Fix Keras 3 version check

[sampathweb]

keras-nlp has a function to check if TensorFlow uses Keras 3 via -
https://github.com/keras-team/keras-nlp/blob/master/keras_nlp/backend/config.py#L66

This works when using tf-nightly or TensorFlow 2.15, but breaks on TensorFlow 2.14 which is the default in Colab. To fix it, use import keras instead of from tensorflow import keras when doing this check.

Colab to repro the issue:
https://colab.research.google.com/drive/17AwBX5S29d-hjUDPLN8roHSNzrjXricG?usp=sharing

[mattdangerw]

I am not sure this actually keep any use case working, so might not be worth the extra complexity. tensorflow-text will use tf.keras, so if tf.keras is broken, our preprocessing is broken.

https://github.com/tensorflow/text/blob/b32645fbf1e4fd7e81d8d03fa2d2b4872e3a270d/tensorflow_text/python/keras/layers/todense.py#L28

I would somewhat consider this an expected fail state, and maybe instead do something like

try:
    from tensorflow import keras
except:
    raise ValueError(
        "Unable to import `keras` with `tensorflow`.  Please check your Keras and "
        "Tensorflow version are compatible; Keras 3 requires TensorFlow 2.15 or later. "
        "See keras.io/getting_started for more information on installing Keras."
    )

[sampathweb]

Addressed it. Can you check if the updates make sense?

[mattdangerw]

Looks fine!

I think we could also just keep the try/catch logic I posted above for simplicity. No version parsing. Might catch other weird edge cases we haven't thought of and still give the helpful link to keras.io. But ok with this too. Is there a reason we are avoiding trying from tensorflow import keras here?

[sampathweb]

Makes sense. Reverted back to try/except with from tensorflow import keras

[mattdangerw]

Looks fine!

I think we could also just keep the try/catch logic I posted above for simplicity. No version parsing. Might catch other weird edge cases we haven't thought of and still give the helpful link to keras.io. But ok with this too. Is there a reason we are avoiding trying from tensorflow import keras here?

[mattdangerw]

lgtm! thank you