Does not work with eager mode

[qinghao1]

To reproduce:

1. Go to tutorials/Classification_Privacy.ipynb
2. Run the following

tf.compat.v1.enable_eager_execution()
assert tf.executing_eagerly()

3. Run the notebook. Accuracy does not increase with training, stays at 0.1. If I remove the above two lines and rerun, accuracy increases as expected.

[skeydan]

I can confirm on my own data. Running TF 2 or TF nightly, whether using subclassing or Keras, the losses and weights are all nan. I have to disable_v2_behavior to make it work.

It would be great if this could be fixed, thanks!

[carlodavid012]

what's the update to this? does tf-privacy still doesn't work on eager mode?