clear plot path every start

image-inpainting/src/architecture.py

@@ -18,20 +18,6 @@ def init_weights(m):
     elif isinstance(m, nn.BatchNorm2d):
         nn.init.constant_(m.weight, 1)
         nn.init.constant_(m.bias, 0)
-    elif isinstance(m, nn.GroupNorm):
-        if m.weight is not None:
-            nn.init.constant_(m.weight, 1)
-        if m.bias is not None:
-            nn.init.constant_(m.bias, 0)
-
-
-def _make_norm(num_channels: int) -> nn.Module:
-    """Batch-size independent normalization (works well for batch_size=1 eval)."""
-    # Choose a group count that divides num_channels.
-    num_groups = min(32, num_channels)
-    while num_groups > 1 and (num_channels % num_groups) != 0:
-        num_groups //= 2
-    return nn.GroupNorm(num_groups=num_groups, num_channels=num_channels)
 
 
 class ChannelAttention(nn.Module):
@@ -82,12 +68,52 @@ 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):
         super().__init__()
         self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
-        self.bn = _make_norm(out_channels)
+        self.bn = nn.BatchNorm2d(out_channels)
         self.relu = nn.LeakyReLU(0.1, inplace=True)
         self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
     
@@ -99,9 +125,9 @@ class ResidualConvBlock(nn.Module):
     def __init__(self, channels, dropout=0.0):
         super().__init__()
         self.conv1 = nn.Conv2d(channels, channels, 3, padding=1)
-        self.bn1 = _make_norm(channels)
+        self.bn1 = nn.BatchNorm2d(channels)
         self.conv2 = nn.Conv2d(channels, channels, 3, padding=1)
-        self.bn2 = _make_norm(channels)
+        self.bn2 = nn.BatchNorm2d(channels)
         self.relu = nn.LeakyReLU(0.1, inplace=True)
         self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
     
@@ -114,26 +140,6 @@ class ResidualConvBlock(nn.Module):
         return self.relu(out)
 
 
-class GatedConvBlock(nn.Module):
-    """Gated convolution block (helps the network condition on the mask channel)."""
-    def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dropout=0.0):
-        super().__init__()
-        self.feature = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
-        self.gate = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
-        self.norm = _make_norm(out_channels)
-        self.act = nn.LeakyReLU(0.1, inplace=True)
-        self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
-
-    def forward(self, x):
-        feat = self.feature(x)
-        gate = torch.sigmoid(self.gate(x))
-        out = feat * gate
-        out = self.norm(out)
-        out = self.act(out)
-        out = self.dropout(out)
-        return 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):
@@ -181,7 +187,7 @@ class MyModel(nn.Module):
         
         # Initial convolution with larger receptive field
         self.init_conv = nn.Sequential(
-            GatedConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3, dropout=dropout),
+            ConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3),
             ConvBlock(base_channels, base_channels),
             ResidualConvBlock(base_channels)
         )
@@ -192,10 +198,11 @@ 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
+        # Bottleneck with multiple residual blocks and multi-scale features
         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),
             ResidualConvBlock(base_channels * 16, dropout=dropout),
             CBAM(base_channels * 16)

image-inpainting/src/datasets.py

@@ -10,11 +10,50 @@ import numpy as np
 import random
 import glob
 import os
-from PIL import Image
+from PIL import Image, ImageEnhance, ImageFilter
 
 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)
@@ -32,36 +71,19 @@ def resize(img: Image):
         transforms.CenterCrop((IMAGE_DIMENSION, IMAGE_DIMENSION))
     ])
     return resize_transforms(img)
-
-
-def augment_geometric(img: Image.Image) -> Image.Image:
-    """Lightweight, label-preserving augmentation (safe for train/val/test splits)."""
-    # Horizontal flip
-    if random.random() < 0.5:
-        img = img.transpose(Image.Transpose.FLIP_LEFT_RIGHT)
-    # Vertical flip (less frequent)
-    if random.random() < 0.2:
-        img = img.transpose(Image.Transpose.FLIP_TOP_BOTTOM)
-    # 90-degree rotations (no interpolation artifacts)
-    r = random.random()
-    if r < 0.25:
-        img = img.transpose(Image.Transpose.ROTATE_90)
-    elif r < 0.5:
-        img = img.transpose(Image.Transpose.ROTATE_180)
-    elif r < 0.75:
-        img = img.transpose(Image.Transpose.ROTATE_270)
-    return img
 def preprocess(input_array: np.ndarray):
     input_array = np.asarray(input_array, dtype=np.float32) / 255.0
     return input_array
 
 class ImageDataset(torch.utils.data.Dataset):
     """
-    Dataset class for loading images from a folder
+    Dataset class for loading images from a folder with augmentation
     """
 
-    def __init__(self, datafolder: str):
+    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)
@@ -69,22 +91,28 @@ class ImageDataset(torch.utils.data.Dataset):
     def __getitem__(self, idx: int):
         index = int(idx)
         
-        image = Image.open(self.imagefiles[index]).convert("RGB")
+        image = Image.open(self.imagefiles[index]).convert('RGB')
-        image = augment_geometric(image)
+        
-        image = np.asarray(resize(image))
+        # Apply augmentation before resize
+        if self.augment and self.augmentation is not None:
+            image = self.augmentation(image)
+        
+        image = resize(image)
+        image = np.asarray(image)
         image = preprocess(image)
         
-        # Sample a grid-mask similar in density to the challenge testset (~8% known pixels).
+        # More varied spacing for better generalization
-        # IMPORTANT: offset ranges must be tied to spacing to avoid accidental distribution shift.
+        spacing_x = random.randint(2, 8)
-        spacing_x = random.randint(4, 6)
+        spacing_y = random.randint(2, 8)
-        spacing_y = random.randint(2, 4)
+        offset_x = random.randint(0, min(spacing_x - 1, 8))
-        offset_x = random.randint(0, spacing_x - 1)
+        offset_y = random.randint(0, min(spacing_y - 1, 8))
-        offset_y = random.randint(0, spacing_y - 1)
         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

@@ -24,11 +24,11 @@ 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'] = 3e-4  # Optimal learning rate for AdamW
+    config_dict['learningrate'] = 2e-4  # Slightly lower for stable training
-    config_dict['weight_decay'] = 1e-4  # Slightly higher for better regularization
+    config_dict['weight_decay'] = 5e-5  # Reduced weight decay
-    config_dict['n_updates'] = 5000  # More updates for better convergence
+    config_dict['n_updates'] = 8000  # More updates for better convergence
     config_dict['batchsize'] = 8  # Smaller batch for better gradient estimates
-    config_dict['early_stopping_patience'] = 10  # More patience for complex model
+    config_dict['early_stopping_patience'] = 15  # More patience for complex model
     config_dict['use_wandb'] = False
 
     config_dict['print_train_stats_at'] = 10
@@ -38,8 +38,8 @@ if __name__ == '__main__':
 
     network_config = {
         'n_in_channels': 4,
-        'base_channels': 48,  # Good balance between capacity and memory
+        'base_channels': 56,  # Increased capacity for better feature learning
-        'dropout': 0.1  # Regularization
+        'dropout': 0.08  # Slightly less dropout with augmentation
     }
     
     config_dict['network_config'] = network_config

image-inpainting/src/train.py

@@ -10,6 +10,7 @@ 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
 
@@ -20,15 +21,58 @@ 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-like perceptual component"""
+    """Combined loss: MSE + L1 + SSIM + Edge for comprehensive image reconstruction"""
-    def __init__(self, mse_weight=1.0, l1_weight=0.5, edge_weight=0.1):
+    def __init__(self, mse_weight=1.0, l1_weight=0.5, edge_weight=0.15, ssim_weight=0.3):
         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)
@@ -38,10 +82,10 @@ class CombinedLoss(nn.Module):
     
     def edge_loss(self, pred, target):
         """Compute edge-aware loss using Sobel filters"""
-        pred_edge_x = torch.nn.functional.conv2d(pred, self.sobel_x, padding=1, groups=3)
+        pred_edge_x = F.conv2d(pred, self.sobel_x, padding=1, groups=3)
-        pred_edge_y = torch.nn.functional.conv2d(pred, self.sobel_y, padding=1, groups=3)
+        pred_edge_y = F.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_x = F.conv2d(target, self.sobel_x, padding=1, groups=3)
-        target_edge_y = torch.nn.functional.conv2d(target, self.sobel_y, padding=1, groups=3)
+        target_edge_y = F.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
@@ -50,8 +94,12 @@ 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
+        total_loss = (self.mse_weight * mse_loss + 
+                      self.l1_weight * l1_loss + 
+                      self.edge_weight * edge_loss +
+                      self.ssim_weight * ssim_loss)
         return total_loss
 
 
@@ -112,7 +160,7 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
     network.train()
 
     # defining the loss - combined loss for better reconstruction
-    combined_loss = CombinedLoss(mse_weight=1.0, l1_weight=0.5, edge_weight=0.1).to(device)
+    combined_loss = CombinedLoss(mse_weight=1.0, l1_weight=0.5, edge_weight=0.15, ssim_weight=0.3).to(device)
     mse_loss = torch.nn.MSELoss()  # Keep for evaluation
 
     # defining the optimizer with AdamW for better weight decay handling

image-inpainting/src/utils.py

@@ -81,9 +81,42 @@ def read_compressed_file(file_path: str):
     return input_arrays, known_arrays
 
 
-def create_predictions(model_config, state_dict_path, testset_path, device, save_path, plot_path, plot_at=20, rmse_value=None):
+def apply_tta(model, input_tensor, device):
     """
-    Here, one might needs to adjust the code based on the used preprocessing
+    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.
     """
 
     if device is None:
@@ -94,7 +127,7 @@ 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))
+    model.load_state_dict(torch.load(state_dict_path, weights_only=True))
     model.to(device)
     model.eval()
 
@@ -111,9 +144,14 @@ 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(
+            input_array = torch.from_numpy(input_arrays[i]).to(device)
-                device)
+            input_tensor = input_array.unsqueeze(0) if input_array.dim() == 3 else input_array
-            output = model(input_array.unsqueeze(0) if hasattr(input_array, 'dim') and input_array.dim() == 3 else input_array)
+            
+            if use_tta:
+                output = apply_tta(model, input_tensor, device)
+            else:
+                output = model(input_tensor)
+            
             output = output.cpu().numpy()
             predictions.append(output)