Spaces:
Running
on
Zero
Running
on
Zero
File size: 19,934 Bytes
ceeabec |
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 |
import logging
from dataclasses import dataclass, field
from typing import Optional
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
import numpy as np
import random
import os
import sys
from fairseq.data.data_utils import compute_mask_indices
from fairseq.models import BaseFairseqModel, register_model
from fairseq.models.wav2vec import (
Wav2Vec2Config,
TransformerEncoder,
)
# Debug print to show where Wav2Vec2Config is defined
print(f"Wav2Vec2Config is imported from: {Wav2Vec2Config.__module__}")
print(f"Full path: {sys.modules[Wav2Vec2Config.__module__].__file__}")
from fairseq.modules import (
LayerNorm,
)
logger = logging.getLogger(__name__)
@dataclass
class SignHubertConfig(Wav2Vec2Config):
# pos_conv_kernel: int = field(default=32)
conv_pos: int = field(default=32)
discrete: bool = field(default=False)
codebook_size: int = field(default=256)
channels_embed_dim: int = field(default=384)
channels_pose_embed_dim: int = field(default=14)
intermediate_dim: int = field(default=1024) # This will be overridden if needed
mask_strategy: str = field(default="random")
channels: str = field(default="face,left_hand,right_hand,body_posture")
@register_model("signhubert_onlyhands", dataclass=SignHubertConfig)
class SignHubertModel(BaseFairseqModel):
def __init__(self, cfg: SignHubertConfig):
super().__init__()
self.cfg = cfg
# print(cfg)
self.discrete = cfg.discrete # since it's hubert this will always be discrete anyways
self.embed = cfg.encoder_embed_dim # whether it is small(384), base(768), large, etc.
self.channel_embed = cfg.channels_embed_dim # embedding dimension for face, left_hand and right_hand (default: 384)
self.channel_pose_embed = cfg.channels_pose_embed_dim # embedding dimension for pose (default: 14)
self.intermediate_dim = cfg.intermediate_dim # intermediate dimension before the projection layer to encoder_embed_dim (default: 1024)
self.channels = cfg.channels.split(",")
self.post_extract_proj = nn.Linear(cfg.intermediate_dim, cfg.encoder_embed_dim) # 4 channels concatenated
self.mask_prob = cfg.mask_prob
self.mask_selection = cfg.mask_selection
self.mask_strategy = cfg.mask_strategy
self.mask_other = cfg.mask_other
self.mask_length = cfg.mask_length
self.no_mask_overlap = cfg.no_mask_overlap
self.mask_min_space = cfg.mask_min_space
self.mask_channel_prob = cfg.mask_channel_prob
self.mask_channel_before = cfg.mask_channel_before
self.mask_channel_selection = cfg.mask_channel_selection
self.mask_channel_other = cfg.mask_channel_other
self.mask_channel_length = cfg.mask_channel_length
self.no_mask_channel_overlap = cfg.no_mask_channel_overlap
self.mask_channel_min_space = cfg.mask_channel_min_space
self.dropout_input = nn.Dropout(cfg.dropout_input)
self.dropout_features = nn.Dropout(cfg.dropout_features)
self.feature_grad_mult = cfg.feature_grad_mult
self.mask_emb = nn.Parameter(
torch.FloatTensor(1, 1, 1, cfg.intermediate_dim // len(self.channels)).uniform_()
)
self.encoder = TransformerEncoder(cfg)
self.layer_norm = LayerNorm(self.channel_embed * len(self.channels))
if "face" in self.channels:
self.layer_norm_face = LayerNorm(self.channel_embed)
self.face_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // len(self.channels))
if "left_hand" in self.channels:
self.layer_norm_lhand = LayerNorm(self.channel_embed)
self.left_hand_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // len(self.channels))
if "right_hand" in self.channels:
self.layer_norm_rhand = LayerNorm(self.channel_embed)
self.right_hand_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // len(self.channels))
if "body_posture" in self.channels:
self.layer_norm_body = LayerNorm(self.channel_pose_embed)
self.body_posture_proj = nn.Linear(self.channel_pose_embed, cfg.intermediate_dim // len(self.channels))
self.codebook_size = cfg.codebook_size # number of codebook vectors
self.heads = []
for i in range(len(self.channels)):
self.heads.append(nn.Linear(cfg.encoder_embed_dim, cfg.codebook_size))
self.heads = torch.nn.ModuleList(self.heads)
# self.heads = torch.nn.ModuleList([
# nn.Linear(cfg.encoder_embed_dim, cfg.codebook_size) ,
# nn.Linear(cfg.encoder_embed_dim, cfg.codebook_size),
# nn.Linear(cfg.encoder_embed_dim, cfg.codebook_size),
# ]
# )
# # Define separate linear layers for each channel
# self.face_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // 4)
# self.left_hand_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // 4)
# self.right_hand_proj = nn.Linear(self.channel_embed, cfg.intermediate_dim // 4)
# self.body_posture_proj = nn.Linear(self.channel_pose_embed, cfg.intermediate_dim // 4)
def state_dict(self, destination=None, prefix="", keep_vars=False):
state = super().state_dict(destination, prefix, keep_vars)
return state
@classmethod
def build_model(cls, cfg: SignHubertConfig, task=None):
"""Build a new model instance."""
return cls(cfg)
def apply_mask(
self,
x,
padding_mask,
mask_indices=None,
mask_channel_indices=None,
):
B, T, C, D = x.shape
# Initialize a mask vector with ones (same shape as x)
mask = torch.ones_like(x)
# channel masking
if self.mask_prob > 0 and self.mask_strategy == "channel":
if mask_indices is None:
mask_indices = torch.zeros_like(x[:,:,:,0], dtype=bool)
num_channels_to_mask = int(C * self.mask_prob)
num_channels_to_mask = max(1, num_channels_to_mask)
for i in range(B):
channels_to_mask = np.random.choice(C, num_channels_to_mask, replace=False)
mask_indices[i, :, channels_to_mask] = True
mask[mask_indices.unsqueeze(-1).expand(-1, -1, -1, D)] = 0
# gloss/time masking
elif self.mask_prob > 0 and self.mask_strategy == "gloss":
if mask_indices is None:
mask_indices_channel = compute_mask_indices(
(B, T),
padding_mask,
self.mask_prob,
self.mask_length,
self.mask_selection,
self.mask_other,
min_masks=1,
no_overlap=self.no_mask_channel_overlap,
min_space=self.mask_min_space,
require_same_masks=self.cfg.require_same_masks,
mask_dropout=self.cfg.mask_dropout,
)
mask_indices_channel = torch.from_numpy(mask_indices_channel).to(x.device)
# Apply the same mask to all channels
mask_indices = mask_indices_channel.unsqueeze(2).expand(-1, -1, C)
mask_indices = mask_indices.unsqueeze(3).expand(-1, -1, -1, D)
mask[mask_indices] = 0
# random masking
elif self.mask_prob > 0 and self.mask_strategy == "random":
if mask_indices is None:
mask_indices = compute_mask_indices(
(B, T*C), # Note: T*C instead of T
padding_mask,
self.mask_prob,
self.mask_length,
self.mask_selection,
self.mask_other,
min_masks=1,
no_overlap=self.no_mask_channel_overlap,
min_space=self.mask_min_space,
require_same_masks=self.cfg.require_same_masks,
mask_dropout=self.cfg.mask_dropout,
)
mask_indices = torch.from_numpy(mask_indices).to(x.device)
mask_indices = mask_indices.view(B, T, C)
mask_indices = mask_indices.unsqueeze(3).expand(-1, -1, -1, D)
mask[mask_indices] = 0
else:
raise ValueError(f"unknown mask strategy {self.mask_strategy}")
# Apply the mask to x and return the masked tensor with the same shape as x
# x = x * mask
x = x * mask + self.mask_emb * (1 - mask)
return x, mask
# mask is a tensor of shape BxTx4x256 where 0 means the value is masked and 1 means the value is not masked
def forward(
self,
source,
padding_mask=None,
mask=True,
features_only=False,
layer=None,
mask_indices=None,
mask_channel_indices=None,
padding_count=None,
kmeans_labels=None,
):
channels_to_use = []
for c in self.channels:
if c in source[0]:
channels_to_use.append(c)
for c in channels_to_use:
if c == "face":
face_features_list = []
label_face_features_list = []
elif c == "left_hand":
left_hand_features_list = []
label_left_hand_features_list = []
elif c == "right_hand":
right_hand_features_list = []
label_right_hand_features_list = []
elif c == "body_posture":
body_posture_features_list = []
label_body_posture_features_list = []
# # source is a list of dictionaries with keys "face", "left_hand", "right_hand", "body_posture"
# face_features_list = []
# left_hand_features_list = []
# right_hand_features_list = []
# body_posture_features_list = []
# label_face_features_list = []
# label_left_hand_features_list = []
# label_right_hand_features_list = []
# label_body_posture_features_list = []
# for sample in source:
# face_features_list.append(sample["face"]) # Tx384
# left_hand_features_list.append(sample["left_hand"]) # Tx384
# right_hand_features_list.append(sample["right_hand"]) # Tx384
# body_posture_features_list.append(sample["body_posture"]) # Tx14
# label_face_features_list.append(sample["label_face"]) # Tx1
# label_left_hand_features_list.append(sample["label_left_hand"]) # Tx1
# label_right_hand_features_list.append(sample["label_right_hand"]) # Tx1
# label_body_posture_features_list.append(sample["label_body_posture"]) # Tx1
for sample in source:
for c in channels_to_use:
if c == "face":
face_features_list.append(sample["face"]) # Tx384
label_face_features_list.append(sample["label_face"]) # Tx1
elif c == "left_hand":
left_hand_features_list.append(sample["left_hand"]) # Tx384
label_left_hand_features_list.append(sample["label_left_hand"]) # Tx1
elif c == "right_hand":
right_hand_features_list.append(sample["right_hand"]) # Tx384
label_right_hand_features_list.append(sample["label_right_hand"]) # Tx1
elif c == "body_posture":
body_posture_features_list.append(sample["body_posture"]) # Tx14
label_body_posture_features_list.append(sample["label_body_posture"]) # Tx1
# face_features = torch.stack(face_features_list) # BxTx384
# left_hand_features = torch.stack(left_hand_features_list) # BxTx384
# right_hand_features = torch.stack(right_hand_features_list) # BxTx384
# body_posture_features = torch.stack(body_posture_features_list) # BxTx14
# face_labels = torch.stack(label_face_features_list) # BxTx1
# left_hand_labels = torch.stack(label_left_hand_features_list) # BxTx1
# right_hand_labels = torch.stack(label_right_hand_features_list) # BxTx1
# body_posture_labels = torch.stack(label_body_posture_features_list) # BxTx1
# # Apply layer normalization to each part
# face_features = self.layer_norm_face(face_features) # BxTx384
# left_hand_features = self.layer_norm_lhand(left_hand_features) # BxTx384
# right_hand_features = self.layer_norm_rhand(right_hand_features) # BxTx384
# body_posture_features = self.layer_norm_body(body_posture_features) # BxTx14
# # Apply separate linear projections for each channel
# face_features = self.face_proj(face_features) # BxTx256
# left_hand_features = self.left_hand_proj(left_hand_features) # BxTx256
# right_hand_features = self.right_hand_proj(right_hand_features) # BxTx256
# body_posture_features = self.body_posture_proj(body_posture_features) # BxTx256
features_list = []
labels_list = []
for c in channels_to_use:
if c == "face":
face_features = torch.stack(face_features_list) # BxTx384
face_labels = torch.stack(label_face_features_list) # BxTx1
face_features = self.layer_norm_face(face_features) # BxTx384
face_features = self.face_proj(face_features) # BxTx256
features_list.append(face_features)
labels_list.append(face_labels)
elif c == "left_hand":
left_hand_features = torch.stack(left_hand_features_list) # BxTx384
left_hand_labels = torch.stack(label_left_hand_features_list) # BxTx1
left_hand_features = self.layer_norm_lhand(left_hand_features) # BxTx384
left_hand_features = self.left_hand_proj(left_hand_features) # BxTx256
features_list.append(left_hand_features)
labels_list.append(left_hand_labels)
elif c == "right_hand":
right_hand_features = torch.stack(right_hand_features_list) # BxTx384
right_hand_labels = torch.stack(label_right_hand_features_list) # BxTx1
right_hand_features = self.layer_norm_rhand(right_hand_features) # BxTx384
right_hand_features = self.right_hand_proj(right_hand_features) # BxTx256
features_list.append(right_hand_features)
labels_list.append(right_hand_labels)
elif c == "body_posture":
body_posture_features = torch.stack(body_posture_features_list) # BxTx14
body_posture_labels = torch.stack(label_body_posture_features_list) # BxTx1
body_posture_features = self.layer_norm_body(body_posture_features) # BxTx14
body_posture_features = self.body_posture_proj(body_posture_features) # BxTx256
features_list.append(body_posture_features)
labels_list.append(body_posture_labels)
# concatenate the projected features to have dimension BxTxCxD where C=4 and D=256
# features = torch.stack(
# [
# face_features,
# left_hand_features,
# right_hand_features,
# body_posture_features
# ],
# dim=2) # BxTx4x256
features = torch.stack(features_list, dim=2) # BxTx4x256
if mask:
x, mask_indices = self.apply_mask(
features,
padding_mask,
mask_indices=mask_indices,
mask_channel_indices=mask_channel_indices,
)
# mask_indices is a tensor of shape BxTx4x256 where 0 means the value is masked and 1 means the value is not masked
else:
x = features
mask_indices = None
x = self.dropout_input(x) # BxTx4x256
x = x.view(x.size(0), x.size(1), -1) # BxTx1024
if self.post_extract_proj is not None:
x = self.post_extract_proj(x) # BxTx768
x, layer_results = self.encoder(
x,
padding_mask=padding_mask,
layer=layer,
)
if features_only:
return {
"x": x,
"padding_mask": padding_mask,
"layer_results": layer_results,
}
result = {
"losses": {},
}
# use linear heads to compute the discrete prediction for each channel and make it into a single tensor of shape BxTxCxcodebook_size
predictions = []
for i, head in enumerate(self.heads):
channel_pred = head(x) # BxTxcodebook_size
predictions.append(channel_pred)
predictions = torch.stack(predictions, dim=2) # BxTx4xcodebook_size
# labels = torch.stack(
# [
# face_labels,
# left_hand_labels,
# right_hand_labels,
# body_posture_labels
# ],
# dim=2) # BxTx4x1
labels = torch.stack(labels_list, dim=2) # BxTx4x1
# print(f"predictions shape: {predictions.shape} and labels shape: {labels.shape}")
predictions_flat = predictions.view(-1, self.codebook_size) # Shape: (B * T * C, codebook_size)
labels_flat = labels.view(-1) # Shape: (B * T * C)
# Ensure labels are of correct shape
labels_flat = labels_flat.squeeze(-1) # Remove the last dimension if it's size 1
# Correct the mask_indices to match the shape of predictions_flat
mask_indices_reduced = mask_indices.any(dim=-1) # Reduce mask to (B, T, C) by collapsing last dimension
mask_indices_flat = mask_indices_reduced.view(-1) # Flatten to match the shape of (B * T * C)
# Calculate the loss only for the masked positions (where mask_indices_flat is zero)
masked_loss = F.cross_entropy(
predictions_flat[mask_indices_flat == 0],
labels_flat[mask_indices_flat == 0],
reduction='none'
)
# Store the result
result['losses']['kmeans_loss'] = masked_loss
if "sample_size" not in result:
result['sample_size'] = masked_loss.numel()
return result
@staticmethod
def compute_var(y):
y = y.view(-1, y.size(-1))
if dist.is_initialized():
zc = torch.tensor(y.size(0)).cuda()
zs = y.sum(dim=0)
zss = (y ** 2).sum(dim=0)
dist.all_reduce(zc)
dist.all_reduce(zs)
dist.all_reduce(zss)
var = zss / (zc - 1) - (zs ** 2) / (zc * (zc - 1))
return torch.sqrt(var + 1e-6).mean()
else:
return torch.sqrt(y.var(dim=0) + 1e-6).mean()
def extract_features(
self, source, padding_mask, kmeans_labels, mask=False, layer=None
):
res = self.forward(
source,
padding_mask,
mask=mask,
features_only=True,
layer=layer,
kmeans_labels=kmeans_labels,
)
return res
def remove_pretraining_modules(self, last_layer=None):
self.heads = None
self.final_proj = None
if last_layer is not None:
self.encoder.layers = nn.ModuleList(
l for i, l in enumerate(self.encoder.layers) if i <= last_layer
)
|