search

torch / torch7

http://torch.ch
Other
8.97k stars 2.38k forks source link

How to compute d(gradInput)/d(weight)? (or second order derivative) #998

Open lifelongeek opened 7 years ago

lifelongeek commented 7 years ago

In "Improved Training of Wasserstein GAN", they use gradient norm penalty term. (https://arxiv.org/pdf/1704.00028.pdf)

This requires computing gradient of norm(Model.gradInput) w.r.t. Model parameters, and then accumulate to the gradParams of the Model.

Is this computation available with current torch's implementation? Or do you plan to update any related features?

Thanks for developer's hard work.