-
Notifications
You must be signed in to change notification settings - Fork 29
/
train_stage_2.py
390 lines (323 loc) · 13.1 KB
/
train_stage_2.py
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
import torch
import torch.optim as optim
from torch.utils.data import DataLoader
from omegaconf import OmegaConf
import wandb
from pathlib import Path
from dataset import VASADataset
from VASA import VASAFaceEncoder, VASADiffusionTransformer
import torch
from torch.utils.data import DataLoader
import torch.nn.functional as F
from einops import rearrange
import torch
from torch.utils.data import DataLoader
from torch.optim import AdamW
from torch.optim.lr_scheduler import CosineAnnealingLR
from accelerate import Accelerator
from omegaconf import OmegaConf
import wandb
from pathlib import Path
from collections import defaultdict
from tqdm import tqdm
import torch.nn.functional as F
from typing import Dict, Any, Optional
from accelerate.logging import get_logger
logger = get_logger(__name__)
class VASAStage2Trainer:
"""
Complete implementation of Stage 2 VASA training - Diffusion Transformer for motion generation
with distributed training, logging, and checkpoint management
"""
def __init__(
self,
face_encoder,
diffusion_model,
config,
):
self.config = config
self.face_encoder = face_encoder
self.diffusion_model = diffusion_model
# Initialize accelerator
self.accelerator = Accelerator(
gradient_accumulation_steps=config.training.gradient_accumulation_steps,
log_with="wandb",
mixed_precision=None
)
# Initialize trackers
self.accelerator.init_trackers(
project_name=config.project_name,
config=OmegaConf.to_container(config, resolve=True),
init_kwargs={"wandb": {
"name": config.experiment_name,
"dir": config.training.log_dir
}}
)
# Freeze face encoder
for param in self.face_encoder.parameters():
param.requires_grad = False
# Setup optimizer and scheduler
self.optimizer = AdamW(
self.diffusion_model.parameters(),
lr=config.training.lr,
betas=(config.training.beta1, config.training.beta2),
weight_decay=config.training.weight_decay
)
self.scheduler = CosineAnnealingLR(
self.optimizer,
T_max=config.training.num_epochs,
eta_min=config.training.min_lr
)
# Training state
self.current_epoch = 0
self.global_step = 0
self.best_metric = float('inf')
self.metric_history = []
def train_step(self, batch: Dict[str, torch.Tensor]) -> Dict[str, float]:
"""Single training step"""
# Extract data
images = batch['frames']
audio_features = batch['audio_features']
gaze = batch['gaze']
emotion = batch['emotion']
# Extract facial representations using frozen encoder
with torch.no_grad():
facial_reps = self.face_encoder.encode_holistic(
images[:, 0], # Source frame
gaze=gaze[:, 0],
emotion=emotion[:, 0]
)
# Get facial dynamics sequence
facial_dynamics = facial_reps['facial_dynamics']
# Sample timestep
batch_size = facial_dynamics.shape[0]
t = torch.randint(0, self.config.num_steps, (batch_size,),
device=self.accelerator.device)
# Add noise
noise = torch.randn_like(facial_dynamics)
noisy_dynamics = self.diffusion_model.add_noise(facial_dynamics, t, noise)
# Prepare conditions
conditions = {
'audio': audio_features,
'gaze': gaze,
'emotion': emotion
}
# Forward pass with CFG
with self.accelerator.autocast():
model_output = self.diffusion_model(
noisy_dynamics,
t,
conditions=conditions,
use_cfg=self.training
)
# Compute losses
losses = {}
# Base diffusion loss
losses['diffusion'] = F.mse_loss(
model_output['predicted_noise'],
noise
)
# CFG losses
if self.training and self.config.cfg_scales is not None:
for cond_type, scale in self.config.cfg_scales.items():
if scale > 0 and f'masked_{cond_type}' in model_output:
losses[f'cfg_{cond_type}'] = F.mse_loss(
model_output[f'masked_{cond_type}'],
model_output['uncond']
)
# Total loss
total_loss = losses['diffusion'] + sum(
self.config.lambda_cfg * losses[f'cfg_{k}']
for k in ['audio', 'gaze']
if f'cfg_{k}' in losses
)
# Backward pass
self.accelerator.backward(total_loss)
if self.accelerator.sync_gradients:
self.accelerator.clip_grad_norm_(
self.diffusion_model.parameters(),
self.config.max_grad_norm
)
return {
'total_loss': total_loss.item(),
**{k: v.item() for k, v in losses.items()}
}
def train_epoch(self, train_loader: DataLoader, epoch: int) -> Dict[str, float]:
"""Train for one epoch"""
self.diffusion_model.train()
metrics = defaultdict(float)
progress_bar = tqdm(
total=len(train_loader),
desc=f"Epoch {epoch+1}",
disable=not self.accelerator.is_local_main_process
)
for batch_idx, batch in enumerate(train_loader):
with self.accelerator.accumulate(self.diffusion_model):
step_metrics = self.train_step(batch)
# Update metrics
for k, v in step_metrics.items():
metrics[k] += v
self.optimizer.step()
self.optimizer.zero_grad()
# Log samples periodically
if batch_idx % self.config.sample_interval == 0:
self.log_samples(batch, self.global_step)
progress_bar.update(1)
self.global_step += 1
# Average metrics
metrics = {k: v / len(train_loader) for k, v in metrics.items()}
progress_bar.close()
return metrics
@torch.no_grad()
def validate(self, val_loader: DataLoader) -> Dict[str, float]:
"""Run validation"""
self.diffusion_model.eval()
metrics = defaultdict(float)
for batch in tqdm(val_loader, desc="Validation"):
step_metrics = self.train_step(batch)
for k, v in step_metrics.items():
metrics[k] += v
metrics = {k: v / len(val_loader) for k, v in metrics.items()}
return metrics
def train(self, train_loader: DataLoader, val_loader: DataLoader):
"""Complete training loop"""
# Prepare for training
(
self.face_encoder,
self.diffusion_model,
self.optimizer,
self.scheduler,
train_loader,
val_loader
) = self.accelerator.prepare(
self.face_encoder,
self.diffusion_model,
self.optimizer,
self.scheduler,
train_loader,
val_loader
)
try:
for epoch in range(self.current_epoch, self.config.training.num_epochs):
# Training
train_metrics = self.train_epoch(train_loader, epoch)
# Validation
val_metrics = self.validate(val_loader)
# Update scheduler
self.scheduler.step()
# Log metrics
metrics = {
"train/" + k: v for k, v in train_metrics.items()
}
metrics.update({
"val/" + k: v for k, v in val_metrics.items()
})
metrics["learning_rate"] = self.scheduler.get_last_lr()[0]
self.accelerator.log(metrics, step=self.global_step)
# Save checkpoint
if (epoch + 1) % self.config.save_interval == 0:
self.save_checkpoint(
epoch=epoch,
metrics=metrics,
is_best=val_metrics['total_loss'] < self.best_metric
)
# Update best metric
if val_metrics['total_loss'] < self.best_metric:
self.best_metric = val_metrics['total_loss']
# Early stopping
if not self.check_improvement(val_metrics['total_loss']):
print("Early stopping triggered!")
break
self.current_epoch = epoch + 1
except Exception as e:
print(f"Training error: {str(e)}")
raise
finally:
self.cleanup()
def log_samples(self, batch: Dict[str, torch.Tensor], step: int):
"""Log sample generations"""
if self.accelerator.is_local_main_process:
with torch.no_grad():
# Generate motion sequence
generated_motion = self.generate_motion(
batch['audio_features'][:1],
{
'gaze': batch['gaze'][:1],
'emotion': batch['emotion'][:1]
}
)
# Log samples
self.accelerator.log({
"samples/motion": wandb.Histogram(
generated_motion.cpu().numpy()
),
"samples/audio": wandb.Histogram(
batch['audio_features'][:1].cpu().numpy()
)
}, step=step)
@torch.no_grad()
def generate_motion(self, audio_features, conditions, cfg_scales=None):
"""Generate motion sequence"""
self.diffusion_model.eval()
# Initialize from noise
motion = torch.randn(
(1, self.config.sequence_length, self.config.motion_dim),
device=self.accelerator.device
)
# Iterative denoising
for t in reversed(range(self.config.num_steps)):
timesteps = torch.full((1,), t, device=self.accelerator.device)
model_output = self.diffusion_model(
motion,
timesteps,
conditions=conditions,
cfg_scales=cfg_scales
)
motion = self.diffusion_model.update_sample(
motion,
model_output['predicted_noise'],
timesteps
)
return motion
def save_checkpoint(self, epoch: int, metrics: Dict[str, float], is_best: bool = False):
"""Save training checkpoint"""
unwrapped_model = self.accelerator.unwrap_model(self.diffusion_model)
checkpoint = {
'epoch': epoch,
'global_step': self.global_step,
'model_state_dict': unwrapped_model.state_dict(),
'optimizer_state_dict': self.optimizer.state_dict(),
'scheduler_state_dict': self.scheduler.state_dict(),
'metrics': metrics,
'best_metric': self.best_metric,
'metric_history': self.metric_history,
'config': self.config
}
save_dir = Path(self.config.training.checkpoint_dir)
save_dir.mkdir(parents=True, exist_ok=True)
if is_best:
path = save_dir / 'best_model.pt'
else:
path = save_dir / f'checkpoint_epoch_{epoch}.pt'
self.accelerator.save(checkpoint, path)
def load_checkpoint(self, path: str):
"""Load training checkpoint"""
checkpoint = torch.load(path, map_location=self.accelerator.device)
self.diffusion_model.load_state_dict(checkpoint['model_state_dict'])
self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
self.scheduler.load_state_dict(checkpoint['scheduler_state_dict'])
self.current_epoch = checkpoint['epoch'] + 1
self.global_step = checkpoint['global_step']
self.best_metric = checkpoint['best_metric']
self.metric_history = checkpoint['metric_history']
def check_improvement(self, current_metric: float) -> bool:
"""Check for improvement (for early stopping)"""
self.metric_history.append(current_metric)
if len(self.metric_history) > self.config.patience:
best_recent = min(self.metric_history[-self.config.patience:])
if best_recent >= min(self.metric_history[:-self.config.patience]):
return False
return True
def cleanup(self):
"""Cleanup resources"""
self.accelerator.end_training()