Oops, must ensure that tf_keras is installed.

Have I mentioned how much I dislike our non-hermetic Python setup in Bazel? I might fix that soon.

Thanks Ryan. Should we also add this for all other files that use Keras? Or does issue only occur for this one?

Also is it not an issue for the CI because we have only one keras version in those testing environments?

Thanks Ryan. Should we also add this for all other files that use Keras? Or does issue only occur for this one?

AFACT, the problem only rears its head when Keras >= 3 is visible alongside (older versions of?) Tensorflow. I'm not 100% sure of the circumstances in which this happens. Perhaps it only happens with some of our Bazel targets due to varying dependency sets and Bazel's sandboxing 🤷‍♂️.

I had Keras 3 in my environment, because I was doing model compression with TensorFlow Model Optimization Toolkit, which has a dependency on Keras 3.

Maybe the workaround in this PR isn't the right solution. I can think of at least two other possibilities:

1. Upgrade the version of TF Bazel's Python environment brings in, and see if that fixes the incompatibility.
2. Configure the Bazel Python toolchain to be hermetic, so the user's environment can't affect it.

Those seemed like changes with possibly farther-reaching impact, hence my affinity for the workaround. I'll experiment with it a bit more and follow up.

Ugh.

Turns out I was wrong about the underlying problem. The workaround does indeed avoid importing Keras, but that "fixes" a different underlying problem. The underlying problem is that Keras imports the module psutil for whatever reason, and psutil has a crashing bug. See the backtrace below.

psutil is imported if and only if it happens to be in the environment, and it happened to be in my environment. Uninstalling psutil fixes everything without the TF_USE_LEGACY_KERAS=1 workaround.

There appears to be no problem with Tensorflow coexisting with Keras >= 3. In fact, Keras >= 3 has been in our Bazel Python dependencies alongside Tensorflow 2.18 for some time now without any issues.

I will withdraw this PR, and open or reopen an issue about configuring Bazel's Python toolchain to be hermetic.

Backtrace:

[....]
  File "/tmp/bazel-working-directory/tflite_micro/bazel-out/k8-fastbuild/bin/python/tflite_micro/runtime_test.runfiles/tflm_pip_deps_keras/site-packages/keras/src/saving/saving_lib.py", line 37, in
 <module>
    import psutil
  File "/home/rkuester/.local/lib/python3.10/site-packages/psutil/__init__.py", line 95, in <module>
    from . import _pslinux as _psplatform
  File "/home/rkuester/.local/lib/python3.10/site-packages/psutil/_pslinux.py", line 25, in <module>
    from . import _psposix
  File "/home/rkuester/.local/lib/python3.10/site-packages/psutil/_psposix.py", line 50, in <module>
    'Negsignal', {x.name: -x.value for x in signal.Signals}
AttributeError: module 'signal' has no attribute 'Signals'