Hi @wyli @SachidanandAlle ,

There is a potential issue in this PR that I am not very sure how to fix and want to discuss with you guys here:
As DeepGrow is slightly different from regular training, it needs to run several iterations inference / interactions, so we need to inference -> decollate -> click transforms -> collate -> inference as what I changed in this PR.
But some random click transforms may output data with different size, so we can't collate them into a batch for next inference, like the "guidance" generated by AddRandomGuidanced here: https://github.com/Project-MONAI/tutorials/blob/master/deepgrow/ignite/train.py#L98

Can we pad the data and collate? Or any other idea to avoid collate?

Thanks in advance.

/black

This fix looks ok to me.. we can Vish to verify sample 2D deepgrow training.. for accuracy

This fix looks ok to me.. we can Vish to verify sample 2D deepgrow training.. for accuracy

Hi @SachidanandAlle ,

Thanks for your review, do you mean to use pad_collate to solve the size issue when collating again?
@wyli @rijobro what do you guys think?

Thanks in advance.

/black

Hi @wyli ,

Thanks for @SachidanandAlle 's analysis and help, this PR is ready for review now.
Could you please help also review it?

Thanks in advance.