Added baseline

image-inpainting/src/architecture.py

@@ -5,7 +5,109 @@
 """
 
 import torch
+import torch.nn as nn
 
-class MyModel(torch.nn.Module):
-    # TODO: Implement the model architecture.
-    pass
+class ConvBlock(nn.Module):
+    """Convolutional block with Conv2d -> BatchNorm -> ReLU"""
+    def __init__(self, in_channels, out_channels, kernel_size=3, padding=1):
+        super().__init__()
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
+        self.bn = nn.BatchNorm2d(out_channels)
+        self.relu = nn.ReLU(inplace=True)
+    
+    def forward(self, x):
+        return self.relu(self.bn(self.conv(x)))
+
+class DownBlock(nn.Module):
+    """Downsampling block with two conv blocks and max pooling"""
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        self.conv1 = ConvBlock(in_channels, out_channels)
+        self.conv2 = ConvBlock(out_channels, out_channels)
+        self.pool = nn.MaxPool2d(2)
+    
+    def forward(self, x):
+        skip = self.conv2(self.conv1(x))
+        return self.pool(skip), skip
+
+class UpBlock(nn.Module):
+    """Upsampling block with transposed conv and two conv blocks"""
+    def __init__(self, in_channels, out_channels):
+        super().__init__()
+        self.up = nn.ConvTranspose2d(in_channels, out_channels, kernel_size=2, stride=2)
+        self.conv1 = ConvBlock(in_channels, out_channels)  # in_channels because of concatenation
+        self.conv2 = ConvBlock(out_channels, out_channels)
+    
+    def forward(self, x, skip):
+        x = self.up(x)
+        # Handle dimension mismatch by interpolating x to match skip's size
+        if x.shape[2:] != skip.shape[2:]:
+            x = nn.functional.interpolate(x, size=skip.shape[2:], mode='bilinear', align_corners=False)
+        x = torch.cat([x, skip], dim=1)
+        x = self.conv1(x)
+        x = self.conv2(x)
+        return x
+
+class MyModel(nn.Module):
+    """U-Net style architecture for image inpainting"""
+    def __init__(self, n_in_channels: int, base_channels: int = 64):
+        super().__init__()
+        
+        # Initial convolution
+        self.init_conv = ConvBlock(n_in_channels, base_channels)
+        
+        # Encoder (downsampling path)
+        self.down1 = DownBlock(base_channels, base_channels * 2)
+        self.down2 = DownBlock(base_channels * 2, base_channels * 4)
+        self.down3 = DownBlock(base_channels * 4, base_channels * 8)
+        
+        # Bottleneck
+        self.bottleneck1 = ConvBlock(base_channels * 8, base_channels * 16)
+        self.bottleneck2 = ConvBlock(base_channels * 16, base_channels * 16)
+        
+        # Decoder (upsampling path)
+        self.up1 = UpBlock(base_channels * 16, base_channels * 8)
+        self.up2 = UpBlock(base_channels * 8, base_channels * 4)
+        self.up3 = UpBlock(base_channels * 4, base_channels * 2)
+        
+        # Final upsampling and output
+        self.final_up = nn.ConvTranspose2d(base_channels * 2, base_channels, kernel_size=2, stride=2)
+        self.final_conv1 = ConvBlock(base_channels * 2, base_channels)
+        self.final_conv2 = ConvBlock(base_channels, base_channels)
+        
+        # Output layer
+        self.output = nn.Conv2d(base_channels, 3, kernel_size=1)
+        self.sigmoid = nn.Sigmoid()  # To ensure output is in [0, 1] range
+    
+    def forward(self, x):
+        # Initial convolution
+        x0 = self.init_conv(x)
+        
+        # Encoder
+        x1, skip1 = self.down1(x0)
+        x2, skip2 = self.down2(x1)
+        x3, skip3 = self.down3(x2)
+        
+        # Bottleneck
+        x = self.bottleneck1(x3)
+        x = self.bottleneck2(x)
+        
+        # Decoder with skip connections
+        x = self.up1(x, skip3)
+        x = self.up2(x, skip2)
+        x = self.up3(x, skip1)
+        
+        # Final layers
+        x = self.final_up(x)
+        # Handle dimension mismatch for final concatenation
+        if x.shape[2:] != x0.shape[2:]:
+            x = nn.functional.interpolate(x, size=x0.shape[2:], mode='bilinear', align_corners=False)
+        x = torch.cat([x, x0], dim=1)
+        x = self.final_conv1(x)
+        x = self.final_conv2(x)
+        
+        # Output
+        x = self.output(x)
+        x = self.sigmoid(x)
+        
+        return x

image-inpainting/src/datasets.py

@@ -4,6 +4,7 @@
     datasets.py
 """
 
+from torchvision import transforms
 import torch
 import numpy as np
 import random
@@ -15,16 +16,25 @@ IMAGE_DIMENSION = 100
 
 
 def create_arrays_from_image(image_array: np.ndarray, offset: tuple, spacing: tuple) -> tuple[np.ndarray, np.ndarray]:
-    image_array, known_array = None, None
+    image_array = np.transpose(image_array, (2, 0, 1))
+    known_array = np.zeros_like(image_array)
 
-    # TODO: Implement the logic to create input and known arrays based on offset and spacing
+    known_array[:, offset[1]::spacing[1], offset[0]::spacing[0]] = 1
+
+    image_array[known_array == 0] = 0
+    known_array = known_array[0:1]
 
     return image_array, known_array
 
 def resize(img: Image):
-    pass
+    resize_transforms = transforms.Compose([
+        transforms.Resize((IMAGE_DIMENSION, IMAGE_DIMENSION)),
+        transforms.CenterCrop((IMAGE_DIMENSION, IMAGE_DIMENSION))
+    ])
+    return resize_transforms(img)
 def preprocess(input_array: np.ndarray):
-    pass
+    input_array = np.asarray(input_array, dtype=np.float32) / 255.0
+    return input_array
 
 class ImageDataset(torch.utils.data.Dataset):
     """
@@ -38,6 +48,20 @@ class ImageDataset(torch.utils.data.Dataset):
         return len(self.imagefiles)
         
     def __getitem__(self, idx:int):
-        pass
+        index = int(idx)
         
-    # TODO: Implement the __init__, __len__, and __getitem__ methods
+        image = Image.open(self.imagefiles[index])
+        image = np.asarray(resize(image))
+        image = preprocess(image)
+        spacing_x = random.randint(2,6)
+        spacing_y = random.randint(2,6)
+        offset_x = random.randint(0,8)
+        offset_y = random.randint(0,8)
+        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

@@ -17,30 +17,34 @@ if __name__ == '__main__':
     config_dict['seed'] = 42
     config_dict['testset_ratio'] = 0.1
     config_dict['validset_ratio'] = 0.1
-    config_dict['results_path'] = os.path.join("results")
-    config_dict['data_path'] = os.path.join("data", "dataset")
+    # Get the absolute path based on the script's location
+    script_dir = os.path.dirname(os.path.abspath(__file__))
+    project_root = os.path.dirname(script_dir)
+    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'] = 1e-3
+    config_dict['learningrate'] = 5e-4  # Slightly lower for more stable training
     config_dict['weight_decay'] = 1e-5 # default is 0
-    config_dict['n_updates'] = 50000
-    config_dict['batchsize'] = 32
-    config_dict['early_stopping_patience'] = 3
+    config_dict['n_updates'] = 200
+    config_dict['batchsize'] = 16  # Reduced due to larger model
+    config_dict['early_stopping_patience'] = 5  # 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'] = 100
+    config_dict['plot_at'] = 10
     config_dict['validate_at'] = 100
 
     network_config = {
-        'n_in_channels': 4
+        'n_in_channels': 4,
+        'base_channels': 32  # Start with 32, can increase to 64 for even better results
     }
     
     config_dict['network_config'] = network_config
 
     train(**config_dict)
     
-    testset_path = os.path.join("data", "challenge_testset.npz")
+    testset_path = os.path.join(project_root, "data", "challenge_testset.npz")
     state_dict_path = os.path.join(config_dict['results_path'], "best_model.pt")
     save_path = os.path.join(config_dict['results_path'], "testset", "my_submission_name.npz")
     plot_path = os.path.join(config_dict['results_path'], "testset", "plots")

image-inpainting/src/utils.py

@@ -18,7 +18,7 @@ def plot(inputs, targets, predictions, path, update):
     os.makedirs(path, exist_ok=True)
     fig, axes = plt.subplots(ncols=3, figsize=(15, 5))
 
-    for i in range(len(inputs)):
+    for i in range(5):
         for ax, data, title in zip(axes, [inputs, targets, predictions], ["Input", "Target", "Prediction"]):
             ax.clear()
             ax.set_title(title)