check compatibility with Mooncake

[CarloLucibello]

Check that all our layers and operators are compatible with Mooncake.jl AD

[Parvm1102]

Hi @CarloLucibello ,

I have done some progress on this: I've set up the test suite for the spatial convolutional layers comparing Mooncake and Zygote.

Results: 16 / 20 conv layers are compatible (935 passed tests).
However, there are errors with these two layers:

• DConv (3 errors): Fails during reverse-mode differentiation (might be related to Wrong implementation DConv #592).
• ChebConv (1 error): Crashes during gradient cache preparation before the test executes.

2 layers were excluded: CGConv and GMMConv, same as existing tests.

How would you like me to proceed? Should I dig into finding workarounds for DConv and ChebConv, or should I move on to testing the other layer categories (pool.jl, heteroconv.jl, etc.)? I can also open a PR if you want to see the code.

[CarloLucibello]

Hi @Parvm1102, thanks.
I suggest you implement something similar to the test infrastructure of Flux (see the test_gradients function)
https://github.com/FluxML/Flux.jl/blob/master/test/test_utils.jl#L82

Then open a PR with just the conv layers tested and we will discuss there.