Can I easily create a TensorFlowV2Classifier that uses the same training step as a KerasClassifier?

[christian-westbrook]

Hello everyone!

I want to attack a model defended by an ART defense with attacks from three different frameworks: ART, CleverHans, and Foolbox.

The KerasClassifier class is the easiest for me to work with. but it requires eager execution to be disabled. ART attacks are compatible with this, and CleverHans attacks are compatible with this, but Foolbox attacks require eager execution.

The solution I have in mind would be to use a TensorFlowV2Classifier, as this seems to support eager execution, but I notice that you have to define your own training step for a TensorFlowV2Classifier. I'd really like to use a logically equivalent training step to Keras or a KerasClassifier, but I'm not sure how to do that. If anyone could help me to create a TensorFlowV2Classifier with a training step that closely resembles or is equivalent to the Keras or KerasClassifier training step, I would be really grateful!

[christian-westbrook]

Hello!

I wanted to add that I've recently discovered this discussion: #1906

I had no idea that art.estimators.classification.TensorFlowV2Classifier was created to support eager execution, or that KerasClassifier is only maintained for backward compatibility. It would have been pretty useful to know that upfront, and I'd recommend updating documentation to reflect that.