File size: 56,909 Bytes
d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 531e776 d72b2c3 0a8807e d72b2c3 531e776 d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 531e776 d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 0a8807e d72b2c3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 |
# ========================= From conditioners.py
import soundfile
from collections import defaultdict
from copy import deepcopy
from dataclasses import dataclass, field
from itertools import chain
import logging
import math
from pathlib import Path
import random
import re
import typing as tp
import warnings
import einops
from num2words import num2words
import spacy
from transformers import T5EncoderModel, T5Tokenizer # type: ignore
import torch
import torch.nn.functional as F
from torch.nn.utils.rnn import pad_sequence
from audiocraft.streaming import StreamingModule
from audiocraft.transformer import create_sin_embedding
from audiocraft.utils.autocast import TorchAutocast
from audiocraft.utils.utils import collate, length_to_mask
from audiocraft.transformer import StreamingTransformer, create_norm_fn
from dataclasses import dataclass
from functools import partial
import logging
import math
import typing as tp
from torch import nn
from audiocraft.utils import utils
from audiocraft.codebooks_patterns import CodebooksPatternProvider
from audiocraft.activations import get_activation_fn
logger = logging.getLogger(__name__)
TextCondition = tp.Optional[str] # a text condition can be a string or None (if doesn't exist)
ConditionType = tp.Tuple[torch.Tensor, torch.Tensor] # condition, mask
class WavCondition(tp.NamedTuple):
wav: torch.Tensor
length: torch.Tensor
sample_rate: tp.List[int]
path: tp.List[tp.Optional[str]] = []
seek_time: tp.List[tp.Optional[float]] = []
class JointEmbedCondition(tp.NamedTuple):
wav: torch.Tensor
text: tp.List[tp.Optional[str]]
length: torch.Tensor
sample_rate: tp.List[int]
path: tp.List[tp.Optional[str]] = []
seek_time: tp.List[tp.Optional[float]] = []
@dataclass
class ConditioningAttributes:
text: tp.Dict[str, tp.Optional[str]] = field(default_factory=dict)
wav: tp.Dict[str, WavCondition] = field(default_factory=dict)
joint_embed: tp.Dict[str, JointEmbedCondition] = field(default_factory=dict)
def __getitem__(self, item):
return getattr(self, item)
@property
def text_attributes(self):
return self.text.keys()
@property
def wav_attributes(self):
return self.wav.keys()
@property
def joint_embed_attributes(self):
return self.joint_embed.keys()
@property
def attributes(self):
return {
"text": self.text_attributes,
"wav": self.wav_attributes,
"joint_embed": self.joint_embed_attributes,
}
def to_flat_dict(self):
return {
**{f"text.{k}": v for k, v in self.text.items()},
**{f"wav.{k}": v for k, v in self.wav.items()},
**{f"joint_embed.{k}": v for k, v in self.joint_embed.items()}
}
@classmethod
def from_flat_dict(cls, x):
out = cls()
for k, v in x.items():
kind, att = k.split(".")
out[kind][att] = v
return out
def nullify_condition(condition: ConditionType, dim: int = 1):
"""Transform an input condition to a null condition.
The way it is done by converting it to a single zero vector similarly
to how it is done inside WhiteSpaceTokenizer and NoopTokenizer.
Args:
condition (ConditionType): A tuple of condition and mask (tuple[torch.Tensor, torch.Tensor])
dim (int): The dimension that will be truncated (should be the time dimension)
WARNING!: dim should not be the batch dimension!
Returns:
ConditionType: A tuple of null condition and mask
"""
assert dim != 0, "dim cannot be the batch dimension!"
assert isinstance(condition, tuple) and \
isinstance(condition[0], torch.Tensor) and \
isinstance(condition[1], torch.Tensor), "'nullify_condition' got an unexpected input type!"
cond, mask = condition
B = cond.shape[0]
last_dim = cond.dim() - 1
out = cond.transpose(dim, last_dim)
out = 0. * out[..., :1]
out = out.transpose(dim, last_dim)
mask = torch.zeros((B, 1), device=out.device).int()
assert cond.dim() == out.dim()
return out, mask
def nullify_wav(cond: WavCondition) -> WavCondition:
"""Transform a WavCondition to a nullified WavCondition.
It replaces the wav by a null tensor, forces its length to 0, and replaces metadata by dummy attributes.
Args:
cond (WavCondition): Wav condition with wav, tensor of shape [B, T].
Returns:
WavCondition: Nullified wav condition.
"""
null_wav, _ = nullify_condition((cond.wav, torch.zeros_like(cond.wav)), dim=cond.wav.dim() - 1)
return WavCondition(
wav=null_wav,
length=torch.tensor([0] * cond.wav.shape[0], device=cond.wav.device),
sample_rate=cond.sample_rate,
path=[None] * cond.wav.shape[0],
seek_time=[None] * cond.wav.shape[0],
)
def nullify_joint_embed(embed: JointEmbedCondition) -> JointEmbedCondition:
"""Nullify the joint embedding condition by replacing it by a null tensor, forcing its length to 0,
and replacing metadata by dummy attributes.
Args:
cond (JointEmbedCondition): Joint embedding condition with wav and text, wav tensor of shape [B, C, T].
"""
null_wav, _ = nullify_condition((embed.wav, torch.zeros_like(embed.wav)), dim=embed.wav.dim() - 1)
return JointEmbedCondition(
wav=null_wav, text=[None] * len(embed.text),
length=torch.LongTensor([0]).to(embed.wav.device),
sample_rate=embed.sample_rate,
path=[None] * embed.wav.shape[0],
seek_time=[0] * embed.wav.shape[0],
)
class Tokenizer:
"""Base tokenizer implementation
(in case we want to introduce more advances tokenizers in the future).
"""
def __call__(self, texts: tp.List[tp.Optional[str]]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
raise NotImplementedError()
class WhiteSpaceTokenizer(Tokenizer):
"""This tokenizer should be used for natural language descriptions.
For example:
["he didn't, know he's going home.", 'shorter sentence'] =>
[[78, 62, 31, 4, 78, 25, 19, 34],
[59, 77, 0, 0, 0, 0, 0, 0]]
"""
PUNCTUATION = "?:!.,;"
def __init__(self, n_bins: int, pad_idx: int = 0, language: str = "en_core_web_sm",
lemma: bool = True, stopwords: bool = True) -> None:
self.n_bins = n_bins
self.pad_idx = pad_idx
self.lemma = lemma
self.stopwords = stopwords
try:
self.nlp = spacy.load(language)
except IOError:
spacy.cli.download(language) # type: ignore
self.nlp = spacy.load(language)
@tp.no_type_check
def __call__(self, texts: tp.List[tp.Optional[str]],
return_text: bool = False) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""Take a list of strings and convert them to a tensor of indices.
Args:
texts (list[str]): List of strings.
return_text (bool, optional): Whether to return text as additional tuple item. Defaults to False.
Returns:
tuple[torch.Tensor, torch.Tensor]:
- Indices of words in the LUT.
- And a mask indicating where the padding tokens are
"""
output, lengths = [], []
texts = deepcopy(texts)
for i, text in enumerate(texts):
# if current sample doesn't have a certain attribute, replace with pad token
if text is None:
output.append(torch.Tensor([self.pad_idx]))
lengths.append(0)
continue
# convert numbers to words
text = re.sub(r"(\d+)", lambda x: num2words(int(x.group(0))), text) # type: ignore
# normalize text
text = self.nlp(text) # type: ignore
# remove stopwords
if self.stopwords:
text = [w for w in text if not w.is_stop] # type: ignore
# remove punctuation
text = [w for w in text if w.text not in self.PUNCTUATION] # type: ignore
# lemmatize if needed
text = [getattr(t, "lemma_" if self.lemma else "text") for t in text] # type: ignore
texts[i] = " ".join(text)
lengths.append(len(text))
# convert to tensor
tokens = torch.Tensor([hash_trick(w, self.n_bins) for w in text])
output.append(tokens)
mask = length_to_mask(torch.IntTensor(lengths)).int()
padded_output = pad_sequence(output, padding_value=self.pad_idx).int().t()
if return_text:
return padded_output, mask, texts # type: ignore
return padded_output, mask
class NoopTokenizer(Tokenizer):
"""This tokenizer should be used for global conditioners such as: artist, genre, key, etc.
The difference between this and WhiteSpaceTokenizer is that NoopTokenizer does not split
strings, so "Jeff Buckley" will get it's own index. Whereas WhiteSpaceTokenizer will
split it to ["Jeff", "Buckley"] and return an index per word.
For example:
["Queen", "ABBA", "Jeff Buckley"] => [43, 55, 101]
["Metal", "Rock", "Classical"] => [0, 223, 51]
"""
def __init__(self, n_bins: int, pad_idx: int = 0):
self.n_bins = n_bins
self.pad_idx = pad_idx
def __call__(self, texts: tp.List[tp.Optional[str]]) -> tp.Tuple[torch.Tensor, torch.Tensor]:
output, lengths = [], []
for text in texts:
# if current sample doesn't have a certain attribute, replace with pad token
if text is None:
output.append(self.pad_idx)
lengths.append(0)
else:
output.append(hash_trick(text, self.n_bins))
lengths.append(1)
tokens = torch.LongTensor(output).unsqueeze(1)
mask = length_to_mask(torch.IntTensor(lengths)).int()
return tokens, mask
class BaseConditioner(nn.Module):
"""Base model for all conditioner modules.
We allow the output dim to be different than the hidden dim for two reasons:
1) keep our LUTs small when the vocab is large;
2) make all condition dims consistent.
Args:
dim (int): Hidden dim of the model.
output_dim (int): Output dim of the conditioner.
"""
def __init__(self, dim: int, output_dim: int):
super().__init__()
self.dim = dim
self.output_dim = output_dim
self.output_proj = nn.Linear(dim, output_dim)
def forward(self, inputs: tp.Any) -> ConditionType:
"""Gets input that should be used as conditioning (e.g, genre, description or a waveform).
Outputs a ConditionType, after the input data was embedded as a dense vector.
Returns:
ConditionType:
- A tensor of size [B, T, D] where B is the batch size, T is the length of the
output embedding and D is the dimension of the embedding.
- And a mask indicating where the padding tokens.
"""
raise NotImplementedError()
class TextConditioner(BaseConditioner):
...
class T5Conditioner(TextConditioner):
"""T5-based TextConditioner.
Args:
name (str): Name of the T5 model.
output_dim (int): Output dim of the conditioner.
finetune (bool): Whether to fine-tune T5 at train time.
device (str): Device for T5 Conditioner.
autocast_dtype (tp.Optional[str], optional): Autocast dtype.
word_dropout (float, optional): Word dropout probability.
normalize_text (bool, optional): Whether to apply text normalization.
"""
MODELS = ["t5-small", "t5-base", "t5-large", "t5-3b", "t5-11b",
"google/flan-t5-small", "google/flan-t5-base", "google/flan-t5-large",
"google/flan-t5-xl", "google/flan-t5-xxl"]
MODELS_DIMS = {
"t5-small": 512,
"t5-base": 768,
"t5-large": 1024,
"t5-3b": 1024,
"t5-11b": 1024,
"google/flan-t5-small": 512,
"google/flan-t5-base": 768,
"google/flan-t5-large": 1024,
"google/flan-t5-3b": 1024,
"google/flan-t5-11b": 1024,
}
def __init__(self, name: str, output_dim: int, finetune: bool, device: str,
autocast_dtype: tp.Optional[str] = 'float32', word_dropout: float = 0.,
normalize_text: bool = False):
assert name in self.MODELS, f"Unrecognized t5 model name (should in {self.MODELS})"
super().__init__(self.MODELS_DIMS[name], output_dim)
self.device = device
self.name = name
self.finetune = finetune
self.word_dropout = word_dropout
if autocast_dtype is None or self.device == 'cpu':
self.autocast = TorchAutocast(enabled=False)
if self.device != 'cpu':
logger.warning("T5 has no autocast, this might lead to NaN")
else:
dtype = getattr(torch, autocast_dtype)
assert isinstance(dtype, torch.dtype)
logger.info(f"T5 will be evaluated with autocast as {autocast_dtype}")
self.autocast = TorchAutocast(enabled=True, device_type=self.device, dtype=dtype)
# Let's disable logging temporarily because T5 will vomit some errors otherwise.
# thanks https://gist.github.com/simon-weber/7853144
previous_level = logging.root.manager.disable
logging.disable(logging.ERROR)
with warnings.catch_warnings():
warnings.simplefilter("ignore")
try:
self.t5_tokenizer = T5Tokenizer.from_pretrained(name)
t5 = T5EncoderModel.from_pretrained(name).train(mode=finetune)
finally:
logging.disable(previous_level)
if finetune:
self.t5 = t5
else:
# this makes sure that the t5 models is not part
# of the saved checkpoint
self.__dict__['t5'] = t5.to(device)
self.normalize_text = normalize_text
if normalize_text:
self.text_normalizer = WhiteSpaceTokenizer(1, lemma=True, stopwords=True)
def tokenize(self, x: tp.List[tp.Optional[str]]) -> tp.Dict[str, torch.Tensor]:
# if current sample doesn't have a certain attribute, replace with empty string
entries: tp.List[str] = [xi if xi is not None else "" for xi in x]
if self.normalize_text:
_, _, entries = self.text_normalizer(entries, return_text=True)
if self.word_dropout > 0. and self.training:
new_entries = []
for entry in entries:
words = [word for word in entry.split(" ") if random.random() >= self.word_dropout]
new_entries.append(" ".join(words))
entries = new_entries
empty_idx = torch.LongTensor([i for i, xi in enumerate(entries) if xi == ""])
inputs = self.t5_tokenizer(entries, return_tensors='pt', padding=True).to(self.device)
mask = inputs['attention_mask']
mask[empty_idx, :] = 0 # zero-out index where the input is non-existant
return inputs
def forward(self, inputs: tp.Dict[str, torch.Tensor]) -> ConditionType:
mask = inputs['attention_mask']
with torch.set_grad_enabled(self.finetune), self.autocast:
embeds = self.t5(**inputs).last_hidden_state
embeds = self.output_proj(embeds.to(self.output_proj.weight))
embeds = (embeds * mask.unsqueeze(-1))
return embeds, mask
class JointEmbeddingConditioner(BaseConditioner):
"""Joint embedding conditioning supporting both audio or text conditioning.
Args:
dim (int): Dimension.
output_dim (int): Output dimension.
device (str): Device.
attribute (str): Attribute used by the conditioner.
autocast_dtype (str): Autocast for the conditioner.
quantize (bool): Whether to quantize the CLAP embedding.
n_q (int): Number of residual quantizers (used if quantize is true).
bins (int): Quantizers' codebooks size (used if quantize is true).
kwargs: Additional parameters for residual vector quantizer.
"""
def __init__(self, dim: int, output_dim: int, device: str, attribute: str,
autocast_dtype: tp.Optional[str] = 'float32', quantize: bool = True,
n_q: int = 12, bins: int = 1024, **kwargs):
super().__init__(dim=dim, output_dim=output_dim)
self.device = device
self.attribute = attribute
if autocast_dtype is None or device == 'cpu':
self.autocast = TorchAutocast(enabled=False)
logger.warning("JointEmbeddingConditioner has no autocast, this might lead to NaN.")
else:
dtype = getattr(torch, autocast_dtype)
assert isinstance(dtype, torch.dtype)
logger.info(f"JointEmbeddingConditioner will be evaluated with autocast as {autocast_dtype}.")
self.autocast = TorchAutocast(enabled=True, device_type=self.device, dtype=dtype)
# residual vector quantizer to discretize the conditioned embedding
self.quantizer=None
if quantize:
print('\n\n\n\nWANTS TO QUANTIZE on Inference\n\n\n\n')
# self.quantizer = ResidualVectorQuantizer(dim, n_q=n_q, bins=bins, **kwargs)
def _get_embed(self, x: JointEmbedCondition) -> tp.Tuple[torch.Tensor, torch.Tensor]:
"""Get joint embedding in latent space from the inputs.
Returns:
tuple[torch.Tensor, torch.Tensor]: Tensor for the latent embedding
and corresponding empty indexes.
"""
raise NotImplementedError()
def forward(self, x: JointEmbedCondition) -> ConditionType:
with self.autocast:
embed, empty_idx = self._get_embed(x)
if self.quantizer is not None:
embed = embed.view(-1, self.dim, 1)
q_res = self.quantizer(embed, frame_rate=1)
out_embed = q_res.x.view(-1, self.dim)
else:
out_embed = embed
out_embed = self.output_proj(out_embed).view(-1, 1, self.output_dim)
mask = torch.ones(*out_embed.shape[:2], device=out_embed.device)
mask[empty_idx, :] = 0 # zero-out index where the input is non-existant
out_embed = (out_embed * mask.unsqueeze(-1))
return out_embed, mask
def tokenize(self, x: JointEmbedCondition) -> JointEmbedCondition:
return x
class ConditioningProvider(nn.Module):
"""Prepare and provide conditions given all the supported conditioners.
Args:
conditioners (dict): Dictionary of conditioners.
device (torch.device or str, optional): Device for conditioners and output condition types.
"""
def __init__(self, conditioners: tp.Dict[str, BaseConditioner], device: tp.Union[torch.device, str] = "cpu"):
super().__init__()
self.device = device
self.conditioners = nn.ModuleDict(conditioners)
@property
def joint_embed_conditions(self):
return [m.attribute for m in self.conditioners.values() if isinstance(m, JointEmbeddingConditioner)]
@property
def has_joint_embed_conditions(self):
return len(self.joint_embed_conditions) > 0
@property
def text_conditions(self):
return [k for k, v in self.conditioners.items() if isinstance(v, TextConditioner)]
@property
def wav_conditions(self):
return [k for k, v in self.conditioners.items() if isinstance(v, WaveformConditioner)]
@property
def has_wav_condition(self):
return len(self.wav_conditions) > 0
def forward(self, tokenized: tp.Dict[str, tp.Any]) -> tp.Dict[str, ConditionType]:
"""Compute pairs of `(embedding, mask)` using the configured conditioners and the tokenized representations.
The output is for example:
{
"genre": (torch.Tensor([B, 1, D_genre]), torch.Tensor([B, 1])),
"description": (torch.Tensor([B, T_desc, D_desc]), torch.Tensor([B, T_desc])),
...
}
Args:
tokenized (dict): Dict of tokenized representations as returned by `tokenize()`.
"""
output = {}
for attribute, inputs in tokenized.items():
condition, mask = self.conditioners[attribute](inputs)
output[attribute] = (condition, mask)
return output
def _collate_text(self, samples: tp.List[ConditioningAttributes]) -> tp.Dict[str, tp.List[tp.Optional[str]]]:
"""Given a list of ConditioningAttributes objects, compile a dictionary where the keys
are the attributes and the values are the aggregated input per attribute.
For example:
Input:
[
ConditioningAttributes(text={"genre": "Rock", "description": "A rock song with a guitar solo"}, wav=...),
ConditioningAttributes(text={"genre": "Hip-hop", "description": "A hip-hop verse"}, wav=...),
]
Output:
{
"genre": ["Rock", "Hip-hop"],
"description": ["A rock song with a guitar solo", "A hip-hop verse"]
}
Args:
samples (list of ConditioningAttributes): List of ConditioningAttributes samples.
Returns:
dict[str, list[str, optional]]: A dictionary mapping an attribute name to text batch.
"""
out: tp.Dict[str, tp.List[tp.Optional[str]]] = defaultdict(list)
texts = [x.text for x in samples]
for text in texts:
for condition in self.text_conditions:
out[condition].append(text[condition])
return out
def _collate_wavs(self, samples: tp.List[ConditioningAttributes]) -> tp.Dict[str, WavCondition]:
"""Generate a dict where the keys are attributes by which we fetch similar wavs,
and the values are Tensors of wavs according to said attributes.
*Note*: by the time the samples reach this function, each sample should have some waveform
inside the "wav" attribute. It should be either:
1. A real waveform
2. A null waveform due to the sample having no similar waveforms (nullified by the dataset)
3. A null waveform due to it being dropped in a dropout module (nullified by dropout)
Args:
samples (list of ConditioningAttributes): List of ConditioningAttributes samples.
Returns:
dict[str, WavCondition]: A dictionary mapping an attribute name to wavs.
"""
wavs = defaultdict(list)
lengths = defaultdict(list)
sample_rates = defaultdict(list)
paths = defaultdict(list)
seek_times = defaultdict(list)
out: tp.Dict[str, WavCondition] = {}
for sample in samples:
for attribute in self.wav_conditions:
wav, length, sample_rate, path, seek_time = sample.wav[attribute]
assert wav.dim() == 3, f"Got wav with dim={wav.dim()}, but expected 3 [1, C, T]"
assert wav.size(0) == 1, f"Got wav [B, C, T] with shape={wav.shape}, but expected B == 1"
# mono-channel conditioning
wav = wav.mean(1, keepdim=True) # [1, 1, T]
wavs[attribute].append(wav.flatten()) # [T]
lengths[attribute].append(length)
sample_rates[attribute].extend(sample_rate)
paths[attribute].extend(path)
seek_times[attribute].extend(seek_time)
# stack all wavs to a single tensor
for attribute in self.wav_conditions:
stacked_wav, _ = collate(wavs[attribute], dim=0)
out[attribute] = WavCondition(
stacked_wav.unsqueeze(1), torch.cat(lengths[attribute]), sample_rates[attribute],
paths[attribute], seek_times[attribute])
return out
def _collate_joint_embeds(self, samples: tp.List[ConditioningAttributes]) -> tp.Dict[str, JointEmbedCondition]:
"""Generate a dict where the keys are attributes by which we compute joint embeddings,
and the values are Tensors of pre-computed embeddings and the corresponding text attributes.
Args:
samples (list[ConditioningAttributes]): List of ConditioningAttributes samples.
Returns:
A dictionary mapping an attribute name to joint embeddings.
"""
texts = defaultdict(list)
wavs = defaultdict(list)
lengths = defaultdict(list)
sample_rates = defaultdict(list)
paths = defaultdict(list)
seek_times = defaultdict(list)
channels: int = 0
out = {}
for sample in samples:
for attribute in self.joint_embed_conditions:
wav, text, length, sample_rate, path, seek_time = sample.joint_embed[attribute]
assert wav.dim() == 3
if channels == 0:
channels = wav.size(1)
else:
assert channels == wav.size(1), "not all audio has same number of channels in batch"
assert wav.size(0) == 1, "Expecting single-wav batch in the collate method"
wav = einops.rearrange(wav, "b c t -> (b c t)") # [1, C, T] => [C * T]
wavs[attribute].append(wav)
texts[attribute].extend(text)
lengths[attribute].append(length)
sample_rates[attribute].extend(sample_rate)
paths[attribute].extend(path)
seek_times[attribute].extend(seek_time)
for attribute in self.joint_embed_conditions:
stacked_texts = texts[attribute]
stacked_paths = paths[attribute]
stacked_seek_times = seek_times[attribute]
stacked_wavs = pad_sequence(wavs[attribute]).to(self.device)
stacked_wavs = einops.rearrange(stacked_wavs, "(c t) b -> b c t", c=channels)
stacked_sample_rates = sample_rates[attribute]
stacked_lengths = torch.cat(lengths[attribute]).to(self.device)
assert stacked_lengths.size(0) == stacked_wavs.size(0)
assert len(stacked_sample_rates) == stacked_wavs.size(0)
assert len(stacked_texts) == stacked_wavs.size(0)
out[attribute] = JointEmbedCondition(
text=stacked_texts, wav=stacked_wavs,
length=stacked_lengths, sample_rate=stacked_sample_rates,
path=stacked_paths, seek_time=stacked_seek_times)
return out
class ConditionFuser(StreamingModule):
"""Condition fuser handles the logic to combine the different conditions
to the actual model input.
Args:
fuse2cond (tp.Dict[str, str]): A dictionary that says how to fuse
each condition. For example:
{
"prepend": ["description"],
"sum": ["genre", "bpm"],
"cross": ["description"],
}
cross_attention_pos_emb (bool, optional): Use positional embeddings in cross attention.
cross_attention_pos_emb_scale (int): Scale for positional embeddings in cross attention if used.
"""
FUSING_METHODS = ["sum", "prepend", "cross", "input_interpolate"]
def __init__(self, fuse2cond: tp.Dict[str, tp.List[str]], cross_attention_pos_emb: bool = False,
cross_attention_pos_emb_scale: float = 1.0):
super().__init__()
assert all(
[k in self.FUSING_METHODS for k in fuse2cond.keys()]
), f"Got invalid fuse method, allowed methods: {self.FUSING_METHODS}"
self.cross_attention_pos_emb = cross_attention_pos_emb
self.cross_attention_pos_emb_scale = cross_attention_pos_emb_scale
self.fuse2cond: tp.Dict[str, tp.List[str]] = fuse2cond
self.cond2fuse: tp.Dict[str, str] = {}
for fuse_method, conditions in fuse2cond.items():
for condition in conditions:
self.cond2fuse[condition] = fuse_method
def forward(
self,
input: torch.Tensor,
conditions: tp.Dict[str, ConditionType]
) -> tp.Tuple[torch.Tensor, tp.Optional[torch.Tensor]]:
"""Fuse the conditions to the provided model input.
Args:
input (torch.Tensor): Transformer input.
conditions (dict[str, ConditionType]): Dict of conditions.
Returns:
tuple[torch.Tensor, torch.Tensor]: The first tensor is the transformer input
after the conditions have been fused. The second output tensor is the tensor
used for cross-attention or None if no cross attention inputs exist.
"""
B, T, _ = input.shape
if 'offsets' in self._streaming_state:
first_step = False
offsets = self._streaming_state['offsets']
else:
first_step = True
offsets = torch.zeros(input.shape[0], dtype=torch.long, device=input.device)
assert set(conditions.keys()).issubset(set(self.cond2fuse.keys())), \
f"given conditions contain unknown attributes for fuser, " \
f"expected {self.cond2fuse.keys()}, got {conditions.keys()}"
cross_attention_output = None
for cond_type, (cond, cond_mask) in conditions.items():
op = self.cond2fuse[cond_type]
if op == 'sum':
input += cond
elif op == 'input_interpolate':
cond = einops.rearrange(cond, "b t d -> b d t")
cond = F.interpolate(cond, size=input.shape[1])
input += einops.rearrange(cond, "b d t -> b t d")
elif op == 'prepend':
if first_step:
input = torch.cat([cond, input], dim=1)
elif op == 'cross':
if cross_attention_output is not None:
cross_attention_output = torch.cat([cross_attention_output, cond], dim=1)
else:
cross_attention_output = cond
else:
raise ValueError(f"unknown op ({op})")
if self.cross_attention_pos_emb and cross_attention_output is not None:
print('SIN EMBED')
positions = torch.arange(
cross_attention_output.shape[1],
device=cross_attention_output.device
).view(1, -1, 1)
pos_emb = create_sin_embedding(positions, cross_attention_output.shape[-1])
cross_attention_output = cross_attention_output + self.cross_attention_pos_emb_scale * pos_emb
if self._is_streaming:
self._streaming_state['offsets'] = offsets + T
return input, cross_attention_output
# ============================================== From LM.py
logger = logging.getLogger(__name__)
ConditionTensors = tp.Dict[str, ConditionType]
CFGConditions = tp.Union[ConditionTensors, tp.Tuple[ConditionTensors, ConditionTensors]]
def get_init_fn(method: str, input_dim: int, init_depth: tp.Optional[int] = None):
"""LM layer initialization.
Inspired from xlformers: https://github.com/fairinternal/xlformers
Args:
method (str): Method name for init function. Valid options are:
'gaussian', 'uniform'.
input_dim (int): Input dimension of the initialized module.
init_depth (int, optional): Optional init depth value used to rescale
the standard deviation if defined.
"""
# Compute std
std = 1 / math.sqrt(input_dim)
# Rescale with depth
if init_depth is not None:
std = std / math.sqrt(2 * init_depth)
if method == 'gaussian':
return partial(
torch.nn.init.trunc_normal_, mean=0.0, std=std, a=-3 * std, b=3 * std
)
elif method == 'uniform':
bound = math.sqrt(3) * std # ensure the standard deviation is `std`
return partial(torch.nn.init.uniform_, a=-bound, b=bound)
else:
raise ValueError("Unsupported layer initialization method")
def init_layer(m: nn.Module,
method: str,
init_depth: tp.Optional[int] = None,
zero_bias_init: bool = False):
"""Wrapper around ``get_init_fn`` for proper initialization of LM modules.
Args:
m (nn.Module): Module to initialize.
method (str): Method name for the init function.
init_depth (int, optional): Optional init depth value used to rescale
the standard deviation if defined.
zero_bias_init (bool): Whether to initialize the bias to 0 or not.
"""
if isinstance(m, nn.Linear):
init_fn = get_init_fn(method, m.in_features, init_depth=init_depth)
if m.weight.device.type == 'cpu' and m.weight.dtype == torch.float16:
weight = m.weight.float()
init_fn(weight)
m.weight.data[:] = weight.half()
else:
init_fn(m.weight)
if zero_bias_init and m.bias is not None:
nn.init.constant_(m.bias, 0)
elif isinstance(m, nn.Embedding):
init_fn = get_init_fn(method, m.embedding_dim, init_depth=None)
if m.weight.device.type == 'cpu' and m.weight.dtype == torch.float16:
weight = m.weight.float()
init_fn(weight)
m.weight.data[:] = weight.half()
else:
init_fn(m.weight)
class ScaledEmbedding(nn.Embedding):
"""Boost learning rate for embeddings (with `scale`).
"""
def __init__(self, *args, lr=None, **kwargs):
super().__init__(*args, **kwargs)
self.lr = lr
def make_optim_group(self):
group = {"params": list(self.parameters())}
if self.lr is not None:
group["lr"] = self.lr
return group
@dataclass
class LMOutput:
# The logits are already re-aligned with the input codes
# hence no extra shift is required, e.g. when computing CE
logits: torch.Tensor # [B, K, T, card]
mask: torch.Tensor # [B, K, T]
class LMModel(StreamingModule):
"""Transformer-based language model on multiple streams of codes.
Args:
pattern_provider (CodebooksPatternProvider): Pattern provider for codebook interleaving.
condition_provider (MusicConditioningProvider): Conditioning provider from metadata.
fuser (ConditionFuser): Fuser handling the fusing of conditions with language model input.
n_q (int): Number of parallel streams to model.
card (int): Cardinality, vocabulary size.
dim (int): Dimension of the transformer encoder.
num_heads (int): Number of heads for the transformer encoder.
hidden_scale (int): Scale for hidden feed forward dimension of the transformer encoder.
norm (str): Normalization method.
norm_first (bool): Use pre-norm instead of post-norm.
emb_lr (float, optional): Embedding-specific learning rate.
bias_proj (bool): Use bias for output projections.
weight_init (str, optional): Method for weight initialization.
depthwise_init (str, optional): Method for depthwise weight initialization.
zero_bias_init (bool): If true and bias in Linears, initialize bias to zeros.
cfg_dropout (float): Classifier-free guidance dropout.
cfg_coef (float): Classifier-free guidance coefficient.
attribute_dropout (dict): Attribute dropout probabilities.
two_step_cfg (bool): Whether to run classifier free-guidance with 2 distinct steps.
**kwargs: Additional parameters for the transformer encoder.
"""
def __init__(self, pattern_provider: CodebooksPatternProvider, condition_provider: ConditioningProvider,
fuser: ConditionFuser, n_q: int = 8, card: int = 1024, dim: int = 128, num_heads: int = 8,
hidden_scale: int = 4, norm: str = 'layer_norm', norm_first: bool = False,
emb_lr: tp.Optional[float] = None, bias_proj: bool = True,
weight_init: tp.Optional[str] = None, depthwise_init: tp.Optional[str] = None,
zero_bias_init: bool = False, cfg_dropout: float = 0, cfg_coef: float = 1.0,
attribute_dropout: tp.Dict[str, tp.Dict[str, float]] = {}, two_step_cfg: bool = False,
**kwargs):
super().__init__()
self.cfg_coef = cfg_coef
self.condition_provider = condition_provider
self.fuser = fuser
self.card = card # 2048 ?
embed_dim = self.card + 1
self.n_q = n_q
self.dim = dim
self.pattern_provider = pattern_provider
self.two_step_cfg = two_step_cfg
self.emb = nn.ModuleList([ScaledEmbedding(embed_dim, dim, lr=emb_lr) for _ in range(n_q)])
if 'activation' in kwargs:
kwargs['activation'] = get_activation_fn(kwargs['activation'])
self.transformer = StreamingTransformer(
d_model=dim, num_heads=num_heads, dim_feedforward=int(hidden_scale * dim),
norm=norm, norm_first=norm_first, **kwargs)
self.out_norm: tp.Optional[nn.Module] = None
if norm_first:
self.out_norm = create_norm_fn(norm, dim)
self.linears = nn.ModuleList([nn.Linear(dim, self.card, bias=bias_proj) for _ in range(n_q)])
self._init_weights(weight_init, depthwise_init, zero_bias_init)
self._fsdp: tp.Optional[nn.Module]
self.__dict__['_fsdp'] = None
def _init_weights(self, weight_init: tp.Optional[str], depthwise_init: tp.Optional[str], zero_bias_init: bool):
"""Initialization of the transformer module weights.
Args:
weight_init (str, optional): Weight initialization strategy. See ``get_init_fn`` for valid options.
depthwise_init (str, optional): Depthwise initialization strategy. The following options are valid:
'current' where the depth corresponds to the current layer index or 'global' where the total number
of layer is used as depth. If not set, no depthwise initialization strategy is used.
zero_bias_init (bool): Whether to initialize bias to zero or not.
"""
assert depthwise_init is None or depthwise_init in ['current', 'global']
assert depthwise_init is None or weight_init is not None, \
"If 'depthwise_init' is defined, a 'weight_init' method should be provided."
assert not zero_bias_init or weight_init is not None, \
"If 'zero_bias_init', a 'weight_init' method should be provided"
if weight_init is None:
return
for emb_layer in self.emb:
init_layer(emb_layer, method=weight_init, init_depth=None, zero_bias_init=zero_bias_init)
for layer_idx, tr_layer in enumerate(self.transformer.layers):
depth = None
if depthwise_init == 'current':
depth = layer_idx + 1
elif depthwise_init == 'global':
depth = len(self.transformer.layers)
init_fn = partial(init_layer, method=weight_init, init_depth=depth, zero_bias_init=zero_bias_init)
tr_layer.apply(init_fn)
for linear in self.linears:
init_layer(linear, method=weight_init, init_depth=None, zero_bias_init=zero_bias_init)
@property
def special_token_id(self) -> int:
return self.card
@property
def num_codebooks(self) -> int:
return self.n_q
def forward(self, sequence: torch.Tensor,
conditions: tp.List[ConditioningAttributes],
condition_tensors: tp.Optional[ConditionTensors] = None,
stage: int = -1) -> torch.Tensor:
"""Apply language model on sequence and conditions.
Given a tensor of sequence of shape [B, K, S] with K the number of codebooks and
S the sequence steps, return the logits with shape [B, card, K, S].
Args:
indices (torch.Tensor): Indices of the codes to model.
conditions (list of ConditioningAttributes): Conditions to use when modeling
the given codes. Note that when evaluating multiple time with the same conditioning
you should pre-compute those and pass them as `condition_tensors`.
condition_tensors (dict[str, ConditionType], optional): Pre-computed conditioning
tensors, see `conditions`.
stage (int): The codebook level that is being predicted. Relevant for MAGNeT
in which prediction is done in a codebook-by-codebook manner.
Takes values in range(n_q), and ignored by default.
Returns:
torch.Tensor: Logits.
"""
B, K, S = sequence.shape
assert K == self.num_codebooks, "Sequence shape must match the specified number of codebooks"
input_ = sum([self.emb[k](sequence[:, k]) for k in range(K)])
if condition_tensors is None:
assert not self._is_streaming, "Conditions tensors should be precomputed when streaming."
# encode conditions and fuse, both have a streaming cache to not recompute when generating.
condition_tensors = self.condition_provider(tokenized)
else:
assert not conditions, "Shouldn't pass both conditions and condition_tensors."
input_, cross_attention_input = self.fuser(input_, condition_tensors)
out = self.transformer(input_, cross_attention_src=cross_attention_input,
src_mask=(self.attn_mask_per_stage[stage] if stage >= 0 else None))
if self.out_norm:
out = self.out_norm(out)
logits = torch.stack([self.linears[k](out) for k in range(K)], dim=1) # [B, K, S, card]
# remove the prefix from the model outputs
if len(self.fuser.fuse2cond['prepend']) > 0:
logits = logits[:, :, -S:]
print('PRESFIX')
return logits # [B, K, S, card]
def compute_predictions(
self, codes: torch.Tensor,
conditions: tp.List[ConditioningAttributes],
condition_tensors: tp.Optional[ConditionTensors] = None,
stage: int = -1,
keep_only_valid_steps: bool = True) -> LMOutput:
"""Given an input tensor of codes [B, K, T] and list of conditions, runs the model
forward using the specified codes interleaving pattern.
Args:
codes (torch.Tensor): Input codes of shape [B, K, T] with B the batch size,
K the number of codebooks and T the number of timesteps.
conditions (list of ConditioningAttributes): conditionings to use when modeling
the given codes. Note that when evaluating multiple time with the same conditioning
you should pre-compute those and pass them as `condition_tensors`.
condition_tensors (dict[str, ConditionType], optional): pre-computed conditioning
tensors, see `conditions`.
stage (int): The codebook level that is being predicted. Relevant for MAGNeT
in which prediction is done in a codebook-by-codebook manner.
Takes values in range(n_q), and ignored by default.
keep_only_valid_steps (bool): Build a sequence from the pattern up to valid (= fully defined) steps.
Steps that are beyond valid steps will be replaced by the special_token in that case.
Returns:
LMOutput: Language model outputs
logits (torch.Tensor) of shape [B, K, T, card] corresponding to the provided codes,
i.e. the first item corresponds to logits to predict the first code, meaning that
no additional shifting of codes and logits is required.
mask (torch.Tensor) of shape [B, K, T], mask over valid and invalid positions.
Given the specified interleaving strategies, parts of the logits and codes should
not be considered as valid predictions because of invalid context.
"""
B, K, T = codes.shape
codes = codes.contiguous()
# map codes [B, K, T] into pattern sequence [B, K, S] using special_token_id for masked tokens
# what is the T is it 2048 ?
# and then what is pattern -> another function?
pattern = self.pattern_provider.get_pattern(T)
sequence_codes, sequence_indexes, sequence_mask = pattern.build_pattern_sequence(
codes, self.special_token_id, keep_only_valid_steps=keep_only_valid_steps,
)
# apply model on pattern sequence
model = self if self._fsdp is None else self._fsdp
logits = model(sequence_codes, conditions, condition_tensors, stage=stage) # [B, K, S, card]
# map back the logits on pattern sequence to logits on original codes: [B, K, S, card] -> [B, K, T, card]
# and provide the corresponding mask over invalid positions of tokens
logits = logits.permute(0, 3, 1, 2) # [B, card, K, S]
# note: we use nans as special token to make it obvious if we feed unexpected logits
logits, logits_indexes, logits_mask = pattern.revert_pattern_logits(
logits, float('nan'), keep_only_valid_steps=keep_only_valid_steps
)
logits = logits.permute(0, 2, 3, 1) # [B, K, T, card]
logits_mask = logits_mask[None, :, :].expand(B, -1, -1) # [K, T] -> [B, K, T]
return LMOutput(logits, logits_mask)
def _sample_next_token(self,
sequence,
cfg_conditions,
unconditional_state,
use_sampling=False,
temp: float = 1.0,
top_k: int = 0,
top_p: float = 0.0,
cfg_coef: tp.Optional[float] = None,
two_step_cfg: tp.Optional[bool] = None) -> torch.Tensor:
"""Sample next token from the model given a sequence and a set of conditions. The model supports
multiple sampling strategies (greedy sampling, softmax, top-k, top-p...).
Args:
sequence (torch.Tensor): Current sequence of shape [B, K, S]
with K corresponding to the number of codebooks and S the number of sequence steps.
S = 1 in streaming mode, except for the first step that contains a bigger prompt.
condition_tensors (dict[str, ConditionType): Set of conditions. If CFG is used,
should be twice the batch size, being the concatenation of the conditions + null conditions.
use_sampling (bool): Whether to use a sampling strategy or not.
temp (float): Sampling temperature.
top_k (int): K for "top-k" sampling.
top_p (float): P for "top-p" sampling.
cfg_coef (float, optional): classifier free guidance coefficient
Returns:
next_token (torch.Tensor): Next token tensor of shape [B, K, 1].
"""
B = sequence.shape[0]
cfg_coef = self.cfg_coef if cfg_coef is None else cfg_coef
model = self if self._fsdp is None else self._fsdp
two_step_cfg = self.two_step_cfg if two_step_cfg is None else two_step_cfg
if two_step_cfg and cfg_conditions != {}:
print('\nNOT HERE\n')
else:
print('C')
assert isinstance(cfg_conditions, dict)
condition_tensors = cfg_conditions
if condition_tensors:
# print('\nD\n')
# Preparing for CFG, predicting both conditional and unconditional logits.
sequence = torch.cat([sequence, sequence], dim=0)
all_logits = model(
sequence,
conditions=[], condition_tensors=condition_tensors)
if condition_tensors:
cond_logits, uncond_logits = all_logits.split(B, dim=0) #torch.Size([2, 4, 1, 2048])
# logits = uncond_logits + (cond_logits - uncond_logits) * cfg_coef
# logits = 3 * cond_logits - 2.4 * uncond_logits
logits = 2 * cond_logits - 1.4 * uncond_logits
else:
print('\nF!\n')
logits = logits.permute(0, 1, 3, 2) # [B, K, card, T]
logits = logits[..., -1] # [B x K x card]
# Apply softmax for sampling if temp > 0. Else, do greedy sampling to avoid zero division error.
if use_sampling and temp > 0.0:
# print(f'\nR {temp=} {top_p=} {top_k=}\n') -------------> R temp=1.0 top_p=0.0 top_k=250
probs = torch.softmax(logits / temp, dim=-1)
if top_p > 0.0:
next_token = utils.sample_top_p(probs, p=top_p)
elif top_k > 0:
next_token = utils.sample_top_k(probs, k=top_k)
else:
next_token = utils.multinomial(probs, num_samples=1)
else:
#
print('\nNeverHere\n')
return next_token
@torch.no_grad()
def generate(self,
prompt: tp.Optional[torch.Tensor] = None,
conditions: tp.List[ConditioningAttributes] = [],
num_samples: tp.Optional[int] = None,
max_gen_len: int = 256,
use_sampling: bool = True,
temp: float = 1.0,
top_k: int = 250,
top_p: float = 0.0,
cfg_coef: tp.Optional[float] = None,
two_step_cfg: tp.Optional[bool] = None,
remove_prompts: bool = False,
check: bool = False,
callback: tp.Optional[tp.Callable[[int, int], None]] = None,
**kwargs) -> torch.Tensor:
"""Generate tokens sampling from the model given a prompt or unconditionally. Generation can
be performed in a greedy fashion or using sampling with top K and top P strategies.
Args:
prompt (torch.Tensor, optional): Prompt tokens of shape [B, K, T].
conditions_tensors (list of ConditioningAttributes, optional): List of conditions.
num_samples (int, optional): Number of samples to generate when no prompt and no conditions are given.
max_gen_len (int): Maximum generation length.
use_sampling (bool): Whether to use a sampling strategy or not.
temp (float): Sampling temperature.
top_k (int): K for "top-k" sampling.
top_p (float): P for "top-p" sampling.
cfg_coeff (float, optional): Classifier-free guidance coefficient.
two_step_cfg (bool, optional): Whether to perform classifier-free guidance with two steps generation.
remove_prompts (bool): Whether to remove prompts from generation or not.
check (bool): Whether to apply further checks on generated sequence.
callback (Callback, optional): Callback function to report generation progress.
Returns:
torch.Tensor: Generated tokens.
"""
assert not self.training, "generation shouldn't be used in training mode."
first_param = next(iter(self.parameters()))
device = first_param.device
# Checking all input shapes are consistent.
possible_num_samples = []
if num_samples is not None:
possible_num_samples.append(num_samples)
elif prompt is not None:
possible_num_samples.append(prompt.shape[0])
elif conditions:
possible_num_samples.append(len(conditions))
else:
possible_num_samples.append(1)
assert [x == possible_num_samples[0] for x in possible_num_samples], "Inconsistent inputs shapes"
num_samples = possible_num_samples[0]
# below we create set of conditions: one conditional and one unconditional
# to do that we merge the regular condition together with the null condition
# we then do 1 forward pass instead of 2.
# the reason for that is two-fold:
# 1. it is about x2 faster than doing 2 forward passes
# 2. avoid the streaming API treating the 2 passes as part of different time steps
# We also support doing two different passes, in particular to ensure that
# the padding structure is exactly the same between train and test.
# With a batch size of 1, this can be slower though.
cfg_conditions: CFGConditions
two_step_cfg = self.two_step_cfg if two_step_cfg is None else two_step_cfg
if conditions:
null_conditions = conditions
if two_step_cfg:
cfg_conditions = (
self.condition_provider(self.condition_provider.tokenize(conditions)),
self.condition_provider(self.condition_provider.tokenize(null_conditions)),
)
else:
conditions = conditions + null_conditions
tokenized = self.condition_provider.tokenize(conditions)
cfg_conditions = self.condition_provider(tokenized)
else:
cfg_conditions = {}
if prompt is None:
assert num_samples > 0
prompt = torch.zeros((num_samples, self.num_codebooks, 0), dtype=torch.long, device=device)
B, K, T = prompt.shape
start_offset = T
assert start_offset < max_gen_len
pattern = self.pattern_provider.get_pattern(max_gen_len)
# this token is used as default value for codes that are not generated yet
unknown_token = -1
gen_codes = torch.full((B, K, max_gen_len), unknown_token, dtype=torch.long, device=device)
gen_codes[..., :start_offset] = prompt
# create the gen_sequence with proper interleaving from the pattern: [B, K, S]
gen_sequence, indexes, mask = pattern.build_pattern_sequence(gen_codes, self.special_token_id)
# retrieve the start_offset in the sequence:
# it is the first sequence step that contains the `start_offset` timestep
start_offset_sequence = pattern.get_first_step_with_timesteps(start_offset)
assert start_offset_sequence is not None
with self.streaming():
unconditional_state = self.get_streaming_state()
prev_offset = 0
gen_sequence_len = gen_sequence.shape[-1] # gen_sequence shape is [B, K, S]
for offset in range(start_offset_sequence, gen_sequence_len):
# get current sequence (note that the streaming API is providing the caching over previous offsets)
curr_sequence = gen_sequence[..., prev_offset:offset]
curr_mask = mask[None, ..., prev_offset:offset].expand(B, -1, -1)
if check:
# check coherence between mask and sequence
assert (curr_sequence == torch.where(curr_mask, curr_sequence, self.special_token_id)).all()
# should never happen as gen_sequence is filled progressively
assert not (curr_sequence == unknown_token).any()
# sample next token from the model, next token shape is [B, K, 1]
next_token = self._sample_next_token(
curr_sequence, cfg_conditions, unconditional_state, use_sampling, temp, top_k, top_p,
cfg_coef=cfg_coef, two_step_cfg=two_step_cfg)
# ensure the tokens that should be masked are properly set to special_token_id
# as the model never output special_token_id
valid_mask = mask[..., offset:offset+1].expand(B, -1, -1)
# next_token[~valid_mask] = self.special_token_id
# print(f'{unconditional_state=} \n
# print('Set All to Special')
# next_token[:] = self.special_token_id
# ensure we don't overwrite prompt tokens, we only write over unknown tokens
gen_sequence[..., offset:offset+1] = torch.where(
gen_sequence[..., offset:offset+1] == unknown_token,
next_token, gen_sequence[..., offset:offset+1]
)
prev_offset = offset
if callback is not None:
callback(1 + offset - start_offset_sequence, gen_sequence_len - start_offset_sequence)
unconditional_state.clear()
out_codes, out_indexes, out_mask = pattern.revert_pattern_sequence(gen_sequence, special_token=unknown_token)
out_start_offset = start_offset if remove_prompts else 0
out_codes = out_codes[..., out_start_offset:max_gen_len]
# ensure the returned codes are all valid
# assert (out_codes >= 0).all() and (out_codes <= self.card).all()
return out_codes
|