Questions about the code

[baumgach]

Dear Orobix team,

I am the first author of the VAGAN paper. I would like to thank you one more time for porting our code to PyTorch this has been immensely useful to the community working with these types of approaches! In fact, since TensorFlow 1 has become more or less obsolete people from my own team have started working with your code. This also motivated me to have a closer look at it. I have one potential issue and a question that I would like to bring to your attention:

1. If I read the code correctly, you are re-using the last batch from the discriminator update for the generator update. While I don't think this would be a real issue in practice, I think to be completely correct and avoid unwanted dependencies between the two optimisation steps, a new batch should be used for the generator update. (https://github.com/orobix/Visual-Feature-Attribution-Using-Wasserstein-GANs-Pytorch/blob/master/src/train.py#L204)

2. I do not understand why you set the requires_grad=True for the input images. This may be down to my lack of experience with PyTorch, though. What is the motivation for this?

Happy, to make a PR if you agree with my points.

Best wishes,
Christian

[dnlcrl]

Hey Christian,

First of all thank you very much! I am more than honored to learn that your team is using my code, I remember it took some time to really understand what was going from the paper and your code :D

I also understand your points and really don’t remember why I set that requires_grad=True, I think it should be fixed as you suggest.

The problem is that as I am no longer working at Orobix, probably I also cannot accept any PR as I am no longer in the Orobix organization. But my suggestion would be to send the PR, I will try to reach someone that will look into and accept it, I don’t think it would be a problem for anyone.

Thank you very much again!
Daniele

[baumgach]

Thanks a lot for the lightning fast response. I will do as you suggested.
Best of luck on your next endeavour!
Christian

[dnlcrl]

you're welcome, thanks a lot to you, also for your amazing work! :)

see you soon online!

Daniele