added result, 21.3905

image-inpainting/src/architecture.py

@@ -68,46 +68,6 @@ class CBAM(nn.Module):
         return x
 
 
-class MultiScaleFeatureExtraction(nn.Module):
-    """Multi-scale feature extraction using dilated convolutions"""
-    def __init__(self, channels):
-        super().__init__()
-        self.branch1 = nn.Sequential(
-            nn.Conv2d(channels, channels // 4, 1),
-            nn.BatchNorm2d(channels // 4),
-            nn.LeakyReLU(0.1, inplace=True)
-        )
-        self.branch2 = nn.Sequential(
-            nn.Conv2d(channels, channels // 4, 3, padding=2, dilation=2),
-            nn.BatchNorm2d(channels // 4),
-            nn.LeakyReLU(0.1, inplace=True)
-        )
-        self.branch3 = nn.Sequential(
-            nn.Conv2d(channels, channels // 4, 3, padding=4, dilation=4),
-            nn.BatchNorm2d(channels // 4),
-            nn.LeakyReLU(0.1, inplace=True)
-        )
-        self.branch4 = nn.Sequential(
-            nn.Conv2d(channels, channels // 4, 3, padding=8, dilation=8),
-            nn.BatchNorm2d(channels // 4),
-            nn.LeakyReLU(0.1, inplace=True)
-        )
-        self.fusion = nn.Sequential(
-            nn.Conv2d(channels, channels, 1),
-            nn.BatchNorm2d(channels),
-            nn.LeakyReLU(0.1, inplace=True)
-        )
-    
-    def forward(self, x):
-        b1 = self.branch1(x)
-        b2 = self.branch2(x)
-        b3 = self.branch3(x)
-        b4 = self.branch4(x)
-        out = torch.cat([b1, b2, b3, b4], dim=1)
-        out = self.fusion(out)
-        return out + x  # Residual connection
-
-
 class ConvBlock(nn.Module):
     """Convolutional block with Conv2d -> BatchNorm -> LeakyReLU"""
     def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dropout=0.0):
@@ -140,6 +100,26 @@ class ResidualConvBlock(nn.Module):
         return self.relu(out)
 
 
+class DilatedResidualBlock(nn.Module):
+    """Residual block with dilated convolutions for larger receptive field"""
+    def __init__(self, channels, dilation=2, dropout=0.0):
+        super().__init__()
+        self.conv1 = nn.Conv2d(channels, channels, 3, padding=dilation, dilation=dilation)
+        self.bn1 = nn.BatchNorm2d(channels)
+        self.conv2 = nn.Conv2d(channels, channels, 3, padding=dilation, dilation=dilation)
+        self.bn2 = nn.BatchNorm2d(channels)
+        self.relu = nn.LeakyReLU(0.1, inplace=True)
+        self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
+    
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.bn1(self.conv1(x)))
+        out = self.dropout(out)
+        out = self.bn2(self.conv2(out))
+        out = out + residual
+        return self.relu(out)
+
+
 class DownBlock(nn.Module):
     """Downsampling block with conv blocks, residual connection, attention, and max pooling"""
     def __init__(self, in_channels, out_channels, dropout=0.1):
@@ -198,12 +178,12 @@ class MyModel(nn.Module):
         self.down3 = DownBlock(base_channels * 4, base_channels * 8, dropout=dropout)
         self.down4 = DownBlock(base_channels * 8, base_channels * 16, dropout=dropout)
         
-        # Bottleneck with multiple residual blocks and multi-scale features
+        # Bottleneck with multi-scale dilated convolutions (ASPP-style)
         self.bottleneck = nn.Sequential(
             ConvBlock(base_channels * 16, base_channels * 16, dropout=dropout),
             ResidualConvBlock(base_channels * 16, dropout=dropout),
-            MultiScaleFeatureExtraction(base_channels * 16),
-            ResidualConvBlock(base_channels * 16, dropout=dropout),
+            DilatedResidualBlock(base_channels * 16, dilation=2, dropout=dropout),
+            DilatedResidualBlock(base_channels * 16, dilation=4, dropout=dropout),
             ResidualConvBlock(base_channels * 16, dropout=dropout),
             CBAM(base_channels * 16)
         )

image-inpainting/src/datasets.py

@@ -10,50 +10,11 @@ import numpy as np
 import random
 import glob
 import os
-from PIL import Image, ImageEnhance, ImageFilter
+from PIL import Image, ImageEnhance
 
 IMAGE_DIMENSION = 100
 
 
-class DataAugmentation:
-    """Data augmentation pipeline for improved generalization"""
-    
-    def __init__(self, p=0.5):
-        self.p = p
-    
-    def __call__(self, image: Image.Image) -> Image.Image:
-        # Random horizontal flip
-        if random.random() < self.p:
-            image = image.transpose(Image.FLIP_LEFT_RIGHT)
-        
-        # Random vertical flip
-        if random.random() < self.p * 0.5:
-            image = image.transpose(Image.FLIP_TOP_BOTTOM)
-        
-        # Random rotation (90 degree increments)
-        if random.random() < self.p * 0.3:
-            angle = random.choice([90, 180, 270])
-            image = image.rotate(angle)
-        
-        # Color jittering
-        if random.random() < self.p * 0.4:
-            # Brightness
-            enhancer = ImageEnhance.Brightness(image)
-            image = enhancer.enhance(random.uniform(0.85, 1.15))
-        
-        if random.random() < self.p * 0.4:
-            # Contrast
-            enhancer = ImageEnhance.Contrast(image)
-            image = enhancer.enhance(random.uniform(0.85, 1.15))
-        
-        if random.random() < self.p * 0.3:
-            # Saturation
-            enhancer = ImageEnhance.Color(image)
-            image = enhancer.enhance(random.uniform(0.85, 1.15))
-        
-        return image
-
-
 def create_arrays_from_image(image_array: np.ndarray, offset: tuple, spacing: tuple) -> tuple[np.ndarray, np.ndarray]:
     image_array = np.transpose(image_array, (2, 0, 1))
     known_array = np.zeros_like(image_array)
@@ -77,42 +38,74 @@ def preprocess(input_array: np.ndarray):
 
 class ImageDataset(torch.utils.data.Dataset):
     """
-    Dataset class for loading images from a folder with augmentation
+    Dataset class for loading images from a folder with data augmentation
     """
 
     def __init__(self, datafolder: str, augment: bool = True):
         self.imagefiles = sorted(glob.glob(os.path.join(datafolder,"**","*.jpg"),recursive=True))
         self.augment = augment
-        self.augmentation = DataAugmentation(p=0.5) if augment else None
 
     def __len__(self):
         return len(self.imagefiles)
         
+    def augment_image(self, image: Image) -> Image:
+        """Apply random augmentations to image"""
+        # Random horizontal flip
+        if random.random() > 0.5:
+            image = image.transpose(Image.FLIP_LEFT_RIGHT)
+        
+        # Random vertical flip
+        if random.random() > 0.5:
+            image = image.transpose(Image.FLIP_TOP_BOTTOM)
+        
+        # Random rotation (90, 180, 270 degrees)
+        if random.random() > 0.5:
+            angle = random.choice([90, 180, 270])
+            image = image.rotate(angle)
+        
+        # Random brightness adjustment
+        if random.random() > 0.5:
+            enhancer = ImageEnhance.Brightness(image)
+            factor = random.uniform(0.8, 1.2)
+            image = enhancer.enhance(factor)
+        
+        # Random contrast adjustment
+        if random.random() > 0.5:
+            enhancer = ImageEnhance.Contrast(image)
+            factor = random.uniform(0.8, 1.2)
+            image = enhancer.enhance(factor)
+        
+        # Random color adjustment
+        if random.random() > 0.5:
+            enhancer = ImageEnhance.Color(image)
+            factor = random.uniform(0.8, 1.2)
+            image = enhancer.enhance(factor)
+        
+        return image
+    
     def __getitem__(self, idx:int):
         index = int(idx)
         
-        image = Image.open(self.imagefiles[index]).convert('RGB')
-        
-        # Apply augmentation before resize
-        if self.augment and self.augmentation is not None:
-            image = self.augmentation(image)
-        
+        image = Image.open(self.imagefiles[index])
         image = resize(image)
+        
+        # Apply augmentation if enabled
+        if self.augment:
+            image = self.augment_image(image)
+        
         image = np.asarray(image)
         image = preprocess(image)
         
-        # More varied spacing for better generalization
-        spacing_x = random.randint(2, 8)
-        spacing_y = random.randint(2, 8)
-        offset_x = random.randint(0, min(spacing_x - 1, 8))
-        offset_y = random.randint(0, min(spacing_y - 1, 8))
+        # Vary spacing and offset more for additional diversity
+        spacing_x = random.randint(2,7)
+        spacing_y = random.randint(2,7)
+        offset_x = random.randint(0,10)
+        offset_y = random.randint(0,10)
         spacing = (spacing_x, spacing_y)
         offset = (offset_x, offset_y)
-        
         input_array, known_array = create_arrays_from_image(image.copy(), offset, spacing)
         target_image = torch.from_numpy(np.transpose(image, (2,0,1)))
         input_array = torch.from_numpy(input_array)
         known_array = torch.from_numpy(known_array)
         input_array = torch.cat((input_array, known_array), dim=0)
-        
         return input_array, target_image

image-inpainting/src/main.py

@@ -24,22 +24,22 @@ if __name__ == '__main__':
     config_dict['results_path'] = os.path.join(project_root, "results")
     config_dict['data_path'] = os.path.join(project_root, "data", "dataset")
     config_dict['device'] = None
-    config_dict['learningrate'] = 2e-4  # Slightly lower for stable training
-    config_dict['weight_decay'] = 5e-5  # Reduced weight decay
+    config_dict['learningrate'] = 2e-4  # Slightly lower for more stable training
+    config_dict['weight_decay'] = 5e-5  # Reduced for less aggressive regularization
     config_dict['n_updates'] = 8000  # More updates for better convergence
-    config_dict['batchsize'] = 8  # Smaller batch for better gradient estimates
+    config_dict['batchsize'] = 12  # Larger batch for more stable gradients
     config_dict['early_stopping_patience'] = 15  # More patience for complex model
     config_dict['use_wandb'] = False
 
     config_dict['print_train_stats_at'] = 10
     config_dict['print_stats_at'] = 100
-    config_dict['plot_at'] = 300
-    config_dict['validate_at'] = 300  # Validate more frequently
+    config_dict['plot_at'] = 400
+    config_dict['validate_at'] = 200  # Validate frequently but not too often
 
     network_config = {
         'n_in_channels': 4,
-        'base_channels': 56,  # Increased capacity for better feature learning
-        'dropout': 0.08  # Slightly less dropout with augmentation
+        'base_channels': 32,  # Smaller base for efficiency, depth compensates
+        'dropout': 0.15  # Slightly more regularization with augmentation
     }
     
     config_dict['network_config'] = network_config

image-inpainting/src/train.py

@@ -10,7 +10,6 @@ from utils import plot, evaluate_model
 
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
 import numpy as np
 import os
 
@@ -21,58 +20,15 @@ from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
 import wandb
 
 
-def gaussian_kernel(window_size=11, sigma=1.5):
-    """Create a Gaussian kernel for SSIM computation"""
-    x = torch.arange(window_size).float() - window_size // 2
-    gauss = torch.exp(-x.pow(2) / (2 * sigma ** 2))
-    kernel = gauss / gauss.sum()
-    kernel_2d = kernel.unsqueeze(1) * kernel.unsqueeze(0)
-    return kernel_2d.unsqueeze(0).unsqueeze(0)
-
-
-class SSIMLoss(nn.Module):
-    """Structural Similarity Index Loss for perceptual quality"""
-    def __init__(self, window_size=11, sigma=1.5):
-        super().__init__()
-        self.window_size = window_size
-        kernel = gaussian_kernel(window_size, sigma)
-        self.register_buffer('kernel', kernel)
-        self.C1 = 0.01 ** 2
-        self.C2 = 0.03 ** 2
-    
-    def forward(self, pred, target):
-        # Apply to each channel
-        channels = pred.shape[1]
-        kernel = self.kernel.repeat(channels, 1, 1, 1)
-        
-        mu_pred = F.conv2d(pred, kernel, padding=self.window_size // 2, groups=channels)
-        mu_target = F.conv2d(target, kernel, padding=self.window_size // 2, groups=channels)
-        
-        mu_pred_sq = mu_pred.pow(2)
-        mu_target_sq = mu_target.pow(2)
-        mu_pred_target = mu_pred * mu_target
-        
-        sigma_pred_sq = F.conv2d(pred * pred, kernel, padding=self.window_size // 2, groups=channels) - mu_pred_sq
-        sigma_target_sq = F.conv2d(target * target, kernel, padding=self.window_size // 2, groups=channels) - mu_target_sq
-        sigma_pred_target = F.conv2d(pred * target, kernel, padding=self.window_size // 2, groups=channels) - mu_pred_target
-        
-        ssim = ((2 * mu_pred_target + self.C1) * (2 * sigma_pred_target + self.C2)) / \
-               ((mu_pred_sq + mu_target_sq + self.C1) * (sigma_pred_sq + sigma_target_sq + self.C2))
-        
-        return 1 - ssim.mean()
-
-
 class CombinedLoss(nn.Module):
-    """Combined loss: MSE + L1 + SSIM + Edge for comprehensive image reconstruction"""
-    def __init__(self, mse_weight=1.0, l1_weight=0.5, edge_weight=0.15, ssim_weight=0.3):
+    """Combined loss: MSE + L1 + Edge-aware component for better reconstruction"""
+    def __init__(self, mse_weight=0.7, l1_weight=0.8, edge_weight=0.2):
         super().__init__()
         self.mse_weight = mse_weight
         self.l1_weight = l1_weight
         self.edge_weight = edge_weight
-        self.ssim_weight = ssim_weight
         self.mse = nn.MSELoss()
         self.l1 = nn.L1Loss()
-        self.ssim = SSIMLoss(window_size=7)
         
         # Sobel filters for edge detection
         sobel_x = torch.tensor([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]], dtype=torch.float32).view(1, 1, 3, 3)
@@ -82,10 +38,10 @@ class CombinedLoss(nn.Module):
     
     def edge_loss(self, pred, target):
         """Compute edge-aware loss using Sobel filters"""
-        pred_edge_x = F.conv2d(pred, self.sobel_x, padding=1, groups=3)
-        pred_edge_y = F.conv2d(pred, self.sobel_y, padding=1, groups=3)
-        target_edge_x = F.conv2d(target, self.sobel_x, padding=1, groups=3)
-        target_edge_y = F.conv2d(target, self.sobel_y, padding=1, groups=3)
+        pred_edge_x = torch.nn.functional.conv2d(pred, self.sobel_x, padding=1, groups=3)
+        pred_edge_y = torch.nn.functional.conv2d(pred, self.sobel_y, padding=1, groups=3)
+        target_edge_x = torch.nn.functional.conv2d(target, self.sobel_x, padding=1, groups=3)
+        target_edge_y = torch.nn.functional.conv2d(target, self.sobel_y, padding=1, groups=3)
         
         edge_loss = self.l1(pred_edge_x, target_edge_x) + self.l1(pred_edge_y, target_edge_y)
         return edge_loss
@@ -94,12 +50,8 @@ class CombinedLoss(nn.Module):
         mse_loss = self.mse(pred, target)
         l1_loss = self.l1(pred, target)
         edge_loss = self.edge_loss(pred, target)
-        ssim_loss = self.ssim(pred, target)
         
-        total_loss = (self.mse_weight * mse_loss + 
-                      self.l1_weight * l1_loss + 
-                      self.edge_weight * edge_loss +
-                      self.ssim_weight * ssim_loss)
+        total_loss = self.mse_weight * mse_loss + self.l1_weight * l1_loss + self.edge_weight * edge_loss
         return total_loss
 
 
@@ -132,20 +84,24 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
     plotpath = os.path.join(results_path, "plots")
     os.makedirs(plotpath, exist_ok=True)
 
-    image_dataset = datasets.ImageDataset(datafolder=data_path)
+    # Create dataset with augmentation for training, without for validation/test
+    image_dataset_full = datasets.ImageDataset(datafolder=data_path, augment=False)
 
-    n_total = len(image_dataset)
+    n_total = len(image_dataset_full)
     n_test = int(n_total * testset_ratio)
     n_valid = int(n_total * validset_ratio)
     n_train = n_total - n_test - n_valid
     indices = np.random.permutation(n_total)
-    dataset_train = Subset(image_dataset, indices=indices[0:n_train])
-    dataset_valid = Subset(image_dataset, indices=indices[n_train:n_train + n_valid])
-    dataset_test = Subset(image_dataset, indices=indices[n_train + n_valid:n_total])
     
-    assert len(image_dataset) == len(dataset_train) + len(dataset_test) + len(dataset_valid)
+    # Create augmented dataset for training
+    image_dataset_train = datasets.ImageDataset(datafolder=data_path, augment=True)
+    dataset_train = Subset(image_dataset_train, indices=indices[0:n_train])
+    dataset_valid = Subset(image_dataset_full, indices=indices[n_train:n_train + n_valid])
+    dataset_test = Subset(image_dataset_full, indices=indices[n_train + n_valid:n_total])
 
-    del image_dataset
+    assert n_total == len(dataset_train) + len(dataset_test) + len(dataset_valid)
+
+    del image_dataset_full, image_dataset_train
 
     dataloader_train = DataLoader(dataset=dataset_train, batch_size=batchsize,
                                   num_workers=0, shuffle=True)
@@ -159,15 +115,19 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
     network.to(device)
     network.train()
 
-    # defining the loss - combined loss for better reconstruction
-    combined_loss = CombinedLoss(mse_weight=1.0, l1_weight=0.5, edge_weight=0.15, ssim_weight=0.3).to(device)
+    # defining the loss - combined loss with optimized weights
+    combined_loss = CombinedLoss(mse_weight=0.7, l1_weight=0.8, edge_weight=0.2).to(device)
     mse_loss = torch.nn.MSELoss()  # Keep for evaluation
 
     # defining the optimizer with AdamW for better weight decay handling
     optimizer = torch.optim.AdamW(network.parameters(), lr=learningrate, weight_decay=weight_decay)
     
-    # Learning rate scheduler for better convergence
-    scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=50, T_mult=2, eta_min=1e-6)
+    # Learning rate scheduler with better configuration
+    scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=100, T_mult=2, eta_min=1e-7)
+    
+    # Mixed precision training for faster computation and lower memory usage
+    scaler = torch.cuda.amp.GradScaler() if device.type == 'cuda' else None
+    use_amp = scaler is not None
 
     if use_wandb:
         wandb.watch(network, mse_loss, log="all", log_freq=10)
@@ -176,10 +136,13 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
     counter = 0
     best_validation_loss = np.inf
     loss_list = []
+    accumulation_steps = 2  # Gradient accumulation for effective larger batch size
 
     saved_model_path = os.path.join(results_path, "best_model.pt")
 
     print(f"Started training on device {device}")
+    print(f"Using mixed precision: {use_amp}")
+    print(f"Gradient accumulation steps: {accumulation_steps}")
 
     while i < n_updates:
 
@@ -190,21 +153,33 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
             if (i + 1) % print_train_stats_at == 0:
                 print(f'Update Step {i + 1} of {n_updates}: Current loss: {loss_list[-1]}')
 
-            optimizer.zero_grad()
-
+            # Use mixed precision if available
+            if use_amp:
+                with torch.cuda.amp.autocast():
                     output = network(input)
-
                     loss = combined_loss(output, target)
-
+                loss = loss / accumulation_steps
+                scaler.scale(loss).backward()
+            else:
+                output = network(input)
+                loss = combined_loss(output, target)
+                loss = loss / accumulation_steps
                 loss.backward()
             
-            # Gradient clipping for training stability
+            # Gradient accumulation - update weights every accumulation_steps
+            if (i + 1) % accumulation_steps == 0:
+                if use_amp:
+                    scaler.unscale_(optimizer)
+                    torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
+                    scaler.step(optimizer)
+                    scaler.update()
+                else:
                     torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
-
                     optimizer.step()
-            scheduler.step(i + len(loss_list) / len(dataloader_train))
+                optimizer.zero_grad()
+                scheduler.step(i / n_updates)
 
-            loss_list.append(loss.item())
+            loss_list.append(loss.item() * accumulation_steps)
 
             # writing the stats to wandb
             if use_wandb and (i+1) % print_stats_at == 0:
@@ -213,7 +188,9 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
             # plotting
             if (i + 1) % plot_at == 0:
                 print(f"Plotting images, current update {i + 1}")
-                plot(input.cpu().numpy(), target.detach().cpu().numpy(), output.detach().cpu().numpy(), plotpath, i)
+                # Convert to float32 for matplotlib compatibility (mixed precision may produce float16)
+                plot(input.float().cpu().numpy(), target.detach().float().cpu().numpy(), 
+                     output.detach().float().cpu().numpy(), plotpath, i)
 
             # evaluating model every validate_at sample
             if (i + 1) % validate_at == 0:

image-inpainting/src/utils.py

@@ -81,42 +81,9 @@ def read_compressed_file(file_path: str):
     return input_arrays, known_arrays
 
 
-def apply_tta(model, input_tensor, device):
+def create_predictions(model_config, state_dict_path, testset_path, device, save_path, plot_path, plot_at=20, rmse_value=None):
     """
-    Apply Test-Time Augmentation for better predictions.
-    Averages predictions from original and augmented versions.
-    """
-    outputs = []
-    
-    # Original
-    out = model(input_tensor)
-    outputs.append(out)
-    
-    # Horizontal flip
-    flipped_h = torch.flip(input_tensor, dims=[3])
-    out_h = model(flipped_h)
-    out_h = torch.flip(out_h, dims=[3])
-    outputs.append(out_h)
-    
-    # Vertical flip
-    flipped_v = torch.flip(input_tensor, dims=[2])
-    out_v = model(flipped_v)
-    out_v = torch.flip(out_v, dims=[2])
-    outputs.append(out_v)
-    
-    # Both flips
-    flipped_hv = torch.flip(input_tensor, dims=[2, 3])
-    out_hv = model(flipped_hv)
-    out_hv = torch.flip(out_hv, dims=[2, 3])
-    outputs.append(out_hv)
-    
-    # Average all predictions
-    return torch.stack(outputs, dim=0).mean(dim=0)
-
-
-def create_predictions(model_config, state_dict_path, testset_path, device, save_path, plot_path, plot_at=20, rmse_value=None, use_tta=True):
-    """
-    Create predictions with optional Test-Time Augmentation for improved results.
+    Here, one might needs to adjust the code based on the used preprocessing
     """
 
     if device is None:
@@ -127,7 +94,7 @@ def create_predictions(model_config, state_dict_path, testset_path, device, save
         device = torch.device(device)
 
     model = MyModel(**model_config)
-    model.load_state_dict(torch.load(state_dict_path, weights_only=True))
+    model.load_state_dict(torch.load(state_dict_path))
     model.to(device)
     model.eval()
 
@@ -144,14 +111,9 @@ def create_predictions(model_config, state_dict_path, testset_path, device, save
     with torch.no_grad():
         for i in range(len(input_arrays)):
             print(f"Processing image {i + 1}/{len(input_arrays)}")
-            input_array = torch.from_numpy(input_arrays[i]).to(device)
-            input_tensor = input_array.unsqueeze(0) if input_array.dim() == 3 else input_array
-            
-            if use_tta:
-                output = apply_tta(model, input_tensor, device)
-            else:
-                output = model(input_tensor)
-            
+            input_array = torch.from_numpy(input_arrays[i]).to(
+                device)
+            output = model(input_array.unsqueeze(0) if hasattr(input_array, 'dim') and input_array.dim() == 3 else input_array)
             output = output.cpu().numpy()
             predictions.append(output)