Support for SmoothGrad

[vinutah]

Hi,
Can we expect to see an implementation of SmoothGrad.
https://arxiv.org/abs/1706.03825

Thanks!

[ruthcfong]

Hi @vinutah, sorry for the delayed reply.

We will add it to a list of todos for future work (unfortunately, I'm busy atm wrapping up my PhD).

If you'd like to implement, feel free to start a PR.

A minimal function should be pretty easy to write and have a definition like the following:

def smoothgrad(saliency_func, model, x, target, sigma=0.1, num_samples=50, **kwargs):
    # 
    # Generate noise tensor (i.e., x_ = x + noise)
    # Call saliency_func on the new noisy input, possibly replacing the gradient_to_saliency function with the identity function
    # Average returned saliency map / tensors