search

igul222 / improved_wgan_training

Code for reproducing experiments in "Improved Training of Wasserstein GANs"
MIT License
2.35k stars 668 forks source link

reproducing inception score on gan_cifar.py #92

Closed vibss2397 closed 3 years ago

vibss2397 commented 3 years ago

Hi, I am trying to reproduce the figure in the paper for inception score. After running your repository these are the results i get.

inception_score

Sure its reaching the score of 6 as shown in the paper, but it's not the same. So my question is does the random seed also matter, or was that figure the best you got after you ran a bunch of experiments.

igul222 commented 3 years ago

Looks like the same plot as the paper to me (https://arxiv.org/pdf/1704.00028.pdf Figure 3, left subplot, orange line). Note that the axis scales are different here.

vibss2397 commented 3 years ago

yeah close enough, actually I am trying to make an implementation in pytorch for this script and the inception score and so far with different initializations provided by pytorch I have been able to reach 5.6 score max. However, I have saved the weights you used in this script and hopefully i should reach up to 6. Thanks :)