WGAN-gp still not working


I’m trying to implement the gradient penalty in wgans. From what I understood, once (https://github.com/pytorch/pytorch/pull/1643) was merged with master it should be working. I installed from source today, and I’m getting a segmentation fault when I’m calling torch.autograd.grad. That being said, I’ve never used said function before, so maybe I’m doing it wrong. Any case, any help is welcomed :slight_smile:


(code) https://github.com/pclucas14/WassersteinGAN/blob/master/main.py

if you can create a branch with dummy data, so that i can clone and reproduce the segmentation fault, i will take a look.
Thank you for reporting this.

Hi Soumith,

Thanks for replying. I got away with it by using someone else’s wgan-gp implementation. However, if you still want to investigate the segmentation fault, let me know and I’ll make a branch.