Anton Smirnov
Anton Smirnov
My bad, I forgot that you don't need `KernelGradients` now, so it installed an old version. With the updated code it works.
Superseded by #627 as it covers more cases and fixes CI. Thanks!
Kind ping on the status of this PR as it will unblock support for AMDGPU backend (and we can specialize batchnorm using MIOpen for which I plan to open a...
Yeah, should've done that at the beginning :) I've figured that if both encoder blocks and decoder blocks in my code end with BatchNorm, I might as well construct everything...
Indeed, it comes from the `push!`. However, if you replace `identity` activation with any other activation function (e.g. `relu`), the error disappears. But in `MBConv` the last `BatchNorm` has no...
Similar thing happens if you replace `BatchNorm` with `Conv`. And the error disappears if you specify non-identity activation function. MWE: ```julia using Flux function encode(encoder, x) features = typeof(x)[] for...
Yes, I've changed the feature extraction part (encoder) for the model to use `map` instead of loops and now can take gradients. Although, having support for `push!`, in this case,...
This is likely because of atomic operations, which rely on Atomix: https://github.com/FluxML/NNlib.jl/blob/83df6426d5a669754399bc4a9d8920c4b52e1a76/src/scatter.jl#L112 IIUC, Metal does not yet support Atomix.
> and I would wait for KA 0.9 That was my next question :) Then we also need to wait on https://github.com/JuliaGPU/AMDGPU.jl/pull/374 for AMDGPU.
`AMDGPU#master` now supports KA 0.9, I'll attempt this.