Fix tests.

[dan-zheng]

WIP, one remaining test to fix: testUpSampling3D.

[Shashi456]

@dan-zheng is there no simple fix for this? Doesn't that mean upsampling 3d doesn't work effectively?

[dan-zheng]

There's no trivial fix, unfortunately. Please see TF-525 for more info, including a crash reproducer and a link to Keras's implementation. We can copy Keras to avoid the unsupported broadcasting error.

[Shashi456]

@dan-zheng so for using resize-volumes, would we have to write a vjp again? I see that the keras implementation uses resize_volumes, which again uses repeat elements, It might not be hard to fix, just looking at a way to do it properly.

[dan-zheng]

@dan-zheng so for using resize-volumes, would we have to write a vjp again? I see that the keras implementation uses resize_volumes, which again uses repeat elements, It might not be hard to fix, just looking at a way to do it properly.

Yes, it will be necessary to define some kind of "repeat elements" helper function that is differentiable.
Are you interested in taking this on? Let me share a quick prototype with you soon.

[Shashi456]

@dan-zheng yes I would be willing to, but I'm not accustomed to writing vjps for custom functions so that might act as some kind of bottleneck for me there. Still, have a lot more to learn.

Also i personally believe we should comment out Upsampling3d from layer.swift as well. No point keeping it open, when the functionality isn't available. Should i put a PR in for that? Might be quicker for you though.