Hi @ericspod @wyli @Nic-Ma , could you please help to look at this draft PR? The forward function of BasicUNetDecoder requires a sequence of tensors (represents all feature maps from the encoder) as input, and the sequence length is variable.

However, this implementation does not meet the requirements of torchscript. Do you have any suggestions to fix this issue? Thanks!

Hi @yiheng-wang-nv the basic unet is only a simple implementation with a focus on the code readability for the basic architecture, I feel this PR will add a bit too much complexity

I think you can create a separate standalone module called MultiscaleDecoder as a network block.

Hi @yiheng-wang-nv the basic unet is only a simple implementation with a focus on the code readability for the basic architecture, I feel this PR will add a bit too much complexity

I think you can create a separate standalone module called MultiscaleDecoder as a network block.

Thanks @wyli for the suggestion, I removed the decoder and I will submit a new PR for the implementation. Now this PR is only used to fix #2759 #2760 and add some upsample related parameters into UpCat (these are non-breaking changes).