bhack

I suppose this is very similar to your initial post or not: https://colab.research.google.com/gist/bhack/f15b04a8181774c72254a8f72485fc4f/untitled129.ipynb

Yes it was just to understand your use case, but what I meant why you need to use `super().__init__(inputs=[], outputs=[])`? As it is seems to me unrelated from the ISSUE...

I think 1 run cause init accept kwargs. But I don't find a subclassed model definition like your (1) also in tests: https://github.com/keras-team/keras/blob/master/keras/tests/model_subclassing_test.py

Yes but the problem is that "de-facto" many Bazel based projects rely on that export: https://github.com/tensorflow/tensorflow/blob/master/.bazelrc#L158

> We could add a symlink to ~/bin and add that to PATH to keep that stable. The problem is that also in this case the cache will be invalidated...

I suppose that the Bazel/Tensorflow logic about exporting `PATH` env is that It could be a too sensitive env for reproducible builds E.g. see step 6 on Nvidia official guide:...

/cc @bamurtaugh This is going also to invalidate the prepared docker nighty Bazel cache produced by our CI and published RO on GCS when an user will try to consume...

It could be interesting to official support https://github.com/monder/goofys-docker directly in this repository.

I've opened a specific issue at https://github.com/naibaf7/libdnn/issues/18

I don't know if AMD could be interested to add a transparent backend for [HSA like this](https://github.com/libocca/occa/blob/master/src/HSAV2.cpp). Actually libdnn rely on viennacl, and there was some [official HSA backend initiative](https://github.com/HSAFoundation/viennacl-dev)...