Unable to load_model !!!! about adversarial-robustness-toolbox HOT 5 CLOSED

Hi, @Luke06! Please see the following comment about how to download h5 files without corrupting their integrity: h5py/h5py#757 (comment) Let me know if this solves the issue. Also, the model was created with keras==2.2.2; please consider using the same version to make sure that you can load it.

from adversarial-robustness-toolbox.

Hi
I looked at the comment but I don't see how it fix my problem , Iam using the std keras library to load the model and the file is they one from the ART repo that I retrieve thru a wget in my python notebook. the downloaded file seems to be correct. I am in Keras 2.2.4 ! but cannot move back to 2.2.2 due to others incompatibilities in my code :(. in fact my goal is to retrain my model using your defence training libraries but it was not working neither so....

from adversarial-robustness-toolbox.

@Luke06 I have tested your scenario. Loading the model does not work for me neither if I download it using wget. Actually, the model from the repo and the one downloaded with wget don't even have the same size on disk (I get 44357 for wget, just the same as you, and 2737408 for the model downloaded properly). That is what the comment I pointed you towards explains. You could clone the repo or use the solution in the comment to make sure you get a valid version of the file.

On a different note, if your goal is to train an architecture that you already have with the defenses from the library, you don't necessarily need to download the pre-trained models that we use as examples.

from adversarial-robustness-toolbox.

OK I understand the problem with the downloaded h5 file.
Yes my initial goal was to retrain my model using the defences from the libraries but I was encountering some issues probably due to invalid usage of the library, here is the code I tried to use
I have defined a CNN model called model and I create a Classifier wrapper :
Original_Classifier = KerasClassifier((0,1),model,use_logits=False)
Original_Classifier.fit(x_train, y_train, batch_size=128, nb_epochs=10)

Then I tried this lines that I found in the notebook sample called adversarial-training-mnist.ipynb

attacks = BasicIterativeMethod(Original_Classifier,eps=0.3,eps_step=0.01, max_iter=40)
trainer = AdversarialTrainer(Original_Classifier,attacks,ratio=1.0)
trainer.fit(x_train,y_train, nb_epochs=10,batch_size=50)

but the training ended after the 1 epochs with very strange results.
Can you pls provide me with the right way & sample code of doing that for my training against Fast Gradient Method attack ?
by advance thanks a lot

from adversarial-robustness-toolbox.

I would say that your original problem (.h5 files) is solved, so I will close this issue.

I don't see any obvious error in your code example. Would you mind opening a new issue for it, where you would copy the code example, as well as the full stack trace of the error you are getting when running it?

from adversarial-robustness-toolbox.