Resize Convolution

Use resize-convolution instead of transposed convolution for upsampling in the U-Net decoder.

The U-Net architecture requires upsampling in its decoder pathway. This option controls how that upsampling is performed.

Resize Convolution (Enabled - Default)

First resizes the feature maps using interpolation, then applies a convolution. This approach avoids checkerboard artifacts that can occur with transposed convolutions.

Transposed Convolution (Disabled)

Uses learned upsampling through transposed (deconvolution) layers. Can produce sharper results but may introduce checkerboard patterns.

Recommendations

• Keep enabled (default) for most cases

• The resize-convolution approach generally produces smoother, artifact-free results

• Only disable if you observe overly smooth results and want to experiment