Example `nlp/lstm_seq2seq.py` doesn't train with JAX backend

[keras-team]

Example `nlp/lstm_seq2seq.py` doesn't train with JAX backend

[fchollet]

Absolutely! It's a pretty hard issue I think. The problem is very non-obvious, everything looks good piecewise. TF and torch train well, but JAX trains very poorly (while still training anyway). The code is here:

https://gist.github.com/fchollet/f0c84ecbed8441e54820df8366a5a629

[qlzh727]

ok, I was able to verify my fix in #888 to fix this issue.

See https://colab.corp.google.com/drive/1z_QDD0uX9ApLJdFTxhHYxJNejUdgowkG#scrollTo=kxjMd749C1nA. Will send a PR very soon.

[fchollet]

I verified that the two backends give identical forward pass numerics for the model.

The saving issue was unrelated and I've now fixed it.

So there must be something different about either initializers or the optimization process somehow.

[fchollet]

I ruled out initialization. So it can only really be an optimizer issue or trainer issue.

[fchollet]

It trains fine in torch as well.

[shivance]

Hi @fchollet, can I contribute here ?

[ariG23498]

While working on the code the "tensorflow" backend threw an error.

Here is the Gist to reproduce the error: https://gist.github.com/ariG23498/b8b4c0912a0a19dfe2ef8b29b3160943

[fchollet]

The error message tells you that cuDNN can't be compiled to XLA, basically. It's somewhat tricky to solve on our side. Either we disable jit_compile if there's a cuDNN-enabled layer, or we don't use cuDNN if we detect we're tracing for XLA?

[fchollet]

To work around this you can just pass jit_compile=False to compile() when using TF.

[ariG23498]

To work around this you can just pass jit_compile=False to compile() when using TF.

Right! Now it trains using tf backend.

The same issue persists when loading the saved model it seems. Do I also need to pass compile=False when loading the saved model (when using TensorFlow as backend)?

[fchollet]

Do I also need to pass compile=False when loading the saved model (when using TensorFlow as backend)?

You can just set jit_compile = False on the model I think.

[fchollet]

@ariG23498 I have fixed it in this commit, please check that it works for you. c9bce12