TimmEncoderFixed¶

class TimmEncoderFixed(name, in_channels=3, depth=5, output_stride=32, drop_rate=0.5, drop_path_rate=0.0, *, pretrained=True)[source]¶

Fixed version of TIMM encoder that handles drop_path_rate parameter properly.

This encoder wraps TIMM models to provide consistent feature extraction interface for segmentation tasks. It extracts features at multiple scales from the encoder backbone.

Initialize TimmEncoderFixed.

Parameters:

  • name (str) – Name of the TIMM model to use as backbone.

  • in_channels (int) – Number of input channels. Default is 3.

  • depth (int) – Number of encoder stages to extract features from. Default is 5.

  • output_stride (int) – Output stride of the encoder. Default is 32.

  • drop_rate (float) – Dropout rate. Default is 0.5.

  • drop_path_rate (float | None) – Drop path rate of the encoder. Default is 0.0.

  • pretrained (bool) – Whether to use pretrained weights. Default is True.

Methods

forward

Forward pass through the encoder.

Attributes

out_channels

Get output channels for each feature level.

output_stride

Get the output stride of the encoder.

training

forward(x)[source]¶

Forward pass through the encoder.

Parameters:

x (torch.Tensor) – Input tensor of shape (B, C, H, W)

Returns:

List of feature tensors at different scales, including the input as the first element

Return type:

list[torch.Tensor]

property out_channels: list[int]¶

Get output channels for each feature level.

Returns:

Number of channels at each feature level

Return type:

list[int]

property output_stride: int¶

Get the output stride of the encoder.

Returns:

Output stride value

Return type:

int