# train.py import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.data import Dataset, DataLoader from PIL import Image import numpy as np import cv2 import os import argparse import logging from datetime import datetime # Import project modules import config from models.rord import RoRD from utils.data_utils import get_transform # Setup logging def setup_logging(save_dir): """Setup training logging""" if not os.path.exists(save_dir): os.makedirs(save_dir) log_file = os.path.join(save_dir, f'training_{datetime.now().strftime("%Y%m%d_%H%M%S")}.log') logging.basicConfig( level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s', handlers=[ logging.FileHandler(log_file), logging.StreamHandler() ] ) return logging.getLogger(__name__) # --- (Modified) Training-specific dataset class --- class ICLayoutTrainingDataset(Dataset): def __init__(self, image_dir, patch_size=256, transform=None, scale_range=(1.0, 1.0)): self.image_dir = image_dir self.image_paths = [os.path.join(image_dir, f) for f in os.listdir(image_dir) if f.endswith('.png')] self.patch_size = patch_size self.transform = transform self.scale_range = scale_range # New scale range parameter def __len__(self): return len(self.image_paths) def __getitem__(self, index): img_path = self.image_paths[index] image = Image.open(img_path).convert('L') W, H = image.size # --- New: Scale jittering data augmentation --- # 1. Randomly select a scaling factor scale = np.random.uniform(self.scale_range[0], self.scale_range[1]) # 2. Calculate crop size from original image based on scaling factor crop_size = int(self.patch_size / scale) # 确保裁剪尺寸不超过图像边界 if crop_size > min(W, H): crop_size = min(W, H) # 3. Random cropping x = np.random.randint(0, W - crop_size + 1) y = np.random.randint(0, H - crop_size + 1) patch = image.crop((x, y, x + crop_size, y + crop_size)) # 4. Resize cropped patch back to standard patch_size patch = patch.resize((self.patch_size, self.patch_size), Image.Resampling.LANCZOS) # --- Scale jittering end --- # --- New: Additional data augmentation --- # Brightness adjustment if np.random.random() < 0.5: brightness_factor = np.random.uniform(0.8, 1.2) patch = patch.point(lambda x: int(x * brightness_factor)) # Contrast adjustment if np.random.random() < 0.5: contrast_factor = np.random.uniform(0.8, 1.2) patch = patch.point(lambda x: int(((x - 128) * contrast_factor) + 128)) # Add noise if np.random.random() < 0.3: patch_np = np.array(patch, dtype=np.float32) noise = np.random.normal(0, 5, patch_np.shape) patch_np = np.clip(patch_np + noise, 0, 255) patch = Image.fromarray(patch_np.astype(np.uint8)) # --- Additional data augmentation end --- patch_np = np.array(patch) # Implement 8-direction discrete geometric transformations (this logic remains unchanged) theta_deg = np.random.choice([0, 90, 180, 270]) is_mirrored = np.random.choice([True, False]) cx, cy = self.patch_size / 2.0, self.patch_size / 2.0 M = cv2.getRotationMatrix2D((cx, cy), theta_deg, 1) if is_mirrored: T1 = np.array([[1, 0, -cx], [0, 1, -cy], [0, 0, 1]]) Flip = np.array([[-1, 0, 0], [0, 1, 0], [0, 0, 1]]) T2 = np.array([[1, 0, cx], [0, 1, cy], [0, 0, 1]]) M_mirror_3x3 = T2 @ Flip @ T1 M_3x3 = np.vstack([M, [0, 0, 1]]) H = (M_3x3 @ M_mirror_3x3).astype(np.float32) else: H = np.vstack([M, [0, 0, 1]]).astype(np.float32) transformed_patch_np = cv2.warpPerspective(patch_np, H, (self.patch_size, self.patch_size)) transformed_patch = Image.fromarray(transformed_patch_np) if self.transform: patch = self.transform(patch) transformed_patch = self.transform(transformed_patch) H_tensor = torch.from_numpy(H[:2, :]).float() return patch, transformed_patch, H_tensor # --- (Modified) Feature map transformation and loss functions (improved version) --- def warp_feature_map(feature_map, H_inv): B, C, H, W = feature_map.size() grid = F.affine_grid(H_inv, feature_map.size(), align_corners=False).to(feature_map.device) return F.grid_sample(feature_map, grid, align_corners=False) def compute_detection_loss(det_original, det_rotated, H): """Improved detection loss: use BCE loss instead of MSE""" with torch.no_grad(): H_inv = torch.inverse(torch.cat([H, torch.tensor([0.0, 0.0, 1.0]).view(1, 1, 3).repeat(H.shape[0], 1, 1)], dim=1))[:, :2, :] warped_det_rotated = warp_feature_map(det_rotated, H_inv) # Use BCE loss, more suitable for binary classification problems bce_loss = F.binary_cross_entropy(det_original, warped_det_rotated) # Add smooth L1 loss as auxiliary smooth_l1_loss = F.smooth_l1_loss(det_original, warped_det_rotated) return bce_loss + 0.1 * smooth_l1_loss def compute_description_loss(desc_original, desc_rotated, H, margin=1.0): """IC layout-specific geometric-aware descriptor loss: encodes Manhattan geometric features""" B, C, H_feat, W_feat = desc_original.size() # Manhattan geometric-aware sampling: focus on edge and corner regions num_samples = 200 # Generate Manhattan-aligned sampling grid (horizontal and vertical priority) h_coords = torch.linspace(-1, 1, int(np.sqrt(num_samples)), device=desc_original.device) w_coords = torch.linspace(-1, 1, int(np.sqrt(num_samples)), device=desc_original.device) # Increase sampling density in Manhattan directions manhattan_h = torch.cat([h_coords, torch.zeros_like(h_coords)]) manhattan_w = torch.cat([torch.zeros_like(w_coords), w_coords]) manhattan_coords = torch.stack([manhattan_h, manhattan_w], dim=1).unsqueeze(0).repeat(B, 1, 1) # Sample anchor points anchor = F.grid_sample(desc_original, manhattan_coords.unsqueeze(1), align_corners=False).squeeze(2).transpose(1, 2) # Calculate corresponding positive samples coords_hom = torch.cat([manhattan_coords, torch.ones(B, manhattan_coords.size(1), 1, device=manhattan_coords.device)], dim=2) M_inv = torch.inverse(torch.cat([H, torch.tensor([0.0, 0.0, 1.0]).view(1, 1, 3).repeat(H.shape[0], 1, 1)], dim=1)) coords_transformed = (coords_hom @ M_inv.transpose(1, 2))[:, :, :2] positive = F.grid_sample(desc_rotated, coords_transformed.unsqueeze(1), align_corners=False).squeeze(2).transpose(1, 2) # IC layout-specific negative sample strategy: consider repetitive structures with torch.no_grad(): # 1. Geometric-aware negative samples: different regions after Manhattan transformation neg_coords = [] for b in range(B): # Generate coordinates after Manhattan transformation (90-degree rotation, etc.) angles = [0, 90, 180, 270] for angle in angles: if angle != 0: theta = torch.tensor([angle * np.pi / 180]) rot_matrix = torch.tensor([ [torch.cos(theta), -torch.sin(theta), 0], [torch.sin(theta), torch.cos(theta), 0] ]) rotated_coords = manhattan_coords[b] @ rot_matrix[:2, :2].T neg_coords.append(rotated_coords) neg_coords = torch.stack(neg_coords[:B*num_samples//2]).reshape(B, -1, 2) # 2. Feature space hard negative samples negative_candidates = F.grid_sample(desc_rotated, neg_coords, align_corners=False).squeeze(2).transpose(1, 2) # 3. Manhattan distance constrained hard sample selection anchor_expanded = anchor.unsqueeze(2).expand(-1, -1, negative_candidates.size(1), -1) negative_expanded = negative_candidates.unsqueeze(1).expand(-1, anchor.size(1), -1, -1) # Use Manhattan distance instead of Euclidean distance manhattan_dist = torch.sum(torch.abs(anchor_expanded - negative_expanded), dim=3) hard_indices = torch.topk(manhattan_dist, k=anchor.size(1)//2, largest=False)[1] negative = torch.gather(negative_candidates, 1, hard_indices) # IC layout-specific geometric consistency loss # 1. Manhattan direction consistency loss manhattan_loss = 0 for i in range(anchor.size(1)): # Calculate geometric consistency in horizontal and vertical directions anchor_norm = F.normalize(anchor[:, i], p=2, dim=1) positive_norm = F.normalize(positive[:, i], p=2, dim=1) # Encourage descriptor invariance to Manhattan transformations cos_sim = torch.sum(anchor_norm * positive_norm, dim=1) manhattan_loss += torch.mean(1 - cos_sim) # 2. Sparsity regularization (IC layout features are sparse) sparsity_loss = torch.mean(torch.abs(anchor)) + torch.mean(torch.abs(positive)) # 3. Binary feature distance (handles binary input) binary_loss = torch.mean(torch.abs(torch.sign(anchor) - torch.sign(positive))) # Combined loss triplet_loss = nn.TripletMarginLoss(margin=margin, p=1, reduction='mean') # Use L1 distance geometric_triplet = triplet_loss(anchor, positive, negative) return geometric_triplet + 0.1 * manhattan_loss + 0.01 * sparsity_loss + 0.05 * binary_loss # --- (Modified) Main function and command-line interface --- def main(args): # Setup logging logger = setup_logging(args.save_dir) logger.info("--- Starting RoRD model training ---") logger.info(f"Training parameters: Epochs={args.epochs}, Batch Size={args.batch_size}, LR={args.lr}") logger.info(f"Data directory: {args.data_dir}") logger.info(f"Save directory: {args.save_dir}") transform = get_transform() # Pass scale jittering range during dataset initialization dataset = ICLayoutTrainingDataset( args.data_dir, patch_size=config.PATCH_SIZE, transform=transform, scale_range=config.SCALE_JITTER_RANGE ) logger.info(f"Dataset size: {len(dataset)}") # Split training and validation sets train_size = int(0.8 * len(dataset)) val_size = len(dataset) - train_size train_dataset, val_dataset = torch.utils.data.random_split(dataset, [train_size, val_size]) logger.info(f"Training set size: {len(train_dataset)}, Validation set size: {len(val_dataset)}") train_dataloader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=4) val_dataloader = DataLoader(val_dataset, batch_size=args.batch_size, shuffle=False, num_workers=4) model = RoRD().cuda() logger.info(f"Model parameter count: {sum(p.numel() for p in model.parameters()):,}") optimizer = torch.optim.Adam(model.parameters(), lr=args.lr, weight_decay=1e-4) # Add learning rate scheduler scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode='min', factor=0.5, patience=5 ) # Early stopping mechanism best_val_loss = float('inf') patience_counter = 0 patience = 10 for epoch in range(args.epochs): # Training phase model.train() total_train_loss = 0 total_det_loss = 0 total_desc_loss = 0 for i, (original, rotated, H) in enumerate(train_dataloader): original, rotated, H = original.cuda(), rotated.cuda(), H.cuda() det_original, desc_original = model(original) det_rotated, desc_rotated = model(rotated) det_loss = compute_detection_loss(det_original, det_rotated, H) desc_loss = compute_description_loss(desc_original, desc_rotated, H) loss = det_loss + desc_loss optimizer.zero_grad() loss.backward() # Gradient clipping to prevent gradient explosion torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0) optimizer.step() total_train_loss += loss.item() total_det_loss += det_loss.item() total_desc_loss += desc_loss.item() if i % 10 == 0: logger.info(f"Epoch {epoch+1}, Batch {i}, Total Loss: {loss.item():.4f}, " f"Det Loss: {det_loss.item():.4f}, Desc Loss: {desc_loss.item():.4f}") avg_train_loss = total_train_loss / len(train_dataloader) avg_det_loss = total_det_loss / len(train_dataloader) avg_desc_loss = total_desc_loss / len(train_dataloader) # Validation phase model.eval() total_val_loss = 0 total_val_det_loss = 0 total_val_desc_loss = 0 with torch.no_grad(): for original, rotated, H in val_dataloader: original, rotated, H = original.cuda(), rotated.cuda(), H.cuda() det_original, desc_original = model(original) det_rotated, desc_rotated = model(rotated) val_det_loss = compute_detection_loss(det_original, det_rotated, H) val_desc_loss = compute_description_loss(desc_original, desc_rotated, H) val_loss = val_det_loss + val_desc_loss total_val_loss += val_loss.item() total_val_det_loss += val_det_loss.item() total_val_desc_loss += val_desc_loss.item() avg_val_loss = total_val_loss / len(val_dataloader) avg_val_det_loss = total_val_det_loss / len(val_dataloader) avg_val_desc_loss = total_val_desc_loss / len(val_dataloader) # Learning rate scheduling scheduler.step(avg_val_loss) logger.info(f"--- Epoch {epoch+1} completed ---") logger.info(f"Training - Total: {avg_train_loss:.4f}, Det: {avg_det_loss:.4f}, Desc: {avg_desc_loss:.4f}") logger.info(f"Validation - Total: {avg_val_loss:.4f}, Det: {avg_val_det_loss:.4f}, Desc: {avg_val_desc_loss:.4f}") logger.info(f"Learning rate: {optimizer.param_groups[0]['lr']:.2e}") # Early stopping check if avg_val_loss < best_val_loss: best_val_loss = avg_val_loss patience_counter = 0 # Save best model if not os.path.exists(args.save_dir): os.makedirs(args.save_dir) save_path = os.path.join(args.save_dir, 'rord_model_best.pth') torch.save({ 'epoch': epoch, 'model_state_dict': model.state_dict(), 'optimizer_state_dict': optimizer.state_dict(), 'best_val_loss': best_val_loss, 'config': { 'learning_rate': args.lr, 'batch_size': args.batch_size, 'epochs': args.epochs } }, save_path) logger.info(f"Best model saved to: {save_path}") else: patience_counter += 1 if patience_counter >= patience: logger.info(f"Early stopping triggered! No improvement for {patience} epochs") break # Save final model save_path = os.path.join(args.save_dir, 'rord_model_final.pth') torch.save({ 'epoch': args.epochs, 'model_state_dict': model.state_dict(), 'optimizer_state_dict': optimizer.state_dict(), 'final_val_loss': avg_val_loss, 'config': { 'learning_rate': args.lr, 'batch_size': args.batch_size, 'epochs': args.epochs } }, save_path) logger.info(f"Final model saved to: {save_path}") logger.info("Training completed!") if __name__ == "__main__": parser = argparse.ArgumentParser(description="Train RoRD model") parser.add_argument('--data_dir', type=str, default=config.LAYOUT_DIR) parser.add_argument('--save_dir', type=str, default=config.SAVE_DIR) parser.add_argument('--epochs', type=int, default=config.NUM_EPOCHS) parser.add_argument('--batch_size', type=int, default=config.BATCH_SIZE) parser.add_argument('--lr', type=float, default=config.LEARNING_RATE) main(parser.parse_args())