added result, 18.5276

image-inpainting/src/architecture.py

@@ -15,13 +15,25 @@ def init_weights(m):
         nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
         if m.bias is not None:
             nn.init.constant_(m.bias, 0)
-    elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d)):
+    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):
     """Channel attention module (squeeze-and-excitation style)"""
     def __init__(self, channels, reduction=16):
@@ -71,36 +83,55 @@ class CBAM(nn.Module):
 
 
 class ConvBlock(nn.Module):
-    """Convolutional block with Conv2d -> InstanceNorm2d -> GELU"""
-    def __init__(self, in_channels, out_channels, kernel_size=3, padding=1, dropout=0.0, dilation=1):
+    """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, dilation=dilation)
-        # InstanceNorm is preferred for style/inpainting tasks
-        self.bn = nn.InstanceNorm2d(out_channels, affine=True)
-        self.act = nn.GELU()
+        self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, padding=padding)
+        self.bn = _make_norm(out_channels)
+        self.relu = nn.LeakyReLU(0.1, inplace=True)
         self.dropout = nn.Dropout2d(dropout) if dropout > 0 else nn.Identity()
     
     def forward(self, x):
-        return self.dropout(self.act(self.bn(self.conv(x))))
+        return self.dropout(self.relu(self.bn(self.conv(x))))
 
 class ResidualConvBlock(nn.Module):
     """Residual convolutional block for better gradient flow"""
-    def __init__(self, channels, dropout=0.0, dilation=1):
+    def __init__(self, channels, dropout=0.0):
         super().__init__()
-        self.conv1 = nn.Conv2d(channels, channels, 3, padding=dilation, dilation=dilation)
-        self.bn1 = nn.InstanceNorm2d(channels, affine=True)
-        self.conv2 = nn.Conv2d(channels, channels, 3, padding=dilation, dilation=dilation)
-        self.bn2 = nn.InstanceNorm2d(channels, affine=True)
-        self.act = nn.GELU()
+        self.conv1 = nn.Conv2d(channels, channels, 3, padding=1)
+        self.bn1 = _make_norm(channels)
+        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.act(self.bn1(self.conv1(x)))
+        out = self.relu(self.bn1(self.conv1(x)))
         out = self.dropout(out)
         out = self.bn2(self.conv2(out))
         out = out + residual
-        return self.act(out)
+        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):
@@ -150,7 +181,7 @@ class MyModel(nn.Module):
         
         # Initial convolution with larger receptive field
         self.init_conv = nn.Sequential(
-            ConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3),
+            GatedConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3, dropout=dropout),
             ConvBlock(base_channels, base_channels),
             ResidualConvBlock(base_channels)
         )
@@ -164,9 +195,9 @@ class MyModel(nn.Module):
         # Bottleneck with multiple residual blocks
         self.bottleneck = nn.Sequential(
             ConvBlock(base_channels * 16, base_channels * 16, dropout=dropout),
-            ResidualConvBlock(base_channels * 16, dropout=dropout, dilation=2),
-            ResidualConvBlock(base_channels * 16, dropout=dropout, dilation=4),
-            ResidualConvBlock(base_channels * 16, dropout=dropout, dilation=8),
+            ResidualConvBlock(base_channels * 16, dropout=dropout),
+            ResidualConvBlock(base_channels * 16, dropout=dropout),
+            ResidualConvBlock(base_channels * 16, dropout=dropout),
             CBAM(base_channels * 16)
         )
         
@@ -186,7 +217,7 @@ class MyModel(nn.Module):
         # Output layer with smooth transition
         self.output = nn.Sequential(
             nn.Conv2d(base_channels, base_channels // 2, kernel_size=3, padding=1),
-            nn.GELU(),
+            nn.LeakyReLU(0.1, inplace=True),
             nn.Conv2d(base_channels // 2, 3, kernel_size=1),
             nn.Sigmoid()  # Ensure output is in [0, 1] range
         )

image-inpainting/src/datasets.py

@@ -26,22 +26,31 @@ def create_arrays_from_image(image_array: np.ndarray, offset: tuple, spacing: tu
 
     return image_array, known_array
 
-def resize(img: Image, augment: bool = False):
-    transforms_list = [
+def resize(img: Image):
+    resize_transforms = transforms.Compose([
         transforms.Resize((IMAGE_DIMENSION, IMAGE_DIMENSION)),
         transforms.CenterCrop((IMAGE_DIMENSION, IMAGE_DIMENSION))
-    ]
-    
-    if augment:
-        transforms_list = [
-            transforms.RandomHorizontalFlip(),
-            transforms.RandomVerticalFlip(),
-            transforms.ColorJitter(brightness=0.1, contrast=0.1, saturation=0.1, hue=0.05),
-            transforms.RandomRotation(10),
-        ] + transforms_list
-        
-    resize_transforms = transforms.Compose(transforms_list)
+    ])
     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
@@ -51,9 +60,8 @@ class ImageDataset(torch.utils.data.Dataset):
     Dataset class for loading images from a folder
     """
 
-    def __init__(self, datafolder: str, augment: bool = False):
+    def __init__(self, datafolder: str):
         self.imagefiles = sorted(glob.glob(os.path.join(datafolder,"**","*.jpg"),recursive=True))
-        self.augment = augment
 
     def __len__(self):
         return len(self.imagefiles)
@@ -61,13 +69,17 @@ class ImageDataset(torch.utils.data.Dataset):
     def __getitem__(self, idx:int):
         index = int(idx)
         
-        image = Image.open(self.imagefiles[index])
-        image = np.asarray(resize(image, self.augment))
+        image = Image.open(self.imagefiles[index]).convert("RGB")
+        image = augment_geometric(image)
+        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)
+
+        # Sample a grid-mask similar in density to the challenge testset (~8% known pixels).
+        # IMPORTANT: offset ranges must be tied to spacing to avoid accidental distribution shift.
+        spacing_x = random.randint(4, 6)
+        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/train.py

@@ -84,21 +84,16 @@ 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, augment=False)
+    image_dataset = datasets.ImageDataset(datafolder=data_path)
 
     n_total = len(image_dataset)
     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)
-    
-    # Create datasets with and without augmentation
-    train_dataset_source = datasets.ImageDataset(datafolder=data_path, augment=True)
-    val_test_dataset_source = datasets.ImageDataset(datafolder=data_path, augment=False)
-    
-    dataset_train = Subset(train_dataset_source, indices=indices[0:n_train])
-    dataset_valid = Subset(val_test_dataset_source, indices=indices[n_train:n_train + n_valid])
-    dataset_test = Subset(val_test_dataset_source, indices=indices[n_train + n_valid: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)