[MRG] Sparse L1 Descent for PyTorch

[iamgroot42]

Implementing the Sparse L-1 Descent Algorithm in PyTorch.

[iamgroot42]

Implementing the Sparse L-1 Descent Algorithm in PyTorch, ported from the TF version.
All tests ported from TF version.

[nick-jia]

Hello @iamgroot42, thanks a lot for yout contribution and sorry for the late reply. I read your codes and it looks great! Just a few things I want to discuss with you:

1. Regarding sanity check, do you mind including the assertion at line 51 of cleverhans/future/torch/attacks/sparse_l1_descent.py in the asserts list, and set default sanity_checks to False?
2. I would suggest add grad_sparsity = torch.tensor(grad_sparsity) after line 61, since grad_sparsity may not have attribute 'shape'
3. Comparing to the tf version (see line 495 of cleverhans/utils_tf.py), the random initialization at line 88 of cleverhans/future/torch/attacks/sparse_l1_descent.py seems to miss the scaling step.

Thank you again!