search

gtegner / mine-pytorch

Mutual Information Neural Estimation in Pytorch
MIT License
294 stars 56 forks source link

-inf or Zero loss values #3

Open Mojzaar opened 2 years ago

Mojzaar commented 2 years ago

Hi,

Thank you for providing your paper's code. I'm quite curious about how I will use MINE in my code. I add the output of the MINE torch model to the adversarial loss to train the generator (with Wasserstein loss). But my issue is that I either receive -inf (if I don't utilize gradient normalization) or values that are so near to zero (1e-5), implying that the generator output and the real image are utterly misaligned (am I right?) And this small loss will never differ from zero. Do you have any ideas about why this is occurring? Should Increase the complexity of the MINE network or iterate more on MINE loss?

anonymous102410 commented 2 years ago

Hi, Did you solve your problem? Could you tell me how to solve it? thanks