DynUNet needs updating for deep_supervision

[rijobro]

Describe the bug
Following this PR to DynUNet, the corresponding notebook needs updating.

I tried simply removing deep_supervision = True from the constructor, but this wasn't enough.

The subsequent error is attached here: dynunet.txt.

@Nic-Ma I think you might best know what needs to be changed here.

[saruarlive]

Dear @Nic-Ma @rijobro et al,

The dynunet returns list (containing the output tensor). Had it returned a tensor in validation mode, it would have been easier to use.
So I changed a bit in the sliding window inference code to make it work. What do you think?

  #seg_prob = predictor(window_data, *args, **kwargs).to(device)  # batched patch segmentation
  seg_problist2tensor = predictor(window_data, *args, **kwargs)
  seg_prob = seg_problist2tensor[0] if isinstance(seg_problist2tensor, list) else seg_problist2tensor
  seg_prob = seg_prob.to(device)

[Nic-Ma]

I will have a try to run the notebook tomorrow, thanks for your issue report.

Thanks.

[rijobro]

Possibly best to make any changes in this PR, as it will already have PEP8 modifications (and git merge is horrible wit jupyter notebooks)

[Nic-Ma]

Hi @saruarlive and @rijobro ,

I just submitted a commit to fix this issue in that branch: c61cfd1
@wyli , should we try to merge the notebook format PR ASAP? Otherwise, it's hard to develop new PRs and resolve conflicts.

Thanks.

[wyli]

Hi @saruarlive and @rijobro ,

I just submitted a commit to fix this issue in that branch: c61cfd1

@wyli , should we try to merge the notebook format PR ASAP? Otherwise, it's hard to develop new PRs and resolve conflicts.

Thanks.

sure, could you please help review the changes? @Nic-Ma @yiheng-wang-nv @ericspod
the review request is there for 5days #107

[zedtran]

Dear @Nic-Ma @rijobro et al,

The dynunet returns list (containing the output tensor). Had it returned a tensor in validation mode, it would have been easier to use.
So I changed a bit in the sliding window inference code to make it work. What do you think?

  #seg_prob = predictor(window_data, *args, **kwargs).to(device)  # batched patch segmentation
  seg_problist2tensor = predictor(window_data, *args, **kwargs)
  seg_prob = seg_problist2tensor[0] if isinstance(seg_problist2tensor, list) else seg_problist2tensor
  seg_prob = seg_prob.to(device)

@saruarlive This was helpful for me. I was finally able to get DynUNet working for the specified tutorial.

[yiheng-wang-nv]

Hi all, I just submitted a PR to update the DynUNet tutorial to meet the changes here. Now, the forward function will only return a single tensor (for both training and eval mode, with or without deep supervision).