File size: 28,611 Bytes
5e88ff2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 | import asyncio
import hashlib
import logging
import math
import os
import re
import shutil
import subprocess
import sys
import time
import uuid
from pathlib import Path
import numpy as np
import soundfile as sf
import torch
from aiohttp import web
import comfy.model_management as mm
import comfy.model_patcher
import comfy.utils
import folder_paths
from server import PromptServer
from .robin import Robin, pieces
from .robin import train as trainer
from .robin.vae import load_vae
from .robin.pipeline import SPEECH, VAE, WEIGHTS
folder_paths.add_model_folder_path('robin', os.path.join(folder_paths.models_dir, 'robin'))
REPO, MODEL, LORA = 'FixedSeed/Robin', 'robin', 'robin_character_lora.safetensors'
MODEL_FILES = (WEIGHTS, 'config.json', 'tokenizer.json', SPEECH)
DTYPES = dict(bf16=torch.bfloat16, fp32=torch.float32)
LANGUAGES = {'English': 'en', 'Japanese': 'ja', 'English + Japanese': 'mix'}
UPLOADS = 'robin_datasets'
SOUND_FILES = ('.wav', '.flac', '.mp3', '.ogg', '.opus', '.aiff', '.aif')
MAX_SOURCE_SECONDS = 45
WORKING_MEMORY = 2 * 1024 ** 3
TRAINING_MEMORY = 4 * 1024 ** 3
NAME = re.compile(r'[A-Za-z0-9][A-Za-z0-9_.-]{0,63}')
SAVED = re.compile(r'step_\d+')
def model_folders():
found = set()
for base in folder_paths.get_folder_paths('robin'):
base = Path(base)
if base.is_dir():
for config in base.rglob('config.json'):
if config.parent != base and (config.parent/WEIGHTS).is_file():
found.add(config.parent.relative_to(base).as_posix())
return [MODEL] + sorted(found - {MODEL})
def model_path(name):
for base in folder_paths.get_folder_paths('robin'):
path = Path(base)/name
if (path/'config.json').is_file():
return path
return Path(folder_paths.get_folder_paths('robin')[0])/name
def shown(path):
try:
return Path(path).resolve().relative_to(Path(folder_paths.base_path).resolve()).as_posix()
except ValueError:
return str(path)
def own_vae(vae):
return vae == VAE and not folder_paths.get_full_path('vae', VAE)
def model_status(model, vae=VAE):
path = model_path(model)
files = MODEL_FILES + ((VAE,) if own_vae(vae) else ())
present = [path/f for f in files if (path/f).is_file()]
return dict(model=model, ready=len(present) == len(files), size=sum(f.stat().st_size for f in present),
voice=bool(folder_paths.get_full_path('loras', LORA)))
def download(file, folder):
from huggingface_hub import hf_hub_download
logging.info(f'Robin: downloading {file} from Hugging Face ({REPO}) into {folder}')
try:
hf_hub_download(REPO, file, local_dir=str(folder))
except Exception as e:
raise RuntimeError(f'Robin could not download {file} from Hugging Face ({REPO}): {e}') from e
def lora_path():
found = folder_paths.get_full_path('loras', LORA)
if found:
return found
folder = Path(folder_paths.get_folder_paths('loras')[0])
download(LORA, folder)
return str(folder/LORA)
def loras():
found = [f for f in folder_paths.get_filename_list('loras') if f.endswith('.safetensors') and '/cache/' not in f.replace(os.sep, '/')]
return [LORA] + [f for f in found if f != LORA]
def input_audio():
folder = Path(folder_paths.get_input_directory())
return sorted(f.name for f in folder.iterdir() if f.is_file() and f.suffix.lower() in SOUND_FILES) if folder.is_dir() else []
def dataset_path(folder):
path = Path(folder.strip()).expanduser()
return path if path.is_absolute() else Path(folder_paths.get_input_directory())/path
def folder_signature(folder):
path = dataset_path(folder)
if not folder.strip() or not path.is_dir():
return ''
files = sorted(f for f in path.rglob('*') if f.is_file())
return hashlib.sha256('\n'.join(f'{f}:{f.stat().st_size}:{f.stat().st_mtime_ns}' for f in files).encode()).hexdigest()
def check_folder(folder, language):
if not folder.strip():
return None, 'Choose a folder of clips, or type its path.'
path = dataset_path(folder)
if not path.is_dir():
return None, f'There is no folder at {path}.'
lang = LANGUAGES[language]
found = trainer.scan(path, lang)
found['folder'] = shown(path)
lines = []
try:
trainer.describe(found, log=lines.append)
except SystemExit as e:
lines.append(str(e))
return dict(folder=str(path), lang=lang), '\n'.join(lines)
def tell(event, node, **data):
server = PromptServer.instance
data = {k: None if isinstance(v, float) and not math.isfinite(v) else v for k, v in data.items()}
server.send_sync(event, dict(node=node, **data), server.client_id)
def preview(path, subfolder):
folder = Path(folder_paths.get_temp_directory())/subfolder
folder.mkdir(parents=True, exist_ok=True)
shutil.copyfile(path, folder/Path(path).name)
return dict(filename=Path(path).name, subfolder=subfolder, type='temp')
def take(audio, seed):
folder = Path(folder_paths.get_temp_directory())/'robin'
folder.mkdir(parents=True, exist_ok=True)
name = f'take_{uuid.uuid4().hex[:12]}.flac'
sf.write(str(folder/name), np.clip(audio, -1, 1), pieces.RATE, subtype='PCM_24')
return dict(filename=name, subfolder='robin', type='temp', seconds=round(len(audio) / pieces.RATE, 2), seed=seed)
def trainable(robin):
for net in (robin.planner, robin.decoder):
for module in net.modules():
for key, value in list(module._parameters.items()):
if value is not None and value.is_inference():
module._parameters[key] = torch.nn.Parameter(value.clone(), requires_grad=False)
for key, value in list(module._buffers.items()):
if value is not None and value.is_inference():
module._buffers[key] = value.clone()
for key, value in list(vars(module).items()):
if isinstance(value, torch.Tensor) and value.is_inference():
setattr(module, key, value.clone())
def devices():
return ['default', 'cpu'] + ([f'cuda:{i}' for i in range(torch.cuda.device_count())] if torch.cuda.device_count() > 1 else [])
class Core:
def __init__(self, robin, load_device, offload_device):
self.robin, self.load_device, self.offload_device = robin, load_device, offload_device
bundle = torch.nn.ModuleDict(dict(planner=robin.planner, decoder=robin.decoder, vae=robin.vae))
self.patcher = self.patch(bundle)
self.speech_patcher = self.lora_patcher = None
self.lora_file = None
def patch(self, module):
return comfy.model_patcher.ModelPatcher(module, load_device=self.load_device, offload_device=self.offload_device)
class RobinModel:
def __init__(self, core, lora=None):
self.core, self.lora = core, lora
def ready(self, speech=False, memory=WORKING_MEMORY):
core, robin = self.core, self.core.robin
if self.lora is None:
if robin.lora is not None:
robin.remove_lora(); core.lora_patcher = core.lora_file = None
else:
path, scale = self.lora
if core.lora_file != path:
robin.load_lora(path, scale)
core.lora_file, core.lora_patcher = path, core.patch(robin.lora)
robin.set_lora_scale(scale)
patchers = [core.patcher] + ([core.lora_patcher] if core.lora_patcher else [])
if speech:
tokenizer = robin.speech
if core.speech_patcher is None:
core.speech_patcher = core.patch(torch.nn.ModuleDict(dict(speech=tokenizer.model)))
patchers.append(core.speech_patcher)
mm.load_models_gpu(patchers, memory_required=memory, force_full_load=True)
return robin
class Progress:
def __init__(self, renders):
self.bar, self.total, self.done = comfy.utils.ProgressBar(100 * renders), renders, 0
self.expected, self.plan, self.decode = 1, 0., 0.
def start(self, seconds):
self.expected, self.plan, self.decode = max(1., seconds * 25), 0., 0.
self.show()
def __call__(self, stage, step, total):
mm.throw_exception_if_processing_interrupted()
if stage == 'plan':
self.plan = min(1., step / self.expected)
if step % 25: return
else:
self.plan, self.decode = 1., step / total
self.show()
def finish(self, extra=0):
self.done += 1; self.total += extra; self.show()
def show(self):
within = .8 * self.plan + .2 * self.decode
self.bar.update_absolute(int(100 * (self.done + (within if self.done < self.total else 0))), 100 * self.total)
def mono(audio):
wave = audio['waveform']
return wave[0].float().mean(0).cpu().numpy(), int(audio['sample_rate'])
def to_audio(x, rate=pieces.RATE):
return {'waveform': torch.from_numpy(np.ascontiguousarray(x, dtype=np.float32))[None, None], 'sample_rate': rate}
SAMPLING = {
'seed': ('INT', {'default': 0, 'min': 0, 'max': 0xffffffffffffffff, 'control_after_generate': True}),
'temperature': ('FLOAT', {'default': .9, 'min': 0., 'max': 2., 'step': .01, 'tooltip': 'Higher is livelier and less predictable; 0 always picks the likeliest sound.'}),
'top_k': ('INT', {'default': 25, 'min': 0, 'max': 500, 'tooltip': 'Choose among only this many likeliest sounds at each step (0 = no limit).'}),
'top_p': ('FLOAT', {'default': .8, 'min': .01, 'max': 1., 'step': .01, 'tooltip': 'Choose among only the likeliest sounds that add up to this share.'}),
}
RETRIES = ('INT', {'default': 2, 'min': 0, 'max': 6, 'advanced': True,
'tooltip': 'Tries a piece again with a new seed when it comes out silent or the wrong length.'})
class RobinLoader:
DESCRIPTION = ('Loads Robin from a folder in models/robin. The first run downloads whatever is missing from Hugging Face, '
f'along with Robin\'s own voice ({LORA}) for the Robin LoRA node.')
RETURN_TYPES = ('ROBIN_MODEL',)
RETURN_NAMES = ('robin',)
FUNCTION = 'load'
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'model': (model_folders(), {'default': MODEL, 'tooltip': f'A Robin folder in models/robin. "{MODEL}" is downloaded on the first run if it is not there yet.'}),
'vae': ([VAE] + [f for f in folder_paths.get_filename_list('vae') if f != VAE],
{'default': VAE, 'tooltip': f'The VAE weights Robin decodes with. {VAE} is downloaded into the Robin folder on the '
'first run, unless models/vae already has it.'}),
'device': (devices(), {'default': 'default'}),
'dtype': (list(DTYPES), {'default': 'bf16', 'tooltip': 'bf16 needs about 8 GB of GPU memory; fp32 about twice that.'}),
},
'hidden': {'unique_id': 'UNIQUE_ID'},
}
def load(self, model, vae, device, dtype, unique_id=None):
path = model_path(model)
missing = [f for f in MODEL_FILES + ((VAE,) if own_vae(vae) else ()) if not (path/f).is_file()]
for file in missing:
tell('robin.loader', unique_id, stage='download', file=file)
download(file, path)
if not folder_paths.get_full_path('loras', LORA):
tell('robin.loader', unique_id, stage='download', file=LORA)
lora_path()
tell('robin.loader', unique_id, stage='load')
load_device = mm.get_torch_device() if device == 'default' else torch.device(device)
offload_device = load_device if load_device.type == 'cpu' else mm.unet_offload_device()
vae_file = path/VAE if own_vae(vae) else folder_paths.get_full_path_or_raise('vae', vae)
try:
decoder = load_vae(vae_file, offload_device)
except (RuntimeError, KeyError) as e:
raise ValueError(f'{vae} is not a VAE Robin can decode with.') from e
robin = Robin.from_pretrained(str(path), vae=decoder, device=offload_device, dtype=DTYPES[dtype])
robin.device = load_device
status = dict(model_status(model, vae), loaded=dtype, downloaded=bool(missing))
return {'ui': {'robin_status': [status]}, 'result': (RobinModel(Core(robin, load_device, offload_device)),)}
class RobinLoRA:
DESCRIPTION = (f'Applies a voice LoRA trained with Robin. {LORA} is Robin\'s own voice. One LoRA at a time: a second LoRA node '
'replaces the first.')
RETURN_TYPES = ('ROBIN_MODEL',)
RETURN_NAMES = ('robin',)
FUNCTION = 'apply'
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'robin': ('ROBIN_MODEL',),
'lora': (loras(), {'default': LORA}),
'strength': ('FLOAT', {'default': 1., 'min': 0., 'max': 2., 'step': .05, 'tooltip': '1 is the voice as trained, 0 is Robin without the LoRA.'}),
},
}
def apply(self, robin, lora, strength):
path = lora_path() if lora == LORA else folder_paths.get_full_path_or_raise('loras', lora)
from safetensors import safe_open
with safe_open(path, framework='pt', device='cpu') as f:
if 'robin_lora' not in (f.metadata() or {}):
raise ValueError(f'{lora} is not a Robin LoRA.')
return (RobinModel(robin.core, (path, strength)),)
class RobinSpeak:
DESCRIPTION = ('Reads text aloud in a voice you describe. Long text is read in pieces that carry the voice from one to the next. '
'[laughs] and [coughs] can go anywhere in the text. The last take plays on the node; connect Save Audio to keep '
'every take in the output folder.')
RETURN_TYPES = ('AUDIO',)
RETURN_NAMES = ('audio',)
FUNCTION = 'speak'
OUTPUT_NODE = True
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'robin': ('ROBIN_MODEL',),
'text': ('STRING', {'multiline': True, 'default': 'Hello there! [laughs] It is so good to finally hear my own voice.',
'tooltip': 'What to say. [laughs] and [coughs] can go anywhere in it.'}),
'voice': ('STRING', {'multiline': True, 'default': 'A warm, low voice, unhurried and close to the mic.',
'tooltip': 'The voice in plain words: age, pitch, pace, mood, how close to the mic. Leave it empty to let Robin choose.'}),
**SAMPLING,
'cfg': ('FLOAT', {'default': 1., 'min': 1., 'max': 10., 'step': .1,
'tooltip': 'Above 1 follows the voice description more closely (try 3).'}),
},
'optional': {
'retries': RETRIES,
'max_words': ('INT', {'default': 35, 'min': 8, 'max': 60, 'advanced': True, 'tooltip': 'Long text is read in pieces of about this many words.'}),
'language': (list(LANGUAGES), {'default': 'English', 'tooltip': 'The language of the text. English + Japanese is for lines that mix the two.'}),
},
}
def speak(self, robin, text, voice, seed, temperature, top_k, top_p, cfg, retries=2, max_words=35, language='English'):
lang = LANGUAGES[language]
chunks = pieces.split(text, max_words, lang)
if not chunks:
raise ValueError('There is no text to read.')
model = robin.ready(speech=len(chunks) > 1)
audio = pieces.read(model, text, voice, seed=seed, cfg=cfg, retries=retries, max_words=max_words, lang=lang,
progress=Progress(len(chunks)), temperature=temperature, top_k=top_k, top_p=top_p)
return {'ui': {'robin_take': [take(audio, seed)]}, 'result': (to_audio(audio),)}
class RobinVoiceChanger:
DESCRIPTION = ('Performs a recording again in another voice: the same words, timing and delivery. The new voice comes from the '
'description, and from a Robin LoRA when one is connected. The last take plays on the node; connect Save Audio to '
'keep every take in the output folder.')
RETURN_TYPES = ('AUDIO',)
RETURN_NAMES = ('audio',)
FUNCTION = 'change'
OUTPUT_NODE = True
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'robin': ('ROBIN_MODEL',),
'source_audio': ('AUDIO', {'tooltip': f'The performance to keep, up to {MAX_SOURCE_SECONDS} seconds.'}),
'source_text': ('STRING', {'multiline': True, 'default': '', 'tooltip': 'Exactly what the source clip says.'}),
'voice': ('STRING', {'multiline': True, 'default': '',
'tooltip': 'The new voice in plain words. It can stay empty when a Robin LoRA is connected.'}),
**SAMPLING,
},
'optional': {
'retries': RETRIES,
'language': (list(LANGUAGES), {'default': 'English', 'tooltip': 'The language of the recording. English + Japanese is for lines that mix the two.'}),
},
}
def change(self, robin, source_audio, source_text, voice, seed, temperature, top_k, top_p, retries=2, language='English'):
if not source_text.strip():
raise ValueError('Fill in source_text with exactly what the source clip says.')
if not voice.strip() and robin.lora is None:
raise ValueError('Describe the new voice, or connect a Robin LoRA.')
source, source_rate = mono(source_audio)
seconds = len(source) / source_rate
if seconds > MAX_SOURCE_SECONDS:
raise ValueError(f'The source clip is {seconds:.0f} s long; the Voice Changer takes up to {MAX_SOURCE_SECONDS} s at a time.')
model = robin.ready(speech=True)
audio = pieces.change(model, source, source_rate, source_text, voice=voice, seed=seed, retries=retries, lang=LANGUAGES[language],
progress=Progress(2), temperature=temperature, top_k=top_k, top_p=top_p)
return {'ui': {'robin_take': [take(audio, seed)]}, 'result': (to_audio(audio),)}
class RobinLoadAudio:
DESCRIPTION = ('A recording for the Voice Changer: record one with your microphone, or choose an audio file. Recordings are '
'trimmed to the line and kept in ComfyUI\'s input folder.')
RETURN_TYPES = ('AUDIO',)
RETURN_NAMES = ('audio',)
FUNCTION = 'load'
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'audio': (input_audio(), {'tooltip': 'A recording in ComfyUI\'s input folder. Record and Choose file add new ones.'}),
},
}
@classmethod
def IS_CHANGED(cls, audio):
path = Path(folder_paths.get_annotated_filepath(audio))
return hashlib.sha256(path.read_bytes()).hexdigest() if path.is_file() else ''
@classmethod
def VALIDATE_INPUTS(cls, audio):
if not audio or not folder_paths.exists_annotated_filepath(audio):
return 'Record a line or choose an audio file first.' if not audio else f'There is no {audio} in ComfyUI\'s input folder.'
return True
def load(self, audio):
x, rate = sf.read(folder_paths.get_annotated_filepath(audio), dtype='float32', always_2d=True)
return ({'waveform': torch.from_numpy(np.ascontiguousarray(x.T))[None], 'sample_rate': int(rate)},)
class RobinCheckDataset:
DESCRIPTION = ('Checks a folder of clips for training a voice: each audio file needs a .txt transcript with the same name, and a '
'voice.txt can describe the voice. Lists the clips it will leave out and anything worth fixing, in plain words. '
f'Choose folder uploads a folder from your computer into input/{UPLOADS}.')
RETURN_TYPES = ('ROBIN_DATASET', 'STRING')
RETURN_NAMES = ('dataset', 'report')
FUNCTION = 'check'
OUTPUT_NODE = True
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'folder': ('STRING', {'default': '', 'tooltip': 'The folder of clips: audio files, each with a .txt transcript of the same name. '
'A relative path is inside ComfyUI\'s input folder.'}),
'language': (list(LANGUAGES), {'default': 'English', 'tooltip': 'The language of the recordings.'}),
},
}
@classmethod
def IS_CHANGED(cls, folder, language):
return folder_signature(folder) + language
def check(self, folder, language):
dataset, report = check_folder(folder, language)
if dataset is None:
raise ValueError(report)
return {'ui': {'text': [report]}, 'result': (dataset, report)}
class RobinTrainLoRA:
DESCRIPTION = ('Trains a voice LoRA on a checked dataset, with the same recipe as the Robin app. Every save writes a checkpoint and '
'reads the sample line in the new voice, into models/loras/robin/<name>. Cancel the queue to stop; running it again '
'continues from the last save.')
RETURN_TYPES = ('ROBIN_MODEL', 'STRING')
RETURN_NAMES = ('robin', 'lora_name')
FUNCTION = 'train'
OUTPUT_NODE = True
CATEGORY = 'Robin'
@classmethod
def INPUT_TYPES(cls):
return {
'required': {
'robin': ('ROBIN_MODEL',),
'dataset': ('ROBIN_DATASET',),
'name': ('STRING', {'default': 'my_voice', 'tooltip': 'The LoRA is saved as models/loras/robin/<name>/lora.safetensors.'}),
'steps': ('INT', {'default': 1000, 'min': 10, 'max': 20000, 'step': 10}),
'save_every': ('INT', {'default': 200, 'min': 10, 'max': 5000, 'step': 10, 'tooltip': 'How often to save a checkpoint and read the sample line.'}),
'sample_text': ('STRING', {'multiline': True, 'default': trainer.SAMPLE_TEXT['en'],
'tooltip': 'Read in the new voice at every save, so you can pick a step by ear.'}),
'restart': ('BOOLEAN', {'default': False, 'tooltip': 'Start over instead of continuing an unfinished run with this name.'}),
},
'optional': {
'rank_planner': ('INT', {'default': 32, 'min': 0, 'max': 256, 'step': 8, 'advanced': True}),
'rank_decoder': ('INT', {'default': 64, 'min': 0, 'max': 256, 'step': 8, 'advanced': True}),
'lr_planner': ('FLOAT', {'default': 5e-5, 'min': 1e-6, 'max': 1e-2, 'step': 1e-6, 'round': False, 'advanced': True}),
'lr_decoder': ('FLOAT', {'default': 2e-4, 'min': 1e-6, 'max': 1e-2, 'step': 1e-6, 'round': False, 'advanced': True}),
'clips_per_step': ('INT', {'default': 8, 'min': 1, 'max': 64, 'advanced': True}),
'seed': ('INT', {'default': 0, 'min': 0, 'max': 2 ** 31 - 1, 'control_after_generate': False, 'advanced': True}),
},
'hidden': {'unique_id': 'UNIQUE_ID'},
}
def train(self, robin, dataset, name, steps, save_every, sample_text, restart, rank_planner=32, rank_decoder=64, lr_planner=5e-5,
lr_decoder=2e-4, clips_per_step=8, seed=0, unique_id=None):
name = name.strip()
if not NAME.fullmatch(name):
raise ValueError('Name the LoRA with letters, digits, dashes, dots or underscores.')
base = Path(folder_paths.get_folder_paths('loras')[0])
out = base/'robin'/name
bar = comfy.utils.ProgressBar(steps)
listen, samples = f'robin/{name}_{uuid.uuid4().hex[:8]}', []
def progress(stage, step, total, path=None, **info):
if stage == 'sample':
samples.append(dict(preview(path, listen), step=step))
info = dict(sample=samples[-1])
else:
bar.update_absolute(step, total)
tell('robin.train', unique_id, stage=stage, step=step, total=total, **info)
tell('robin.train', unique_id, stage='start', step=0, total=steps, save_every=save_every)
if not restart:
for f in sorted((out/'samples').glob('*.wav')):
if f.stem == 'base' or SAVED.fullmatch(f.stem):
progress('sample', 0 if f.stem == 'base' else int(f.stem[5:]), steps, path=f)
model = RobinModel(robin.core).ready(speech=True, memory=TRAINING_MEMORY)
started = time.time()
with torch.inference_mode(False), torch.enable_grad():
trainable(model)
try:
path = trainer.train(model, dataset['folder'], out, lang=dataset['lang'], restart=restart, log=lambda m: logging.info(f'Robin: {m}'),
stop=mm.processing_interrupted, progress=progress, steps=steps, save_every=save_every,
rank_planner=rank_planner, rank_decoder=rank_decoder, lr_planner=lr_planner, lr_decoder=lr_decoder,
accum=clips_per_step, sample_text=sample_text.strip() or None, seed=seed)
except SystemExit as e:
tell('robin.train', unique_id, stage='stopped' if mm.processing_interrupted() else 'failed')
mm.throw_exception_if_processing_interrupted()
raise ValueError(str(e)) from None
if path is None or mm.processing_interrupted():
tell('robin.train', unique_id, stage='stopped')
mm.throw_exception_if_processing_interrupted()
raise ValueError('Training stopped before it finished.')
done = dict(stage='done', step=steps, total=steps, minutes=(time.time() - started) / 60, samples=samples)
tell('robin.train', unique_id, **done)
return {'ui': {'robin_train': [done]}, 'result': (RobinModel(robin.core, (str(path), 1.)), Path(path).relative_to(base).as_posix())}
routes = PromptServer.instance.routes
@routes.get('/robin/status')
async def status_route(request):
model = request.query.get('model', MODEL)
return web.json_response(model_status(model if model in model_folders() else MODEL, request.query.get('vae', VAE)))
@routes.get('/robin/uploaded')
async def uploaded_route(request):
base = (Path(folder_paths.get_input_directory())/UPLOADS).resolve()
folder = (base/request.query.get('name', '')).resolve()
if folder.parent != base:
return web.Response(status=400)
files = {f.relative_to(folder).as_posix(): f.stat().st_size for f in folder.rglob('*') if f.is_file()} if folder.is_dir() else {}
return web.json_response(dict(files=files))
@routes.post('/robin/open_recordings')
async def open_recordings_route(request):
folder = folder_paths.get_input_directory()
here = request.remote in ('127.0.0.1', '::1', '::ffff:127.0.0.1') and request.host.rsplit(':', 1)[-1] == str(PromptServer.instance.port)
desktop = sys.platform in ('darwin', 'win32') or os.environ.get('DISPLAY') or os.environ.get('WAYLAND_DISPLAY')
if not (here and desktop):
return web.json_response(dict(opened=False, folder=folder))
if sys.platform == 'win32':
os.startfile(folder)
else:
subprocess.Popen(['open' if sys.platform == 'darwin' else 'xdg-open', folder], stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL,
start_new_session=True)
return web.json_response(dict(opened=True, folder=folder))
@routes.post('/robin/check')
async def check_route(request):
body = await request.json()
language = body.get('language') if body.get('language') in LANGUAGES else 'English'
_, report = await asyncio.to_thread(check_folder, str(body.get('folder', '')), language)
return web.json_response(dict(report=report))
NODE_CLASS_MAPPINGS = {
'RobinLoader': RobinLoader,
'RobinLoRA': RobinLoRA,
'RobinSpeak': RobinSpeak,
'RobinVoiceChanger': RobinVoiceChanger,
'RobinLoadAudio': RobinLoadAudio,
'RobinCheckDataset': RobinCheckDataset,
'RobinTrainLoRA': RobinTrainLoRA,
}
NODE_DISPLAY_NAME_MAPPINGS = {
'RobinLoader': 'Robin Loader',
'RobinLoRA': 'Robin LoRA',
'RobinSpeak': 'Robin Speak',
'RobinVoiceChanger': 'Robin Voice Changer',
'RobinLoadAudio': 'Robin Load Audio',
'RobinCheckDataset': 'Robin Check Dataset',
'RobinTrainLoRA': 'Robin Train LoRA',
}
|