RandZoom on batch

[bhashemian]

Is your feature request related to a problem? Please describe.
RandZoom uses Zoom which uses torch.nn.functional.interpolate under the hood to extract randomly zoomed images. Although torch.nn.functional.interpolate is capable of being applied on batch x channels x [optional depth] x [optional height] x width, Zoom and RandZoom don't.

Instead, they treat the channel dimension as spatial dimension. This is causing an issue since there is no error and warning if user apply it on a batched dataset.

This requires to distinguish between 3D multi-channel images vs 2D multi-channel batched images. Although an automatic logic would be hard to achieve, a simple flag can do the work.

[wyli]

the concept of minibatch is part of the model optimization using stochastic gradient descent, anything that dealing with minibatches should be in the 'monai.networks' module of monai.

The price for mixing the concepts of batch and nonbatch is not just a simple flag, it requires a full set of dimensionality input validation, and potentially reduces the usability/code readability.

[bhashemian]

Hi @wyli, isn't it a too restrictive assumption?
What about validation and testing (inference in general)? There is no optimization there but mini-batch is required for computational efficiency.

[wyli]

interestingly in pytorch 1.10 torch.nn module starts to support no batch dimensions see (github/pytorch/issues/47149)

[bhashemian]

interestingly in pytorch 1.10 torch.nn module starts to support no batch dimensions see (github/pytorch/issues/47149)

Yes, that's interesting but it is the opposite of the problem that we have here.

[bhashemian]

Do we have any decision here? Is there any technical issue that prevents us from supporting batch transform for RandZoom?
Maybe we can have a discussion after we are done with this release.
Thanks

[bhashemian]

I'm closing this since random zoom is implemented in cuCIM so the need is satisfied for now although the general discussion remains unresolved and can be discussed in the future if needed.