(\"{} already extracted.\".format(file_path))\n if not overwrite:\n continue\n zfile.extract(file_, to_path)\n files = [f for f in files if os.path.isfile(f)]\n logging.info(\"Finished extracting zip file {}.\".format(from_path))\n return files\n\n elif from_path.endswith(\".gz\"):\n logging.info(\"Opening gz file {} to {}.\".format(from_path, to_path))\n default_block_size = 65536\n filename = from_path[:-3]\n files = [filename]\n with gzip.open(from_path, \"rb\") as gzfile, open(filename, \"wb\") as d_file:\n while True:\n block = gzfile.read(default_block_size)\n if not block:\n break\n else:\n d_file.write(block)\n d_file.write(block)\n logging.info(\"Finished extracting gz file {}.\".format(from_path))\n return files\n\n else:\n raise NotImplementedError(\n \"We currently only support tar.gz, .tgz, .gz and zip achives.\"\n )\n\n\ndef save_frames_grid(img_array, out_path):\n import torch\n from PIL import Image\n from torchvision.utils import make_grid\n\n if len(img_array.shape) == 3:\n img_array = img_array.unsqueeze(0)\n elif len(img_array.shape) == 5:\n b, t, c, h, w = img_array.shape\n img_array = img_array.view(-1, c, h, w)\n elif len(img_array.shape) == 4:\n pass\n else:\n raise NotImplementedError(\n \"Supports only (b,t,c,h,w)-shaped inputs. First two dimensions can be ignored.\"\n )\n\n assert img_array.shape[1] == 3, \"Exepcting input shape of (H, W, 3), i.e. RGB-only.\"\n\n grid = make_grid(img_array)\n ndarr = grid.permute(1, 2, 0).to(\"cpu\", torch.uint8).numpy()\n\n img = Image.fromarray(ndarr)\n\n img.save(out_path)","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.readFlow","uri":"program://CREMA/function/lavis.datasets.data_utils.readFlow#L33-L66","kind":"function","name":"readFlow","path":"lavis/datasets/data_utils.py","language":"python","start_line":33,"end_line":66,"context_start_line":13,"context_end_line":86,"code":"import random as rnd\nimport tarfile\nimport zipfile\n\nimport decord\nimport webdataset as wds\nimport numpy as np\nimport torch\nfrom torch.utils.data.dataset import IterableDataset, ChainDataset\nfrom decord import VideoReader\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.base_dataset import ConcatDataset\nfrom tqdm import tqdm\n\nfrom PIL import Image\n\ndecord.bridge.set_bridge(\"torch\")\nMAX_INT = registry.get(\"MAX_INT\")\n\n\ndef readFlow(filename, new_size=(256, 256)):\n \"\"\" Read optical flow from file.\n \n The function reads optical flow from a .flo file and returns it as a numpy array.\n \n Args:\n filename (str): The path to the .flo file.\n \n Returns:\n numpy.ndarray: A numpy array containing the optical flow with u and v components stacked in depth.\n \"\"\"\n # Open the file in binary mode\n with open(filename, 'rb') as f:\n # Check the magic number in the header matches the expected .flo format\n magic = np.fromfile(f, np.float32, count=1)[0]\n assert magic == 202021.25, 'Invalid .flo file'\n\n # Read the width and height of the flow field\n width = np.fromfile(f, np.int32, count=1)[0]\n height = np.fromfile(f, np.int32, count=1)[0]\n\n # Read the flow field data\n flow_data = np.fromfile(f, np.float32, count=2 * width * height)\n\n # Reshape the flow field data into a 3D numpy array with shape (height, width, 2)\n flow = np.reshape(flow_data, (height, width, 2))\n flow = np.transpose(flow, (1, 0, 2))\n\n flow_resized = cv2.resize(flow, new_size[::-1], interpolation=cv2.INTER_LINEAR)\n # min_val = flow_resized.min()\n # max_val = flow_resized.max()\n # flow_resized = 2 * (flow_resized - min_val) / (max_val - min_val) - 1\n\n return flow_resized\n\ndef load_flow(frames_dir, indices,\n height=-1, width=-1):\n\n #print(frames_dir)\n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['flo']])\n frms = []\n for idx in indices:\n # print(idx, frame_files[idx])\n flow = readFlow(frame_files[idx], (height, width))\n min_val = flow.min()\n max_val = flow.max()\n flow = 2 * (flow - min_val) / (max_val - min_val) - 1\n frms.append(np.asarray(flow))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n # print('frms',frms.shape)\n # print(frms.shape)\n frms = torch.from_numpy(frms) #.unsqueeze(1) # (T, 1, H, W)\n frms = frms.permute(0,3,1,2)","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.load_flow","uri":"program://CREMA/function/lavis.datasets.data_utils.load_flow#L68-L87","kind":"function","name":"load_flow","path":"lavis/datasets/data_utils.py","language":"python","start_line":68,"end_line":87,"context_start_line":48,"context_end_line":107,"code":" assert magic == 202021.25, 'Invalid .flo file'\n\n # Read the width and height of the flow field\n width = np.fromfile(f, np.int32, count=1)[0]\n height = np.fromfile(f, np.int32, count=1)[0]\n\n # Read the flow field data\n flow_data = np.fromfile(f, np.float32, count=2 * width * height)\n\n # Reshape the flow field data into a 3D numpy array with shape (height, width, 2)\n flow = np.reshape(flow_data, (height, width, 2))\n flow = np.transpose(flow, (1, 0, 2))\n\n flow_resized = cv2.resize(flow, new_size[::-1], interpolation=cv2.INTER_LINEAR)\n # min_val = flow_resized.min()\n # max_val = flow_resized.max()\n # flow_resized = 2 * (flow_resized - min_val) / (max_val - min_val) - 1\n\n return flow_resized\n\ndef load_flow(frames_dir, indices,\n height=-1, width=-1):\n\n #print(frames_dir)\n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['flo']])\n frms = []\n for idx in indices:\n # print(idx, frame_files[idx])\n flow = readFlow(frame_files[idx], (height, width))\n min_val = flow.min()\n max_val = flow.max()\n flow = 2 * (flow - min_val) / (max_val - min_val) - 1\n frms.append(np.asarray(flow))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n # print('frms',frms.shape)\n # print(frms.shape)\n frms = torch.from_numpy(frms) #.unsqueeze(1) # (T, 1, H, W)\n frms = frms.permute(0,3,1,2)\n return frms\n\ndef load_depth(frames_dir, indices,\n height=-1, width=-1, invalid_val=-99):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print(frame_files)\n # print(len(indices), indices)\n frms = []\n for idx in indices:\n depth = Image.open(frame_files[idx])\n depth = depth.resize((width, height))\n depth = np.asarray(depth)\n\n invalid_mask = depth == invalid_val\n mask = np.logical_not(invalid_mask)\n vmin = np.percentile(depth[mask],2) \n vmax = np.percentile(depth[mask],85)\n if vmin != vmax:\n depth = (depth - vmin) / (vmax - vmin) \n else:","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.load_depth","uri":"program://CREMA/function/lavis.datasets.data_utils.load_depth#L89-L115","kind":"function","name":"load_depth","path":"lavis/datasets/data_utils.py","language":"python","start_line":89,"end_line":115,"context_start_line":69,"context_end_line":135,"code":" height=-1, width=-1):\n\n #print(frames_dir)\n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['flo']])\n frms = []\n for idx in indices:\n # print(idx, frame_files[idx])\n flow = readFlow(frame_files[idx], (height, width))\n min_val = flow.min()\n max_val = flow.max()\n flow = 2 * (flow - min_val) / (max_val - min_val) - 1\n frms.append(np.asarray(flow))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n # print('frms',frms.shape)\n # print(frms.shape)\n frms = torch.from_numpy(frms) #.unsqueeze(1) # (T, 1, H, W)\n frms = frms.permute(0,3,1,2)\n return frms\n\ndef load_depth(frames_dir, indices,\n height=-1, width=-1, invalid_val=-99):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print(frame_files)\n # print(len(indices), indices)\n frms = []\n for idx in indices:\n depth = Image.open(frame_files[idx])\n depth = depth.resize((width, height))\n depth = np.asarray(depth)\n\n invalid_mask = depth == invalid_val\n mask = np.logical_not(invalid_mask)\n vmin = np.percentile(depth[mask],2) \n vmax = np.percentile(depth[mask],85)\n if vmin != vmax:\n depth = (depth - vmin) / (vmax - vmin) \n else:\n depth = depth * 0.\n\n frms.append(np.asarray(depth))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms).unsqueeze(1) # (1, C, T, H, W)\n\n return frms\n\ndef load_frames(frames_dir, n_frms=MAX_INT, height=-1, width=-1, \n sampling=\"uniform\", clip_proposal=None, type='rgb', indices=None):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print('frame_files', len(frame_files))\n vlen = len(frame_files) - 1 # for flow\n n_frms = n_frms #min(n_frms, vlen)\n \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*vlen), int(clip_proposal[1]*vlen)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = [(intervals[i], intervals[i + 1]) for i in range(len(intervals) - 1)]","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.load_frames","uri":"program://CREMA/function/lavis.datasets.data_utils.load_frames#L117-L181","kind":"function","name":"load_frames","path":"lavis/datasets/data_utils.py","language":"python","start_line":117,"end_line":181,"context_start_line":97,"context_end_line":201,"code":" depth = Image.open(frame_files[idx])\n depth = depth.resize((width, height))\n depth = np.asarray(depth)\n\n invalid_mask = depth == invalid_val\n mask = np.logical_not(invalid_mask)\n vmin = np.percentile(depth[mask],2) \n vmax = np.percentile(depth[mask],85)\n if vmin != vmax:\n depth = (depth - vmin) / (vmax - vmin) \n else:\n depth = depth * 0.\n\n frms.append(np.asarray(depth))\n \n frms = np.stack(frms).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms).unsqueeze(1) # (1, C, T, H, W)\n\n return frms\n\ndef load_frames(frames_dir, n_frms=MAX_INT, height=-1, width=-1, \n sampling=\"uniform\", clip_proposal=None, type='rgb', indices=None):\n \n frame_files = sorted([os.path.join(frames_dir, f) for f in os.listdir(frames_dir) if f.split('.')[-1] in ['jpg', 'png']])\n # print('frame_files', len(frame_files))\n vlen = len(frame_files) - 1 # for flow\n n_frms = n_frms #min(n_frms, vlen)\n \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*vlen), int(clip_proposal[1]*vlen)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = [(intervals[i], intervals[i + 1]) for i in range(len(intervals) - 1)]\n \n if indices is None:\n if sampling == 'random':\n indices_ = [rnd.choice(range(x[0], x[1])) if x[0] != x[1] else x[0] for x in ranges]\n elif sampling == 'uniform':\n indices_ = [(x[0] + x[1]) // 2 for x in ranges]\n # elif sampling == \"headtail\":\n # indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n # indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n # indices = indices_h + indices_t\n else:\n raise NotImplementedError\n else:\n indices_ = indices\n \n # print(len(indices_) , n_frms, indices_)\n\n if len(indices_) < n_frms:\n extra = [indices_[-1] for i in range(n_frms - len(indices_))]\n indices_.extend(extra)\n frms = []\n # debug = [frame_files[idx] for idx in indices]\n # print(debug)\n # print('frame_files',len(frame_files))\n # print('indices_', indices_)\n for idx in indices_:\n # print(frame_files[idx])\n with Image.open(frame_files[idx]) as img:\n if type == 'norm':\n img = img.convert('RGB')\n if type == 'flow':\n img = img.transpose(Image.FLIP_LEFT_RIGHT)\n img = img.transpose(2) # ROTATE_90\n # print('here')\n if type == 'depth':\n img = img.convert('RGB')\n \n if height > 0 and width > 0:\n img = img.resize((width, height))\n frms.append(np.asarray(img))\n \n frms = np.stack(frms).transpose(3, 0, 1, 2).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms)\n # print('frms', frms.shape) # 3, 4, 224, 224 for rgb, 4, 4, 224, 224 for norm, 3, 4, 224, 224 for norm.convert('RGB')\n\n return frms, indices_\n\n# add for loading video\ndef load_video(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.load_video","uri":"program://CREMA/function/lavis.datasets.data_utils.load_video#L184-L227","kind":"function","name":"load_video","path":"lavis/datasets/data_utils.py","language":"python","start_line":184,"end_line":227,"context_start_line":164,"context_end_line":247,"code":" if type == 'norm':\n img = img.convert('RGB')\n if type == 'flow':\n img = img.transpose(Image.FLIP_LEFT_RIGHT)\n img = img.transpose(2) # ROTATE_90\n # print('here')\n if type == 'depth':\n img = img.convert('RGB')\n \n if height > 0 and width > 0:\n img = img.resize((width, height))\n frms.append(np.asarray(img))\n \n frms = np.stack(frms).transpose(3, 0, 1, 2).astype(np.float32) # (C, T, H, W)\n frms = torch.from_numpy(frms)\n # print('frms', frms.shape) # 3, 4, 224, 224 for rgb, 4, 4, 224, 224 for norm, 3, 4, 224, 224 for norm.convert('RGB')\n\n return frms, indices_\n\n# add for loading video\ndef load_video(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))\n\n if sampling == 'random':\n indices = []\n for x in ranges:\n if x[0] == x[1]:\n indices.append(x[0])\n else:\n indices.append(rnd.choice(range(x[0], x[1])))\n elif sampling == 'uniform':\n \n indices = [(x[0] + x[1]) // 2 for x in ranges]\n\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n frms = vr.get_batch(indices).permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps\n\ndef load_video_demo(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))\n","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.load_video_demo","uri":"program://CREMA/function/lavis.datasets.data_utils.load_video_demo#L229-L276","kind":"function","name":"load_video_demo","path":"lavis/datasets/data_utils.py","language":"python","start_line":229,"end_line":276,"context_start_line":209,"context_end_line":296,"code":" indices.append(rnd.choice(range(x[0], x[1])))\n elif sampling == 'uniform':\n \n indices = [(x[0] + x[1]) // 2 for x in ranges]\n\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n frms = vr.get_batch(indices).permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps\n\ndef load_video_demo(video_path, n_frms=MAX_INT, height=-1, width=-1, sampling=\"uniform\", clip_proposal=None):\n vr = VideoReader(uri=video_path, height=height, width=width)\n vlen = len(vr)\n n_frms = min(n_frms, vlen)\n fps = vr.get_avg_fps() \n if clip_proposal is None:\n start, end = 0, vlen\n else:\n start, end = int(clip_proposal[0]*fps), int(clip_proposal[1]*fps)\n if start < 0:\n start = 0\n if end > vlen:\n end = vlen\n\n intervals = np.linspace(start=start, stop=end, num=n_frms + 1).astype(int)\n ranges = []\n for idx, interv in enumerate(intervals[:-1]):\n ranges.append((interv, intervals[idx + 1]))\n\n if sampling == 'random':\n indices = []\n for x in ranges:\n if x[0] == x[1]:\n indices.append(x[0])\n else:\n indices.append(rnd.choice(range(x[0], x[1])))\n elif sampling == 'uniform':\n \n indices = [(x[0] + x[1]) // 2 for x in ranges]\n\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n \n frms = vr.get_batch(indices)\n frms = frms.asnumpy()\n frms = torch.from_numpy(frms)\n frms = frms.permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.apply_to_sample","uri":"program://CREMA/function/lavis.datasets.data_utils.apply_to_sample#L278-L292","kind":"function","name":"apply_to_sample","path":"lavis/datasets/data_utils.py","language":"python","start_line":278,"end_line":292,"context_start_line":258,"context_end_line":312,"code":"\n elif sampling == \"headtail\":\n indices_h = sorted(rnd.sample(range(vlen // 2), n_frms // 2))\n indices_t = sorted(rnd.sample(range(vlen // 2, vlen), n_frms // 2))\n indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n \n frms = vr.get_batch(indices)\n frms = frms.asnumpy()\n frms = torch.from_numpy(frms)\n frms = frms.permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.move_to_cuda","uri":"program://CREMA/function/lavis.datasets.data_utils.move_to_cuda#L295-L299","kind":"function","name":"move_to_cuda","path":"lavis/datasets/data_utils.py","language":"python","start_line":295,"end_line":299,"context_start_line":275,"context_end_line":319,"code":"\n return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"\n Organizes datasets by split.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by name.\n\n Returns:\n Dict of datasets by split {split_name: List[Datasets]}.","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.prepare_sample","uri":"program://CREMA/function/lavis.datasets.data_utils.prepare_sample#L302-L308","kind":"function","name":"prepare_sample","path":"lavis/datasets/data_utils.py","language":"python","start_line":302,"end_line":308,"context_start_line":282,"context_end_line":328,"code":" def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"\n Organizes datasets by split.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by name.\n\n Returns:\n Dict of datasets by split {split_name: List[Datasets]}.\n \"\"\"\n # if len(datasets) == 1:\n # return datasets[list(datasets.keys())[0]]\n # else:\n reorg_datasets = dict()\n\n # reorganize by split\n for _, dataset in datasets.items():\n for split_name, dataset_split in dataset.items():","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.reorg_datasets_by_split","uri":"program://CREMA/function/lavis.datasets.data_utils.reorg_datasets_by_split#L311-L334","kind":"function","name":"reorg_datasets_by_split","path":"lavis/datasets/data_utils.py","language":"python","start_line":311,"end_line":334,"context_start_line":291,"context_end_line":354,"code":"\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"\n Organizes datasets by split.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by name.\n\n Returns:\n Dict of datasets by split {split_name: List[Datasets]}.\n \"\"\"\n # if len(datasets) == 1:\n # return datasets[list(datasets.keys())[0]]\n # else:\n reorg_datasets = dict()\n\n # reorganize by split\n for _, dataset in datasets.items():\n for split_name, dataset_split in dataset.items():\n if split_name not in reorg_datasets:\n reorg_datasets[split_name] = [dataset_split]\n else:\n reorg_datasets[split_name].append(dataset_split)\n\n return reorg_datasets\n\n\ndef concat_datasets(datasets):\n \"\"\"\n Concatenates multiple datasets into a single dataset.\n\n It supports may-style datasets and DataPipeline from WebDataset. Currently, does not support\n generic IterableDataset because it requires creating separate samplers.\n\n Now only supports conctenating training datasets and assuming validation and testing\n have only a single dataset. This is because metrics should not be computed on the concatenated\n datasets.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by split.\n\n Returns:\n Dict of concatenated datasets by split, \"train\" is the concatenation of multiple datasets,\n \"val\" and \"test\" remain the same.\n","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.concat_datasets","uri":"program://CREMA/function/lavis.datasets.data_utils.concat_datasets#L337-L401","kind":"function","name":"concat_datasets","path":"lavis/datasets/data_utils.py","language":"python","start_line":337,"end_line":401,"context_start_line":317,"context_end_line":421,"code":"\n Returns:\n Dict of datasets by split {split_name: List[Datasets]}.\n \"\"\"\n # if len(datasets) == 1:\n # return datasets[list(datasets.keys())[0]]\n # else:\n reorg_datasets = dict()\n\n # reorganize by split\n for _, dataset in datasets.items():\n for split_name, dataset_split in dataset.items():\n if split_name not in reorg_datasets:\n reorg_datasets[split_name] = [dataset_split]\n else:\n reorg_datasets[split_name].append(dataset_split)\n\n return reorg_datasets\n\n\ndef concat_datasets(datasets):\n \"\"\"\n Concatenates multiple datasets into a single dataset.\n\n It supports may-style datasets and DataPipeline from WebDataset. Currently, does not support\n generic IterableDataset because it requires creating separate samplers.\n\n Now only supports conctenating training datasets and assuming validation and testing\n have only a single dataset. This is because metrics should not be computed on the concatenated\n datasets.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by split.\n\n Returns:\n Dict of concatenated datasets by split, \"train\" is the concatenation of multiple datasets,\n \"val\" and \"test\" remain the same.\n\n If the input training datasets contain both map-style and DataPipeline datasets, returns\n a tuple, where the first element is a concatenated map-style dataset and the second\n element is a chained DataPipeline dataset.\n\n \"\"\"\n # concatenate datasets in the same split\n for split_name in datasets:\n if split_name != \"train\":\n assert (\n len(datasets[split_name]) == 1\n ), \"Do not support multiple {} datasets.\".format(split_name)\n datasets[split_name] = datasets[split_name][0]\n else:\n iterable_datasets, map_datasets = [], []\n for dataset in datasets[split_name]:\n if isinstance(dataset, wds.DataPipeline):\n logging.info(\n \"Dataset {} is IterableDataset, can't be concatenated.\".format(\n dataset\n )\n )\n iterable_datasets.append(dataset)\n elif isinstance(dataset, IterableDataset):\n raise NotImplementedError(\n \"Do not support concatenation of generic IterableDataset.\"\n )\n else:\n map_datasets.append(dataset)\n\n # if len(iterable_datasets) > 0:\n # concatenate map-style datasets and iterable-style datasets separately\n chained_datasets = (\n ChainDataset(iterable_datasets) if len(iterable_datasets) > 0 else None\n )\n concat_datasets = (\n ConcatDataset(map_datasets) if len(map_datasets) > 0 else None\n )\n\n train_datasets = concat_datasets, chained_datasets\n train_datasets = tuple([x for x in train_datasets if x is not None])\n train_datasets = (\n train_datasets[0] if len(train_datasets) == 1 else train_datasets\n )\n\n datasets[split_name] = train_datasets\n\n return datasets\n\n\ndef extract_archive(from_path, to_path=None, overwrite=False):\n \"\"\"Extract archive.\n\n Args:\n from_path: the path of the archive.\n to_path: the root path of the extracted files (directory of from_path)\n overwrite: overwrite existing files (False)\n\n Returns:\n List of paths to extracted files even if not overwritten.\n\n Examples:\n >>> url = 'http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz'\n >>> from_path = './validation.tar.gz'\n >>> to_path = './'\n >>> torchtext.utils.download_from_url(url, from_path)\n >>> torchtext.utils.extract_archive(from_path, to_path)\n >>> ['.data/val.de', '.data/val.en']","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.extract_archive","uri":"program://CREMA/function/lavis.datasets.data_utils.extract_archive#L404-L483","kind":"function","name":"extract_archive","path":"lavis/datasets/data_utils.py","language":"python","start_line":404,"end_line":483,"context_start_line":384,"context_end_line":503,"code":" # if len(iterable_datasets) > 0:\n # concatenate map-style datasets and iterable-style datasets separately\n chained_datasets = (\n ChainDataset(iterable_datasets) if len(iterable_datasets) > 0 else None\n )\n concat_datasets = (\n ConcatDataset(map_datasets) if len(map_datasets) > 0 else None\n )\n\n train_datasets = concat_datasets, chained_datasets\n train_datasets = tuple([x for x in train_datasets if x is not None])\n train_datasets = (\n train_datasets[0] if len(train_datasets) == 1 else train_datasets\n )\n\n datasets[split_name] = train_datasets\n\n return datasets\n\n\ndef extract_archive(from_path, to_path=None, overwrite=False):\n \"\"\"Extract archive.\n\n Args:\n from_path: the path of the archive.\n to_path: the root path of the extracted files (directory of from_path)\n overwrite: overwrite existing files (False)\n\n Returns:\n List of paths to extracted files even if not overwritten.\n\n Examples:\n >>> url = 'http://www.quest.dcs.shef.ac.uk/wmt16_files_mmt/validation.tar.gz'\n >>> from_path = './validation.tar.gz'\n >>> to_path = './'\n >>> torchtext.utils.download_from_url(url, from_path)\n >>> torchtext.utils.extract_archive(from_path, to_path)\n >>> ['.data/val.de', '.data/val.en']\n >>> torchtext.utils.download_from_url(url, from_path)\n >>> torchtext.utils.extract_archive(from_path, to_path)\n >>> ['.data/val.de', '.data/val.en']\n\n \"\"\"\n\n if to_path is None:\n to_path = os.path.dirname(from_path)\n\n if from_path.endswith((\".tar.gz\", \".tgz\")):\n logging.info(\"Opening tar file {} to {}.\".format(from_path, to_path))\n with tarfile.open(from_path, \"r\") as tar:\n files = []\n for file_ in tqdm(tar):\n file_path = os.path.join(to_path, file_.name)\n if file_.isfile():\n files.append(file_path)\n if os.path.exists(file_path):\n logging.info(\"{} already extracted.\".format(file_path))\n if not overwrite:\n continue\n tar.extract(file_, to_path)\n logging.info(\"Finished extracting tar file {}.\".format(from_path))\n return files\n\n elif from_path.endswith(\".zip\"):\n assert zipfile.is_zipfile(from_path), from_path\n logging.info(\"Opening zip file {} to {}.\".format(from_path, to_path))\n with zipfile.ZipFile(from_path, \"r\") as zfile:\n files = []\n for file_ in tqdm(zfile.namelist()):\n file_path = os.path.join(to_path, file_)\n files.append(file_path)\n if os.path.exists(file_path):\n logging.info(\"{} already extracted.\".format(file_path))\n if not overwrite:\n continue\n zfile.extract(file_, to_path)\n files = [f for f in files if os.path.isfile(f)]\n logging.info(\"Finished extracting zip file {}.\".format(from_path))\n return files\n\n elif from_path.endswith(\".gz\"):\n logging.info(\"Opening gz file {} to {}.\".format(from_path, to_path))\n default_block_size = 65536\n filename = from_path[:-3]\n files = [filename]\n with gzip.open(from_path, \"rb\") as gzfile, open(filename, \"wb\") as d_file:\n while True:\n block = gzfile.read(default_block_size)\n if not block:\n break\n else:\n d_file.write(block)\n d_file.write(block)\n logging.info(\"Finished extracting gz file {}.\".format(from_path))\n return files\n\n else:\n raise NotImplementedError(\n \"We currently only support tar.gz, .tgz, .gz and zip achives.\"\n )\n\n\ndef save_frames_grid(img_array, out_path):\n import torch\n from PIL import Image\n from torchvision.utils import make_grid\n\n if len(img_array.shape) == 3:\n img_array = img_array.unsqueeze(0)\n elif len(img_array.shape) == 5:\n b, t, c, h, w = img_array.shape\n img_array = img_array.view(-1, c, h, w)\n elif len(img_array.shape) == 4:\n pass\n else:\n raise NotImplementedError(\n \"Supports only (b,t,c,h,w)-shaped inputs. First two dimensions can be ignored.\"\n )\n\n assert img_array.shape[1] == 3, \"Exepcting input shape of (H, W, 3), i.e. RGB-only.\"","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils.save_frames_grid","uri":"program://CREMA/function/lavis.datasets.data_utils.save_frames_grid#L486-L510","kind":"function","name":"save_frames_grid","path":"lavis/datasets/data_utils.py","language":"python","start_line":486,"end_line":510,"context_start_line":466,"context_end_line":510,"code":" default_block_size = 65536\n filename = from_path[:-3]\n files = [filename]\n with gzip.open(from_path, \"rb\") as gzfile, open(filename, \"wb\") as d_file:\n while True:\n block = gzfile.read(default_block_size)\n if not block:\n break\n else:\n d_file.write(block)\n d_file.write(block)\n logging.info(\"Finished extracting gz file {}.\".format(from_path))\n return files\n\n else:\n raise NotImplementedError(\n \"We currently only support tar.gz, .tgz, .gz and zip achives.\"\n )\n\n\ndef save_frames_grid(img_array, out_path):\n import torch\n from PIL import Image\n from torchvision.utils import make_grid\n\n if len(img_array.shape) == 3:\n img_array = img_array.unsqueeze(0)\n elif len(img_array.shape) == 5:\n b, t, c, h, w = img_array.shape\n img_array = img_array.view(-1, c, h, w)\n elif len(img_array.shape) == 4:\n pass\n else:\n raise NotImplementedError(\n \"Supports only (b,t,c,h,w)-shaped inputs. First two dimensions can be ignored.\"\n )\n\n assert img_array.shape[1] == 3, \"Exepcting input shape of (H, W, 3), i.e. RGB-only.\"\n\n grid = make_grid(img_array)\n ndarr = grid.permute(1, 2, 0).to(\"cpu\", torch.uint8).numpy()\n\n img = Image.fromarray(ndarr)\n\n img.save(out_path)","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils._apply","uri":"program://CREMA/function/lavis.datasets.data_utils._apply#L282-L290","kind":"function","name":"_apply","path":"lavis/datasets/data_utils.py","language":"python","start_line":282,"end_line":290,"context_start_line":262,"context_end_line":310,"code":" indices = indices_h + indices_t\n else:\n raise NotImplementedError\n \n if len(indices) < n_frms:\n rest = [indices[-1] for i in range(n_frms - len(indices))]\n indices = indices + rest \n # get_batch -> T, H, W, C\n \n frms = vr.get_batch(indices)\n frms = frms.asnumpy()\n frms = torch.from_numpy(frms)\n frms = frms.permute(3, 0, 1, 2).float() # (C, T, H, W)\n\n return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.data_utils._move_to_cuda","uri":"program://CREMA/function/lavis.datasets.data_utils._move_to_cuda#L296-L297","kind":"function","name":"_move_to_cuda","path":"lavis/datasets/data_utils.py","language":"python","start_line":296,"end_line":297,"context_start_line":276,"context_end_line":317,"code":" return frms, indices, fps, vlen\n\ndef apply_to_sample(f, sample):\n if len(sample) == 0:\n return {}\n\n def _apply(x):\n if torch.is_tensor(x):\n return f(x)\n elif isinstance(x, dict):\n return {key: _apply(value) for key, value in x.items()}\n elif isinstance(x, list):\n return [_apply(x) for x in x]\n else:\n return x\n\n return _apply(sample)\n\n\ndef move_to_cuda(sample):\n def _move_to_cuda(tensor):\n return tensor.cuda()\n\n return apply_to_sample(_move_to_cuda, sample)\n\n\ndef prepare_sample(samples, cuda_enabled=True):\n if cuda_enabled:\n samples = move_to_cuda(samples)\n\n # TODO fp16 support\n\n return samples\n\n\ndef reorg_datasets_by_split(datasets):\n \"\"\"\n Organizes datasets by split.\n\n Args:\n datasets: dict of torch.utils.data.Dataset objects by name.\n","source_hash":"5357d6f7a73e67412b049d21cdd6b20f33d1ad1cc936c269da0084248b5e51a0","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msrvtt","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_msrvtt#L1-L105","kind":"module","name":"lavis.datasets.download_scripts.download_msrvtt","path":"lavis/datasets/download_scripts/download_msrvtt.py","language":"python","start_line":1,"end_line":105,"context_start_line":1,"context_end_line":105,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\n# TODO\n# 1. Go to https://www.mediafire.com/file/czh8sezbo9s4692/test_videos.zip/file\n# and https://www.mediafire.com/file/x3rrbe4hwp04e6w/train_val_videos.zip/file\n# 2. Right-click the Download button and copy the link address\n# e.g.\n# DATA_URL = {\n# \"train\": \"https://download1602.mediafire.com/xxxxxxxxxxxx/x3rrbe4hwp04e6w/train_val_videos.zip\",\n# \"test\": \"https://download2390.mediafire.com/xxxxxxxxxxxx/czh8sezbo9s4692/test_videos.zip\",\n# }\n# 3. Paste the link address to DATA_URL\n\nDATA_URL = {\n \"train\": \"https://download2295.mediafire.com/4bb7p74xrbgg/x3rrbe4hwp04e6w/train_val_videos.zip\",\n \"test\": \"https://download2390.mediafire.com/79hfq3592lqg/czh8sezbo9s4692/test_videos.zip\",\n}\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef merge_datasets(download_path, storage_path):\n \"\"\"\n Merge datasets in download_path to storage_path\n \"\"\"\n\n # Merge train and test datasets\n train_path = os.path.join(download_path, \"TrainValVideo\")\n test_path = os.path.join(download_path, \"TestVideo\")\n train_test_path = storage_path\n\n print(\"Merging to {}\".format(train_test_path))\n\n os.makedirs(train_test_path, exist_ok=True)\n\n for file_name in os.listdir(train_path):\n os.rename(\n os.path.join(train_path, file_name),\n os.path.join(train_test_path, file_name),\n )\n\n for file_name in os.listdir(test_path):\n os.rename(\n os.path.join(test_path, file_name),\n os.path.join(train_test_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/msrvtt/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.msrvtt_cap.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))\n download_datasets(download_dir, v)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n try:\n merge_datasets(download_dir, storage_dir)\n except Exception as e:\n # remove storage dir if failed\n cleanup_dir(download_dir)\n cleanup_dir(storage_dir)\n print(\"Failed to merging datasets. Aborting.\")\n\n cleanup_dir(download_dir)","source_hash":"709740ed1d91d51fddb1f16b02a087f354f642ebbaaeaf3866141d0a752195b2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msrvtt.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_msrvtt.download_datasets#L38-L43","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_msrvtt.py","language":"python","start_line":38,"end_line":43,"context_start_line":18,"context_end_line":63,"code":")\n\n\n# TODO\n# 1. Go to https://www.mediafire.com/file/czh8sezbo9s4692/test_videos.zip/file\n# and https://www.mediafire.com/file/x3rrbe4hwp04e6w/train_val_videos.zip/file\n# 2. Right-click the Download button and copy the link address\n# e.g.\n# DATA_URL = {\n# \"train\": \"https://download1602.mediafire.com/xxxxxxxxxxxx/x3rrbe4hwp04e6w/train_val_videos.zip\",\n# \"test\": \"https://download2390.mediafire.com/xxxxxxxxxxxx/czh8sezbo9s4692/test_videos.zip\",\n# }\n# 3. Paste the link address to DATA_URL\n\nDATA_URL = {\n \"train\": \"https://download2295.mediafire.com/4bb7p74xrbgg/x3rrbe4hwp04e6w/train_val_videos.zip\",\n \"test\": \"https://download2390.mediafire.com/79hfq3592lqg/czh8sezbo9s4692/test_videos.zip\",\n}\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef merge_datasets(download_path, storage_path):\n \"\"\"\n Merge datasets in download_path to storage_path\n \"\"\"\n\n # Merge train and test datasets\n train_path = os.path.join(download_path, \"TrainValVideo\")\n test_path = os.path.join(download_path, \"TestVideo\")\n train_test_path = storage_path\n\n print(\"Merging to {}\".format(train_test_path))\n\n os.makedirs(train_test_path, exist_ok=True)\n\n for file_name in os.listdir(train_path):\n os.rename(\n os.path.join(train_path, file_name),\n os.path.join(train_test_path, file_name),","source_hash":"709740ed1d91d51fddb1f16b02a087f354f642ebbaaeaf3866141d0a752195b2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msrvtt.merge_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_msrvtt.merge_datasets#L46-L70","kind":"function","name":"merge_datasets","path":"lavis/datasets/download_scripts/download_msrvtt.py","language":"python","start_line":46,"end_line":70,"context_start_line":26,"context_end_line":90,"code":"# DATA_URL = {\n# \"train\": \"https://download1602.mediafire.com/xxxxxxxxxxxx/x3rrbe4hwp04e6w/train_val_videos.zip\",\n# \"test\": \"https://download2390.mediafire.com/xxxxxxxxxxxx/czh8sezbo9s4692/test_videos.zip\",\n# }\n# 3. Paste the link address to DATA_URL\n\nDATA_URL = {\n \"train\": \"https://download2295.mediafire.com/4bb7p74xrbgg/x3rrbe4hwp04e6w/train_val_videos.zip\",\n \"test\": \"https://download2390.mediafire.com/79hfq3592lqg/czh8sezbo9s4692/test_videos.zip\",\n}\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef merge_datasets(download_path, storage_path):\n \"\"\"\n Merge datasets in download_path to storage_path\n \"\"\"\n\n # Merge train and test datasets\n train_path = os.path.join(download_path, \"TrainValVideo\")\n test_path = os.path.join(download_path, \"TestVideo\")\n train_test_path = storage_path\n\n print(\"Merging to {}\".format(train_test_path))\n\n os.makedirs(train_test_path, exist_ok=True)\n\n for file_name in os.listdir(train_path):\n os.rename(\n os.path.join(train_path, file_name),\n os.path.join(train_test_path, file_name),\n )\n\n for file_name in os.listdir(test_path):\n os.rename(\n os.path.join(test_path, file_name),\n os.path.join(train_test_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/msrvtt/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.msrvtt_cap.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))","source_hash":"709740ed1d91d51fddb1f16b02a087f354f642ebbaaeaf3866141d0a752195b2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_flickr","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_flickr#L1-L78","kind":"module","name":"lavis.datasets.download_scripts.download_flickr","path":"lavis/datasets/download_scripts/download_flickr.py","language":"python","start_line":1,"end_line":78,"context_start_line":1,"context_end_line":78,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n get_abs_path,\n get_cache_path,\n)\n\nimport opendatasets as od\n\n\nDATA_URL = \"https://www.kaggle.com/datasets/hsankesara/flickr-image-dataset\"\n\nprint(\n \"\"\"\n To download the dataset, you need to have a Kaggle account and the associated key.\n See https://www.kaggle.com/docs/api to create account and a new API token.\n \"\"\"\n)\n\n\ndef move_directory(src_dir, dst_dir):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(dst_dir))\n\n os.makedirs(dst_dir, exist_ok=True)\n\n for file_name in os.listdir(src_dir):\n os.rename(\n os.path.join(src_dir, file_name),\n os.path.join(dst_dir, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/flickr30k/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.flickr30k.build_info.images.storage\n\n storage_dir = Path(get_cache_path(storage_dir))\n download_dir = storage_dir.parent / \"download\"\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n os.makedirs(download_dir)\n\n try:\n print(\"Downloading {} to {}\".format(DATA_URL, download_dir))\n od.download(DATA_URL, download_dir)\n except Exception as e:\n print(e)\n # remove download dir if failed\n cleanup_dir(download_dir)\n exit(1)\n\n move_directory(\n download_dir / \"flickr-image-dataset\" / \"flickr30k_images\" / \"flickr30k_images\",\n storage_dir / \"flickr30k-images\",\n )\n\n cleanup_dir(download_dir)","source_hash":"57b474f161868593352dae9acd81d4db37c8bd7afbd7728b948d2adc6a21fc93","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_flickr.move_directory","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_flickr.move_directory#L32-L44","kind":"function","name":"move_directory","path":"lavis/datasets/download_scripts/download_flickr.py","language":"python","start_line":32,"end_line":44,"context_start_line":12,"context_end_line":64,"code":"\nfrom lavis.common.utils import (\n cleanup_dir,\n get_abs_path,\n get_cache_path,\n)\n\nimport opendatasets as od\n\n\nDATA_URL = \"https://www.kaggle.com/datasets/hsankesara/flickr-image-dataset\"\n\nprint(\n \"\"\"\n To download the dataset, you need to have a Kaggle account and the associated key.\n See https://www.kaggle.com/docs/api to create account and a new API token.\n \"\"\"\n)\n\n\ndef move_directory(src_dir, dst_dir):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(dst_dir))\n\n os.makedirs(dst_dir, exist_ok=True)\n\n for file_name in os.listdir(src_dir):\n os.rename(\n os.path.join(src_dir, file_name),\n os.path.join(dst_dir, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/flickr30k/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.flickr30k.build_info.images.storage\n\n storage_dir = Path(get_cache_path(storage_dir))\n download_dir = storage_dir.parent / \"download\"\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n os.makedirs(download_dir)\n\n try:","source_hash":"57b474f161868593352dae9acd81d4db37c8bd7afbd7728b948d2adc6a21fc93","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_sbu","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_sbu#L1-L82","kind":"module","name":"lavis.datasets.download_scripts.download_sbu","path":"lavis/datasets/download_scripts/download_sbu.py","language":"python","start_line":1,"end_line":82,"context_start_line":1,"context_end_line":82,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport io\nimport os\nimport pathlib\nimport urllib\nimport tqdm\n\nfrom concurrent.futures import ThreadPoolExecutor\n\nfrom lavis.common.utils import get_abs_path, get_cache_path\nfrom lavis.datasets.builders import load_dataset\nfrom omegaconf import OmegaConf\nfrom PIL import Image\n\n# DATA_URL = {\"train\": \"http://www.cs.rice.edu/~vo9/sbucaptions/sbu_images.tar\"}\n\nUSER_AGENT = (\n \"Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:15.0) Gecko/20100101 Firefox/15.0.1\"\n)\n\n\ndef fetch_single_image(image_url, timeout=None, retries=0):\n for _ in range(retries + 1):\n try:\n request = urllib.request.Request(\n image_url,\n data=None,\n headers={\"user-agent\": USER_AGENT},\n )\n with urllib.request.urlopen(request, timeout=timeout) as req:\n image = Image.open(io.BytesIO(req.read()))\n break\n except Exception:\n image = None\n return image\n\n\ndef download_and_save_image(ann, save_dir, timeout=None, retries=0):\n image = fetch_single_image(ann[\"url\"], timeout=timeout, retries=retries)\n\n if image is not None:\n image_path = os.path.join(save_dir, ann[\"image\"])\n print(image_path)\n image.save(image_path)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/sbu_caption/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.sbu_caption.build_info.images.storage\n\n storage_dir = pathlib.Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n storage_dir.mkdir(parents=True, exist_ok=True)\n\n num_threads = 20\n dset = load_dataset(\"sbu_caption\")[\"train\"].annotation\n\n print(\"Downloading dataset...\")\n # multiprocessing\n with ThreadPoolExecutor(max_workers=num_threads) as executor:\n for ann in tqdm.tqdm(dset):\n executor.submit(\n download_and_save_image,\n ann,\n storage_dir,\n timeout=30,\n retries=10,\n )","source_hash":"ec80f88837dbe2d2b42abe647ed9b43e38a6dbfbf60c2755bad7ea212d76330e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_sbu.fetch_single_image","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_sbu.fetch_single_image#L28-L41","kind":"function","name":"fetch_single_image","path":"lavis/datasets/download_scripts/download_sbu.py","language":"python","start_line":28,"end_line":41,"context_start_line":8,"context_end_line":61,"code":"import io\nimport os\nimport pathlib\nimport urllib\nimport tqdm\n\nfrom concurrent.futures import ThreadPoolExecutor\n\nfrom lavis.common.utils import get_abs_path, get_cache_path\nfrom lavis.datasets.builders import load_dataset\nfrom omegaconf import OmegaConf\nfrom PIL import Image\n\n# DATA_URL = {\"train\": \"http://www.cs.rice.edu/~vo9/sbucaptions/sbu_images.tar\"}\n\nUSER_AGENT = (\n \"Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:15.0) Gecko/20100101 Firefox/15.0.1\"\n)\n\n\ndef fetch_single_image(image_url, timeout=None, retries=0):\n for _ in range(retries + 1):\n try:\n request = urllib.request.Request(\n image_url,\n data=None,\n headers={\"user-agent\": USER_AGENT},\n )\n with urllib.request.urlopen(request, timeout=timeout) as req:\n image = Image.open(io.BytesIO(req.read()))\n break\n except Exception:\n image = None\n return image\n\n\ndef download_and_save_image(ann, save_dir, timeout=None, retries=0):\n image = fetch_single_image(ann[\"url\"], timeout=timeout, retries=retries)\n\n if image is not None:\n image_path = os.path.join(save_dir, ann[\"image\"])\n print(image_path)\n image.save(image_path)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/sbu_caption/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.sbu_caption.build_info.images.storage\n\n storage_dir = pathlib.Path(get_cache_path(storage_dir))","source_hash":"ec80f88837dbe2d2b42abe647ed9b43e38a6dbfbf60c2755bad7ea212d76330e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_sbu.download_and_save_image","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_sbu.download_and_save_image#L44-L50","kind":"function","name":"download_and_save_image","path":"lavis/datasets/download_scripts/download_sbu.py","language":"python","start_line":44,"end_line":50,"context_start_line":24,"context_end_line":70,"code":" \"Mozilla/5.0 (X11; Ubuntu; Linux x86_64; rv:15.0) Gecko/20100101 Firefox/15.0.1\"\n)\n\n\ndef fetch_single_image(image_url, timeout=None, retries=0):\n for _ in range(retries + 1):\n try:\n request = urllib.request.Request(\n image_url,\n data=None,\n headers={\"user-agent\": USER_AGENT},\n )\n with urllib.request.urlopen(request, timeout=timeout) as req:\n image = Image.open(io.BytesIO(req.read()))\n break\n except Exception:\n image = None\n return image\n\n\ndef download_and_save_image(ann, save_dir, timeout=None, retries=0):\n image = fetch_single_image(ann[\"url\"], timeout=timeout, retries=retries)\n\n if image is not None:\n image_path = os.path.join(save_dir, ann[\"image\"])\n print(image_path)\n image.save(image_path)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/sbu_caption/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.sbu_caption.build_info.images.storage\n\n storage_dir = pathlib.Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n storage_dir.mkdir(parents=True, exist_ok=True)\n\n num_threads = 20\n dset = load_dataset(\"sbu_caption\")[\"train\"].annotation","source_hash":"ec80f88837dbe2d2b42abe647ed9b43e38a6dbfbf60c2755bad7ea212d76330e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_didemo","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_didemo#L1-L70","kind":"module","name":"lavis.datasets.download_scripts.download_didemo","path":"lavis/datasets/download_scripts/download_didemo.py","language":"python","start_line":1,"end_line":70,"context_start_line":1,"context_end_line":70,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\nDATA_URL = \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/didemo/didemo_videos.tar.gz\"\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/didemo/defaults_ret.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.didemo_retrieval.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {} to {}\".format(DATA_URL, download_dir))\n download_datasets(download_dir, DATA_URL)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n move_files(download_dir / \"videos\", storage_dir)\n cleanup_dir(download_dir)","source_hash":"c872f9ee39db03a7eda39d6e480a1412159d68d2a275361103aeb795d01a1f07","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_didemo.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_didemo.download_datasets#L23-L28","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_didemo.py","language":"python","start_line":23,"end_line":28,"context_start_line":3,"context_end_line":48,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\nDATA_URL = \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/didemo/didemo_videos.tar.gz\"\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/didemo/defaults_ret.yaml\")","source_hash":"c872f9ee39db03a7eda39d6e480a1412159d68d2a275361103aeb795d01a1f07","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_didemo.move_files","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_didemo.move_files#L31-L43","kind":"function","name":"move_files","path":"lavis/datasets/download_scripts/download_didemo.py","language":"python","start_line":31,"end_line":43,"context_start_line":11,"context_end_line":63,"code":"from omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\nDATA_URL = \"https://storage.googleapis.com/sfr-vision-language-research/LAVIS/datasets/didemo/didemo_videos.tar.gz\"\n\n\ndef download_datasets(root, url):\n \"\"\"\n Download the Imagenet-R dataset archives and expand them\n in the folder provided as parameter\n \"\"\"\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/didemo/defaults_ret.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.didemo_retrieval.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {} to {}\".format(DATA_URL, download_dir))\n download_datasets(download_dir, DATA_URL)","source_hash":"c872f9ee39db03a7eda39d6e480a1412159d68d2a275361103aeb795d01a1f07","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msvd","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_msvd#L1-L67","kind":"module","name":"lavis.datasets.download_scripts.download_msvd","path":"lavis/datasets/download_scripts/download_msvd.py","language":"python","start_line":1,"end_line":67,"context_start_line":1,"context_end_line":67,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = \"https://www.cs.utexas.edu/users/ml/clamp/videoDescription/YouTubeClips.tar\"\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/msvd/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.msvd_cap.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {}\".format(DATA_URL))\n download_datasets(download_dir, DATA_URL)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n move_files(download_dir / \"YouTubeClips\", storage_dir)\n cleanup_dir(download_dir)","source_hash":"8d87d539b0b7046e524b9e3bbfaadebc86450a4d37bd1a850c0946ba9d29d8a6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msvd.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_msvd.download_datasets#L24-L25","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_msvd.py","language":"python","start_line":24,"end_line":25,"context_start_line":4,"context_end_line":45,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = \"https://www.cs.utexas.edu/users/ml/clamp/videoDescription/YouTubeClips.tar\"\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/msvd/defaults_cap.yaml\")","source_hash":"8d87d539b0b7046e524b9e3bbfaadebc86450a4d37bd1a850c0946ba9d29d8a6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_msvd.move_files","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_msvd.move_files#L28-L40","kind":"function","name":"move_files","path":"lavis/datasets/download_scripts/download_msvd.py","language":"python","start_line":28,"end_line":40,"context_start_line":8,"context_end_line":60,"code":"import os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = \"https://www.cs.utexas.edu/users/ml/clamp/videoDescription/YouTubeClips.tar\"\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root)\n\n\ndef move_files(download_path, storage_path):\n \"\"\"\n Move files from download_path to storage_path\n \"\"\"\n print(\"Moving to {}\".format(storage_path))\n\n os.makedirs(storage_path, exist_ok=True)\n\n for file_name in os.listdir(download_path):\n os.rename(\n os.path.join(download_path, file_name),\n os.path.join(storage_path, file_name),\n )\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/msvd/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.msvd_cap.build_info.videos.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {}\".format(DATA_URL))\n download_datasets(download_dir, DATA_URL)","source_hash":"8d87d539b0b7046e524b9e3bbfaadebc86450a4d37bd1a850c0946ba9d29d8a6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_vg","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_vg#L1-L55","kind":"module","name":"lavis.datasets.download_scripts.download_vg","path":"lavis/datasets/download_scripts/download_vg.py","language":"python","start_line":1,"end_line":55,"context_start_line":1,"context_end_line":55,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = {\n \"train\": \"https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip\",\n \"train2\": \"https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip\",\n}\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/vg/defaults_caption.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.vg_caption.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))\n download_datasets(download_dir, v)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n cleanup_dir(download_dir)","source_hash":"a68b4630dc28d3510d591cfb5cb509fbc9b9727180d36dc5fb7058aa2243c2b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_vg.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_vg.download_datasets#L27-L28","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_vg.py","language":"python","start_line":27,"end_line":28,"context_start_line":7,"context_end_line":48,"code":"\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = {\n \"train\": \"https://cs.stanford.edu/people/rak248/VG_100K_2/images.zip\",\n \"train2\": \"https://cs.stanford.edu/people/rak248/VG_100K_2/images2.zip\",\n}\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/vg/defaults_caption.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.vg_caption.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))","source_hash":"a68b4630dc28d3510d591cfb5cb509fbc9b9727180d36dc5fb7058aa2243c2b5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_gqa","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_gqa#L1-L51","kind":"module","name":"lavis.datasets.download_scripts.download_gqa","path":"lavis/datasets/download_scripts/download_gqa.py","language":"python","start_line":1,"end_line":51,"context_start_line":1,"context_end_line":51,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = \"https://downloads.cs.stanford.edu/nlp/data/gqa/images.zip\"\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir.parent)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/gqa/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.gqa.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {}\".format(DATA_URL))\n download_datasets(download_dir, DATA_URL)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n cleanup_dir(download_dir)","source_hash":"42c5da6c7805a8a63c747d6711ef0719cde599bcfd211ded86affce80771ee43","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_gqa.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_gqa.download_datasets#L24-L25","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_gqa.py","language":"python","start_line":24,"end_line":25,"context_start_line":4,"context_end_line":45,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = \"https://downloads.cs.stanford.edu/nlp/data/gqa/images.zip\"\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir.parent)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/gqa/defaults.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.gqa.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n print(\"Downloading {}\".format(DATA_URL))\n download_datasets(download_dir, DATA_URL)","source_hash":"42c5da6c7805a8a63c747d6711ef0719cde599bcfd211ded86affce80771ee43","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_coco","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_coco#L1-L57","kind":"module","name":"lavis.datasets.download_scripts.download_coco","path":"lavis/datasets/download_scripts/download_coco.py","language":"python","start_line":1,"end_line":57,"context_start_line":1,"context_end_line":57,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = {\n \"train\": \"http://images.cocodataset.org/zips/train2014.zip\", # md5: 0da8c0bd3d6becc4dcb32757491aca88\n \"val\": \"http://images.cocodataset.org/zips/val2014.zip\", # md5: a3d79f5ed8d289b7a7554ce06a5782b3\n \"test\": \"http://images.cocodataset.org/zips/test2014.zip\", # md5: 04127eef689ceac55e3a572c2c92f264\n \"test2015\": \"http://images.cocodataset.org/zips/test2015.zip\", # md5: 04127eef689ceac55e3a572c2c92f264\n}\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/coco/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.coco_caption.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))\n download_datasets(download_dir, v)\n except Exception as e:\n # remove download dir if failed\n cleanup_dir(download_dir)\n print(\"Failed to download or extracting datasets. Aborting.\")\n\n cleanup_dir(download_dir)","source_hash":"57bda023bf11d0efbebff2053cfe192a1ea103d1c4bd0f48ca4c259e7b48b73e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_coco.download_datasets","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_coco.download_datasets#L29-L30","kind":"function","name":"download_datasets","path":"lavis/datasets/download_scripts/download_coco.py","language":"python","start_line":29,"end_line":30,"context_start_line":9,"context_end_line":50,"code":"from pathlib import Path\n\nfrom omegaconf import OmegaConf\n\nfrom lavis.common.utils import (\n cleanup_dir,\n download_and_extract_archive,\n get_abs_path,\n get_cache_path,\n)\n\n\nDATA_URL = {\n \"train\": \"http://images.cocodataset.org/zips/train2014.zip\", # md5: 0da8c0bd3d6becc4dcb32757491aca88\n \"val\": \"http://images.cocodataset.org/zips/val2014.zip\", # md5: a3d79f5ed8d289b7a7554ce06a5782b3\n \"test\": \"http://images.cocodataset.org/zips/test2014.zip\", # md5: 04127eef689ceac55e3a572c2c92f264\n \"test2015\": \"http://images.cocodataset.org/zips/test2015.zip\", # md5: 04127eef689ceac55e3a572c2c92f264\n}\n\n\ndef download_datasets(root, url):\n download_and_extract_archive(url=url, download_root=root, extract_root=storage_dir)\n\n\nif __name__ == \"__main__\":\n\n config_path = get_abs_path(\"configs/datasets/coco/defaults_cap.yaml\")\n\n storage_dir = OmegaConf.load(\n config_path\n ).datasets.coco_caption.build_info.images.storage\n\n download_dir = Path(get_cache_path(storage_dir)).parent / \"download\"\n storage_dir = Path(get_cache_path(storage_dir))\n\n if storage_dir.exists():\n print(f\"Dataset already exists at {storage_dir}. Aborting.\")\n exit(0)\n\n try:\n for k, v in DATA_URL.items():\n print(\"Downloading {} to {}\".format(v, k))","source_hash":"57bda023bf11d0efbebff2053cfe192a1ea103d1c4bd0f48ca4c259e7b48b73e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_nocaps","uri":"program://CREMA/module/lavis.datasets.download_scripts.download_nocaps#L1-L134","kind":"module","name":"lavis.datasets.download_scripts.download_nocaps","path":"lavis/datasets/download_scripts/download_nocaps.py","language":"python","start_line":1,"end_line":134,"context_start_line":1,"context_end_line":134,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport logging\nimport os\nimport time\nfrom multiprocessing import Pool\n\nimport numpy as np\nimport requests\nimport tqdm\nfrom lavis.common.utils import cleanup_dir, get_abs_path, get_cache_path\nfrom omegaconf import OmegaConf\n\nheader_mzl = {\n \"User-Agent\": \"Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36\",\n # \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n # \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\nheader_gbot = {\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n}\n\nheaders = [header_mzl, header_gbot]\n\n# Setup\nlogging.basicConfig(filename=\"download_nocaps.log\", filemode=\"w\", level=logging.INFO)\nrequests.packages.urllib3.disable_warnings(\n requests.packages.urllib3.exceptions.InsecureRequestWarning\n)\n\n\ndef download_file(url, filename):\n max_retries = 20\n cur_retries = 0\n\n header = headers[0]\n\n while cur_retries < max_retries:\n try:\n r = requests.get(url, headers=header, timeout=10)\n with open(filename, \"wb\") as f:\n f.write(r.content)\n\n break\n except Exception as e:\n logging.info(\" \".join(repr(e).splitlines()))\n logging.error(url)\n cur_retries += 1\n\n # random sample a header from headers\n header = headers[np.random.randint(0, len(headers))]\n\n time.sleep(3 + cur_retries * 2)\n\n\ndef download_image_from_url_val(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"val\", basename)\n\n download_file(url, filename)\n\n\ndef download_image_from_url_test(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"test\", basename)\n\n download_file(url, filename)\n\n\nif __name__ == \"__main__\":\n os.makedirs(\"tmp\", exist_ok=True)\n\n # storage dir\n config_path = get_abs_path(\"configs/datasets/nocaps/defaults.yaml\")\n\n storage_dir = OmegaConf.load(config_path).datasets.nocaps.build_info.images.storage\n storage_dir = get_cache_path(storage_dir)\n # make sure the storage dir exists\n os.makedirs(storage_dir, exist_ok=True)\n print(\"Storage dir:\", storage_dir)\n\n # make sure the storage dir for val and test exists\n os.makedirs(os.path.join(storage_dir, \"val\"), exist_ok=True)\n os.makedirs(os.path.join(storage_dir, \"test\"), exist_ok=True)\n\n # download annotations\n val_url = \"https://nocaps.s3.amazonaws.com/nocaps_val_4500_captions.json\"\n tst_url = \"https://s3.amazonaws.com/nocaps/nocaps_test_image_info.json\"\n\n print(\"Downloading validation annotations from %s\" % val_url)\n download_file(val_url, \"tmp/nocaps_val_ann.json\")\n print(\"Downloading testing annotations from %s\" % tst_url)\n download_file(tst_url, \"tmp/nocaps_tst_ann.json\")\n\n # open annotations\n val_ann = json.load(open(\"tmp/nocaps_val_ann.json\"))\n tst_ann = json.load(open(\"tmp/nocaps_tst_ann.json\"))\n\n # collect image urls\n val_info = val_ann[\"images\"]\n tst_info = tst_ann[\"images\"]\n\n val_urls = [info[\"coco_url\"] for info in val_info]\n tst_urls = [info[\"coco_url\"] for info in tst_info]\n\n # setup multiprocessing\n # large n_procs possibly causes server to reject requests\n n_procs = 16\n\n with Pool(n_procs) as pool:\n print(\"Downloading validation images...\")\n list(\n tqdm.tqdm(\n pool.imap(download_image_from_url_val, val_urls), total=len(val_urls)\n )\n )\n\n with Pool(n_procs) as pool:\n print(\"Downloading test images...\")\n list(\n tqdm.tqdm(\n pool.imap(download_image_from_url_test, tst_urls), total=len(tst_urls)\n )\n )\n\n # clean tmp\n cleanup_dir(\"tmp\")","source_hash":"bad5e9e9e38e9001c05a04bad66598cc3c05037cc5d506ee633804f12d3cfff3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_nocaps.download_file","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_nocaps.download_file#L39-L60","kind":"function","name":"download_file","path":"lavis/datasets/download_scripts/download_nocaps.py","language":"python","start_line":39,"end_line":60,"context_start_line":19,"context_end_line":80,"code":"\nheader_mzl = {\n \"User-Agent\": \"Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36\",\n # \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n # \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\nheader_gbot = {\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n}\n\nheaders = [header_mzl, header_gbot]\n\n# Setup\nlogging.basicConfig(filename=\"download_nocaps.log\", filemode=\"w\", level=logging.INFO)\nrequests.packages.urllib3.disable_warnings(\n requests.packages.urllib3.exceptions.InsecureRequestWarning\n)\n\n\ndef download_file(url, filename):\n max_retries = 20\n cur_retries = 0\n\n header = headers[0]\n\n while cur_retries < max_retries:\n try:\n r = requests.get(url, headers=header, timeout=10)\n with open(filename, \"wb\") as f:\n f.write(r.content)\n\n break\n except Exception as e:\n logging.info(\" \".join(repr(e).splitlines()))\n logging.error(url)\n cur_retries += 1\n\n # random sample a header from headers\n header = headers[np.random.randint(0, len(headers))]\n\n time.sleep(3 + cur_retries * 2)\n\n\ndef download_image_from_url_val(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"val\", basename)\n\n download_file(url, filename)\n\n\ndef download_image_from_url_test(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"test\", basename)\n\n download_file(url, filename)\n\n\nif __name__ == \"__main__\":\n os.makedirs(\"tmp\", exist_ok=True)\n\n # storage dir","source_hash":"bad5e9e9e38e9001c05a04bad66598cc3c05037cc5d506ee633804f12d3cfff3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_nocaps.download_image_from_url_val","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_nocaps.download_image_from_url_val#L63-L67","kind":"function","name":"download_image_from_url_val","path":"lavis/datasets/download_scripts/download_nocaps.py","language":"python","start_line":63,"end_line":67,"context_start_line":43,"context_end_line":87,"code":" header = headers[0]\n\n while cur_retries < max_retries:\n try:\n r = requests.get(url, headers=header, timeout=10)\n with open(filename, \"wb\") as f:\n f.write(r.content)\n\n break\n except Exception as e:\n logging.info(\" \".join(repr(e).splitlines()))\n logging.error(url)\n cur_retries += 1\n\n # random sample a header from headers\n header = headers[np.random.randint(0, len(headers))]\n\n time.sleep(3 + cur_retries * 2)\n\n\ndef download_image_from_url_val(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"val\", basename)\n\n download_file(url, filename)\n\n\ndef download_image_from_url_test(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"test\", basename)\n\n download_file(url, filename)\n\n\nif __name__ == \"__main__\":\n os.makedirs(\"tmp\", exist_ok=True)\n\n # storage dir\n config_path = get_abs_path(\"configs/datasets/nocaps/defaults.yaml\")\n\n storage_dir = OmegaConf.load(config_path).datasets.nocaps.build_info.images.storage\n storage_dir = get_cache_path(storage_dir)\n # make sure the storage dir exists\n os.makedirs(storage_dir, exist_ok=True)\n print(\"Storage dir:\", storage_dir)","source_hash":"bad5e9e9e38e9001c05a04bad66598cc3c05037cc5d506ee633804f12d3cfff3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.download_nocaps.download_image_from_url_test","uri":"program://CREMA/function/lavis.datasets.download_scripts.download_nocaps.download_image_from_url_test#L70-L74","kind":"function","name":"download_image_from_url_test","path":"lavis/datasets/download_scripts/download_nocaps.py","language":"python","start_line":70,"end_line":74,"context_start_line":50,"context_end_line":94,"code":"\n break\n except Exception as e:\n logging.info(\" \".join(repr(e).splitlines()))\n logging.error(url)\n cur_retries += 1\n\n # random sample a header from headers\n header = headers[np.random.randint(0, len(headers))]\n\n time.sleep(3 + cur_retries * 2)\n\n\ndef download_image_from_url_val(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"val\", basename)\n\n download_file(url, filename)\n\n\ndef download_image_from_url_test(url):\n basename = os.path.basename(url)\n filename = os.path.join(storage_dir, \"test\", basename)\n\n download_file(url, filename)\n\n\nif __name__ == \"__main__\":\n os.makedirs(\"tmp\", exist_ok=True)\n\n # storage dir\n config_path = get_abs_path(\"configs/datasets/nocaps/defaults.yaml\")\n\n storage_dir = OmegaConf.load(config_path).datasets.nocaps.build_info.images.storage\n storage_dir = get_cache_path(storage_dir)\n # make sure the storage dir exists\n os.makedirs(storage_dir, exist_ok=True)\n print(\"Storage dir:\", storage_dir)\n\n # make sure the storage dir for val and test exists\n os.makedirs(os.path.join(storage_dir, \"val\"), exist_ok=True)\n os.makedirs(os.path.join(storage_dir, \"test\"), exist_ok=True)\n\n # download annotations\n val_url = \"https://nocaps.s3.amazonaws.com/nocaps_val_4500_captions.json\"","source_hash":"bad5e9e9e38e9001c05a04bad66598cc3c05037cc5d506ee633804f12d3cfff3","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m","uri":"program://CREMA/module/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m#L1-L229","kind":"module","name":"lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":1,"end_line":229,"context_start_line":1,"context_end_line":229,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport glob\nfrom pathlib import Path\nimport time\nfrom omegaconf import OmegaConf\nimport pandas as pd\nimport numpy as np\nimport requests\nimport zlib\nimport os\nimport io\nimport shelve\nfrom lavis.common.utils import get_abs_path, get_cache_path\nimport magic # pip install python-magic\nimport json\nfrom multiprocessing import Pool\nfrom tqdm import tqdm\nfrom PIL import Image\nfrom torchvision.transforms import functional as TF\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )\n print(\n int(len(df) / chunk_size),\n \"parts.\",\n chunk_size,\n \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"caption\", \"url\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_3m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_3m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))\n\nos.makedirs(storage_dir, exist_ok=True)\n\n# number of processes in the pool can be larger than cores\nnum_processes = 32\n# chunk_size is how many images per chunk per process - changing this resets progress when restarting.\nimages_per_part = 100\n\ndata_name = \"cc3m\"\ndf = open_tsv(\"/nas-ssd2/ziyang/data/dl-cc3m/Validation_GCC-1.1.0-Validation.tsv\", data_name)\ndf_multiprocess(\n df=df,\n processes=num_processes,\n chunk_size=images_per_part,\n func=download_image,\n dataset_name=data_name,\n)\ndf = df_from_shelve(\n chunk_size=images_per_part, func=download_image, dataset_name=data_name\n)\ndf.to_csv(\n \"downloaded_%s_report.tsv.gz\" % data_name,\n compression=\"gzip\",\n sep=\"\\t\",\n header=False,\n index=False,\n)\nprint(\"Saved.\")","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m._df_split_apply","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m._df_split_apply#L34-L37","kind":"function","name":"_df_split_apply","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":34,"end_line":37,"context_start_line":14,"context_end_line":57,"code":"import requests\nimport zlib\nimport os\nimport io\nimport shelve\nfrom lavis.common.utils import get_abs_path, get_cache_path\nimport magic # pip install python-magic\nimport json\nfrom multiprocessing import Pool\nfrom tqdm import tqdm\nfrom PIL import Image\nfrom torchvision.transforms import functional as TF\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.df_multiprocess","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.df_multiprocess#L40-L78","kind":"function","name":"df_multiprocess","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":40,"end_line":78,"context_start_line":20,"context_end_line":98,"code":"import magic # pip install python-magic\nimport json\nfrom multiprocessing import Pool\nfrom tqdm import tqdm\nfrom PIL import Image\nfrom torchvision.transforms import functional as TF\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )\n print(\n int(len(df) / chunk_size),\n \"parts.\",\n chunk_size,\n \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m._file_name","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m._file_name#L82-L93","kind":"function","name":"_file_name","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":82,"end_line":93,"context_start_line":62,"context_end_line":113,"code":" \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.check_mimetype","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.check_mimetype#L97-L101","kind":"function","name":"check_mimetype","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":97,"end_line":101,"context_start_line":77,"context_end_line":121,"code":" print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.check_download","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.check_download#L106-L121","kind":"function","name":"check_download","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":106,"end_line":121,"context_start_line":86,"context_end_line":141,"code":" # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.resize_img","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.resize_img#L124-L131","kind":"function","name":"resize_img","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":124,"end_line":131,"context_start_line":104,"context_end_line":151,"code":"# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.download_image","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.download_image#L134-L171","kind":"function","name":"download_image","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":134,"end_line":171,"context_start_line":114,"context_end_line":191,"code":" row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"caption\", \"url\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.open_tsv","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.open_tsv#L174-L181","kind":"function","name":"open_tsv","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":174,"end_line":181,"context_start_line":154,"context_end_line":201,"code":" return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"caption\", \"url\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_3m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_3m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.df_from_shelve","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc3m.df_from_shelve#L184-L191","kind":"function","name":"df_from_shelve","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc3m.py","language":"python","start_line":184,"end_line":191,"context_start_line":164,"context_end_line":211,"code":" row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"caption\", \"url\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_3m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_3m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))\n\nos.makedirs(storage_dir, exist_ok=True)\n\n# number of processes in the pool can be larger than cores\nnum_processes = 32\n# chunk_size is how many images per chunk per process - changing this resets progress when restarting.\nimages_per_part = 100\n\ndata_name = \"cc3m\"\ndf = open_tsv(\"/nas-ssd2/ziyang/data/dl-cc3m/Validation_GCC-1.1.0-Validation.tsv\", data_name)","source_hash":"0a749d01e80a13d9bd9db7d3deeb07adf938035ecd0518ec8e90c468121a4163","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m","uri":"program://CREMA/module/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m#L1-L232","kind":"module","name":"lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":1,"end_line":232,"context_start_line":1,"context_end_line":232,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport time\nfrom PIL import Image\nfrom lavis.common.utils import get_abs_path, get_cache_path\nfrom multiprocessing import Pool\nfrom omegaconf import OmegaConf\nfrom pathlib import Path\nfrom torchvision.transforms import functional as TF\nfrom tqdm import tqdm\nimport glob\nimport io\nimport json\nimport magic # pip install python-magic\nimport numpy as np\nimport os\nimport pandas as pd\nimport requests\nimport shelve\nimport zlib\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )\n print(\n int(len(df) / chunk_size),\n \"parts.\",\n chunk_size,\n \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"url\", \"caption\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_12m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_12m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))\n\nos.makedirs(storage_dir, exist_ok=True)\n\n# number of processes in the pool can be larger than cores\nnum_processes = 96\n# num_processes = 1\n# chunk_size is how many images per chunk per process - changing this resets progress when restarting.\nimages_per_part = 100\n\ndata_name = \"cc12m\"\n# os.makedirs(data_name, exist_ok=True)\n\ndf = open_tsv(\"cc12m.tsv\", data_name)\ndf_multiprocess(\n df=df,\n processes=num_processes,\n chunk_size=images_per_part,\n func=download_image,\n dataset_name=data_name,\n)\ndf = df_from_shelve(\n chunk_size=images_per_part, func=download_image, dataset_name=data_name\n)\ndf.to_csv(\n \"downloaded_%s_report.tsv.gz\" % data_name,\n compression=\"gzip\",\n sep=\"\\t\",\n header=False,\n index=False,\n)\nprint(\"Saved.\")","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m._df_split_apply","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m._df_split_apply#L34-L37","kind":"function","name":"_df_split_apply","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":34,"end_line":37,"context_start_line":14,"context_end_line":57,"code":"from torchvision.transforms import functional as TF\nfrom tqdm import tqdm\nimport glob\nimport io\nimport json\nimport magic # pip install python-magic\nimport numpy as np\nimport os\nimport pandas as pd\nimport requests\nimport shelve\nimport zlib\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.df_multiprocess","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.df_multiprocess#L40-L78","kind":"function","name":"df_multiprocess","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":40,"end_line":78,"context_start_line":20,"context_end_line":98,"code":"import numpy as np\nimport os\nimport pandas as pd\nimport requests\nimport shelve\nimport zlib\n\nheaders = {\n #'User-Agent': 'Mozilla/5.0 (X11; Linux x86_64) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/71.0.3578.98 Safari/537.36',\n \"User-Agent\": \"Googlebot-Image/1.0\", # Pretend to be googlebot\n \"X-Forwarded-For\": \"64.18.15.200\",\n}\n\n\ndef _df_split_apply(tup_arg):\n split_ind, subset, func = tup_arg\n r = subset.apply(func, axis=1)\n return (split_ind, r)\n\n\ndef df_multiprocess(df, processes, chunk_size, func, dataset_name):\n print(\"Generating parts...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n\n pbar = tqdm(total=len(df), position=0)\n # Resume:\n finished_chunks = set([int(k) for k in results.keys()])\n pbar.desc = \"Resuming\"\n for k in results.keys():\n pbar.update(len(results[str(k)][1]))\n\n pool_data = (\n (index, df[i : i + chunk_size], func)\n for index, i in enumerate(range(0, len(df), chunk_size))\n if index not in finished_chunks\n )\n print(\n int(len(df) / chunk_size),\n \"parts.\",\n chunk_size,\n \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m._file_name","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m._file_name#L82-L93","kind":"function","name":"_file_name","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":82,"end_line":93,"context_start_line":62,"context_end_line":113,"code":" \"per part.\",\n \"Using\",\n processes,\n \"processes\",\n )\n\n pbar.desc = \"Downloading\"\n with Pool(processes) as pool:\n for i, result in enumerate(\n pool.imap_unordered(_df_split_apply, pool_data, 2)\n ):\n results[str(result[0])] = result\n pbar.update(len(result[1]))\n pbar.close()\n\n print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.check_mimetype","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.check_mimetype#L97-L101","kind":"function","name":"check_mimetype","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":97,"end_line":101,"context_start_line":77,"context_end_line":121,"code":" print(\"Finished Downloading.\")\n return\n\n\n# Unique name based on url\ndef _file_name(row):\n name = (\n \"%s/%s_%s\"\n % (\n # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.check_download","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.check_download#L106-L121","kind":"function","name":"check_download","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":106,"end_line":121,"context_start_line":86,"context_end_line":141,"code":" # row[\"folder\"],\n storage_dir,\n row.name,\n (zlib.crc32(row[\"url\"].encode(\"utf-8\")) & 0xFFFFFFFF),\n )\n + \".jpg\"\n )\n return name\n\n\n# For checking mimetypes separately without download\ndef check_mimetype(row):\n if os.path.isfile(str(row[\"file\"])):\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n\n# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.resize_img","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.resize_img#L124-L131","kind":"function","name":"resize_img","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":124,"end_line":131,"context_start_line":104,"context_end_line":151,"code":"# Don't download image, just check with a HEAD request, can't resume.\n# Can use this instead of download_image to get HTTP status codes.\ndef check_download(row):\n fname = _file_name(row)\n try:\n # not all sites will support HEAD\n response = requests.head(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.download_image","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.download_image#L134-L171","kind":"function","name":"download_image","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":134,"end_line":171,"context_start_line":114,"context_end_line":191,"code":" row[\"headers\"] = dict(response.headers)\n except:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n if response.ok:\n row[\"file\"] = fname\n return row\n\n\ndef resize_img(req):\n image = Image.open(req).convert(\"RGB\")\n image = TF.resize(\n # image, size=(resize_size, resize_size)\n image,\n size=resize_size,\n ) # , interpolation=Image.LANCZOS)\n return image\n\n\ndef download_image(row):\n fname = _file_name(row)\n # Skip Already downloaded, retry others later\n if os.path.isfile(fname):\n row[\"status\"] = 200\n row[\"file\"] = fname\n row[\"mimetype\"] = magic.from_file(row[\"file\"], mime=True)\n row[\"size\"] = os.stat(row[\"file\"]).st_size\n return row\n\n try:\n # use smaller timeout to skip errors, but can result in failed downloads\n response = requests.get(\n row[\"url\"], stream=False, timeout=5, allow_redirects=True, headers=headers\n )\n row[\"status\"] = response.status_code\n # row['headers'] = dict(response.headers)\n except Exception as e:\n # log errors later, set error as 408 timeout\n row[\"status\"] = 408\n return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"url\", \"caption\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.open_tsv","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.open_tsv#L174-L181","kind":"function","name":"open_tsv","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":174,"end_line":181,"context_start_line":154,"context_end_line":201,"code":" return row\n\n if response.ok:\n try:\n # some sites respond with gzip transport encoding\n response.raw.decode_content = True\n img = resize_img(io.BytesIO(response.content))\n img.save(fname)\n\n row[\"mimetype\"] = magic.from_file(fname, mime=True)\n row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"url\", \"caption\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_12m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_12m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.df_from_shelve","uri":"program://CREMA/function/lavis.datasets.download_scripts.DownloadConceptualCaptions.download_data_cc12m.df_from_shelve#L184-L191","kind":"function","name":"df_from_shelve","path":"lavis/datasets/download_scripts/DownloadConceptualCaptions/download_data_cc12m.py","language":"python","start_line":184,"end_line":191,"context_start_line":164,"context_end_line":211,"code":" row[\"size\"] = os.stat(fname).st_size\n\n except Exception as e:\n # # This is if it times out during a download or decode\n row[\"status\"] = 408\n\n row[\"file\"] = fname\n return row\n\n\ndef open_tsv(fname, folder):\n print(\"Opening %s Data File...\" % fname)\n df = pd.read_csv(\n fname, sep=\"\\t\", names=[\"url\", \"caption\"]\n ) # , usecols=range(1, 2))\n df[\"folder\"] = folder\n print(\"Processing\", len(df), \" Images:\")\n return df\n\n\ndef df_from_shelve(chunk_size, func, dataset_name):\n print(\"Generating Dataframe from results...\")\n with shelve.open(\n \"%s_%s_%s_results.tmp\" % (dataset_name, func.__name__, chunk_size)\n ) as results:\n keylist = sorted([int(k) for k in results.keys()])\n df = pd.concat([results[str(k)][1] for k in keylist], sort=True)\n return df\n\n\nresize_size = 384\n\nconfig_path = get_abs_path(\"configs/datasets/conceptual_caption/defaults_12m.yaml\")\n\nstorage_dir = OmegaConf.load(\n config_path\n).datasets.conceptual_caption_12m.build_info.images.storage\nstorage_dir = Path(get_cache_path(storage_dir))\n\nos.makedirs(storage_dir, exist_ok=True)\n\n# number of processes in the pool can be larger than cores\nnum_processes = 96\n# num_processes = 1\n# chunk_size is how many images per chunk per process - changing this resets progress when restarting.\nimages_per_part = 100\n\ndata_name = \"cc12m\"","source_hash":"ea65c6b7a204e1f3db6fd488781e344c8eb858a61c3fcd28818eb935d8adf192","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.laion_dataset","uri":"program://CREMA/module/lavis.datasets.datasets.laion_dataset#L1-L62","kind":"module","name":"lavis.datasets.datasets.laion_dataset","path":"lavis/datasets/datasets/laion_dataset.py","language":"python","start_line":1,"end_line":62,"context_start_line":1,"context_end_line":62,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport webdataset as wds\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass LaionDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, location):\n super().__init__(vis_processor=vis_processor, text_processor=text_processor)\n\n self.inner_dataset = wds.DataPipeline(\n wds.ResampledShards(location),\n wds.tarfile_to_samples(handler=wds.warn_and_continue),\n wds.shuffle(1000, handler=wds.warn_and_continue),\n wds.decode(\"pilrgb\", handler=wds.warn_and_continue),\n wds.to_tuple(\"jpg\", \"json\", handler=wds.warn_and_continue),\n wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),\n wds.map(self.to_dict, handler=wds.warn_and_continue),\n )\n\n def to_dict(self, sample):\n return {\n \"image\": sample[0],\n \"text_input\": self.text_processor(sample[1][\"caption\"]),\n }\n\n\nif __name__ == \"__main__\":\n from torchvision import transforms\n\n def to_image_text_pair(sample):\n return sample[0], sample[1][\"caption\"]\n\n normalize = transforms.Normalize(\n (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)\n )\n\n transform_train = transforms.Compose(\n [\n transforms.RandomResizedCrop(256, scale=(0.2, 1.0)),\n transforms.RandomHorizontalFlip(),\n transforms.ToTensor(),\n normalize,\n ]\n )\n\n dataset = LaionDataset(\n vis_processor=transform_train,\n text_processor=lambda x: x,\n location=\"/export/laion/laion2B-multi/part-00000/{00000..01743}.tar\",\n )\n\n import torch\n\n loader = torch.utils.data.DataLoader(dataset.inner_dataset, batch_size=2)\n\n print(next(iter(loader))[\"text_input\"])","source_hash":"1de8547a223c98889ee22417b5bd2dfe70446793ad2011ac06c3b7704bb32629","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.laion_dataset.LaionDataset","uri":"program://CREMA/class/lavis.datasets.datasets.laion_dataset.LaionDataset#L12-L30","kind":"class","name":"LaionDataset","path":"lavis/datasets/datasets/laion_dataset.py","language":"python","start_line":12,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport webdataset as wds\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass LaionDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, location):\n super().__init__(vis_processor=vis_processor, text_processor=text_processor)\n\n self.inner_dataset = wds.DataPipeline(\n wds.ResampledShards(location),\n wds.tarfile_to_samples(handler=wds.warn_and_continue),\n wds.shuffle(1000, handler=wds.warn_and_continue),\n wds.decode(\"pilrgb\", handler=wds.warn_and_continue),\n wds.to_tuple(\"jpg\", \"json\", handler=wds.warn_and_continue),\n wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),\n wds.map(self.to_dict, handler=wds.warn_and_continue),\n )\n\n def to_dict(self, sample):\n return {\n \"image\": sample[0],\n \"text_input\": self.text_processor(sample[1][\"caption\"]),\n }\n\n\nif __name__ == \"__main__\":\n from torchvision import transforms\n\n def to_image_text_pair(sample):\n return sample[0], sample[1][\"caption\"]\n\n normalize = transforms.Normalize(\n (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)\n )\n\n transform_train = transforms.Compose(\n [\n transforms.RandomResizedCrop(256, scale=(0.2, 1.0)),\n transforms.RandomHorizontalFlip(),\n transforms.ToTensor(),\n normalize,\n ]\n )","source_hash":"1de8547a223c98889ee22417b5bd2dfe70446793ad2011ac06c3b7704bb32629","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.laion_dataset.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.laion_dataset.__init__#L13-L24","kind":"function","name":"__init__","path":"lavis/datasets/datasets/laion_dataset.py","language":"python","start_line":13,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport webdataset as wds\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass LaionDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, location):\n super().__init__(vis_processor=vis_processor, text_processor=text_processor)\n\n self.inner_dataset = wds.DataPipeline(\n wds.ResampledShards(location),\n wds.tarfile_to_samples(handler=wds.warn_and_continue),\n wds.shuffle(1000, handler=wds.warn_and_continue),\n wds.decode(\"pilrgb\", handler=wds.warn_and_continue),\n wds.to_tuple(\"jpg\", \"json\", handler=wds.warn_and_continue),\n wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),\n wds.map(self.to_dict, handler=wds.warn_and_continue),\n )\n\n def to_dict(self, sample):\n return {\n \"image\": sample[0],\n \"text_input\": self.text_processor(sample[1][\"caption\"]),\n }\n\n\nif __name__ == \"__main__\":\n from torchvision import transforms\n\n def to_image_text_pair(sample):\n return sample[0], sample[1][\"caption\"]\n\n normalize = transforms.Normalize(\n (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)\n )\n\n transform_train = transforms.Compose(\n [","source_hash":"1de8547a223c98889ee22417b5bd2dfe70446793ad2011ac06c3b7704bb32629","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.laion_dataset.to_dict","uri":"program://CREMA/function/lavis.datasets.datasets.laion_dataset.to_dict#L26-L30","kind":"function","name":"to_dict","path":"lavis/datasets/datasets/laion_dataset.py","language":"python","start_line":26,"end_line":30,"context_start_line":6,"context_end_line":50,"code":"\"\"\"\n\nimport webdataset as wds\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass LaionDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, location):\n super().__init__(vis_processor=vis_processor, text_processor=text_processor)\n\n self.inner_dataset = wds.DataPipeline(\n wds.ResampledShards(location),\n wds.tarfile_to_samples(handler=wds.warn_and_continue),\n wds.shuffle(1000, handler=wds.warn_and_continue),\n wds.decode(\"pilrgb\", handler=wds.warn_and_continue),\n wds.to_tuple(\"jpg\", \"json\", handler=wds.warn_and_continue),\n wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),\n wds.map(self.to_dict, handler=wds.warn_and_continue),\n )\n\n def to_dict(self, sample):\n return {\n \"image\": sample[0],\n \"text_input\": self.text_processor(sample[1][\"caption\"]),\n }\n\n\nif __name__ == \"__main__\":\n from torchvision import transforms\n\n def to_image_text_pair(sample):\n return sample[0], sample[1][\"caption\"]\n\n normalize = transforms.Normalize(\n (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)\n )\n\n transform_train = transforms.Compose(\n [\n transforms.RandomResizedCrop(256, scale=(0.2, 1.0)),\n transforms.RandomHorizontalFlip(),\n transforms.ToTensor(),\n normalize,\n ]\n )","source_hash":"1de8547a223c98889ee22417b5bd2dfe70446793ad2011ac06c3b7704bb32629","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.laion_dataset.to_image_text_pair","uri":"program://CREMA/function/lavis.datasets.datasets.laion_dataset.to_image_text_pair#L36-L37","kind":"function","name":"to_image_text_pair","path":"lavis/datasets/datasets/laion_dataset.py","language":"python","start_line":36,"end_line":37,"context_start_line":16,"context_end_line":57,"code":" self.inner_dataset = wds.DataPipeline(\n wds.ResampledShards(location),\n wds.tarfile_to_samples(handler=wds.warn_and_continue),\n wds.shuffle(1000, handler=wds.warn_and_continue),\n wds.decode(\"pilrgb\", handler=wds.warn_and_continue),\n wds.to_tuple(\"jpg\", \"json\", handler=wds.warn_and_continue),\n wds.map_tuple(self.vis_processor, handler=wds.warn_and_continue),\n wds.map(self.to_dict, handler=wds.warn_and_continue),\n )\n\n def to_dict(self, sample):\n return {\n \"image\": sample[0],\n \"text_input\": self.text_processor(sample[1][\"caption\"]),\n }\n\n\nif __name__ == \"__main__\":\n from torchvision import transforms\n\n def to_image_text_pair(sample):\n return sample[0], sample[1][\"caption\"]\n\n normalize = transforms.Normalize(\n (0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711)\n )\n\n transform_train = transforms.Compose(\n [\n transforms.RandomResizedCrop(256, scale=(0.2, 1.0)),\n transforms.RandomHorizontalFlip(),\n transforms.ToTensor(),\n normalize,\n ]\n )\n\n dataset = LaionDataset(\n vis_processor=transform_train,\n text_processor=lambda x: x,\n location=\"/export/laion/laion2B-multi/part-00000/{00000..01743}.tar\",\n )\n","source_hash":"1de8547a223c98889ee22417b5bd2dfe70446793ad2011ac06c3b7704bb32629","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.retrieval_datasets#L1-L162","kind":"module","name":"lavis.datasets.datasets.retrieval_datasets","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":1,"end_line":162,"context_start_line":1,"context_end_line":162,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n visual_key = \"image\" if \"image\" in ann else \"video\"\n\n return OrderedDict(\n {\n \"file\": ann[visual_key],\n \"caption\": ann[\"caption\"],\n visual_key: sample[visual_key],\n }\n )\n\n\nclass RetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass RetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"image\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n\n image_path = os.path.join(self.vis_root, self.annotation[index][\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\"image\": image, \"index\": index}\n\n\nclass VideoRetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"video\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n\n video = self.vis_processor(vpath)\n caption = self.text_processor(ann[\"caption\"])\n\n # return image, caption, self.img_ids[ann['image_id']]\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"video\"]],\n }\n\n\nclass VideoRetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"video\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n video = self.vis_processor(vpath)\n\n return {\"video\": video, \"index\": index}","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.retrieval_datasets.__DisplMixin#L15-L26","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":15,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n visual_key = \"image\" if \"image\" in ann else \"video\"\n\n return OrderedDict(\n {\n \"file\": ann[visual_key],\n \"caption\": ann[\"caption\"],\n visual_key: sample[visual_key],\n }\n )\n\n\nclass RetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.RetrievalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.retrieval_datasets.RetrievalDataset#L29-L60","kind":"class","name":"RetrievalDataset","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":29,"end_line":60,"context_start_line":9,"context_end_line":80,"code":"from collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n visual_key = \"image\" if \"image\" in ann else \"video\"\n\n return OrderedDict(\n {\n \"file\": ann[visual_key],\n \"caption\": ann[\"caption\"],\n visual_key: sample[visual_key],\n }\n )\n\n\nclass RetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass RetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"image\"])","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.RetrievalEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.retrieval_datasets.RetrievalEvalDataset#L63-L95","kind":"class","name":"RetrievalEvalDataset","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":63,"end_line":95,"context_start_line":43,"context_end_line":115,"code":" n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass RetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"image\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n\n image_path = os.path.join(self.vis_root, self.annotation[index][\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\"image\": image, \"index\": index}\n\n\nclass VideoRetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"video\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.VideoRetrievalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.retrieval_datasets.VideoRetrievalDataset#L98-L128","kind":"class","name":"VideoRetrievalDataset","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":98,"end_line":128,"context_start_line":78,"context_end_line":148,"code":" txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"image\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n\n image_path = os.path.join(self.vis_root, self.annotation[index][\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\"image\": image, \"index\": index}\n\n\nclass VideoRetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"video\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n\n video = self.vis_processor(vpath)\n caption = self.text_processor(ann[\"caption\"])\n\n # return image, caption, self.img_ids[ann['image_id']]\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"video\"]],\n }\n\n\nclass VideoRetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"video\"])","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.VideoRetrievalEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.retrieval_datasets.VideoRetrievalEvalDataset#L131-L162","kind":"class","name":"VideoRetrievalEvalDataset","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":131,"end_line":162,"context_start_line":111,"context_end_line":162,"code":" self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n\n video = self.vis_processor(vpath)\n caption = self.text_processor(ann[\"caption\"])\n\n # return image, caption, self.img_ids[ann['image_id']]\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"video\"]],\n }\n\n\nclass VideoRetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"video\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n video = self.vis_processor(vpath)\n\n return {\"video\": video, \"index\": index}","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.retrieval_datasets.displ_item#L16-L26","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":16,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n visual_key = \"image\" if \"image\" in ann else \"video\"\n\n return OrderedDict(\n {\n \"file\": ann[visual_key],\n \"caption\": ann[\"caption\"],\n visual_key: sample[visual_key],\n }\n )\n\n\nclass RetrievalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.retrieval_datasets.__init__#L132-L154","kind":"function","name":"__init__","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":132,"end_line":154,"context_start_line":112,"context_end_line":162,"code":" n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n\n video = self.vis_processor(vpath)\n caption = self.text_processor(ann[\"caption\"])\n\n # return image, caption, self.img_ids[ann['image_id']]\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"video\"]],\n }\n\n\nclass VideoRetrievalEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of videos.\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"video\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n video = self.vis_processor(vpath)\n\n return {\"video\": video, \"index\": index}","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.retrieval_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.retrieval_datasets.__getitem__#L156-L162","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/retrieval_datasets.py","language":"python","start_line":156,"end_line":162,"context_start_line":136,"context_end_line":162,"code":" split (string): val or test\n \"\"\"\n\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.text = []\n self.image = []\n self.txt2img = {}\n self.img2txt = {}\n\n txt_id = 0\n for img_id, ann in enumerate(self.annotation):\n self.image.append(ann[\"video\"])\n self.img2txt[img_id] = []\n for i, caption in enumerate(ann[\"caption\"]):\n self.text.append(self.text_processor(caption))\n self.img2txt[img_id].append(txt_id)\n self.txt2img[txt_id] = img_id\n txt_id += 1\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n vpath = os.path.join(self.vis_root, ann[\"video\"])\n video = self.vis_processor(vpath)\n\n return {\"video\": video, \"index\": index}","source_hash":"b1c4d861abc5f359a265636a32979433e92c764b5072ea96f409f66fd376e87f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.image_text_pair_datasets#L1-L47","kind":"module","name":"lavis.datasets.datasets.image_text_pair_datasets","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":1,"end_line":47,"context_start_line":1,"context_end_line":47,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\"image\": image, \"text_input\": caption}","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.image_text_pair_datasets.__DisplMixin#L15-L25","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":15,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets.ImageTextPairDataset","uri":"program://CREMA/class/lavis.datasets.datasets.image_text_pair_datasets.ImageTextPairDataset#L28-L47","kind":"class","name":"ImageTextPairDataset","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":28,"end_line":47,"context_start_line":8,"context_end_line":47,"code":"import os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\"image\": image, \"text_input\": caption}","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.image_text_pair_datasets.displ_item#L16-L25","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":16,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.image_text_pair_datasets.__init__#L29-L34","kind":"function","name":"__init__","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":29,"end_line":34,"context_start_line":9,"context_end_line":47,"code":"from collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\"image\": image, \"text_input\": caption}","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.image_text_pair_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.image_text_pair_datasets.__getitem__#L36-L47","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/image_text_pair_datasets.py","language":"python","start_line":36,"end_line":47,"context_start_line":16,"context_end_line":47,"code":" def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass ImageTextPairDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\"image\": image, \"text_input\": caption}","source_hash":"cc7aebfe7ad9810a0a286cc5fb1f860a8d59bc128fea7a8c3cf6e97dca91a718","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.music_avqa_datasets#L1-L139","kind":"module","name":"lavis.datasets.datasets.music_avqa_datasets","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":1,"end_line":139,"context_start_line":1,"context_end_line":139,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport copy\nimport os\nimport random\nimport json\nimport ast\nimport re\nfrom PIL import Image\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nclass MusicAVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n for modality in self.modalities:\n if 'image' in modality:\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:\n ann['rgb'] = self.vis_processor(Image.open(ann[f\"images_path\"]))\n else:\n if modality == 'video':\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n ann['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices, clip = None, None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n ann['rgb'] = rgb\n \n if modality == 'flow':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n flow, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n ann['flow'] = flow\n \n if modality == 'audio':\n if 'id' in ann: # aduio pt data\n ann[f\"{modality}_path\"] = self.get_pt_audio_path(ann)\n ann[modality] = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"]).to(torch.float32)\n\n if 'id' not in ann:\n ann[\"sample_id\"] = ann[\"video_id\"]\n if len(ann['templ_values']) != 0:\n question = ann['question_content']\n templ_values = ast.literal_eval(ann['templ_values'])\n matches = re.findall(r'<(.*?)>', question)\n for k, v in zip(matches, templ_values):\n question = question.replace('<'+k+'>', v)\n else:\n question = ann['question_content']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames and audio information, answer the question using a single word or phase.'\n ann[\"question_id\"] = ann['question_id']\n answers = ann['anser']\n if '_' in answers:\n answers = answers.replace('_', ' ')\n ann['answers'] = answers #ann['anser']\n\n else:\n ann[\"sample_id\"] = ann[\"id\"]\n ann['answers'] = ann['answer']\n ann[\"question_id\"] = ann['id']\n question = ann['question']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the audio information, answer the question using a single word or phase.'\n \n return ann\n \nclass MusicAVQAInstructDataset(MusicAVQADataset):\n def __getitem__(self, index):\n data = super().__getitem__(index)\n if data != None:\n data['answer'] = data[\"answers\"] # needed to use gqa task\n data['qa_output'] = data[\"answers\"]\n \n return data","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.MusicAVQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.music_avqa_datasets.MusicAVQADataset#L18-L130","kind":"class","name":"MusicAVQADataset","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":18,"end_line":130,"context_start_line":1,"context_end_line":139,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport copy\nimport os\nimport random\nimport json\nimport ast\nimport re\nfrom PIL import Image\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nclass MusicAVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n for modality in self.modalities:\n if 'image' in modality:\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:\n ann['rgb'] = self.vis_processor(Image.open(ann[f\"images_path\"]))\n else:\n if modality == 'video':\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n ann['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices, clip = None, None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n ann['rgb'] = rgb\n \n if modality == 'flow':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n flow, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n ann['flow'] = flow\n \n if modality == 'audio':\n if 'id' in ann: # aduio pt data\n ann[f\"{modality}_path\"] = self.get_pt_audio_path(ann)\n ann[modality] = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"]).to(torch.float32)\n\n if 'id' not in ann:\n ann[\"sample_id\"] = ann[\"video_id\"]\n if len(ann['templ_values']) != 0:\n question = ann['question_content']\n templ_values = ast.literal_eval(ann['templ_values'])\n matches = re.findall(r'<(.*?)>', question)\n for k, v in zip(matches, templ_values):\n question = question.replace('<'+k+'>', v)\n else:\n question = ann['question_content']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames and audio information, answer the question using a single word or phase.'\n ann[\"question_id\"] = ann['question_id']\n answers = ann['anser']\n if '_' in answers:\n answers = answers.replace('_', ' ')\n ann['answers'] = answers #ann['anser']\n\n else:\n ann[\"sample_id\"] = ann[\"id\"]\n ann['answers'] = ann['answer']\n ann[\"question_id\"] = ann['id']\n question = ann['question']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the audio information, answer the question using a single word or phase.'\n \n return ann\n \nclass MusicAVQAInstructDataset(MusicAVQADataset):\n def __getitem__(self, index):\n data = super().__getitem__(index)\n if data != None:\n data['answer'] = data[\"answers\"] # needed to use gqa task\n data['qa_output'] = data[\"answers\"]\n \n return data","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.MusicAVQAInstructDataset","uri":"program://CREMA/class/lavis.datasets.datasets.music_avqa_datasets.MusicAVQAInstructDataset#L132-L139","kind":"class","name":"MusicAVQAInstructDataset","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":132,"end_line":139,"context_start_line":112,"context_end_line":139,"code":" for k, v in zip(matches, templ_values):\n question = question.replace('<'+k+'>', v)\n else:\n question = ann['question_content']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames and audio information, answer the question using a single word or phase.'\n ann[\"question_id\"] = ann['question_id']\n answers = ann['anser']\n if '_' in answers:\n answers = answers.replace('_', ' ')\n ann['answers'] = answers #ann['anser']\n\n else:\n ann[\"sample_id\"] = ann[\"id\"]\n ann['answers'] = ann['answer']\n ann[\"question_id\"] = ann['id']\n question = ann['question']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the audio information, answer the question using a single word or phase.'\n \n return ann\n \nclass MusicAVQAInstructDataset(MusicAVQADataset):\n def __getitem__(self, index):\n data = super().__getitem__(index)\n if data != None:\n data['answer'] = data[\"answers\"] # needed to use gqa task\n data['qa_output'] = data[\"answers\"]\n \n return data","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.__init__#L19-L38","kind":"function","name":"__init__","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":19,"end_line":38,"context_start_line":1,"context_end_line":58,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nimport copy\nimport os\nimport random\nimport json\nimport ast\nimport re\nfrom PIL import Image\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nclass MusicAVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n for modality in self.modalities:\n if 'image' in modality:\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_existing_audio_annotations","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_existing_audio_annotations#L40-L41","kind":"function","name":"get_existing_audio_annotations","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":40,"end_line":41,"context_start_line":20,"context_end_line":61,"code":" super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n for modality in self.modalities:\n if 'image' in modality:\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n ","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_existing_frame_annotations","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_existing_frame_annotations#L43-L44","kind":"function","name":"get_existing_frame_annotations","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":43,"end_line":44,"context_start_line":23,"context_end_line":64,"code":"\n for modality in self.modalities:\n if 'image' in modality:\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n ","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_existing_flow_annotations","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_existing_flow_annotations#L46-L47","kind":"function","name":"get_existing_flow_annotations","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":46,"end_line":47,"context_start_line":26,"context_end_line":67,"code":" setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n continue\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n self.sample_ids = set.intersection(*[set(getattr(self, f\"existing_{modality}_annotation\")) for modality in self.modalities])\n\n try:\n self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n ","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_existing_video_annotations","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_existing_video_annotations#L55-L56","kind":"function","name":"get_existing_video_annotations","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":55,"end_line":56,"context_start_line":35,"context_end_line":76,"code":" self.annotation = [ann for ann in self.annotation if ann['video_id'] in self.sample_ids]\n except:\n self.sample_ids = set.intersection(*[set(self.get_existing_audio_pt_annotations()) for modality in self.modalities])\n self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_audio_path","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_audio_path#L58-L60","kind":"function","name":"get_audio_path","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":58,"end_line":60,"context_start_line":38,"context_end_line":80,"code":" self.annotation = [ann for ann in self.annotation if ann['id'] in self.sample_ids]\n \n def get_existing_audio_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.audio_root)]\n \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_video_path","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_video_path#L62-L63","kind":"function","name":"get_video_path","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":62,"end_line":63,"context_start_line":42,"context_end_line":83,"code":" \n def get_existing_frame_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.frame_root)]\n \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:\n ann['rgb'] = self.vis_processor(Image.open(ann[f\"images_path\"]))\n else:\n if modality == 'video':","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_flow_path","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_flow_path#L65-L66","kind":"function","name":"get_flow_path","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":65,"end_line":66,"context_start_line":45,"context_end_line":86,"code":" \n def get_existing_flow_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.flow_root)]\n\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:\n ann['rgb'] = self.vis_processor(Image.open(ann[f\"images_path\"]))\n else:\n if modality == 'video':\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n ann['rgb'] = rgb.to(torch.float32)\n ","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.get_frame_path","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.get_frame_path#L68-L69","kind":"function","name":"get_frame_path","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":68,"end_line":69,"context_start_line":48,"context_end_line":89,"code":"\n # def get_existing_audio_pt_annotations(self):\n # return [f.split('.')[0] for f in os.listdir('/nas-ssd2/shoubin/datasets/audios')]\n\n # def get_pt_audio_path(self, ann):\n # return os.path.join('/nas-ssd2/shoubin/datasets/audios', f'{str(ann[\"id\"])}.wav')\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_audio_path(self, ann):\n # return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.flac')\n return os.path.join(self.audio_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"video_id\"]}.mp4')\n \n def get_flow_path(self, ann):\n return os.path.join(self.flow_root, f'{ann[\"video_id\"]}/')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"video_id\"]}/')\n\n def __getitem__(self, index):\n ann = copy.deepcopy(self.annotation[index])\n\n for modality in self.modalities:\n if 'id' not in ann:\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n if type(ann[f\"{modality}_path\"]) == list:\n ann[f\"{modality}_path\"] = random.choice(ann[f\"{modality}_path\"])\n\n if 'image' in modality:\n ann['rgb'] = self.vis_processor(Image.open(ann[f\"images_path\"]))\n else:\n if modality == 'video':\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n ann['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices, clip = None, None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.music_avqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.music_avqa_datasets.__getitem__#L133-L139","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/music_avqa_datasets.py","language":"python","start_line":133,"end_line":139,"context_start_line":113,"context_end_line":139,"code":" question = question.replace('<'+k+'>', v)\n else:\n question = ann['question_content']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames and audio information, answer the question using a single word or phase.'\n ann[\"question_id\"] = ann['question_id']\n answers = ann['anser']\n if '_' in answers:\n answers = answers.replace('_', ' ')\n ann['answers'] = answers #ann['anser']\n\n else:\n ann[\"sample_id\"] = ann[\"id\"]\n ann['answers'] = ann['answer']\n ann[\"question_id\"] = ann['id']\n question = ann['question']\n ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the audio information, answer the question using a single word or phase.'\n \n return ann\n \nclass MusicAVQAInstructDataset(MusicAVQADataset):\n def __getitem__(self, index):\n data = super().__getitem__(index)\n if data != None:\n data['answer'] = data[\"answers\"] # needed to use gqa task\n data['qa_output'] = data[\"answers\"]\n \n return data","source_hash":"f66975386bada248492326bbda9ceb4dedcb12773a887c306d81af4f4fd5cd3e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.dialogue_datasets#L1-L141","kind":"module","name":"lavis.datasets.datasets.dialogue_datasets","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":1,"end_line":141,"context_start_line":1,"context_end_line":141,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nimport json\nimport copy\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"dialogue\": ann[\"dialogue\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass DialogueDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = []\n for turn in all_turns:\n dialog_instance = copy.deepcopy(dialog)\n question = turn[\"question\"]\n answer = turn[\"answer\"]\n\n dialog_instance[\"dialog\"] = copy.deepcopy(dialogue_context)\n dialog_instance[\"question\"] = question\n dialog_instance[\"answer\"] = answer\n self.annotation.append(dialog_instance)\n dialogue_context.append(turn)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass DialogueEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = all_turns[:-1]\n last_turn = all_turns[-1]\n\n question = last_turn[\"question\"]\n answer = last_turn[\"answer\"]\n\n dialog[\"dialog\"] = dialogue_context\n dialog[\"question\"] = question\n dialog[\"answer\"] = answer\n\n self.annotation.append(dialog)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.dialogue_datasets.__DisplMixin#L19-L29","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":19,"end_line":29,"context_start_line":1,"context_end_line":49,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nimport json\nimport copy\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"dialogue\": ann[\"dialogue\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass DialogueDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = []\n for turn in all_turns:\n dialog_instance = copy.deepcopy(dialog)\n question = turn[\"question\"]","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.DialogueDataset","uri":"program://CREMA/class/lavis.datasets.datasets.dialogue_datasets.DialogueDataset#L32-L85","kind":"class","name":"DialogueDataset","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":32,"end_line":85,"context_start_line":12,"context_end_line":105,"code":"\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nimport json\nimport copy\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"dialogue\": ann[\"dialogue\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass DialogueDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = []\n for turn in all_turns:\n dialog_instance = copy.deepcopy(dialog)\n question = turn[\"question\"]\n answer = turn[\"answer\"]\n\n dialog_instance[\"dialog\"] = copy.deepcopy(dialogue_context)\n dialog_instance[\"question\"] = question\n dialog_instance[\"answer\"] = answer\n self.annotation.append(dialog_instance)\n dialogue_context.append(turn)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass DialogueEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = all_turns[:-1]\n last_turn = all_turns[-1]\n","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.DialogueEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.dialogue_datasets.DialogueEvalDataset#L88-L141","kind":"class","name":"DialogueEvalDataset","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":88,"end_line":141,"context_start_line":68,"context_end_line":141,"code":" self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass DialogueEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = all_turns[:-1]\n last_turn = all_turns[-1]\n\n question = last_turn[\"question\"]\n answer = last_turn[\"answer\"]\n\n dialog[\"dialog\"] = dialogue_context\n dialog[\"question\"] = question\n dialog[\"answer\"] = answer\n\n self.annotation.append(dialog)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.dialogue_datasets.displ_item#L20-L29","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":20,"end_line":29,"context_start_line":1,"context_end_line":49,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nimport json\nimport copy\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"dialogue\": ann[\"dialogue\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass DialogueDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = []\n for turn in all_turns:\n dialog_instance = copy.deepcopy(dialog)\n question = turn[\"question\"]","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.dialogue_datasets.__init__#L89-L126","kind":"function","name":"__init__","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":89,"end_line":126,"context_start_line":69,"context_end_line":141,"code":" n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass DialogueEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = []\n for ann_path in ann_paths:\n dialogs = json.load(open(ann_path, \"r\"))[\"dialogs\"]\n for dialog in dialogs:\n all_turns = dialog[\"dialog\"]\n dialogue_context = all_turns[:-1]\n last_turn = all_turns[-1]\n\n question = last_turn[\"question\"]\n answer = last_turn[\"answer\"]\n\n dialog[\"dialog\"] = dialogue_context\n dialog[\"question\"] = question\n dialog[\"answer\"] = answer\n\n self.annotation.append(dialog)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dialogue_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.dialogue_datasets.__getitem__#L128-L141","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/dialogue_datasets.py","language":"python","start_line":128,"end_line":141,"context_start_line":108,"context_end_line":141,"code":"\n dialog[\"dialog\"] = dialogue_context\n dialog[\"question\"] = question\n dialog[\"answer\"] = answer\n\n self.annotation.append(dialog)\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"cf2d33690e47af74e758edb7dd1a4078533f0208ad00fc8974ea5b4d261f6261","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.nlvr_datasets#L1-L94","kind":"module","name":"lavis.datasets.datasets.nlvr_datasets","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":1,"end_line":94,"context_start_line":1,"context_end_line":94,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport random\n\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:\n if random.random() < 0.5:\n sentence = sentence.replace(\"left\", \"[TEMP_TOKEN]\")\n sentence = sentence.replace(\"right\", \"left\")\n sentence = sentence.replace(\"[TEMP_TOKEN]\", \"right\")\n\n image0, image1 = image1, image0\n\n samples[\"text_input\"] = sentence\n samples[\"image0\"] = image0\n samples[\"image1\"] = image1\n\n return samples\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image0_path = os.path.join(self.vis_root, ann[\"images\"][0])\n image0 = Image.open(image0_path).convert(\"RGB\")\n image0 = self.vis_processor(image0)\n\n image1_path = os.path.join(self.vis_root, ann[\"images\"][1])\n image1 = Image.open(image1_path).convert(\"RGB\")\n image1 = self.vis_processor(image1)\n\n sentence = self.text_processor(ann[\"sentence\"])\n label = self.class_labels[ann[\"label\"]]\n\n return self._flip(\n {\n \"image0\": image0,\n \"image1\": image1,\n \"text_input\": sentence,\n \"label\": label,\n # \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n )\n\n\nclass NLVREvalDataset(NLVRDataset):\n @staticmethod\n def _flip(samples):\n return samples","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.nlvr_datasets.__DisplMixin#L19-L31","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":19,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport random\n\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.NLVRDataset","uri":"program://CREMA/class/lavis.datasets.datasets.nlvr_datasets.NLVRDataset#L34-L88","kind":"class","name":"NLVRDataset","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":34,"end_line":88,"context_start_line":14,"context_end_line":94,"code":" MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:\n if random.random() < 0.5:\n sentence = sentence.replace(\"left\", \"[TEMP_TOKEN]\")\n sentence = sentence.replace(\"right\", \"left\")\n sentence = sentence.replace(\"[TEMP_TOKEN]\", \"right\")\n\n image0, image1 = image1, image0\n\n samples[\"text_input\"] = sentence\n samples[\"image0\"] = image0\n samples[\"image1\"] = image1\n\n return samples\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image0_path = os.path.join(self.vis_root, ann[\"images\"][0])\n image0 = Image.open(image0_path).convert(\"RGB\")\n image0 = self.vis_processor(image0)\n\n image1_path = os.path.join(self.vis_root, ann[\"images\"][1])\n image1 = Image.open(image1_path).convert(\"RGB\")\n image1 = self.vis_processor(image1)\n\n sentence = self.text_processor(ann[\"sentence\"])\n label = self.class_labels[ann[\"label\"]]\n\n return self._flip(\n {\n \"image0\": image0,\n \"image1\": image1,\n \"text_input\": sentence,\n \"label\": label,\n # \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n )\n\n\nclass NLVREvalDataset(NLVRDataset):\n @staticmethod\n def _flip(samples):\n return samples","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.NLVREvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.nlvr_datasets.NLVREvalDataset#L91-L94","kind":"class","name":"NLVREvalDataset","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":91,"end_line":94,"context_start_line":71,"context_end_line":94,"code":"\n image1_path = os.path.join(self.vis_root, ann[\"images\"][1])\n image1 = Image.open(image1_path).convert(\"RGB\")\n image1 = self.vis_processor(image1)\n\n sentence = self.text_processor(ann[\"sentence\"])\n label = self.class_labels[ann[\"label\"]]\n\n return self._flip(\n {\n \"image0\": image0,\n \"image1\": image1,\n \"text_input\": sentence,\n \"label\": label,\n # \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n )\n\n\nclass NLVREvalDataset(NLVRDataset):\n @staticmethod\n def _flip(samples):\n return samples","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.nlvr_datasets.displ_item#L20-L31","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":20,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport random\n\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.nlvr_datasets.__init__#L35-L38","kind":"function","name":"__init__","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":35,"end_line":38,"context_start_line":15,"context_end_line":58,"code":")\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:\n if random.random() < 0.5:\n sentence = sentence.replace(\"left\", \"[TEMP_TOKEN]\")\n sentence = sentence.replace(\"right\", \"left\")\n sentence = sentence.replace(\"[TEMP_TOKEN]\", \"right\")\n\n image0, image1 = image1, image0\n","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets._build_class_labels","uri":"program://CREMA/function/lavis.datasets.datasets.nlvr_datasets._build_class_labels#L40-L41","kind":"function","name":"_build_class_labels","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":40,"end_line":41,"context_start_line":20,"context_end_line":61,"code":" def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file_L\": ann[\"images\"][0],\n \"file_R\": ann[\"images\"][1],\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": [sample[\"image0\"], sample[\"image1\"]],\n }\n )\n\n\nclass NLVRDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"False\": 0, \"True\": 1}\n\n @staticmethod\n def _flip(samples):\n sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:\n if random.random() < 0.5:\n sentence = sentence.replace(\"left\", \"[TEMP_TOKEN]\")\n sentence = sentence.replace(\"right\", \"left\")\n sentence = sentence.replace(\"[TEMP_TOKEN]\", \"right\")\n\n image0, image1 = image1, image0\n\n samples[\"text_input\"] = sentence\n samples[\"image0\"] = image0\n samples[\"image1\"] = image1","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets._flip","uri":"program://CREMA/function/lavis.datasets.datasets.nlvr_datasets._flip#L93-L94","kind":"function","name":"_flip","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":93,"end_line":94,"context_start_line":73,"context_end_line":94,"code":" image1 = Image.open(image1_path).convert(\"RGB\")\n image1 = self.vis_processor(image1)\n\n sentence = self.text_processor(ann[\"sentence\"])\n label = self.class_labels[ann[\"label\"]]\n\n return self._flip(\n {\n \"image0\": image0,\n \"image1\": image1,\n \"text_input\": sentence,\n \"label\": label,\n # \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n )\n\n\nclass NLVREvalDataset(NLVRDataset):\n @staticmethod\n def _flip(samples):\n return samples","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.nlvr_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.nlvr_datasets.__getitem__#L65-L88","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/nlvr_datasets.py","language":"python","start_line":65,"end_line":88,"context_start_line":45,"context_end_line":94,"code":" sentence = samples[\"text_input\"]\n image0, image1 = samples[\"image0\"], samples[\"image1\"]\n\n if \"left\" not in sentence and \"right\" not in sentence:\n if random.random() < 0.5:\n image0, image1 = image1, image0\n else:\n if random.random() < 0.5:\n sentence = sentence.replace(\"left\", \"[TEMP_TOKEN]\")\n sentence = sentence.replace(\"right\", \"left\")\n sentence = sentence.replace(\"[TEMP_TOKEN]\", \"right\")\n\n image0, image1 = image1, image0\n\n samples[\"text_input\"] = sentence\n samples[\"image0\"] = image0\n samples[\"image1\"] = image1\n\n return samples\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image0_path = os.path.join(self.vis_root, ann[\"images\"][0])\n image0 = Image.open(image0_path).convert(\"RGB\")\n image0 = self.vis_processor(image0)\n\n image1_path = os.path.join(self.vis_root, ann[\"images\"][1])\n image1 = Image.open(image1_path).convert(\"RGB\")\n image1 = self.vis_processor(image1)\n\n sentence = self.text_processor(ann[\"sentence\"])\n label = self.class_labels[ann[\"label\"]]\n\n return self._flip(\n {\n \"image0\": image0,\n \"image1\": image1,\n \"text_input\": sentence,\n \"label\": label,\n # \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n )\n\n\nclass NLVREvalDataset(NLVRDataset):\n @staticmethod\n def _flip(samples):\n return samples","source_hash":"43ba3e5b9485e14ad067bcdd91ddf2b0a73ba5d6152ed3f4a2fd9cee07857b48","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_caption_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.coco_caption_datasets#L1-L70","kind":"module","name":"lavis.datasets.datasets.coco_caption_datasets","path":"lavis/datasets/datasets/coco_caption_datasets.py","language":"python","start_line":1,"end_line":70,"context_start_line":1,"context_end_line":70,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\nfrom PIL import ImageFile\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\n\nfrom lavis.datasets.datasets.caption_datasets import CaptionDataset, CaptionEvalDataset\n\nCOCOCapDataset = CaptionDataset\n\n\nclass COCOCapEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"image\"].split(\"/\")[-1].strip(\".jpg\").split(\"_\")[-1]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass NoCapsEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"img_id\"]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"a32c5b04eee3052fb6c8288ffecf3b3dcb1f65f5cc2d3bdabeaea1e10a74d35f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_caption_datasets.COCOCapEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.coco_caption_datasets.COCOCapEvalDataset#L21-L44","kind":"class","name":"COCOCapEvalDataset","path":"lavis/datasets/datasets/coco_caption_datasets.py","language":"python","start_line":21,"end_line":44,"context_start_line":1,"context_end_line":64,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\nfrom PIL import ImageFile\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\n\nfrom lavis.datasets.datasets.caption_datasets import CaptionDataset, CaptionEvalDataset\n\nCOCOCapDataset = CaptionDataset\n\n\nclass COCOCapEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"image\"].split(\"/\")[-1].strip(\".jpg\").split(\"_\")[-1]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass NoCapsEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"img_id\"]","source_hash":"a32c5b04eee3052fb6c8288ffecf3b3dcb1f65f5cc2d3bdabeaea1e10a74d35f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_caption_datasets.NoCapsEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.coco_caption_datasets.NoCapsEvalDataset#L47-L70","kind":"class","name":"NoCapsEvalDataset","path":"lavis/datasets/datasets/coco_caption_datasets.py","language":"python","start_line":47,"end_line":70,"context_start_line":27,"context_end_line":70,"code":" \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"image\"].split(\"/\")[-1].strip(\".jpg\").split(\"_\")[-1]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass NoCapsEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"img_id\"]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"a32c5b04eee3052fb6c8288ffecf3b3dcb1f65f5cc2d3bdabeaea1e10a74d35f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_caption_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.coco_caption_datasets.__init__#L48-L54","kind":"function","name":"__init__","path":"lavis/datasets/datasets/coco_caption_datasets.py","language":"python","start_line":48,"end_line":54,"context_start_line":28,"context_end_line":70,"code":" super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"image\"].split(\"/\")[-1].strip(\".jpg\").split(\"_\")[-1]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass NoCapsEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"img_id\"]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"a32c5b04eee3052fb6c8288ffecf3b3dcb1f65f5cc2d3bdabeaea1e10a74d35f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_caption_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.coco_caption_datasets.__getitem__#L56-L70","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/coco_caption_datasets.py","language":"python","start_line":56,"end_line":70,"context_start_line":36,"context_end_line":70,"code":" image = self.vis_processor(image)\n\n img_id = ann[\"image\"].split(\"/\")[-1].strip(\".jpg\").split(\"_\")[-1]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }\n\n\nclass NoCapsEvalDataset(CaptionEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n img_id = ann[\"img_id\"]\n\n return {\n \"image\": image,\n \"image_id\": img_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"a32c5b04eee3052fb6c8288ffecf3b3dcb1f65f5cc2d3bdabeaea1e10a74d35f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.avsd_dialogue_datasets#L1-L166","kind":"module","name":"lavis.datasets.datasets.avsd_dialogue_datasets","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":1,"end_line":166,"context_start_line":1,"context_end_line":166,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.datasets.datasets.dialogue_datasets import (\n DialogueDataset,\n DialogueEvalDataset,\n)\n\n\nclass AVSDDialDataset(DialogueDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n\n labels = self.text_processor.padding(\n labels, -1\n ) # ignore token indice -1 by default\n video_fts = self.vis_processor.padding(video_fts)\n\n token_type_ids = self.text_processor.padding(token_type_ids)\n video_token_type_ids = self.text_processor.padding(video_token_type_ids)\n token_type_ids = torch.cat([video_token_type_ids, token_type_ids], dim=1)\n\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples\n\n\nclass AVSDDialEvalDataset(DialogueEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n\n labels = self.text_processor.padding(\n labels, -1\n ) # ignore token indice -1 by default\n video_fts = self.vis_processor.padding(video_fts)\n\n token_type_ids = self.text_processor.padding(token_type_ids)\n video_token_type_ids = self.text_processor.padding(video_token_type_ids)\n token_type_ids = torch.cat([video_token_type_ids, token_type_ids], dim=1)\n\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets.AVSDDialDataset","uri":"program://CREMA/class/lavis.datasets.datasets.avsd_dialogue_datasets.AVSDDialDataset#L15-L89","kind":"class","name":"AVSDDialDataset","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":15,"end_line":89,"context_start_line":1,"context_end_line":109,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.datasets.datasets.dialogue_datasets import (\n DialogueDataset,\n DialogueEvalDataset,\n)\n\n\nclass AVSDDialDataset(DialogueDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n\n labels = self.text_processor.padding(\n labels, -1\n ) # ignore token indice -1 by default\n video_fts = self.vis_processor.padding(video_fts)\n\n token_type_ids = self.text_processor.padding(token_type_ids)\n video_token_type_ids = self.text_processor.padding(video_token_type_ids)\n token_type_ids = torch.cat([video_token_type_ids, token_type_ids], dim=1)\n\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples\n\n\nclass AVSDDialEvalDataset(DialogueEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets.AVSDDialEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.avsd_dialogue_datasets.AVSDDialEvalDataset#L92-L166","kind":"class","name":"AVSDDialEvalDataset","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":92,"end_line":166,"context_start_line":72,"context_end_line":166,"code":"\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples\n\n\nclass AVSDDialEvalDataset(DialogueEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n\n labels = self.text_processor.padding(\n labels, -1\n ) # ignore token indice -1 by default\n video_fts = self.vis_processor.padding(video_fts)\n\n token_type_ids = self.text_processor.padding(token_type_ids)\n video_token_type_ids = self.text_processor.padding(video_token_type_ids)\n token_type_ids = torch.cat([video_token_type_ids, token_type_ids], dim=1)\n\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.avsd_dialogue_datasets.__init__#L93-L99","kind":"function","name":"__init__","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":93,"end_line":99,"context_start_line":73,"context_end_line":119,"code":" attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples\n\n\nclass AVSDDialEvalDataset(DialogueEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.avsd_dialogue_datasets.__getitem__#L101-L120","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":101,"end_line":120,"context_start_line":81,"context_end_line":140,"code":"\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples\n\n\nclass AVSDDialEvalDataset(DialogueEvalDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.avsd_dialogue_datasets.collater","uri":"program://CREMA/function/lavis.datasets.datasets.avsd_dialogue_datasets.collater#L122-L166","kind":"function","name":"collater","path":"lavis/datasets/datasets/avsd_dialogue_datasets.py","language":"python","start_line":122,"end_line":166,"context_start_line":102,"context_end_line":166,"code":"\n ann = self.annotation[index]\n\n vname = ann[\"image_id\"]\n\n video = self.vis_processor(self.vis_root, vname)\n\n dialogue = self.text_processor(ann)\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video_fts\": video[\"video_fts\"],\n \"video_token_type_ids\": video[\"token_type_ids\"],\n \"input_ids\": dialogue[\"input_ids\"],\n \"token_type_ids\": dialogue[\"token_type_ids\"],\n \"labels\": dialogue[\"labels\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def collater(self, samples):\n\n input_ids, token_type_ids, labels, video_fts, video_token_type_ids = (\n [],\n [],\n [],\n [],\n [],\n )\n\n for i in samples:\n input_ids.append(i[\"input_ids\"])\n token_type_ids.append(i[\"token_type_ids\"])\n labels.append(i[\"labels\"])\n video_fts.append(i[\"video_fts\"])\n video_token_type_ids.append(i[\"video_token_type_ids\"])\n\n input_ids = self.text_processor.padding(input_ids)\n\n labels = self.text_processor.padding(\n labels, -1\n ) # ignore token indice -1 by default\n video_fts = self.vis_processor.padding(video_fts)\n\n token_type_ids = self.text_processor.padding(token_type_ids)\n video_token_type_ids = self.text_processor.padding(video_token_type_ids)\n token_type_ids = torch.cat([video_token_type_ids, token_type_ids], dim=1)\n\n attn_mask = self.text_processor.get_attention_mask(input_ids)\n video_mask = self.vis_processor.get_attention_mask(video_fts)\n attn_mask = torch.cat([video_mask, attn_mask], dim=1)\n\n video_labels = (\n torch.ones((video_fts.size(0), video_fts.size(1))).long() * -1\n ) # ignore token indice -1 by default\n labels = torch.cat([video_labels, labels], dim=1)\n\n samples = {}\n samples[\"input_ids\"] = input_ids\n samples[\"token_type_ids\"] = token_type_ids\n samples[\"labels\"] = labels\n samples[\"video_fts\"] = video_fts\n samples[\"attn_mask\"] = attn_mask\n\n return samples","source_hash":"436abc29cf7177c1498c153686d1a3e1a167db7732d03d52e84478fdab33204b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.caption_datasets#L1-L84","kind":"module","name":"lavis.datasets.datasets.caption_datasets","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":1,"end_line":84,"context_start_line":1,"context_end_line":84,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass CaptionDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass CaptionEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.caption_datasets.__DisplMixin#L15-L25","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":15,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass CaptionDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.CaptionDataset","uri":"program://CREMA/class/lavis.datasets.datasets.caption_datasets.CaptionDataset#L28-L59","kind":"class","name":"CaptionDataset","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":28,"end_line":59,"context_start_line":8,"context_end_line":79,"code":"import os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass CaptionDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass CaptionEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.CaptionEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.caption_datasets.CaptionEvalDataset#L62-L84","kind":"class","name":"CaptionEvalDataset","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":62,"end_line":84,"context_start_line":42,"context_end_line":84,"code":" n += 1\n\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass CaptionEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.caption_datasets.displ_item#L16-L25","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":16,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"caption\": ann[\"caption\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass CaptionDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.img_ids = {}\n n = 0\n for ann in self.annotation:\n img_id = ann[\"image_id\"]\n if img_id not in self.img_ids.keys():\n self.img_ids[img_id] = n\n n += 1\n\n def __getitem__(self, index):\n","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.caption_datasets.__init__#L63-L69","kind":"function","name":"__init__","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":63,"end_line":69,"context_start_line":43,"context_end_line":84,"code":"\n def __getitem__(self, index):\n\n # TODO this assumes image input, not general enough\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass CaptionEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.caption_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.caption_datasets.__getitem__#L71-L84","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/caption_datasets.py","language":"python","start_line":71,"end_line":84,"context_start_line":51,"context_end_line":84,"code":"\n image = self.vis_processor(image)\n caption = self.text_processor(ann[\"caption\"])\n\n return {\n \"image\": image,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass CaptionEvalDataset(BaseDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"03d165a2c789444be87190f45623f79405f294605cd6d430697ca750bb791663","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset","uri":"program://CREMA/module/lavis.datasets.datasets.imagefolder_dataset#L1-L59","kind":"module","name":"lavis.datasets.datasets.imagefolder_dataset","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":1,"end_line":59,"context_start_line":1,"context_end_line":59,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\nfrom torchvision import datasets\n\n\nclass ImageFolderDataset(BaseDataset):\n def __init__(self, vis_processor, vis_root, classnames=[], **kwargs):\n super().__init__(vis_processor=vis_processor, vis_root=vis_root)\n\n self.inner_dataset = datasets.ImageFolder(vis_root)\n\n self.annotation = [\n {\"image\": elem[0], \"label\": elem[1], \"image_id\": elem[0]}\n for elem in self.inner_dataset.imgs\n ]\n\n self.classnames = classnames\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"label\": self.classnames[ann[\"label\"]],\n \"image\": sample[\"image\"],\n }\n )","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset.ImageFolderDataset","uri":"program://CREMA/class/lavis.datasets.datasets.imagefolder_dataset.ImageFolderDataset#L16-L59","kind":"class","name":"ImageFolderDataset","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":16,"end_line":59,"context_start_line":1,"context_end_line":59,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\nfrom torchvision import datasets\n\n\nclass ImageFolderDataset(BaseDataset):\n def __init__(self, vis_processor, vis_root, classnames=[], **kwargs):\n super().__init__(vis_processor=vis_processor, vis_root=vis_root)\n\n self.inner_dataset = datasets.ImageFolder(vis_root)\n\n self.annotation = [\n {\"image\": elem[0], \"label\": elem[1], \"image_id\": elem[0]}\n for elem in self.inner_dataset.imgs\n ]\n\n self.classnames = classnames\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"label\": self.classnames[ann[\"label\"]],\n \"image\": sample[\"image\"],\n }\n )","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.imagefolder_dataset.__init__#L17-L29","kind":"function","name":"__init__","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":17,"end_line":29,"context_start_line":1,"context_end_line":49,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\nfrom torchvision import datasets\n\n\nclass ImageFolderDataset(BaseDataset):\n def __init__(self, vis_processor, vis_root, classnames=[], **kwargs):\n super().__init__(vis_processor=vis_processor, vis_root=vis_root)\n\n self.inner_dataset = datasets.ImageFolder(vis_root)\n\n self.annotation = [\n {\"image\": elem[0], \"label\": elem[1], \"image_id\": elem[0]}\n for elem in self.inner_dataset.imgs\n ]\n\n self.classnames = classnames\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset.__len__","uri":"program://CREMA/function/lavis.datasets.datasets.imagefolder_dataset.__len__#L31-L32","kind":"function","name":"__len__","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":31,"end_line":32,"context_start_line":11,"context_end_line":52,"code":"from lavis.datasets.datasets.base_dataset import BaseDataset\nfrom PIL import Image\nfrom torchvision import datasets\n\n\nclass ImageFolderDataset(BaseDataset):\n def __init__(self, vis_processor, vis_root, classnames=[], **kwargs):\n super().__init__(vis_processor=vis_processor, vis_root=vis_root)\n\n self.inner_dataset = datasets.ImageFolder(vis_root)\n\n self.annotation = [\n {\"image\": elem[0], \"label\": elem[1], \"image_id\": elem[0]}\n for elem in self.inner_dataset.imgs\n ]\n\n self.classnames = classnames\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.imagefolder_dataset.__getitem__#L34-L48","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":34,"end_line":48,"context_start_line":14,"context_end_line":59,"code":"\n\nclass ImageFolderDataset(BaseDataset):\n def __init__(self, vis_processor, vis_root, classnames=[], **kwargs):\n super().__init__(vis_processor=vis_processor, vis_root=vis_root)\n\n self.inner_dataset = datasets.ImageFolder(vis_root)\n\n self.annotation = [\n {\"image\": elem[0], \"label\": elem[1], \"image_id\": elem[0]}\n for elem in self.inner_dataset.imgs\n ]\n\n self.classnames = classnames\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"label\": self.classnames[ann[\"label\"]],\n \"image\": sample[\"image\"],\n }\n )","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.imagefolder_dataset.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.imagefolder_dataset.displ_item#L50-L59","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/imagefolder_dataset.py","language":"python","start_line":50,"end_line":59,"context_start_line":30,"context_end_line":59,"code":"\n def __len__(self):\n return len(self.inner_dataset)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n img_fn = ann[\"image\"]\n image_path = os.path.join(self.vis_root, img_fn)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n\n return {\n \"image\": image,\n \"label\": ann[\"label\"],\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }\n\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"label\": self.classnames[ann[\"label\"]],\n \"image\": sample[\"image\"],\n }\n )","source_hash":"9746d9bcbeb3429a8f8c31f83c601ee63e8292335c5568e2b4119dbb274c0577","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.snli_ve_datasets#L1-L56","kind":"module","name":"lavis.datasets.datasets.snli_ve_datasets","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":1,"end_line":56,"context_start_line":1,"context_end_line":56,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])\n\n return {\n \"image\": image,\n \"text_input\": sentence,\n \"label\": self.class_labels[ann[\"label\"]],\n \"image_id\": image_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.snli_ve_datasets.__DisplMixin#L17-L28","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":17,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets.SNLIVisualEntialmentDataset","uri":"program://CREMA/class/lavis.datasets.datasets.snli_ve_datasets.SNLIVisualEntialmentDataset#L31-L56","kind":"class","name":"SNLIVisualEntialmentDataset","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":31,"end_line":56,"context_start_line":11,"context_end_line":56,"code":"from lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])\n\n return {\n \"image\": image,\n \"text_input\": sentence,\n \"label\": self.class_labels[ann[\"label\"]],\n \"image_id\": image_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.snli_ve_datasets.displ_item#L18-L28","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":18,"end_line":28,"context_start_line":1,"context_end_line":48,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.snli_ve_datasets.__init__#L32-L35","kind":"function","name":"__init__","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":32,"end_line":35,"context_start_line":12,"context_end_line":55,"code":" MultimodalClassificationDataset,\n)\nfrom PIL import Image\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])\n\n return {\n \"image\": image,\n \"text_input\": sentence,\n \"label\": self.class_labels[ann[\"label\"]],\n \"image_id\": image_id,\n \"instance_id\": ann[\"instance_id\"],","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets._build_class_labels","uri":"program://CREMA/function/lavis.datasets.datasets.snli_ve_datasets._build_class_labels#L37-L38","kind":"function","name":"_build_class_labels","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":37,"end_line":38,"context_start_line":17,"context_end_line":56,"code":"class __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])\n\n return {\n \"image\": image,\n \"text_input\": sentence,\n \"label\": self.class_labels[ann[\"label\"]],\n \"image_id\": image_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.snli_ve_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.snli_ve_datasets.__getitem__#L40-L56","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/snli_ve_datasets.py","language":"python","start_line":40,"end_line":56,"context_start_line":20,"context_end_line":56,"code":"\n return OrderedDict(\n {\n \"file\": os.path.basename(ann[\"image\"]),\n \"sentence\": ann[\"sentence\"],\n \"label\": ann[\"label\"],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass SNLIVisualEntialmentDataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n self.class_labels = self._build_class_labels()\n\n def _build_class_labels(self):\n return {\"contradiction\": 0, \"neutral\": 1, \"entailment\": 2}\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_id = ann[\"image\"]\n image_path = os.path.join(self.vis_root, \"%s.jpg\" % image_id)\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n sentence = self.text_processor(ann[\"sentence\"])\n\n return {\n \"image\": image,\n \"text_input\": sentence,\n \"label\": self.class_labels[ann[\"label\"]],\n \"image_id\": image_id,\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"97507d7f928e296b9ed214aa7a8530f586d34a36a2638e6195bacc9223af943f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.vqa_datasets#L1-L44","kind":"module","name":"lavis.datasets.datasets.vqa_datasets","path":"lavis/datasets/datasets/vqa_datasets.py","language":"python","start_line":1,"end_line":44,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass VQADataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def collater(self, samples):\n image_list, question_list, answer_list, weight_list = [], [], [], []\n\n num_answers = []\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n\n weight_list.extend(sample[\"weights\"])\n\n answers = sample[\"answers\"]\n\n answer_list.extend(answers)\n num_answers.append(len(answers))\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"answer\": answer_list,\n \"weight\": torch.Tensor(weight_list),\n \"n_answers\": torch.LongTensor(num_answers),\n }\n\n\nclass VQAEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)","source_hash":"8f6836389d6dd482a972e204ad8477ab2cfca1649b40f2e379cc981aa9a728f6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vqa_datasets.VQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.vqa_datasets.VQADataset#L13-L39","kind":"class","name":"VQADataset","path":"lavis/datasets/datasets/vqa_datasets.py","language":"python","start_line":13,"end_line":39,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass VQADataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def collater(self, samples):\n image_list, question_list, answer_list, weight_list = [], [], [], []\n\n num_answers = []\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n\n weight_list.extend(sample[\"weights\"])\n\n answers = sample[\"answers\"]\n\n answer_list.extend(answers)\n num_answers.append(len(answers))\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"answer\": answer_list,\n \"weight\": torch.Tensor(weight_list),\n \"n_answers\": torch.LongTensor(num_answers),\n }\n\n\nclass VQAEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)","source_hash":"8f6836389d6dd482a972e204ad8477ab2cfca1649b40f2e379cc981aa9a728f6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vqa_datasets.VQAEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.vqa_datasets.VQAEvalDataset#L42-L44","kind":"class","name":"VQAEvalDataset","path":"lavis/datasets/datasets/vqa_datasets.py","language":"python","start_line":42,"end_line":44,"context_start_line":22,"context_end_line":44,"code":" for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n\n weight_list.extend(sample[\"weights\"])\n\n answers = sample[\"answers\"]\n\n answer_list.extend(answers)\n num_answers.append(len(answers))\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"answer\": answer_list,\n \"weight\": torch.Tensor(weight_list),\n \"n_answers\": torch.LongTensor(num_answers),\n }\n\n\nclass VQAEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)","source_hash":"8f6836389d6dd482a972e204ad8477ab2cfca1649b40f2e379cc981aa9a728f6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.vqa_datasets.__init__#L43-L44","kind":"function","name":"__init__","path":"lavis/datasets/datasets/vqa_datasets.py","language":"python","start_line":43,"end_line":44,"context_start_line":23,"context_end_line":44,"code":" image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n\n weight_list.extend(sample[\"weights\"])\n\n answers = sample[\"answers\"]\n\n answer_list.extend(answers)\n num_answers.append(len(answers))\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"answer\": answer_list,\n \"weight\": torch.Tensor(weight_list),\n \"n_answers\": torch.LongTensor(num_answers),\n }\n\n\nclass VQAEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)","source_hash":"8f6836389d6dd482a972e204ad8477ab2cfca1649b40f2e379cc981aa9a728f6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vqa_datasets.collater","uri":"program://CREMA/function/lavis.datasets.datasets.vqa_datasets.collater#L17-L39","kind":"function","name":"collater","path":"lavis/datasets/datasets/vqa_datasets.py","language":"python","start_line":17,"end_line":39,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\n\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass VQADataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def collater(self, samples):\n image_list, question_list, answer_list, weight_list = [], [], [], []\n\n num_answers = []\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n\n weight_list.extend(sample[\"weights\"])\n\n answers = sample[\"answers\"]\n\n answer_list.extend(answers)\n num_answers.append(len(answers))\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"answer\": answer_list,\n \"weight\": torch.Tensor(weight_list),\n \"n_answers\": torch.LongTensor(num_answers),\n }\n\n\nclass VQAEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)","source_hash":"8f6836389d6dd482a972e204ad8477ab2cfca1649b40f2e379cc981aa9a728f6","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_caption_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.video_caption_datasets#L1-L63","kind":"module","name":"lavis.datasets.datasets.video_caption_datasets","path":"lavis/datasets/datasets/video_caption_datasets.py","language":"python","start_line":1,"end_line":63,"context_start_line":1,"context_end_line":63,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nfrom lavis.datasets.datasets.caption_datasets import CaptionDataset\n\n\nclass VideoCaptionDataset(CaptionDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n caption = self.text_processor(ann[\"caption\"])\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass VideoCaptionEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n\n return {\n \"video\": video,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"61e6181f407be1592e59bc6eba3c5da67781c5061d07279167dc90db07faa245","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_caption_datasets.VideoCaptionDataset","uri":"program://CREMA/class/lavis.datasets.datasets.video_caption_datasets.VideoCaptionDataset#L14-L38","kind":"class","name":"VideoCaptionDataset","path":"lavis/datasets/datasets/video_caption_datasets.py","language":"python","start_line":14,"end_line":38,"context_start_line":1,"context_end_line":58,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\nfrom lavis.datasets.datasets.caption_datasets import CaptionDataset\n\n\nclass VideoCaptionDataset(CaptionDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n caption = self.text_processor(ann[\"caption\"])\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass VideoCaptionEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n","source_hash":"61e6181f407be1592e59bc6eba3c5da67781c5061d07279167dc90db07faa245","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_caption_datasets.VideoCaptionEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.video_caption_datasets.VideoCaptionEvalDataset#L41-L63","kind":"class","name":"VideoCaptionEvalDataset","path":"lavis/datasets/datasets/video_caption_datasets.py","language":"python","start_line":41,"end_line":63,"context_start_line":21,"context_end_line":63,"code":" super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n caption = self.text_processor(ann[\"caption\"])\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass VideoCaptionEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n\n return {\n \"video\": video,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"61e6181f407be1592e59bc6eba3c5da67781c5061d07279167dc90db07faa245","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_caption_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.video_caption_datasets.__init__#L42-L48","kind":"function","name":"__init__","path":"lavis/datasets/datasets/video_caption_datasets.py","language":"python","start_line":42,"end_line":48,"context_start_line":22,"context_end_line":63,"code":"\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n caption = self.text_processor(ann[\"caption\"])\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass VideoCaptionEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n\n return {\n \"video\": video,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"61e6181f407be1592e59bc6eba3c5da67781c5061d07279167dc90db07faa245","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_caption_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.video_caption_datasets.__getitem__#L50-L63","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/video_caption_datasets.py","language":"python","start_line":50,"end_line":63,"context_start_line":30,"context_end_line":63,"code":" video = self.vis_processor(video_path)\n caption = self.text_processor(ann[\"caption\"])\n\n # \"image_id\" is kept to stay compatible with the COCO evaluation format\n return {\n \"video\": video,\n \"text_input\": caption,\n \"image_id\": self.img_ids[ann[\"image_id\"]],\n }\n\n\nclass VideoCaptionEvalDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n split (string): val or test\n \"\"\"\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n video_path = os.path.join(self.vis_root, vname)\n\n video = self.vis_processor(video_path)\n\n return {\n \"video\": video,\n \"image_id\": ann[\"image_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"61e6181f407be1592e59bc6eba3c5da67781c5061d07279167dc90db07faa245","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils","uri":"program://CREMA/module/lavis.datasets.datasets.dataloader_utils#L1-L162","kind":"module","name":"lavis.datasets.datasets.dataloader_utils","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":1,"end_line":162,"context_start_line":1,"context_end_line":162,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport time\nimport random\nimport torch\nfrom lavis.datasets.data_utils import move_to_cuda\nfrom torch.utils.data import DataLoader\n\n\nclass MultiIterLoader:\n \"\"\"\n A simple wrapper for iterating over multiple iterators.\n\n Args:\n loaders (List[Loader]): List of Iterator loaders.\n ratios (List[float]): List of ratios to sample from each loader. If None, all loaders are sampled uniformly.\n \"\"\"\n\n def __init__(self, loaders, ratios=None):\n # assert all loaders has __next__ method\n for loader in loaders:\n assert hasattr(\n loader, \"__next__\"\n ), \"Loader {} has no __next__ method.\".format(loader)\n\n if ratios is None:\n ratios = [1.0] * len(loaders)\n else:\n assert len(ratios) == len(loaders)\n ratios = [float(ratio) / sum(ratios) for ratio in ratios]\n\n self.loaders = loaders\n self.ratios = ratios\n\n def __next__(self):\n # random sample from each loader by ratio\n loader_idx = random.choices(range(len(self.loaders)), self.ratios, k=1)[0]\n return next(self.loaders[loader_idx])\n\n\nclass PrefetchLoader(object):\n \"\"\"\n Modified from https://github.com/ChenRocks/UNITER.\n\n overlap compute and cuda data transfer\n (copied and then modified from nvidia apex)\n \"\"\"\n\n def __init__(self, loader):\n self.loader = loader\n self.stream = torch.cuda.Stream()\n\n def __iter__(self):\n loader_it = iter(self.loader)\n self.preload(loader_it)\n batch = self.next(loader_it)\n while batch is not None:\n is_tuple = isinstance(batch, tuple)\n if is_tuple:\n task, batch = batch\n\n if is_tuple:\n yield task, batch\n else:\n yield batch\n batch = self.next(loader_it)\n\n def __len__(self):\n return len(self.loader)\n\n def preload(self, it):\n try:\n self.batch = next(it)\n except StopIteration:\n self.batch = None\n return\n # if record_stream() doesn't work, another option is to make sure\n # device inputs are created on the main stream.\n # self.next_input_gpu = torch.empty_like(self.next_input,\n # device='cuda')\n # self.next_target_gpu = torch.empty_like(self.next_target,\n # device='cuda')\n # Need to make sure the memory allocated for next_* is not still in use\n # by the main stream at the time we start copying to next_*:\n # self.stream.wait_stream(torch.cuda.current_stream())\n with torch.cuda.stream(self.stream):\n self.batch = move_to_cuda(self.batch)\n # more code for the alternative if record_stream() doesn't work:\n # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.MultiIterLoader","uri":"program://CREMA/class/lavis.datasets.datasets.dataloader_utils.MultiIterLoader#L15-L43","kind":"class","name":"MultiIterLoader","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":15,"end_line":43,"context_start_line":1,"context_end_line":63,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport time\nimport random\nimport torch\nfrom lavis.datasets.data_utils import move_to_cuda\nfrom torch.utils.data import DataLoader\n\n\nclass MultiIterLoader:\n \"\"\"\n A simple wrapper for iterating over multiple iterators.\n\n Args:\n loaders (List[Loader]): List of Iterator loaders.\n ratios (List[float]): List of ratios to sample from each loader. If None, all loaders are sampled uniformly.\n \"\"\"\n\n def __init__(self, loaders, ratios=None):\n # assert all loaders has __next__ method\n for loader in loaders:\n assert hasattr(\n loader, \"__next__\"\n ), \"Loader {} has no __next__ method.\".format(loader)\n\n if ratios is None:\n ratios = [1.0] * len(loaders)\n else:\n assert len(ratios) == len(loaders)\n ratios = [float(ratio) / sum(ratios) for ratio in ratios]\n\n self.loaders = loaders\n self.ratios = ratios\n\n def __next__(self):\n # random sample from each loader by ratio\n loader_idx = random.choices(range(len(self.loaders)), self.ratios, k=1)[0]\n return next(self.loaders[loader_idx])\n\n\nclass PrefetchLoader(object):\n \"\"\"\n Modified from https://github.com/ChenRocks/UNITER.\n\n overlap compute and cuda data transfer\n (copied and then modified from nvidia apex)\n \"\"\"\n\n def __init__(self, loader):\n self.loader = loader\n self.stream = torch.cuda.Stream()\n\n def __iter__(self):\n loader_it = iter(self.loader)\n self.preload(loader_it)\n batch = self.next(loader_it)\n while batch is not None:\n is_tuple = isinstance(batch, tuple)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.PrefetchLoader","uri":"program://CREMA/class/lavis.datasets.datasets.dataloader_utils.PrefetchLoader#L46-L111","kind":"class","name":"PrefetchLoader","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":46,"end_line":111,"context_start_line":26,"context_end_line":131,"code":" for loader in loaders:\n assert hasattr(\n loader, \"__next__\"\n ), \"Loader {} has no __next__ method.\".format(loader)\n\n if ratios is None:\n ratios = [1.0] * len(loaders)\n else:\n assert len(ratios) == len(loaders)\n ratios = [float(ratio) / sum(ratios) for ratio in ratios]\n\n self.loaders = loaders\n self.ratios = ratios\n\n def __next__(self):\n # random sample from each loader by ratio\n loader_idx = random.choices(range(len(self.loaders)), self.ratios, k=1)[0]\n return next(self.loaders[loader_idx])\n\n\nclass PrefetchLoader(object):\n \"\"\"\n Modified from https://github.com/ChenRocks/UNITER.\n\n overlap compute and cuda data transfer\n (copied and then modified from nvidia apex)\n \"\"\"\n\n def __init__(self, loader):\n self.loader = loader\n self.stream = torch.cuda.Stream()\n\n def __iter__(self):\n loader_it = iter(self.loader)\n self.preload(loader_it)\n batch = self.next(loader_it)\n while batch is not None:\n is_tuple = isinstance(batch, tuple)\n if is_tuple:\n task, batch = batch\n\n if is_tuple:\n yield task, batch\n else:\n yield batch\n batch = self.next(loader_it)\n\n def __len__(self):\n return len(self.loader)\n\n def preload(self, it):\n try:\n self.batch = next(it)\n except StopIteration:\n self.batch = None\n return\n # if record_stream() doesn't work, another option is to make sure\n # device inputs are created on the main stream.\n # self.next_input_gpu = torch.empty_like(self.next_input,\n # device='cuda')\n # self.next_target_gpu = torch.empty_like(self.next_target,\n # device='cuda')\n # Need to make sure the memory allocated for next_* is not still in use\n # by the main stream at the time we start copying to next_*:\n # self.stream.wait_stream(torch.cuda.current_stream())\n with torch.cuda.stream(self.stream):\n self.batch = move_to_cuda(self.batch)\n # more code for the alternative if record_stream() doesn't work:\n # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.record_cuda_stream","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.record_cuda_stream#L114-L124","kind":"function","name":"record_cuda_stream","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":114,"end_line":124,"context_start_line":94,"context_end_line":144,"code":" # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.IterLoader","uri":"program://CREMA/class/lavis.datasets.datasets.dataloader_utils.IterLoader#L127-L162","kind":"class","name":"IterLoader","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":127,"end_line":162,"context_start_line":107,"context_end_line":162,"code":" return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.__init__#L135-L139","kind":"function","name":"__init__","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":135,"end_line":139,"context_start_line":115,"context_end_line":159,"code":" if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.__next__","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.__next__#L145-L156","kind":"function","name":"__next__","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":145,"end_line":156,"context_start_line":125,"context_end_line":162,"code":"\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.__iter__","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.__iter__#L158-L159","kind":"function","name":"__iter__","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":158,"end_line":159,"context_start_line":138,"context_end_line":162,"code":" self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.__len__","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.__len__#L161-L162","kind":"function","name":"__len__","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":161,"end_line":162,"context_start_line":141,"context_end_line":162,"code":" @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.preload","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.preload#L76-L92","kind":"function","name":"preload","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":76,"end_line":92,"context_start_line":56,"context_end_line":112,"code":" self.stream = torch.cuda.Stream()\n\n def __iter__(self):\n loader_it = iter(self.loader)\n self.preload(loader_it)\n batch = self.next(loader_it)\n while batch is not None:\n is_tuple = isinstance(batch, tuple)\n if is_tuple:\n task, batch = batch\n\n if is_tuple:\n yield task, batch\n else:\n yield batch\n batch = self.next(loader_it)\n\n def __len__(self):\n return len(self.loader)\n\n def preload(self, it):\n try:\n self.batch = next(it)\n except StopIteration:\n self.batch = None\n return\n # if record_stream() doesn't work, another option is to make sure\n # device inputs are created on the main stream.\n # self.next_input_gpu = torch.empty_like(self.next_input,\n # device='cuda')\n # self.next_target_gpu = torch.empty_like(self.next_target,\n # device='cuda')\n # Need to make sure the memory allocated for next_* is not still in use\n # by the main stream at the time we start copying to next_*:\n # self.stream.wait_stream(torch.cuda.current_stream())\n with torch.cuda.stream(self.stream):\n self.batch = move_to_cuda(self.batch)\n # more code for the alternative if record_stream() doesn't work:\n # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.next","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.next#L101-L107","kind":"function","name":"next","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":101,"end_line":107,"context_start_line":81,"context_end_line":127,"code":" return\n # if record_stream() doesn't work, another option is to make sure\n # device inputs are created on the main stream.\n # self.next_input_gpu = torch.empty_like(self.next_input,\n # device='cuda')\n # self.next_target_gpu = torch.empty_like(self.next_target,\n # device='cuda')\n # Need to make sure the memory allocated for next_* is not still in use\n # by the main stream at the time we start copying to next_*:\n # self.stream.wait_stream(torch.cuda.current_stream())\n with torch.cuda.stream(self.stream):\n self.batch = move_to_cuda(self.batch)\n # more code for the alternative if record_stream() doesn't work:\n # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.__getattr__","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.__getattr__#L109-L111","kind":"function","name":"__getattr__","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":109,"end_line":111,"context_start_line":89,"context_end_line":131,"code":" # by the main stream at the time we start copying to next_*:\n # self.stream.wait_stream(torch.cuda.current_stream())\n with torch.cuda.stream(self.stream):\n self.batch = move_to_cuda(self.batch)\n # more code for the alternative if record_stream() doesn't work:\n # copy_ will record the use of the pinned source tensor in this\n # side stream.\n # self.next_input_gpu.copy_(self.next_input, non_blocking=True)\n # self.next_target_gpu.copy_(self.next_target, non_blocking=True)\n # self.next_input = self.next_input_gpu\n # self.next_target = self.next_target_gpu\n\n def next(self, it):\n torch.cuda.current_stream().wait_stream(self.stream)\n batch = self.batch\n if batch is not None:\n record_cuda_stream(batch)\n self.preload(it)\n return batch\n\n def __getattr__(self, name):\n method = self.loader.__getattribute__(name)\n return method\n\n\ndef record_cuda_stream(batch):\n if isinstance(batch, torch.Tensor):\n batch.record_stream(torch.cuda.current_stream())\n elif isinstance(batch, list) or isinstance(batch, tuple):\n for t in batch:\n record_cuda_stream(t)\n elif isinstance(batch, dict):\n for t in batch.values():\n record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.dataloader_utils.epoch","uri":"program://CREMA/function/lavis.datasets.datasets.dataloader_utils.epoch#L142-L143","kind":"function","name":"epoch","path":"lavis/datasets/datasets/dataloader_utils.py","language":"python","start_line":142,"end_line":143,"context_start_line":122,"context_end_line":162,"code":" record_cuda_stream(t)\n else:\n pass\n\n\nclass IterLoader:\n \"\"\"\n A wrapper to convert DataLoader as an infinite iterator.\n\n Modified from:\n https://github.com/open-mmlab/mmcv/blob/master/mmcv/runner/iter_based_runner.py\n \"\"\"\n\n def __init__(self, dataloader: DataLoader, use_distributed: bool = False):\n self._dataloader = dataloader\n self.iter_loader = iter(self._dataloader)\n self._use_distributed = use_distributed\n self._epoch = 0\n\n @property\n def epoch(self) -> int:\n return self._epoch\n\n def __next__(self):\n try:\n data = next(self.iter_loader)\n except StopIteration:\n self._epoch += 1\n if hasattr(self._dataloader.sampler, \"set_epoch\") and self._use_distributed:\n self._dataloader.sampler.set_epoch(self._epoch)\n time.sleep(2) # Prevent possible deadlock during epoch transition\n self.iter_loader = iter(self._dataloader)\n data = next(self.iter_loader)\n\n return data\n\n def __iter__(self):\n return self\n\n def __len__(self):\n return len(self._dataloader)","source_hash":"d8bb08e45b75d65e0388a6598a034e6c786e36d0cec25232f445e6ef2b4dd4df","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vg_vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.vg_vqa_datasets#L1-L37","kind":"module","name":"lavis.datasets.datasets.vg_vqa_datasets","path":"lavis/datasets/datasets/vg_vqa_datasets.py","language":"python","start_line":1,"end_line":37,"context_start_line":1,"context_end_line":37,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset\n\n\nclass VGVQADataset(VQADataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n # TODO this should be configured better\n weights = [0.2]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }","source_hash":"a28ed0a66cd031f312e45e302568bf0f8050ab5c5b4dc27900220d96b91c7491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vg_vqa_datasets.VGVQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.vg_vqa_datasets.VGVQADataset#L15-L37","kind":"class","name":"VGVQADataset","path":"lavis/datasets/datasets/vg_vqa_datasets.py","language":"python","start_line":15,"end_line":37,"context_start_line":1,"context_end_line":37,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset\n\n\nclass VGVQADataset(VQADataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n # TODO this should be configured better\n weights = [0.2]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }","source_hash":"a28ed0a66cd031f312e45e302568bf0f8050ab5c5b4dc27900220d96b91c7491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vg_vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.vg_vqa_datasets.__init__#L16-L17","kind":"function","name":"__init__","path":"lavis/datasets/datasets/vg_vqa_datasets.py","language":"python","start_line":16,"end_line":17,"context_start_line":1,"context_end_line":37,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset\n\n\nclass VGVQADataset(VQADataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n # TODO this should be configured better\n weights = [0.2]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }","source_hash":"a28ed0a66cd031f312e45e302568bf0f8050ab5c5b4dc27900220d96b91c7491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.vg_vqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.vg_vqa_datasets.__getitem__#L19-L37","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/vg_vqa_datasets.py","language":"python","start_line":19,"end_line":37,"context_start_line":1,"context_end_line":37,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset\n\n\nclass VGVQADataset(VQADataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n # TODO this should be configured better\n weights = [0.2]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }","source_hash":"a28ed0a66cd031f312e45e302568bf0f8050ab5c5b4dc27900220d96b91c7491","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset","uri":"program://CREMA/module/lavis.datasets.datasets.base_dataset#L1-L85","kind":"module","name":"lavis.datasets.datasets.base_dataset","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":1,"end_line":85,"context_start_line":1,"context_end_line":85,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport pandas as pd\n\nfrom typing import Iterable\nfrom torch.utils.data import Dataset, ConcatDataset\nfrom torch.utils.data.dataloader import default_collate\n\n\nclass BaseDataset(Dataset):\n def __init__(\n self, vis_processor=None, text_processor=None, vis_root=None, ann_paths=[], modality_type=None,\n ):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n self.vis_root = vis_root\n self.modality_type = modality_type\n\n self.annotation = []\n\n # print(ann_paths)\n\n for ann_path in ann_paths:\n #print('herrrr', ann_path)\n if '.json' in ann_path:\n self.annotation.extend(json.load(open(ann_path, \"r\")))\n if 'train' in ann_path: \n self.data_type = 'train'\n else:\n self.data_type = 'val'\n else:\n raise AttributeError('Undefined data type')\n \n #self.annotation = self.annotation[:100] \n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})\n\n return self.datasets[0].collater(samples_shared_keys)","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.BaseDataset","uri":"program://CREMA/class/lavis.datasets.datasets.base_dataset.BaseDataset#L16-L63","kind":"class","name":"BaseDataset","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":16,"end_line":63,"context_start_line":1,"context_end_line":83,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport pandas as pd\n\nfrom typing import Iterable\nfrom torch.utils.data import Dataset, ConcatDataset\nfrom torch.utils.data.dataloader import default_collate\n\n\nclass BaseDataset(Dataset):\n def __init__(\n self, vis_processor=None, text_processor=None, vis_root=None, ann_paths=[], modality_type=None,\n ):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n self.vis_root = vis_root\n self.modality_type = modality_type\n\n self.annotation = []\n\n # print(ann_paths)\n\n for ann_path in ann_paths:\n #print('herrrr', ann_path)\n if '.json' in ann_path:\n self.annotation.extend(json.load(open(ann_path, \"r\")))\n if 'train' in ann_path: \n self.data_type = 'train'\n else:\n self.data_type = 'val'\n else:\n raise AttributeError('Undefined data type')\n \n #self.annotation = self.annotation[:100] \n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.ConcatDataset","uri":"program://CREMA/class/lavis.datasets.datasets.base_dataset.ConcatDataset#L66-L85","kind":"class","name":"ConcatDataset","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":66,"end_line":85,"context_start_line":46,"context_end_line":85,"code":" self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})\n\n return self.datasets[0].collater(samples_shared_keys)","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.base_dataset.__init__#L67-L68","kind":"function","name":"__init__","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":67,"end_line":68,"context_start_line":47,"context_end_line":85,"code":"\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})\n\n return self.datasets[0].collater(samples_shared_keys)","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.__len__","uri":"program://CREMA/function/lavis.datasets.datasets.base_dataset.__len__#L48-L49","kind":"function","name":"__len__","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":48,"end_line":49,"context_start_line":28,"context_end_line":69,"code":"\n # print(ann_paths)\n\n for ann_path in ann_paths:\n #print('herrrr', ann_path)\n if '.json' in ann_path:\n self.annotation.extend(json.load(open(ann_path, \"r\")))\n if 'train' in ann_path: \n self.data_type = 'train'\n else:\n self.data_type = 'val'\n else:\n raise AttributeError('Undefined data type')\n \n #self.annotation = self.annotation[:100] \n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.collater","uri":"program://CREMA/function/lavis.datasets.datasets.base_dataset.collater#L70-L85","kind":"function","name":"collater","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":70,"end_line":85,"context_start_line":50,"context_end_line":85,"code":"\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})\n\n return self.datasets[0].collater(samples_shared_keys)","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset.set_processors","uri":"program://CREMA/function/lavis.datasets.datasets.base_dataset.set_processors#L54-L56","kind":"function","name":"set_processors","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":54,"end_line":56,"context_start_line":34,"context_end_line":76,"code":" self.annotation.extend(json.load(open(ann_path, \"r\")))\n if 'train' in ann_path: \n self.data_type = 'train'\n else:\n self.data_type = 'val'\n else:\n raise AttributeError('Undefined data type')\n \n #self.annotation = self.annotation[:100] \n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.base_dataset._add_instance_ids","uri":"program://CREMA/function/lavis.datasets.datasets.base_dataset._add_instance_ids#L58-L63","kind":"function","name":"_add_instance_ids","path":"lavis/datasets/datasets/base_dataset.py","language":"python","start_line":58,"end_line":63,"context_start_line":38,"context_end_line":83,"code":" self.data_type = 'val'\n else:\n raise AttributeError('Undefined data type')\n \n #self.annotation = self.annotation[:100] \n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __len__(self):\n return len(self.annotation)\n\n def collater(self, samples):\n return default_collate(samples)\n\n def set_processors(self, vis_processor, text_processor):\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n def _add_instance_ids(self, key=\"instance_id\"):\n for idx, ann in enumerate(self.annotation): \n if isinstance(ann, str):\n pass\n else:\n ann[key] = str(idx)\n\n\nclass ConcatDataset(ConcatDataset):\n def __init__(self, datasets: Iterable[Dataset]) -> None:\n super().__init__(datasets)\n\n def collater(self, samples):\n # TODO For now only supports datasets with same underlying collater implementations\n\n all_keys = set()\n for s in samples:\n all_keys.update(s)\n\n shared_keys = all_keys\n for s in samples:\n shared_keys = shared_keys & set(s.keys())\n\n samples_shared_keys = []\n for s in samples:\n samples_shared_keys.append({k: s[k] for k in s.keys() if k in shared_keys})","source_hash":"e36ca4815ec8bc7e185402acef50e7e0044fb2b0146874a2a3abe9a5f6e0496f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.threedvqa_datasets#L1-L282","kind":"module","name":"lavis.datasets.datasets.threedvqa_datasets","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":1,"end_line":282,"context_start_line":1,"context_end_line":282,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\nimport torch\nimport numpy as np\nimport copy\n\nfrom PIL import Image\nfrom PIL import ImageFile\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\ndef get_qclass(question):\n lques = question\n if 'What' in lques:\n return 'What'\n if 'How' in lques:\n return 'How'\n if 'Can' in lques:\n return 'Can'\n if 'Is' in lques:\n return 'Is'\n if 'Which' in lques:\n return 'Which'\n return 'Other'\n\nclass ThreeDVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation\n \n\n for modality in self.modalities:\n if 'pc' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n self.pc_feat_root = self.pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n self.voxel_root = self.pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n \n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n\n out = {\n \"qa_input\": qa_input,\n \"qa_output\": answer,\n \"scene_id\": self.scene_ids[ann[\"scene_id\"]],\n \"question_id\": question_id,\n }\n\n for modality in self.modalities:\n \n if modality == 'pc':\n pc_feat = torch.load(os.path.join(self.pc_feat_root, f\"{scene_id}.pt\"), map_location=\"cpu\")\n if isinstance(pc_feat, np.ndarray):\n pc_feat = torch.tensor(pc_feat).float()\n pc = np.load(os.path.join(self.voxel_root, f\"{scene_id}.npy\"))\n pc = torch.tensor(pc).float().cpu()\n if pc_feat.shape[0] > 5000:\n idxes = torch.sort(torch.randperm(pc_feat.shape[0])[:5000])[1]\n pc_feat = pc_feat[idxes]\n pc = pc[idxes]\n else:\n pc_feat = torch.cat([pc_feat, torch.zeros(5000 - pc_feat.shape[0], 1408)], dim=0)\n pc = torch.cat([pc, torch.zeros(5000 - pc.shape[0], 3)], dim=0)\n out[\"pc_feat\"] = pc_feat\n out['pc'] = pc\n \n if modality == 'video':\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)\n\n\nclass ThreeDVQAEvalDataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation\n \n for modality in self.modalities:\n if 'pc' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n self.pc_feat_root = self.pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n self.voxel_root = self.pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n # pc_root = '/nas-ssd2/shoubin/datasets/scannet_feat/'\n # self.pc_feat_root = pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n # self.voxel_root = pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n \n\n out = {\n \"qa_input\": qa_input,\n \"qa_output\": answer,\n \"scene_id\": self.scene_ids[ann[\"scene_id\"]],\n \"question_id\": question_id,\n }\n\n for modality in self.modalities:\n \n if modality == 'pc':\n pc_feat = torch.load(os.path.join(self.pc_feat_root, f\"{scene_id}.pt\"), map_location=\"cpu\") # [N, 1408]\n if isinstance(pc_feat, np.ndarray):\n pc_feat = torch.tensor(pc_feat).float()\n pc = np.load(os.path.join(self.voxel_root, f\"{scene_id}.npy\"))\n pc = torch.tensor(pc).float().cpu()\n # sample 10000 points: [N, 1408] -> [10000, 1408]\n if pc_feat.shape[0] > 5000:\n idxes = torch.sort(torch.randperm(pc_feat.shape[0])[:5000])[1]\n pc_feat = pc_feat[idxes]\n pc = pc[idxes]\n else:\n pc_feat = torch.cat([pc_feat, torch.zeros(5000 - pc_feat.shape[0], 1408)], dim=0)\n pc = torch.cat([pc, torch.zeros(5000 - pc.shape[0], 3)], dim=0)\n\n out[\"pc_feat\"] = pc_feat\n out['pc'] = pc\n if modality == 'video':\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.get_qclass","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.get_qclass#L21-L33","kind":"function","name":"get_qclass","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":21,"end_line":33,"context_start_line":1,"context_end_line":53,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\nimport torch\nimport numpy as np\nimport copy\n\nfrom PIL import Image\nfrom PIL import ImageFile\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\ndef get_qclass(question):\n lques = question\n if 'What' in lques:\n return 'What'\n if 'How' in lques:\n return 'How'\n if 'Can' in lques:\n return 'Can'\n if 'Is' in lques:\n return 'Is'\n if 'Which' in lques:\n return 'Which'\n return 'Other'\n\nclass ThreeDVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.ThreeDVQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.threedvqa_datasets.ThreeDVQADataset#L35-L154","kind":"class","name":"ThreeDVQADataset","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":35,"end_line":154,"context_start_line":15,"context_end_line":174,"code":"from PIL import ImageFile\n\nImageFile.LOAD_TRUNCATED_IMAGES = True\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\ndef get_qclass(question):\n lques = question\n if 'What' in lques:\n return 'What'\n if 'How' in lques:\n return 'How'\n if 'Can' in lques:\n return 'Can'\n if 'Is' in lques:\n return 'Is'\n if 'Which' in lques:\n return 'Which'\n return 'Other'\n\nclass ThreeDVQADataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation\n \n\n for modality in self.modalities:\n if 'pc' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n self.pc_feat_root = self.pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n self.voxel_root = self.pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n \n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n \n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n\n out = {\n \"qa_input\": qa_input,\n \"qa_output\": answer,\n \"scene_id\": self.scene_ids[ann[\"scene_id\"]],\n \"question_id\": question_id,\n }\n\n for modality in self.modalities:\n \n if modality == 'pc':\n pc_feat = torch.load(os.path.join(self.pc_feat_root, f\"{scene_id}.pt\"), map_location=\"cpu\")\n if isinstance(pc_feat, np.ndarray):\n pc_feat = torch.tensor(pc_feat).float()\n pc = np.load(os.path.join(self.voxel_root, f\"{scene_id}.npy\"))\n pc = torch.tensor(pc).float().cpu()\n if pc_feat.shape[0] > 5000:\n idxes = torch.sort(torch.randperm(pc_feat.shape[0])[:5000])[1]\n pc_feat = pc_feat[idxes]\n pc = pc[idxes]\n else:\n pc_feat = torch.cat([pc_feat, torch.zeros(5000 - pc_feat.shape[0], 1408)], dim=0)\n pc = torch.cat([pc, torch.zeros(5000 - pc.shape[0], 3)], dim=0)\n out[\"pc_feat\"] = pc_feat\n out['pc'] = pc\n \n if modality == 'video':\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)\n\n\nclass ThreeDVQAEvalDataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.ThreeDVQAEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.threedvqa_datasets.ThreeDVQAEvalDataset#L157-L282","kind":"class","name":"ThreeDVQAEvalDataset","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":157,"end_line":282,"context_start_line":137,"context_end_line":282,"code":" indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)\n\n\nclass ThreeDVQAEvalDataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation\n \n for modality in self.modalities:\n if 'pc' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n self.pc_feat_root = self.pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n self.voxel_root = self.pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n # pc_root = '/nas-ssd2/shoubin/datasets/scannet_feat/'\n # self.pc_feat_root = pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n # self.voxel_root = pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n \n\n out = {\n \"qa_input\": qa_input,\n \"qa_output\": answer,\n \"scene_id\": self.scene_ids[ann[\"scene_id\"]],\n \"question_id\": question_id,\n }\n\n for modality in self.modalities:\n \n if modality == 'pc':\n pc_feat = torch.load(os.path.join(self.pc_feat_root, f\"{scene_id}.pt\"), map_location=\"cpu\") # [N, 1408]\n if isinstance(pc_feat, np.ndarray):\n pc_feat = torch.tensor(pc_feat).float()\n pc = np.load(os.path.join(self.voxel_root, f\"{scene_id}.npy\"))\n pc = torch.tensor(pc).float().cpu()\n # sample 10000 points: [N, 1408] -> [10000, 1408]\n if pc_feat.shape[0] > 5000:\n idxes = torch.sort(torch.randperm(pc_feat.shape[0])[:5000])[1]\n pc_feat = pc_feat[idxes]\n pc = pc[idxes]\n else:\n pc_feat = torch.cat([pc_feat, torch.zeros(5000 - pc_feat.shape[0], 1408)], dim=0)\n pc = torch.cat([pc, torch.zeros(5000 - pc.shape[0], 3)], dim=0)\n\n out[\"pc_feat\"] = pc_feat\n out['pc'] = pc\n if modality == 'video':\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.__init__#L158-L200","kind":"function","name":"__init__","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":158,"end_line":200,"context_start_line":138,"context_end_line":220,"code":" clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)\n\n\nclass ThreeDVQAEvalDataset(BaseDataset):\n def __init__(self, **kwargs):\n super().__init__(kwargs['vis_processor'], kwargs['text_processor'], kwargs['vis_root'], kwargs['ann_paths'], kwargs['modalities'])\n\n self.modalities = kwargs['modalities']\n\n self.scene_ids = {}\n n = 0\n new_annotation = []\n for ann in self.annotation:\n try:\n img_id = ann[\"scene_id\"]\n if img_id not in self.scene_ids.keys():\n self.scene_ids[img_id] = n\n n += 1\n new_annotation.append(ann)\n except:\n pass\n self.annotation = new_annotation\n \n for modality in self.modalities:\n if 'pc' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n self.pc_feat_root = self.pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n self.voxel_root = self.pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n # pc_root = '/nas-ssd2/shoubin/datasets/scannet_feat/'\n # self.pc_feat_root = pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n # self.voxel_root = pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.get_existing_video_annotations","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.get_existing_video_annotations#L202-L203","kind":"function","name":"get_existing_video_annotations","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":202,"end_line":203,"context_start_line":182,"context_end_line":223,"code":" # pc_root = '/nas-ssd2/shoubin/datasets/scannet_feat/'\n # self.pc_feat_root = pc_root + '/voxelized_features_sam_nonzero_preprocess/'\n # self.voxel_root = pc_root + '/voxelized_voxels_sam_nonzero_preprocess/'\n self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.get_video_path","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.get_video_path#L205-L206","kind":"function","name":"get_video_path","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":205,"end_line":206,"context_start_line":185,"context_end_line":226,"code":" self.annotation = [\n ann for ann in self.annotation if os.path.exists(os.path.join(self.pc_feat_root, str(ann[\"scene_id\"]) + \".pt\"))\n ]\n if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.get_frame_path","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.get_frame_path#L208-L209","kind":"function","name":"get_frame_path","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":208,"end_line":209,"context_start_line":188,"context_end_line":229,"code":" if 'video' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.get_depth_path","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.get_depth_path#L211-L212","kind":"function","name":"get_depth_path","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":211,"end_line":212,"context_start_line":191,"context_end_line":232,"code":" setattr(self, f\"existing_{modality}_annotation\",getattr(self, f'get_existing_{modality}_annotations')())\n\n if 'frame' in modality:\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n \n","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.__getitem__#L214-L279","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":214,"end_line":279,"context_start_line":194,"context_end_line":282,"code":" setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n\n if 'depth' in modality:\n # todo\n setattr(self, f\"{modality}_root\", kwargs[f\"{modality}_root\"])\n setattr(self, f\"{modality}_processor\", kwargs[f\"{modality}_processor\"])\n \n def get_existing_video_annotations(self):\n return [f.split('.')[0] for f in os.listdir(self.video_root)]\n\n def get_video_path(self, ann):\n return os.path.join(self.video_root, f'{ann[\"scene_id\"]}.mp4')\n \n def get_frame_path(self, ann):\n return os.path.join(self.frame_root, f'{ann[\"scene_id\"]}/')\n \n def get_depth_path(self, ann):\n return os.path.join(self.depth_root, f'{ann[\"scene_id\"]}/')\n\n def __getitem__(self, index):\n\n ann = copy.deepcopy(self.annotation[index])\n if 'question_id' in ann.keys(): # 3dqa data\n qa_input = self.text_processor(ann['situation']) + '. Question: ' + self.text_processor(ann[\"question\"]) + ' Based on the frames and 3D Model information, answer the question using a single word or phase.'\n qtype = get_qclass(ann['question'])\n question_id = ann['question_id']\n answer = ann[\"answers\"][0]\n question_id = qtype + '_' + str(question_id)\n else: # pre-training data\n # ann['qa_input'] = 'Question: ' + self.text_processor(question) + ' Based on the frames information, answer the question using a single word or phase.'\n qa_input = 'Question: ' + self.text_processor(ann[\"question\"]) + ' Based on 3D Model information, answer the question using a single word or phase.'\n answer = ann[\"answers\"][0]\n answer = self.text_processor(answer)\n question_id = str(ann[\"scene_id\"])\n\n scene_id = str(ann[\"scene_id\"])\n \n\n out = {\n \"qa_input\": qa_input,\n \"qa_output\": answer,\n \"scene_id\": self.scene_ids[ann[\"scene_id\"]],\n \"question_id\": question_id,\n }\n\n for modality in self.modalities:\n \n if modality == 'pc':\n pc_feat = torch.load(os.path.join(self.pc_feat_root, f\"{scene_id}.pt\"), map_location=\"cpu\") # [N, 1408]\n if isinstance(pc_feat, np.ndarray):\n pc_feat = torch.tensor(pc_feat).float()\n pc = np.load(os.path.join(self.voxel_root, f\"{scene_id}.npy\"))\n pc = torch.tensor(pc).float().cpu()\n # sample 10000 points: [N, 1408] -> [10000, 1408]\n if pc_feat.shape[0] > 5000:\n idxes = torch.sort(torch.randperm(pc_feat.shape[0])[:5000])[1]\n pc_feat = pc_feat[idxes]\n pc = pc[idxes]\n else:\n pc_feat = torch.cat([pc_feat, torch.zeros(5000 - pc_feat.shape[0], 1408)], dim=0)\n pc = torch.cat([pc, torch.zeros(5000 - pc.shape[0], 3)], dim=0)\n\n out[\"pc_feat\"] = pc_feat\n out['pc'] = pc\n if modality == 'video':\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.threedvqa_datasets.__len__","uri":"program://CREMA/function/lavis.datasets.datasets.threedvqa_datasets.__len__#L281-L282","kind":"function","name":"__len__","path":"lavis/datasets/datasets/threedvqa_datasets.py","language":"python","start_line":281,"end_line":282,"context_start_line":261,"context_end_line":282,"code":" rgb, indices, fps = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"])\n out['rgb'] = rgb.to(torch.float32)\n \n if modality == 'frame':\n indices = None\n clip = None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n frms, indices = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n assert len(rgb) == getattr(self, f\"{modality}_processor\").n_frms\n out['rgb'] = rgb\n\n if modality == 'depth':\n assert indices is not None\n ann[f\"{modality}_path\"] = getattr(self, f\"get_{modality}_path\")(ann)\n depth, _ = getattr(self, f\"{modality}_processor\")(ann[f\"{modality}_path\"], clip_proposal=clip, indices=indices, type='depth')\n out['depth'] = depth\n\n return out\n\n def __len__(self):\n return len(self.annotation)","source_hash":"ef1e65ebb5b923b01f4ebfe20b68dd16b523a4d56c5f85b57505e6d34bad910f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.rgbd_vqa_datasets#L1-L130","kind":"module","name":"lavis.datasets.datasets.rgbd_vqa_datasets","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":1,"end_line":130,"context_start_line":1,"context_end_line":130,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport torch\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n\n if 'start' in ann:\n start, end = float(ann['start']), float(ann['end'])\n clip = [start, end]\n else:\n clip = None \n\n # for QA\n prompt = 'Question: ' + q\n for j in range(ann['num_option']):\n a = ann['a{}'.format(j)]\n prompt += ' Option {}: '.format(ANS_MAPPING[j])\n prompt += a\n qa_prompt = prompt + ' Considering the information presented in the frame, select the correct answer from the options.'\n # loc_prompt = 'Question: ' + q + ' ' + hints + ' Does the information within the frame provide the necessary details to accurately answer the given question?' \n answers = 'Option ' + ANS_MAPPING[int(ann['answer'])]\n duration = 1\n\n # print(self.modality_type)\n indices = None\n \n if 'rgb' in self.modality_type:\n vpath = os.path.join(self.vis_root, str(ann['video']))\n frms, indices = self.vis_processor(vpath, clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n # print(indices)\n assert len(rgb) == self.vis_processor.n_frms\n out['rgb'] = rgb\n \n if 'depth' in self.modality_type:\n depth_root = self.vis_root[:-1] + '_depth/'\n # if 'nas-hdd' in depth_root:\n # depth_root = depth_root.replace('nas-hdd','nas-ssd2')\n depth_path = os.path.join(depth_root, str(ann['video']))\n depth, indices_ = self.vis_processor(depth_path, clip_proposal=clip, indices=indices, type='depth')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n \n out['depth'] = depth\n \n if 'flow' in self.modality_type:\n flow_root = self.vis_root[:-1] + '_flow/'\n # if 'nas-hdd' in flow_root:\n # flow_root = flow_root.replace('nas-hdd','nas-ssd2')\n flow_path = os.path.join(flow_root, str(ann['video']))\n try: \n flow, indices_ = self.vis_processor(flow_path, clip_proposal=clip, indices=indices, type='flow')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['flow'] = flow\n result = True\n flow_flag = False\n except Exception as e:\n print(f\"Error while read flow file idx\")\n print(\"video is: {}\".format(ann['video']))\n index = random.randint(0, len(self.annotation) - 1)\n flow_flag = True \n \n if 'norm' in self.modality_type:\n norm_root = self.vis_root[:-1] + '_norm/'\n norm_path = os.path.join(norm_root, str(ann['video']))\n norm, indices_ = self.vis_processor(norm_path, clip_proposal=clip, indices=indices, type='norm')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['norm'] = norm\n\n if not flow_flag:\n result = True\n \n out['qa_input'] = qa_prompt\n out['qa_output'] = answers\n out['question_id'] = qid\n out['duration'] = duration\n \n return out","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.rgbd_vqa_datasets.__DisplMixin#L18-L26","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":18,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport torch\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets.MCVideoQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.rgbd_vqa_datasets.MCVideoQADataset#L31-L130","kind":"class","name":"MCVideoQADataset","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":31,"end_line":130,"context_start_line":11,"context_end_line":130,"code":"from collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n\n if 'start' in ann:\n start, end = float(ann['start']), float(ann['end'])\n clip = [start, end]\n else:\n clip = None \n\n # for QA\n prompt = 'Question: ' + q\n for j in range(ann['num_option']):\n a = ann['a{}'.format(j)]\n prompt += ' Option {}: '.format(ANS_MAPPING[j])\n prompt += a\n qa_prompt = prompt + ' Considering the information presented in the frame, select the correct answer from the options.'\n # loc_prompt = 'Question: ' + q + ' ' + hints + ' Does the information within the frame provide the necessary details to accurately answer the given question?' \n answers = 'Option ' + ANS_MAPPING[int(ann['answer'])]\n duration = 1\n\n # print(self.modality_type)\n indices = None\n \n if 'rgb' in self.modality_type:\n vpath = os.path.join(self.vis_root, str(ann['video']))\n frms, indices = self.vis_processor(vpath, clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n # print(indices)\n assert len(rgb) == self.vis_processor.n_frms\n out['rgb'] = rgb\n \n if 'depth' in self.modality_type:\n depth_root = self.vis_root[:-1] + '_depth/'\n # if 'nas-hdd' in depth_root:\n # depth_root = depth_root.replace('nas-hdd','nas-ssd2')\n depth_path = os.path.join(depth_root, str(ann['video']))\n depth, indices_ = self.vis_processor(depth_path, clip_proposal=clip, indices=indices, type='depth')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n \n out['depth'] = depth\n \n if 'flow' in self.modality_type:\n flow_root = self.vis_root[:-1] + '_flow/'\n # if 'nas-hdd' in flow_root:\n # flow_root = flow_root.replace('nas-hdd','nas-ssd2')\n flow_path = os.path.join(flow_root, str(ann['video']))\n try: \n flow, indices_ = self.vis_processor(flow_path, clip_proposal=clip, indices=indices, type='flow')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['flow'] = flow\n result = True\n flow_flag = False\n except Exception as e:\n print(f\"Error while read flow file idx\")\n print(\"video is: {}\".format(ann['video']))\n index = random.randint(0, len(self.annotation) - 1)\n flow_flag = True \n \n if 'norm' in self.modality_type:\n norm_root = self.vis_root[:-1] + '_norm/'\n norm_path = os.path.join(norm_root, str(ann['video']))\n norm, indices_ = self.vis_processor(norm_path, clip_proposal=clip, indices=indices, type='norm')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['norm'] = norm\n\n if not flow_flag:\n result = True\n \n out['qa_input'] = qa_prompt\n out['qa_output'] = answers\n out['question_id'] = qid\n out['duration'] = duration\n \n return out","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.rgbd_vqa_datasets.displ_item#L19-L26","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":19,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport torch\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.rgbd_vqa_datasets.__init__#L32-L33","kind":"function","name":"__init__","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":32,"end_line":33,"context_start_line":12,"context_end_line":53,"code":"\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets._load_auxiliary_mappings","uri":"program://CREMA/function/lavis.datasets.datasets.rgbd_vqa_datasets._load_auxiliary_mappings#L35-L36","kind":"function","name":"_load_auxiliary_mappings","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":35,"end_line":36,"context_start_line":15,"context_end_line":56,"code":")\nimport random\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n\n if 'start' in ann:\n start, end = float(ann['start']), float(ann['end'])\n clip = [start, end]","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets._get_answer_label","uri":"program://CREMA/function/lavis.datasets.datasets.rgbd_vqa_datasets._get_answer_label#L38-L42","kind":"function","name":"_get_answer_label","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":38,"end_line":42,"context_start_line":18,"context_end_line":62,"code":"class __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n\n if 'start' in ann:\n start, end = float(ann['start']), float(ann['end'])\n clip = [start, end]\n else:\n clip = None \n\n # for QA\n prompt = 'Question: ' + q\n for j in range(ann['num_option']):","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.rgbd_vqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.rgbd_vqa_datasets.__getitem__#L44-L130","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/rgbd_vqa_datasets.py","language":"python","start_line":44,"end_line":130,"context_start_line":24,"context_end_line":130,"code":" return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\nANS_MAPPING = {0:'A',1:'B',2:'C',3:'D',4:'E'}\n\n# NextQA\nclass MCVideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n def _load_auxiliary_mappings(self):\n pass\n \n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n \n result, flow_flag = None, False\n out = {}\n\n while result is None:\n ann = self.annotation[index]\n qid = ann['qid'] \n q = ann['question']\n\n if 'start' in ann:\n start, end = float(ann['start']), float(ann['end'])\n clip = [start, end]\n else:\n clip = None \n\n # for QA\n prompt = 'Question: ' + q\n for j in range(ann['num_option']):\n a = ann['a{}'.format(j)]\n prompt += ' Option {}: '.format(ANS_MAPPING[j])\n prompt += a\n qa_prompt = prompt + ' Considering the information presented in the frame, select the correct answer from the options.'\n # loc_prompt = 'Question: ' + q + ' ' + hints + ' Does the information within the frame provide the necessary details to accurately answer the given question?' \n answers = 'Option ' + ANS_MAPPING[int(ann['answer'])]\n duration = 1\n\n # print(self.modality_type)\n indices = None\n \n if 'rgb' in self.modality_type:\n vpath = os.path.join(self.vis_root, str(ann['video']))\n frms, indices = self.vis_processor(vpath, clip_proposal=clip, indices=indices)\n rgb = frms.permute(1, 0, 2, 3)\n # print(indices)\n assert len(rgb) == self.vis_processor.n_frms\n out['rgb'] = rgb\n \n if 'depth' in self.modality_type:\n depth_root = self.vis_root[:-1] + '_depth/'\n # if 'nas-hdd' in depth_root:\n # depth_root = depth_root.replace('nas-hdd','nas-ssd2')\n depth_path = os.path.join(depth_root, str(ann['video']))\n depth, indices_ = self.vis_processor(depth_path, clip_proposal=clip, indices=indices, type='depth')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n \n out['depth'] = depth\n \n if 'flow' in self.modality_type:\n flow_root = self.vis_root[:-1] + '_flow/'\n # if 'nas-hdd' in flow_root:\n # flow_root = flow_root.replace('nas-hdd','nas-ssd2')\n flow_path = os.path.join(flow_root, str(ann['video']))\n try: \n flow, indices_ = self.vis_processor(flow_path, clip_proposal=clip, indices=indices, type='flow')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['flow'] = flow\n result = True\n flow_flag = False\n except Exception as e:\n print(f\"Error while read flow file idx\")\n print(\"video is: {}\".format(ann['video']))\n index = random.randint(0, len(self.annotation) - 1)\n flow_flag = True \n \n if 'norm' in self.modality_type:\n norm_root = self.vis_root[:-1] + '_norm/'\n norm_path = os.path.join(norm_root, str(ann['video']))\n norm, indices_ = self.vis_processor(norm_path, clip_proposal=clip, indices=indices, type='norm')\n if indices is not None:\n assert indices == indices_\n indices = indices_\n out['norm'] = norm\n\n if not flow_flag:\n result = True\n \n out['qa_input'] = qa_prompt\n out['qa_output'] = answers\n out['question_id'] = qid\n out['duration'] = duration\n \n return out","source_hash":"8d07ac6597dd46c979309f48f8be53f47fcc6dbea6332bea67d3672bcdd60c70","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.coco_vqa_datasets#L1-L107","kind":"module","name":"lavis.datasets.datasets.coco_vqa_datasets","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":1,"end_line":107,"context_start_line":1,"context_end_line":107,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass COCOVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:\n answer_weight[answer] = 1 / len(ann[\"answer\"])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass COCOVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.coco_vqa_datasets.__DisplMixin#L18-L30","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":18,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass COCOVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.COCOVQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.coco_vqa_datasets.COCOVQADataset#L33-L61","kind":"class","name":"COCOVQADataset","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":33,"end_line":61,"context_start_line":13,"context_end_line":81,"code":"from lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass COCOVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:\n answer_weight[answer] = 1 / len(ann[\"answer\"])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass COCOVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.COCOVQAEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.coco_vqa_datasets.COCOVQAEvalDataset#L64-L107","kind":"class","name":"COCOVQAEvalDataset","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":64,"end_line":107,"context_start_line":44,"context_end_line":107,"code":" question = self.text_processor(ann[\"question\"])\n\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:\n answer_weight[answer] = 1 / len(ann[\"answer\"])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass COCOVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.coco_vqa_datasets.displ_item#L19-L30","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":19,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass COCOVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.coco_vqa_datasets.__init__#L65-L91","kind":"function","name":"__init__","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":65,"end_line":91,"context_start_line":45,"context_end_line":107,"code":"\n answer_weight = {}\n for answer in ann[\"answer\"]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[\"answer\"])\n else:\n answer_weight[answer] = 1 / len(ann[\"answer\"])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass COCOVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.coco_vqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.coco_vqa_datasets.__getitem__#L93-L107","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/coco_vqa_datasets.py","language":"python","start_line":93,"end_line":107,"context_start_line":73,"context_end_line":107,"code":" self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"827e687b4f366bab92a3f5b0ea83bceb8f3fa920ef0c2c1ec873dd8e9a26acc5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.gqa_datasets#L1-L101","kind":"module","name":"lavis.datasets.datasets.gqa_datasets","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":1,"end_line":101,"context_start_line":1,"context_end_line":101,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass GQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass GQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. gqa/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n ## TODO: support inference method == 'ranking'\n answer_list_path = ann_paths[1] if len(ann_paths) > 1 else ''\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n if \"answer\" in ann:\n # answer is a string\n answer = ann[\"answer\"]\n else:\n answer = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answer\": answer,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.gqa_datasets.__DisplMixin#L18-L30","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":18,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass GQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.GQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.gqa_datasets.GQADataset#L33-L54","kind":"class","name":"GQADataset","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":33,"end_line":54,"context_start_line":13,"context_end_line":74,"code":"from lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass GQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass GQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. gqa/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n ## TODO: support inference method == 'ranking'\n answer_list_path = ann_paths[1] if len(ann_paths) > 1 else ''\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.GQAEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.gqa_datasets.GQAEvalDataset#L57-L101","kind":"class","name":"GQAEvalDataset","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":57,"end_line":101,"context_start_line":37,"context_end_line":101,"code":" def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass GQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. gqa/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n ## TODO: support inference method == 'ranking'\n answer_list_path = ann_paths[1] if len(ann_paths) > 1 else ''\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n if \"answer\" in ann:\n # answer is a string\n answer = ann[\"answer\"]\n else:\n answer = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answer\": answer,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.gqa_datasets.displ_item#L19-L30","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":19,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nimport json\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\nfrom collections import OrderedDict\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"answers\": \"; \".join(ann[\"answer\"]),\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass GQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.gqa_datasets.__init__#L58-L78","kind":"function","name":"__init__","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":58,"end_line":78,"context_start_line":38,"context_end_line":98,"code":" ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answers = [ann[\"answer\"]]\n weights = [1]\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass GQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. gqa/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n ## TODO: support inference method == 'ranking'\n answer_list_path = ann_paths[1] if len(ann_paths) > 1 else ''\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n if \"answer\" in ann:\n # answer is a string\n answer = ann[\"answer\"]\n else:\n answer = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answer\": answer,","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.gqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.gqa_datasets.__getitem__#L80-L101","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/gqa_datasets.py","language":"python","start_line":80,"end_line":101,"context_start_line":60,"context_end_line":101,"code":" vis_root (string): Root directory of images (e.g. gqa/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n ## TODO: support inference method == 'ranking'\n answer_list_path = ann_paths[1] if len(ann_paths) > 1 else ''\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n if \"answer\" in ann:\n # answer is a string\n answer = ann[\"answer\"]\n else:\n answer = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answer\": answer,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"b70361d7242360fe71fcc5186ca6d897a012420780122fa88a5da289370d86c4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.video_vqa_datasets#L1-L62","kind":"module","name":"lavis.datasets.datasets.video_vqa_datasets","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":1,"end_line":62,"context_start_line":1,"context_end_line":62,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n frms = self.vis_processor(vpath)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"video\": frms,\n \"text_input\": question,\n \"answers\": self._get_answer_label(ann[\"answer\"]),\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.video_vqa_datasets.__DisplMixin#L17-L25","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":17,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets.VideoQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.video_vqa_datasets.VideoQADataset#L28-L62","kind":"class","name":"VideoQADataset","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":28,"end_line":62,"context_start_line":8,"context_end_line":62,"code":"import json\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n frms = self.vis_processor(vpath)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"video\": frms,\n \"text_input\": question,\n \"answers\": self._get_answer_label(ann[\"answer\"]),\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.video_vqa_datasets.displ_item#L18-L25","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":18,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.video_vqa_datasets.__init__#L29-L30","kind":"function","name":"__init__","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":29,"end_line":30,"context_start_line":9,"context_end_line":50,"code":"import os\nfrom collections import OrderedDict\n\nfrom lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets._build_class_labels","uri":"program://CREMA/function/lavis.datasets.datasets.video_vqa_datasets._build_class_labels#L32-L35","kind":"function","name":"_build_class_labels","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":32,"end_line":35,"context_start_line":12,"context_end_line":55,"code":"from lavis.datasets.datasets.multimodal_classification_datasets import (\n MultimodalClassificationDataset,\n)\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n frms = self.vis_processor(vpath)\n question = self.text_processor(ann[\"question\"])\n","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets._get_answer_label","uri":"program://CREMA/function/lavis.datasets.datasets.video_vqa_datasets._get_answer_label#L37-L41","kind":"function","name":"_get_answer_label","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":37,"end_line":41,"context_start_line":17,"context_end_line":61,"code":"class __DisplMixin:\n def displ_item(self, index):\n ann = self.annotation[index]\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n frms = self.vis_processor(vpath)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"video\": frms,\n \"text_input\": question,\n \"answers\": self._get_answer_label(ann[\"answer\"]),\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.video_vqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.video_vqa_datasets.__getitem__#L43-L62","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/video_vqa_datasets.py","language":"python","start_line":43,"end_line":62,"context_start_line":23,"context_end_line":62,"code":" return OrderedDict(\n {\"file\": vpath, \"question\": ann[\"question\"], \"answer\": ann[\"answer\"]}\n )\n\n\nclass VideoQADataset(MultimodalClassificationDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def _build_class_labels(self, ans_path):\n ans2label = json.load(open(ans_path))\n\n self.class_labels = ans2label\n\n def _get_answer_label(self, answer):\n if answer in self.class_labels:\n return self.class_labels[answer]\n else:\n return len(self.class_labels)\n\n def __getitem__(self, index):\n assert (\n self.class_labels\n ), f\"class_labels of {__class__.__name__} is not built yet.\"\n\n ann = self.annotation[index]\n\n vname = ann[\"video\"]\n vpath = os.path.join(self.vis_root, vname)\n\n frms = self.vis_processor(vpath)\n question = self.text_processor(ann[\"question\"])\n\n return {\n \"video\": frms,\n \"text_input\": question,\n \"answers\": self._get_answer_label(ann[\"answer\"]),\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n }","source_hash":"5f3b7a0201e218f74a5f584a300f5e47e5e8eb558ba0a8f91d937588c7d2cc4e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.multimodal_classification_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.multimodal_classification_datasets#L1-L25","kind":"module","name":"lavis.datasets.datasets.multimodal_classification_datasets","path":"lavis/datasets/datasets/multimodal_classification_datasets.py","language":"python","start_line":1,"end_line":25,"context_start_line":1,"context_end_line":25,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom abc import abstractmethod\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass MultimodalClassificationDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n self.class_labels = None\n\n @abstractmethod\n def _build_class_labels(self):\n pass\n\n @abstractmethod\n def _load_auxiliary_mappings(self):\n pass\n","source_hash":"51ea7c61e9449050f8ae0b06043ea2fac149b91aa07e362dfdb393a4b0656ab8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.multimodal_classification_datasets.MultimodalClassificationDataset","uri":"program://CREMA/class/lavis.datasets.datasets.multimodal_classification_datasets.MultimodalClassificationDataset#L12-L24","kind":"class","name":"MultimodalClassificationDataset","path":"lavis/datasets/datasets/multimodal_classification_datasets.py","language":"python","start_line":12,"end_line":24,"context_start_line":1,"context_end_line":25,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom abc import abstractmethod\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass MultimodalClassificationDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n self.class_labels = None\n\n @abstractmethod\n def _build_class_labels(self):\n pass\n\n @abstractmethod\n def _load_auxiliary_mappings(self):\n pass\n","source_hash":"51ea7c61e9449050f8ae0b06043ea2fac149b91aa07e362dfdb393a4b0656ab8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.multimodal_classification_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.multimodal_classification_datasets.__init__#L13-L16","kind":"function","name":"__init__","path":"lavis/datasets/datasets/multimodal_classification_datasets.py","language":"python","start_line":13,"end_line":16,"context_start_line":1,"context_end_line":25,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom abc import abstractmethod\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass MultimodalClassificationDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n self.class_labels = None\n\n @abstractmethod\n def _build_class_labels(self):\n pass\n\n @abstractmethod\n def _load_auxiliary_mappings(self):\n pass\n","source_hash":"51ea7c61e9449050f8ae0b06043ea2fac149b91aa07e362dfdb393a4b0656ab8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.multimodal_classification_datasets._build_class_labels","uri":"program://CREMA/function/lavis.datasets.datasets.multimodal_classification_datasets._build_class_labels#L19-L20","kind":"function","name":"_build_class_labels","path":"lavis/datasets/datasets/multimodal_classification_datasets.py","language":"python","start_line":19,"end_line":20,"context_start_line":1,"context_end_line":25,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom abc import abstractmethod\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass MultimodalClassificationDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n self.class_labels = None\n\n @abstractmethod\n def _build_class_labels(self):\n pass\n\n @abstractmethod\n def _load_auxiliary_mappings(self):\n pass\n","source_hash":"51ea7c61e9449050f8ae0b06043ea2fac149b91aa07e362dfdb393a4b0656ab8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.multimodal_classification_datasets._load_auxiliary_mappings","uri":"program://CREMA/function/lavis.datasets.datasets.multimodal_classification_datasets._load_auxiliary_mappings#L23-L24","kind":"function","name":"_load_auxiliary_mappings","path":"lavis/datasets/datasets/multimodal_classification_datasets.py","language":"python","start_line":23,"end_line":24,"context_start_line":3,"context_end_line":25,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom abc import abstractmethod\nfrom lavis.datasets.datasets.base_dataset import BaseDataset\n\n\nclass MultimodalClassificationDataset(BaseDataset):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths, modality_type):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths, modality_type)\n\n self.class_labels = None\n\n @abstractmethod\n def _build_class_labels(self):\n pass\n\n @abstractmethod\n def _load_auxiliary_mappings(self):\n pass\n","source_hash":"51ea7c61e9449050f8ae0b06043ea2fac149b91aa07e362dfdb393a4b0656ab8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets","uri":"program://CREMA/module/lavis.datasets.datasets.aok_vqa_datasets#L1-L154","kind":"module","name":"lavis.datasets.datasets.aok_vqa_datasets","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":1,"end_line":154,"context_start_line":1,"context_end_line":154,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom collections import OrderedDict\nimport json\nimport os\nimport torch\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"direct_answers\": \"; \".join(ann[\"direct_answers\"]),\n \"choices\": \"; \".join(ann[\"choices\"]),\n \"correct_choice\": ann[\"choices\"][ann[\"correct_choice_idx\"]],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass AOKVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_key = \"direct_answers\"\n\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[answer_key])\n else:\n answer_weight[answer] = 1 / len(ann[answer_key])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass AOKVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def collater(self, samples):\n (\n image_list,\n question_list,\n question_id_list,\n instance_id_list,\n choices_list,\n correct_choice_idx_list,\n direct_answers_list,\n ) = ([], [], [], [], [], [], [])\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n question_id_list.append(sample[\"question_id\"])\n instance_id_list.append(sample[\"instance_id\"])\n choices_list.append(sample[\"choices\"])\n correct_choice_idx_list.append(sample[\"correct_choice_idx\"])\n direct_answers_list.append(sample[\"direct_answers\"])\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"question_id\": question_id_list,\n \"instance_id\": instance_id_list,\n \"choices\": choices_list,\n \"correct_choice_idx\": correct_choice_idx_list,\n \"direct_answers\": direct_answers_list,\n }\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n choices = ann[\"choices\"]\n if \"correct_choice_idx\" in ann:\n correct_choice_idx = ann[\"correct_choice_idx\"]\n else:\n correct_choice_idx = None\n\n if \"direct_answers\" in ann:\n direct_answers = ann[\"direct_answers\"]\n else:\n direct_answers = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n \"choices\": choices,\n \"correct_choice_idx\": correct_choice_idx,\n \"direct_answers\": direct_answers,\n }","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.__DisplMixin","uri":"program://CREMA/class/lavis.datasets.datasets.aok_vqa_datasets.__DisplMixin#L18-L31","kind":"class","name":"__DisplMixin","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":18,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom collections import OrderedDict\nimport json\nimport os\nimport torch\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"direct_answers\": \"; \".join(ann[\"direct_answers\"]),\n \"choices\": \"; \".join(ann[\"choices\"]),\n \"correct_choice\": ann[\"choices\"][ann[\"correct_choice_idx\"]],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass AOKVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_key = \"direct_answers\"\n\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.AOKVQADataset","uri":"program://CREMA/class/lavis.datasets.datasets.aok_vqa_datasets.AOKVQADataset#L34-L64","kind":"class","name":"AOKVQADataset","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":34,"end_line":64,"context_start_line":14,"context_end_line":84,"code":"\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"direct_answers\": \"; \".join(ann[\"direct_answers\"]),\n \"choices\": \"; \".join(ann[\"choices\"]),\n \"correct_choice\": ann[\"choices\"][ann[\"correct_choice_idx\"]],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass AOKVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_key = \"direct_answers\"\n\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[answer_key])\n else:\n answer_weight[answer] = 1 / len(ann[answer_key])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass AOKVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.AOKVQAEvalDataset","uri":"program://CREMA/class/lavis.datasets.datasets.aok_vqa_datasets.AOKVQAEvalDataset#L67-L154","kind":"class","name":"AOKVQAEvalDataset","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":67,"end_line":154,"context_start_line":47,"context_end_line":154,"code":" answer_key = \"direct_answers\"\n\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[answer_key])\n else:\n answer_weight[answer] = 1 / len(ann[answer_key])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass AOKVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def collater(self, samples):\n (\n image_list,\n question_list,\n question_id_list,\n instance_id_list,\n choices_list,\n correct_choice_idx_list,\n direct_answers_list,\n ) = ([], [], [], [], [], [], [])\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n question_id_list.append(sample[\"question_id\"])\n instance_id_list.append(sample[\"instance_id\"])\n choices_list.append(sample[\"choices\"])\n correct_choice_idx_list.append(sample[\"correct_choice_idx\"])\n direct_answers_list.append(sample[\"direct_answers\"])\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"question_id\": question_id_list,\n \"instance_id\": instance_id_list,\n \"choices\": choices_list,\n \"correct_choice_idx\": correct_choice_idx_list,\n \"direct_answers\": direct_answers_list,\n }\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n choices = ann[\"choices\"]\n if \"correct_choice_idx\" in ann:\n correct_choice_idx = ann[\"correct_choice_idx\"]\n else:\n correct_choice_idx = None\n\n if \"direct_answers\" in ann:\n direct_answers = ann[\"direct_answers\"]\n else:\n direct_answers = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n \"choices\": choices,\n \"correct_choice_idx\": correct_choice_idx,\n \"direct_answers\": direct_answers,\n }","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.displ_item","uri":"program://CREMA/function/lavis.datasets.datasets.aok_vqa_datasets.displ_item#L19-L31","kind":"function","name":"displ_item","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":19,"end_line":31,"context_start_line":1,"context_end_line":51,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom collections import OrderedDict\nimport json\nimport os\nimport torch\n\nfrom PIL import Image\n\nfrom lavis.datasets.datasets.vqa_datasets import VQADataset, VQAEvalDataset\n\n\nclass __DisplMixin:\n def displ_item(self, index):\n sample, ann = self.__getitem__(index), self.annotation[index]\n return OrderedDict(\n {\n \"file\": ann[\"image\"],\n \"question\": ann[\"question\"],\n \"question_id\": ann[\"question_id\"],\n \"direct_answers\": \"; \".join(ann[\"direct_answers\"]),\n \"choices\": \"; \".join(ann[\"choices\"]),\n \"correct_choice\": ann[\"choices\"][ann[\"correct_choice_idx\"]],\n \"image\": sample[\"image\"],\n }\n )\n\n\nclass AOKVQADataset(VQADataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n super().__init__(vis_processor, text_processor, vis_root, ann_paths)\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n answer_key = \"direct_answers\"\n\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.__init__","uri":"program://CREMA/function/lavis.datasets.datasets.aok_vqa_datasets.__init__#L68-L94","kind":"function","name":"__init__","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":68,"end_line":94,"context_start_line":48,"context_end_line":114,"code":"\n answer_weight = {}\n for answer in ann[answer_key]:\n if answer in answer_weight.keys():\n answer_weight[answer] += 1 / len(ann[answer_key])\n else:\n answer_weight[answer] = 1 / len(ann[answer_key])\n\n answers = list(answer_weight.keys())\n weights = list(answer_weight.values())\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"answers\": answers,\n \"weights\": weights,\n }\n\n\nclass AOKVQAEvalDataset(VQAEvalDataset, __DisplMixin):\n def __init__(self, vis_processor, text_processor, vis_root, ann_paths):\n \"\"\"\n vis_root (string): Root directory of images (e.g. coco/images/)\n ann_root (string): directory to store the annotation file\n \"\"\"\n\n self.vis_root = vis_root\n\n self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def collater(self, samples):\n (\n image_list,\n question_list,\n question_id_list,\n instance_id_list,\n choices_list,\n correct_choice_idx_list,\n direct_answers_list,\n ) = ([], [], [], [], [], [], [])\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n question_id_list.append(sample[\"question_id\"])\n instance_id_list.append(sample[\"instance_id\"])\n choices_list.append(sample[\"choices\"])\n correct_choice_idx_list.append(sample[\"correct_choice_idx\"])\n direct_answers_list.append(sample[\"direct_answers\"])","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.__getitem__","uri":"program://CREMA/function/lavis.datasets.datasets.aok_vqa_datasets.__getitem__#L126-L154","kind":"function","name":"__getitem__","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":126,"end_line":154,"context_start_line":106,"context_end_line":154,"code":"\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n question_id_list.append(sample[\"question_id\"])\n instance_id_list.append(sample[\"instance_id\"])\n choices_list.append(sample[\"choices\"])\n correct_choice_idx_list.append(sample[\"correct_choice_idx\"])\n direct_answers_list.append(sample[\"direct_answers\"])\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"question_id\": question_id_list,\n \"instance_id\": instance_id_list,\n \"choices\": choices_list,\n \"correct_choice_idx\": correct_choice_idx_list,\n \"direct_answers\": direct_answers_list,\n }\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n choices = ann[\"choices\"]\n if \"correct_choice_idx\" in ann:\n correct_choice_idx = ann[\"correct_choice_idx\"]\n else:\n correct_choice_idx = None\n\n if \"direct_answers\" in ann:\n direct_answers = ann[\"direct_answers\"]\n else:\n direct_answers = None\n\n return {\n \"image\": image,\n \"text_input\": question,\n \"question_id\": ann[\"question_id\"],\n \"instance_id\": ann[\"instance_id\"],\n \"choices\": choices,\n \"correct_choice_idx\": correct_choice_idx,\n \"direct_answers\": direct_answers,\n }","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.datasets.aok_vqa_datasets.collater","uri":"program://CREMA/function/lavis.datasets.datasets.aok_vqa_datasets.collater#L96-L124","kind":"function","name":"collater","path":"lavis/datasets/datasets/aok_vqa_datasets.py","language":"python","start_line":96,"end_line":124,"context_start_line":76,"context_end_line":144,"code":" self.annotation = json.load(open(ann_paths[0]))\n\n answer_list_path = ann_paths[1]\n if os.path.exists(answer_list_path):\n self.answer_list = json.load(open(answer_list_path))\n else:\n self.answer_list = None\n\n try:\n self.coco_fmt_qust_file = ann_paths[2]\n self.coco_fmt_anno_file = ann_paths[3]\n except IndexError:\n self.coco_fmt_qust_file = None\n self.coco_fmt_anno_file = None\n\n self.vis_processor = vis_processor\n self.text_processor = text_processor\n\n self._add_instance_ids()\n\n def collater(self, samples):\n (\n image_list,\n question_list,\n question_id_list,\n instance_id_list,\n choices_list,\n correct_choice_idx_list,\n direct_answers_list,\n ) = ([], [], [], [], [], [], [])\n\n for sample in samples:\n image_list.append(sample[\"image\"])\n question_list.append(sample[\"text_input\"])\n question_id_list.append(sample[\"question_id\"])\n instance_id_list.append(sample[\"instance_id\"])\n choices_list.append(sample[\"choices\"])\n correct_choice_idx_list.append(sample[\"correct_choice_idx\"])\n direct_answers_list.append(sample[\"direct_answers\"])\n\n return {\n \"image\": torch.stack(image_list, dim=0),\n \"text_input\": question_list,\n \"question_id\": question_id_list,\n \"instance_id\": instance_id_list,\n \"choices\": choices_list,\n \"correct_choice_idx\": correct_choice_idx_list,\n \"direct_answers\": direct_answers_list,\n }\n\n def __getitem__(self, index):\n ann = self.annotation[index]\n\n image_path = os.path.join(self.vis_root, ann[\"image\"])\n image = Image.open(image_path).convert(\"RGB\")\n\n image = self.vis_processor(image)\n question = self.text_processor(ann[\"question\"])\n\n choices = ann[\"choices\"]\n if \"correct_choice_idx\" in ann:\n correct_choice_idx = ann[\"correct_choice_idx\"]\n else:\n correct_choice_idx = None\n\n if \"direct_answers\" in ann:\n direct_answers = ann[\"direct_answers\"]\n else:\n direct_answers = None","source_hash":"6dd9ac0645366a016b350ef7c44eaf277bbe68afcf271ccc8e39246ae3b9e2a8","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.classification_builder","uri":"program://CREMA/module/lavis.datasets.builders.classification_builder#L1-L27","kind":"module","name":"lavis.datasets.builders.classification_builder","path":"lavis/datasets/builders/classification_builder.py","language":"python","start_line":1,"end_line":27,"context_start_line":1,"context_end_line":27,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.nlvr_datasets import NLVRDataset, NLVREvalDataset\nfrom lavis.datasets.datasets.snli_ve_datasets import SNLIVisualEntialmentDataset\n\n\n@registry.register_builder(\"nlvr\")\nclass NLVRBuilder(BaseDatasetBuilder):\n train_dataset_cls = NLVRDataset\n eval_dataset_cls = NLVREvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/nlvr/defaults.yaml\"}\n\n\n@registry.register_builder(\"snli_ve\")\nclass SNLIVisualEntailmentBuilder(BaseDatasetBuilder):\n train_dataset_cls = SNLIVisualEntialmentDataset\n eval_dataset_cls = SNLIVisualEntialmentDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/snli_ve/defaults.yaml\"}","source_hash":"07deb88cd820f300f8d9b8578338b56ba0ecddc0f4ed9ca881f99196feda0ad9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.classification_builder.NLVRBuilder","uri":"program://CREMA/class/lavis.datasets.builders.classification_builder.NLVRBuilder#L15-L19","kind":"class","name":"NLVRBuilder","path":"lavis/datasets/builders/classification_builder.py","language":"python","start_line":15,"end_line":19,"context_start_line":1,"context_end_line":27,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.nlvr_datasets import NLVRDataset, NLVREvalDataset\nfrom lavis.datasets.datasets.snli_ve_datasets import SNLIVisualEntialmentDataset\n\n\n@registry.register_builder(\"nlvr\")\nclass NLVRBuilder(BaseDatasetBuilder):\n train_dataset_cls = NLVRDataset\n eval_dataset_cls = NLVREvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/nlvr/defaults.yaml\"}\n\n\n@registry.register_builder(\"snli_ve\")\nclass SNLIVisualEntailmentBuilder(BaseDatasetBuilder):\n train_dataset_cls = SNLIVisualEntialmentDataset\n eval_dataset_cls = SNLIVisualEntialmentDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/snli_ve/defaults.yaml\"}","source_hash":"07deb88cd820f300f8d9b8578338b56ba0ecddc0f4ed9ca881f99196feda0ad9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.classification_builder.SNLIVisualEntailmentBuilder","uri":"program://CREMA/class/lavis.datasets.builders.classification_builder.SNLIVisualEntailmentBuilder#L23-L27","kind":"class","name":"SNLIVisualEntailmentBuilder","path":"lavis/datasets/builders/classification_builder.py","language":"python","start_line":23,"end_line":27,"context_start_line":3,"context_end_line":27,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.nlvr_datasets import NLVRDataset, NLVREvalDataset\nfrom lavis.datasets.datasets.snli_ve_datasets import SNLIVisualEntialmentDataset\n\n\n@registry.register_builder(\"nlvr\")\nclass NLVRBuilder(BaseDatasetBuilder):\n train_dataset_cls = NLVRDataset\n eval_dataset_cls = NLVREvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/nlvr/defaults.yaml\"}\n\n\n@registry.register_builder(\"snli_ve\")\nclass SNLIVisualEntailmentBuilder(BaseDatasetBuilder):\n train_dataset_cls = SNLIVisualEntialmentDataset\n eval_dataset_cls = SNLIVisualEntialmentDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/snli_ve/defaults.yaml\"}","source_hash":"07deb88cd820f300f8d9b8578338b56ba0ecddc0f4ed9ca881f99196feda0ad9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder","uri":"program://CREMA/module/lavis.datasets.builders.image_text_pair_builder#L1-L77","kind":"module","name":"lavis.datasets.builders.image_text_pair_builder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":1,"end_line":77,"context_start_line":1,"context_end_line":77,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom lavis.common.registry import registry\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.image_text_pair_datasets import ImageTextPairDataset\nfrom lavis.datasets.datasets.laion_dataset import LaionDataset\n\n\n@registry.register_builder(\"conceptual_caption_3m\")\nclass ConceptualCaption3MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_3m.yaml\"\n }\n\n\n@registry.register_builder(\"conceptual_caption_12m\")\nclass ConceptualCaption12MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()\n split = \"train\" # laion dataset only has train split\n\n # create datasets\n # [NOTE] return inner_datasets (wds.DataPipeline)\n dataset_cls = self.train_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=self.vis_processors[split],\n text_processor=self.text_processors[split],\n location=build_info.storage,\n ).inner_dataset\n\n return datasets","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.ConceptualCaption3MBuilder","uri":"program://CREMA/class/lavis.datasets.builders.image_text_pair_builder.ConceptualCaption3MBuilder#L17-L22","kind":"class","name":"ConceptualCaption3MBuilder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":17,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\nfrom lavis.common.registry import registry\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.image_text_pair_datasets import ImageTextPairDataset\nfrom lavis.datasets.datasets.laion_dataset import LaionDataset\n\n\n@registry.register_builder(\"conceptual_caption_3m\")\nclass ConceptualCaption3MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_3m.yaml\"\n }\n\n\n@registry.register_builder(\"conceptual_caption_12m\")\nclass ConceptualCaption12MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.ConceptualCaption12MBuilder","uri":"program://CREMA/class/lavis.datasets.builders.image_text_pair_builder.ConceptualCaption12MBuilder#L26-L31","kind":"class","name":"ConceptualCaption12MBuilder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":26,"end_line":31,"context_start_line":6,"context_end_line":51,"code":"\"\"\"\n\nimport os\nfrom lavis.common.registry import registry\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.image_text_pair_datasets import ImageTextPairDataset\nfrom lavis.datasets.datasets.laion_dataset import LaionDataset\n\n\n@registry.register_builder(\"conceptual_caption_3m\")\nclass ConceptualCaption3MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_3m.yaml\"\n }\n\n\n@registry.register_builder(\"conceptual_caption_12m\")\nclass ConceptualCaption12MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.SBUCaptionBuilder","uri":"program://CREMA/class/lavis.datasets.builders.image_text_pair_builder.SBUCaptionBuilder#L35-L38","kind":"class","name":"SBUCaptionBuilder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":35,"end_line":38,"context_start_line":15,"context_end_line":58,"code":"\n@registry.register_builder(\"conceptual_caption_3m\")\nclass ConceptualCaption3MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_3m.yaml\"\n }\n\n\n@registry.register_builder(\"conceptual_caption_12m\")\nclass ConceptualCaption12MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.VGCaptionBuilder","uri":"program://CREMA/class/lavis.datasets.builders.image_text_pair_builder.VGCaptionBuilder#L42-L45","kind":"class","name":"VGCaptionBuilder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":42,"end_line":45,"context_start_line":22,"context_end_line":65,"code":" }\n\n\n@registry.register_builder(\"conceptual_caption_12m\")\nclass ConceptualCaption12MBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.Laion2BMultiBuilder","uri":"program://CREMA/class/lavis.datasets.builders.image_text_pair_builder.Laion2BMultiBuilder#L49-L77","kind":"class","name":"Laion2BMultiBuilder","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":49,"end_line":77,"context_start_line":29,"context_end_line":77,"code":" DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/conceptual_caption/defaults_12m.yaml\"\n }\n\n\n@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()\n split = \"train\" # laion dataset only has train split\n\n # create datasets\n # [NOTE] return inner_datasets (wds.DataPipeline)\n dataset_cls = self.train_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=self.vis_processors[split],\n text_processor=self.text_processors[split],\n location=build_info.storage,\n ).inner_dataset\n\n return datasets","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder._download_ann","uri":"program://CREMA/function/lavis.datasets.builders.image_text_pair_builder._download_ann#L54-L55","kind":"function","name":"_download_ann","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":54,"end_line":55,"context_start_line":34,"context_end_line":75,"code":"@registry.register_builder(\"sbu_caption\")\nclass SBUCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()\n split = \"train\" # laion dataset only has train split\n\n # create datasets\n # [NOTE] return inner_datasets (wds.DataPipeline)\n dataset_cls = self.train_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=self.vis_processors[split],\n text_processor=self.text_processors[split],\n location=build_info.storage,\n ).inner_dataset","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder._download_vis","uri":"program://CREMA/function/lavis.datasets.builders.image_text_pair_builder._download_vis#L57-L58","kind":"function","name":"_download_vis","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":57,"end_line":58,"context_start_line":37,"context_end_line":77,"code":"\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sbu_caption/defaults.yaml\"}\n\n\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()\n split = \"train\" # laion dataset only has train split\n\n # create datasets\n # [NOTE] return inner_datasets (wds.DataPipeline)\n dataset_cls = self.train_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=self.vis_processors[split],\n text_processor=self.text_processors[split],\n location=build_info.storage,\n ).inner_dataset\n\n return datasets","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.image_text_pair_builder.build","uri":"program://CREMA/function/lavis.datasets.builders.image_text_pair_builder.build#L60-L77","kind":"function","name":"build","path":"lavis/datasets/builders/image_text_pair_builder.py","language":"python","start_line":60,"end_line":77,"context_start_line":40,"context_end_line":77,"code":"\n@registry.register_builder(\"vg_caption\")\nclass VGCaptionBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageTextPairDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_caption.yaml\"}\n\n\n@registry.register_builder(\"laion2B_multi\")\nclass Laion2BMultiBuilder(BaseDatasetBuilder):\n train_dataset_cls = LaionDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/laion/defaults_2B_multi.yaml\"}\n\n def _download_ann(self):\n pass\n\n def _download_vis(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n datasets = dict()\n split = \"train\" # laion dataset only has train split\n\n # create datasets\n # [NOTE] return inner_datasets (wds.DataPipeline)\n dataset_cls = self.train_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=self.vis_processors[split],\n text_processor=self.text_processors[split],\n location=build_info.storage,\n ).inner_dataset\n\n return datasets","source_hash":"c626b6a8ead05fad6b2757e6d92ace8363569a8a8401f7812ba65b04fa4dc518","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder","uri":"program://CREMA/module/lavis.datasets.builders.base_dataset_builder#L1-L329","kind":"module","name":"lavis.datasets.builders.base_dataset_builder","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":1,"end_line":329,"context_start_line":1,"context_end_line":329,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\nimport shutil\nimport warnings\n\nimport lavis.common.utils as utils\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import is_dist_avail_and_initialized, is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import extract_archive\nfrom lavis.processors.base_processor import BaseProcessor\nfrom omegaconf import OmegaConf\nfrom torchvision.datasets.utils import download_url\n\n\nclass BaseDatasetBuilder:\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__()\n\n if cfg is None:\n # help to create datasets from default config.\n self.config = load_dataset_config(self.default_config_path())\n elif isinstance(cfg, str):\n self.config = load_dataset_config(cfg)\n else:\n # when called from task.build_dataset()\n self.config = cfg\n\n self.data_type = self.config.data_type\n try:\n self.modality_type = self.config.modality_type\n except:\n self.modality_type = None\n \n self.vis_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n self.text_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n\n def build_datasets(self):\n # download, split, etc...\n # only called on 1 GPU/TPU in distributed\n\n if is_main_process():\n self._download_data()\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n # at this point, all the annotations and image/videos should be all downloaded to the specified locations.\n logging.info(\"Building datasets...\")\n datasets = self.build() # dataset['train'/'val'/'test']\n\n return datasets\n\n def build_processors(self):\n vis_proc_cfg = self.config.get(\"vis_processor\")\n txt_proc_cfg = self.config.get(\"text_processor\")\n\n if vis_proc_cfg is not None:\n vis_train_cfg = vis_proc_cfg.get(\"train\")\n vis_eval_cfg = vis_proc_cfg.get(\"eval\")\n\n self.vis_processors[\"train\"] = self._build_proc_from_cfg(vis_train_cfg)\n self.vis_processors[\"eval\"] = self._build_proc_from_cfg(vis_eval_cfg)\n\n if txt_proc_cfg is not None:\n txt_train_cfg = txt_proc_cfg.get(\"train\")\n txt_eval_cfg = txt_proc_cfg.get(\"eval\")\n\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod\n def _build_proc_from_cfg(cfg):\n return (\n registry.get_processor_class(cfg.name).from_config(cfg)\n if cfg is not None\n else None\n )\n\n @classmethod\n def default_config_path(cls, type=\"default\"):\n return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])\n\n def _download_data(self):\n self._download_ann()\n self._download_vis()\n\n def _download_ann(self):\n \"\"\"\n Download annotation files if necessary.\n All the vision-language datasets should have annotations of unified format.\n\n storage_path can be:\n (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.\n (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.\n\n Local annotation paths should be relative.\n \"\"\"\n anns = self.config.build_info.annotations\n\n splits = anns.keys()\n\n cache_root = registry.get_path(\"cache_root\")\n\n for split in splits:\n info = anns[split]\n \n urls, storage_paths = info.get(\"url\", None), info.storage\n\n if isinstance(urls, str):\n urls = [urls]\n if isinstance(storage_paths, str):\n storage_paths = [storage_paths]\n\n # print('urls', urls)\n # print('storage_paths', storage_paths)\n\n assert len(urls) == len(storage_paths)\n\n for url_or_filename, storage_path in zip(urls, storage_paths):\n # if storage_path is relative, make it full by prefixing with cache_root.\n if not os.path.isabs(storage_path):\n storage_path = os.path.join(cache_root, storage_path)\n\n dirname = os.path.dirname(storage_path)\n if not os.path.exists(dirname):\n os.makedirs(dirname)\n\n if os.path.isfile(url_or_filename):\n src, dst = url_or_filename, storage_path\n if not os.path.exists(dst):\n shutil.copyfile(src=src, dst=dst)\n else:\n logging.info(\"Using existing file {}.\".format(dst))\n else:\n if os.path.isdir(storage_path):\n # if only dirname is provided, suffix with basename of URL.\n raise ValueError(\n \"Expecting storage_path to be a file path, got directory {}\".format(\n storage_path\n )\n )\n else:\n filename = os.path.basename(storage_path)\n\n download_url(url=url_or_filename, root=dirname, filename=filename)\n\n def _download_vis(self):\n\n storage_path = self.config.build_info.get(self.data_type).storage\n storage_path = utils.get_cache_path(storage_path)\n\n if not os.path.exists(storage_path):\n warnings.warn(\n f\"\"\"\n The specified path {storage_path} for visual inputs does not exist.\n Please provide a correct path to the visual inputs or\n refer to datasets/download_scripts/README.md for downloading instructions.\n \"\"\"\n )\n\n def build(self):\n \"\"\"\n Create by split datasets inheriting torch.utils.data.Datasets.\n\n # build() can be dataset-specific. Overwrite to customize.\n \"\"\"\n self.build_processors()\n\n build_info = self.config.build_info\n\n ann_info = build_info.annotations\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in ann_info.keys():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n\n # processors\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train\n else self.vis_processors[\"eval\"]\n )\n text_processor = (\n self.text_processors[\"train\"]\n if is_train\n else self.text_processors[\"eval\"]\n )\n\n # annotation path\n ann_paths = ann_info.get(split).storage\n if isinstance(ann_paths, str):\n ann_paths = [ann_paths]\n\n abs_ann_paths = []\n for ann_path in ann_paths:\n if not os.path.isabs(ann_path):\n ann_path = utils.get_cache_path(ann_path)\n abs_ann_paths.append(ann_path)\n ann_paths = abs_ann_paths\n\n # visual data storage path\n vis_path = vis_info.storage\n #print('vis_path',vis_path)\n if not os.path.isabs(vis_path):\n # vis_path = os.path.join(utils.get_cache_path(), vis_path)\n vis_path = utils.get_cache_path(vis_path)\n #print('vis_path2', vis_path)\n if not os.path.exists(vis_path):\n warnings.warn(\"storage path {} does not exist.\".format(vis_path))\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n text_processor=text_processor,\n ann_paths=ann_paths,\n vis_root=vis_path,\n modality_type=self.modality_type,\n )\n\n return datasets\n\n\nclass MultiModalDatasetBuilder(BaseDatasetBuilder):\n \"\"\"\n MultiModalDatasetBuilder is a utility class designed to construct datasets\n suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)\n\n return datasets\n\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))\n \n for modality in self.data_type:\n proc_name = f\"{'vis' if 'image' in modality else modality}_processor\"\n dataset_args[proc_name] = self.processors[\"train\" if is_train else \"eval\"][modality]\n mm_path = self._get_absolute_path(self.config.build_info.get(modality).storage)\n dataset_args[f\"{'vis' if 'image' in modality else modality}_root\"] = mm_path\n \n dataset_args['text_processor'] = self.text_processors[\"train\" if is_train else \"eval\"]\n dataset_args[\"ann_paths\"] = [self._get_absolute_path(path) for path in info.storage]\n dataset_args['modalities'] = self.data_type\n \n # Conform to base\n for key in ['vis_processor', 'vis_root', 'test_processor']:\n dataset_args.setdefault(key, None)\n \n return dataset_args\n \ndef load_dataset_config(cfg_path):\n cfg = OmegaConf.load(cfg_path).datasets\n cfg = cfg[list(cfg.keys())[0]]\n\n return cfg","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.BaseDatasetBuilder","uri":"program://CREMA/class/lavis.datasets.builders.base_dataset_builder.BaseDatasetBuilder#L23-L235","kind":"class","name":"BaseDatasetBuilder","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":23,"end_line":235,"context_start_line":3,"context_end_line":255,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\nimport shutil\nimport warnings\n\nimport lavis.common.utils as utils\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import is_dist_avail_and_initialized, is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import extract_archive\nfrom lavis.processors.base_processor import BaseProcessor\nfrom omegaconf import OmegaConf\nfrom torchvision.datasets.utils import download_url\n\n\nclass BaseDatasetBuilder:\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__()\n\n if cfg is None:\n # help to create datasets from default config.\n self.config = load_dataset_config(self.default_config_path())\n elif isinstance(cfg, str):\n self.config = load_dataset_config(cfg)\n else:\n # when called from task.build_dataset()\n self.config = cfg\n\n self.data_type = self.config.data_type\n try:\n self.modality_type = self.config.modality_type\n except:\n self.modality_type = None\n \n self.vis_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n self.text_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n\n def build_datasets(self):\n # download, split, etc...\n # only called on 1 GPU/TPU in distributed\n\n if is_main_process():\n self._download_data()\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n # at this point, all the annotations and image/videos should be all downloaded to the specified locations.\n logging.info(\"Building datasets...\")\n datasets = self.build() # dataset['train'/'val'/'test']\n\n return datasets\n\n def build_processors(self):\n vis_proc_cfg = self.config.get(\"vis_processor\")\n txt_proc_cfg = self.config.get(\"text_processor\")\n\n if vis_proc_cfg is not None:\n vis_train_cfg = vis_proc_cfg.get(\"train\")\n vis_eval_cfg = vis_proc_cfg.get(\"eval\")\n\n self.vis_processors[\"train\"] = self._build_proc_from_cfg(vis_train_cfg)\n self.vis_processors[\"eval\"] = self._build_proc_from_cfg(vis_eval_cfg)\n\n if txt_proc_cfg is not None:\n txt_train_cfg = txt_proc_cfg.get(\"train\")\n txt_eval_cfg = txt_proc_cfg.get(\"eval\")\n\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod\n def _build_proc_from_cfg(cfg):\n return (\n registry.get_processor_class(cfg.name).from_config(cfg)\n if cfg is not None\n else None\n )\n\n @classmethod\n def default_config_path(cls, type=\"default\"):\n return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])\n\n def _download_data(self):\n self._download_ann()\n self._download_vis()\n\n def _download_ann(self):\n \"\"\"\n Download annotation files if necessary.\n All the vision-language datasets should have annotations of unified format.\n\n storage_path can be:\n (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.\n (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.\n\n Local annotation paths should be relative.\n \"\"\"\n anns = self.config.build_info.annotations\n\n splits = anns.keys()\n\n cache_root = registry.get_path(\"cache_root\")\n\n for split in splits:\n info = anns[split]\n \n urls, storage_paths = info.get(\"url\", None), info.storage\n\n if isinstance(urls, str):\n urls = [urls]\n if isinstance(storage_paths, str):\n storage_paths = [storage_paths]\n\n # print('urls', urls)\n # print('storage_paths', storage_paths)\n\n assert len(urls) == len(storage_paths)\n\n for url_or_filename, storage_path in zip(urls, storage_paths):\n # if storage_path is relative, make it full by prefixing with cache_root.\n if not os.path.isabs(storage_path):\n storage_path = os.path.join(cache_root, storage_path)\n\n dirname = os.path.dirname(storage_path)\n if not os.path.exists(dirname):\n os.makedirs(dirname)\n\n if os.path.isfile(url_or_filename):\n src, dst = url_or_filename, storage_path\n if not os.path.exists(dst):\n shutil.copyfile(src=src, dst=dst)\n else:\n logging.info(\"Using existing file {}.\".format(dst))\n else:\n if os.path.isdir(storage_path):\n # if only dirname is provided, suffix with basename of URL.\n raise ValueError(\n \"Expecting storage_path to be a file path, got directory {}\".format(\n storage_path\n )\n )\n else:\n filename = os.path.basename(storage_path)\n\n download_url(url=url_or_filename, root=dirname, filename=filename)\n\n def _download_vis(self):\n\n storage_path = self.config.build_info.get(self.data_type).storage\n storage_path = utils.get_cache_path(storage_path)\n\n if not os.path.exists(storage_path):\n warnings.warn(\n f\"\"\"\n The specified path {storage_path} for visual inputs does not exist.\n Please provide a correct path to the visual inputs or\n refer to datasets/download_scripts/README.md for downloading instructions.\n \"\"\"\n )\n\n def build(self):\n \"\"\"\n Create by split datasets inheriting torch.utils.data.Datasets.\n\n # build() can be dataset-specific. Overwrite to customize.\n \"\"\"\n self.build_processors()\n\n build_info = self.config.build_info\n\n ann_info = build_info.annotations\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in ann_info.keys():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n\n # processors\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train\n else self.vis_processors[\"eval\"]\n )\n text_processor = (\n self.text_processors[\"train\"]\n if is_train\n else self.text_processors[\"eval\"]\n )\n\n # annotation path\n ann_paths = ann_info.get(split).storage\n if isinstance(ann_paths, str):\n ann_paths = [ann_paths]\n\n abs_ann_paths = []\n for ann_path in ann_paths:\n if not os.path.isabs(ann_path):\n ann_path = utils.get_cache_path(ann_path)\n abs_ann_paths.append(ann_path)\n ann_paths = abs_ann_paths\n\n # visual data storage path\n vis_path = vis_info.storage\n #print('vis_path',vis_path)\n if not os.path.isabs(vis_path):\n # vis_path = os.path.join(utils.get_cache_path(), vis_path)\n vis_path = utils.get_cache_path(vis_path)\n #print('vis_path2', vis_path)\n if not os.path.exists(vis_path):\n warnings.warn(\"storage path {} does not exist.\".format(vis_path))\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n text_processor=text_processor,\n ann_paths=ann_paths,\n vis_root=vis_path,\n modality_type=self.modality_type,\n )\n\n return datasets\n\n\nclass MultiModalDatasetBuilder(BaseDatasetBuilder):\n \"\"\"\n MultiModalDatasetBuilder is a utility class designed to construct datasets\n suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) ","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.MultiModalDatasetBuilder","uri":"program://CREMA/class/lavis.datasets.builders.base_dataset_builder.MultiModalDatasetBuilder#L238-L323","kind":"class","name":"MultiModalDatasetBuilder","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":238,"end_line":323,"context_start_line":218,"context_end_line":329,"code":" if not os.path.isabs(vis_path):\n # vis_path = os.path.join(utils.get_cache_path(), vis_path)\n vis_path = utils.get_cache_path(vis_path)\n #print('vis_path2', vis_path)\n if not os.path.exists(vis_path):\n warnings.warn(\"storage path {} does not exist.\".format(vis_path))\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n text_processor=text_processor,\n ann_paths=ann_paths,\n vis_root=vis_path,\n modality_type=self.modality_type,\n )\n\n return datasets\n\n\nclass MultiModalDatasetBuilder(BaseDatasetBuilder):\n \"\"\"\n MultiModalDatasetBuilder is a utility class designed to construct datasets\n suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)\n\n return datasets\n\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))\n \n for modality in self.data_type:\n proc_name = f\"{'vis' if 'image' in modality else modality}_processor\"\n dataset_args[proc_name] = self.processors[\"train\" if is_train else \"eval\"][modality]\n mm_path = self._get_absolute_path(self.config.build_info.get(modality).storage)\n dataset_args[f\"{'vis' if 'image' in modality else modality}_root\"] = mm_path\n \n dataset_args['text_processor'] = self.text_processors[\"train\" if is_train else \"eval\"]\n dataset_args[\"ann_paths\"] = [self._get_absolute_path(path) for path in info.storage]\n dataset_args['modalities'] = self.data_type\n \n # Conform to base\n for key in ['vis_processor', 'vis_root', 'test_processor']:\n dataset_args.setdefault(key, None)\n \n return dataset_args\n \ndef load_dataset_config(cfg_path):\n cfg = OmegaConf.load(cfg_path).datasets\n cfg = cfg[list(cfg.keys())[0]]\n\n return cfg","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.load_dataset_config","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.load_dataset_config#L325-L329","kind":"function","name":"load_dataset_config","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":325,"end_line":329,"context_start_line":305,"context_end_line":329,"code":"\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))\n \n for modality in self.data_type:\n proc_name = f\"{'vis' if 'image' in modality else modality}_processor\"\n dataset_args[proc_name] = self.processors[\"train\" if is_train else \"eval\"][modality]\n mm_path = self._get_absolute_path(self.config.build_info.get(modality).storage)\n dataset_args[f\"{'vis' if 'image' in modality else modality}_root\"] = mm_path\n \n dataset_args['text_processor'] = self.text_processors[\"train\" if is_train else \"eval\"]\n dataset_args[\"ann_paths\"] = [self._get_absolute_path(path) for path in info.storage]\n dataset_args['modalities'] = self.data_type\n \n # Conform to base\n for key in ['vis_processor', 'vis_root', 'test_processor']:\n dataset_args.setdefault(key, None)\n \n return dataset_args\n \ndef load_dataset_config(cfg_path):\n cfg = OmegaConf.load(cfg_path).datasets\n cfg = cfg[list(cfg.keys())[0]]\n\n return cfg","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.__init__","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.__init__#L247-L250","kind":"function","name":"__init__","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":247,"end_line":250,"context_start_line":227,"context_end_line":270,"code":" datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n text_processor=text_processor,\n ann_paths=ann_paths,\n vis_root=vis_path,\n modality_type=self.modality_type,\n )\n\n return datasets\n\n\nclass MultiModalDatasetBuilder(BaseDatasetBuilder):\n \"\"\"\n MultiModalDatasetBuilder is a utility class designed to construct datasets\n suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.build_datasets","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.build_datasets#L47-L61","kind":"function","name":"build_datasets","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":47,"end_line":61,"context_start_line":27,"context_end_line":81,"code":" super().__init__()\n\n if cfg is None:\n # help to create datasets from default config.\n self.config = load_dataset_config(self.default_config_path())\n elif isinstance(cfg, str):\n self.config = load_dataset_config(cfg)\n else:\n # when called from task.build_dataset()\n self.config = cfg\n\n self.data_type = self.config.data_type\n try:\n self.modality_type = self.config.modality_type\n except:\n self.modality_type = None\n \n self.vis_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n self.text_processors = {\"train\": BaseProcessor(), \"eval\": BaseProcessor()}\n\n def build_datasets(self):\n # download, split, etc...\n # only called on 1 GPU/TPU in distributed\n\n if is_main_process():\n self._download_data()\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n # at this point, all the annotations and image/videos should be all downloaded to the specified locations.\n logging.info(\"Building datasets...\")\n datasets = self.build() # dataset['train'/'val'/'test']\n\n return datasets\n\n def build_processors(self):\n vis_proc_cfg = self.config.get(\"vis_processor\")\n txt_proc_cfg = self.config.get(\"text_processor\")\n\n if vis_proc_cfg is not None:\n vis_train_cfg = vis_proc_cfg.get(\"train\")\n vis_eval_cfg = vis_proc_cfg.get(\"eval\")\n\n self.vis_processors[\"train\"] = self._build_proc_from_cfg(vis_train_cfg)\n self.vis_processors[\"eval\"] = self._build_proc_from_cfg(vis_eval_cfg)\n\n if txt_proc_cfg is not None:\n txt_train_cfg = txt_proc_cfg.get(\"train\")\n txt_eval_cfg = txt_proc_cfg.get(\"eval\")\n\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.build_processors","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.build_processors#L261-L272","kind":"function","name":"build_processors","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":261,"end_line":272,"context_start_line":241,"context_end_line":292,"code":" suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._build_proc_from_cfg","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._build_proc_from_cfg#L82-L87","kind":"function","name":"_build_proc_from_cfg","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":82,"end_line":87,"context_start_line":62,"context_end_line":107,"code":"\n def build_processors(self):\n vis_proc_cfg = self.config.get(\"vis_processor\")\n txt_proc_cfg = self.config.get(\"text_processor\")\n\n if vis_proc_cfg is not None:\n vis_train_cfg = vis_proc_cfg.get(\"train\")\n vis_eval_cfg = vis_proc_cfg.get(\"eval\")\n\n self.vis_processors[\"train\"] = self._build_proc_from_cfg(vis_train_cfg)\n self.vis_processors[\"eval\"] = self._build_proc_from_cfg(vis_eval_cfg)\n\n if txt_proc_cfg is not None:\n txt_train_cfg = txt_proc_cfg.get(\"train\")\n txt_eval_cfg = txt_proc_cfg.get(\"eval\")\n\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod\n def _build_proc_from_cfg(cfg):\n return (\n registry.get_processor_class(cfg.name).from_config(cfg)\n if cfg is not None\n else None\n )\n\n @classmethod\n def default_config_path(cls, type=\"default\"):\n return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])\n\n def _download_data(self):\n self._download_ann()\n self._download_vis()\n\n def _download_ann(self):\n \"\"\"\n Download annotation files if necessary.\n All the vision-language datasets should have annotations of unified format.\n\n storage_path can be:\n (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.\n (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.\n\n Local annotation paths should be relative.\n \"\"\"","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.default_config_path","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.default_config_path#L90-L91","kind":"function","name":"default_config_path","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":90,"end_line":91,"context_start_line":70,"context_end_line":111,"code":"\n self.vis_processors[\"train\"] = self._build_proc_from_cfg(vis_train_cfg)\n self.vis_processors[\"eval\"] = self._build_proc_from_cfg(vis_eval_cfg)\n\n if txt_proc_cfg is not None:\n txt_train_cfg = txt_proc_cfg.get(\"train\")\n txt_eval_cfg = txt_proc_cfg.get(\"eval\")\n\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod\n def _build_proc_from_cfg(cfg):\n return (\n registry.get_processor_class(cfg.name).from_config(cfg)\n if cfg is not None\n else None\n )\n\n @classmethod\n def default_config_path(cls, type=\"default\"):\n return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])\n\n def _download_data(self):\n self._download_ann()\n self._download_vis()\n\n def _download_ann(self):\n \"\"\"\n Download annotation files if necessary.\n All the vision-language datasets should have annotations of unified format.\n\n storage_path can be:\n (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.\n (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.\n\n Local annotation paths should be relative.\n \"\"\"\n anns = self.config.build_info.annotations\n\n splits = anns.keys()\n","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._download_data","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._download_data#L279-L282","kind":"function","name":"_download_data","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":279,"end_line":282,"context_start_line":259,"context_end_line":302,"code":" }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._download_ann","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._download_ann#L97-L155","kind":"function","name":"_download_ann","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":97,"end_line":155,"context_start_line":77,"context_end_line":175,"code":"\n self.text_processors[\"train\"] = self._build_proc_from_cfg(txt_train_cfg)\n self.text_processors[\"eval\"] = self._build_proc_from_cfg(txt_eval_cfg)\n\n @staticmethod\n def _build_proc_from_cfg(cfg):\n return (\n registry.get_processor_class(cfg.name).from_config(cfg)\n if cfg is not None\n else None\n )\n\n @classmethod\n def default_config_path(cls, type=\"default\"):\n return utils.get_abs_path(cls.DATASET_CONFIG_DICT[type])\n\n def _download_data(self):\n self._download_ann()\n self._download_vis()\n\n def _download_ann(self):\n \"\"\"\n Download annotation files if necessary.\n All the vision-language datasets should have annotations of unified format.\n\n storage_path can be:\n (1) relative/absolute: will be prefixed with env.cache_root to make full path if relative.\n (2) basename/dirname: will be suffixed with base name of URL if dirname is provided.\n\n Local annotation paths should be relative.\n \"\"\"\n anns = self.config.build_info.annotations\n\n splits = anns.keys()\n\n cache_root = registry.get_path(\"cache_root\")\n\n for split in splits:\n info = anns[split]\n \n urls, storage_paths = info.get(\"url\", None), info.storage\n\n if isinstance(urls, str):\n urls = [urls]\n if isinstance(storage_paths, str):\n storage_paths = [storage_paths]\n\n # print('urls', urls)\n # print('storage_paths', storage_paths)\n\n assert len(urls) == len(storage_paths)\n\n for url_or_filename, storage_path in zip(urls, storage_paths):\n # if storage_path is relative, make it full by prefixing with cache_root.\n if not os.path.isabs(storage_path):\n storage_path = os.path.join(cache_root, storage_path)\n\n dirname = os.path.dirname(storage_path)\n if not os.path.exists(dirname):\n os.makedirs(dirname)\n\n if os.path.isfile(url_or_filename):\n src, dst = url_or_filename, storage_path\n if not os.path.exists(dst):\n shutil.copyfile(src=src, dst=dst)\n else:\n logging.info(\"Using existing file {}.\".format(dst))\n else:\n if os.path.isdir(storage_path):\n # if only dirname is provided, suffix with basename of URL.\n raise ValueError(\n \"Expecting storage_path to be a file path, got directory {}\".format(\n storage_path\n )\n )\n else:\n filename = os.path.basename(storage_path)\n\n download_url(url=url_or_filename, root=dirname, filename=filename)\n\n def _download_vis(self):\n\n storage_path = self.config.build_info.get(self.data_type).storage\n storage_path = utils.get_cache_path(storage_path)\n\n if not os.path.exists(storage_path):\n warnings.warn(\n f\"\"\"\n The specified path {storage_path} for visual inputs does not exist.\n Please provide a correct path to the visual inputs or\n refer to datasets/download_scripts/README.md for downloading instructions.\n \"\"\"\n )\n\n def build(self):\n \"\"\"\n Create by split datasets inheriting torch.utils.data.Datasets.\n\n # build() can be dataset-specific. Overwrite to customize.","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._download_vis","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._download_vis#L157-L169","kind":"function","name":"_download_vis","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":157,"end_line":169,"context_start_line":137,"context_end_line":189,"code":"\n if os.path.isfile(url_or_filename):\n src, dst = url_or_filename, storage_path\n if not os.path.exists(dst):\n shutil.copyfile(src=src, dst=dst)\n else:\n logging.info(\"Using existing file {}.\".format(dst))\n else:\n if os.path.isdir(storage_path):\n # if only dirname is provided, suffix with basename of URL.\n raise ValueError(\n \"Expecting storage_path to be a file path, got directory {}\".format(\n storage_path\n )\n )\n else:\n filename = os.path.basename(storage_path)\n\n download_url(url=url_or_filename, root=dirname, filename=filename)\n\n def _download_vis(self):\n\n storage_path = self.config.build_info.get(self.data_type).storage\n storage_path = utils.get_cache_path(storage_path)\n\n if not os.path.exists(storage_path):\n warnings.warn(\n f\"\"\"\n The specified path {storage_path} for visual inputs does not exist.\n Please provide a correct path to the visual inputs or\n refer to datasets/download_scripts/README.md for downloading instructions.\n \"\"\"\n )\n\n def build(self):\n \"\"\"\n Create by split datasets inheriting torch.utils.data.Datasets.\n\n # build() can be dataset-specific. Overwrite to customize.\n \"\"\"\n self.build_processors()\n\n build_info = self.config.build_info\n\n ann_info = build_info.annotations\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in ann_info.keys():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder.build","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder.build#L289-L304","kind":"function","name":"build","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":289,"end_line":304,"context_start_line":269,"context_end_line":324,"code":" for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)\n\n return datasets\n\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))\n \n for modality in self.data_type:\n proc_name = f\"{'vis' if 'image' in modality else modality}_processor\"\n dataset_args[proc_name] = self.processors[\"train\" if is_train else \"eval\"][modality]\n mm_path = self._get_absolute_path(self.config.build_info.get(modality).storage)\n dataset_args[f\"{'vis' if 'image' in modality else modality}_root\"] = mm_path\n \n dataset_args['text_processor'] = self.text_processors[\"train\" if is_train else \"eval\"]\n dataset_args[\"ann_paths\"] = [self._get_absolute_path(path) for path in info.storage]\n dataset_args['modalities'] = self.data_type\n \n # Conform to base\n for key in ['vis_processor', 'vis_root', 'test_processor']:\n dataset_args.setdefault(key, None)\n \n return dataset_args\n ","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._build_processor","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._build_processor#L252-L259","kind":"function","name":"_build_processor","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":252,"end_line":259,"context_start_line":232,"context_end_line":279,"code":" modality_type=self.modality_type,\n )\n\n return datasets\n\n\nclass MultiModalDatasetBuilder(BaseDatasetBuilder):\n \"\"\"\n MultiModalDatasetBuilder is a utility class designed to construct datasets\n suitable for multi-modal tasks. This class simplifies the creation of \n datasets that incorporate data of multiple modalities, such as text, \n images, video, or audio.\n \"\"\"\n train_dataset_cls, eval_dataset_cls = None, None\n\n def __init__(self, cfg=None):\n super().__init__(cfg)\n if isinstance(self.data_type, str):\n self.data_type = [self.data_type]\n\n def _build_processor(self, cfg_name):\n cfg = self.config.get(cfg_name)\n return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._download_multimodal","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._download_multimodal#L274-L277","kind":"function","name":"_download_multimodal","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":274,"end_line":277,"context_start_line":254,"context_end_line":297,"code":" return {\n split: self._build_proc_from_cfg(cfg.get(split)) \n if cfg is not None \n else None\n for split in ['train', 'eval']\n }\n\n def build_processors(self):\n self.text_processors = self._build_processor(\"text_processor\")\n \n self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._get_absolute_path","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._get_absolute_path#L284-L287","kind":"function","name":"_get_absolute_path","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":284,"end_line":287,"context_start_line":264,"context_end_line":307,"code":" self.processors = {\n split: {\n modality: self._build_proc_from_cfg(\n self.config.get(f\"{'vis' if 'image' in modality else modality}_processor\").get(split)\n )\n for modality in self.data_type\n }\n for split in ['train', 'eval']\n }\n\n def _download_multimodal(self, modality):\n storage_path = utils.get_cache_path(self.config.build_info.get(modality).storage)\n if not os.path.exists(storage_path):\n warnings.warn(f\"The specified path {storage_path} for {modality} inputs does not exist.\")\n\n def _download_data(self):\n self._download_ann()\n for modality in self.data_type:\n self._download_multimodal(modality)\n\n def _get_absolute_path(self, path):\n if not os.path.isabs(path):\n return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)\n\n return datasets\n\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.base_dataset_builder._get_dataset_args","uri":"program://CREMA/function/lavis.datasets.builders.base_dataset_builder._get_dataset_args#L306-L323","kind":"function","name":"_get_dataset_args","path":"lavis/datasets/builders/base_dataset_builder.py","language":"python","start_line":306,"end_line":323,"context_start_line":286,"context_end_line":329,"code":" return utils.get_cache_path(path)\n return path\n\n def build(self):\n self.build_processors()\n build_info = self.config.build_info\n datasets = {}\n \n for split, info in build_info.annotations.items():\n if split not in [\"train\", \"val\", \"test\"]:\n continue\n\n is_train = split == \"train\"\n dataset_args = self._get_dataset_args(info, is_train)\n \n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(**dataset_args)\n\n return datasets\n\n def _get_dataset_args(self, info, is_train):\n dataset_args = dict(self.config.build_info.get('kwargs', {}))\n \n for modality in self.data_type:\n proc_name = f\"{'vis' if 'image' in modality else modality}_processor\"\n dataset_args[proc_name] = self.processors[\"train\" if is_train else \"eval\"][modality]\n mm_path = self._get_absolute_path(self.config.build_info.get(modality).storage)\n dataset_args[f\"{'vis' if 'image' in modality else modality}_root\"] = mm_path\n \n dataset_args['text_processor'] = self.text_processors[\"train\" if is_train else \"eval\"]\n dataset_args[\"ann_paths\"] = [self._get_absolute_path(path) for path in info.storage]\n dataset_args['modalities'] = self.data_type\n \n # Conform to base\n for key in ['vis_processor', 'vis_root', 'test_processor']:\n dataset_args.setdefault(key, None)\n \n return dataset_args\n \ndef load_dataset_config(cfg_path):\n cfg = OmegaConf.load(cfg_path).datasets\n cfg = cfg[list(cfg.keys())[0]]\n\n return cfg","source_hash":"37c51d31a78f720367b4a6b90235335cd7de8e8be000f155fa1d6678b2aafaa9","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder","uri":"program://CREMA/module/lavis.datasets.builders.video_qa_builder#L1-L97","kind":"module","name":"lavis.datasets.builders.video_qa_builder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":1,"end_line":97,"context_start_line":1,"context_end_line":97,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_cache_path\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder, MultiModalDatasetBuilder\nfrom lavis.datasets.datasets.video_vqa_datasets import VideoQADataset\n\nfrom lavis.datasets.datasets.rgbd_vqa_datasets import MCVideoQADataset\nfrom lavis.datasets.datasets.music_avqa_datasets import MusicAVQAInstructDataset, MusicAVQADataset\n\nfrom lavis.datasets.datasets.threedvqa_datasets import ThreeDVQADataset, ThreeDVQAEvalDataset\n\nclass VideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoQADataset\n eval_dataset_cls = VideoQADataset\n\n def build(self):\n datasets = super().build()\n\n ans2label = self.config.build_info.annotations.get(\"ans2label\")\n if ans2label is None:\n raise ValueError(\"ans2label is not specified in build_info.\")\n\n ans2label = get_cache_path(ans2label.storage)\n\n for split in datasets:\n datasets[split]._build_class_labels(ans2label)\n\n return datasets\n \nclass MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",\n# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset\n eval_dataset_cls = MusicAVQADataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa.yaml\"}\n\n@registry.register_builder(\"musicavqa_mm_instruct\")\nclass MusicAVQAInstructBuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQAInstructDataset\n eval_dataset_cls = MusicAVQAInstructDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa_instruct.yaml\"}\n\n\n@registry.register_builder(\"sqa3d\")\nclass ThreeDVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = ThreeDVQADataset\n eval_dataset_cls = ThreeDVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sqa3d/defaults.yaml\"}","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.VideoQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.VideoQABuilder#L18-L34","kind":"class","name":"VideoQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":18,"end_line":34,"context_start_line":1,"context_end_line":54,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.common.utils import get_cache_path\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder, MultiModalDatasetBuilder\nfrom lavis.datasets.datasets.video_vqa_datasets import VideoQADataset\n\nfrom lavis.datasets.datasets.rgbd_vqa_datasets import MCVideoQADataset\nfrom lavis.datasets.datasets.music_avqa_datasets import MusicAVQAInstructDataset, MusicAVQADataset\n\nfrom lavis.datasets.datasets.threedvqa_datasets import ThreeDVQADataset, ThreeDVQAEvalDataset\n\nclass VideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoQADataset\n eval_dataset_cls = VideoQADataset\n\n def build(self):\n datasets = super().build()\n\n ans2label = self.config.build_info.annotations.get(\"ans2label\")\n if ans2label is None:\n raise ValueError(\"ans2label is not specified in build_info.\")\n\n ans2label = get_cache_path(ans2label.storage)\n\n for split in datasets:\n datasets[split]._build_class_labels(ans2label)\n\n return datasets\n \nclass MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.MCVideoQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.MCVideoQABuilder#L36-L46","kind":"class","name":"MCVideoQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":36,"end_line":46,"context_start_line":16,"context_end_line":66,"code":"from lavis.datasets.datasets.threedvqa_datasets import ThreeDVQADataset, ThreeDVQAEvalDataset\n\nclass VideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoQADataset\n eval_dataset_cls = VideoQADataset\n\n def build(self):\n datasets = super().build()\n\n ans2label = self.config.build_info.annotations.get(\"ans2label\")\n if ans2label is None:\n raise ValueError(\"ans2label is not specified in build_info.\")\n\n ans2label = get_cache_path(ans2label.storage)\n\n for split in datasets:\n datasets[split]._build_class_labels(ans2label)\n\n return datasets\n \nclass MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.MSRVTTQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.MSRVTTQABuilder#L49-L52","kind":"class","name":"MSRVTTQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":49,"end_line":52,"context_start_line":29,"context_end_line":72,"code":" ans2label = get_cache_path(ans2label.storage)\n\n for split in datasets:\n datasets[split]._build_class_labels(ans2label)\n\n return datasets\n \nclass MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.MSVDQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.MSVDQABuilder#L56-L59","kind":"class","name":"MSVDQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":56,"end_line":59,"context_start_line":36,"context_end_line":79,"code":"class MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",\n# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.NextQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.NextQABuilder#L64-L67","kind":"class","name":"NextQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":64,"end_line":67,"context_start_line":44,"context_end_line":87,"code":" datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",\n# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset\n eval_dataset_cls = MusicAVQADataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa.yaml\"}\n\n@registry.register_builder(\"musicavqa_mm_instruct\")\nclass MusicAVQAInstructBuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQAInstructDataset\n eval_dataset_cls = MusicAVQAInstructDataset","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.MusicAVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.MusicAVQABuilder#L78-L82","kind":"class","name":"MusicAVQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":78,"end_line":82,"context_start_line":58,"context_end_line":97,"code":" \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",\n# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset\n eval_dataset_cls = MusicAVQADataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa.yaml\"}\n\n@registry.register_builder(\"musicavqa_mm_instruct\")\nclass MusicAVQAInstructBuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQAInstructDataset\n eval_dataset_cls = MusicAVQAInstructDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa_instruct.yaml\"}\n\n\n@registry.register_builder(\"sqa3d\")\nclass ThreeDVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = ThreeDVQADataset\n eval_dataset_cls = ThreeDVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sqa3d/defaults.yaml\"}","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.MusicAVQAInstructBuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.MusicAVQAInstructBuilder#L85-L89","kind":"class","name":"MusicAVQAInstructBuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":85,"end_line":89,"context_start_line":65,"context_end_line":97,"code":" DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",\n }\n \n# @registry.register_builder(\"perception_test\")\n# class NextQA3DBuilder(MCVideoQA3DBuilder):\n# DATASET_CONFIG_DICT = {\n# \"default\": \"configs/datasets/perception/defaults_qa.yaml\",\n# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset\n eval_dataset_cls = MusicAVQADataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa.yaml\"}\n\n@registry.register_builder(\"musicavqa_mm_instruct\")\nclass MusicAVQAInstructBuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQAInstructDataset\n eval_dataset_cls = MusicAVQAInstructDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa_instruct.yaml\"}\n\n\n@registry.register_builder(\"sqa3d\")\nclass ThreeDVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = ThreeDVQADataset\n eval_dataset_cls = ThreeDVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sqa3d/defaults.yaml\"}","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.ThreeDVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.video_qa_builder.ThreeDVQABuilder#L93-L97","kind":"class","name":"ThreeDVQABuilder","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":93,"end_line":97,"context_start_line":73,"context_end_line":97,"code":"# }\n \n# open-ended QA\n\n@registry.register_builder(\"musicavqa_mm\")\nclass MusicAVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQADataset\n eval_dataset_cls = MusicAVQADataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa.yaml\"}\n\n@registry.register_builder(\"musicavqa_mm_instruct\")\nclass MusicAVQAInstructBuilder(MultiModalDatasetBuilder):\n train_dataset_cls = MusicAVQAInstructDataset\n eval_dataset_cls = MusicAVQAInstructDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/music_avqa/defaults_mm_qa_instruct.yaml\"}\n\n\n@registry.register_builder(\"sqa3d\")\nclass ThreeDVQABuilder(MultiModalDatasetBuilder):\n train_dataset_cls = ThreeDVQADataset\n eval_dataset_cls = ThreeDVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/sqa3d/defaults.yaml\"}","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.video_qa_builder.build","uri":"program://CREMA/function/lavis.datasets.builders.video_qa_builder.build#L40-L46","kind":"function","name":"build","path":"lavis/datasets/builders/video_qa_builder.py","language":"python","start_line":40,"end_line":46,"context_start_line":20,"context_end_line":66,"code":" eval_dataset_cls = VideoQADataset\n\n def build(self):\n datasets = super().build()\n\n ans2label = self.config.build_info.annotations.get(\"ans2label\")\n if ans2label is None:\n raise ValueError(\"ans2label is not specified in build_info.\")\n\n ans2label = get_cache_path(ans2label.storage)\n\n for split in datasets:\n datasets[split]._build_class_labels(ans2label)\n\n return datasets\n \nclass MCVideoQABuilder(BaseDatasetBuilder):\n train_dataset_cls = MCVideoQADataset\n eval_dataset_cls = MCVideoQADataset\n\n def build(self):\n datasets = super().build()\n\n for split in datasets:\n datasets[split]._load_auxiliary_mappings()\n\n return datasets\n\n@registry.register_builder(\"msrvtt_qa\")\nclass MSRVTTQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_qa.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_qa\")\nclass MSVDQABuilder(VideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_qa.yaml\",\n }\n\n# multi-choice videoqa\n# to do update it \n@registry.register_builder(\"nextqa\")\nclass NextQABuilder(MCVideoQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nextqa/defaults_qa.yaml\",","source_hash":"5e92b5cf9b0ad00fc0c02efd865ef2d16234d8b880ca7f24d6a7cb4324d9bded","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.caption_builder","uri":"program://CREMA/module/lavis.datasets.builders.caption_builder#L1-L68","kind":"module","name":"lavis.datasets.builders.caption_builder","path":"lavis/datasets/builders/caption_builder.py","language":"python","start_line":1,"end_line":68,"context_start_line":1,"context_end_line":68,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.coco_caption_datasets import (\n COCOCapDataset,\n COCOCapEvalDataset,\n NoCapsEvalDataset,\n)\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.video_caption_datasets import (\n VideoCaptionDataset,\n VideoCaptionEvalDataset,\n)\n\n@registry.register_builder(\"coco_caption\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOCapDataset\n eval_dataset_cls = COCOCapEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"nocaps\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n eval_dataset_cls = NoCapsEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nocaps/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"msrvtt_caption\")\nclass MSRVTTCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_caption\")\nclass MSVDCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"vatex_caption\")\nclass VATEXCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/vatex/defaults_cap.yaml\",\n }\n","source_hash":"5a58a7aa38550fc96a66b81a4de9dde2e1370ca32583ca5b59677da39dc86263","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.caption_builder.COCOCapBuilder","uri":"program://CREMA/class/lavis.datasets.builders.caption_builder.COCOCapBuilder#L32-L37","kind":"class","name":"COCOCapBuilder","path":"lavis/datasets/builders/caption_builder.py","language":"python","start_line":32,"end_line":37,"context_start_line":12,"context_end_line":57,"code":" NoCapsEvalDataset,\n)\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.video_caption_datasets import (\n VideoCaptionDataset,\n VideoCaptionEvalDataset,\n)\n\n@registry.register_builder(\"coco_caption\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOCapDataset\n eval_dataset_cls = COCOCapEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"nocaps\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n eval_dataset_cls = NoCapsEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nocaps/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"msrvtt_caption\")\nclass MSRVTTCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_caption\")\nclass MSVDCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_cap.yaml\",\n }","source_hash":"5a58a7aa38550fc96a66b81a4de9dde2e1370ca32583ca5b59677da39dc86263","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.caption_builder.MSRVTTCapBuilder","uri":"program://CREMA/class/lavis.datasets.builders.caption_builder.MSRVTTCapBuilder#L41-L47","kind":"class","name":"MSRVTTCapBuilder","path":"lavis/datasets/builders/caption_builder.py","language":"python","start_line":41,"end_line":47,"context_start_line":21,"context_end_line":67,"code":"@registry.register_builder(\"coco_caption\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOCapDataset\n eval_dataset_cls = COCOCapEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"nocaps\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n eval_dataset_cls = NoCapsEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nocaps/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"msrvtt_caption\")\nclass MSRVTTCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_caption\")\nclass MSVDCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"vatex_caption\")\nclass VATEXCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/vatex/defaults_cap.yaml\",\n }","source_hash":"5a58a7aa38550fc96a66b81a4de9dde2e1370ca32583ca5b59677da39dc86263","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.caption_builder.MSVDCapBuilder","uri":"program://CREMA/class/lavis.datasets.builders.caption_builder.MSVDCapBuilder#L51-L57","kind":"class","name":"MSVDCapBuilder","path":"lavis/datasets/builders/caption_builder.py","language":"python","start_line":51,"end_line":57,"context_start_line":31,"context_end_line":68,"code":"@registry.register_builder(\"nocaps\")\nclass COCOCapBuilder(BaseDatasetBuilder):\n eval_dataset_cls = NoCapsEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/nocaps/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"msrvtt_caption\")\nclass MSRVTTCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_caption\")\nclass MSVDCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"vatex_caption\")\nclass VATEXCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/vatex/defaults_cap.yaml\",\n }\n","source_hash":"5a58a7aa38550fc96a66b81a4de9dde2e1370ca32583ca5b59677da39dc86263","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.caption_builder.VATEXCapBuilder","uri":"program://CREMA/class/lavis.datasets.builders.caption_builder.VATEXCapBuilder#L61-L67","kind":"class","name":"VATEXCapBuilder","path":"lavis/datasets/builders/caption_builder.py","language":"python","start_line":61,"end_line":67,"context_start_line":41,"context_end_line":68,"code":"class MSRVTTCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msrvtt/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"msvd_caption\")\nclass MSVDCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/msvd/defaults_cap.yaml\",\n }\n\n\n@registry.register_builder(\"vatex_caption\")\nclass VATEXCapBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoCaptionDataset\n eval_dataset_cls = VideoCaptionEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/vatex/defaults_cap.yaml\",\n }\n","source_hash":"5a58a7aa38550fc96a66b81a4de9dde2e1370ca32583ca5b59677da39dc86263","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.retrieval_builder","uri":"program://CREMA/module/lavis.datasets.builders.retrieval_builder#L1-L48","kind":"module","name":"lavis.datasets.builders.retrieval_builder","path":"lavis/datasets/builders/retrieval_builder.py","language":"python","start_line":1,"end_line":48,"context_start_line":1,"context_end_line":48,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.retrieval_datasets import (\n RetrievalDataset,\n RetrievalEvalDataset,\n VideoRetrievalDataset,\n VideoRetrievalEvalDataset,\n)\n\nfrom lavis.common.registry import registry\n\n\n@registry.register_builder(\"msrvtt_retrieval\")\nclass MSRVTTRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/msrvtt/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"didemo_retrieval\")\nclass DiDeMoRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/didemo/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"coco_retrieval\")\nclass COCORetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/coco/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"flickr30k\")\nclass Flickr30kBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/flickr30k/defaults.yaml\"}","source_hash":"440a946283135f054733ed112f6f256d1c35b3aba154640ff355f3b8845af27a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.retrieval_builder.MSRVTTRetrievalBuilder","uri":"program://CREMA/class/lavis.datasets.builders.retrieval_builder.MSRVTTRetrievalBuilder#L20-L24","kind":"class","name":"MSRVTTRetrievalBuilder","path":"lavis/datasets/builders/retrieval_builder.py","language":"python","start_line":20,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.retrieval_datasets import (\n RetrievalDataset,\n RetrievalEvalDataset,\n VideoRetrievalDataset,\n VideoRetrievalEvalDataset,\n)\n\nfrom lavis.common.registry import registry\n\n\n@registry.register_builder(\"msrvtt_retrieval\")\nclass MSRVTTRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/msrvtt/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"didemo_retrieval\")\nclass DiDeMoRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/didemo/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"coco_retrieval\")\nclass COCORetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/coco/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"flickr30k\")\nclass Flickr30kBuilder(BaseDatasetBuilder):","source_hash":"440a946283135f054733ed112f6f256d1c35b3aba154640ff355f3b8845af27a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.retrieval_builder.DiDeMoRetrievalBuilder","uri":"program://CREMA/class/lavis.datasets.builders.retrieval_builder.DiDeMoRetrievalBuilder#L28-L32","kind":"class","name":"DiDeMoRetrievalBuilder","path":"lavis/datasets/builders/retrieval_builder.py","language":"python","start_line":28,"end_line":32,"context_start_line":8,"context_end_line":48,"code":"from lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.retrieval_datasets import (\n RetrievalDataset,\n RetrievalEvalDataset,\n VideoRetrievalDataset,\n VideoRetrievalEvalDataset,\n)\n\nfrom lavis.common.registry import registry\n\n\n@registry.register_builder(\"msrvtt_retrieval\")\nclass MSRVTTRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/msrvtt/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"didemo_retrieval\")\nclass DiDeMoRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/didemo/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"coco_retrieval\")\nclass COCORetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/coco/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"flickr30k\")\nclass Flickr30kBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/flickr30k/defaults.yaml\"}","source_hash":"440a946283135f054733ed112f6f256d1c35b3aba154640ff355f3b8845af27a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.retrieval_builder.COCORetrievalBuilder","uri":"program://CREMA/class/lavis.datasets.builders.retrieval_builder.COCORetrievalBuilder#L36-L40","kind":"class","name":"COCORetrievalBuilder","path":"lavis/datasets/builders/retrieval_builder.py","language":"python","start_line":36,"end_line":40,"context_start_line":16,"context_end_line":48,"code":"from lavis.common.registry import registry\n\n\n@registry.register_builder(\"msrvtt_retrieval\")\nclass MSRVTTRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/msrvtt/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"didemo_retrieval\")\nclass DiDeMoRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/didemo/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"coco_retrieval\")\nclass COCORetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/coco/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"flickr30k\")\nclass Flickr30kBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/flickr30k/defaults.yaml\"}","source_hash":"440a946283135f054733ed112f6f256d1c35b3aba154640ff355f3b8845af27a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.retrieval_builder.Flickr30kBuilder","uri":"program://CREMA/class/lavis.datasets.builders.retrieval_builder.Flickr30kBuilder#L44-L48","kind":"class","name":"Flickr30kBuilder","path":"lavis/datasets/builders/retrieval_builder.py","language":"python","start_line":44,"end_line":48,"context_start_line":24,"context_end_line":48,"code":" DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/msrvtt/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"didemo_retrieval\")\nclass DiDeMoRetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = VideoRetrievalDataset\n eval_dataset_cls = VideoRetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/didemo/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"coco_retrieval\")\nclass COCORetrievalBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/coco/defaults_ret.yaml\"}\n\n\n@registry.register_builder(\"flickr30k\")\nclass Flickr30kBuilder(BaseDatasetBuilder):\n train_dataset_cls = RetrievalDataset\n eval_dataset_cls = RetrievalEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/flickr30k/defaults.yaml\"}","source_hash":"440a946283135f054733ed112f6f256d1c35b3aba154640ff355f3b8845af27a","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder","uri":"program://CREMA/module/lavis.datasets.builders.vqa_builder#L1-L58","kind":"module","name":"lavis.datasets.builders.vqa_builder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":1,"end_line":58,"context_start_line":1,"context_end_line":58,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.aok_vqa_datasets import AOKVQADataset, AOKVQAEvalDataset\nfrom lavis.datasets.datasets.coco_vqa_datasets import COCOVQADataset, COCOVQAEvalDataset\nfrom lavis.datasets.datasets.vg_vqa_datasets import VGVQADataset\nfrom lavis.datasets.datasets.gqa_datasets import GQADataset, GQAEvalDataset\n\n\n@registry.register_builder(\"coco_vqa\")\nclass COCOVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOVQADataset\n eval_dataset_cls = COCOVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_vqa.yaml\",\n \"eval\": \"configs/datasets/coco/eval_vqa.yaml\",\n }\n\n\n@registry.register_builder(\"vg_vqa\")\nclass VGVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/aokvqa/defaults.yaml\"}\n\n\n@registry.register_builder(\"gqa\")\nclass GQABuilder(BaseDatasetBuilder):\n train_dataset_cls = GQADataset\n eval_dataset_cls = GQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/gqa/defaults.yaml\",\n \"balanced_val\": \"configs/datasets/gqa/balanced_val.yaml\",\n \"balanced_testdev\": \"configs/datasets/gqa/balanced_testdev.yaml\",\n }","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder.COCOVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.vqa_builder.COCOVQABuilder#L18-L25","kind":"class","name":"COCOVQABuilder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":18,"end_line":25,"context_start_line":1,"context_end_line":45,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.aok_vqa_datasets import AOKVQADataset, AOKVQAEvalDataset\nfrom lavis.datasets.datasets.coco_vqa_datasets import COCOVQADataset, COCOVQAEvalDataset\nfrom lavis.datasets.datasets.vg_vqa_datasets import VGVQADataset\nfrom lavis.datasets.datasets.gqa_datasets import GQADataset, GQAEvalDataset\n\n\n@registry.register_builder(\"coco_vqa\")\nclass COCOVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOVQADataset\n eval_dataset_cls = COCOVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_vqa.yaml\",\n \"eval\": \"configs/datasets/coco/eval_vqa.yaml\",\n }\n\n\n@registry.register_builder(\"vg_vqa\")\nclass VGVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder.VGVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.vqa_builder.VGVQABuilder#L29-L31","kind":"class","name":"VGVQABuilder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":29,"end_line":31,"context_start_line":9,"context_end_line":51,"code":"\nfrom lavis.common.registry import registry\nfrom lavis.datasets.datasets.aok_vqa_datasets import AOKVQADataset, AOKVQAEvalDataset\nfrom lavis.datasets.datasets.coco_vqa_datasets import COCOVQADataset, COCOVQAEvalDataset\nfrom lavis.datasets.datasets.vg_vqa_datasets import VGVQADataset\nfrom lavis.datasets.datasets.gqa_datasets import GQADataset, GQAEvalDataset\n\n\n@registry.register_builder(\"coco_vqa\")\nclass COCOVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOVQADataset\n eval_dataset_cls = COCOVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_vqa.yaml\",\n \"eval\": \"configs/datasets/coco/eval_vqa.yaml\",\n }\n\n\n@registry.register_builder(\"vg_vqa\")\nclass VGVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/aokvqa/defaults.yaml\"}\n\n\n@registry.register_builder(\"gqa\")\nclass GQABuilder(BaseDatasetBuilder):\n train_dataset_cls = GQADataset","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder.OKVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.vqa_builder.OKVQABuilder#L35-L38","kind":"class","name":"OKVQABuilder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":35,"end_line":38,"context_start_line":15,"context_end_line":58,"code":"\n\n@registry.register_builder(\"coco_vqa\")\nclass COCOVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = COCOVQADataset\n eval_dataset_cls = COCOVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_vqa.yaml\",\n \"eval\": \"configs/datasets/coco/eval_vqa.yaml\",\n }\n\n\n@registry.register_builder(\"vg_vqa\")\nclass VGVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/aokvqa/defaults.yaml\"}\n\n\n@registry.register_builder(\"gqa\")\nclass GQABuilder(BaseDatasetBuilder):\n train_dataset_cls = GQADataset\n eval_dataset_cls = GQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/gqa/defaults.yaml\",\n \"balanced_val\": \"configs/datasets/gqa/balanced_val.yaml\",\n \"balanced_testdev\": \"configs/datasets/gqa/balanced_testdev.yaml\",\n }","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder.AOKVQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.vqa_builder.AOKVQABuilder#L42-L46","kind":"class","name":"AOKVQABuilder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":42,"end_line":46,"context_start_line":22,"context_end_line":58,"code":" DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/coco/defaults_vqa.yaml\",\n \"eval\": \"configs/datasets/coco/eval_vqa.yaml\",\n }\n\n\n@registry.register_builder(\"vg_vqa\")\nclass VGVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/aokvqa/defaults.yaml\"}\n\n\n@registry.register_builder(\"gqa\")\nclass GQABuilder(BaseDatasetBuilder):\n train_dataset_cls = GQADataset\n eval_dataset_cls = GQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/gqa/defaults.yaml\",\n \"balanced_val\": \"configs/datasets/gqa/balanced_val.yaml\",\n \"balanced_testdev\": \"configs/datasets/gqa/balanced_testdev.yaml\",\n }","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.vqa_builder.GQABuilder","uri":"program://CREMA/class/lavis.datasets.builders.vqa_builder.GQABuilder#L50-L58","kind":"class","name":"GQABuilder","path":"lavis/datasets/builders/vqa_builder.py","language":"python","start_line":50,"end_line":58,"context_start_line":30,"context_end_line":58,"code":" train_dataset_cls = VGVQADataset\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/vg/defaults_vqa.yaml\"}\n\n\n@registry.register_builder(\"ok_vqa\")\nclass OKVQABuilder(COCOVQABuilder):\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/okvqa/defaults.yaml\",\n }\n\n\n@registry.register_builder(\"aok_vqa\")\nclass AOKVQABuilder(BaseDatasetBuilder):\n train_dataset_cls = AOKVQADataset\n eval_dataset_cls = AOKVQAEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/aokvqa/defaults.yaml\"}\n\n\n@registry.register_builder(\"gqa\")\nclass GQABuilder(BaseDatasetBuilder):\n train_dataset_cls = GQADataset\n eval_dataset_cls = GQAEvalDataset\n\n DATASET_CONFIG_DICT = {\n \"default\": \"configs/datasets/gqa/defaults.yaml\",\n \"balanced_val\": \"configs/datasets/gqa/balanced_val.yaml\",\n \"balanced_testdev\": \"configs/datasets/gqa/balanced_testdev.yaml\",\n }","source_hash":"062a71acca24c81bf1506a65e3951c0dd3f43feb02143a2e34a9e0f4b5a1e79e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.dialogue_builder","uri":"program://CREMA/module/lavis.datasets.builders.dialogue_builder#L1-L21","kind":"module","name":"lavis.datasets.builders.dialogue_builder","path":"lavis/datasets/builders/dialogue_builder.py","language":"python","start_line":1,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.avsd_dialogue_datasets import (\n AVSDDialDataset,\n AVSDDialEvalDataset,\n)\n\n\n@registry.register_builder(\"avsd_dialogue\")\nclass AVSDDialBuilder(BaseDatasetBuilder):\n train_dataset_cls = AVSDDialDataset\n eval_dataset_cls = AVSDDialEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/avsd/defaults_dial.yaml\"}","source_hash":"9b946f21f6fff9ee95b0b3b1ab2cfe48d30e1baf54059bb0d4cee30e9825c0a4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.dialogue_builder.AVSDDialBuilder","uri":"program://CREMA/class/lavis.datasets.builders.dialogue_builder.AVSDDialBuilder#L17-L21","kind":"class","name":"AVSDDialBuilder","path":"lavis/datasets/builders/dialogue_builder.py","language":"python","start_line":17,"end_line":21,"context_start_line":1,"context_end_line":21,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.avsd_dialogue_datasets import (\n AVSDDialDataset,\n AVSDDialEvalDataset,\n)\n\n\n@registry.register_builder(\"avsd_dialogue\")\nclass AVSDDialBuilder(BaseDatasetBuilder):\n train_dataset_cls = AVSDDialDataset\n eval_dataset_cls = AVSDDialEvalDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/avsd/defaults_dial.yaml\"}","source_hash":"9b946f21f6fff9ee95b0b3b1ab2cfe48d30e1baf54059bb0d4cee30e9825c0a4","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.imagefolder_builder","uri":"program://CREMA/module/lavis.datasets.builders.imagefolder_builder#L1-L1061","kind":"module","name":"lavis.datasets.builders.imagefolder_builder","path":"lavis/datasets/builders/imagefolder_builder.py","language":"python","start_line":1,"end_line":1061,"context_start_line":1,"context_end_line":1061,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.imagefolder_dataset import ImageFolderDataset\n\n\n@registry.register_builder(\"imagenet\")\nclass ImageNetBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageFolderDataset\n eval_dataset_cls = ImageFolderDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/imagenet/defaults.yaml\"}\n\n def _download_ann(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in build_info.splits:\n assert split in [\n \"train\",\n \"val\",\n ], \"Invalid split name {}, must be one of 'train', 'val' and 'test'.\"\n\n is_train = split == \"train\"\n\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train\n else self.vis_processors[\"eval\"]\n )\n\n vis_path = os.path.join(vis_info.storage, split)\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n vis_root=vis_path,\n classnames=imagenet_classnames,\n )\n\n return datasets\n\n\nimagenet_classnames = [\n \"tench\",\n \"goldfish\",\n \"great white shark\",\n \"tiger shark\",\n \"hammerhead shark\",\n \"electric ray\",\n \"stingray\",\n \"rooster\",\n \"hen\",\n \"ostrich\",\n \"brambling\",\n \"goldfinch\",\n \"house finch\",\n \"junco\",\n \"indigo bunting\",\n \"American robin\",\n \"bulbul\",\n \"jay\",\n \"magpie\",\n \"chickadee\",\n \"American dipper\",\n \"kite (bird of prey)\",\n \"bald eagle\",\n \"vulture\",\n \"great grey owl\",\n \"fire salamander\",\n \"smooth newt\",\n \"newt\",\n \"spotted salamander\",\n \"axolotl\",\n \"American bullfrog\",\n \"tree frog\",\n \"tailed frog\",\n \"loggerhead sea turtle\",\n \"leatherback sea turtle\",\n \"mud turtle\",\n \"terrapin\",\n \"box turtle\",\n \"banded gecko\",\n \"green iguana\",\n \"Carolina anole\",\n \"desert grassland whiptail lizard\",\n \"agama\",\n \"frilled-necked lizard\",\n \"alligator lizard\",\n \"Gila monster\",\n \"European green lizard\",\n \"chameleon\",\n \"Komodo dragon\",\n \"Nile crocodile\",\n \"American alligator\",\n \"triceratops\",\n \"worm snake\",\n \"ring-necked snake\",\n \"eastern hog-nosed snake\",\n \"smooth green snake\",\n \"kingsnake\",\n \"garter snake\",\n \"water snake\",\n \"vine snake\",\n \"night snake\",\n \"boa constrictor\",\n \"African rock python\",\n \"Indian cobra\",\n \"green mamba\",\n \"sea snake\",\n \"Saharan horned viper\",\n \"eastern diamondback rattlesnake\",\n \"sidewinder rattlesnake\",\n \"trilobite\",\n \"harvestman\",\n \"scorpion\",\n \"yellow garden spider\",\n \"barn spider\",\n \"European garden spider\",\n \"southern black widow\",\n \"tarantula\",\n \"wolf spider\",\n \"tick\",\n \"centipede\",\n \"black grouse\",\n \"ptarmigan\",\n \"ruffed grouse\",\n \"prairie grouse\",\n \"peafowl\",\n \"quail\",\n \"partridge\",\n \"african grey parrot\",\n \"macaw\",\n \"sulphur-crested cockatoo\",\n \"lorikeet\",\n \"coucal\",\n \"bee eater\",\n \"hornbill\",\n \"hummingbird\",\n \"jacamar\",\n \"toucan\",\n \"duck\",\n \"red-breasted merganser\",\n \"goose\",\n \"black swan\",\n \"tusker\",\n \"echidna\",\n \"platypus\",\n \"wallaby\",\n \"koala\",\n \"wombat\",\n \"jellyfish\",\n \"sea anemone\",\n \"brain coral\",\n \"flatworm\",\n \"nematode\",\n \"conch\",\n \"snail\",\n \"slug\",\n \"sea slug\",\n \"chiton\",\n \"chambered nautilus\",\n \"Dungeness crab\",\n \"rock crab\",\n \"fiddler crab\",\n \"red king crab\",\n \"American lobster\",\n \"spiny lobster\",\n \"crayfish\",\n \"hermit crab\",\n \"isopod\",\n \"white stork\",\n \"black stork\",\n \"spoonbill\",\n \"flamingo\",\n \"little blue heron\",\n \"great egret\",\n \"bittern bird\",\n \"crane bird\",\n \"limpkin\",\n \"common gallinule\",\n \"American coot\",\n \"bustard\",\n \"ruddy turnstone\",\n \"dunlin\",\n \"common redshank\",\n \"dowitcher\",\n \"oystercatcher\",\n \"pelican\",\n \"king penguin\",\n \"albatross\",\n \"grey whale\",\n \"killer whale\",\n \"dugong\",\n \"sea lion\",\n \"Chihuahua\",\n \"Japanese Chin\",\n \"Maltese\",\n \"Pekingese\",\n \"Shih Tzu\",\n \"King Charles Spaniel\",\n \"Papillon\",\n \"toy terrier\",\n \"Rhodesian Ridgeback\",\n \"Afghan Hound\",\n \"Basset Hound\",\n \"Beagle\",\n \"Bloodhound\",\n \"Bluetick Coonhound\",\n \"Black and Tan Coonhound\",\n \"Treeing Walker Coonhound\",\n \"English foxhound\",\n \"Redbone Coonhound\",\n \"borzoi\",\n \"Irish Wolfhound\",\n \"Italian Greyhound\",\n \"Whippet\",\n \"Ibizan Hound\",\n \"Norwegian Elkhound\",\n \"Otterhound\",\n \"Saluki\",\n \"Scottish Deerhound\",\n \"Weimaraner\",\n \"Staffordshire Bull Terrier\",\n \"American Staffordshire Terrier\",\n \"Bedlington Terrier\",\n \"Border Terrier\",\n \"Kerry Blue Terrier\",\n \"Irish Terrier\",\n \"Norfolk Terrier\",\n \"Norwich Terrier\",\n \"Yorkshire Terrier\",\n \"Wire Fox Terrier\",\n \"Lakeland Terrier\",\n \"Sealyham Terrier\",\n \"Airedale Terrier\",\n \"Cairn Terrier\",\n \"Australian Terrier\",\n \"Dandie Dinmont Terrier\",\n \"Boston Terrier\",\n \"Miniature Schnauzer\",\n \"Giant Schnauzer\",\n \"Standard Schnauzer\",\n \"Scottish Terrier\",\n \"Tibetan Terrier\",\n \"Australian Silky Terrier\",\n \"Soft-coated Wheaten Terrier\",\n \"West Highland White Terrier\",\n \"Lhasa Apso\",\n \"Flat-Coated Retriever\",\n \"Curly-coated Retriever\",\n \"Golden Retriever\",\n \"Labrador Retriever\",\n \"Chesapeake Bay Retriever\",\n \"German Shorthaired Pointer\",\n \"Vizsla\",\n \"English Setter\",\n \"Irish Setter\",\n \"Gordon Setter\",\n \"Brittany dog\",\n \"Clumber Spaniel\",\n \"English Springer Spaniel\",\n \"Welsh Springer Spaniel\",\n \"Cocker Spaniel\",\n \"Sussex Spaniel\",\n \"Irish Water Spaniel\",\n \"Kuvasz\",\n \"Schipperke\",\n \"Groenendael dog\",\n \"Malinois\",\n \"Briard\",\n \"Australian Kelpie\",\n \"Komondor\",\n \"Old English Sheepdog\",\n \"Shetland Sheepdog\",\n \"collie\",\n \"Border Collie\",\n \"Bouvier des Flandres dog\",\n \"Rottweiler\",\n \"German Shepherd Dog\",\n \"Dobermann\",\n \"Miniature Pinscher\",\n \"Greater Swiss Mountain Dog\",\n \"Bernese Mountain Dog\",\n \"Appenzeller Sennenhund\",\n \"Entlebucher Sennenhund\",\n \"Boxer\",\n \"Bullmastiff\",\n \"Tibetan Mastiff\",\n \"French Bulldog\",\n \"Great Dane\",\n \"St. Bernard\",\n \"husky\",\n \"Alaskan Malamute\",\n \"Siberian Husky\",\n \"Dalmatian\",\n \"Affenpinscher\",\n \"Basenji\",\n \"pug\",\n \"Leonberger\",\n \"Newfoundland dog\",\n \"Great Pyrenees dog\",\n \"Samoyed\",\n \"Pomeranian\",\n \"Chow Chow\",\n \"Keeshond\",\n \"brussels griffon\",\n \"Pembroke Welsh Corgi\",\n \"Cardigan Welsh Corgi\",\n \"Toy Poodle\",\n \"Miniature Poodle\",\n \"Standard Poodle\",\n \"Mexican hairless dog (xoloitzcuintli)\",\n \"grey wolf\",\n \"Alaskan tundra wolf\",\n \"red wolf or maned wolf\",\n \"coyote\",\n \"dingo\",\n \"dhole\",\n \"African wild dog\",\n \"hyena\",\n \"red fox\",\n \"kit fox\",\n \"Arctic fox\",\n \"grey fox\",\n \"tabby cat\",\n \"tiger cat\",\n \"Persian cat\",\n \"Siamese cat\",\n \"Egyptian Mau\",\n \"cougar\",\n \"lynx\",\n \"leopard\",\n \"snow leopard\",\n \"jaguar\",\n \"lion\",\n \"tiger\",\n \"cheetah\",\n \"brown bear\",\n \"American black bear\",\n \"polar bear\",\n \"sloth bear\",\n \"mongoose\",\n \"meerkat\",\n \"tiger beetle\",\n \"ladybug\",\n \"ground beetle\",\n \"longhorn beetle\",\n \"leaf beetle\",\n \"dung beetle\",\n \"rhinoceros beetle\",\n \"weevil\",\n \"fly\",\n \"bee\",\n \"ant\",\n \"grasshopper\",\n \"cricket insect\",\n \"stick insect\",\n \"cockroach\",\n \"praying mantis\",\n \"cicada\",\n \"leafhopper\",\n \"lacewing\",\n \"dragonfly\",\n \"damselfly\",\n \"red admiral butterfly\",\n \"ringlet butterfly\",\n \"monarch butterfly\",\n \"small white butterfly\",\n \"sulphur butterfly\",\n \"gossamer-winged butterfly\",\n \"starfish\",\n \"sea urchin\",\n \"sea cucumber\",\n \"cottontail rabbit\",\n \"hare\",\n \"Angora rabbit\",\n \"hamster\",\n \"porcupine\",\n \"fox squirrel\",\n \"marmot\",\n \"beaver\",\n \"guinea pig\",\n \"common sorrel horse\",\n \"zebra\",\n \"pig\",\n \"wild boar\",\n \"warthog\",\n \"hippopotamus\",\n \"ox\",\n \"water buffalo\",\n \"bison\",\n \"ram (adult male sheep)\",\n \"bighorn sheep\",\n \"Alpine ibex\",\n \"hartebeest\",\n \"impala (antelope)\",\n \"gazelle\",\n \"arabian camel\",\n \"llama\",\n \"weasel\",\n \"mink\",\n \"European polecat\",\n \"black-footed ferret\",\n \"otter\",\n \"skunk\",\n \"badger\",\n \"armadillo\",\n \"three-toed sloth\",\n \"orangutan\",\n \"gorilla\",\n \"chimpanzee\",\n \"gibbon\",\n \"siamang\",\n \"guenon\",\n \"patas monkey\",\n \"baboon\",\n \"macaque\",\n \"langur\",\n \"black-and-white colobus\",\n \"proboscis monkey\",\n \"marmoset\",\n \"white-headed capuchin\",\n \"howler monkey\",\n \"titi monkey\",\n \"Geoffroy's spider monkey\",\n \"common squirrel monkey\",\n \"ring-tailed lemur\",\n \"indri\",\n \"Asian elephant\",\n \"African bush elephant\",\n \"red panda\",\n \"giant panda\",\n \"snoek fish\",\n \"eel\",\n \"silver salmon\",\n \"rock beauty fish\",\n \"clownfish\",\n \"sturgeon\",\n \"gar fish\",\n \"lionfish\",\n \"pufferfish\",\n \"abacus\",\n \"abaya\",\n \"academic gown\",\n \"accordion\",\n \"acoustic guitar\",\n \"aircraft carrier\",\n \"airliner\",\n \"airship\",\n \"altar\",\n \"ambulance\",\n \"amphibious vehicle\",\n \"analog clock\",\n \"apiary\",\n \"apron\",\n \"trash can\",\n \"assault rifle\",\n \"backpack\",\n \"bakery\",\n \"balance beam\",\n \"balloon\",\n \"ballpoint pen\",\n \"Band-Aid\",\n \"banjo\",\n \"baluster / handrail\",\n \"barbell\",\n \"barber chair\",\n \"barbershop\",\n \"barn\",\n \"barometer\",\n \"barrel\",\n \"wheelbarrow\",\n \"baseball\",\n \"basketball\",\n \"bassinet\",\n \"bassoon\",\n \"swimming cap\",\n \"bath towel\",\n \"bathtub\",\n \"station wagon\",\n \"lighthouse\",\n \"beaker\",\n \"military hat (bearskin or shako)\",\n \"beer bottle\",\n \"beer glass\",\n \"bell tower\",\n \"baby bib\",\n \"tandem bicycle\",\n \"bikini\",\n \"ring binder\",\n \"binoculars\",\n \"birdhouse\",\n \"boathouse\",\n \"bobsleigh\",\n \"bolo tie\",\n \"poke bonnet\",\n \"bookcase\",\n \"bookstore\",\n \"bottle cap\",\n \"hunting bow\",\n \"bow tie\",\n \"brass memorial plaque\",\n \"bra\",\n \"breakwater\",\n \"breastplate\",\n \"broom\",\n \"bucket\",\n \"buckle\",\n \"bulletproof vest\",\n \"high-speed train\",\n \"butcher shop\",\n \"taxicab\",\n \"cauldron\",\n \"candle\",\n \"cannon\",\n \"canoe\",\n \"can opener\",\n \"cardigan\",\n \"car mirror\",\n \"carousel\",\n \"tool kit\",\n \"cardboard box / carton\",\n \"car wheel\",\n \"automated teller machine\",\n \"cassette\",\n \"cassette player\",\n \"castle\",\n \"catamaran\",\n \"CD player\",\n \"cello\",\n \"mobile phone\",\n \"chain\",\n \"chain-link fence\",\n \"chain mail\",\n \"chainsaw\",\n \"storage chest\",\n \"chiffonier\",\n \"bell or wind chime\",\n \"china cabinet\",\n \"Christmas stocking\",\n \"church\",\n \"movie theater\",\n \"cleaver\",\n \"cliff dwelling\",\n \"cloak\",\n \"clogs\",\n \"cocktail shaker\",\n \"coffee mug\",\n \"coffeemaker\",\n \"spiral or coil\",\n \"combination lock\",\n \"computer keyboard\",\n \"candy store\",\n \"container ship\",\n \"convertible\",\n \"corkscrew\",\n \"cornet\",\n \"cowboy boot\",\n \"cowboy hat\",\n \"cradle\",\n \"construction crane\",\n \"crash helmet\",\n \"crate\",\n \"infant bed\",\n \"Crock Pot\",\n \"croquet ball\",\n \"crutch\",\n \"cuirass\",\n \"dam\",\n \"desk\",\n \"desktop computer\",\n \"rotary dial telephone\",\n \"diaper\",\n \"digital clock\",\n \"digital watch\",\n \"dining table\",\n \"dishcloth\",\n \"dishwasher\",\n \"disc brake\",\n \"dock\",\n \"dog sled\",\n \"dome\",\n \"doormat\",\n \"drilling rig\",\n \"drum\",\n \"drumstick\",\n \"dumbbell\",\n \"Dutch oven\",\n \"electric fan\",\n \"electric guitar\",\n \"electric locomotive\",\n \"entertainment center\",\n \"envelope\",\n \"espresso machine\",\n \"face powder\",\n \"feather boa\",\n \"filing cabinet\",\n \"fireboat\",\n \"fire truck\",\n \"fire screen\",\n \"flagpole\",\n \"flute\",\n \"folding chair\",\n \"football helmet\",\n \"forklift\",\n \"fountain\",\n \"fountain pen\",\n \"four-poster bed\",\n \"freight car\",\n \"French horn\",\n \"frying pan\",\n \"fur coat\",\n \"garbage truck\",\n \"gas mask or respirator\",\n \"gas pump\",\n \"goblet\",\n \"go-kart\",\n \"golf ball\",\n \"golf cart\",\n \"gondola\",\n \"gong\",\n \"gown\",\n \"grand piano\",\n \"greenhouse\",\n \"radiator grille\",\n \"grocery store\",\n \"guillotine\",\n \"hair clip\",\n \"hair spray\",\n \"half-track\",\n \"hammer\",\n \"hamper\",\n \"hair dryer\",\n \"hand-held computer\",\n \"handkerchief\",\n \"hard disk drive\",\n \"harmonica\",\n \"harp\",\n \"combine harvester\",\n \"hatchet\",\n \"holster\",\n \"home theater\",\n \"honeycomb\",\n \"hook\",\n \"hoop skirt\",\n \"gymnastic horizontal bar\",\n \"horse-drawn vehicle\",\n \"hourglass\",\n \"iPod\",\n \"clothes iron\",\n \"carved pumpkin\",\n \"jeans\",\n \"jeep\",\n \"T-shirt\",\n \"jigsaw puzzle\",\n \"rickshaw\",\n \"joystick\",\n \"kimono\",\n \"knee pad\",\n \"knot\",\n \"lab coat\",\n \"ladle\",\n \"lampshade\",\n \"laptop computer\",\n \"lawn mower\",\n \"lens cap\",\n \"letter opener\",\n \"library\",\n \"lifeboat\",\n \"lighter\",\n \"limousine\",\n \"ocean liner\",\n \"lipstick\",\n \"slip-on shoe\",\n \"lotion\",\n \"music speaker\",\n \"loupe magnifying glass\",\n \"sawmill\",\n \"magnetic compass\",\n \"messenger bag\",\n \"mailbox\",\n \"tights\",\n \"one-piece bathing suit\",\n \"manhole cover\",\n \"maraca\",\n \"marimba\",\n \"mask\",\n \"matchstick\",\n \"maypole\",\n \"maze\",\n \"measuring cup\",\n \"medicine cabinet\",\n \"megalith\",\n \"microphone\",\n \"microwave oven\",\n \"military uniform\",\n \"milk can\",\n \"minibus\",\n \"miniskirt\",\n \"minivan\",\n \"missile\",\n \"mitten\",\n \"mixing bowl\",\n \"mobile home\",\n \"ford model t\",\n \"modem\",\n \"monastery\",\n \"monitor\",\n \"moped\",\n \"mortar and pestle\",\n \"graduation cap\",\n \"mosque\",\n \"mosquito net\",\n \"vespa\",\n \"mountain bike\",\n \"tent\",\n \"computer mouse\",\n \"mousetrap\",\n \"moving van\",\n \"muzzle\",\n \"metal nail\",\n \"neck brace\",\n \"necklace\",\n \"baby pacifier\",\n \"notebook computer\",\n \"obelisk\",\n \"oboe\",\n \"ocarina\",\n \"odometer\",\n \"oil filter\",\n \"pipe organ\",\n \"oscilloscope\",\n \"overskirt\",\n \"bullock cart\",\n \"oxygen mask\",\n \"product packet / packaging\",\n \"paddle\",\n \"paddle wheel\",\n \"padlock\",\n \"paintbrush\",\n \"pajamas\",\n \"palace\",\n \"pan flute\",\n \"paper towel\",\n \"parachute\",\n \"parallel bars\",\n \"park bench\",\n \"parking meter\",\n \"railroad car\",\n \"patio\",\n \"payphone\",\n \"pedestal\",\n \"pencil case\",\n \"pencil sharpener\",\n \"perfume\",\n \"Petri dish\",\n \"photocopier\",\n \"plectrum\",\n \"Pickelhaube\",\n \"picket fence\",\n \"pickup truck\",\n \"pier\",\n \"piggy bank\",\n \"pill bottle\",\n \"pillow\",\n \"ping-pong ball\",\n \"pinwheel\",\n \"pirate ship\",\n \"drink pitcher\",\n \"block plane\",\n \"planetarium\",\n \"plastic bag\",\n \"plate rack\",\n \"farm plow\",\n \"plunger\",\n \"Polaroid camera\",\n \"pole\",\n \"police van\",\n \"poncho\",\n \"pool table\",\n \"soda bottle\",\n \"plant pot\",\n \"potter's wheel\",\n \"power drill\",\n \"prayer rug\",\n \"printer\",\n \"prison\",\n \"missile\",\n \"projector\",\n \"hockey puck\",\n \"punching bag\",\n \"purse\",\n \"quill\",\n \"quilt\",\n \"race car\",\n \"racket\",\n \"radiator\",\n \"radio\",\n \"radio telescope\",\n \"rain barrel\",\n \"recreational vehicle\",\n \"fishing casting reel\",\n \"reflex camera\",\n \"refrigerator\",\n \"remote control\",\n \"restaurant\",\n \"revolver\",\n \"rifle\",\n \"rocking chair\",\n \"rotisserie\",\n \"eraser\",\n \"rugby ball\",\n \"ruler measuring stick\",\n \"sneaker\",\n \"safe\",\n \"safety pin\",\n \"salt shaker\",\n \"sandal\",\n \"sarong\",\n \"saxophone\",\n \"scabbard\",\n \"weighing scale\",\n \"school bus\",\n \"schooner\",\n \"scoreboard\",\n \"CRT monitor\",\n \"screw\",\n \"screwdriver\",\n \"seat belt\",\n \"sewing machine\",\n \"shield\",\n \"shoe store\",\n \"shoji screen / room divider\",\n \"shopping basket\",\n \"shopping cart\",\n \"shovel\",\n \"shower cap\",\n \"shower curtain\",\n \"ski\",\n \"balaclava ski mask\",\n \"sleeping bag\",\n \"slide rule\",\n \"sliding door\",\n \"slot machine\",\n \"snorkel\",\n \"snowmobile\",\n \"snowplow\",\n \"soap dispenser\",\n \"soccer ball\",\n \"sock\",\n \"solar thermal collector\",\n \"sombrero\",\n \"soup bowl\",\n \"keyboard space bar\",\n \"space heater\",\n \"space shuttle\",\n \"spatula\",\n \"motorboat\",\n \"spider web\",\n \"spindle\",\n \"sports car\",\n \"spotlight\",\n \"stage\",\n \"steam locomotive\",\n \"through arch bridge\",\n \"steel drum\",\n \"stethoscope\",\n \"scarf\",\n \"stone wall\",\n \"stopwatch\",\n \"stove\",\n \"strainer\",\n \"tram\",\n \"stretcher\",\n \"couch\",\n \"stupa\",\n \"submarine\",\n \"suit\",\n \"sundial\",\n \"sunglasses\",\n \"sunglasses\",\n \"sunscreen\",\n \"suspension bridge\",\n \"mop\",\n \"sweatshirt\",\n \"swim trunks / shorts\",\n \"swing\",\n \"electrical switch\",\n \"syringe\",\n \"table lamp\",\n \"tank\",\n \"tape player\",\n \"teapot\",\n \"teddy bear\",\n \"television\",\n \"tennis ball\",\n \"thatched roof\",\n \"front curtain\",\n \"thimble\",\n \"threshing machine\",\n \"throne\",\n \"tile roof\",\n \"toaster\",\n \"tobacco shop\",\n \"toilet seat\",\n \"torch\",\n \"totem pole\",\n \"tow truck\",\n \"toy store\",\n \"tractor\",\n \"semi-trailer truck\",\n \"tray\",\n \"trench coat\",\n \"tricycle\",\n \"trimaran\",\n \"tripod\",\n \"triumphal arch\",\n \"trolleybus\",\n \"trombone\",\n \"hot tub\",\n \"turnstile\",\n \"typewriter keyboard\",\n \"umbrella\",\n \"unicycle\",\n \"upright piano\",\n \"vacuum cleaner\",\n \"vase\",\n \"vaulted or arched ceiling\",\n \"velvet fabric\",\n \"vending machine\",\n \"vestment\",\n \"viaduct\",\n \"violin\",\n \"volleyball\",\n \"waffle iron\",\n \"wall clock\",\n \"wallet\",\n \"wardrobe\",\n \"military aircraft\",\n \"sink\",\n \"washing machine\",\n \"water bottle\",\n \"water jug\",\n \"water tower\",\n \"whiskey jug\",\n \"whistle\",\n \"hair wig\",\n \"window screen\",\n \"window shade\",\n \"Windsor tie\",\n \"wine bottle\",\n \"airplane wing\",\n \"wok\",\n \"wooden spoon\",\n \"wool\",\n \"split-rail fence\",\n \"shipwreck\",\n \"sailboat\",\n \"yurt\",\n \"website\",\n \"comic book\",\n \"crossword\",\n \"traffic or street sign\",\n \"traffic light\",\n \"dust jacket\",\n \"menu\",\n \"plate\",\n \"guacamole\",\n \"consomme\",\n \"hot pot\",\n \"trifle\",\n \"ice cream\",\n \"popsicle\",\n \"baguette\",\n \"bagel\",\n \"pretzel\",\n \"cheeseburger\",\n \"hot dog\",\n \"mashed potatoes\",\n \"cabbage\",\n \"broccoli\",\n \"cauliflower\",\n \"zucchini\",\n \"spaghetti squash\",\n \"acorn squash\",\n \"butternut squash\",\n \"cucumber\",\n \"artichoke\",\n \"bell pepper\",\n \"cardoon\",\n \"mushroom\",\n \"Granny Smith apple\",\n \"strawberry\",\n \"orange\",\n \"lemon\",\n \"fig\",\n \"pineapple\",\n \"banana\",\n \"jackfruit\",\n \"cherimoya (custard apple)\",\n \"pomegranate\",\n \"hay\",\n \"carbonara\",\n \"chocolate syrup\",\n \"dough\",\n \"meatloaf\",\n \"pizza\",\n \"pot pie\",\n \"burrito\",\n \"red wine\",\n \"espresso\",\n \"tea cup\",\n \"eggnog\",\n# ... truncated ...","source_hash":"fe061ecdefa5455c7e7d574223dafa753028d63cad312d733c625ebcf5472cdb","truncated":true} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.imagefolder_builder.ImageNetBuilder","uri":"program://CREMA/class/lavis.datasets.builders.imagefolder_builder.ImageNetBuilder#L16-L57","kind":"class","name":"ImageNetBuilder","path":"lavis/datasets/builders/imagefolder_builder.py","language":"python","start_line":16,"end_line":57,"context_start_line":1,"context_end_line":77,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.imagefolder_dataset import ImageFolderDataset\n\n\n@registry.register_builder(\"imagenet\")\nclass ImageNetBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageFolderDataset\n eval_dataset_cls = ImageFolderDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/imagenet/defaults.yaml\"}\n\n def _download_ann(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in build_info.splits:\n assert split in [\n \"train\",\n \"val\",\n ], \"Invalid split name {}, must be one of 'train', 'val' and 'test'.\"\n\n is_train = split == \"train\"\n\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train\n else self.vis_processors[\"eval\"]\n )\n\n vis_path = os.path.join(vis_info.storage, split)\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n vis_root=vis_path,\n classnames=imagenet_classnames,\n )\n\n return datasets\n\n\nimagenet_classnames = [\n \"tench\",\n \"goldfish\",\n \"great white shark\",\n \"tiger shark\",\n \"hammerhead shark\",\n \"electric ray\",\n \"stingray\",\n \"rooster\",\n \"hen\",\n \"ostrich\",\n \"brambling\",\n \"goldfinch\",\n \"house finch\",\n \"junco\",\n \"indigo bunting\",\n \"American robin\",\n \"bulbul\",","source_hash":"fe061ecdefa5455c7e7d574223dafa753028d63cad312d733c625ebcf5472cdb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.imagefolder_builder._download_ann","uri":"program://CREMA/function/lavis.datasets.builders.imagefolder_builder._download_ann#L22-L23","kind":"function","name":"_download_ann","path":"lavis/datasets/builders/imagefolder_builder.py","language":"python","start_line":22,"end_line":23,"context_start_line":2,"context_end_line":43,"code":" Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.imagefolder_dataset import ImageFolderDataset\n\n\n@registry.register_builder(\"imagenet\")\nclass ImageNetBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageFolderDataset\n eval_dataset_cls = ImageFolderDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/imagenet/defaults.yaml\"}\n\n def _download_ann(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in build_info.splits:\n assert split in [\n \"train\",\n \"val\",\n ], \"Invalid split name {}, must be one of 'train', 'val' and 'test'.\"\n\n is_train = split == \"train\"\n\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train","source_hash":"fe061ecdefa5455c7e7d574223dafa753028d63cad312d733c625ebcf5472cdb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.datasets.builders.imagefolder_builder.build","uri":"program://CREMA/function/lavis.datasets.builders.imagefolder_builder.build#L25-L57","kind":"function","name":"build","path":"lavis/datasets/builders/imagefolder_builder.py","language":"python","start_line":25,"end_line":57,"context_start_line":5,"context_end_line":77,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport os\n\nfrom lavis.common.registry import registry\nfrom lavis.datasets.builders.base_dataset_builder import BaseDatasetBuilder\nfrom lavis.datasets.datasets.imagefolder_dataset import ImageFolderDataset\n\n\n@registry.register_builder(\"imagenet\")\nclass ImageNetBuilder(BaseDatasetBuilder):\n train_dataset_cls = ImageFolderDataset\n eval_dataset_cls = ImageFolderDataset\n\n DATASET_CONFIG_DICT = {\"default\": \"configs/datasets/imagenet/defaults.yaml\"}\n\n def _download_ann(self):\n pass\n\n def build(self):\n self.build_processors()\n\n build_info = self.config.build_info\n\n vis_info = build_info.get(self.data_type)\n\n datasets = dict()\n for split in build_info.splits:\n assert split in [\n \"train\",\n \"val\",\n ], \"Invalid split name {}, must be one of 'train', 'val' and 'test'.\"\n\n is_train = split == \"train\"\n\n vis_processor = (\n self.vis_processors[\"train\"]\n if is_train\n else self.vis_processors[\"eval\"]\n )\n\n vis_path = os.path.join(vis_info.storage, split)\n\n # create datasets\n dataset_cls = self.train_dataset_cls if is_train else self.eval_dataset_cls\n datasets[split] = dataset_cls(\n vis_processor=vis_processor,\n vis_root=vis_path,\n classnames=imagenet_classnames,\n )\n\n return datasets\n\n\nimagenet_classnames = [\n \"tench\",\n \"goldfish\",\n \"great white shark\",\n \"tiger shark\",\n \"hammerhead shark\",\n \"electric ray\",\n \"stingray\",\n \"rooster\",\n \"hen\",\n \"ostrich\",\n \"brambling\",\n \"goldfinch\",\n \"house finch\",\n \"junco\",\n \"indigo bunting\",\n \"American robin\",\n \"bulbul\",","source_hash":"fe061ecdefa5455c7e7d574223dafa753028d63cad312d733c625ebcf5472cdb","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue","uri":"program://CREMA/module/lavis.tasks.dialogue#L1-L127","kind":"module","name":"lavis.tasks.dialogue","path":"lavis/tasks/dialogue.py","language":"python","start_line":1,"end_line":127,"context_start_line":1,"context_end_line":127,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport numpy as np\n\n\n@registry.register_task(\"dialogue\")\nclass DialogueTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_dialogue_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_dialogue_eval(coco_gt_root, results_file, split):\n\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])\n\n # create coco object and coco_result object\n coco = COCO(annotation_file)\n coco_result = coco.loadRes(results_file)\n\n # create coco_eval object by taking coco and coco_result\n coco_eval = COCOEvalCap(coco, coco_result)\n\n # evaluate on a subset of images by setting\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n # please remove this line when evaluating the full validation set\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n\n # evaluate results\n # SPICE will take a few minutes the first time, but speeds up due to caching\n coco_eval.evaluate()\n\n # print output evaluation scores\n for metric, score in coco_eval.eval.items():\n print(f\"{metric}: {score:.3f}\")\n\n return coco_eval","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.DialogueTask","uri":"program://CREMA/class/lavis.tasks.dialogue.DialogueTask#L21-L84","kind":"class","name":"DialogueTask","path":"lavis/tasks/dialogue.py","language":"python","start_line":21,"end_line":84,"context_start_line":1,"context_end_line":104,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport numpy as np\n\n\n@registry.register_task(\"dialogue\")\nclass DialogueTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_dialogue_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_dialogue_eval(coco_gt_root, results_file, split):\n\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.coco_dialogue_eval","uri":"program://CREMA/function/lavis.tasks.dialogue.coco_dialogue_eval#L93-L127","kind":"function","name":"coco_dialogue_eval","path":"lavis/tasks/dialogue.py","language":"python","start_line":93,"end_line":127,"context_start_line":73,"context_end_line":127,"code":" agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_dialogue_eval(coco_gt_root, results_file, split):\n\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])\n\n # create coco object and coco_result object\n coco = COCO(annotation_file)\n coco_result = coco.loadRes(results_file)\n\n # create coco_eval object by taking coco and coco_result\n coco_eval = COCOEvalCap(coco, coco_result)\n\n # evaluate on a subset of images by setting\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n # please remove this line when evaluating the full validation set\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n\n # evaluate results\n # SPICE will take a few minutes the first time, but speeds up due to caching\n coco_eval.evaluate()\n\n # print output evaluation scores\n for metric, score in coco_eval.eval.items():\n print(f\"{metric}: {score:.3f}\")\n\n return coco_eval","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.__init__","uri":"program://CREMA/function/lavis.tasks.dialogue.__init__#L22-L30","kind":"function","name":"__init__","path":"lavis/tasks/dialogue.py","language":"python","start_line":22,"end_line":30,"context_start_line":2,"context_end_line":50,"code":" Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.logger import MetricLogger\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport numpy as np\n\n\n@registry.register_task(\"dialogue\")\nclass DialogueTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.setup_task","uri":"program://CREMA/function/lavis.tasks.dialogue.setup_task#L33-L49","kind":"function","name":"setup_task","path":"lavis/tasks/dialogue.py","language":"python","start_line":33,"end_line":49,"context_start_line":13,"context_end_line":69,"code":"from lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport numpy as np\n\n\n@registry.register_task(\"dialogue\")\nclass DialogueTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.valid_step","uri":"program://CREMA/function/lavis.tasks.dialogue.valid_step#L51-L55","kind":"function","name":"valid_step","path":"lavis/tasks/dialogue.py","language":"python","start_line":51,"end_line":55,"context_start_line":31,"context_end_line":75,"code":"\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_dialogue_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue.after_evaluation","uri":"program://CREMA/function/lavis.tasks.dialogue.after_evaluation#L57-L65","kind":"function","name":"after_evaluation","path":"lavis/tasks/dialogue.py","language":"python","start_line":57,"end_line":65,"context_start_line":37,"context_end_line":85,"code":" max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_dialogue_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.dialogue._report_metrics","uri":"program://CREMA/function/lavis.tasks.dialogue._report_metrics#L68-L84","kind":"function","name":"_report_metrics","path":"lavis/tasks/dialogue.py","language":"python","start_line":68,"end_line":84,"context_start_line":48,"context_end_line":104,"code":" report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n loss = model(samples)[\"loss\"].item()\n\n return [loss]\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n\n if self.report_metric:\n avg_loss = np.mean(val_result)\n metrics = {\"agg_metrics\": avg_loss}\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_dialogue_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_dialogue_eval(coco_gt_root, results_file, split):\n\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)","source_hash":"514ed9b2b3fc033cfd12858643edb8e619a96539b154f7491098038aa916e20f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa","uri":"program://CREMA/module/lavis.tasks.vqa#L1-L611","kind":"module","name":"lavis.tasks.vqa","path":"lavis/tasks/vqa.py","language":"python","start_line":1,"end_line":611,"context_start_line":1,"context_end_line":611,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport json\nimport os\nimport torch\nimport numpy as np\nimport random\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.registry import registry\nfrom lavis.common.vqa_tools.vqa import VQA\nfrom lavis.common.vqa_tools.vqa_eval import VQAEval\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.common.dist_utils import main_process\n\nnumber_mapping = {'1': 'one', '2': 'two', '3': 'three', '4': 'four', '5': 'five', '6':'six', '7':'seven', '8':'eight', '9':'nine', '10':'ten'}\n\n@registry.register_task(\"vqa\")\nclass VQATask(BaseTask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n prompt=\"\",\n ):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n\n self.evaluate = evaluate\n self.inference_method = inference_method\n self.num_ans_candidates = num_ans_candidates\n self.prompt = prompt\n\n self.answer_list = None\n\n self.ques_files = dict()\n self.anno_files = dict()\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n prompt = run_cfg.get(\"prompt\", \"\")\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n prompt=prompt,\n )\n\n def build_datasets(self, cfg):\n datasets = super().build_datasets(cfg)\n\n # get question file, annotation file and anwser list in COCO format\n for dataset in datasets.values():\n for split in dataset:\n if (\n hasattr(dataset[split], \"coco_fmt_qust_file\")\n and dataset[split].coco_fmt_qust_file is not None\n ):\n self.ques_files[split] = dataset[split].coco_fmt_qust_file\n self.anno_files[split] = dataset[split].coco_fmt_anno_file\n\n try:\n self.answer_list = dataset[split].answer_list\n except AttributeError:\n # if answer_list is not provided, then set it to None\n pass\n\n if len(self.ques_files) > 0:\n assert len(self.ques_files) == len(\n self.anno_files\n ), \"Only support one split for evaluation.\"\n\n return datasets\n\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n for answer, ques_id in zip(answers, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n\n return pred_qa_pairs\n\n def after_evaluation(self, val_result, split_name, **kwargs):\n result_file = self.save_result(\n val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n Use official VQA evaluation script to report metrics.\n \"\"\"\n metrics = {}\n\n if split in self.ques_files and split in self.anno_files:\n vqa = VQA(self.anno_files[split], self.ques_files[split])\n vqa_result = vqa.loadRes(\n resFile=result_file, quesFile=self.ques_files[split]\n )\n\n # create vqaEval object by taking vqa and vqaRes\n # n is precision of accuracy (number of places after decimal), default is 2\n vqa_scorer = VQAEval(vqa, vqa_result, n=2)\n logging.info(\"Start VQA evaluation.\")\n vqa_scorer.evaluate()\n\n # print accuracies\n overall_acc = vqa_scorer.accuracy[\"overall\"]\n metrics[\"agg_metrics\"] = overall_acc\n\n logging.info(\"Overall Accuracy is: %.02f\\n\" % overall_acc)\n logging.info(\"Per Answer Type Accuracy is the following:\")\n\n for ans_type in vqa_scorer.accuracy[\"perAnswerType\"]:\n logging.info(\n \"%s : %.02f\"\n % (ans_type, vqa_scorer.accuracy[\"perAnswerType\"][ans_type])\n )\n metrics[ans_type] = vqa_scorer.accuracy[\"perAnswerType\"][ans_type]\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n return metrics\n\n@registry.register_task(\"gqa\")\nclass GQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n\n # print('gt_answers', gt_answers)\n # print('answers', answers)\n \n for answer, ques_id, gt_answer in zip(answers, question_id, gt_answers):\n ques_id = int(ques_id.item())\n if answer in number_mapping:\n answer = number_mapping[answer]\n if answer in gt_answer: # guitar -> acoustic_guitar, in ---> indoor for zs test\n answer = gt_answer\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": answer, \"gt_ans\": gt_answer})\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQAEval()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n\n # if self.inference_method == \"generate\":\n # pred = vqa_tool.processPunctuation(pred)\n # pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n \n\n@registry.register_task(\"aok_vqa\")\nclass AOKVQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"direct_answers\"]\n\n for pred_answer, ques_id, gt_answer in zip(answers, question_id, gt_answers):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer}\n )\n\n return pred_qa_pairs\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n Implementing accuracy computation for AOKVQA, see\n https://github.com/allenai/aokvqa/blob/main/evaluation/eval_predictions.py#L45 for details.\n \"\"\"\n # TODO add evaluation for multi-choice\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n pred = res[\"pred_ans\"]\n gt_ans = res[\"gt_ans\"]\n\n num_match = sum([pred == gt for gt in gt_ans])\n vqa_acc = min(1.0, num_match / 3.0)\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n\n [TODO] add support for multi-choice.\n \"\"\"\n result_leaderboard = dict()\n for res in results:\n result_leaderboard[res[\"question_id\"]] = {\n \"direct_answer\": res[\"pred_ans\"],\n \"multiple_choice\": \"\",\n }\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")\n \n\n \n\n@registry.register_task(\"videoqa\")\nclass VideoQA(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n\n answer = outputs[\"answer\"]\n qid = outputs[\"qid\"]\n output_text = outputs['output_text']\n # need_extra = outputs['need_extra_info']\n # print('need_extra', need_extra)\n if 'frame_idx' in outputs:\n frame_idx = outputs['frame_idx']\n else:\n frame_idx = [0 for i in range(len(qid))]\n \n # if 'dis_acc' in outputs:\n # dis_acc = outputs['dis_acc']\n # else:\n # dis_acc = [0 for i in range(len(qid))]\n\n # print(qid)\n # print(len(output_text), output_text)\n assert len(qid)==len(output_text)\n assert len(qid)==len(answer) \n \n for a, q, o, f, d in zip(answer, qid, output_text, frame_idx, dis_acc):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q,\n \"prediction\": o,\n \"target\": self.ANS_MAPPING[a[-1]],\n \"frame_idx\": f,\n # \"difficult\": n,\n 'dis_acc': d\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch)\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n total_num = len(results)\n acc = 0\n qtype_correct_dict = {}\n qtype_total_dict = {}\n\n # difficult_acc, difficult_total = 0, 0\n dis_acc = 0\n for r in results: \n qtype = r['qid'].split('_')[0]\n dis_acc += r['dis_acc']\n if qtype not in qtype_total_dict:\n qtype_total_dict[qtype] = 1\n else:\n qtype_total_dict[qtype] += 1 \n # if r['difficult'] == 1:\n # difficult_total += 1\n\n if r['prediction'] == r['target']:\n acc += 1\n if qtype not in qtype_correct_dict:\n qtype_correct_dict[qtype] = 1\n else:\n qtype_correct_dict[qtype] += 1 \n\n # if r['difficult'] == 1:\n # difficult_acc += 1\n \n metrics = {\"agg_metrics\": acc/total_num , 'total':total_num}\n \n for qtype in qtype_total_dict:\n metrics[qtype] = qtype_correct_dict[qtype] / qtype_total_dict[qtype] * 100\n \n # metrics['dis_acc'] = dis_acc/total_num*100\n \n # metrics['difficult_num'] = difficult_total\n # if difficult_total == 0:\n # metrics['difficult_acc'] = 0\n # else:\n # metrics['difficult_acc'] = difficult_acc / difficult_total * 100\n\n # for STAR\n if ('Interaction' in metrics) and ('Sequence' in metrics) and ('Prediction' in metrics) and ('Feasibility' in metrics):\n metrics[\"agg_metrics\"] = (metrics['Interaction'] + metrics['Sequence'] + metrics['Prediction'] + metrics['Feasibility']) / 4\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n@registry.register_task(\"moment_retrieval\")\nclass MR(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'no': 0, 'yes': 1}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n answer = outputs['answer']\n qid = outputs['qid']\n score = outputs['yes_score']\n pred = outputs['pred_ans']\n assert len(qid)==len(answer)\n assert len(qid)==len(score)\n assert len(qid)==len(pred) \n \n i = 0\n for a, q, s, p in zip(answer, qid, score, pred):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q + '_' + str(i),\n \"prediction\": p,\n \"target\": self.ANS_MAPPING[a],\n 'score': s\n }\n )\n i += 1\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch)\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n total_num = len(results)\n acc = 0\n for r in results:\n if r['prediction'] == r['target']:\n acc += 1\n metrics = {\"agg_metrics\": acc / total_num, 'total': total_num}\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n\n\n\n \n\ndef evals_json(gold_data, preds):\n score_list = ['Top1 (EM)']\n score = {s:[] for s in score_list}\n \n for ref, pred in zip(gold_data, preds):\n if pred in ref:\n score['Top1 (EM)'].append(1)\n else:\n #scores=[tokens_unigram_f_value(answer,ref) for ref in ref_answers]\n score['Top1 (EM)'].append(0)\n \n rlt={}\n for k,v in score.items():\n assert len(v)==len(gold_data),len(v)\n rlt[k]=np.mean(v)*100\n \n return rlt\n\n# calculate the Top1(EM) for each question type\nSQA3D_QT=['All', 'What', 'How', 'Can', 'Is', 'Which', 'Other']\n@registry.register_task(\"sqa3d\")\nclass ThreeDVQATask(VQATask):\n \n def valid_step(self, model, samples):\n\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n pred_anwers = answers\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n questions = samples[\"qa_input\"]\n\n for pred_answer, ques_id, gt_answer, q in zip(pred_anwers, question_id, gt_answers, questions):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer, 'question': q}\n )\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n # for sqa3d-test: it is fine\n # for sqa3d-val: some 3d scene features are missing, need to filter the data/gt\n results = json.load(open(result_file, \"r\"))\n preds_={k:[] for k in SQA3D_QT}\n golds_={k:[] for k in SQA3D_QT}\n\n metrics = {}\n\n for data in results:\n qid = data['question_id']\n qtype = qid.split('_')[0]\n pred = data['pred_ans']\n gt = data['gt_ans']\n # qtype = self.qclass1(pred['question'])\n pred = pred.replace(\",\", '').replace(\"\", '').replace(\"\", \"\").strip()\n pred = pred.strip().split('\\n')[0]\n pred = pred.lower()\n if len(gt.split(' ')) == 1:\n pred = pred.split(' ')[-1]\n if pred in number_mapping:\n pred = number_mapping[pred]\n preds_[qtype].append(pred)\n golds_[qtype].append(gt)\n\n preds_['All'].append(pred)\n golds_['All'].append(gt)\n\n for qt in SQA3D_QT:\n score = evals_json(golds_[qt], preds_[qt])\n if qt == 'All':\n metrics['agg_metrics'] = score['Top1 (EM)']\n else:\n metrics[qt] = score['Top1 (EM)']\n \n with open(os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\") as f:\n f.write(json.dumps(metrics) + \"\\n\")\n logging.info(metrics)\n \n return metrics\n\n\n\n \n\n ","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.VQATask","uri":"program://CREMA/class/lavis.tasks.vqa.VQATask#L25-L172","kind":"class","name":"VQATask","path":"lavis/tasks/vqa.py","language":"python","start_line":25,"end_line":172,"context_start_line":5,"context_end_line":192,"code":" For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport json\nimport os\nimport torch\nimport numpy as np\nimport random\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.registry import registry\nfrom lavis.common.vqa_tools.vqa import VQA\nfrom lavis.common.vqa_tools.vqa_eval import VQAEval\nfrom lavis.tasks.base_task import BaseTask\nfrom lavis.common.dist_utils import main_process\n\nnumber_mapping = {'1': 'one', '2': 'two', '3': 'three', '4': 'four', '5': 'five', '6':'six', '7':'seven', '8':'eight', '9':'nine', '10':'ten'}\n\n@registry.register_task(\"vqa\")\nclass VQATask(BaseTask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n prompt=\"\",\n ):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n\n self.evaluate = evaluate\n self.inference_method = inference_method\n self.num_ans_candidates = num_ans_candidates\n self.prompt = prompt\n\n self.answer_list = None\n\n self.ques_files = dict()\n self.anno_files = dict()\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n prompt = run_cfg.get(\"prompt\", \"\")\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n prompt=prompt,\n )\n\n def build_datasets(self, cfg):\n datasets = super().build_datasets(cfg)\n\n # get question file, annotation file and anwser list in COCO format\n for dataset in datasets.values():\n for split in dataset:\n if (\n hasattr(dataset[split], \"coco_fmt_qust_file\")\n and dataset[split].coco_fmt_qust_file is not None\n ):\n self.ques_files[split] = dataset[split].coco_fmt_qust_file\n self.anno_files[split] = dataset[split].coco_fmt_anno_file\n\n try:\n self.answer_list = dataset[split].answer_list\n except AttributeError:\n # if answer_list is not provided, then set it to None\n pass\n\n if len(self.ques_files) > 0:\n assert len(self.ques_files) == len(\n self.anno_files\n ), \"Only support one split for evaluation.\"\n\n return datasets\n\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n for answer, ques_id in zip(answers, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n\n return pred_qa_pairs\n\n def after_evaluation(self, val_result, split_name, **kwargs):\n result_file = self.save_result(\n val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n Use official VQA evaluation script to report metrics.\n \"\"\"\n metrics = {}\n\n if split in self.ques_files and split in self.anno_files:\n vqa = VQA(self.anno_files[split], self.ques_files[split])\n vqa_result = vqa.loadRes(\n resFile=result_file, quesFile=self.ques_files[split]\n )\n\n # create vqaEval object by taking vqa and vqaRes\n # n is precision of accuracy (number of places after decimal), default is 2\n vqa_scorer = VQAEval(vqa, vqa_result, n=2)\n logging.info(\"Start VQA evaluation.\")\n vqa_scorer.evaluate()\n\n # print accuracies\n overall_acc = vqa_scorer.accuracy[\"overall\"]\n metrics[\"agg_metrics\"] = overall_acc\n\n logging.info(\"Overall Accuracy is: %.02f\\n\" % overall_acc)\n logging.info(\"Per Answer Type Accuracy is the following:\")\n\n for ans_type in vqa_scorer.accuracy[\"perAnswerType\"]:\n logging.info(\n \"%s : %.02f\"\n % (ans_type, vqa_scorer.accuracy[\"perAnswerType\"][ans_type])\n )\n metrics[ans_type] = vqa_scorer.accuracy[\"perAnswerType\"][ans_type]\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n return metrics\n\n@registry.register_task(\"gqa\")\nclass GQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n\n # print('gt_answers', gt_answers)","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.GQATask","uri":"program://CREMA/class/lavis.tasks.vqa.GQATask#L175-L242","kind":"class","name":"GQATask","path":"lavis/tasks/vqa.py","language":"python","start_line":175,"end_line":242,"context_start_line":155,"context_end_line":262,"code":" metrics[\"agg_metrics\"] = overall_acc\n\n logging.info(\"Overall Accuracy is: %.02f\\n\" % overall_acc)\n logging.info(\"Per Answer Type Accuracy is the following:\")\n\n for ans_type in vqa_scorer.accuracy[\"perAnswerType\"]:\n logging.info(\n \"%s : %.02f\"\n % (ans_type, vqa_scorer.accuracy[\"perAnswerType\"][ans_type])\n )\n metrics[ans_type] = vqa_scorer.accuracy[\"perAnswerType\"][ans_type]\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n return metrics\n\n@registry.register_task(\"gqa\")\nclass GQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n\n # print('gt_answers', gt_answers)\n # print('answers', answers)\n \n for answer, ques_id, gt_answer in zip(answers, question_id, gt_answers):\n ques_id = int(ques_id.item())\n if answer in number_mapping:\n answer = number_mapping[answer]\n if answer in gt_answer: # guitar -> acoustic_guitar, in ---> indoor for zs test\n answer = gt_answer\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": answer, \"gt_ans\": gt_answer})\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQAEval()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n\n # if self.inference_method == \"generate\":\n # pred = vqa_tool.processPunctuation(pred)\n # pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n \n\n@registry.register_task(\"aok_vqa\")\nclass AOKVQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"direct_answers\"]\n","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.AOKVQATask","uri":"program://CREMA/class/lavis.tasks.vqa.AOKVQATask#L246-L326","kind":"class","name":"AOKVQATask","path":"lavis/tasks/vqa.py","language":"python","start_line":246,"end_line":326,"context_start_line":226,"context_end_line":346,"code":" # pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n \n\n@registry.register_task(\"aok_vqa\")\nclass AOKVQATask(VQATask):\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"direct_answers\"]\n\n for pred_answer, ques_id, gt_answer in zip(answers, question_id, gt_answers):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer}\n )\n\n return pred_qa_pairs\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n Implementing accuracy computation for AOKVQA, see\n https://github.com/allenai/aokvqa/blob/main/evaluation/eval_predictions.py#L45 for details.\n \"\"\"\n # TODO add evaluation for multi-choice\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n pred = res[\"pred_ans\"]\n gt_ans = res[\"gt_ans\"]\n\n num_match = sum([pred == gt for gt in gt_ans])\n vqa_acc = min(1.0, num_match / 3.0)\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n\n [TODO] add support for multi-choice.\n \"\"\"\n result_leaderboard = dict()\n for res in results:\n result_leaderboard[res[\"question_id\"]] = {\n \"direct_answer\": res[\"pred_ans\"],\n \"multiple_choice\": \"\",\n }\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")\n \n\n \n\n@registry.register_task(\"videoqa\")\nclass VideoQA(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n\n answer = outputs[\"answer\"]\n qid = outputs[\"qid\"]\n output_text = outputs['output_text']\n # need_extra = outputs['need_extra_info']\n # print('need_extra', need_extra)","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.VideoQA","uri":"program://CREMA/class/lavis.tasks.vqa.VideoQA#L332-L445","kind":"class","name":"VideoQA","path":"lavis/tasks/vqa.py","language":"python","start_line":332,"end_line":445,"context_start_line":312,"context_end_line":465,"code":" [TODO] add support for multi-choice.\n \"\"\"\n result_leaderboard = dict()\n for res in results:\n result_leaderboard[res[\"question_id\"]] = {\n \"direct_answer\": res[\"pred_ans\"],\n \"multiple_choice\": \"\",\n }\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")\n \n\n \n\n@registry.register_task(\"videoqa\")\nclass VideoQA(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n\n answer = outputs[\"answer\"]\n qid = outputs[\"qid\"]\n output_text = outputs['output_text']\n # need_extra = outputs['need_extra_info']\n # print('need_extra', need_extra)\n if 'frame_idx' in outputs:\n frame_idx = outputs['frame_idx']\n else:\n frame_idx = [0 for i in range(len(qid))]\n \n # if 'dis_acc' in outputs:\n # dis_acc = outputs['dis_acc']\n # else:\n # dis_acc = [0 for i in range(len(qid))]\n\n # print(qid)\n # print(len(output_text), output_text)\n assert len(qid)==len(output_text)\n assert len(qid)==len(answer) \n \n for a, q, o, f, d in zip(answer, qid, output_text, frame_idx, dis_acc):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q,\n \"prediction\": o,\n \"target\": self.ANS_MAPPING[a[-1]],\n \"frame_idx\": f,\n # \"difficult\": n,\n 'dis_acc': d\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch)\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n total_num = len(results)\n acc = 0\n qtype_correct_dict = {}\n qtype_total_dict = {}\n\n # difficult_acc, difficult_total = 0, 0\n dis_acc = 0\n for r in results: \n qtype = r['qid'].split('_')[0]\n dis_acc += r['dis_acc']\n if qtype not in qtype_total_dict:\n qtype_total_dict[qtype] = 1\n else:\n qtype_total_dict[qtype] += 1 \n # if r['difficult'] == 1:\n # difficult_total += 1\n\n if r['prediction'] == r['target']:\n acc += 1\n if qtype not in qtype_correct_dict:\n qtype_correct_dict[qtype] = 1\n else:\n qtype_correct_dict[qtype] += 1 \n\n # if r['difficult'] == 1:\n # difficult_acc += 1\n \n metrics = {\"agg_metrics\": acc/total_num , 'total':total_num}\n \n for qtype in qtype_total_dict:\n metrics[qtype] = qtype_correct_dict[qtype] / qtype_total_dict[qtype] * 100\n \n # metrics['dis_acc'] = dis_acc/total_num*100\n \n # metrics['difficult_num'] = difficult_total\n # if difficult_total == 0:\n # metrics['difficult_acc'] = 0\n # else:\n # metrics['difficult_acc'] = difficult_acc / difficult_total * 100\n\n # for STAR\n if ('Interaction' in metrics) and ('Sequence' in metrics) and ('Prediction' in metrics) and ('Feasibility' in metrics):\n metrics[\"agg_metrics\"] = (metrics['Interaction'] + metrics['Sequence'] + metrics['Prediction'] + metrics['Feasibility']) / 4\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n@registry.register_task(\"moment_retrieval\")\nclass MR(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'no': 0, 'yes': 1}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n answer = outputs['answer']\n qid = outputs['qid']\n score = outputs['yes_score']\n pred = outputs['pred_ans']\n assert len(qid)==len(answer)\n assert len(qid)==len(score)\n assert len(qid)==len(pred) \n \n i = 0","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.MR","uri":"program://CREMA/class/lavis.tasks.vqa.MR#L448-L510","kind":"class","name":"MR","path":"lavis/tasks/vqa.py","language":"python","start_line":448,"end_line":510,"context_start_line":428,"context_end_line":530,"code":" # if difficult_total == 0:\n # metrics['difficult_acc'] = 0\n # else:\n # metrics['difficult_acc'] = difficult_acc / difficult_total * 100\n\n # for STAR\n if ('Interaction' in metrics) and ('Sequence' in metrics) and ('Prediction' in metrics) and ('Feasibility' in metrics):\n metrics[\"agg_metrics\"] = (metrics['Interaction'] + metrics['Sequence'] + metrics['Prediction'] + metrics['Feasibility']) / 4\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n@registry.register_task(\"moment_retrieval\")\nclass MR(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'no': 0, 'yes': 1}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n answer = outputs['answer']\n qid = outputs['qid']\n score = outputs['yes_score']\n pred = outputs['pred_ans']\n assert len(qid)==len(answer)\n assert len(qid)==len(score)\n assert len(qid)==len(pred) \n \n i = 0\n for a, q, s, p in zip(answer, qid, score, pred):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q + '_' + str(i),\n \"prediction\": p,\n \"target\": self.ANS_MAPPING[a],\n 'score': s\n }\n )\n i += 1\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch)\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n total_num = len(results)\n acc = 0\n for r in results:\n if r['prediction'] == r['target']:\n acc += 1\n metrics = {\"agg_metrics\": acc / total_num, 'total': total_num}\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n\n\n\n \n\ndef evals_json(gold_data, preds):\n score_list = ['Top1 (EM)']\n score = {s:[] for s in score_list}\n \n for ref, pred in zip(gold_data, preds):\n if pred in ref:\n score['Top1 (EM)'].append(1)\n else:\n #scores=[tokens_unigram_f_value(answer,ref) for ref in ref_answers]\n score['Top1 (EM)'].append(0)\n \n rlt={}\n for k,v in score.items():\n assert len(v)==len(gold_data),len(v)","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.evals_json","uri":"program://CREMA/function/lavis.tasks.vqa.evals_json#L517-L533","kind":"function","name":"evals_json","path":"lavis/tasks/vqa.py","language":"python","start_line":517,"end_line":533,"context_start_line":497,"context_end_line":553,"code":" acc = 0\n for r in results:\n if r['prediction'] == r['target']:\n acc += 1\n metrics = {\"agg_metrics\": acc / total_num, 'total': total_num}\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n\n\n\n \n\ndef evals_json(gold_data, preds):\n score_list = ['Top1 (EM)']\n score = {s:[] for s in score_list}\n \n for ref, pred in zip(gold_data, preds):\n if pred in ref:\n score['Top1 (EM)'].append(1)\n else:\n #scores=[tokens_unigram_f_value(answer,ref) for ref in ref_answers]\n score['Top1 (EM)'].append(0)\n \n rlt={}\n for k,v in score.items():\n assert len(v)==len(gold_data),len(v)\n rlt[k]=np.mean(v)*100\n \n return rlt\n\n# calculate the Top1(EM) for each question type\nSQA3D_QT=['All', 'What', 'How', 'Can', 'Is', 'Which', 'Other']\n@registry.register_task(\"sqa3d\")\nclass ThreeDVQATask(VQATask):\n \n def valid_step(self, model, samples):\n\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n pred_anwers = answers","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.ThreeDVQATask","uri":"program://CREMA/class/lavis.tasks.vqa.ThreeDVQATask#L538-L605","kind":"class","name":"ThreeDVQATask","path":"lavis/tasks/vqa.py","language":"python","start_line":538,"end_line":605,"context_start_line":518,"context_end_line":611,"code":" score_list = ['Top1 (EM)']\n score = {s:[] for s in score_list}\n \n for ref, pred in zip(gold_data, preds):\n if pred in ref:\n score['Top1 (EM)'].append(1)\n else:\n #scores=[tokens_unigram_f_value(answer,ref) for ref in ref_answers]\n score['Top1 (EM)'].append(0)\n \n rlt={}\n for k,v in score.items():\n assert len(v)==len(gold_data),len(v)\n rlt[k]=np.mean(v)*100\n \n return rlt\n\n# calculate the Top1(EM) for each question type\nSQA3D_QT=['All', 'What', 'How', 'Can', 'Is', 'Which', 'Other']\n@registry.register_task(\"sqa3d\")\nclass ThreeDVQATask(VQATask):\n \n def valid_step(self, model, samples):\n\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n pred_anwers = answers\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n questions = samples[\"qa_input\"]\n\n for pred_answer, ques_id, gt_answer, q in zip(pred_anwers, question_id, gt_answers, questions):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer, 'question': q}\n )\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n # for sqa3d-test: it is fine\n # for sqa3d-val: some 3d scene features are missing, need to filter the data/gt\n results = json.load(open(result_file, \"r\"))\n preds_={k:[] for k in SQA3D_QT}\n golds_={k:[] for k in SQA3D_QT}\n\n metrics = {}\n\n for data in results:\n qid = data['question_id']\n qtype = qid.split('_')[0]\n pred = data['pred_ans']\n gt = data['gt_ans']\n # qtype = self.qclass1(pred['question'])\n pred = pred.replace(\",\", '').replace(\"\", '').replace(\"\", \"\").strip()\n pred = pred.strip().split('\\n')[0]\n pred = pred.lower()\n if len(gt.split(' ')) == 1:\n pred = pred.split(' ')[-1]\n if pred in number_mapping:\n pred = number_mapping[pred]\n preds_[qtype].append(pred)\n golds_[qtype].append(gt)\n\n preds_['All'].append(pred)\n golds_['All'].append(gt)\n\n for qt in SQA3D_QT:\n score = evals_json(golds_[qt], preds_[qt])\n if qt == 'All':\n metrics['agg_metrics'] = score['Top1 (EM)']\n else:\n metrics[qt] = score['Top1 (EM)']\n \n with open(os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\") as f:\n f.write(json.dumps(metrics) + \"\\n\")\n logging.info(metrics)\n \n return metrics\n\n\n\n \n\n ","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.__init__","uri":"program://CREMA/function/lavis.tasks.vqa.__init__#L449-L451","kind":"function","name":"__init__","path":"lavis/tasks/vqa.py","language":"python","start_line":449,"end_line":451,"context_start_line":429,"context_end_line":471,"code":" # metrics['difficult_acc'] = 0\n # else:\n # metrics['difficult_acc'] = difficult_acc / difficult_total * 100\n\n # for STAR\n if ('Interaction' in metrics) and ('Sequence' in metrics) and ('Prediction' in metrics) and ('Feasibility' in metrics):\n metrics[\"agg_metrics\"] = (metrics['Interaction'] + metrics['Sequence'] + metrics['Prediction'] + metrics['Feasibility']) / 4\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n \n@registry.register_task(\"moment_retrieval\")\nclass MR(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'no': 0, 'yes': 1}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n answer = outputs['answer']\n qid = outputs['qid']\n score = outputs['yes_score']\n pred = outputs['pred_ans']\n assert len(qid)==len(answer)\n assert len(qid)==len(score)\n assert len(qid)==len(pred) \n \n i = 0\n for a, q, s, p in zip(answer, qid, score, pred):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q + '_' + str(i),\n \"prediction\": p,","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.setup_task","uri":"program://CREMA/function/lavis.tasks.vqa.setup_task#L53-L74","kind":"function","name":"setup_task","path":"lavis/tasks/vqa.py","language":"python","start_line":53,"end_line":74,"context_start_line":33,"context_end_line":94,"code":" inference_method=\"rank\",\n prompt=\"\",\n ):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n\n self.evaluate = evaluate\n self.inference_method = inference_method\n self.num_ans_candidates = num_ans_candidates\n self.prompt = prompt\n\n self.answer_list = None\n\n self.ques_files = dict()\n self.anno_files = dict()\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n prompt = run_cfg.get(\"prompt\", \"\")\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n prompt=prompt,\n )\n\n def build_datasets(self, cfg):\n datasets = super().build_datasets(cfg)\n\n # get question file, annotation file and anwser list in COCO format\n for dataset in datasets.values():\n for split in dataset:\n if (\n hasattr(dataset[split], \"coco_fmt_qust_file\")\n and dataset[split].coco_fmt_qust_file is not None\n ):\n self.ques_files[split] = dataset[split].coco_fmt_qust_file\n self.anno_files[split] = dataset[split].coco_fmt_anno_file\n\n try:\n self.answer_list = dataset[split].answer_list\n except AttributeError:\n # if answer_list is not provided, then set it to None\n pass\n","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.build_datasets","uri":"program://CREMA/function/lavis.tasks.vqa.build_datasets#L76-L100","kind":"function","name":"build_datasets","path":"lavis/tasks/vqa.py","language":"python","start_line":76,"end_line":100,"context_start_line":56,"context_end_line":120,"code":" num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n prompt = run_cfg.get(\"prompt\", \"\")\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n prompt=prompt,\n )\n\n def build_datasets(self, cfg):\n datasets = super().build_datasets(cfg)\n\n # get question file, annotation file and anwser list in COCO format\n for dataset in datasets.values():\n for split in dataset:\n if (\n hasattr(dataset[split], \"coco_fmt_qust_file\")\n and dataset[split].coco_fmt_qust_file is not None\n ):\n self.ques_files[split] = dataset[split].coco_fmt_qust_file\n self.anno_files[split] = dataset[split].coco_fmt_anno_file\n\n try:\n self.answer_list = dataset[split].answer_list\n except AttributeError:\n # if answer_list is not provided, then set it to None\n pass\n\n if len(self.ques_files) > 0:\n assert len(self.ques_files) == len(\n self.anno_files\n ), \"Only support one split for evaluation.\"\n\n return datasets\n\n def valid_step(self, model, samples):\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n prompt=self.prompt,\n )\n pred_qa_pairs = []\n\n question_id = samples[\"question_id\"]\n for answer, ques_id in zip(answers, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n\n return pred_qa_pairs","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.valid_step","uri":"program://CREMA/function/lavis.tasks.vqa.valid_step#L540-L563","kind":"function","name":"valid_step","path":"lavis/tasks/vqa.py","language":"python","start_line":540,"end_line":563,"context_start_line":520,"context_end_line":583,"code":" \n for ref, pred in zip(gold_data, preds):\n if pred in ref:\n score['Top1 (EM)'].append(1)\n else:\n #scores=[tokens_unigram_f_value(answer,ref) for ref in ref_answers]\n score['Top1 (EM)'].append(0)\n \n rlt={}\n for k,v in score.items():\n assert len(v)==len(gold_data),len(v)\n rlt[k]=np.mean(v)*100\n \n return rlt\n\n# calculate the Top1(EM) for each question type\nSQA3D_QT=['All', 'What', 'How', 'Can', 'Is', 'Which', 'Other']\n@registry.register_task(\"sqa3d\")\nclass ThreeDVQATask(VQATask):\n \n def valid_step(self, model, samples):\n\n answers = model.predict_answers(\n samples=samples,\n answer_list=self.answer_list,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n pred_anwers = answers\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n questions = samples[\"qa_input\"]\n\n for pred_answer, ques_id, gt_answer, q in zip(pred_anwers, question_id, gt_answers, questions):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer, 'question': q}\n )\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n # for sqa3d-test: it is fine\n # for sqa3d-val: some 3d scene features are missing, need to filter the data/gt\n results = json.load(open(result_file, \"r\"))\n preds_={k:[] for k in SQA3D_QT}\n golds_={k:[] for k in SQA3D_QT}\n\n metrics = {}\n\n for data in results:\n qid = data['question_id']\n qtype = qid.split('_')[0]\n pred = data['pred_ans']\n gt = data['gt_ans']\n # qtype = self.qclass1(pred['question'])\n pred = pred.replace(\",\", '').replace(\"\", '').replace(\"\", \"\").strip()\n pred = pred.strip().split('\\n')[0]\n pred = pred.lower()","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa.after_evaluation","uri":"program://CREMA/function/lavis.tasks.vqa.after_evaluation#L480-L491","kind":"function","name":"after_evaluation","path":"lavis/tasks/vqa.py","language":"python","start_line":480,"end_line":491,"context_start_line":460,"context_end_line":511,"code":" pred = outputs['pred_ans']\n assert len(qid)==len(answer)\n assert len(qid)==len(score)\n assert len(qid)==len(pred) \n \n i = 0\n for a, q, s, p in zip(answer, qid, score, pred):\n # l = l[self.ANS_MAPPING[a[-1]]]\n results.append(\n {\n \"qid\": q + '_' + str(i),\n \"prediction\": p,\n \"target\": self.ANS_MAPPING[a],\n 'score': s\n }\n )\n i += 1\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch)\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n total_num = len(results)\n acc = 0\n for r in results:\n if r['prediction'] == r['target']:\n acc += 1\n metrics = {\"agg_metrics\": acc / total_num, 'total': total_num}\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics\n ","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa._report_metrics","uri":"program://CREMA/function/lavis.tasks.vqa._report_metrics#L566-L605","kind":"function","name":"_report_metrics","path":"lavis/tasks/vqa.py","language":"python","start_line":566,"end_line":605,"context_start_line":546,"context_end_line":611,"code":" num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n num_ans_candidates=self.num_ans_candidates,\n )\n\n pred_qa_pairs = []\n pred_anwers = answers\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"qa_output\"]\n questions = samples[\"qa_input\"]\n\n for pred_answer, ques_id, gt_answer, q in zip(pred_anwers, question_id, gt_answers, questions):\n pred_qa_pairs.append(\n {\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer, 'question': q}\n )\n\n return pred_qa_pairs\n \n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n # for sqa3d-test: it is fine\n # for sqa3d-val: some 3d scene features are missing, need to filter the data/gt\n results = json.load(open(result_file, \"r\"))\n preds_={k:[] for k in SQA3D_QT}\n golds_={k:[] for k in SQA3D_QT}\n\n metrics = {}\n\n for data in results:\n qid = data['question_id']\n qtype = qid.split('_')[0]\n pred = data['pred_ans']\n gt = data['gt_ans']\n # qtype = self.qclass1(pred['question'])\n pred = pred.replace(\",\", '').replace(\"\", '').replace(\"\", \"\").strip()\n pred = pred.strip().split('\\n')[0]\n pred = pred.lower()\n if len(gt.split(' ')) == 1:\n pred = pred.split(' ')[-1]\n if pred in number_mapping:\n pred = number_mapping[pred]\n preds_[qtype].append(pred)\n golds_[qtype].append(gt)\n\n preds_['All'].append(pred)\n golds_['All'].append(gt)\n\n for qt in SQA3D_QT:\n score = evals_json(golds_[qt], preds_[qt])\n if qt == 'All':\n metrics['agg_metrics'] = score['Top1 (EM)']\n else:\n metrics[qt] = score['Top1 (EM)']\n \n with open(os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\") as f:\n f.write(json.dumps(metrics) + \"\\n\")\n logging.info(metrics)\n \n return metrics\n\n\n\n \n\n ","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa._save_result_leaderboard","uri":"program://CREMA/function/lavis.tasks.vqa._save_result_leaderboard#L308-L326","kind":"function","name":"_save_result_leaderboard","path":"lavis/tasks/vqa.py","language":"python","start_line":308,"end_line":326,"context_start_line":288,"context_end_line":346,"code":" gt_ans = res[\"gt_ans\"]\n\n num_match = sum([pred == gt for gt in gt_ans])\n vqa_acc = min(1.0, num_match / 3.0)\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n\n [TODO] add support for multi-choice.\n \"\"\"\n result_leaderboard = dict()\n for res in results:\n result_leaderboard[res[\"question_id\"]] = {\n \"direct_answer\": res[\"pred_ans\"],\n \"multiple_choice\": \"\",\n }\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")\n \n\n \n\n@registry.register_task(\"videoqa\")\nclass VideoQA(BaseTask):\n def __init__(self):\n super().__init__()\n self.ANS_MAPPING = {'A':0, 'B':1, 'C':2, 'D':3, 'E':4}\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.generate(samples)\n\n answer = outputs[\"answer\"]\n qid = outputs[\"qid\"]\n output_text = outputs['output_text']\n # need_extra = outputs['need_extra_info']\n # print('need_extra', need_extra)","source_hash":"62824d1b13d070d7f1325306624a122597efc9d5b427a48fe7f20e22f0438e5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension","uri":"program://CREMA/module/lavis.tasks.vqa_reading_comprehension#L1-L248","kind":"module","name":"lavis.tasks.vqa_reading_comprehension","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":1,"end_line":248,"context_start_line":1,"context_end_line":248,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport json\nimport os\nimport torch\nimport torch.distributed as dist\nfrom itertools import chain\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.common.vqa_tools.vqa_eval import VQAEval as VQATool\nfrom lavis.tasks.vqa import VQATask\n\n\n@registry.register_task(\"vqa_reading_comprehension\")\nclass VQARCTask(VQATask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n **kwargs,\n ):\n super().__init__(num_beams, max_len, min_len, evaluate, num_ans_candidates, inference_method)\n\n self.config = kwargs.get('config')\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n config=run_cfg,\n )\n\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n for answer, caption, gradcam, ques_id in zip(answers, captions, gradcams, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n def after_evaluation(self, val_result, split_name, **kwargs):\n result_ = list(chain(*val_result[0::3]))\n result_file = self.save_gradcam(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_gradcam_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[1::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_caption_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[2::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n def save_gradcam(self, result, result_dir, filename, remove_duplicate=\"\"):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, get_rank()))\n final_result_file = os.path.join(result_dir, '%s.pth' % filename)\n torch.save({'result': result}, result_file)\n\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, rank))\n res_ckpt = torch.load(result_file, map_location='cpu')\n res = res_ckpt['result']\n\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n torch.save({'result': result}, final_result_file)\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file\n\n\n@registry.register_task(\"gqa_reading_comprehension\")\nclass GQARCTask(VQARCTask):\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"answer\"]\n\n for pred_answer, caption, gradcam, ques_id, gt_answer in zip(answers, captions, gradcams, question_id, gt_answers):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQATool()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n\n if self.inference_method == \"generate\":\n pred = vqa_tool.processPunctuation(pred)\n pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n \"\"\"\n result_leaderboard = []\n for res in results:\n result_leaderboard.append({\n \"questionId\": str(res['question_id']),\n \"prediction\": str(res[\"pred_ans\"]),\n })\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.VQARCTask","uri":"program://CREMA/class/lavis.tasks.vqa_reading_comprehension.VQARCTask#L23-L153","kind":"class","name":"VQARCTask","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":23,"end_line":153,"context_start_line":3,"context_end_line":173,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport json\nimport os\nimport torch\nimport torch.distributed as dist\nfrom itertools import chain\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.common.vqa_tools.vqa_eval import VQAEval as VQATool\nfrom lavis.tasks.vqa import VQATask\n\n\n@registry.register_task(\"vqa_reading_comprehension\")\nclass VQARCTask(VQATask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n **kwargs,\n ):\n super().__init__(num_beams, max_len, min_len, evaluate, num_ans_candidates, inference_method)\n\n self.config = kwargs.get('config')\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n config=run_cfg,\n )\n\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n for answer, caption, gradcam, ques_id in zip(answers, captions, gradcams, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n def after_evaluation(self, val_result, split_name, **kwargs):\n result_ = list(chain(*val_result[0::3]))\n result_file = self.save_gradcam(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_gradcam_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[1::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_caption_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[2::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n def save_gradcam(self, result, result_dir, filename, remove_duplicate=\"\"):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, get_rank()))\n final_result_file = os.path.join(result_dir, '%s.pth' % filename)\n torch.save({'result': result}, result_file)\n\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, rank))\n res_ckpt = torch.load(result_file, map_location='cpu')\n res = res_ckpt['result']\n\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n torch.save({'result': result}, final_result_file)\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file\n\n\n@registry.register_task(\"gqa_reading_comprehension\")\nclass GQARCTask(VQARCTask):\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.GQARCTask","uri":"program://CREMA/class/lavis.tasks.vqa_reading_comprehension.GQARCTask#L157-L248","kind":"class","name":"GQARCTask","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":157,"end_line":248,"context_start_line":137,"context_end_line":248,"code":" res = res_ckpt['result']\n\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n torch.save({'result': result}, final_result_file)\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file\n\n\n@registry.register_task(\"gqa_reading_comprehension\")\nclass GQARCTask(VQARCTask):\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"answer\"]\n\n for pred_answer, caption, gradcam, ques_id, gt_answer in zip(answers, captions, gradcams, question_id, gt_answers):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQATool()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n\n if self.inference_method == \"generate\":\n pred = vqa_tool.processPunctuation(pred)\n pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n \"\"\"\n result_leaderboard = []\n for res in results:\n result_leaderboard.append({\n \"questionId\": str(res['question_id']),\n \"prediction\": str(res[\"pred_ans\"]),\n })\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.__init__","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension.__init__#L24-L36","kind":"function","name":"__init__","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":24,"end_line":36,"context_start_line":4,"context_end_line":56,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport json\nimport os\nimport torch\nimport torch.distributed as dist\nfrom itertools import chain\n\nimport lavis.common.dist_utils as dist_utils\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.common.vqa_tools.vqa_eval import VQAEval as VQATool\nfrom lavis.tasks.vqa import VQATask\n\n\n@registry.register_task(\"vqa_reading_comprehension\")\nclass VQARCTask(VQATask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n **kwargs,\n ):\n super().__init__(num_beams, max_len, min_len, evaluate, num_ans_candidates, inference_method)\n\n self.config = kwargs.get('config')\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.setup_task","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension.setup_task#L39-L59","kind":"function","name":"setup_task","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":39,"end_line":59,"context_start_line":19,"context_end_line":79,"code":"from lavis.tasks.vqa import VQATask\n\n\n@registry.register_task(\"vqa_reading_comprehension\")\nclass VQARCTask(VQATask):\n def __init__(\n self,\n num_beams,\n max_len,\n min_len,\n evaluate,\n num_ans_candidates,\n inference_method=\"rank\",\n **kwargs,\n ):\n super().__init__(num_beams, max_len, min_len, evaluate, num_ans_candidates, inference_method)\n\n self.config = kwargs.get('config')\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.get(\"num_beams\", 3)\n max_len = run_cfg.get(\"max_len\", 10)\n min_len = run_cfg.get(\"min_len\", 1)\n\n evaluate = run_cfg.get(\"evaluate\", False)\n\n inference_method = run_cfg.get(\"inference_method\", \"rank\")\n num_ans_candidates = run_cfg.get(\"num_ans_candidates\", 128)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n num_ans_candidates=num_ans_candidates,\n inference_method=inference_method,\n config=run_cfg,\n )\n\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.valid_step","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension.valid_step#L158-L190","kind":"function","name":"valid_step","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":158,"end_line":190,"context_start_line":138,"context_end_line":210,"code":"\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n torch.save({'result': result}, final_result_file)\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file\n\n\n@registry.register_task(\"gqa_reading_comprehension\")\nclass GQARCTask(VQARCTask):\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"answer\"]\n\n for pred_answer, caption, gradcam, ques_id, gt_answer in zip(answers, captions, gradcams, question_id, gt_answers):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQATool()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.after_evaluation","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension.after_evaluation#L93-L120","kind":"function","name":"after_evaluation","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":93,"end_line":120,"context_start_line":73,"context_end_line":140,"code":" top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n for answer, caption, gradcam, ques_id in zip(answers, captions, gradcams, question_id):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"answer\": answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n def after_evaluation(self, val_result, split_name, **kwargs):\n result_ = list(chain(*val_result[0::3]))\n result_file = self.save_gradcam(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_gradcam_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[1::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_caption_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[2::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n def save_gradcam(self, result, result_dir, filename, remove_duplicate=\"\"):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, get_rank()))\n final_result_file = os.path.join(result_dir, '%s.pth' % filename)\n torch.save({'result': result}, result_file)\n\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, rank))\n res_ckpt = torch.load(result_file, map_location='cpu')\n res = res_ckpt['result']\n\n result += res\n","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension.save_gradcam","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension.save_gradcam#L122-L153","kind":"function","name":"save_gradcam","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":122,"end_line":153,"context_start_line":102,"context_end_line":173,"code":" result_ = list(chain(*val_result[1::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_caption_result\",\n remove_duplicate=\"question_id\",\n )\n\n result_ = list(chain(*val_result[2::3]))\n result_file = self.save_result(\n result_,\n result_dir=registry.get_path(\"result_dir\"),\n filename=f\"{split_name}_vqa_result\",\n remove_duplicate=\"question_id\",\n )\n\n metrics = self._report_metrics(result_file=result_file, split=split_name)\n\n return metrics\n\n def save_gradcam(self, result, result_dir, filename, remove_duplicate=\"\"):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, get_rank()))\n final_result_file = os.path.join(result_dir, '%s.pth' % filename)\n torch.save({'result': result}, result_file)\n\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(result_dir, '%s_rank%d.pth' % (filename, rank))\n res_ckpt = torch.load(result_file, map_location='cpu')\n res = res_ckpt['result']\n\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n torch.save({'result': result}, final_result_file)\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file\n\n\n@registry.register_task(\"gqa_reading_comprehension\")\nclass GQARCTask(VQARCTask):\n def valid_step(self, model, samples):\n answers, captions, gradcams = model.predict_answers(\n samples=samples,\n inference_method=self.inference_method,\n num_beams=self.num_beams,\n max_len=self.max_len,\n min_len=self.min_len,\n internal_bsz_fid=self.config['internal_bsz_fid'],\n num_captions=self.config['num_captions'],\n num_captions_fid=self.config['num_captions_fid'],\n cap_max_length=self.config['cap_max_length'],\n cap_min_length=self.config['cap_min_length'],\n top_k=self.config['top_k'],\n top_p=self.config['top_p'],\n repetition_penalty=self.config['repetition_penalty'],\n num_patches=self.config['num_patches'],","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension._report_metrics","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension._report_metrics#L193-L229","kind":"function","name":"_report_metrics","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":193,"end_line":229,"context_start_line":173,"context_end_line":248,"code":" num_patches=self.config['num_patches'],\n block_num=self.config['block_num'],\n )\n\n pred_qa_pairs = []\n sample_captions = []\n sample_gradcams = []\n\n question_id = samples[\"question_id\"]\n gt_answers = samples[\"answer\"]\n\n for pred_answer, caption, gradcam, ques_id, gt_answer in zip(answers, captions, gradcams, question_id, gt_answers):\n ques_id = int(ques_id.item())\n pred_qa_pairs.append({\"question_id\": ques_id, \"pred_ans\": pred_answer, \"gt_ans\": gt_answer})\n sample_captions.append({\"question_id\": ques_id, \"caption\": caption})\n sample_gradcams.append({\"question_id\": ques_id, \"gradcam\": gradcam})\n\n return [sample_gradcams, sample_captions, pred_qa_pairs]\n\n @dist_utils.main_process\n def _report_metrics(self, result_file, split):\n \"\"\"\n TODO: add other evaluation metrics for GQA\n \"\"\"\n\n results = json.load(open(result_file, \"r\"))\n acc = []\n vqa_tool = VQATool()\n\n for res in results:\n if res[\"gt_ans\"] is None:\n # prepare test results for leaderboard evaluation\n self._save_result_leaderboard(results)\n return\n\n gt_ans = res[\"gt_ans\"]\n pred = res[\"pred_ans\"]\n\n if self.inference_method == \"generate\":\n pred = vqa_tool.processPunctuation(pred)\n pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n \"\"\"\n result_leaderboard = []\n for res in results:\n result_leaderboard.append({\n \"questionId\": str(res['question_id']),\n \"prediction\": str(res[\"pred_ans\"]),\n })\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.vqa_reading_comprehension._save_result_leaderboard","uri":"program://CREMA/function/lavis.tasks.vqa_reading_comprehension._save_result_leaderboard#L232-L248","kind":"function","name":"_save_result_leaderboard","path":"lavis/tasks/vqa_reading_comprehension.py","language":"python","start_line":232,"end_line":248,"context_start_line":212,"context_end_line":248,"code":" pred = vqa_tool.processPunctuation(pred)\n pred = vqa_tool.processDigitArticle(pred)\n\n vqa_acc = 1 if pred == gt_ans else 0\n\n acc.append(vqa_acc)\n\n accuracy = sum(acc) / len(acc) * 100\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(metrics) + \"\\n\")\n\n logging.info(metrics)\n\n return metrics\n\n @dist_utils.main_process\n def _save_result_leaderboard(self, results):\n \"\"\"\n Saving the results in the format required for leaderboard evaluation.\n \"\"\"\n result_leaderboard = []\n for res in results:\n result_leaderboard.append({\n \"questionId\": str(res['question_id']),\n \"prediction\": str(res[\"pred_ans\"]),\n })\n\n result_file = registry.get_path(\"result_dir\") + \"_leaderboard.json\"\n\n with open(result_file, \"w\") as f:\n json.dump(result_leaderboard, f)\n\n logging.info(f\"Saved results for leaderboard evaluation at {result_file}\")","source_hash":"426e33ad2f7cd7e875d6db1fa066d5fda0ed96516e8d45da0cf97d061e1cf52e","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning","uri":"program://CREMA/module/lavis.tasks.captioning#L1-L142","kind":"module","name":"lavis.tasks.captioning","path":"lavis/tasks/captioning.py","language":"python","start_line":1,"end_line":142,"context_start_line":1,"context_end_line":142,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"captioning\")\nclass CaptionTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n\n # run_cfg = slf.cfg.run_cfg\n captions = model.generate(\n samples,\n use_nucleus_sampling=False,\n num_beams=self.num_beams,\n max_length=self.max_len,\n min_length=self.min_len,\n )\n\n img_ids = samples[\"image_id\"]\n for caption, img_id in zip(captions, img_ids):\n results.append({\"caption\": caption, \"image_id\": int(img_id)})\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=\"image_id\",\n )\n\n if self.report_metric:\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_caption_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_caption_eval(coco_gt_root, results_file, split):\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])\n\n # create coco object and coco_result object\n coco = COCO(annotation_file)\n coco_result = coco.loadRes(results_file)\n\n # create coco_eval object by taking coco and coco_result\n coco_eval = COCOEvalCap(coco, coco_result)\n\n # evaluate on a subset of images by setting\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n # please remove this line when evaluating the full validation set\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n\n # evaluate results\n # SPICE will take a few minutes the first time, but speeds up due to caching\n coco_eval.evaluate()\n\n # print output evaluation scores\n for metric, score in coco_eval.eval.items():\n print(f\"{metric}: {score:.3f}\")\n\n return coco_eval","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.CaptionTask","uri":"program://CREMA/class/lavis.tasks.captioning.CaptionTask#L17-L100","kind":"class","name":"CaptionTask","path":"lavis/tasks/captioning.py","language":"python","start_line":17,"end_line":100,"context_start_line":1,"context_end_line":120,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"captioning\")\nclass CaptionTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n\n # run_cfg = slf.cfg.run_cfg\n captions = model.generate(\n samples,\n use_nucleus_sampling=False,\n num_beams=self.num_beams,\n max_length=self.max_len,\n min_length=self.min_len,\n )\n\n img_ids = samples[\"image_id\"]\n for caption, img_id in zip(captions, img_ids):\n results.append({\"caption\": caption, \"image_id\": int(img_id)})\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=\"image_id\",\n )\n\n if self.report_metric:\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_caption_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_caption_eval(coco_gt_root, results_file, split):\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.coco_caption_eval","uri":"program://CREMA/function/lavis.tasks.captioning.coco_caption_eval#L109-L142","kind":"function","name":"coco_caption_eval","path":"lavis/tasks/captioning.py","language":"python","start_line":109,"end_line":142,"context_start_line":89,"context_end_line":142,"code":" agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_caption_eval(coco_gt_root, results_file, split):\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])\n\n # create coco object and coco_result object\n coco = COCO(annotation_file)\n coco_result = coco.loadRes(results_file)\n\n # create coco_eval object by taking coco and coco_result\n coco_eval = COCOEvalCap(coco, coco_result)\n\n # evaluate on a subset of images by setting\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n # please remove this line when evaluating the full validation set\n # coco_eval.params['image_id'] = coco_result.getImgIds()\n\n # evaluate results\n # SPICE will take a few minutes the first time, but speeds up due to caching\n coco_eval.evaluate()\n\n # print output evaluation scores\n for metric, score in coco_eval.eval.items():\n print(f\"{metric}: {score:.3f}\")\n\n return coco_eval","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.__init__","uri":"program://CREMA/function/lavis.tasks.captioning.__init__#L18-L26","kind":"function","name":"__init__","path":"lavis/tasks/captioning.py","language":"python","start_line":18,"end_line":26,"context_start_line":1,"context_end_line":46,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"captioning\")\nclass CaptionTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.setup_task","uri":"program://CREMA/function/lavis.tasks.captioning.setup_task#L29-L45","kind":"function","name":"setup_task","path":"lavis/tasks/captioning.py","language":"python","start_line":29,"end_line":45,"context_start_line":9,"context_end_line":65,"code":"import os\n\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"captioning\")\nclass CaptionTask(BaseTask):\n def __init__(self, num_beams, max_len, min_len, evaluate, report_metric=True):\n super().__init__()\n\n self.num_beams = num_beams\n self.max_len = max_len\n self.min_len = min_len\n self.evaluate = evaluate\n\n self.report_metric = report_metric\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n\n # run_cfg = slf.cfg.run_cfg\n captions = model.generate(\n samples,\n use_nucleus_sampling=False,\n num_beams=self.num_beams,\n max_length=self.max_len,\n min_length=self.min_len,\n )\n\n img_ids = samples[\"image_id\"]\n for caption, img_id in zip(captions, img_ids):\n results.append({\"caption\": caption, \"image_id\": int(img_id)})\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.valid_step","uri":"program://CREMA/function/lavis.tasks.captioning.valid_step#L47-L63","kind":"function","name":"valid_step","path":"lavis/tasks/captioning.py","language":"python","start_line":47,"end_line":63,"context_start_line":27,"context_end_line":83,"code":"\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n num_beams = run_cfg.num_beams\n max_len = run_cfg.max_len\n min_len = run_cfg.min_len\n evaluate = run_cfg.evaluate\n\n report_metric = run_cfg.get(\"report_metric\", True)\n\n return cls(\n num_beams=num_beams,\n max_len=max_len,\n min_len=min_len,\n evaluate=evaluate,\n report_metric=report_metric,\n )\n\n def valid_step(self, model, samples):\n results = []\n\n # run_cfg = slf.cfg.run_cfg\n captions = model.generate(\n samples,\n use_nucleus_sampling=False,\n num_beams=self.num_beams,\n max_length=self.max_len,\n min_length=self.min_len,\n )\n\n img_ids = samples[\"image_id\"]\n for caption, img_id in zip(captions, img_ids):\n results.append({\"caption\": caption, \"image_id\": int(img_id)})\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=\"image_id\",\n )\n\n if self.report_metric:\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning.after_evaluation","uri":"program://CREMA/function/lavis.tasks.captioning.after_evaluation#L65-L80","kind":"function","name":"after_evaluation","path":"lavis/tasks/captioning.py","language":"python","start_line":65,"end_line":80,"context_start_line":45,"context_end_line":100,"code":" )\n\n def valid_step(self, model, samples):\n results = []\n\n # run_cfg = slf.cfg.run_cfg\n captions = model.generate(\n samples,\n use_nucleus_sampling=False,\n num_beams=self.num_beams,\n max_length=self.max_len,\n min_length=self.min_len,\n )\n\n img_ids = samples[\"image_id\"]\n for caption, img_id in zip(captions, img_ids):\n results.append({\"caption\": caption, \"image_id\": int(img_id)})\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=\"image_id\",\n )\n\n if self.report_metric:\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_caption_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.captioning._report_metrics","uri":"program://CREMA/function/lavis.tasks.captioning._report_metrics#L83-L100","kind":"function","name":"_report_metrics","path":"lavis/tasks/captioning.py","language":"python","start_line":83,"end_line":100,"context_start_line":63,"context_end_line":120,"code":" return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=\"image_id\",\n )\n\n if self.report_metric:\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n else:\n metrics = {\"agg_metrics\": 0.0}\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n\n # TODO better way to define this\n coco_gt_root = os.path.join(registry.get_path(\"cache_root\"), \"coco_gt\")\n coco_val = coco_caption_eval(coco_gt_root, eval_result_file, split_name)\n\n agg_metrics = coco_val.eval[\"CIDEr\"] + coco_val.eval[\"Bleu_4\"]\n log_stats = {split_name: {k: v for k, v in coco_val.eval.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n coco_res = {k: v for k, v in coco_val.eval.items()}\n coco_res[\"agg_metrics\"] = agg_metrics\n\n return coco_res\n\n\n# TODO better structure for this.\nfrom pycocoevalcap.eval import COCOEvalCap\nfrom pycocotools.coco import COCO\nfrom torchvision.datasets.utils import download_url\n\n\ndef coco_caption_eval(coco_gt_root, results_file, split):\n urls = {\n \"val\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_val_gt.json\",\n \"test\": \"https://storage.googleapis.com/sfr-vision-language-research/datasets/coco_karpathy_test_gt.json\",\n }\n filenames = {\n \"val\": \"coco_karpathy_val_gt.json\",\n \"test\": \"coco_karpathy_test_gt.json\",\n }\n\n download_url(urls[split], coco_gt_root)\n annotation_file = os.path.join(coco_gt_root, filenames[split])","source_hash":"feebaeffc70eea1b5eb6832299d8ebc52fbad7cb6aedf4ef1f2cf5474e626111","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.image_text_pretrain","uri":"program://CREMA/module/lavis.tasks.image_text_pretrain#L1-L18","kind":"module","name":"lavis.tasks.image_text_pretrain","path":"lavis/tasks/image_text_pretrain.py","language":"python","start_line":1,"end_line":18,"context_start_line":1,"context_end_line":18,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"image_text_pretrain\")\nclass ImageTextPretrainTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n pass","source_hash":"a9353c27e3e19dddfc3a5c1230e3d96b48fc7847e32af0b1e59d61ab941647c7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.image_text_pretrain.ImageTextPretrainTask","uri":"program://CREMA/class/lavis.tasks.image_text_pretrain.ImageTextPretrainTask#L13-L18","kind":"class","name":"ImageTextPretrainTask","path":"lavis/tasks/image_text_pretrain.py","language":"python","start_line":13,"end_line":18,"context_start_line":1,"context_end_line":18,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"image_text_pretrain\")\nclass ImageTextPretrainTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n pass","source_hash":"a9353c27e3e19dddfc3a5c1230e3d96b48fc7847e32af0b1e59d61ab941647c7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.image_text_pretrain.__init__","uri":"program://CREMA/function/lavis.tasks.image_text_pretrain.__init__#L14-L15","kind":"function","name":"__init__","path":"lavis/tasks/image_text_pretrain.py","language":"python","start_line":14,"end_line":15,"context_start_line":1,"context_end_line":18,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"image_text_pretrain\")\nclass ImageTextPretrainTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n pass","source_hash":"a9353c27e3e19dddfc3a5c1230e3d96b48fc7847e32af0b1e59d61ab941647c7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.image_text_pretrain.evaluation","uri":"program://CREMA/function/lavis.tasks.image_text_pretrain.evaluation#L17-L18","kind":"function","name":"evaluation","path":"lavis/tasks/image_text_pretrain.py","language":"python","start_line":17,"end_line":18,"context_start_line":1,"context_end_line":18,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"image_text_pretrain\")\nclass ImageTextPretrainTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n pass","source_hash":"a9353c27e3e19dddfc3a5c1230e3d96b48fc7847e32af0b1e59d61ab941647c7","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval","uri":"program://CREMA/module/lavis.tasks.retrieval#L1-L107","kind":"module","name":"lavis.tasks.retrieval","path":"lavis/tasks/retrieval.py","language":"python","start_line":1,"end_line":107,"context_start_line":1,"context_end_line":107,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"retrieval\")\nclass RetrievalTask(BaseTask):\n def __init__(self, cfg):\n super().__init__()\n\n self.cfg = cfg\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result\n\n @staticmethod\n @torch.no_grad()\n def _report_metrics(scores_i2t, scores_t2i, txt2img, img2txt):\n\n # Images->Text\n ranks = np.zeros(scores_i2t.shape[0])\n for index, score in enumerate(scores_i2t):\n inds = np.argsort(score)[::-1]\n # Score\n rank = 1e20\n for i in img2txt[index]:\n tmp = np.where(inds == i)[0][0]\n if tmp < rank:\n rank = tmp\n ranks[index] = rank\n\n # Compute metrics\n tr1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n tr5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n tr10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n # Text->Images\n ranks = np.zeros(scores_t2i.shape[0])\n\n for index, score in enumerate(scores_t2i):\n inds = np.argsort(score)[::-1]\n ranks[index] = np.where(inds == txt2img[index])[0][0]\n\n # Compute metrics\n ir1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n ir5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n ir10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n tr_mean = (tr1 + tr5 + tr10) / 3\n ir_mean = (ir1 + ir5 + ir10) / 3\n r_mean = (tr_mean + ir_mean) / 2\n\n agg_metrics = (tr1 + tr5 + tr10) / 3\n\n eval_result = {\n \"txt_r1\": tr1,\n \"txt_r5\": tr5,\n \"txt_r10\": tr10,\n \"txt_r_mean\": tr_mean,\n \"img_r1\": ir1,\n \"img_r5\": ir5,\n \"img_r10\": ir10,\n \"img_r_mean\": ir_mean,\n \"r_mean\": r_mean,\n \"agg_metrics\": agg_metrics,\n }\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(eval_result) + \"\\n\")\n return eval_result","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval.RetrievalTask","uri":"program://CREMA/class/lavis.tasks.retrieval.RetrievalTask#L20-L107","kind":"class","name":"RetrievalTask","path":"lavis/tasks/retrieval.py","language":"python","start_line":20,"end_line":107,"context_start_line":1,"context_end_line":107,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"retrieval\")\nclass RetrievalTask(BaseTask):\n def __init__(self, cfg):\n super().__init__()\n\n self.cfg = cfg\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result\n\n @staticmethod\n @torch.no_grad()\n def _report_metrics(scores_i2t, scores_t2i, txt2img, img2txt):\n\n # Images->Text\n ranks = np.zeros(scores_i2t.shape[0])\n for index, score in enumerate(scores_i2t):\n inds = np.argsort(score)[::-1]\n # Score\n rank = 1e20\n for i in img2txt[index]:\n tmp = np.where(inds == i)[0][0]\n if tmp < rank:\n rank = tmp\n ranks[index] = rank\n\n # Compute metrics\n tr1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n tr5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n tr10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n # Text->Images\n ranks = np.zeros(scores_t2i.shape[0])\n\n for index, score in enumerate(scores_t2i):\n inds = np.argsort(score)[::-1]\n ranks[index] = np.where(inds == txt2img[index])[0][0]\n\n # Compute metrics\n ir1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n ir5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n ir10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n tr_mean = (tr1 + tr5 + tr10) / 3\n ir_mean = (ir1 + ir5 + ir10) / 3\n r_mean = (tr_mean + ir_mean) / 2\n\n agg_metrics = (tr1 + tr5 + tr10) / 3\n\n eval_result = {\n \"txt_r1\": tr1,\n \"txt_r5\": tr5,\n \"txt_r10\": tr10,\n \"txt_r_mean\": tr_mean,\n \"img_r1\": ir1,\n \"img_r5\": ir5,\n \"img_r10\": ir10,\n \"img_r_mean\": ir_mean,\n \"r_mean\": r_mean,\n \"agg_metrics\": agg_metrics,\n }\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(eval_result) + \"\\n\")\n return eval_result","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval.__init__","uri":"program://CREMA/function/lavis.tasks.retrieval.__init__#L21-L24","kind":"function","name":"__init__","path":"lavis/tasks/retrieval.py","language":"python","start_line":21,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"retrieval\")\nclass RetrievalTask(BaseTask):\n def __init__(self, cfg):\n super().__init__()\n\n self.cfg = cfg\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval.setup_task","uri":"program://CREMA/function/lavis.tasks.retrieval.setup_task#L27-L30","kind":"function","name":"setup_task","path":"lavis/tasks/retrieval.py","language":"python","start_line":27,"end_line":30,"context_start_line":7,"context_end_line":50,"code":"\nimport json\nimport logging\nimport os\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"retrieval\")\nclass RetrievalTask(BaseTask):\n def __init__(self, cfg):\n super().__init__()\n\n self.cfg = cfg\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval.evaluation","uri":"program://CREMA/function/lavis.tasks.retrieval.evaluation#L32-L47","kind":"function","name":"evaluation","path":"lavis/tasks/retrieval.py","language":"python","start_line":32,"end_line":47,"context_start_line":12,"context_end_line":67,"code":"import numpy as np\nimport torch\nfrom lavis.common.dist_utils import is_main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"retrieval\")\nclass RetrievalTask(BaseTask):\n def __init__(self, cfg):\n super().__init__()\n\n self.cfg = cfg\n\n @classmethod\n def setup_task(cls, cfg):\n run_cfg = cfg.run_cfg\n\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result\n\n @staticmethod\n @torch.no_grad()\n def _report_metrics(scores_i2t, scores_t2i, txt2img, img2txt):\n\n # Images->Text\n ranks = np.zeros(scores_i2t.shape[0])\n for index, score in enumerate(scores_i2t):\n inds = np.argsort(score)[::-1]\n # Score\n rank = 1e20\n for i in img2txt[index]:\n tmp = np.where(inds == i)[0][0]\n if tmp < rank:\n rank = tmp\n ranks[index] = rank\n","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval.after_evaluation","uri":"program://CREMA/function/lavis.tasks.retrieval.after_evaluation#L49-L50","kind":"function","name":"after_evaluation","path":"lavis/tasks/retrieval.py","language":"python","start_line":49,"end_line":50,"context_start_line":29,"context_end_line":70,"code":"\n return cls(cfg=run_cfg)\n\n def evaluation(self, model, data_loader, **kwargs):\n # score_i2t, score_t2i = model.compute_sim_matrix(model, data_loader)\n score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result\n\n @staticmethod\n @torch.no_grad()\n def _report_metrics(scores_i2t, scores_t2i, txt2img, img2txt):\n\n # Images->Text\n ranks = np.zeros(scores_i2t.shape[0])\n for index, score in enumerate(scores_i2t):\n inds = np.argsort(score)[::-1]\n # Score\n rank = 1e20\n for i in img2txt[index]:\n tmp = np.where(inds == i)[0][0]\n if tmp < rank:\n rank = tmp\n ranks[index] = rank\n\n # Compute metrics\n tr1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n tr5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.retrieval._report_metrics","uri":"program://CREMA/function/lavis.tasks.retrieval._report_metrics#L54-L107","kind":"function","name":"_report_metrics","path":"lavis/tasks/retrieval.py","language":"python","start_line":54,"end_line":107,"context_start_line":34,"context_end_line":107,"code":" score_i2t, score_t2i = model.compute_sim_matrix(data_loader, task_cfg=self.cfg)\n\n if is_main_process():\n eval_result = self._report_metrics(\n score_i2t,\n score_t2i,\n data_loader.dataset.txt2img,\n data_loader.dataset.img2txt,\n )\n logging.info(eval_result)\n else:\n eval_result = None\n\n return eval_result\n\n def after_evaluation(self, val_result, **kwargs):\n return val_result\n\n @staticmethod\n @torch.no_grad()\n def _report_metrics(scores_i2t, scores_t2i, txt2img, img2txt):\n\n # Images->Text\n ranks = np.zeros(scores_i2t.shape[0])\n for index, score in enumerate(scores_i2t):\n inds = np.argsort(score)[::-1]\n # Score\n rank = 1e20\n for i in img2txt[index]:\n tmp = np.where(inds == i)[0][0]\n if tmp < rank:\n rank = tmp\n ranks[index] = rank\n\n # Compute metrics\n tr1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n tr5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n tr10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n # Text->Images\n ranks = np.zeros(scores_t2i.shape[0])\n\n for index, score in enumerate(scores_t2i):\n inds = np.argsort(score)[::-1]\n ranks[index] = np.where(inds == txt2img[index])[0][0]\n\n # Compute metrics\n ir1 = 100.0 * len(np.where(ranks < 1)[0]) / len(ranks)\n ir5 = 100.0 * len(np.where(ranks < 5)[0]) / len(ranks)\n ir10 = 100.0 * len(np.where(ranks < 10)[0]) / len(ranks)\n\n tr_mean = (tr1 + tr5 + tr10) / 3\n ir_mean = (ir1 + ir5 + ir10) / 3\n r_mean = (tr_mean + ir_mean) / 2\n\n agg_metrics = (tr1 + tr5 + tr10) / 3\n\n eval_result = {\n \"txt_r1\": tr1,\n \"txt_r5\": tr5,\n \"txt_r10\": tr10,\n \"txt_r_mean\": tr_mean,\n \"img_r1\": ir1,\n \"img_r5\": ir5,\n \"img_r10\": ir10,\n \"img_r_mean\": ir_mean,\n \"r_mean\": r_mean,\n \"agg_metrics\": agg_metrics,\n }\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(eval_result) + \"\\n\")\n return eval_result","source_hash":"dc43b8ba0b9ae99526cd0fd8edb054202a342b540b3e421d4dc017ed60d83bb5","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task","uri":"program://CREMA/module/lavis.tasks.base_task#L1-L345","kind":"module","name":"lavis.tasks.base_task","path":"lavis/tasks/base_task.py","language":"python","start_line":1,"end_line":345,"context_start_line":1,"context_end_line":345,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process, is_dist_avail_and_initialized\nfrom lavis.common.logger import MetricLogger, SmoothedValue\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n return results\n\n def train_epoch(\n self,\n epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n iters_per_epoch=len(data_loader),\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def train_iters(\n self,\n epoch,\n start_iters,\n iters_per_inner_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n start_iters=start_iters,\n iters_per_epoch=iters_per_inner_epoch,\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def _train_inner_loop(\n self,\n epoch,\n iters_per_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n start_iters=None,\n log_freq=50,\n cuda_enabled=False,\n accum_grad_iters=1,\n ):\n \"\"\"\n An inner training loop compatible with both epoch-based and iter-based training.\n\n When using epoch-based, training stops after one epoch; when using iter-based,\n training stops after #iters_per_epoch iterations.\n \"\"\"\n def check_gradient(module, grad_input, grad_output):\n print(f\"Module: {module}\")\n print(f\"Gradient input: {grad_input}\")\n print(f\"Gradient output: {grad_output}\")\n \n\n use_amp = scaler is not None\n\n if not hasattr(data_loader, \"__next__\"):\n # convert to iterator if not already\n data_loader = iter(data_loader)\n\n metric_logger = MetricLogger(delimiter=\" \")\n metric_logger.add_meter(\"lr\", SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n metric_logger.add_meter(\"loss\", SmoothedValue(window_size=1, fmt=\"{value:.4f}\"))\n\n relu = nn.ReLU(inplace=True)\n tanh = nn.Tanh()\n # if iter-based runner, schedule lr based on inner epoch.\n logging.info(\n \"Start training epoch {}, {} iters per inner epoch.\".format(\n epoch, iters_per_epoch\n )\n )\n header = \"Train: data epoch: [{}]\".format(epoch)\n if start_iters is None:\n # epoch-based runner\n inner_epoch = epoch\n else:\n # In iter-based runner, we schedule the learning rate based on iterations.\n inner_epoch = start_iters // iters_per_epoch\n header = header + \"; inner epoch [{}]\".format(inner_epoch)\n\n for i in metric_logger.log_every(range(iters_per_epoch), log_freq, header):\n\n # hooks = []\n # for layer in model.children():\n # hook = layer.register_backward_hook(check_gradient)\n # hooks.append(hook)\n # if using iter-based runner, we stop after iters_per_epoch iterations.\n if i >= iters_per_epoch:\n break\n\n samples = next(data_loader)\n\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n samples.update(\n {\n \"epoch\": inner_epoch,\n \"num_iters_per_epoch\": iters_per_epoch,\n \"iters\": i,\n }\n )\n\n lr_scheduler.step(cur_epoch=inner_epoch, cur_step=i)\n\n with torch.cuda.amp.autocast(enabled=use_amp):\n loss = self.train_step(model=model, samples=samples)\n # print('loss', loss)\n # after_train_step()\n if use_amp:\n scaler.scale(loss).backward()\n else:\n loss.backward()\n\n # print('here after backward')\n # modalilites = model.module.modalities\n \n # if 'rgb' in modalilites:\n # rgb_grad = model.module.t5_proj_rgb.weight.grad.mean()\n # if 'flow' in modalilites:\n # other_grad = model.module.t5_proj_flow.weight.grad.mean()\n # if 'norm' in modalilites:\n # other_grad = model.module.t5_proj_norm.weight.grad.mean()\n # if 'depth' in modalilites:\n # other_grad = model.module.t5_proj_depth.weight.grad.mean()\n \n # ratio_r = rgb_grad / other_grad\n # ratio_o = 1 / ratio_r\n # # print(model)\n\n # if ratio_r > 1:\n # coeff_r = 1 - tanh(0.1 * relu(ratio_r))\n # coeff_o = 1\n # else:\n # coeff_o = 1 - tanh(0.1 * relu(ratio_o))\n # coeff_r = 1\n\n # if 0 <= epoch <= 10: # bug fixed\n # for layer in model.module.Qformer.bert.encoder.layer:\n # for name, parms in layer.attention.self.named_parameters():\n # if 'rgb' in name: \n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_r \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # if 'norm' in name or 'flow' in name or 'depth' in name:\n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_o \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # else:\n # pass\n\n # print('w_a_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_rgb.weight.grad)\n # print('w_b_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_rgb.weight.grad)\n # print('w_a_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_flow.weight.grad)\n # print('w_b_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_flow.weight.grad)\n # print(type(model))\n\n # update gradients every accum_grad_iters iterations\n if (i + 1) % accum_grad_iters == 0:\n if use_amp:\n scaler.step(optimizer)\n scaler.update() \n else: \n optimizer.step()\n optimizer.zero_grad()\n\n metric_logger.update(loss=loss.item())\n metric_logger.update(lr=optimizer.param_groups[0][\"lr\"])\n\n # for hook in hooks:\n # hook.remove()\n\n # after train_epoch()\n # gather the stats from all processes\n metric_logger.synchronize_between_processes()\n logging.info(\"Averaged stats: \" + str(metric_logger.global_avg()))\n return {\n k: \"{:.3f}\".format(meter.global_avg)\n for k, meter in metric_logger.meters.items()\n }\n\n @staticmethod\n def save_result(result, result_dir, filename, remove_duplicate=\"\"):\n import json\n\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, get_rank())\n )\n final_result_file = os.path.join(result_dir, \"%s.json\" % filename)\n\n json.dump(result, open(result_file, \"w\"))\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, rank)\n )\n res = json.load(open(result_file, \"r\"))\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n json.dump(result, open(final_result_file, \"w\"))\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.BaseTask","uri":"program://CREMA/class/lavis.tasks.base_task.BaseTask#L20-L345","kind":"class","name":"BaseTask","path":"lavis/tasks/base_task.py","language":"python","start_line":20,"end_line":345,"context_start_line":1,"context_end_line":345,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process, is_dist_avail_and_initialized\nfrom lavis.common.logger import MetricLogger, SmoothedValue\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n return results\n\n def train_epoch(\n self,\n epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n iters_per_epoch=len(data_loader),\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def train_iters(\n self,\n epoch,\n start_iters,\n iters_per_inner_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n start_iters=start_iters,\n iters_per_epoch=iters_per_inner_epoch,\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def _train_inner_loop(\n self,\n epoch,\n iters_per_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n start_iters=None,\n log_freq=50,\n cuda_enabled=False,\n accum_grad_iters=1,\n ):\n \"\"\"\n An inner training loop compatible with both epoch-based and iter-based training.\n\n When using epoch-based, training stops after one epoch; when using iter-based,\n training stops after #iters_per_epoch iterations.\n \"\"\"\n def check_gradient(module, grad_input, grad_output):\n print(f\"Module: {module}\")\n print(f\"Gradient input: {grad_input}\")\n print(f\"Gradient output: {grad_output}\")\n \n\n use_amp = scaler is not None\n\n if not hasattr(data_loader, \"__next__\"):\n # convert to iterator if not already\n data_loader = iter(data_loader)\n\n metric_logger = MetricLogger(delimiter=\" \")\n metric_logger.add_meter(\"lr\", SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n metric_logger.add_meter(\"loss\", SmoothedValue(window_size=1, fmt=\"{value:.4f}\"))\n\n relu = nn.ReLU(inplace=True)\n tanh = nn.Tanh()\n # if iter-based runner, schedule lr based on inner epoch.\n logging.info(\n \"Start training epoch {}, {} iters per inner epoch.\".format(\n epoch, iters_per_epoch\n )\n )\n header = \"Train: data epoch: [{}]\".format(epoch)\n if start_iters is None:\n # epoch-based runner\n inner_epoch = epoch\n else:\n # In iter-based runner, we schedule the learning rate based on iterations.\n inner_epoch = start_iters // iters_per_epoch\n header = header + \"; inner epoch [{}]\".format(inner_epoch)\n\n for i in metric_logger.log_every(range(iters_per_epoch), log_freq, header):\n\n # hooks = []\n # for layer in model.children():\n # hook = layer.register_backward_hook(check_gradient)\n # hooks.append(hook)\n # if using iter-based runner, we stop after iters_per_epoch iterations.\n if i >= iters_per_epoch:\n break\n\n samples = next(data_loader)\n\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n samples.update(\n {\n \"epoch\": inner_epoch,\n \"num_iters_per_epoch\": iters_per_epoch,\n \"iters\": i,\n }\n )\n\n lr_scheduler.step(cur_epoch=inner_epoch, cur_step=i)\n\n with torch.cuda.amp.autocast(enabled=use_amp):\n loss = self.train_step(model=model, samples=samples)\n # print('loss', loss)\n # after_train_step()\n if use_amp:\n scaler.scale(loss).backward()\n else:\n loss.backward()\n\n # print('here after backward')\n # modalilites = model.module.modalities\n \n # if 'rgb' in modalilites:\n # rgb_grad = model.module.t5_proj_rgb.weight.grad.mean()\n # if 'flow' in modalilites:\n # other_grad = model.module.t5_proj_flow.weight.grad.mean()\n # if 'norm' in modalilites:\n # other_grad = model.module.t5_proj_norm.weight.grad.mean()\n # if 'depth' in modalilites:\n # other_grad = model.module.t5_proj_depth.weight.grad.mean()\n \n # ratio_r = rgb_grad / other_grad\n # ratio_o = 1 / ratio_r\n # # print(model)\n\n # if ratio_r > 1:\n # coeff_r = 1 - tanh(0.1 * relu(ratio_r))\n # coeff_o = 1\n # else:\n # coeff_o = 1 - tanh(0.1 * relu(ratio_o))\n # coeff_r = 1\n\n # if 0 <= epoch <= 10: # bug fixed\n # for layer in model.module.Qformer.bert.encoder.layer:\n # for name, parms in layer.attention.self.named_parameters():\n # if 'rgb' in name: \n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_r \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # if 'norm' in name or 'flow' in name or 'depth' in name:\n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_o \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # else:\n # pass\n\n # print('w_a_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_rgb.weight.grad)\n # print('w_b_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_rgb.weight.grad)\n # print('w_a_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_flow.weight.grad)\n # print('w_b_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_flow.weight.grad)\n # print(type(model))\n\n # update gradients every accum_grad_iters iterations\n if (i + 1) % accum_grad_iters == 0:\n if use_amp:\n scaler.step(optimizer)\n scaler.update() \n else: \n optimizer.step()\n optimizer.zero_grad()\n\n metric_logger.update(loss=loss.item())\n metric_logger.update(lr=optimizer.param_groups[0][\"lr\"])\n\n # for hook in hooks:\n # hook.remove()\n\n # after train_epoch()\n # gather the stats from all processes\n metric_logger.synchronize_between_processes()\n logging.info(\"Averaged stats: \" + str(metric_logger.global_avg()))\n return {\n k: \"{:.3f}\".format(meter.global_avg)\n for k, meter in metric_logger.meters.items()\n }\n\n @staticmethod\n def save_result(result, result_dir, filename, remove_duplicate=\"\"):\n import json\n\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, get_rank())\n )\n final_result_file = os.path.join(result_dir, \"%s.json\" % filename)\n\n json.dump(result, open(result_file, \"w\"))\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, rank)\n )\n res = json.load(open(result_file, \"r\"))\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n json.dump(result, open(final_result_file, \"w\"))\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.__init__","uri":"program://CREMA/function/lavis.tasks.base_task.__init__#L21-L24","kind":"function","name":"__init__","path":"lavis/tasks/base_task.py","language":"python","start_line":21,"end_line":24,"context_start_line":1,"context_end_line":44,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport logging\nimport os\n\nimport torch\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process, is_dist_avail_and_initialized\nfrom lavis.common.logger import MetricLogger, SmoothedValue\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.setup_task","uri":"program://CREMA/function/lavis.tasks.base_task.setup_task#L27-L28","kind":"function","name":"setup_task","path":"lavis/tasks/base_task.py","language":"python","start_line":27,"end_line":28,"context_start_line":7,"context_end_line":48,"code":"\nimport logging\nimport os\n\nimport torch\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process, is_dist_avail_and_initialized\nfrom lavis.common.logger import MetricLogger, SmoothedValue\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.build_model","uri":"program://CREMA/function/lavis.tasks.base_task.build_model#L30-L34","kind":"function","name":"build_model","path":"lavis/tasks/base_task.py","language":"python","start_line":30,"end_line":34,"context_start_line":10,"context_end_line":54,"code":"\nimport torch\nimport torch.distributed as dist\nfrom lavis.common.dist_utils import get_rank, get_world_size, is_main_process, is_dist_avail_and_initialized\nfrom lavis.common.logger import MetricLogger, SmoothedValue\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.build_datasets","uri":"program://CREMA/function/lavis.tasks.base_task.build_datasets#L36-L60","kind":"function","name":"build_datasets","path":"lavis/tasks/base_task.py","language":"python","start_line":36,"end_line":60,"context_start_line":16,"context_end_line":80,"code":"from lavis.datasets.data_utils import prepare_sample\n\nimport torch.nn as nn\n\nclass BaseTask:\n def __init__(self, **kwargs):\n super().__init__()\n\n self.inst_id_key = \"instance_id\"\n\n @classmethod\n def setup_task(cls, **kwargs):\n return cls()\n\n def build_model(self, cfg):\n model_config = cfg.model_cfg\n\n model_cls = registry.get_model_class(model_config.arch)\n return model_cls.from_config(model_config)\n\n def build_datasets(self, cfg):\n \"\"\"\n Build a dictionary of datasets, keyed by split 'train', 'valid', 'test'.\n Download dataset and annotations automatically if not exist.\n\n Args:\n cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.train_step","uri":"program://CREMA/function/lavis.tasks.base_task.train_step#L62-L64","kind":"function","name":"train_step","path":"lavis/tasks/base_task.py","language":"python","start_line":62,"end_line":64,"context_start_line":42,"context_end_line":84,"code":" cfg (common.config.Config): _description_\n\n Returns:\n dict: Dictionary of torch.utils.data.Dataset objects by split.\n \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.valid_step","uri":"program://CREMA/function/lavis.tasks.base_task.valid_step#L66-L67","kind":"function","name":"valid_step","path":"lavis/tasks/base_task.py","language":"python","start_line":66,"end_line":67,"context_start_line":46,"context_end_line":87,"code":" \"\"\"\n\n datasets = dict()\n\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.before_evaluation","uri":"program://CREMA/function/lavis.tasks.base_task.before_evaluation#L69-L70","kind":"function","name":"before_evaluation","path":"lavis/tasks/base_task.py","language":"python","start_line":69,"end_line":70,"context_start_line":49,"context_end_line":90,"code":"\n datasets_config = cfg.datasets_cfg\n assert len(datasets_config) > 0, \"At least one dataset has to be specified.\"\n\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.after_evaluation","uri":"program://CREMA/function/lavis.tasks.base_task.after_evaluation#L72-L73","kind":"function","name":"after_evaluation","path":"lavis/tasks/base_task.py","language":"python","start_line":72,"end_line":73,"context_start_line":52,"context_end_line":93,"code":"\n for name in datasets_config:\n dataset_config = datasets_config[name]\n builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.inference_step","uri":"program://CREMA/function/lavis.tasks.base_task.inference_step#L75-L76","kind":"function","name":"inference_step","path":"lavis/tasks/base_task.py","language":"python","start_line":75,"end_line":76,"context_start_line":55,"context_end_line":96,"code":" builder = registry.get_builder_class(name)(dataset_config)\n dataset = builder.build_datasets()\n\n datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n return results","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.evaluation","uri":"program://CREMA/function/lavis.tasks.base_task.evaluation#L78-L96","kind":"function","name":"evaluation","path":"lavis/tasks/base_task.py","language":"python","start_line":78,"end_line":96,"context_start_line":58,"context_end_line":116,"code":" datasets[name] = dataset\n\n return datasets\n\n def train_step(self, model, samples):\n loss = model(samples)[\"loss\"]\n return loss\n\n def valid_step(self, model, samples):\n raise NotImplementedError\n\n def before_evaluation(self, model, dataset, **kwargs):\n model.before_evaluation(dataset=dataset, task_type=type(self))\n\n def after_evaluation(self, **kwargs):\n pass\n\n def inference_step(self):\n raise NotImplementedError\n\n def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n return results\n\n def train_epoch(\n self,\n epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n iters_per_epoch=len(data_loader),\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.train_epoch","uri":"program://CREMA/function/lavis.tasks.base_task.train_epoch#L98-L121","kind":"function","name":"train_epoch","path":"lavis/tasks/base_task.py","language":"python","start_line":98,"end_line":121,"context_start_line":78,"context_end_line":141,"code":" def evaluation(self, model, data_loader, cuda_enabled=True):\n metric_logger = MetricLogger(delimiter=\" \")\n header = \"Evaluation\"\n # TODO make it configurable\n print_freq = 10\n\n results = []\n\n for samples in metric_logger.log_every(data_loader, print_freq, header):\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n\n eval_output = self.valid_step(model=model, samples=samples)\n results.extend(eval_output)\n #break\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n return results\n\n def train_epoch(\n self,\n epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n iters_per_epoch=len(data_loader),\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def train_iters(\n self,\n epoch,\n start_iters,\n iters_per_inner_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n start_iters=start_iters,\n iters_per_epoch=iters_per_inner_epoch,\n model=model,","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.train_iters","uri":"program://CREMA/function/lavis.tasks.base_task.train_iters#L123-L149","kind":"function","name":"train_iters","path":"lavis/tasks/base_task.py","language":"python","start_line":123,"end_line":149,"context_start_line":103,"context_end_line":169,"code":" optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n iters_per_epoch=len(data_loader),\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def train_iters(\n self,\n epoch,\n start_iters,\n iters_per_inner_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n start_iters=start_iters,\n iters_per_epoch=iters_per_inner_epoch,\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def _train_inner_loop(\n self,\n epoch,\n iters_per_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n start_iters=None,\n log_freq=50,\n cuda_enabled=False,\n accum_grad_iters=1,\n ):\n \"\"\"\n An inner training loop compatible with both epoch-based and iter-based training.\n\n When using epoch-based, training stops after one epoch; when using iter-based,\n training stops after #iters_per_epoch iterations.","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task._train_inner_loop","uri":"program://CREMA/function/lavis.tasks.base_task._train_inner_loop#L151-L305","kind":"function","name":"_train_inner_loop","path":"lavis/tasks/base_task.py","language":"python","start_line":151,"end_line":305,"context_start_line":131,"context_end_line":325,"code":" lr_scheduler,\n scaler=None,\n cuda_enabled=False,\n log_freq=50,\n accum_grad_iters=1,\n ):\n return self._train_inner_loop(\n epoch=epoch,\n start_iters=start_iters,\n iters_per_epoch=iters_per_inner_epoch,\n model=model,\n data_loader=data_loader,\n optimizer=optimizer,\n scaler=scaler,\n lr_scheduler=lr_scheduler,\n log_freq=log_freq,\n cuda_enabled=cuda_enabled,\n accum_grad_iters=accum_grad_iters,\n )\n\n def _train_inner_loop(\n self,\n epoch,\n iters_per_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n start_iters=None,\n log_freq=50,\n cuda_enabled=False,\n accum_grad_iters=1,\n ):\n \"\"\"\n An inner training loop compatible with both epoch-based and iter-based training.\n\n When using epoch-based, training stops after one epoch; when using iter-based,\n training stops after #iters_per_epoch iterations.\n \"\"\"\n def check_gradient(module, grad_input, grad_output):\n print(f\"Module: {module}\")\n print(f\"Gradient input: {grad_input}\")\n print(f\"Gradient output: {grad_output}\")\n \n\n use_amp = scaler is not None\n\n if not hasattr(data_loader, \"__next__\"):\n # convert to iterator if not already\n data_loader = iter(data_loader)\n\n metric_logger = MetricLogger(delimiter=\" \")\n metric_logger.add_meter(\"lr\", SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n metric_logger.add_meter(\"loss\", SmoothedValue(window_size=1, fmt=\"{value:.4f}\"))\n\n relu = nn.ReLU(inplace=True)\n tanh = nn.Tanh()\n # if iter-based runner, schedule lr based on inner epoch.\n logging.info(\n \"Start training epoch {}, {} iters per inner epoch.\".format(\n epoch, iters_per_epoch\n )\n )\n header = \"Train: data epoch: [{}]\".format(epoch)\n if start_iters is None:\n # epoch-based runner\n inner_epoch = epoch\n else:\n # In iter-based runner, we schedule the learning rate based on iterations.\n inner_epoch = start_iters // iters_per_epoch\n header = header + \"; inner epoch [{}]\".format(inner_epoch)\n\n for i in metric_logger.log_every(range(iters_per_epoch), log_freq, header):\n\n # hooks = []\n # for layer in model.children():\n # hook = layer.register_backward_hook(check_gradient)\n # hooks.append(hook)\n # if using iter-based runner, we stop after iters_per_epoch iterations.\n if i >= iters_per_epoch:\n break\n\n samples = next(data_loader)\n\n samples = prepare_sample(samples, cuda_enabled=cuda_enabled)\n samples.update(\n {\n \"epoch\": inner_epoch,\n \"num_iters_per_epoch\": iters_per_epoch,\n \"iters\": i,\n }\n )\n\n lr_scheduler.step(cur_epoch=inner_epoch, cur_step=i)\n\n with torch.cuda.amp.autocast(enabled=use_amp):\n loss = self.train_step(model=model, samples=samples)\n # print('loss', loss)\n # after_train_step()\n if use_amp:\n scaler.scale(loss).backward()\n else:\n loss.backward()\n\n # print('here after backward')\n # modalilites = model.module.modalities\n \n # if 'rgb' in modalilites:\n # rgb_grad = model.module.t5_proj_rgb.weight.grad.mean()\n # if 'flow' in modalilites:\n # other_grad = model.module.t5_proj_flow.weight.grad.mean()\n # if 'norm' in modalilites:\n # other_grad = model.module.t5_proj_norm.weight.grad.mean()\n # if 'depth' in modalilites:\n # other_grad = model.module.t5_proj_depth.weight.grad.mean()\n \n # ratio_r = rgb_grad / other_grad\n # ratio_o = 1 / ratio_r\n # # print(model)\n\n # if ratio_r > 1:\n # coeff_r = 1 - tanh(0.1 * relu(ratio_r))\n # coeff_o = 1\n # else:\n # coeff_o = 1 - tanh(0.1 * relu(ratio_o))\n # coeff_r = 1\n\n # if 0 <= epoch <= 10: # bug fixed\n # for layer in model.module.Qformer.bert.encoder.layer:\n # for name, parms in layer.attention.self.named_parameters():\n # if 'rgb' in name: \n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_r \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # if 'norm' in name or 'flow' in name or 'depth' in name:\n # if parms.grad is not None:\n # # print(name)\n # parms.grad = parms.grad * coeff_o \n # # + \\\n # # torch.zeros_like(parms.grad).normal_(0, parms.grad.std().item() + 1e-8)\n # else:\n # pass\n\n # print('w_a_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_rgb.weight.grad)\n # print('w_b_rgb', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_rgb.weight.grad)\n # print('w_a_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_a_flow.weight.grad)\n # print('w_b_flow', model.module.Qformer.bert.encoder.layer[-1].attention.self.query.w_b_flow.weight.grad)\n # print(type(model))\n\n # update gradients every accum_grad_iters iterations\n if (i + 1) % accum_grad_iters == 0:\n if use_amp:\n scaler.step(optimizer)\n scaler.update() \n else: \n optimizer.step()\n optimizer.zero_grad()\n\n metric_logger.update(loss=loss.item())\n metric_logger.update(lr=optimizer.param_groups[0][\"lr\"])\n\n # for hook in hooks:\n # hook.remove()\n\n # after train_epoch()\n # gather the stats from all processes\n metric_logger.synchronize_between_processes()\n logging.info(\"Averaged stats: \" + str(metric_logger.global_avg()))\n return {\n k: \"{:.3f}\".format(meter.global_avg)\n for k, meter in metric_logger.meters.items()\n }\n\n @staticmethod\n def save_result(result, result_dir, filename, remove_duplicate=\"\"):\n import json\n\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, get_rank())\n )\n final_result_file = os.path.join(result_dir, \"%s.json\" % filename)\n\n json.dump(result, open(result_file, \"w\"))\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.save_result","uri":"program://CREMA/function/lavis.tasks.base_task.save_result#L308-L345","kind":"function","name":"save_result","path":"lavis/tasks/base_task.py","language":"python","start_line":308,"end_line":345,"context_start_line":288,"context_end_line":345,"code":" else: \n optimizer.step()\n optimizer.zero_grad()\n\n metric_logger.update(loss=loss.item())\n metric_logger.update(lr=optimizer.param_groups[0][\"lr\"])\n\n # for hook in hooks:\n # hook.remove()\n\n # after train_epoch()\n # gather the stats from all processes\n metric_logger.synchronize_between_processes()\n logging.info(\"Averaged stats: \" + str(metric_logger.global_avg()))\n return {\n k: \"{:.3f}\".format(meter.global_avg)\n for k, meter in metric_logger.meters.items()\n }\n\n @staticmethod\n def save_result(result, result_dir, filename, remove_duplicate=\"\"):\n import json\n\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, get_rank())\n )\n final_result_file = os.path.join(result_dir, \"%s.json\" % filename)\n\n json.dump(result, open(result_file, \"w\"))\n\n if is_dist_avail_and_initialized():\n dist.barrier()\n\n if is_main_process():\n logging.warning(\"rank %d starts merging results.\" % get_rank())\n # combine results from all processes\n result = []\n\n for rank in range(get_world_size()):\n result_file = os.path.join(\n result_dir, \"%s_rank%d.json\" % (filename, rank)\n )\n res = json.load(open(result_file, \"r\"))\n result += res\n\n if remove_duplicate:\n result_new = []\n id_list = []\n for res in result:\n if res[remove_duplicate] not in id_list:\n id_list.append(res[remove_duplicate])\n result_new.append(res)\n result = result_new\n\n json.dump(result, open(final_result_file, \"w\"))\n print(\"result file saved to %s\" % final_result_file)\n\n return final_result_file","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.base_task.check_gradient","uri":"program://CREMA/function/lavis.tasks.base_task.check_gradient#L171-L174","kind":"function","name":"check_gradient","path":"lavis/tasks/base_task.py","language":"python","start_line":171,"end_line":174,"context_start_line":151,"context_end_line":194,"code":" def _train_inner_loop(\n self,\n epoch,\n iters_per_epoch,\n model,\n data_loader,\n optimizer,\n lr_scheduler,\n scaler=None,\n start_iters=None,\n log_freq=50,\n cuda_enabled=False,\n accum_grad_iters=1,\n ):\n \"\"\"\n An inner training loop compatible with both epoch-based and iter-based training.\n\n When using epoch-based, training stops after one epoch; when using iter-based,\n training stops after #iters_per_epoch iterations.\n \"\"\"\n def check_gradient(module, grad_input, grad_output):\n print(f\"Module: {module}\")\n print(f\"Gradient input: {grad_input}\")\n print(f\"Gradient output: {grad_output}\")\n \n\n use_amp = scaler is not None\n\n if not hasattr(data_loader, \"__next__\"):\n # convert to iterator if not already\n data_loader = iter(data_loader)\n\n metric_logger = MetricLogger(delimiter=\" \")\n metric_logger.add_meter(\"lr\", SmoothedValue(window_size=1, fmt=\"{value:.6f}\"))\n metric_logger.add_meter(\"loss\", SmoothedValue(window_size=1, fmt=\"{value:.4f}\"))\n\n relu = nn.ReLU(inplace=True)\n tanh = nn.Tanh()\n # if iter-based runner, schedule lr based on inner epoch.\n logging.info(\n \"Start training epoch {}, {} iters per inner epoch.\".format(\n epoch, iters_per_epoch\n )\n )","source_hash":"0d289b8091b17f33204a69e11eb319d7669710afe5a2a078d39652dbd5bcd08c","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification","uri":"program://CREMA/module/lavis.tasks.multimodal_classification#L1-L83","kind":"module","name":"lavis.tasks.multimodal_classification","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":1,"end_line":83,"context_start_line":1,"context_end_line":83,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport logging\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"multimodal_classification\")\nclass MultimodalClassificationTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.predict(samples)\n\n predictions = outputs[\"predictions\"]\n targets = outputs[\"targets\"]\n\n predictions = predictions.max(1)[1].cpu().numpy()\n targets = targets.cpu().numpy()\n\n indices = samples[self.inst_id_key]\n\n for pred, tgt, index in zip(predictions, targets, indices):\n if isinstance(index, torch.Tensor):\n index = index.item()\n\n results.append(\n {\n self.inst_id_key: index,\n \"prediction\": pred.item(),\n \"target\": tgt.item(),\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=self.inst_id_key,\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n\n predictions = np.array([res[\"prediction\"] for res in results])\n targets = np.array([res[\"target\"] for res in results])\n\n accuracy = (targets == predictions).sum() / targets.shape[0]\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification.MultimodalClassificationTask","uri":"program://CREMA/class/lavis.tasks.multimodal_classification.MultimodalClassificationTask#L20-L83","kind":"class","name":"MultimodalClassificationTask","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":20,"end_line":83,"context_start_line":1,"context_end_line":83,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport logging\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"multimodal_classification\")\nclass MultimodalClassificationTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.predict(samples)\n\n predictions = outputs[\"predictions\"]\n targets = outputs[\"targets\"]\n\n predictions = predictions.max(1)[1].cpu().numpy()\n targets = targets.cpu().numpy()\n\n indices = samples[self.inst_id_key]\n\n for pred, tgt, index in zip(predictions, targets, indices):\n if isinstance(index, torch.Tensor):\n index = index.item()\n\n results.append(\n {\n self.inst_id_key: index,\n \"prediction\": pred.item(),\n \"target\": tgt.item(),\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=self.inst_id_key,\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n\n predictions = np.array([res[\"prediction\"] for res in results])\n targets = np.array([res[\"target\"] for res in results])\n\n accuracy = (targets == predictions).sum() / targets.shape[0]\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification.__init__","uri":"program://CREMA/function/lavis.tasks.multimodal_classification.__init__#L21-L22","kind":"function","name":"__init__","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":21,"end_line":22,"context_start_line":1,"context_end_line":42,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport logging\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"multimodal_classification\")\nclass MultimodalClassificationTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.predict(samples)\n\n predictions = outputs[\"predictions\"]\n targets = outputs[\"targets\"]\n\n predictions = predictions.max(1)[1].cpu().numpy()\n targets = targets.cpu().numpy()\n\n indices = samples[self.inst_id_key]\n\n for pred, tgt, index in zip(predictions, targets, indices):\n if isinstance(index, torch.Tensor):\n index = index.item()\n\n results.append(\n {","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification.valid_step","uri":"program://CREMA/function/lavis.tasks.multimodal_classification.valid_step#L24-L49","kind":"function","name":"valid_step","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":24,"end_line":49,"context_start_line":4,"context_end_line":69,"code":" SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport json\nimport os\nimport logging\n\nimport numpy as np\nimport torch\nfrom lavis.common.dist_utils import main_process\nfrom lavis.common.registry import registry\nfrom lavis.tasks.base_task import BaseTask\n\n\n@registry.register_task(\"multimodal_classification\")\nclass MultimodalClassificationTask(BaseTask):\n def __init__(self):\n super().__init__()\n\n def valid_step(self, model, samples):\n results = []\n\n outputs = model.predict(samples)\n\n predictions = outputs[\"predictions\"]\n targets = outputs[\"targets\"]\n\n predictions = predictions.max(1)[1].cpu().numpy()\n targets = targets.cpu().numpy()\n\n indices = samples[self.inst_id_key]\n\n for pred, tgt, index in zip(predictions, targets, indices):\n if isinstance(index, torch.Tensor):\n index = index.item()\n\n results.append(\n {\n self.inst_id_key: index,\n \"prediction\": pred.item(),\n \"target\": tgt.item(),\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=self.inst_id_key,\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n\n predictions = np.array([res[\"prediction\"] for res in results])","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification.after_evaluation","uri":"program://CREMA/function/lavis.tasks.multimodal_classification.after_evaluation#L51-L63","kind":"function","name":"after_evaluation","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":51,"end_line":63,"context_start_line":31,"context_end_line":83,"code":"\n predictions = predictions.max(1)[1].cpu().numpy()\n targets = targets.cpu().numpy()\n\n indices = samples[self.inst_id_key]\n\n for pred, tgt, index in zip(predictions, targets, indices):\n if isinstance(index, torch.Tensor):\n index = index.item()\n\n results.append(\n {\n self.inst_id_key: index,\n \"prediction\": pred.item(),\n \"target\": tgt.item(),\n }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=self.inst_id_key,\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n\n predictions = np.array([res[\"prediction\"] for res in results])\n targets = np.array([res[\"target\"] for res in results])\n\n accuracy = (targets == predictions).sum() / targets.shape[0]\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.tasks.multimodal_classification._report_metrics","uri":"program://CREMA/function/lavis.tasks.multimodal_classification._report_metrics#L66-L83","kind":"function","name":"_report_metrics","path":"lavis/tasks/multimodal_classification.py","language":"python","start_line":66,"end_line":83,"context_start_line":46,"context_end_line":83,"code":" }\n )\n\n return results\n\n def after_evaluation(self, val_result, split_name, epoch, **kwargs):\n eval_result_file = self.save_result(\n result=val_result,\n result_dir=registry.get_path(\"result_dir\"),\n filename=\"{}_epoch{}\".format(split_name, epoch),\n remove_duplicate=self.inst_id_key,\n )\n\n metrics = self._report_metrics(\n eval_result_file=eval_result_file, split_name=split_name\n )\n\n return metrics\n\n @main_process\n def _report_metrics(self, eval_result_file, split_name):\n results = json.load(open(eval_result_file))\n\n predictions = np.array([res[\"prediction\"] for res in results])\n targets = np.array([res[\"target\"] for res in results])\n\n accuracy = (targets == predictions).sum() / targets.shape[0]\n metrics = {\"agg_metrics\": accuracy, \"acc\": accuracy}\n\n log_stats = {split_name: {k: v for k, v in metrics.items()}}\n\n with open(\n os.path.join(registry.get_path(\"output_dir\"), \"evaluate.txt\"), \"a\"\n ) as f:\n f.write(json.dumps(log_stats) + \"\\n\")\n\n logging.info(metrics)\n return metrics","source_hash":"9c52dee7416b923e5bd8e0a15a9aebbedcfa84fd7a2a677598e5cc437c0a0544","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor","uri":"program://CREMA/module/lavis.processors.base_processor#L1-L26","kind":"module","name":"lavis.processors.base_processor","path":"lavis/processors/base_processor.py","language":"python","start_line":1,"end_line":26,"context_start_line":1,"context_end_line":26,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor.BaseProcessor","uri":"program://CREMA/class/lavis.processors.base_processor.BaseProcessor#L11-L26","kind":"class","name":"BaseProcessor","path":"lavis/processors/base_processor.py","language":"python","start_line":11,"end_line":26,"context_start_line":1,"context_end_line":26,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor.__init__","uri":"program://CREMA/function/lavis.processors.base_processor.__init__#L12-L14","kind":"function","name":"__init__","path":"lavis/processors/base_processor.py","language":"python","start_line":12,"end_line":14,"context_start_line":1,"context_end_line":26,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor.__call__","uri":"program://CREMA/function/lavis.processors.base_processor.__call__#L16-L17","kind":"function","name":"__call__","path":"lavis/processors/base_processor.py","language":"python","start_line":16,"end_line":17,"context_start_line":1,"context_end_line":26,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor.from_config","uri":"program://CREMA/function/lavis.processors.base_processor.from_config#L20-L21","kind":"function","name":"from_config","path":"lavis/processors/base_processor.py","language":"python","start_line":20,"end_line":21,"context_start_line":1,"context_end_line":26,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.base_processor.build","uri":"program://CREMA/function/lavis.processors.base_processor.build#L23-L26","kind":"function","name":"build","path":"lavis/processors/base_processor.py","language":"python","start_line":23,"end_line":26,"context_start_line":3,"context_end_line":26,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom omegaconf import OmegaConf\n\n\nclass BaseProcessor:\n def __init__(self):\n self.transform = lambda x: x\n return\n\n def __call__(self, item):\n return self.transform(item)\n\n @classmethod\n def from_config(cls, cfg=None):\n return cls()\n\n def build(self, **kwargs):\n cfg = OmegaConf.create(kwargs)\n\n return self.from_config(cfg)","source_hash":"08c0ed8b96f9a703cc03b0bfa4b44df2c05c8170aa031b03df9dd214d6943b5f","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors","uri":"program://CREMA/module/lavis.processors.alpro_processors#L1-L216","kind":"module","name":"lavis.processors.alpro_processors","path":"lavis/processors/alpro_processors.py","language":"python","start_line":1,"end_line":216,"context_start_line":1,"context_end_line":216,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import load_video\nfrom lavis.processors import transforms_video\nfrom lavis.processors.base_processor import BaseProcessor\nfrom lavis.processors.randaugment import VideoRandomAugment\nfrom lavis.processors import functional_video as F\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\n\nMAX_INT = registry.get(\"MAX_INT\")\n\n\nclass AlproVideoBaseProcessor(BaseProcessor):\n def __init__(self, mean=None, std=None, n_frms=MAX_INT):\n if mean is None:\n mean = (0.48145466, 0.4578275, 0.40821073)\n if std is None:\n std = (0.26862954, 0.26130258, 0.27577711)\n\n self.normalize = transforms_video.NormalizeVideo(mean, std)\n\n self.n_frms = n_frms\n\n\nclass ToUint8(object):\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.to(torch.uint8)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ToTHWC(object):\n \"\"\"\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (C, T, H, W)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (T, H, W, C)\n \"\"\"\n\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.permute(1, 2, 3, 0)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ResizeVideo(object):\n def __init__(self, target_size, interpolation_mode=\"bilinear\"):\n self.target_size = target_size\n self.interpolation_mode = interpolation_mode\n\n def __call__(self, clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: central cropping of video clip. Size is\n (C, T, crop_size, crop_size)\n \"\"\"\n return F.resize(clip, self.target_size, self.interpolation_mode)\n\n def __repr__(self):\n return self.__class__.__name__ + \"(resize_size={0})\".format(self.target_size)\n\n\n@registry.register_processor(\"alpro_video_train\")\nclass AlproVideoTrainProcessor(AlproVideoBaseProcessor):\n def __init__(\n self,\n image_size=384,\n mean=None,\n std=None,\n min_scale=0.5,\n max_scale=1.0,\n n_frms=MAX_INT,\n ):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n self.transform = transforms.Compose(\n [\n # Video size is (C, T, H, W)\n transforms_video.RandomResizedCropVideo(\n image_size,\n scale=(min_scale, max_scale),\n interpolation_mode=\"bicubic\",\n ),\n transforms_video.RandomHorizontalFlipVideo(),\n ToTHWC(), # C, T, H, W -> T, H, W, C\n VideoRandomAugment(\n 2,\n 5,\n augs=[\n \"Identity\",\n \"AutoContrast\",\n \"Brightness\",\n \"Sharpness\",\n \"Equalize\",\n \"ShearX\",\n \"ShearY\",\n \"TranslateX\",\n \"TranslateY\",\n \"Rotate\",\n ],\n ),\n ToUint8(),\n transforms_video.ToTensorVideo(), # T, H, W, C -> C, T, H, W\n self.normalize,\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n sampling=\"headtail\",\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.5)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n n_frms=n_frms,\n )\n\n\n@registry.register_processor(\"alpro_video_eval\")\nclass AlproVideoEvalProcessor(AlproVideoBaseProcessor):\n def __init__(self, image_size=256, mean=None, std=None, n_frms=MAX_INT):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n # Input video size is (C, T, H, W)\n self.transform = transforms.Compose(\n [\n # frames will be resized during decord loading.\n ToUint8(), # C, T, H, W\n ToTHWC(), # T, H, W, C\n transforms_video.ToTensorVideo(), # C, T, H, W\n self.normalize, # C, T, H, W\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(image_size=image_size, mean=mean, std=std, n_frms=n_frms)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.AlproVideoBaseProcessor","uri":"program://CREMA/class/lavis.processors.alpro_processors.AlproVideoBaseProcessor#L21-L30","kind":"class","name":"AlproVideoBaseProcessor","path":"lavis/processors/alpro_processors.py","language":"python","start_line":21,"end_line":30,"context_start_line":1,"context_end_line":50,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport torch\nfrom lavis.common.registry import registry\nfrom lavis.datasets.data_utils import load_video\nfrom lavis.processors import transforms_video\nfrom lavis.processors.base_processor import BaseProcessor\nfrom lavis.processors.randaugment import VideoRandomAugment\nfrom lavis.processors import functional_video as F\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\n\nMAX_INT = registry.get(\"MAX_INT\")\n\n\nclass AlproVideoBaseProcessor(BaseProcessor):\n def __init__(self, mean=None, std=None, n_frms=MAX_INT):\n if mean is None:\n mean = (0.48145466, 0.4578275, 0.40821073)\n if std is None:\n std = (0.26862954, 0.26130258, 0.27577711)\n\n self.normalize = transforms_video.NormalizeVideo(mean, std)\n\n self.n_frms = n_frms\n\n\nclass ToUint8(object):\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.to(torch.uint8)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ToTHWC(object):\n \"\"\"\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (C, T, H, W)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (T, H, W, C)\n \"\"\"","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.ToUint8","uri":"program://CREMA/class/lavis.processors.alpro_processors.ToUint8#L33-L41","kind":"class","name":"ToUint8","path":"lavis/processors/alpro_processors.py","language":"python","start_line":33,"end_line":41,"context_start_line":13,"context_end_line":61,"code":"from lavis.processors.randaugment import VideoRandomAugment\nfrom lavis.processors import functional_video as F\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\n\nMAX_INT = registry.get(\"MAX_INT\")\n\n\nclass AlproVideoBaseProcessor(BaseProcessor):\n def __init__(self, mean=None, std=None, n_frms=MAX_INT):\n if mean is None:\n mean = (0.48145466, 0.4578275, 0.40821073)\n if std is None:\n std = (0.26862954, 0.26130258, 0.27577711)\n\n self.normalize = transforms_video.NormalizeVideo(mean, std)\n\n self.n_frms = n_frms\n\n\nclass ToUint8(object):\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.to(torch.uint8)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ToTHWC(object):\n \"\"\"\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (C, T, H, W)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (T, H, W, C)\n \"\"\"\n\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.permute(1, 2, 3, 0)\n\n def __repr__(self):\n return self.__class__.__name__\n\n","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.ToTHWC","uri":"program://CREMA/class/lavis.processors.alpro_processors.ToTHWC#L44-L59","kind":"class","name":"ToTHWC","path":"lavis/processors/alpro_processors.py","language":"python","start_line":44,"end_line":59,"context_start_line":24,"context_end_line":79,"code":" mean = (0.48145466, 0.4578275, 0.40821073)\n if std is None:\n std = (0.26862954, 0.26130258, 0.27577711)\n\n self.normalize = transforms_video.NormalizeVideo(mean, std)\n\n self.n_frms = n_frms\n\n\nclass ToUint8(object):\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.to(torch.uint8)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ToTHWC(object):\n \"\"\"\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (C, T, H, W)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (T, H, W, C)\n \"\"\"\n\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.permute(1, 2, 3, 0)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ResizeVideo(object):\n def __init__(self, target_size, interpolation_mode=\"bilinear\"):\n self.target_size = target_size\n self.interpolation_mode = interpolation_mode\n\n def __call__(self, clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: central cropping of video clip. Size is\n (C, T, crop_size, crop_size)\n \"\"\"\n return F.resize(clip, self.target_size, self.interpolation_mode)\n\n def __repr__(self):\n return self.__class__.__name__ + \"(resize_size={0})\".format(self.target_size)\n","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.ResizeVideo","uri":"program://CREMA/class/lavis.processors.alpro_processors.ResizeVideo#L62-L78","kind":"class","name":"ResizeVideo","path":"lavis/processors/alpro_processors.py","language":"python","start_line":62,"end_line":78,"context_start_line":42,"context_end_line":98,"code":"\n\nclass ToTHWC(object):\n \"\"\"\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (C, T, H, W)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (T, H, W, C)\n \"\"\"\n\n def __init__(self):\n pass\n\n def __call__(self, tensor):\n return tensor.permute(1, 2, 3, 0)\n\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ResizeVideo(object):\n def __init__(self, target_size, interpolation_mode=\"bilinear\"):\n self.target_size = target_size\n self.interpolation_mode = interpolation_mode\n\n def __call__(self, clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: central cropping of video clip. Size is\n (C, T, crop_size, crop_size)\n \"\"\"\n return F.resize(clip, self.target_size, self.interpolation_mode)\n\n def __repr__(self):\n return self.__class__.__name__ + \"(resize_size={0})\".format(self.target_size)\n\n\n@registry.register_processor(\"alpro_video_train\")\nclass AlproVideoTrainProcessor(AlproVideoBaseProcessor):\n def __init__(\n self,\n image_size=384,\n mean=None,\n std=None,\n min_scale=0.5,\n max_scale=1.0,\n n_frms=MAX_INT,\n ):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n self.transform = transforms.Compose(\n [\n # Video size is (C, T, H, W)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.AlproVideoTrainProcessor","uri":"program://CREMA/class/lavis.processors.alpro_processors.AlproVideoTrainProcessor#L82-L167","kind":"class","name":"AlproVideoTrainProcessor","path":"lavis/processors/alpro_processors.py","language":"python","start_line":82,"end_line":167,"context_start_line":62,"context_end_line":187,"code":"class ResizeVideo(object):\n def __init__(self, target_size, interpolation_mode=\"bilinear\"):\n self.target_size = target_size\n self.interpolation_mode = interpolation_mode\n\n def __call__(self, clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: central cropping of video clip. Size is\n (C, T, crop_size, crop_size)\n \"\"\"\n return F.resize(clip, self.target_size, self.interpolation_mode)\n\n def __repr__(self):\n return self.__class__.__name__ + \"(resize_size={0})\".format(self.target_size)\n\n\n@registry.register_processor(\"alpro_video_train\")\nclass AlproVideoTrainProcessor(AlproVideoBaseProcessor):\n def __init__(\n self,\n image_size=384,\n mean=None,\n std=None,\n min_scale=0.5,\n max_scale=1.0,\n n_frms=MAX_INT,\n ):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n self.transform = transforms.Compose(\n [\n # Video size is (C, T, H, W)\n transforms_video.RandomResizedCropVideo(\n image_size,\n scale=(min_scale, max_scale),\n interpolation_mode=\"bicubic\",\n ),\n transforms_video.RandomHorizontalFlipVideo(),\n ToTHWC(), # C, T, H, W -> T, H, W, C\n VideoRandomAugment(\n 2,\n 5,\n augs=[\n \"Identity\",\n \"AutoContrast\",\n \"Brightness\",\n \"Sharpness\",\n \"Equalize\",\n \"ShearX\",\n \"ShearY\",\n \"TranslateX\",\n \"TranslateY\",\n \"Rotate\",\n ],\n ),\n ToUint8(),\n transforms_video.ToTensorVideo(), # T, H, W, C -> C, T, H, W\n self.normalize,\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n sampling=\"headtail\",\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.5)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n n_frms=n_frms,\n )\n\n\n@registry.register_processor(\"alpro_video_eval\")\nclass AlproVideoEvalProcessor(AlproVideoBaseProcessor):\n def __init__(self, image_size=256, mean=None, std=None, n_frms=MAX_INT):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n # Input video size is (C, T, H, W)\n self.transform = transforms.Compose(\n [\n # frames will be resized during decord loading.\n ToUint8(), # C, T, H, W\n ToTHWC(), # T, H, W, C\n transforms_video.ToTensorVideo(), # C, T, H, W\n self.normalize, # C, T, H, W\n ]\n )\n","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.AlproVideoEvalProcessor","uri":"program://CREMA/class/lavis.processors.alpro_processors.AlproVideoEvalProcessor#L171-L216","kind":"class","name":"AlproVideoEvalProcessor","path":"lavis/processors/alpro_processors.py","language":"python","start_line":171,"end_line":216,"context_start_line":151,"context_end_line":216,"code":"\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.5)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n n_frms=n_frms,\n )\n\n\n@registry.register_processor(\"alpro_video_eval\")\nclass AlproVideoEvalProcessor(AlproVideoBaseProcessor):\n def __init__(self, image_size=256, mean=None, std=None, n_frms=MAX_INT):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n # Input video size is (C, T, H, W)\n self.transform = transforms.Compose(\n [\n # frames will be resized during decord loading.\n ToUint8(), # C, T, H, W\n ToTHWC(), # T, H, W, C\n transforms_video.ToTensorVideo(), # C, T, H, W\n self.normalize, # C, T, H, W\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(image_size=image_size, mean=mean, std=std, n_frms=n_frms)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.__init__","uri":"program://CREMA/function/lavis.processors.alpro_processors.__init__#L172-L186","kind":"function","name":"__init__","path":"lavis/processors/alpro_processors.py","language":"python","start_line":172,"end_line":186,"context_start_line":152,"context_end_line":206,"code":" mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.5)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n n_frms=n_frms,\n )\n\n\n@registry.register_processor(\"alpro_video_eval\")\nclass AlproVideoEvalProcessor(AlproVideoBaseProcessor):\n def __init__(self, image_size=256, mean=None, std=None, n_frms=MAX_INT):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n # Input video size is (C, T, H, W)\n self.transform = transforms.Compose(\n [\n # frames will be resized during decord loading.\n ToUint8(), # C, T, H, W\n ToTHWC(), # T, H, W, C\n transforms_video.ToTensorVideo(), # C, T, H, W\n self.normalize, # C, T, H, W\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.__call__","uri":"program://CREMA/function/lavis.processors.alpro_processors.__call__#L188-L202","kind":"function","name":"__call__","path":"lavis/processors/alpro_processors.py","language":"python","start_line":188,"end_line":202,"context_start_line":168,"context_end_line":216,"code":"\n\n@registry.register_processor(\"alpro_video_eval\")\nclass AlproVideoEvalProcessor(AlproVideoBaseProcessor):\n def __init__(self, image_size=256, mean=None, std=None, n_frms=MAX_INT):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n # Input video size is (C, T, H, W)\n self.transform = transforms.Compose(\n [\n # frames will be resized during decord loading.\n ToUint8(), # C, T, H, W\n ToTHWC(), # T, H, W, C\n transforms_video.ToTensorVideo(), # C, T, H, W\n self.normalize, # C, T, H, W\n ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(image_size=image_size, mean=mean, std=std, n_frms=n_frms)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.__repr__","uri":"program://CREMA/function/lavis.processors.alpro_processors.__repr__#L77-L78","kind":"function","name":"__repr__","path":"lavis/processors/alpro_processors.py","language":"python","start_line":77,"end_line":78,"context_start_line":57,"context_end_line":98,"code":"\n def __repr__(self):\n return self.__class__.__name__\n\n\nclass ResizeVideo(object):\n def __init__(self, target_size, interpolation_mode=\"bilinear\"):\n self.target_size = target_size\n self.interpolation_mode = interpolation_mode\n\n def __call__(self, clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: central cropping of video clip. Size is\n (C, T, crop_size, crop_size)\n \"\"\"\n return F.resize(clip, self.target_size, self.interpolation_mode)\n\n def __repr__(self):\n return self.__class__.__name__ + \"(resize_size={0})\".format(self.target_size)\n\n\n@registry.register_processor(\"alpro_video_train\")\nclass AlproVideoTrainProcessor(AlproVideoBaseProcessor):\n def __init__(\n self,\n image_size=384,\n mean=None,\n std=None,\n min_scale=0.5,\n max_scale=1.0,\n n_frms=MAX_INT,\n ):\n super().__init__(mean=mean, std=std, n_frms=n_frms)\n\n self.image_size = image_size\n\n self.transform = transforms.Compose(\n [\n # Video size is (C, T, H, W)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.alpro_processors.from_config","uri":"program://CREMA/function/lavis.processors.alpro_processors.from_config#L205-L216","kind":"function","name":"from_config","path":"lavis/processors/alpro_processors.py","language":"python","start_line":205,"end_line":216,"context_start_line":185,"context_end_line":216,"code":" ]\n )\n\n def __call__(self, vpath):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n Returns:\n torch.tensor: video clip after transforms. Size is (C, T, size, size).\n \"\"\"\n clip = load_video(\n video_path=vpath,\n n_frms=self.n_frms,\n height=self.image_size,\n width=self.image_size,\n )\n\n return self.transform(clip)\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 256)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n n_frms = cfg.get(\"n_frms\", MAX_INT)\n\n return cls(image_size=image_size, mean=mean, std=std, n_frms=n_frms)","source_hash":"01f3c0f687b33e09f31bd0ea364684bec6ce966a930ee019f5813364497ccf5b","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors","uri":"program://CREMA/module/lavis.processors.clip_processors#L1-L92","kind":"module","name":"lavis.processors.clip_processors","path":"lavis/processors/clip_processors.py","language":"python","start_line":1,"end_line":92,"context_start_line":1,"context_end_line":92,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.processors.blip_processors import BlipImageBaseProcessor\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\nfrom torchvision.transforms.functional import InterpolationMode\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\n@registry.register_processor(\"clip_image_train\")\nclass ClipImageTrainProcessor(BlipImageBaseProcessor):\n def __init__(\n self, image_size=224, mean=None, std=None, min_scale=0.9, max_scale=1.0\n ):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.RandomResizedCrop(\n image_size,\n scale=(min_scale, max_scale),\n interpolation=InterpolationMode.BICUBIC,\n ),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.9)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n )\n\n\n@registry.register_processor(\"clip_image_eval\")\nclass ClipImageEvalProcessor(BlipImageBaseProcessor):\n def __init__(self, image_size=224, mean=None, std=None):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n transforms.CenterCrop(image_size),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n )","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors._convert_to_rgb","uri":"program://CREMA/function/lavis.processors.clip_processors._convert_to_rgb#L15-L16","kind":"function","name":"_convert_to_rgb","path":"lavis/processors/clip_processors.py","language":"python","start_line":15,"end_line":16,"context_start_line":1,"context_end_line":36,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.processors.blip_processors import BlipImageBaseProcessor\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\nfrom torchvision.transforms.functional import InterpolationMode\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\n@registry.register_processor(\"clip_image_train\")\nclass ClipImageTrainProcessor(BlipImageBaseProcessor):\n def __init__(\n self, image_size=224, mean=None, std=None, min_scale=0.9, max_scale=1.0\n ):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.RandomResizedCrop(\n image_size,\n scale=(min_scale, max_scale),\n interpolation=InterpolationMode.BICUBIC,\n ),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors.ClipImageTrainProcessor","uri":"program://CREMA/class/lavis.processors.clip_processors.ClipImageTrainProcessor#L20-L59","kind":"class","name":"ClipImageTrainProcessor","path":"lavis/processors/clip_processors.py","language":"python","start_line":20,"end_line":59,"context_start_line":1,"context_end_line":79,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nfrom lavis.common.registry import registry\nfrom lavis.processors.blip_processors import BlipImageBaseProcessor\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\nfrom torchvision.transforms.functional import InterpolationMode\n\n\ndef _convert_to_rgb(image):\n return image.convert(\"RGB\")\n\n\n@registry.register_processor(\"clip_image_train\")\nclass ClipImageTrainProcessor(BlipImageBaseProcessor):\n def __init__(\n self, image_size=224, mean=None, std=None, min_scale=0.9, max_scale=1.0\n ):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.RandomResizedCrop(\n image_size,\n scale=(min_scale, max_scale),\n interpolation=InterpolationMode.BICUBIC,\n ),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.9)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n )\n\n\n@registry.register_processor(\"clip_image_eval\")\nclass ClipImageEvalProcessor(BlipImageBaseProcessor):\n def __init__(self, image_size=224, mean=None, std=None):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n transforms.CenterCrop(image_size),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors.ClipImageEvalProcessor","uri":"program://CREMA/class/lavis.processors.clip_processors.ClipImageEvalProcessor#L63-L92","kind":"class","name":"ClipImageEvalProcessor","path":"lavis/processors/clip_processors.py","language":"python","start_line":63,"end_line":92,"context_start_line":43,"context_end_line":92,"code":" cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.9)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n )\n\n\n@registry.register_processor(\"clip_image_eval\")\nclass ClipImageEvalProcessor(BlipImageBaseProcessor):\n def __init__(self, image_size=224, mean=None, std=None):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n transforms.CenterCrop(image_size),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n )","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors.__init__","uri":"program://CREMA/function/lavis.processors.clip_processors.__init__#L64-L76","kind":"function","name":"__init__","path":"lavis/processors/clip_processors.py","language":"python","start_line":64,"end_line":76,"context_start_line":44,"context_end_line":92,"code":"\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n min_scale = cfg.get(\"min_scale\", 0.9)\n max_scale = cfg.get(\"max_scale\", 1.0)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n min_scale=min_scale,\n max_scale=max_scale,\n )\n\n\n@registry.register_processor(\"clip_image_eval\")\nclass ClipImageEvalProcessor(BlipImageBaseProcessor):\n def __init__(self, image_size=224, mean=None, std=None):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n transforms.CenterCrop(image_size),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n )","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.clip_processors.from_config","uri":"program://CREMA/function/lavis.processors.clip_processors.from_config#L79-L92","kind":"function","name":"from_config","path":"lavis/processors/clip_processors.py","language":"python","start_line":79,"end_line":92,"context_start_line":59,"context_end_line":92,"code":" )\n\n\n@registry.register_processor(\"clip_image_eval\")\nclass ClipImageEvalProcessor(BlipImageBaseProcessor):\n def __init__(self, image_size=224, mean=None, std=None):\n\n super().__init__(mean=mean, std=std)\n\n self.transform = transforms.Compose(\n [\n transforms.Resize(image_size, interpolation=InterpolationMode.BICUBIC),\n transforms.CenterCrop(image_size),\n _convert_to_rgb,\n transforms.ToTensor(),\n self.normalize,\n ]\n )\n\n @classmethod\n def from_config(cls, cfg=None):\n if cfg is None:\n cfg = OmegaConf.create()\n\n image_size = cfg.get(\"image_size\", 224)\n\n mean = cfg.get(\"mean\", None)\n std = cfg.get(\"std\", None)\n\n return cls(\n image_size=image_size,\n mean=mean,\n std=std,\n )","source_hash":"b088bae35e3af74b5de885162b4399324b4d1992c199ba8b13902e42b84d46ac","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video","uri":"program://CREMA/module/lavis.processors.functional_video#L1-L121","kind":"module","name":"lavis.processors.functional_video","path":"lavis/processors/functional_video.py","language":"python","start_line":1,"end_line":121,"context_start_line":1,"context_end_line":121,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\n\n\ndef _is_tensor_video_clip(clip):\n if not torch.is_tensor(clip):\n raise TypeError(\"clip should be Tensor. Got %s\" % type(clip))\n\n if not clip.ndimension() == 4:\n raise ValueError(\"clip should be 4D. Got %dD\" % clip.dim())\n\n return True\n\n\ndef crop(clip, i, j, h, w):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n \"\"\"\n if len(clip.size()) != 4:\n raise ValueError(\"clip should be a 4D tensor\")\n return clip[..., i : i + h, j : j + w]\n\n\ndef resize(clip, target_size, interpolation_mode):\n if len(target_size) != 2:\n raise ValueError(\n f\"target size should be tuple (height, width), instead got {target_size}\"\n )\n return torch.nn.functional.interpolate(\n clip, size=target_size, mode=interpolation_mode, align_corners=False\n )\n\n\ndef resized_crop(clip, i, j, h, w, size, interpolation_mode=\"bilinear\"):\n \"\"\"\n Do spatial cropping and resizing to the video clip\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n i (int): i in (i,j) i.e coordinates of the upper left corner.\n j (int): j in (i,j) i.e coordinates of the upper left corner.\n h (int): Height of the cropped region.\n w (int): Width of the cropped region.\n size (tuple(int, int)): height and width of resized clip\n Returns:\n clip (torch.tensor): Resized and cropped clip. Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n clip = crop(clip, i, j, h, w)\n clip = resize(clip, size, interpolation_mode)\n return clip\n\n\ndef center_crop(clip, crop_size):\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n h, w = clip.size(-2), clip.size(-1)\n th, tw = crop_size\n if h < th or w < tw:\n raise ValueError(\"height and width must be no smaller than crop_size\")\n\n i = int(round((h - th) / 2.0))\n j = int(round((w - tw) / 2.0))\n return crop(clip, i, j, th, tw)\n\n\ndef to_tensor(clip):\n \"\"\"\n Convert tensor data type from uint8 to float, divide value by 255.0 and\n permute the dimensions of clip tensor\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (T, H, W, C)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (C, T, H, W)\n \"\"\"\n _is_tensor_video_clip(clip)\n if not clip.dtype == torch.uint8:\n raise TypeError(\n \"clip tensor should have data type uint8. Got %s\" % str(clip.dtype)\n )\n return clip.float().permute(3, 0, 1, 2) / 255.0\n\n\ndef normalize(clip, mean, std, inplace=False):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n mean (tuple): pixel RGB mean. Size is (3)\n std (tuple): pixel standard deviation. Size is (3)\n Returns:\n normalized clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n if not inplace:\n clip = clip.clone()\n mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)\n std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)\n clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])\n return clip\n\n\ndef hflip(clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n Returns:\n flipped clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n return clip.flip(-1)","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video._is_tensor_video_clip","uri":"program://CREMA/function/lavis.processors.functional_video._is_tensor_video_clip#L13-L20","kind":"function","name":"_is_tensor_video_clip","path":"lavis/processors/functional_video.py","language":"python","start_line":13,"end_line":20,"context_start_line":1,"context_end_line":40,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\n\n\ndef _is_tensor_video_clip(clip):\n if not torch.is_tensor(clip):\n raise TypeError(\"clip should be Tensor. Got %s\" % type(clip))\n\n if not clip.ndimension() == 4:\n raise ValueError(\"clip should be 4D. Got %dD\" % clip.dim())\n\n return True\n\n\ndef crop(clip, i, j, h, w):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n \"\"\"\n if len(clip.size()) != 4:\n raise ValueError(\"clip should be a 4D tensor\")\n return clip[..., i : i + h, j : j + w]\n\n\ndef resize(clip, target_size, interpolation_mode):\n if len(target_size) != 2:\n raise ValueError(\n f\"target size should be tuple (height, width), instead got {target_size}\"\n )\n return torch.nn.functional.interpolate(\n clip, size=target_size, mode=interpolation_mode, align_corners=False\n )","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.crop","uri":"program://CREMA/function/lavis.processors.functional_video.crop#L23-L30","kind":"function","name":"crop","path":"lavis/processors/functional_video.py","language":"python","start_line":23,"end_line":30,"context_start_line":3,"context_end_line":50,"code":" All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport warnings\n\nimport torch\n\n\ndef _is_tensor_video_clip(clip):\n if not torch.is_tensor(clip):\n raise TypeError(\"clip should be Tensor. Got %s\" % type(clip))\n\n if not clip.ndimension() == 4:\n raise ValueError(\"clip should be 4D. Got %dD\" % clip.dim())\n\n return True\n\n\ndef crop(clip, i, j, h, w):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n \"\"\"\n if len(clip.size()) != 4:\n raise ValueError(\"clip should be a 4D tensor\")\n return clip[..., i : i + h, j : j + w]\n\n\ndef resize(clip, target_size, interpolation_mode):\n if len(target_size) != 2:\n raise ValueError(\n f\"target size should be tuple (height, width), instead got {target_size}\"\n )\n return torch.nn.functional.interpolate(\n clip, size=target_size, mode=interpolation_mode, align_corners=False\n )\n\n\ndef resized_crop(clip, i, j, h, w, size, interpolation_mode=\"bilinear\"):\n \"\"\"\n Do spatial cropping and resizing to the video clip\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n i (int): i in (i,j) i.e coordinates of the upper left corner.\n j (int): j in (i,j) i.e coordinates of the upper left corner.\n h (int): Height of the cropped region.","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.resize","uri":"program://CREMA/function/lavis.processors.functional_video.resize#L33-L40","kind":"function","name":"resize","path":"lavis/processors/functional_video.py","language":"python","start_line":33,"end_line":40,"context_start_line":13,"context_end_line":60,"code":"def _is_tensor_video_clip(clip):\n if not torch.is_tensor(clip):\n raise TypeError(\"clip should be Tensor. Got %s\" % type(clip))\n\n if not clip.ndimension() == 4:\n raise ValueError(\"clip should be 4D. Got %dD\" % clip.dim())\n\n return True\n\n\ndef crop(clip, i, j, h, w):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n \"\"\"\n if len(clip.size()) != 4:\n raise ValueError(\"clip should be a 4D tensor\")\n return clip[..., i : i + h, j : j + w]\n\n\ndef resize(clip, target_size, interpolation_mode):\n if len(target_size) != 2:\n raise ValueError(\n f\"target size should be tuple (height, width), instead got {target_size}\"\n )\n return torch.nn.functional.interpolate(\n clip, size=target_size, mode=interpolation_mode, align_corners=False\n )\n\n\ndef resized_crop(clip, i, j, h, w, size, interpolation_mode=\"bilinear\"):\n \"\"\"\n Do spatial cropping and resizing to the video clip\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n i (int): i in (i,j) i.e coordinates of the upper left corner.\n j (int): j in (i,j) i.e coordinates of the upper left corner.\n h (int): Height of the cropped region.\n w (int): Width of the cropped region.\n size (tuple(int, int)): height and width of resized clip\n Returns:\n clip (torch.tensor): Resized and cropped clip. Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n clip = crop(clip, i, j, h, w)\n clip = resize(clip, size, interpolation_mode)\n return clip","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.resized_crop","uri":"program://CREMA/function/lavis.processors.functional_video.resized_crop#L43-L60","kind":"function","name":"resized_crop","path":"lavis/processors/functional_video.py","language":"python","start_line":43,"end_line":60,"context_start_line":23,"context_end_line":80,"code":"def crop(clip, i, j, h, w):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n \"\"\"\n if len(clip.size()) != 4:\n raise ValueError(\"clip should be a 4D tensor\")\n return clip[..., i : i + h, j : j + w]\n\n\ndef resize(clip, target_size, interpolation_mode):\n if len(target_size) != 2:\n raise ValueError(\n f\"target size should be tuple (height, width), instead got {target_size}\"\n )\n return torch.nn.functional.interpolate(\n clip, size=target_size, mode=interpolation_mode, align_corners=False\n )\n\n\ndef resized_crop(clip, i, j, h, w, size, interpolation_mode=\"bilinear\"):\n \"\"\"\n Do spatial cropping and resizing to the video clip\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n i (int): i in (i,j) i.e coordinates of the upper left corner.\n j (int): j in (i,j) i.e coordinates of the upper left corner.\n h (int): Height of the cropped region.\n w (int): Width of the cropped region.\n size (tuple(int, int)): height and width of resized clip\n Returns:\n clip (torch.tensor): Resized and cropped clip. Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n clip = crop(clip, i, j, h, w)\n clip = resize(clip, size, interpolation_mode)\n return clip\n\n\ndef center_crop(clip, crop_size):\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n h, w = clip.size(-2), clip.size(-1)\n th, tw = crop_size\n if h < th or w < tw:\n raise ValueError(\"height and width must be no smaller than crop_size\")\n\n i = int(round((h - th) / 2.0))\n j = int(round((w - tw) / 2.0))\n return crop(clip, i, j, th, tw)\n\n\ndef to_tensor(clip):\n \"\"\"\n Convert tensor data type from uint8 to float, divide value by 255.0 and\n permute the dimensions of clip tensor\n Args:","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.center_crop","uri":"program://CREMA/function/lavis.processors.functional_video.center_crop#L63-L73","kind":"function","name":"center_crop","path":"lavis/processors/functional_video.py","language":"python","start_line":63,"end_line":73,"context_start_line":43,"context_end_line":93,"code":"def resized_crop(clip, i, j, h, w, size, interpolation_mode=\"bilinear\"):\n \"\"\"\n Do spatial cropping and resizing to the video clip\n Args:\n clip (torch.tensor): Video clip to be cropped. Size is (C, T, H, W)\n i (int): i in (i,j) i.e coordinates of the upper left corner.\n j (int): j in (i,j) i.e coordinates of the upper left corner.\n h (int): Height of the cropped region.\n w (int): Width of the cropped region.\n size (tuple(int, int)): height and width of resized clip\n Returns:\n clip (torch.tensor): Resized and cropped clip. Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n clip = crop(clip, i, j, h, w)\n clip = resize(clip, size, interpolation_mode)\n return clip\n\n\ndef center_crop(clip, crop_size):\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n h, w = clip.size(-2), clip.size(-1)\n th, tw = crop_size\n if h < th or w < tw:\n raise ValueError(\"height and width must be no smaller than crop_size\")\n\n i = int(round((h - th) / 2.0))\n j = int(round((w - tw) / 2.0))\n return crop(clip, i, j, th, tw)\n\n\ndef to_tensor(clip):\n \"\"\"\n Convert tensor data type from uint8 to float, divide value by 255.0 and\n permute the dimensions of clip tensor\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (T, H, W, C)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (C, T, H, W)\n \"\"\"\n _is_tensor_video_clip(clip)\n if not clip.dtype == torch.uint8:\n raise TypeError(\n \"clip tensor should have data type uint8. Got %s\" % str(clip.dtype)\n )\n return clip.float().permute(3, 0, 1, 2) / 255.0\n\n\ndef normalize(clip, mean, std, inplace=False):","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.to_tensor","uri":"program://CREMA/function/lavis.processors.functional_video.to_tensor#L76-L90","kind":"function","name":"to_tensor","path":"lavis/processors/functional_video.py","language":"python","start_line":76,"end_line":90,"context_start_line":56,"context_end_line":110,"code":" if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n clip = crop(clip, i, j, h, w)\n clip = resize(clip, size, interpolation_mode)\n return clip\n\n\ndef center_crop(clip, crop_size):\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n h, w = clip.size(-2), clip.size(-1)\n th, tw = crop_size\n if h < th or w < tw:\n raise ValueError(\"height and width must be no smaller than crop_size\")\n\n i = int(round((h - th) / 2.0))\n j = int(round((w - tw) / 2.0))\n return crop(clip, i, j, th, tw)\n\n\ndef to_tensor(clip):\n \"\"\"\n Convert tensor data type from uint8 to float, divide value by 255.0 and\n permute the dimensions of clip tensor\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (T, H, W, C)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (C, T, H, W)\n \"\"\"\n _is_tensor_video_clip(clip)\n if not clip.dtype == torch.uint8:\n raise TypeError(\n \"clip tensor should have data type uint8. Got %s\" % str(clip.dtype)\n )\n return clip.float().permute(3, 0, 1, 2) / 255.0\n\n\ndef normalize(clip, mean, std, inplace=False):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n mean (tuple): pixel RGB mean. Size is (3)\n std (tuple): pixel standard deviation. Size is (3)\n Returns:\n normalized clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n if not inplace:\n clip = clip.clone()\n mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)\n std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)\n clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])\n return clip\n","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.normalize","uri":"program://CREMA/function/lavis.processors.functional_video.normalize#L93-L109","kind":"function","name":"normalize","path":"lavis/processors/functional_video.py","language":"python","start_line":93,"end_line":109,"context_start_line":73,"context_end_line":121,"code":" return crop(clip, i, j, th, tw)\n\n\ndef to_tensor(clip):\n \"\"\"\n Convert tensor data type from uint8 to float, divide value by 255.0 and\n permute the dimensions of clip tensor\n Args:\n clip (torch.tensor, dtype=torch.uint8): Size is (T, H, W, C)\n Return:\n clip (torch.tensor, dtype=torch.float): Size is (C, T, H, W)\n \"\"\"\n _is_tensor_video_clip(clip)\n if not clip.dtype == torch.uint8:\n raise TypeError(\n \"clip tensor should have data type uint8. Got %s\" % str(clip.dtype)\n )\n return clip.float().permute(3, 0, 1, 2) / 255.0\n\n\ndef normalize(clip, mean, std, inplace=False):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n mean (tuple): pixel RGB mean. Size is (3)\n std (tuple): pixel standard deviation. Size is (3)\n Returns:\n normalized clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n if not inplace:\n clip = clip.clone()\n mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)\n std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)\n clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])\n return clip\n\n\ndef hflip(clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n Returns:\n flipped clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n return clip.flip(-1)","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.functional_video.hflip","uri":"program://CREMA/function/lavis.processors.functional_video.hflip#L112-L121","kind":"function","name":"hflip","path":"lavis/processors/functional_video.py","language":"python","start_line":112,"end_line":121,"context_start_line":92,"context_end_line":121,"code":"\ndef normalize(clip, mean, std, inplace=False):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n mean (tuple): pixel RGB mean. Size is (3)\n std (tuple): pixel standard deviation. Size is (3)\n Returns:\n normalized clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n if not inplace:\n clip = clip.clone()\n mean = torch.as_tensor(mean, dtype=clip.dtype, device=clip.device)\n std = torch.as_tensor(std, dtype=clip.dtype, device=clip.device)\n clip.sub_(mean[:, None, None, None]).div_(std[:, None, None, None])\n return clip\n\n\ndef hflip(clip):\n \"\"\"\n Args:\n clip (torch.tensor): Video clip to be normalized. Size is (C, T, H, W)\n Returns:\n flipped clip (torch.tensor): Size is (C, T, H, W)\n \"\"\"\n if not _is_tensor_video_clip(clip):\n raise ValueError(\"clip should be a 4D torch.tensor\")\n return clip.flip(-1)","source_hash":"772daa8d396bd20e7163810f4392dde4db1b811795c11193d9264af7bc604ec2","truncated":false} {"repo_id":"CREMA","entity_id":"py:lavis.processors.gpt_processors","uri":"program://CREMA/module/lavis.processors.gpt_processors#L1-L171","kind":"module","name":"lavis.processors.gpt_processors","path":"lavis/processors/gpt_processors.py","language":"python","start_line":1,"end_line":171,"context_start_line":1,"context_end_line":171,"code":"\"\"\"\n Copyright (c) 2022, salesforce.com, inc.\n All rights reserved.\n SPDX-License-Identifier: BSD-3-Clause\n For full license text, see the LICENSE file in the repo root or https://opensource.org/licenses/BSD-3-Clause\n\"\"\"\n\nimport re\n\nfrom lavis.common.registry import registry\nfrom lavis.processors.base_processor import BaseProcessor\nfrom lavis.processors.randaugment import RandomAugment\nfrom omegaconf import OmegaConf\nfrom torchvision import transforms\nfrom torchvision.transforms.functional import InterpolationMode\nimport os\nfrom itertools import chain\nimport numpy as np\nimport torch\nfrom transformers import GPT2Tokenizer\n\nSPECIAL_TOKENS_DICT = {\n \"bos_token\": \"\",\n \"eos_token\": \"\",\n \"additional_special_tokens\": [\"\", \"\", \"