About the first experiment about denn HOT 3 CLOSED

Just don't understand this loss term:
wass = lambda y_true, y_pred: torch.mean(y_true * y_pred)
Could you give me some tips, thanks a lot

from denn.

Hi there,

1. The results are very sensitive to initialization (with random seed). But moreover, different computing architectures can differ in how they handle this randomness (e.g. I have observed differences between MacOS and Linux environments). This could be the case for you. I would try various random seeds to see if one of them converges. We have no guarantees for convergence with GANs, unfortunately.

2. The loss term is the Wasserstein loss function (where the labels are -1 and 1 instead of 0 and 1 as with vanilla GAN), see here: https://arxiv.org/abs/1701.07875.

from denn.

Going to close this is this is not a new/unknown issue, but one that has been observed and is detailed in the documentation (i.e. the importance of the platform).

from denn.