DecoderBlock¶

class DecoderBlock(in_channels, skip_channels, out_channels)[source]¶

Decoder block for UNet++ architecture.

This block performs upsampling and feature fusion using skip connections from the encoder. It consists of two convolutional layers with ReLU activation and optional attention mechanisms (not implemented).

conv1¶

First convolutional block applied after concatenating input and skip features.

Type:

Conv2dReLU

conv2¶

Second convolutional block for further refinement.

Type:

Conv2dReLU

attention1¶

Attention mechanism applied before the first convolution (currently Identity).

Type:

nn.Module

attention2¶

Attention mechanism applied after the second convolution (currently Identity).

Type:

nn.Module

Example

>>> block = DecoderBlock(in_channels=128, skip_channels=64, out_channels=64)
>>> input_tensor = torch.randn(1, 128, 64, 64)
>>> skip = torch.randn(1, 64, 128, 128)
>>> output = block(input_tensor, skip)
>>> output.shape
... torch.Size([1, 64, 128, 128])

Initialize DecoderBlock.

Creates two convolutional layers and optional attention modules for feature refinement during decoding.

Parameters:

  • in_channels (int) – Number of input channels from the previous decoder layer.

  • skip_channels (int) – Number of channels from the skip connection.

  • out_channels (int) – Number of output channels for this block.

Methods

forward

Forward pass through the decoder block.

Attributes

training

forward(input_tensor, skip=None)[source]¶

Forward pass through the decoder block.

Upsamples the input tensor, concatenates it with the skip connection (if provided), and applies two convolutional layers with attention.

Parameters:

  • input_tensor (torch.Tensor) – (B, C_in, H, W). Input tensor from the previous decoder layer.

  • skip (torch.Tensor | None) – (B, C_skip, H*2, W*2). Skip connection tensor from the encoder. Defaults to None.

  • self (DecoderBlock)

Returns:

(B, C_out, H*2, W*2). Output tensor after decoding and feature refinement.

Return type:

torch.Tensor