remove pycache

image-inpainting/.gitignore

@@ -1,4 +1,5 @@
 data/*
 *.zip
 *.jpg
-*.pt
+*.pt
+__pycache__/

image-inpainting/results/runtime_config.json

@@ -0,0 +1,16 @@
+{
+  "learningrate": 0.0003,
+  "weight_decay": 1e-05,
+  "n_updates": 150000,
+  "plot_at": 400,
+  "early_stopping_patience": 40,
+  "print_stats_at": 200,
+  "print_train_stats_at": 50,
+  "validate_at": 200,
+  "accumulation_steps": 1,
+  "commands": {
+    "save_checkpoint": false,
+    "run_test_validation": false,
+    "generate_predictions": false
+  }
+}

image-inpainting/src/architecture.py

@@ -18,44 +18,48 @@ 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:
+class GatedSkipConnection(nn.Module):
-    """Batch-size independent normalization (works well for batch_size=1 eval)."""
+    """Gated skip connection for better feature fusion"""
-    # Choose a group count that divides num_channels.
+    def __init__(self, up_channels, skip_channels):
-    num_groups = min(32, num_channels)
+        super().__init__()
-    while num_groups > 1 and (num_channels % num_groups) != 0:
+        self.gate = nn.Sequential(
-        num_groups //= 2
+            nn.Conv2d(up_channels + skip_channels, up_channels, 1),
-    return nn.GroupNorm(num_groups=num_groups, num_channels=num_channels)
+            nn.Sigmoid()
+        )
+        # Project skip to match up_channels if they differ
+        if skip_channels != up_channels:
+            self.skip_proj = nn.Conv2d(skip_channels, up_channels, 1)
+        else:
+            self.skip_proj = nn.Identity()
+    
+    def forward(self, x, skip):
+        skip_proj = self.skip_proj(skip)
+        combined = torch.cat([x, skip], dim=1)
+        gate = self.gate(combined)
+        return x * gate + skip_proj * (1 - gate)
 
 
-class ChannelAttention(nn.Module):
+class EfficientChannelAttention(nn.Module):
-    """Channel attention module (squeeze-and-excitation style)"""
+    """Efficient channel attention without dimensionality reduction"""
-    def __init__(self, channels, reduction=16):
+    def __init__(self, channels):
         super().__init__()
         self.avg_pool = nn.AdaptiveAvgPool2d(1)
-        self.max_pool = nn.AdaptiveMaxPool2d(1)
+        self.conv = nn.Conv1d(1, 1, kernel_size=3, padding=1, bias=False)
-        reduced = max(channels // reduction, 8)
-        self.fc = nn.Sequential(
-            nn.Conv2d(channels, reduced, 1, bias=False),
-            nn.ReLU(inplace=True),
-            nn.Conv2d(reduced, channels, 1, bias=False)
-        )
         self.sigmoid = nn.Sigmoid()
     
     def forward(self, x):
-        avg_out = self.fc(self.avg_pool(x))
+        # Global pooling
-        max_out = self.fc(self.max_pool(x))
+        y = self.avg_pool(x)
-        return x * self.sigmoid(avg_out + max_out)
+        # 1D convolution on channel dimension
+        y = self.conv(y.squeeze(-1).transpose(-1, -2)).transpose(-1, -2).unsqueeze(-1)
+        y = self.sigmoid(y)
+        return x * y.expand_as(x)
 
 
 class SpatialAttention(nn.Module):
-    """Spatial attention module"""
+    """Efficient spatial attention module"""
     def __init__(self, kernel_size=7):
         super().__init__()
         self.conv = nn.Conv2d(2, 1, kernel_size, padding=kernel_size // 2, bias=False)
@@ -69,12 +73,12 @@ class SpatialAttention(nn.Module):
         return x * attn
 
 
-class CBAM(nn.Module):
+class EfficientAttention(nn.Module):
-    """Convolutional Block Attention Module"""
+    """Lightweight attention module combining channel and spatial"""
-    def __init__(self, channels, reduction=16):
+    def __init__(self, channels):
         super().__init__()
-        self.channel_attn = ChannelAttention(channels, reduction)
+        self.channel_attn = EfficientChannelAttention(channels)
-        self.spatial_attn = SpatialAttention()
+        self.spatial_attn = SpatialAttention(kernel_size=5)
     
     def forward(self, x):
         x = self.channel_attn(x)
@@ -84,175 +88,220 @@ class CBAM(nn.Module):
 
 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):
+    def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dilation=1, dropout=0.0, separable=False):
         super().__init__()
-        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
+        if separable and in_channels > 1:
-        self.bn = _make_norm(out_channels)
+            # Depthwise separable convolution for efficiency
-        self.relu = nn.LeakyReLU(0.1, inplace=True)
+            self.conv = nn.Sequential(
+                nn.Conv2d(in_channels, in_channels, kernel_size, padding=padding, dilation=dilation, groups=in_channels),
+                nn.Conv2d(in_channels, out_channels, 1)
+            )
+        else:
+            self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding, dilation=dilation)
+        self.bn = nn.BatchNorm2d(out_channels)
+        self.relu = nn.LeakyReLU(0.2, inplace=True)
         self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
     
     def forward(self, x):
         return self.dropout(self.relu(self.bn(self.conv(x))))
 
+
+class DenseBlock(nn.Module):
+    """Lightweight dense block for better gradient flow"""
+    def __init__(self, channels, growth_rate=8, num_layers=2, dropout=0.0):
+        super().__init__()
+        self.layers = nn.ModuleList()
+        for i in range(num_layers):
+            self.layers.append(ConvBlock(channels + i * growth_rate, growth_rate, dropout=dropout))
+        self.fusion = nn.Conv2d(channels + num_layers * growth_rate, channels, 1)
+        self.bn = nn.BatchNorm2d(channels)
+        self.relu = nn.LeakyReLU(0.2, inplace=True)
+    
+    def forward(self, x):
+        features = [x]
+        for layer in self.layers:
+            out = layer(torch.cat(features, dim=1))
+            features.append(out)
+        out = self.fusion(torch.cat(features, dim=1))
+        out = self.relu(self.bn(out))
+        return out + x  # Residual connection
+
 class ResidualConvBlock(nn.Module):
-    """Residual convolutional block for better gradient flow"""
+    """Improved residual convolutional block with pre-activation"""
     def __init__(self, channels, dropout=0.0):
         super().__init__()
+        self.bn1 = nn.BatchNorm2d(channels)
+        self.relu1 = nn.LeakyReLU(0.2, inplace=True)
         self.conv1 = nn.Conv2d(channels, channels, 3, padding=1)
-        self.bn1 = _make_norm(channels)
+        self.bn2 = nn.BatchNorm2d(channels)
+        self.relu2 = nn.LeakyReLU(0.2, inplace=True)
         self.conv2 = nn.Conv2d(channels, channels, 3, padding=1)
-        self.bn2 = _make_norm(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.relu1(self.bn1(x))
+        out = self.conv1(out)
+        out = self.relu2(self.bn2(out))
         out = self.dropout(out)
-        out = self.bn2(self.conv2(out))
+        out = self.conv2(out)
-        out = out + residual
+        return out + residual
-        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"""
+    """Enhanced downsampling block with dense and residual connections"""
-    def __init__(self, in_channels, out_channels, dropout=0.1):
+    def __init__(self, in_channels, out_channels, dropout=0.1, use_attention=True, use_dense=False):
         super().__init__()
-        self.conv1 = ConvBlock(in_channels, out_channels, dropout=dropout)
+        self.conv1 = ConvBlock(in_channels, out_channels, dropout=dropout, separable=True)
         self.conv2 = ConvBlock(out_channels, out_channels, dropout=dropout)
-        self.residual = ResidualConvBlock(out_channels, dropout=dropout)
+        if use_dense:
-        self.attention = CBAM(out_channels)
+            self.dense = DenseBlock(out_channels, growth_rate=8, num_layers=2, dropout=dropout)
+        else:
+            self.dense = ResidualConvBlock(out_channels, dropout=dropout)
+        self.attention = EfficientAttention(out_channels) if use_attention else nn.Identity()
         self.pool = nn.MaxPool2d(2)
     
     def forward(self, x):
         x = self.conv1(x)
         x = self.conv2(x)
-        x = self.residual(x)
+        x = self.dense(x)
         skip = self.attention(x)
         return self.pool(skip), skip
 
 class UpBlock(nn.Module):
-    """Upsampling block with transposed conv, residual connection, attention, and conv blocks"""
+    """Enhanced upsampling block with gated skip connections"""
-    def __init__(self, in_channels, out_channels, dropout=0.1):
+    def __init__(self, in_channels, out_channels, dropout=0.1, use_attention=True, use_dense=False):
         super().__init__()
         self.up = nn.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2)
-        # After concat: out_channels (from upconv) + in_channels (from skip)
+        # Skip connection has in_channels, upsampled has out_channels
-        self.conv1 = ConvBlock(out_channels + in_channels, out_channels, dropout=dropout)
+        self.gated_skip = GatedSkipConnection(out_channels, in_channels)
+        # After gated skip: out_channels
+        self.conv1 = ConvBlock(out_channels, out_channels, dropout=dropout, separable=True)
         self.conv2 = ConvBlock(out_channels, out_channels, dropout=dropout)
-        self.residual = ResidualConvBlock(out_channels, dropout=dropout)
+        if use_dense:
-        self.attention = CBAM(out_channels)
+            self.dense = DenseBlock(out_channels, growth_rate=8, num_layers=2, dropout=dropout)
+        else:
+            self.dense = ResidualConvBlock(out_channels, dropout=dropout)
+        self.attention = EfficientAttention(out_channels) if use_attention else nn.Identity()
     
     def forward(self, x, skip):
         x = self.up(x)
-        # Handle dimension mismatch by interpolating x to match skip's size
+        # Handle dimension mismatch
         if x.shape[2:] != skip.shape[2:]:
             x = F.interpolate(x, size=skip.shape[2:], mode='bilinear', align_corners=False)
-        x = torch.cat([x, skip], dim=1)
+        x = self.gated_skip(x, skip)
         x = self.conv1(x)
         x = self.conv2(x)
-        x = self.residual(x)
+        x = self.dense(x)
         x = self.attention(x)
         return x
 
 class MyModel(nn.Module):
-    """Improved U-Net style architecture for image inpainting with attention and residual connections"""
+    """Enhanced U-Net architecture with dense connections and efficient attention"""
     def __init__(self, n_in_channels: int, base_channels: int = 64, dropout: float = 0.1):
         super().__init__()
         
-        # Initial convolution with larger receptive field
+        # Separate mask processing for better feature extraction
-        self.init_conv = nn.Sequential(
+        self.mask_conv = nn.Sequential(
-            GatedConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3, dropout=dropout),
+            nn.Conv2d(1, base_channels // 4, 3, padding=1),
-            ConvBlock(base_channels, base_channels),
+            nn.LeakyReLU(0.2, inplace=True),
-            ResidualConvBlock(base_channels)
+            nn.Conv2d(base_channels // 4, base_channels // 4, 3, padding=1),
+            nn.LeakyReLU(0.2, inplace=True)
         )
         
-        # Encoder (downsampling path)
+        # Image processing path
-        self.down1 = DownBlock(base_channels, base_channels * 2, dropout=dropout)
+        self.image_conv = nn.Sequential(
-        self.down2 = DownBlock(base_channels * 2, base_channels * 4, dropout=dropout)
+            ConvBlock(3, base_channels, kernel_size=5, padding=2),
-        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
-        self.bottleneck = nn.Sequential(
-            ConvBlock(base_channels * 16, base_channels * 16, dropout=dropout),
-            ResidualConvBlock(base_channels * 16, dropout=dropout),
-            ResidualConvBlock(base_channels * 16, dropout=dropout),
-            ResidualConvBlock(base_channels * 16, dropout=dropout),
-            CBAM(base_channels * 16)
-        )
-        
-        # Decoder (upsampling path)
-        self.up1 = UpBlock(base_channels * 16, base_channels * 8, dropout=dropout)
-        self.up2 = UpBlock(base_channels * 8, base_channels * 4, dropout=dropout)
-        self.up3 = UpBlock(base_channels * 4, base_channels * 2, dropout=dropout)
-        self.up4 = UpBlock(base_channels * 2, base_channels, dropout=dropout)
-        
-        # Final refinement layers
-        self.final_conv = nn.Sequential(
-            ConvBlock(base_channels * 2, base_channels),
-            ResidualConvBlock(base_channels),
             ConvBlock(base_channels, base_channels)
         )
         
-        # Output layer with smooth transition
+        # Fusion of mask and image features
+        self.fusion = nn.Sequential(
+            nn.Conv2d(base_channels + base_channels // 4, base_channels, 1),
+            nn.BatchNorm2d(base_channels),
+            nn.LeakyReLU(0.2, inplace=True)
+        )
+        
+        # Encoder with progressive feature extraction
+        self.down1 = DownBlock(base_channels, base_channels * 2, dropout=dropout*0.5, use_attention=False, use_dense=False)
+        self.down2 = DownBlock(base_channels * 2, base_channels * 4, dropout=dropout*0.7, use_attention=True, use_dense=True)
+        self.down3 = DownBlock(base_channels * 4, base_channels * 8, dropout=dropout, use_attention=True, use_dense=True)
+        
+        # Enhanced bottleneck with multi-scale features and dense connections
+        self.bottleneck = nn.Sequential(
+            ConvBlock(base_channels * 8, base_channels * 8, dropout=dropout),
+            DenseBlock(base_channels * 8, growth_rate=10, num_layers=3, dropout=dropout),
+            ConvBlock(base_channels * 8, base_channels * 8, dilation=2, padding=2, dropout=dropout),
+            ResidualConvBlock(base_channels * 8, dropout=dropout),
+            EfficientAttention(base_channels * 8)
+        )
+        
+        # Decoder with progressive reconstruction
+        self.up1 = UpBlock(base_channels * 8, base_channels * 4, dropout=dropout, use_attention=True, use_dense=True)
+        self.up2 = UpBlock(base_channels * 4, base_channels * 2, dropout=dropout*0.7, use_attention=True, use_dense=True)
+        self.up3 = UpBlock(base_channels * 2, base_channels, dropout=dropout*0.5, use_attention=False, use_dense=False)
+        
+        # Multi-scale feature fusion with dense connections
+        self.multiscale_fusion = nn.Sequential(
+            ConvBlock(base_channels * 2, base_channels),
+            DenseBlock(base_channels, growth_rate=8, num_layers=2, dropout=dropout//2),
+            ConvBlock(base_channels, base_channels)
+        )
+        
+        # Output with residual connection to input
+        self.pre_output = nn.Sequential(
+            ConvBlock(base_channels, base_channels),
+            ConvBlock(base_channels, base_channels // 2)
+        )
+        
         self.output = nn.Sequential(
-            nn.Conv2d(base_channels, base_channels // 2, kernel_size=3, padding=1),
+            nn.Conv2d(base_channels // 2 + 3, base_channels // 2, 3, padding=1),
-            nn.LeakyReLU(0.1, inplace=True),
+            nn.LeakyReLU(0.2, inplace=True),
-            nn.Conv2d(base_channels // 2, 3, kernel_size=1),
+            nn.Conv2d(base_channels // 2, 3, 1),
-            nn.Sigmoid()  # Ensure output is in [0, 1] range
+            nn.Sigmoid()
         )
         
         # Apply weight initialization
         self.apply(init_weights)
     
     def forward(self, x):
-        # Initial convolution
+        # Split input into image and mask
-        x0 = self.init_conv(x)
+        image = x[:, :3, :, :]
+        mask = x[:, 3:4, :, :]
+        
+        # Process mask and image separately
+        mask_features = self.mask_conv(mask)
+        image_features = self.image_conv(image)
+        
+        # Fuse features
+        x0 = self.fusion(torch.cat([image_features, mask_features], dim=1))
         
         # Encoder
         x1, skip1 = self.down1(x0)
         x2, skip2 = self.down2(x1)
         x3, skip3 = self.down3(x2)
-        x4, skip4 = self.down4(x3)
         
         # Bottleneck
-        x = self.bottleneck(x4)
+        x = self.bottleneck(x3)
         
         # Decoder with skip connections
-        x = self.up1(x, skip4)
+        x = self.up1(x, skip3)
-        x = self.up2(x, skip3)
+        x = self.up2(x, skip2)
-        x = self.up3(x, skip2)
+        x = self.up3(x, skip1)
-        x = self.up4(x, skip1)
         
-        # Handle dimension mismatch for final concatenation
+        # Handle dimension mismatch for final fusion
         if x.shape[2:] != x0.shape[2:]:
             x = F.interpolate(x, size=x0.shape[2:], mode='bilinear', align_corners=False)
         
-        # Concatenate with initial features for better detail preservation
+        # Multi-scale fusion with initial features
         x = torch.cat([x, x0], dim=1)
-        x = self.final_conv(x)
+        x = self.multiscale_fusion(x)
         
-        # Output
+        # Pre-output processing
+        x = self.pre_output(x)
+        
+        # Concatenate with original masked image for residual learning
+        x = torch.cat([x, image], dim=1)
         x = self.output(x)
         
         return x

image-inpainting/src/datasets.py

@@ -10,7 +10,7 @@ import numpy as np
 import random
 import glob
 import os
-from PIL import Image
+from PIL import Image, ImageEnhance
 
 IMAGE_DIMENSION = 100
 
@@ -33,35 +33,51 @@ def resize(img: Image):
     ])
     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
 
+def augment_image(img: Image, strength: float = 0.7) -> Image:
+    """Apply comprehensive data augmentation for better generalization"""
+    # Random horizontal flip
+    if random.random() > 0.5:
+        img = img.transpose(Image.FLIP_LEFT_RIGHT)
+    
+    # Random vertical flip
+    if random.random() > 0.5:
+        img = img.transpose(Image.FLIP_TOP_BOTTOM)
+    
+    # Random rotation (90, 180, 270 degrees)
+    if random.random() > 0.5:
+        angle = random.choice([90, 180, 270])
+        img = img.rotate(angle)
+    
+    # Color augmentation - more aggressive for long training
+    rand = random.random()
+    if rand > 0.75:
+        # Brightness
+        factor = 1.0 + random.uniform(-0.2, 0.2) * strength
+        img = ImageEnhance.Brightness(img).enhance(factor)
+    elif rand > 0.5:
+        # Contrast
+        factor = 1.0 + random.uniform(-0.2, 0.2) * strength
+        img = ImageEnhance.Contrast(img).enhance(factor)
+    elif rand > 0.25:
+        # Saturation
+        factor = 1.0 + random.uniform(-0.15, 0.15) * strength
+        img = ImageEnhance.Color(img).enhance(factor)
+    
+    return img
+
 class ImageDataset(torch.utils.data.Dataset):
     """
-    Dataset class for loading images from a folder
+    Dataset class for loading images from a folder with augmentation support
     """
 
-    def __init__(self, datafolder: str):
+    def __init__(self, datafolder: str, augment: bool = True, augment_strength: float = 0.7):
         self.imagefiles = sorted(glob.glob(os.path.join(datafolder,"**","*.jpg"),recursive=True))
+        self.augment = augment
+        self.augment_strength = augment_strength
 
     def __len__(self):
         return len(self.imagefiles)
@@ -69,17 +85,19 @@ 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])
-        image = augment_geometric(image)
+        image = resize(image)
-        image = np.asarray(resize(image))
+        
+        # Apply augmentation
+        if self.augment:
+            image = augment_image(image, self.augment_strength)
+        
+        image = np.asarray(image)
         image = preprocess(image)
-
+        spacing_x = random.randint(2,6)
-        # Sample a grid-mask similar in density to the challenge testset (~8% known pixels).
+        spacing_y = random.randint(2,6)
-        # IMPORTANT: offset ranges must be tied to spacing to avoid accidental distribution shift.
+        offset_x = random.randint(0,8)
-        spacing_x = random.randint(4, 6)
+        offset_y = random.randint(0,8)
-        spacing_y = random.randint(2, 4)
-        offset_x = random.randint(0, spacing_x - 1)
-        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)

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'] = 3e-4  # Optimal learning rate for AdamW
+    config_dict['learningrate'] = 3e-4  # More stable learning rate
-    config_dict['weight_decay'] = 1e-4  # Slightly higher for better regularization
+    config_dict['weight_decay'] = 1e-4  # Proper regularization
-    config_dict['n_updates'] = 5000  # More updates for better convergence
+    config_dict['n_updates'] = 40000  # Extended training
-    config_dict['batchsize'] = 8  # Smaller batch for better gradient estimates
+    config_dict['batchsize'] = 96  # Maximize batch size for better gradients
-    config_dict['early_stopping_patience'] = 10  # More patience for complex model
+    config_dict['early_stopping_patience'] = 20  # More patience for convergence
     config_dict['use_wandb'] = False
 
-    config_dict['print_train_stats_at'] = 10
+    config_dict['print_train_stats_at'] = 50
-    config_dict['print_stats_at'] = 100
+    config_dict['print_stats_at'] = 200
-    config_dict['plot_at'] = 300
+    config_dict['plot_at'] = 500
-    config_dict['validate_at'] = 300  # Validate more frequently
+    config_dict['validate_at'] = 500  # Regular validation
 
     network_config = {
         'n_in_channels': 4,
-        'base_channels': 48,  # Good balance between capacity and memory
+        'base_channels': 64,
-        'dropout': 0.1  # Regularization
+        'dropout': 0.1  # Proper dropout for regularization
     }
     
     config_dict['network_config'] = network_config

image-inpainting/src/train.py

@@ -10,49 +10,36 @@ 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
 
 from torch.utils.data import DataLoader
 from torch.utils.data import Subset
-from torch.optim.lr_scheduler import CosineAnnealingWarmRestarts
 
 import wandb
 
 
-class CombinedLoss(nn.Module):
+class EnhancedRMSELoss(nn.Module):
-    """Combined loss: MSE + L1 + SSIM-like perceptual component"""
+    """Enhanced RMSE loss with edge weighting for sharper predictions"""
-    def __init__(self, mse_weight=1.0, l1_weight=0.5, edge_weight=0.1):
+    def __init__(self):
         super().__init__()
-        self.mse_weight = mse_weight
-        self.l1_weight = l1_weight
-        self.edge_weight = edge_weight
-        self.mse = nn.MSELoss()
-        self.l1 = nn.L1Loss()
-        
-        # 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)
-        sobel_y = torch.tensor([[-1, -2, -1], [0, 0, 0], [1, 2, 1]], dtype=torch.float32).view(1, 1, 3, 3)
-        self.register_buffer('sobel_x', sobel_x.repeat(3, 1, 1, 1))
-        self.register_buffer('sobel_y', sobel_y.repeat(3, 1, 1, 1))
-    
-    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_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
     
     def forward(self, pred, target):
-        mse_loss = self.mse(pred, target)
+        # Compute per-pixel squared error
-        l1_loss = self.l1(pred, target)
+        se = (pred - target) ** 2
-        edge_loss = self.edge_loss(pred, target)
         
-        total_loss = self.mse_weight * mse_loss + self.l1_weight * l1_loss + self.edge_weight * edge_loss
+        # Weight edges more heavily for sharper results
-        return total_loss
+        edge_weight = 1.0 + 0.3 * torch.abs(target[:, :, 1:, :] - target[:, :, :-1, :]).mean(dim=1, keepdim=True)
+        edge_weight = F.pad(edge_weight, (0, 0, 0, 1), value=1.0)
+        
+        # Apply weighting
+        weighted_se = se * edge_weight
+        
+        # Compute RMSE
+        mse = weighted_se.mean()
+        rmse = torch.sqrt(mse + 1e-8)
+        return rmse
 
 
 def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_stopping_patience, device, learningrate,
@@ -67,6 +54,10 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
 
     if isinstance(device, str):
         device = torch.device(device)
+    
+    # Enable mixed precision training for memory efficiency
+    use_amp = torch.cuda.is_available()
+    scaler = torch.amp.GradScaler('cuda') if use_amp else None
 
     if use_wandb:
         wandb.login()
@@ -111,15 +102,17 @@ 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
+    # defining the loss - Enhanced RMSE for sharper predictions
-    combined_loss = CombinedLoss(mse_weight=1.0, l1_weight=0.5, edge_weight=0.1).to(device)
+    rmse_loss = EnhancedRMSELoss().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)
+    optimizer = torch.optim.AdamW(network.parameters(), lr=learningrate, weight_decay=weight_decay, betas=(0.9, 0.999), eps=1e-8)
     
-    # Learning rate scheduler for better convergence
+    # Cosine annealing with warm restarts for gradual learning rate decay
-    scheduler = CosineAnnealingWarmRestarts(optimizer, T_0=50, T_mult=2, eta_min=1e-6)
+    scheduler = torch.optim.lr_scheduler.CosineAnnealingWarmRestarts(
+        optimizer, T_0=n_updates//4, T_mult=1, eta_min=learningrate/100
+    )
 
     if use_wandb:
         wandb.watch(network, mse_loss, log="all", log_freq=10)
@@ -144,17 +137,31 @@ def train(seed, testset_ratio, validset_ratio, data_path, results_path, early_st
 
             optimizer.zero_grad()
 
-            output = network(input)
+            # Mixed precision training for memory efficiency
-
+            if use_amp:
-            loss = combined_loss(output, target)
+                with torch.amp.autocast('cuda'):
-
+                    output = network(input)
-            loss.backward()
+                    loss = rmse_loss(output, target)
+                
+                scaler.scale(loss).backward()
+                
+                # Gradient clipping for training stability
+                scaler.unscale_(optimizer)
+                torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
+                
+                scaler.step(optimizer)
+                scaler.update()
+            else:
+                output = network(input)
+                loss = rmse_loss(output, target)
+                loss.backward()
+                
+                # Gradient clipping for training stability
+                torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
+                
+                optimizer.step()
             
-            # Gradient clipping for training stability
+            scheduler.step()
-            torch.nn.utils.clip_grad_norm_(network.parameters(), max_norm=1.0)
-
-            optimizer.step()
-            scheduler.step(i + len(loss_list) / len(dataloader_train))
 
             loss_list.append(loss.item())
 
@@ -165,7 +172,11 @@ 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
+                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: