Fix Boundary Attack not working on MPS accelerator

[philswatton]

In the code introduced in #43, we explicitly prevent boundary attacks from being performed on MPS accelerators. This is because of foolbox's dependencies, which at some point create a numpy array (numpy has a default dtype of float64) which is then passed to eagerpy. eagerpy then passes this numpy array to torch.torch.as_tensor(a, device=self.raw.device), where self.raw.device will be "mps" if the input tensor to foolbox was on MPS. torch.as_tensor will inherit the float64 dtype from numpy (and numpy's default cannot be changed), causing an error when it attempts to create a float64 dtype tensor on MPS.

Any of the following could act as solutions:

• finding an alternative to foolbox that does not depend on eagerpy
• numpy allowing the default dtype to be easily altered in a script
• MPS becoming float64-friendly
• foolbox allowing arguments to the numpy array's creation (possibly we could wrap our own class around foolbox)

Links to the various steps to the final error:

• foolbox BoundaryAttack.py page: https://github.com/bethgelab/foolbox/blob/12abe74e2f1ec79edb759454458ad8dd9ce84939/foolbox/attacks/boundary_attack.py
  • the releveant ArrayQueue class (class declaration is on the same page) being created in BoundaryAttack: https://github.com/bethgelab/foolbox/blob/12abe74e2f1ec79edb759454458ad8dd9ce84939/foolbox/attacks/boundary_attack.py#L152
  • theArrayQueue.mean() method which creates the numpy array in question then passes it to eagerpy: https://github.com/bethgelab/foolbox/blob/12abe74e2f1ec79edb759454458ad8dd9ce84939/foolbox/attacks/boundary_attack.py#L317
• eagerpy from_numpy() function: https://github.com/jonasrauber/eagerpy/blob/5b3b37a1e8c49aa3d34b3ed87a5075a20fd50f2f/eagerpy/framework.py#L205
• the eagerpy PyTorchTensor.from_numpy() method which calls torch.as_tensor() where the error occurs: https://github.com/jonasrauber/eagerpy/blob/master/eagerpy/tensor/pytorch.py#L288