WeightNormalization data init fails with Keras experimental_run_tf_function

[seanpmorgan]

Recent Keras change has made it so the data dependent initialization fails:

_SymbolicException: Inputs to eager execution function cannot be Keras symbolic tensors, but found [<tf.Tensor 'sequential_1/weight_normalization/mul:0'

I'm submitting a hotfix which flags the model compile with experimental_run_tf_function=False. Another issue I'm tracking is tensorflow/probability#519 to see how TFP resolves their issue.

[WindQAQ]

This is related to #219. If switching back to .assign(), then it works. However, it won't converge.

[facaiy]

Hi, is it still a problem? I think I have some time to look into it.

[WindQAQ]

@facaiy Hi, it's still a problem. It would be great that you could take over it!

[facaiy]

OK, Tzu-Wei, I'll take a try