Update for DynUNet's changes and fix get_feature_maps issue

[yiheng-wang-nv]

Fixes #106 and #128 .

Description
Due to the forward function's changes of DynUNet, this commit
update the corresponding places, as well as the loss calculation
for trainer.

In addition, the DiceCEloss has been implemented in MONAI, thus the
self-designed loss function part has also been updated.

Signed-off-by: Yiheng Wang vennw@nvidia.com

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[rijobro]

@yiheng-wang-nv thanks! Do you think you could add some comments explaining the shape of the tensor returned by forward? I'm guessing it's the torch.unbind that removes the feature maps? Could you add a comment to that affect, and mention that it's only necessary when deep_supervision==True?

[rijobro]

fixes #128.

[yeeyangtee]

Hi, somehow am getting labels of shape (2, 1, 40, 56, 40) for taskid = 4.

This causes an error with non-matching dimensions between predictions (4) and targets(5).

I'm guessing onehot_y should be = False and need to squeeze labels. Unless some of the decathlon tasks do not have one-hot-labels

[rijobro]

Hi @coranium this seems like a new issue to me. Could you create a new Discussion, preferably with a minimal working example.

[yiheng-wang-nv]

Hi all, I opened a new draft PR, see #131 , since the current tutorial version is too old, which may have many issues, I rewrite a new python file based pipeline, and have tested on all decathlon tasks locally. Since my local version does not have enough doc strings, and used DataParallel rather than DistributedDataParallel for multi-gpu training, to meet the latest MONAI version, the PR is still in processing.
Please take a look at #131 for further updates, thanks!

[wyli]

Hi @yiheng-wang-nv do you have an ETA for the new pipeline? if it'll take some time, is it possible to provide a quick fix for #128 first?

[yiheng-wang-nv]

Hi @yiheng-wang-nv do you have an ETA for the new pipeline? if it'll take some time, is it possible to provide a quick fix for #128 first?

I see, let me reopen this PR, and create another one to distinguish the new pipeline. The new pipeline will be finished next week.

[yiheng-wang-nv]

@yiheng-wang-nv thanks! Do you think you could add some comments explaining the shape of the tensor returned by forward? I'm guessing it's the torch.unbind that removes the feature maps? Could you add a comment to that affect, and mention that it's only necessary when deep_supervision==True?

Hi @rijobro , I've added the comments, could you please help to review it again? As for the new pipeline I mentioned above, I'll open another branch and submit a PR after finishing it.

[wyli]

thanks it looks good to me

[rijobro]

@yiheng-wang-nv added comments look good, thanks!