WIP: Adapt

[davidslater]

No description provided.

[lcadalzo]

I'm running the following code:

model = cifar10_model()
xy_batches, classes = cifar10_data()
attack = PGD_Linf(model)

for batch in xy_batches:
    x, y = batch
    x_adv = attack(x)

and getting the following error:

Traceback (most recent call last):
  File "run.py", line 9, in <module>
    x_adv = attack(x)
  File "/home/lucas/Desktop/gard/git/twosixlabs/armory/armory/adapt/pytorch.py", line 644, in __call__
    self.gradient()
  File "/home/lucas/Desktop/gard/git/twosixlabs/armory/armory/adapt/pytorch.py", line 582, in gradient
    self.loss = self.loss_fn(self.y_pred, self.y_target)
  File "/home/lucas/miniconda3/envs/armory-adapt/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/lucas/Desktop/gard/git/twosixlabs/armory/armory/adapt/pytorch.py", line 511, in forward
    return super().forward(torch.log(input), target)
  File "/home/lucas/miniconda3/envs/armory-adapt/lib/python3.8/site-packages/torch/nn/modules/loss.py", line 211, in forward
    return F.nll_loss(input, target, weight=self.weight, ignore_index=self.ignore_index, reduction=self.reduction)
  File "/home/lucas/miniconda3/envs/armory-adapt/lib/python3.8/site-packages/torch/nn/functional.py", line 2689, in nll_loss
    return torch._C._nn.nll_loss_nd(input, target, weight, _Reduction.get_enum(reduction), ignore_index)
RuntimeError: 0D or 1D target tensor expected, multi-target not supported

which I think occurs because self.y_target is an array of shape torch.Size([1, 10]) (the same one as self.y_pred) rather than an integer

[davidslater]

Thanks, @lcadalzo, I hadn't tested with model predictions.

[lcadalzo]

For the user story Initialize an attack with an interesting input, not sure I'm following exactly what's meant since you can pass whatever to ART's generate() or in the case here,__call__(). Or do you mean the init within the epsilon ball that gets added to the input? If so, it looks like this code restricts you to either random or None.

[davidslater]

Yes, the desire is to enable the init within the epsilon ball. The code needs to be modified to enable passing in an init function instead of those two choices.

[lcadalzo]

just pointing out a challenge here. While setting this task_metric is simple enough for classification, this would break for non-classification tasks and is much less straightforward to define in those cases

[davidslater]

Point taken

[lcadalzo]

is it necessary to have an additional class for patch pgd and random patch pgd, as opposed to this just being something set or defined in the former?

[davidslater]

I would like to collapse as much as possible - but wanted to start from simple attacks, move to complex ones, and then see how to combine them effectively.

[lcadalzo]

it seems like the number of distinct classes needed is multiplying rapidly. Would you also need a class for EoT L2, EoT Patch, EoT random patch, etc.?

[davidslater]

I should have added a note on this one: this was meant to demonstrate how to extend a class to use EoT. I would remove this class, actually. Or, if EoT can be added as a wrapper around an existing class, then that would prevent that sort of class explosion.

I need a better way of handling / representing random components - that would enable specification of what you're performing the expectation over.