Kyle Daruwalla

This goes against #1483, so I don't think we should merge without a discussion. The implementation is good other than @CarloLucibello's suggestion though. Are the errors from #1483 no longer...

I think having at least a warning here is important. It's a more complex issue than the kind of promotion that occurs for convs, because of the type of the...

Should we close this issue? We have the proposed API via Optimisers.jl now which has a mechanism for handling in-place updates within an immutable API.

Yeah upsampling is non-trivial to get right and be performant on the GPU as well (last time I tried it, I had to ask in `#gpu` on Slack to get...

> Btw, I think it would be better to try to remove ONNX.jl from the general registry and use a name like OnnxFlux.jl to clearly state that it translates between...

@gxyd Take a look at [the PR](https://github.com/FluxML/Metalhead.jl/pull/70) linked above. Someone already posted a training script (I haven't had the time to check if it works). I would just ping that...

This will also be important to ease the maintenance burden of libraries with pre-trained weights like Metalhead.jl. Otherwise we'll need to manually translate on any breaking change to Flux that...

These functions already exist in MLDatasets.jl. Metalhead.jl was intentionally stripped down to just the models to reduce feature duplication across packages. Instead, the tutorial should be updated to use MLDatasets.jl.

> using MLDatasets: trainimgs and using MLDatasets: valimgs do not work either This should be `using Metalhead.CIFAR10: traintensor, testtensor` (or `traindata`/`testdata` to get the images + labels. The updated tutorial...

What about gradients here? Not that we were testing this before.