dynamic-network-architectures_translation
This repository adapts the dynamic-network-architectures for image-to-image translation tasks. It is based on the original repository found at MIC-DKFZ/dynamic-network-architectures.
Changes Made
- UNet Decoder Improvements: Added upsampling + convolution in the UNet decoder instead of transposed convolution to help prevent checkerboard artifacts, especially when using perceptual loss. cf this article.
Upsample_TrilinearUpsample_Nearest
Updates
- Added an argument to control the choice of UNet decoder : decoder_type ["standard", "trilinear", "nearest"]
TODO
- Test if
tanhis necessary after the last convolution in decoder (constrain to [-1 ; 1])