adversarial_training_mnist - Problem

[leeahn1223]

Hello.
I have just started to use ART library.
When I run "adversarial_training_mnist" program, I encountered some errors.
Please let me know how to resolve the problem as shown below.

Regards, gyu lee.

[beat-buesser]

Hi @leeahn1223 Which version of Keras and TensorFlow are you using?

[beat-buesser]

I'm not completely sure, but it could be that the loaded model might not work with TensorFlow v2. Could you please try again with tensorflow=1.15.5 and keras==2.3.1?

[leeahn1223]

I have just tried to downgrade my tensorflow and keras as you said.
I am not sure it was done well. The result is as shown below.

After finishing downgrading as shown above, I retried to run "adversarial_training_mnist" but it was also failed.

[leeahn1223]

I have tried many ART programs here as well as "adversarial_training_mnist". However nothing was succesful. I didn't know how difficut using ART is. I wonder if these errors occured only to me?

[beat-buesser]

Theoretically it should not be that difficult to use ART. Are you running on Windows?

[leeahn1223]

Please, give me your frank answer as an engineer who has been using ART.
I have to decide if I study and use ART or not.
If I should fight many errors to use ART, I will give up using ART because I think ART is not ready to meet users.
I have been studying Adversarial AI and searching for tools which can be helpful.

[Foxglove144]

Hi @leeahn1223, I was able to get the adversarial_training_mnist notebook working on Tensorflow 2.12.0 by disabling eager execution.

Add these lines:

import tensorflow as tf
    
tf.compat.v1.disable_eager_execution()