which version of keras is used by latest kws_streaming

[andrewmabc]

I'm trying to run kws_streaming's 00_check_data.ipynb/01_train.ipynb/02_inference.ipynb.

docker image tensorflow/tensorflow:nightly-gpu-jupyter is used to run the demo. And the following packages are installed via pip:
tf-model-optimization-nightly
tf-keras-nightly
tfa-nightly
scipy
pydot
graphviz

The following package is install via apt-get:
apt-get install graphviz

tf._keras_internal.utils.control_flow_util.smart_cond is replaced by tensorflow.python.keras.utils.control_flow_util.smart_cond.
tf._keras_internal.models._clone_layers_and_model_config is replaced by tensorflow.python.keras.models._clone_layers_and_model_config.
tf._keras_internal.models._clone_layer is replaced by tensorflow.python.keras.models._clone_layer.
tf._keras_internal.engine.functional.reconstruct_from_config is replaced by tensorflow.python.keras.engine.functional.reconstruct_from_config.

If TF_USE_LEGACY_KERAS=1, 01_train.ipynb can be run successfully. But the following calling is always running infinitely, i.e. don't stop. Why?

model_non_stream = utils.to_streaming_inference(model_non_stream_batch, flags, Modes.NON_STREAM_INFERENCE)

For the latest kws_streaming, tf.keras.backend.learning_phase() is called in several python files. It seems tf_keras should be used instead of keras 3.

If TF_USE_LEGACY_KERAS=0, 01_train.ipynb/02_inference.ipynb report error.

which version is used, keras 3 or tf_keras?

kws_streaming @rybakov

[Jacob-Bishop]

This may be useful #1972. There's an infinite loop in the clone_model call, i.e. tensorflow/tensorflow#61492

[andrewmabc]

In the issue(tensorflow/tensorflow#61492), the solution is to use keras 3 instead of tf_keras

For the latest kws_streaming, tf.keras.backend.learning_phase() is called in several python files. It seems tf_keras should be used instead of keras 3.

[andrewmabc]

@Jacob-Bishop can you run inference program without error now?