Can I use Feature Adversaries - PyTorch with GPU?

[matanle51]

Hi, I defined a pytorch classfier:

criterion = torch.nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=0.01)
classifier = PyTorchClassifier(
model=model,
clip_values=(0, 1),
loss=criterion,
optimizer=optimizer,
input_shape=(3, 32, 32),
nb_classes=10,
device_type='gpu')

and a Feature Adversaries attack in the following way:

attack = FeatureAdversariesPyTorch(
classifier,
layer=-1,
delta=8/255,
optimizer=None,
step_size=1/255,
max_iter=500,
)

I pass it a source and a guide:

x_test_adv = attack.generate(data, guide)

But when I try to run it, I'm getting the following error:

`13, in FeatureAdversariesPyTorch.generate(self, x, y, **kwargs)
211 guide_batch = torch.tensor(y[begin:end])
212 x_adversarial[begin:end] = self._generate_batch(source_batch, guide_batch).numpy()
--> 213 return np.array(x_adversarial, dtype=x.dtype)

TypeError: Cannot interpret 'torch.float32' as a data type`

It seems that it's looking for numpy data type, but I pass it tensors since I work on GPU.
Any idea on how to solve this issue?

[beat-buesser]

Hi @matanle51 Yes, you should be able to use Feature Adversaries with GPUs. What is the type of your variable data and guide?

[matanle51]

Hi @beat-buesser ,
If I try to pass the data as numpy arrays, I get the following message:
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

This is since the model is on GPU and the arrays are on cpu.

[matanle51]

@beat-buesser
BTW -same error when trying to run your notebook example:
https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/a0411c54f7e0471696f6baeb2c73dea73beedb08/notebooks/attack_feature_adversaries_pytorch.ipynb

[beat-buesser]

Hi @matanle51 I think you have discovered a bug!

These lines in FeatureAdversariesPyTorch are missing a transfer to the device:

adversarial-robustness-toolbox/art/attacks/evasion/feature_adversaries/feature_adversaries_pytorch.py

Lines 210 to 211 in a0411c5

	source_batch = torch.tensor(x[begin:end])
	guide_batch = torch.tensor(y[begin:end])

The solution is to add .to(self.estimator.device) at the end of each line. Can you try if it solves the issue for you too?

I'll open an issue for this bug and schedule it for the next patch release ART 1.13.1.

[matanle51]

@beat-buesser Thank you.

Yes, I changed lines 210-211, and it solved the problem.

source_batch = torch.tensor(x[begin:end]).to(self.estimator.device)
guide_batch = torch.tensor(y[begin:end]).to(self.estimator.device)

[beat-buesser]

Hi @matanle51 Thank you!