search

aelnouby / Text-to-Image-Synthesis

Pytorch implementation of Generative Adversarial Text-to-Image Synthesis paper
GNU General Public License v3.0
405 stars 89 forks source link

The loss of generator #3

Closed pengliu380 closed 6 years ago

pengliu380 commented 6 years ago 
Hello, author. I want to reproduce the experimental results. When I trained the model, I found that the loss of generator was always high and it looked like the generator was much weaker than the discriminator, what should I do?
aelnouby commented 6 years ago

I am working now on better ways to make the training more stable including implementing WGAN, minibatch discrimination and feature matching. I will mention you in the pull request to let you know.

Meanwhile I guess you can to train the generator more than the discriminator if you think it is weaker, this trick is used heavily in WGAN but the discriminator is the one that is trained more because it can be overpowered easily.

aelnouby commented 6 years ago

@pengliu380 I have added WGAN implementation to master, however I am still working on making gradient penalty work but it seems there is a bug in pytorch right now, I have submitted an issue.

Nevertheless, the WGAN implementation should be working but it is quite slow, attached an example of the Wasserstien estimate plot.

newplot