[DOC-FIX] TF autolog and non eager mode

[konstantin-frolov]

Willingness to contribute

No. I cannot contribute a documentation fix at this time.

URL(s) with the issue

https://mlflow.org/docs/latest/python_api/mlflow.tensorflow.html

Description of proposal (what needs changing)

Please add information about work tensorflow.autolog() and non eager mode for TF 2.
autolog() doesn't store any metrics if TF worked in non eager mode (compiled models).
I spend 2 hours for solve problem with not stored metrics. And I have no errors in any logs, just empty metrics in mlflow UI.
by default TF 2 eager mode enabled.

[chenmoneygithub]

@konstantin-frolov Thanks for raising the issue! It shouldn't be the case if you are using Keras model.fit(), for which graph mode is turned on by default, and we are explicitly supporting it. I suspect you are wrapping your whole training loop by tf.function, in which case you cannot get the numerics during training. Checking Keras code for how the compilation works as a reference: link

Could you share a reproducible github gist? We can take a closer look.

[github-actions]

@mlflow/mlflow-team Please assign a maintainer and start triaging this issue.