improve baseline

This commit is contained in:
2026-01-24 16:15:34 +01:00
parent 3b1d3c0497
commit 57026695d4
10 changed files with 228 additions and 63 deletions

View File

@@ -6,78 +6,190 @@
import torch
import torch.nn as nn
import torch.nn.functional as F
def init_weights(m):
"""Initialize weights using Kaiming initialization for better training"""
if isinstance(m, (nn.Conv2d, nn.ConvTranspose2d)):
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.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
class ChannelAttention(nn.Module):
"""Channel attention module (squeeze-and-excitation style)"""
def __init__(self, channels, reduction=16):
super().__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.max_pool = nn.AdaptiveMaxPool2d(1)
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))
max_out = self.fc(self.max_pool(x))
return x * self.sigmoid(avg_out + max_out)
class SpatialAttention(nn.Module):
"""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)
self.sigmoid = nn.Sigmoid()
def forward(self, x):
avg_out = torch.mean(x, dim=1, keepdim=True)
max_out, _ = torch.max(x, dim=1, keepdim=True)
attn = torch.cat([avg_out, max_out], dim=1)
attn = self.sigmoid(self.conv(attn))
return x * attn
class CBAM(nn.Module):
"""Convolutional Block Attention Module"""
def __init__(self, channels, reduction=16):
super().__init__()
self.channel_attn = ChannelAttention(channels, reduction)
self.spatial_attn = SpatialAttention()
def forward(self, x):
x = self.channel_attn(x)
x = self.spatial_attn(x)
return x
class ConvBlock(nn.Module):
"""Convolutional block with Conv2d -> BatchNorm -> ReLU"""
def __init__(self, in_channels, out_channels, kernel_size=3, padding=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)
self.bn = nn.BatchNorm2d(out_channels)
self.relu = nn.ReLU(inplace=True)
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.relu(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):
super().__init__()
self.conv1 = nn.Conv2d(channels, channels, 3, padding=1)
self.bn1 = nn.BatchNorm2d(channels)
self.conv2 = nn.Conv2d(channels, channels, 3, padding=1)
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 two conv blocks and max pooling"""
def __init__(self, in_channels, out_channels):
"""Downsampling block with conv blocks, residual connection, attention, and max pooling"""
def __init__(self, in_channels, out_channels, dropout=0.1):
super().__init__()
self.conv1 = ConvBlock(in_channels, out_channels)
self.conv2 = ConvBlock(out_channels, out_channels)
self.conv1 = ConvBlock(in_channels, out_channels, dropout=dropout)
self.conv2 = ConvBlock(out_channels, out_channels, dropout=dropout)
self.residual = ResidualConvBlock(out_channels, dropout=dropout)
self.attention = CBAM(out_channels)
self.pool = nn.MaxPool2d(2)
def forward(self, x):
skip = self.conv2(self.conv1(x))
x = self.conv1(x)
x = self.conv2(x)
x = self.residual(x)
skip = self.attention(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):
"""Upsampling block with transposed conv, residual connection, attention, and conv blocks"""
def __init__(self, in_channels, out_channels, dropout=0.1):
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)
# After concat: out_channels (from upconv) + in_channels (from skip)
self.conv1 = ConvBlock(out_channels + in_channels, out_channels, dropout=dropout)
self.conv2 = ConvBlock(out_channels, out_channels, dropout=dropout)
self.residual = ResidualConvBlock(out_channels, dropout=dropout)
self.attention = CBAM(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 = F.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)
x = self.residual(x)
x = self.attention(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):
"""Improved U-Net style architecture for image inpainting with attention and residual connections"""
def __init__(self, n_in_channels: int, base_channels: int = 64, dropout: float = 0.1):
super().__init__()
# Initial convolution
self.init_conv = ConvBlock(n_in_channels, base_channels)
# Initial convolution with larger receptive field
self.init_conv = nn.Sequential(
ConvBlock(n_in_channels, base_channels, kernel_size=7, padding=3),
ConvBlock(base_channels, base_channels),
ResidualConvBlock(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)
self.down1 = DownBlock(base_channels, base_channels * 2, dropout=dropout)
self.down2 = DownBlock(base_channels * 2, base_channels * 4, dropout=dropout)
self.down3 = DownBlock(base_channels * 4, base_channels * 8, dropout=dropout)
self.down4 = DownBlock(base_channels * 8, base_channels * 16, dropout=dropout)
# Bottleneck
self.bottleneck1 = ConvBlock(base_channels * 8, base_channels * 16)
self.bottleneck2 = ConvBlock(base_channels * 16, base_channels * 16)
# 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)
self.up2 = UpBlock(base_channels * 8, base_channels * 4)
self.up3 = UpBlock(base_channels * 4, base_channels * 2)
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 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)
# Final refinement layers
self.final_conv = nn.Sequential(
ConvBlock(base_channels * 2, base_channels),
ResidualConvBlock(base_channels),
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
# Output layer with smooth transition
self.output = nn.Sequential(
nn.Conv2d(base_channels, base_channels // 2, kernel_size=3, padding=1),
nn.LeakyReLU(0.1, inplace=True),
nn.Conv2d(base_channels // 2, 3, kernel_size=1),
nn.Sigmoid() # Ensure output is in [0, 1] range
)
# Apply weight initialization
self.apply(init_weights)
def forward(self, x):
# Initial convolution
@@ -87,27 +199,26 @@ class MyModel(nn.Module):
x1, skip1 = self.down1(x0)
x2, skip2 = self.down2(x1)
x3, skip3 = self.down3(x2)
x4, skip4 = self.down4(x3)
# Bottleneck
x = self.bottleneck1(x3)
x = self.bottleneck2(x)
x = self.bottleneck(x4)
# Decoder with skip connections
x = self.up1(x, skip3)
x = self.up2(x, skip2)
x = self.up3(x, skip1)
x = self.up1(x, skip4)
x = self.up2(x, skip3)
x = self.up3(x, skip2)
x = self.up4(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 = F.interpolate(x, size=x0.shape[2:], mode='bilinear', align_corners=False)
# Concatenate with initial features for better detail preservation
x = torch.cat([x, x0], dim=1)
x = self.final_conv1(x)
x = self.final_conv2(x)
x = self.final_conv(x)
# Output
x = self.output(x)
x = self.sigmoid(x)
return x