Performance issue in cifar10_cnn_ict.py

[DLPerf]

Hello! Our static bug checker has found a performance issue in cifar10_cnn_ict.py: Python type argument w is passed to tf.function decorated function train_step instead of tensors. As there is a for loop argument epoch that has effect on w, every time the value of w changes, a new graph will be generated, and that can trigger tf.function retracing warning.

Here is the tensorflow document to support it.

Looking forward to your reply.

[DLPerf]

We are investigating this kind of issues, and your answer will be of great help to our work. Can you take a look? Thank you in advance! @ryancheunggit