File size: 19,387 Bytes
02ca424
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import os
import random
import argparse
from pathlib import Path
from typing import Dict, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import DataLoader
from torch.optim import AdamW

import numpy as np
from PIL import Image
from tqdm import tqdm
from VITON_Dataset import VITONHDTestDataset

# Import your custom modules
from load_model import preload_models_from_standard_weights
from ddpm import DDPMSampler
from utils import check_inputs, get_time_embedding, prepare_image, prepare_mask_image, save_debug_visualization, compute_vae_encodings
from diffusers.utils.torch_utils import randn_tensor


class CatVTONTrainer:
    """Simplified CatVTON Training Class with PEFT, CFG and DREAM support"""
    
    def __init__(
        self,
        models: Dict[str, nn.Module],
        train_dataloader: DataLoader,
        val_dataloader: Optional[DataLoader] = None,
        device: str = "cuda",
        learning_rate: float = 1e-5,
        num_epochs: int = 50,
        save_steps: int = 1000,
        output_dir: str = "./checkpoints",
        cfg_dropout_prob: float = 0.1,
        max_grad_norm: float = 1.0,
        use_peft: bool = True,
        dream_lambda: float = 10.0,
        resume_from_checkpoint: Optional[str] = None,
        use_mixed_precision: bool = True,
        height=512,
        width=384,
    ):
        self.training = True
        self.models = models
        self.train_dataloader = train_dataloader
        self.val_dataloader = val_dataloader
        self.device = device
        self.learning_rate = learning_rate
        self.num_epochs = num_epochs
        self.save_steps = save_steps
        self.output_dir = Path(output_dir)
        self.cfg_dropout_prob = cfg_dropout_prob
        self.max_grad_norm = max_grad_norm
        self.use_peft = use_peft
        self.dream_lambda = dream_lambda
        self.use_mixed_precision = use_mixed_precision
        self.height = height
        self.width = width
        self.generator = torch.Generator(device=device)
        
        # Create output directory
        self.output_dir.mkdir(parents=True, exist_ok=True)

        # Setup mixed precision training
        if self.use_mixed_precision:
            self.scaler = torch.cuda.amp.GradScaler()

        self.weight_dtype = torch.float16 if use_mixed_precision else torch.float32
        
        # Initialize scheduler and sampler
        self.scheduler = DDPMSampler(self.generator, num_training_steps=1000)

        # Resume from checkpoint if provided
        self.global_step = 0
        self.current_epoch = 0
        
        # Setup models and optimizers
        self._setup_training()
        
        if resume_from_checkpoint:
            self._load_checkpoint(resume_from_checkpoint)
            
        
    
        self.encoder = self.models.get('encoder', None)
        self.decoder = self.models.get('decoder', None)
        self.diffusion = self.models.get('diffusion', None)
    
    def _setup_training(self):
        """Setup models for training with PEFT"""
        # Move models to device
        for name, model in self.models.items():
            model.to(self.device)
        
        # Freeze all parameters first
        for model in self.models.values():
            for param in model.parameters():
                param.requires_grad = False
        
        # Enable training for specific layers based on PEFT strategy
        if self.use_peft:
            self._enable_peft_training()
        else:
            # Enable full training for diffusion model
            for param in self.diffusion.parameters():
                param.requires_grad = True
        
        # Collect trainable parameters
        trainable_params = []
        total_params = 0
        trainable_count = 0
        
        for name, model in self.models.items():
            for param_name, param in model.named_parameters():
                total_params += param.numel()
                if param.requires_grad:
                    trainable_params.append(param)
                    trainable_count += param.numel()

        print(f"Total parameters: {total_params:,}")
        print(f"Trainable parameters: {trainable_count:,} ({trainable_count/total_params*100:.2f}%)")
        
        # Setup optimizer - AdamW as per paper
        self.optimizer = AdamW(
            trainable_params,
            lr=self.learning_rate,
            betas=(0.9, 0.999),
            weight_decay=1e-2,
            eps=1e-8
        )
        
        # Setup learning rate scheduler (constant)
        self.lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
            self.optimizer, lr_lambda=lambda epoch: 1.0
        )
    
    def _enable_peft_training(self):
        """Enable PEFT training - only self-attention layers"""
        print("Enabling PEFT training (self-attention layers only)")
        
        unet = self.models['diffusion'].unet
        
        # Enable attention layers in encoders and decoders
        for layers in [unet.encoders, unet.decoders]:
            for layer in layers:
                for module_idx, module in enumerate(layer):
                    for name, param in module.named_parameters():
                        if 'attention_1' in name:
                            param.requires_grad = True
                        
        # Enable attention layers in bottleneck
        for layer in unet.bottleneck:
            for name, param in layer.named_parameters():
                if 'attention_1' in name:
                    param.requires_grad = True
    
    def compute_loss(self, batch: Dict[str, torch.Tensor]) -> torch.Tensor:
        """Compute MSE loss for denoising with DREAM strategy"""
        person_images = batch['person'].to(self.device)
        cloth_images = batch['cloth'].to(self.device)
        masks = batch['mask'].to(self.device)
        
        batch_size = person_images.shape[0]

        concat_dim = -2  # y axis concat
        
        # Prepare inputs
        image, condition_image, mask = check_inputs(person_images, cloth_images, masks, self.width, self.height)
        image = prepare_image(person_images).to(self.device, dtype=self.weight_dtype)
        condition_image = prepare_image(cloth_images).to(self.device, dtype=self.weight_dtype)
        mask = prepare_mask_image(masks).to(self.device, dtype=self.weight_dtype)
        
        # Mask image
        masked_image = image * (mask < 0.5)

        with torch.cuda.amp.autocast(enabled=self.use_mixed_precision):
            # VAE encoding
            masked_latent = compute_vae_encodings(masked_image, self.encoder)
            person_latent = compute_vae_encodings(person_images, self.encoder)
            condition_latent = compute_vae_encodings(condition_image, self.encoder)
            mask_latent = torch.nn.functional.interpolate(mask, size=masked_latent.shape[-2:], mode="nearest")
            
            
            del image, mask, condition_image
            
            # Apply CFG dropout to garment latent (10% chance)
            if self.training and random.random() < self.cfg_dropout_prob:
                condition_latent = torch.zeros_like(condition_latent)
            
            # Concatenate latents
            input_latents = torch.cat([masked_latent, condition_latent], dim=concat_dim)
            mask_input = torch.cat([mask_latent, torch.zeros_like(mask_latent)], dim=concat_dim)
            target_latents = torch.cat([person_latent, condition_latent], dim=concat_dim)

            noise = randn_tensor(
                target_latents.shape,
                generator=self.generator,
                device=target_latents.device,
                dtype=self.weight_dtype,
            )

            # timesteps = torch.randint(1, 1000, size=(1,), device=self.device)[0].long().item()
            # timesteps = torch.tensor(timesteps, device=self.device)
            # timesteps_embedding = get_time_embedding(timesteps).to(self.device, dtype=self.weight_dtype)
            timesteps = torch.randint(1, 1000, size=(batch_size,))
            timesteps_embedding = get_time_embedding(timesteps).to(self.device, dtype=self.weight_dtype)

            # Add noise to latents
            noisy_latents = self.scheduler.add_noise(target_latents, timesteps, noise)

            # UNet(zt ⊙ Mi ⊙ Xi) where ⊙ is channel concatenation
            unet_input = torch.cat([
                input_latents,      # Xi
                mask_input,         # Mi
                noisy_latents,      # zt
            ], dim=1).to(self.device, dtype=self.weight_dtype)  # Channel dimension
            

            # DREAM strategy implementation
            if self.dream_lambda > 0:
                # Get initial noise prediction
                with torch.no_grad():
                    epsilon_theta = self.diffusion(
                        unet_input,
                        timesteps_embedding
                    )
                
                # DREAM noise combination: ε + λ*εθ
                dream_noise = noise + self.dream_lambda * epsilon_theta
                
                # Create new noisy latents with DREAM noise
                dream_noisy_latents = self.scheduler.add_noise(target_latents, timesteps, dream_noise)

                dream_unet_input = torch.cat([
                    input_latents, 
                    mask_input,
                    dream_noisy_latents
                ], dim=1).to(self.device, dtype=self.weight_dtype)

                predicted_noise = self.diffusion(
                    dream_unet_input,
                    timesteps_embedding
                )
                # DREAM loss: |(ε + λεθ) - εθ(ẑt, t)|²
                loss = F.mse_loss(predicted_noise, dream_noise)
            else:
                # Standard training without DREAM
                predicted_noise = self.diffusion(
                    unet_input,
                    timesteps_embedding,
                )
                
                # Standard MSE loss
                loss = F.mse_loss(predicted_noise, noise)
            
            if self.global_step % 1000 == 0:
                save_debug_visualization(
                    person_images=person_images,
                    cloth_images=cloth_images, 
                    masks=masks,
                    masked_image=masked_image,
                    noisy_latents=noisy_latents,
                    predicted_noise=predicted_noise,
                    target_latents=target_latents,
                    decoder=self.decoder,
                    global_step=self.global_step,
                    output_dir=self.output_dir,
                    device=self.device
                )
        return loss
    
    def train_epoch(self) -> float:
        """Train for one epoch - simplified version"""
        self.diffusion.train()
        total_loss = 0.0
        num_batches = len(self.train_dataloader)
        
        # progress_bar = tqdm(self.train_dataloader, desc=f"Epoch {self.current_epoch+1}")
        
        for step, batch in enumerate(self.train_dataloader):
            # Zero gradients
            self.optimizer.zero_grad()
            
            # Forward pass with mixed precision
            if self.use_mixed_precision:
                with torch.cuda.amp.autocast():
                    loss = self.compute_loss(batch)
                
                # Backward pass with scaling
                self.scaler.scale(loss).backward()
                
                # Gradient clipping and optimizer step
                self.scaler.unscale_(self.optimizer)
                torch.nn.utils.clip_grad_norm_(
                    [p for p in self.diffusion.parameters() if p.requires_grad],
                    self.max_grad_norm
                )
                
                self.scaler.step(self.optimizer)
                self.scaler.update()
            else:
                loss = self.compute_loss(batch)
                loss.backward()
                
                # Gradient clipping
                torch.nn.utils.clip_grad_norm_(
                    [p for p in self.diffusion.parameters() if p.requires_grad],
                    self.max_grad_norm
                )
                
                # Optimizer step
                self.optimizer.step()
            
            # Update learning rate
            self.lr_scheduler.step()
            self.global_step += 1
            
            total_loss += loss.item()
            
            # Update progress bar
            # progress_bar.set_postfix({
            #     'loss': loss.item(),
            #     'lr': self.optimizer.param_groups[0]['lr'],
            #     'step': self.global_step
            # })
            
            # Save checkpoint based on steps
            if self.global_step % self.save_steps == 0:
                self._save_checkpoint()
            
            # Clear cache periodically to prevent OOM
            if step % 50 == 0:
                torch.cuda.empty_cache()
        
        return total_loss / num_batches
    
    def train(self):
        """Main training loop - simplified version"""
        print(f"Starting training for {self.num_epochs} epochs")
        print(f"Total training batches per epoch: {len(self.train_dataloader)}")
        print(f"Using DREAM with lambda = {self.dream_lambda}")
        print(f"Mixed precision: {self.use_mixed_precision}")
        
        for epoch in range(self.current_epoch, self.num_epochs):
            self.current_epoch = epoch
            
            # Train one epoch
            train_loss = self.train_epoch()
            
            print(f"Epoch {epoch+1}/{self.num_epochs} - Train Loss: {train_loss:.6f}")
            
            # Save epoch checkpoint
            if (epoch + 1) % 5 == 0:  # Save every 5 epochs
                self._save_checkpoint(epoch_checkpoint=True)
            
            # Clear cache at end of epoch
            torch.cuda.empty_cache()
        
        # Save final checkpoint
        self._save_checkpoint(is_final=True)
        print("Training completed!")
    
    def _save_checkpoint(self, is_best: bool = False, epoch_checkpoint: bool = False, is_final: bool = False):
        """Save model checkpoint"""
        checkpoint = {
            'global_step': self.global_step,
            'current_epoch': self.current_epoch,
            'diffusion_state_dict': self.diffusion.state_dict(),
            'optimizer_state_dict': self.optimizer.state_dict(),
            'lr_scheduler_state_dict': self.lr_scheduler.state_dict(),
            'dream_lambda': self.dream_lambda,
            'use_peft': self.use_peft,
        }
        
        if self.use_mixed_precision:
            checkpoint['scaler_state_dict'] = self.scaler.state_dict()
        
        if is_final:
            checkpoint_path = self.output_dir / "final_model.pth"
        elif is_best:
            checkpoint_path = self.output_dir / "best_model.pth"
        elif epoch_checkpoint:
            checkpoint_path = self.output_dir / f"checkpoint_epoch_{self.current_epoch+1}.pth"
        else:
            checkpoint_path = self.output_dir / f"checkpoint_step_{self.global_step}.pth"
        
        torch.save(checkpoint, checkpoint_path)
        print(f"Checkpoint saved: {checkpoint_path}")
    
    def _load_checkpoint(self, checkpoint_path: str):
        """Load model checkpoint"""
        checkpoint = torch.load(checkpoint_path, map_location=self.device)
        
        self.global_step = checkpoint['global_step']
        self.current_epoch = checkpoint['current_epoch']
        self.dream_lambda = checkpoint.get('dream_lambda', 10.0)
        
        self.models['diffusion'].load_state_dict(checkpoint['diffusion_state_dict'])
        self.optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
        self.lr_scheduler.load_state_dict(checkpoint['lr_scheduler_state_dict'])
        
        if self.use_mixed_precision and 'scaler_state_dict' in checkpoint:
            self.scaler.load_state_dict(checkpoint['scaler_state_dict'])
        
        print(f"Checkpoint loaded: {checkpoint_path}")
        print(f"Resuming from epoch {self.current_epoch}, step {self.global_step}")


def create_dataloaders(args) -> DataLoader:
    """Create training dataloader"""
    if args.dataset_name == "vitonhd":
        dataset = VITONHDTestDataset(args)
    else:
        raise ValueError(f"Invalid dataset name {args.dataset_name}.")
    
    print(f"Dataset {args.dataset_name} loaded, total {len(dataset)} pairs.")
    
    dataloader = DataLoader(
        dataset,
        batch_size=args.batch_size,
        shuffle=True,
        num_workers=8,
        pin_memory=True,
        persistent_workers=True,
        prefetch_factor=2
    )
    return dataloader

def main():
    args = argparse.Namespace()
    args.__dict__ = {
        "base_model_path": "sd-v1-5-inpainting.ckpt",
        "dataset_name": "vitonhd",
        "data_root_path": "./viton-hd-dataset",
        "output_dir": "./checkpoints",
        "resume_from_checkpoint": "./checkpoints/checkpoint_step_50000.pth",
        "seed": 42,
        "batch_size": 2,
        "width": 384,
        "height": 384,
        "repaint": True,
        "eval_pair": True,
        "concat_eval_results": True,
        "concat_axis": 'y',
        "device": "cuda",
        "num_epochs": 50,  
        "learning_rate": 1e-5,
        "max_grad_norm": 1.0,
        "cfg_dropout_prob": 0.1,
        "dream_lambda": 10.0,
        "use_peft": True,
        "use_mixed_precision": True,
        "save_steps": 10000,
        "is_train": True
    }
    
    # Set random seeds
    torch.manual_seed(args.seed)
    np.random.seed(args.seed)
    random.seed(args.seed)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(args.seed)
    
    # Optimize CUDA settings
    torch.backends.cudnn.benchmark = True
    torch.backends.cuda.matmul.allow_tf32 = True 
    torch.backends.cudnn.allow_tf32 = True 
    torch.set_float32_matmul_precision("high")
    
    print("-"*100)

    # Load pretrained models
    print("Loading pretrained models...")
    models = preload_models_from_standard_weights(args.base_model_path, args.device)
    print("Models loaded successfully.")
    
    print("-"*100)
    
    # Create dataloader
    print("Creating dataloader...")
    train_dataloader = create_dataloaders(args)
    
    print(f"Training for {args.num_epochs} epochs")
    print(f"Batches per epoch: {len(train_dataloader)}")
    
    print("-"*100)
    
    # Initialize trainer
    print("Initializing trainer...")    
    trainer = CatVTONTrainer(
        models=models,
        train_dataloader=train_dataloader,
        device=args.device,
        learning_rate=args.learning_rate,
        num_epochs=args.num_epochs,
        save_steps=args.save_steps,
        output_dir=args.output_dir,
        cfg_dropout_prob=args.cfg_dropout_prob,
        max_grad_norm=args.max_grad_norm,
        use_peft=args.use_peft,
        dream_lambda=args.dream_lambda,
        resume_from_checkpoint=args.resume_from_checkpoint,
        use_mixed_precision=args.use_mixed_precision,
        height=args.height,
        width=args.width
    )
    
    # Start training
    print("Starting training...")
    trainer.train() 

if __name__ == "__main__":
    main()