Sean Moriarity

We are likely being more inefficient than TensorFlow somewhere. This might be related: https://github.com/elixir-nx/nx/issues/1003 One thing you can try is mixed precision in all of the models: ```elixir policy =...

Thanks for opening! We support instance norm in Axon so this should be straightforward enough to implement, I will take a look this weekend

I've added instance norm on master, but importing the VAE doesn't work, there's some strange things going on that don't seem to respect the ONNX spec. I will look further...

@jnnks Yes!

I believe it comes with ONNX runtime Python package!

@jnnks @NotQuiteLagom Sorry I let this one fall off. There is an open issue upstream to fix Axon's resize operator. I will add it in later as I have a...

@jnnks If the tests depend on non-constant inputs and raise, then it's okay to ignore them in that case. For a lot of functions we cannot implement the tests because...

We don't really have a choice, Nx does not support dynamic input shapes and so we need to enforce this constraint. In most cases it ends up being fine, there...

@jnnks I think you might be right, but is there a parameter we're not handling that might affect the behavior here?

If you try importing but specifying `mel_spectrogram_dynamic_axes_1: 1`, does the import work: `AxonOnnx.import("model.onnx", mel_spectrogram_dynamic_axes_1: 1)` For some deserializations to work we need to know shapes up front. If this does...