File size: 18,932 Bytes
c8c00f0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import os
import json
import torch
import argparse
import torch.nn.functional as F
import random
import time
from torch.optim import AdamW
from tqdm import tqdm
from diffusers import DDPMScheduler
from model import DefectFillModel, USE_MODELSCOPE
from data_loader import get_data_loaders
from utils import save_checkpoint, load_checkpoint
# TensorBoard support
from torch.utils.tensorboard import SummaryWriter
import datetime


def generate_seed_from_timestamp():
    """Generates a random seed based on the current timestamp"""
    return int(time.time() * 1000) % (2**31)

def train(args):
    # Set up device
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    print(f"Using device: {device}")
    if torch.cuda.is_available():
        print(f"GPU Name: {torch.cuda.get_device_name(0)}")
    else:
        print("CUDA is NOT available. The model will run on CPU, causing severe slowdown.")
    
    # Create output directory structure
    os.makedirs(args.output_dir, exist_ok=True)
    checkpoints_dir = os.path.join(args.output_dir, "checkpoints")
    tensorboard_dir = os.path.join(args.output_dir, "tensorboard")
    os.makedirs(checkpoints_dir, exist_ok=True)
    os.makedirs(tensorboard_dir, exist_ok=True)
    
    # Create log file in the output directory
    log_file_path = os.path.join(args.output_dir, "train_log.txt")
    log_file = open(log_file_path, "a")
    log_file.write(f"\n\n{'='*60}\n")
    log_file.write(f"Training started at {datetime.datetime.now().strftime('%Y-%m-%d %H:%M:%S')}\n")
    log_file.write(f"Object class: {args.object_class}\n")
    log_file.write(f"Defect type: {args.defect_type if args.defect_type else 'all'}\n")
    log_file.write(f"Config name: {args.config_name}\n")
    log_file.write(f"Lambda defect: {args.lambda_defect}\n")
    log_file.write(f"Lambda obj: {args.lambda_obj}\n")
    log_file.write(f"Lambda attn: {args.lambda_attn}\n")
    log_file.write(f"Alpha (obj branch bg weight): {args.alpha}\n")
    log_file.write(f"Gradient accumulation steps: {args.gradient_accumulation_steps}\n")
    log_file.write(f"Random seed: {args.seed}\n")
    log_file.write(f"{'='*60}\n\n")
    
    # Save training configuration to JSON
    train_config = {
        "timestamp": datetime.datetime.now().strftime('%Y-%m-%dT%H:%M:%S'),
        "object_class": args.object_class,
        "defect_type": args.defect_type if args.defect_type else "all",
        "config_name": args.config_name,
        "lambda_defect": args.lambda_defect,
        "lambda_obj": args.lambda_obj,
        "lambda_attn": args.lambda_attn,
        "alpha": args.alpha,
        "batch_size": args.batch_size,
        "max_train_steps": args.max_train_steps,
        "gradient_accumulation_steps": args.gradient_accumulation_steps,
        "lora_rank": args.lora_rank,
        "lora_alpha": args.lora_alpha,
        "text_encoder_lr": args.text_encoder_lr,
        "unet_lr": args.unet_lr,
        "lr_warmup_steps": args.lr_warmup_steps,
        "save_steps": args.save_steps,
        "seed": args.seed
    }
    config_path = os.path.join(args.output_dir, "train_config.json")
    with open(config_path, "w") as f:
        json.dump(train_config, f, indent=4)
    print(f"Training config saved to: {config_path}")
    
    # Initialize TensorBoard writer
    writer = SummaryWriter(tensorboard_dir)
    
    # Load data
    train_loader, test_loader = get_data_loaders(
        root_dir=args.data_dir,
        object_class=args.object_class,
        batch_size=args.batch_size,
        defect_type=args.defect_type,
        dilate_mask=args.dilate_mask,
        mask_kernel_size=args.mask_kernel_size
    )
    
    # Initialize model
    model = DefectFillModel(
        device=device,
        lora_rank=args.lora_rank,
        lora_alpha=args.lora_alpha,
        seed=args.seed
    )
    
    # Set up optimizers with specific learning rates for Text Encoder and UNet
    text_encoder_params = [p for n, p in model.pipeline.text_encoder.named_parameters() if "lora" in n]
    unet_params = [p for n, p in model.pipeline.unet.named_parameters() if "lora" in n]
    
    optimizer = AdamW([
        {"params": text_encoder_params, "lr": args.text_encoder_lr},
        {"params": unet_params, "lr": args.unet_lr}
    ])
    
    # Save original LRs for warmup calculation
    base_lrs = [args.text_encoder_lr, args.unet_lr]
    
    # Set up noise scheduler (handling ModelScope vs HuggingFace)
    hf_model_id = "sd2-community/stable-diffusion-2-inpainting"
    if USE_MODELSCOPE:
        try:
            from modelscope import snapshot_download
            print(f"[ModelScope] Downloading scheduler: {hf_model_id}")
            local_model_path = snapshot_download(hf_model_id)
            noise_scheduler = DDPMScheduler.from_pretrained(local_model_path, subfolder="scheduler")
        except ImportError:
            print("[Warning] modelscope not installed, falling back to HuggingFace...")
            noise_scheduler = DDPMScheduler.from_pretrained(hf_model_id, subfolder="scheduler")
    else:
        noise_scheduler = DDPMScheduler.from_pretrained(hf_model_id, subfolder="scheduler")
    
    # Resume from checkpoint if specified
    start_step = 0
    if args.resume_from:
        start_step = load_checkpoint(model, optimizer, args.resume_from)
        print(f"Resuming from step {start_step}")
        log_file.write(f"Resuming from step {start_step}\n")
    
    # Set models to training mode
    model.pipeline.unet.train()
    model.pipeline.text_encoder.train()
    
    total_steps = args.max_train_steps
    progress_bar = tqdm(range(start_step, total_steps), desc="Training Progress")
    
    global_step = start_step
    accumulation_step = 0

    # Add this BEFORE the while loop
    with torch.no_grad():
        clean_latents_cache = {}
        for batch in train_loader:
            images = batch["image"].to(device, dtype=torch.float16)
            is_defect = batch["is_defect"]
            defect_samples = torch.nonzero(is_defect).squeeze(1)
            if len(defect_samples) > 0:
                defect_images = images[defect_samples]
                # Cache original latents
                latents = model.pipeline.vae.encode(defect_images).latent_dist.sample() * model.pipeline.vae.config.scaling_factor
                clean_latents_cache[tuple(defect_samples.cpu().tolist())] = latents.detach()
    
    while global_step < total_steps:
        for batch in train_loader:
            if global_step >= total_steps:
                break

            # t_start = time.time()

            # Move data to device
            images = batch["image"].to(device, dtype=torch.float16)
            masks = batch["mask"].to(device, dtype=torch.float16)
            backgrounds = batch["background"].to(device, dtype=torch.float16)
            adjusted_masks = batch["adjusted_mask"].to(device, dtype=torch.float16)
            is_defect = batch["is_defect"]
            
            # Ensure we only process defective samples
            defect_samples = torch.nonzero(is_defect).squeeze(1)
            if len(defect_samples) == 0:
                continue  # Skip batch if no defects present
                
            # Extract defect-only samples
            defect_images = images[defect_samples]
            defect_masks = masks[defect_samples]
            defect_backgrounds = backgrounds[defect_samples]
            defect_adjusted_masks = adjusted_masks[defect_samples]
            object_classes = [batch["object_class"][i] for i in defect_samples]

            # t_data = time.time()
            
            # Extract defect type from file path for specific prompting
            defect_types = []
            for i in defect_samples:
                if hasattr(train_loader.dataset, 'images') and i < len(train_loader.dataset.images):
                    img_path = train_loader.dataset.images[i]
                    parts = img_path.split(os.sep)
                    for j, part in enumerate(parts):
                        if part == "defective" and j + 1 < len(parts):
                            defect_types.append(parts[j + 1])
                            break
                    else:
                        defect_types.append("defect")
                else:
                    defect_types.append("defect")
            
            # Learning rate warmup
            if global_step < args.lr_warmup_steps:
                lr_scale = min(1.0, (global_step + 1) / args.lr_warmup_steps)
                for i, param_group in enumerate(optimizer.param_groups):
                    param_group["lr"] = base_lrs[i] * lr_scale
            
            # Reset attention maps
            if hasattr(model, 'attention_maps'):
                model.attention_maps = {}
            
            # ========== PHASE 1: Defect Branch (Defect Texture Learning) ==========
            # Using <defect> learnable token for concept isolation
            # t_text = time.time()
            defect_prompts = [f"A photo of {model.placeholder_token}" for _ in range(len(defect_samples))]
            text_embeddings = model.get_text_embeddings(defect_prompts, enable_grad=True)
            
            # ========== PHASE 1 & 2 OPTIMIZED VAE & MASK PASS ==========
            
            # 1. Single VAE encoding for original defect images
            with torch.no_grad():
                # Instead of re-encoding, load from cache
                latents = clean_latents_cache[tuple(defect_samples.cpu().tolist())]

            # t_vae = time.time()
            
            if len(defect_masks.shape) == 3:
                defect_masks = defect_masks.unsqueeze(1)

            # 2. Fully Vectorized Random Mask Generation directly on GPU (0 Python Loops)
            B, _, H, W = defect_images.shape
            K = 15  # Number of random boxes

            # Create broadcastable coordinate grids (1, 1, H, 1) and (1, 1, 1, W)
            grid_y = torch.arange(H, device=device).view(1, 1, H, 1)
            grid_x = torch.arange(W, device=device).view(1, 1, 1, W)

            # Sample all box parameters for all batch items and box counts simultaneously: shape (B, K, 1, 1)
            rh = torch.randint(max(1, int(H * 0.03)), max(2, int(H * 0.25)), (B, K, 1, 1), device=device)
            rw = torch.randint(max(1, int(W * 0.03)), max(2, int(W * 0.25)), (B, K, 1, 1), device=device)
            ry = torch.randint(0, max(1, H - 20), (B, K, 1, 1), device=device)
            rx = torch.randint(0, max(1, W - 20), (B, K, 1, 1), device=device)

            # Compute rectangle coverage across all (B, K, H, W) elements in a single broadcasted operation
            in_box = (grid_y >= ry) & (grid_y < (ry + rh)) & (grid_x >= rx) & (grid_x < (rx + rw))

            # Merge all K boxes per image using logical OR (.any), shape -> (B, 1, H, W)
            random_masks = in_box.any(dim=1, keepdim=True).to(dtype=defect_images.dtype)

            # 3. Batched VAE Encoding (Combines Phase 1 and Phase 2 into 1 pass)
            with torch.no_grad():
                masked_images = defect_images * (1 - defect_masks)
                rand_masked_images = defect_images * (1 - random_masks)
                
                combined_masked = torch.cat([masked_images, rand_masked_images], dim=0)
                combined_latents = model.pipeline.vae.encode(combined_masked).latent_dist.sample() * model.pipeline.vae.config.scaling_factor
                
                masked_image_latents, random_masked_image_latents = torch.chunk(combined_latents, 2, dim=0)

            # 4. Latent interpolation for masks
            mask_latents = F.interpolate(defect_masks, size=(latents.shape[2], latents.shape[3]))
            random_mask_latents = F.interpolate(random_masks, size=(latents.shape[2], latents.shape[3]))

            # ========== Phase 1 Forward Pass ==========
            noise = torch.randn_like(latents)
            timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (latents.shape[0],), device=device)
            noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)

            outputs = model(
                noisy_latents=noisy_latents,
                masked_image_latents=masked_image_latents,
                mask_latents=mask_latents,
                timesteps=timesteps,
                encoder_hidden_states=text_embeddings
            )

            # t_forward = time.time()
            
            # print(f"[TIMING RESULTS]")
            # print(f" β”œβ”€ Data Fetch & Filter : {t_data - t_start:.2f}s")
            # print(f" β”œβ”€ Text Encoding       : {t_text - t_data:.2f}s")
            # print(f" β”œβ”€ VAE Encoding        : {t_vae - t_text:.2f}s")
            # print(f" └─ UNet Forward & Loss : {t_forward - t_vae:.2f}s")

            
            noise_pred = outputs["noise_pred"]
            defect_loss = model.compute_defect_loss(noise_pred, noise, mask_latents)
            attention_loss = outputs.get("attention_loss", torch.tensor(0.0, device=device))

            # ========== Phase 2 Forward Pass ==========
            obj_prompts = [f"A {obj_class} with {model.placeholder_token}" for obj_class in object_classes]
            obj_text_embeddings = model.get_text_embeddings(obj_prompts, enable_grad=True)

            obj_noise = torch.randn_like(latents)
            obj_timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (latents.shape[0],), device=device)
            obj_noisy_latents = noise_scheduler.add_noise(latents, obj_noise, obj_timesteps)
            
            obj_outputs = model(
                noisy_latents=obj_noisy_latents,
                masked_image_latents=random_masked_image_latents,
                mask_latents=random_mask_latents,
                timesteps=obj_timesteps,
                encoder_hidden_states=obj_text_embeddings
            )
            
            object_loss = model.compute_object_loss(obj_outputs["noise_pred"], obj_noise, random_mask_latents, alpha=args.alpha)
            
            # Total Loss Calculation
            total_loss = args.lambda_defect * defect_loss + args.lambda_obj * object_loss + args.lambda_attn * attention_loss
            total_loss = total_loss / args.gradient_accumulation_steps
            
            # NaN Check
            if torch.isnan(total_loss):
                print(f"Warning: NaN loss detected at step {global_step}")
                log_file.write(f"Warning: NaN loss at step {global_step}\n")
                optimizer.zero_grad()
                continue

            # t_before_backward = time.time()
            total_loss.backward()
            # t_after_backward = time.time()
            # print(f" └─ Backward Pass Time : {t_after_backward - t_before_backward:.2f}s")
            accumulation_step += 1
            
            # Optimization step
            if accumulation_step >= args.gradient_accumulation_steps:
                optimizer.step()
                optimizer.zero_grad()
                accumulation_step = 0
                
                progress_bar.update(1)
                global_step += 1
                
                # Logging and TensorBoard updates
                writer.add_scalar("Loss/Defect", defect_loss.item(), global_step)
                writer.add_scalar("Loss/Object", object_loss.item(), global_step)
                writer.add_scalar("Loss/Attention", attention_loss.item(), global_step)
                writer.add_scalar("Loss/Total", total_loss.item() * args.gradient_accumulation_steps, global_step)
                
                if global_step % 10 == 0:
                    for i, param_group in enumerate(optimizer.param_groups):
                        writer.add_scalar(f"LearningRate/group{i}", param_group["lr"], global_step)
                
                # Periodic checkpointing
                if global_step % args.save_steps == 0 or global_step == total_steps:
                    checkpoint_path = os.path.join(checkpoints_dir, f"checkpoint_{global_step}.pt")
                    save_checkpoint(model, optimizer, global_step, checkpoint_path)
                    log_file.write(f"Checkpoint saved at step {global_step}\n")

    # Save final model
    final_checkpoint_path = os.path.join(checkpoints_dir, "checkpoint_final.pt")
    save_checkpoint(model, optimizer, global_step, final_checkpoint_path)
    print(f"Final model saved to: {final_checkpoint_path}")
    
    writer.close()
    log_file.close()
    return model

if __name__ == "__main__":
    parser = argparse.ArgumentParser(description="Train DefectFill model")
    
    # Paths
    parser.add_argument("--data_dir", type=str, required=True, help="Path to MVTec AD dataset")
    parser.add_argument("--object_class", type=str, required=True, help="Object class to train on")
    parser.add_argument("--output_dir", type=str, default="./output", help="Directory to save models")
    
    # Loss Weights
    parser.add_argument("--lambda_defect", type=float, default=0.5, help="Defect loss weight (L_def)")
    parser.add_argument("--lambda_obj", type=float, default=0.2, help="Object integrity loss weight (L_obj)")
    parser.add_argument("--lambda_attn", type=float, default=0.05, help="Attention loss weight (L_attn)")
    parser.add_argument("--alpha", type=float, default=0.3, help="Background weight for object branch")
    
    # Training Config
    parser.add_argument("--config_name", type=str, default="base", help="Experiment name (base/tex/obj)")
    parser.add_argument("--defect_type", type=str, default=None, help="Specific defect type to train on")
    parser.add_argument("--batch_size", type=int, default=2)
    parser.add_argument("--lora_rank", type=int, default=8)
    parser.add_argument("--lora_alpha", type=int, default=16)
    parser.add_argument("--text_encoder_lr", type=float, default=4e-5)
    parser.add_argument("--unet_lr", type=float, default=2e-4)
    parser.add_argument("--max_train_steps", type=int, default=2000)
    parser.add_argument("--lr_warmup_steps", type=int, default=100)
    parser.add_argument("--save_steps", type=int, default=500)
    parser.add_argument("--seed", type=int, default=-1, help="-1 for timestamp-based seed")
    parser.add_argument("--resume_from", type=str, default=None)
    parser.add_argument("--gradient_accumulation_steps", type=int, default=2)
    parser.add_argument("--dilate_mask", type=str, default="False", help="Whether to dilate masks (True/False)")
    parser.add_argument("--mask_kernel_size", type=int, default=3, help="Size of dilation kernel (must be odd, e.g. 3, 5, 7)")
    
    args = parser.parse_args()

    # Helper to convert string "True" to boolean
    args.dilate_mask = args.dilate_mask.lower() == "true"
    
    # Seed setup
    if args.seed == -1:
        args.seed = generate_seed_from_timestamp()
    
    torch.manual_seed(args.seed)
    random.seed(args.seed)
    
    train(args)