Why even in the era of TF2.x does ART not like eager execution enabled? Why stick to TF1.x style code?

[swapnilsayansaha]

classifier = KerasClassifier(model=model, clip_values=(0, 1), use_logits=False)

This line assumes you have eager execution disabled. However, for custom models, disabling eager execution leads to a lot of error. For example: tensorflow/tensorflow#39908

Why does ART still rely on out-of-date TF1.x style code?

[beat-buesser]

Hi @swapnilsayansaha ART does not rely on out-of-date TF1.x style code as you call it. For the era of TF2.x, also as you call it, and because ART likes eager execution we have created art.estimators.classification.TensorFlowV2Classifier. but kept KerasClassifier for backward-compatibility.

[christian-westbrook]

Hey @beat-buesser! This context is really helpful, and should probably be made more apparent in the documentation, especially if there's no plan to update KerasClassifier to work with TF2. As a beginner, it would have been nice to know that I should use TF2 for a Keras model even though a KerasClassifier class exists.

[beat-buesser]

Hi @christian-westbrook Good point, where would be a good place to provide this information?

[swapnilsayansaha]

Thanks. That function provided a lot of help!