Nondeterministic in `HoVerNet`

[KumoLiu]

In the current HoVerNet implementation of MONAI, it uses UpSample with NONTRAINABLE mode which used torch.nn.functional.interpolate() may cause nondeterministic.

MONAI/monai/networks/nets/hovernet.py

Lines 535 to 537 in d0db5fd

	self.upsample = UpSample(
	2, scale_factor=2, mode=UpsampleMode.NONTRAINABLE, interp_mode=InterpolateMode.BILINEAR, bias=False
	)

https://pytorch.org/docs/stable/generated/torch.use_deterministic_algorithms.html#torch.use_deterministic_algorithms

Describe the solution you'd like
It may need to be explained in the docstring.

[KumoLiu]

cc @JHancox @drbeh, could you please double confirm it?

[JHancox]

@KumoLiu Is this non-determinism specific to Upsample? Also, doesn't it use torch.nn.Upsample rather than torch.nn.functional.interpolate?

[KumoLiu]

Yes, it uses torch.nn.functional.interpolate() in torch.nn.UpSample.
https://github.com/pytorch/pytorch/blob/b136f3f310aa01a8b3c1e63dc0bfda8fd2234b06/torch/nn/modules/upsampling.py#L156

[wyli]

thanks, I'm labelling this a feature request to update the docstring for both hovernet and UpSample