File size: 15,283 Bytes
9c2f1dd | 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 | from collections import OrderedDict
from dataclasses import dataclass
from typing import Optional, Tuple
import torch
import torch.nn as nn
import torch.nn.functional as F
from transformers import PreTrainedModel
from transformers.modeling_outputs import CausalLMOutput
from transformers.utils import ModelOutput, logging
from .modeling_msp_audio import MSPAudioModel
from .modeling_msp_visual import MSPVisualModel
from .configuration_msp import MSPConfig
from .modeling_msp_fusion import MSPFusionModel
logger = logging.get_logger(__name__)
@dataclass
class MSPOutput(ModelOutput):
loss: Optional[torch.FloatTensor] = None
logits: Optional[torch.FloatTensor] = None
audio_logits: Optional[torch.FloatTensor] = None
visual_logits: Optional[torch.FloatTensor] = None
audio_loss: Optional[torch.FloatTensor] = None
visual_loss: Optional[torch.FloatTensor] = None
last_hidden_state: Optional[torch.FloatTensor] = None
audio_hidden_state: Optional[torch.FloatTensor] = None
visual_hidden_state: Optional[torch.FloatTensor] = None
fusion_padding_mask: Optional[torch.Tensor] = None
fusion_input_lengths: Optional[torch.Tensor] = None
audio_input_lengths: Optional[torch.Tensor] = None
visual_input_lengths: Optional[torch.Tensor] = None
attentions: Optional[Tuple[torch.FloatTensor, ...]] = None
class MSPPreTrainedModel(PreTrainedModel):
config_class = MSPConfig
base_model_prefix = "msp"
main_input_name = "input_values"
input_modalities = ["audio", "video"]
supports_gradient_checkpointing = False
all_tied_weights_keys = OrderedDict()
def _init_weights(self, module):
if isinstance(module, nn.Linear):
module.weight.data.normal_(mean=0.0, std=0.02)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
def _apply_modality_dropout(
self,
has_audio: bool,
has_visual: bool,
) -> tuple[bool, bool]:
if not self.training:
return has_audio, has_visual
if not has_audio or not has_visual:
return has_audio, has_visual
if self.config.modality_dropout_prob <= 0.0:
return has_audio, has_visual
if torch.rand(()) >= self.config.modality_dropout_prob:
return has_audio, has_visual
audio_drop_prob = self.config.audio_dropout_prob
visual_drop_prob = self.config.visual_dropout_prob
total = audio_drop_prob + visual_drop_prob
if total <= 0:
return has_audio, has_visual
drop_audio = torch.rand(()) < (audio_drop_prob / total)
if drop_audio:
return False, True
return True, False
def _ctc_loss(
logits: torch.Tensor,
labels: torch.Tensor,
input_lengths: torch.Tensor,
pad_token_id: int,
reduction: str,
zero_infinity: bool,
) -> torch.Tensor:
"""Compute CTC loss from logits, labels, and pre-computed input lengths."""
labels_mask = labels >= 0
target_lengths = labels_mask.sum(-1)
flattened_targets = labels.masked_select(labels_mask)
log_probs = F.log_softmax(logits, dim=-1, dtype=torch.float32).transpose(0, 1)
with torch.backends.cudnn.flags(enabled=False):
return F.ctc_loss(
log_probs,
flattened_targets,
input_lengths,
target_lengths,
blank=pad_token_id,
reduction=reduction,
zero_infinity=zero_infinity,
)
class MSPModel(MSPPreTrainedModel):
def __init__(self, config: MSPConfig):
super().__init__(config)
# Audio encoder and auxiliary CTC head
self.audio_model = MSPAudioModel(config.audio_config)
self.audio_head = nn.Sequential(
nn.Dropout(config.audio_config.final_dropout),
nn.Linear(config.audio_config.hidden_size, config.audio_config.vocab_size),
)
# Visual encoder and auxiliary CTC head
self.visual_model = MSPVisualModel(config.visual_config)
self.visual_head = nn.Sequential(
nn.Dropout(config.visual_config.final_dropout),
nn.Linear(
config.visual_config.hidden_size,
config.visual_config.vocab_size,
),
)
# Bidirectional cross-attention fusion
self.fusion_model = MSPFusionModel(config.msp_fusion_config)
@property
def dummy_inputs(self) -> dict:
return {
"input_values": torch.zeros(1, 16000, dtype=torch.float32),
"pixel_values_videos": torch.zeros(1, 1, 10, 88, 88, dtype=torch.float32),
"padding_mask": torch.ones(1, 16000, dtype=torch.long),
"padding_mask_videos": torch.ones(1, 10, dtype=torch.long),
}
def forward(
self,
input_values: Optional[torch.Tensor] = None,
pixel_values_videos: Optional[torch.Tensor] = None,
padding_mask: Optional[torch.Tensor] = None,
padding_mask_videos: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
**kwargs,
) -> MSPOutput:
has_audio = input_values is not None
has_visual = pixel_values_videos is not None
if not has_audio and not has_visual:
raise ValueError(
"Either input_values or pixel_values_videos must be provided."
)
output_attentions = (
output_attentions if output_attentions is not None else False
)
use_audio, use_visual = self._apply_modality_dropout(has_audio, has_visual)
audio_output, visual_output, fusion_output = None, None, None
audio_hidden_states, visual_hidden_states = None, None
audio_logits, visual_logits = None, None
audio_input_lengths, visual_input_lengths, fusion_input_lengths = (
None,
None,
None,
)
if use_audio:
audio_output = self.audio_model(
input_values=input_values,
padding_mask=padding_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
padding_mask = (
padding_mask
if padding_mask is not None
else torch.ones_like(
input_values, dtype=torch.long, device=input_values.device
)
)
audio_input_lengths = self.audio_model._get_feat_extract_output_lengths(
padding_mask.sum(-1)
).to(torch.long)
audio_hidden_states = audio_output.last_hidden_state
audio_logits = self.audio_head(audio_hidden_states)
if use_visual:
visual_output = self.visual_model(
pixel_values_videos=pixel_values_videos,
padding_mask_videos=padding_mask_videos,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
padding_mask_videos = (
visual_output.padding_mask_videos
if padding_mask_videos is not None
else torch.ones(
(pixel_values_videos.shape[0], pixel_values_videos.shape[2]),
dtype=torch.long,
device=pixel_values_videos.device,
)
)
visual_input_lengths = (
padding_mask_videos.sum(-1)
.to(torch.long)
.to(pixel_values_videos.device)
)
visual_hidden_states = visual_output.last_hidden_state
visual_logits = self.visual_head(visual_hidden_states)
fusion_output = self.fusion_model.forward(
audio_hidden_states=audio_hidden_states,
visual_hidden_states=visual_hidden_states,
audio_key_padding_mask=padding_mask,
visual_key_padding_mask=padding_mask_videos,
output_attentions=output_attentions,
)
fusion_input_lengths = (
fusion_output.fusion_padding_mask.sum(-1)
.to(torch.long)
.to(fusion_output.fusion_padding_mask.device)
if fusion_output.fusion_padding_mask is not None and not use_audio
else audio_input_lengths
)
return MSPOutput(
last_hidden_state=fusion_output.last_hidden_state,
audio_hidden_state=fusion_output.audio_hidden_state,
visual_hidden_state=fusion_output.visual_hidden_state,
fusion_padding_mask=fusion_output.fusion_padding_mask,
audio_logits=audio_logits,
visual_logits=visual_logits,
audio_input_lengths=audio_input_lengths,
visual_input_lengths=visual_input_lengths,
fusion_input_lengths=fusion_input_lengths,
attentions=fusion_output.attentions,
)
class MSPForCTC(MSPPreTrainedModel):
def __init__(self, config: MSPConfig):
super().__init__(config)
if config.vocab_size is None:
raise ValueError(
"vocab_size must be set in MSPConfig to instantiate MSPForCTC."
)
self.msp = MSPModel(config)
# Final CTC head for the fused representation
self.msp_head = nn.Sequential(
nn.Dropout(config.final_dropout),
nn.Linear(config.msp_fusion_config.fusion_hidden_size, config.vocab_size),
)
@property
def dummy_inputs(self) -> dict:
return {
"input_values": torch.zeros(1, 16000, dtype=torch.float32),
"pixel_values_videos": torch.zeros(1, 1, 10, 88, 88, dtype=torch.float32),
"padding_mask": torch.ones(1, 16000, dtype=torch.long),
"padding_mask_videos": torch.ones(1, 10, dtype=torch.long),
"labels": torch.ones(1, 5, dtype=torch.long),
}
# --- Freeze helpers ---
def freeze_feature_encoder(self) -> None:
"""Freeze feature extractors of both encoders (for end-to-end fine-tuning)."""
self.msp.audio_model.feature_extractor._freeze_parameters()
for param in self.msp.visual_model.feature_extractor_video.parameters():
param.requires_grad = False
for param in self.msp.visual_model.feature_extractor_audio.parameters():
param.requires_grad = False
def freeze_base_model(self) -> None:
"""Freeze both encoders (for fusion-only training)."""
for param in self.msp.audio_model.parameters():
param.requires_grad = False
for param in self.msp.visual_model.parameters():
param.requires_grad = False
def freeze_audio_branch(self) -> None:
"""Freeze audio encoder and its CTC head (for fusion-only training)."""
for param in self.msp.audio_model.parameters():
param.requires_grad = False
for param in self.msp.audio_head.parameters():
param.requires_grad = False
def freeze_visual_branch(self) -> None:
"""Freeze visual encoder and its CTC head (for fusion-only training)."""
for param in self.msp.visual_model.parameters():
param.requires_grad = False
for param in self.msp.visual_head.parameters():
param.requires_grad = False
def forward(
self,
input_values: Optional[torch.Tensor] = None,
pixel_values_videos: Optional[torch.Tensor] = None,
padding_mask: Optional[torch.Tensor] = None,
padding_mask_videos: Optional[torch.Tensor] = None,
labels: Optional[torch.Tensor] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
**kwargs,
) -> CausalLMOutput:
if input_values is None and pixel_values_videos is None:
raise ValueError(
"Either input_values or pixel_values_videos must be provided."
)
msp_out = self.msp(
input_values=input_values,
pixel_values_videos=pixel_values_videos,
padding_mask=padding_mask,
padding_mask_videos=padding_mask_videos,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
)
# Final CTC logits from the fused representation
logits = self.msp_head(msp_out.last_hidden_state)
loss = None
if labels is not None:
valid_labels = labels[labels >= 0]
if (
valid_labels.numel() > 0
and valid_labels.max() >= self.config.vocab_size
):
raise ValueError(
f"Label value {valid_labels.max()} >= vocab_size={self.config.vocab_size}."
)
# Audio CTC loss
ctc_audio = None
if (
msp_out.audio_input_lengths is not None
and self.config.ctc_loss_audio_weight != 0.0
):
audio_lengths = msp_out.audio_input_lengths
ctc_audio = _ctc_loss(
logits=msp_out.audio_logits,
labels=labels,
input_lengths=audio_lengths,
pad_token_id=self.config.pad_token_id,
reduction=self.config.ctc_loss_reduction,
zero_infinity=self.config.ctc_zero_infinity,
)
# Visual CTC loss
ctc_visual = None
if (
msp_out.visual_input_lengths is not None
and self.config.ctc_loss_visual_weight != 0.0
):
visual_lengths = msp_out.visual_input_lengths
ctc_visual = _ctc_loss(
logits=msp_out.visual_logits,
labels=labels,
input_lengths=visual_lengths,
pad_token_id=self.config.pad_token_id,
reduction=self.config.ctc_loss_reduction,
zero_infinity=self.config.ctc_zero_infinity,
)
# Fusion CTC loss
ctc_msp = None
if (
msp_out.fusion_input_lengths is not None
and self.config.ctc_loss_msp_weight != 0.0
):
msp_lengths = msp_out.fusion_input_lengths
ctc_msp = _ctc_loss(
logits=logits,
labels=labels,
input_lengths=msp_lengths,
pad_token_id=self.config.pad_token_id,
reduction=self.config.ctc_loss_reduction,
zero_infinity=self.config.ctc_zero_infinity,
)
# Weighted combination
loss = self.config.ctc_loss_msp_weight * ctc_msp
if ctc_audio is not None:
loss = loss + self.config.ctc_loss_audio_weight * ctc_audio
if ctc_visual is not None:
loss = loss + self.config.ctc_loss_visual_weight * ctc_visual
return CausalLMOutput(loss=loss, logits=logits)
|