Quality of cifar-resnet conditioned generated outputs

igul222 / improved_wgan_training

Code for reproducing experiments in "Improved Training of Wasserstein GANs"

MIT License

2.35k stars 668 forks source link

Quality of cifar-resnet conditioned generated outputs #66

Open ysharma1126 opened 6 years ago

ysharma1126 commented 6 years ago

I trained the cifar-resnet for 25,000 iterations, took a few days.

About 30-50% of the generator outputs for each label are quality. Has anyone else experienced this? Any tips on improving that percentage? Has anyone obtained better results by training for the full 100,000 iterations?

a411919924 commented 4 years ago

I trained it for 100k iterations, but the result seems not good. The measured FID score is 10.19.