Sean Moriarity

Yes, definitely! Let me know if you run into any issues or have questions

With the introduction of namespaces and the new custom layer API there are a few additional open questions: 1. How do we deal with duplicate usage of the same param...

Hi @tiagodavi! Axon's implementation of BCE expects `y_true` to have a last dimension of size 1 (there's an explicit check for shape equality between `y_true` and `y_pred`). If you add...

Axon's accuracy should do that thresholding for you. What do you get if you just feed the result of `Axon.predict(model, trained_model, x_test, compiler: EXLA)` into Axon.Metrics.accuracy?

It is probably a bug, please send me the gist!

Overall more test coverage is required for every module. ## Axon - [x] Test layer creation for all layers - [x] Test `Axon.init` - [x] inside defn - [x] outside...

Axon's test coverage has increased greatly here, and despite some shortcomings of specific APIs (Axon.Loop), Axon is tested well against PT in other libraries, so I am closing this :)

@PhillippOhlandt Can you please try saving your training progress with `Axon.Loop.checkpoint` and then resuming training from the checkpoint using `Axon.Loop.from_state` now? If it works, then we can close this :)

@josevalim Most of the loss functions in Axon are not in ONNX either since they are mostly for training and ONNX is mostly focused on inference and converting models between...

@vans163 Just checking in. So we are aware of issues with the RNG and are working for fixes upstream in Nx. Is there anything in particular that is blocking this...