Use cifar10 datasets get score 5.63 ?

sbarratt / inception-score-pytorch

Inception Score for GANs in Pytorch

MIT License

641 stars 125 forks source link

Use cifar10 datasets get score 5.63 ? #3

Closed odie2630463 closed 6 years ago

odie2630463 commented 6 years ago

Hi , i try cifar10 datasets get score 5.63 , but other code get about 12 Is there some thing wrong ? thanks !

sbarratt commented 6 years ago

See http://arxiv.org/abs/1801.01973

simopal6 commented 6 years ago

@sbarratt I read the paper but couldn't find an explanation to @odie2630463's question. Indeed, using your code I obtained an Inception score of 6.19 on CIFAR-10, but even you report results between 9 and 11...

sbarratt commented 6 years ago

See https://github.com/sbarratt/inception-score-pytorch/commit/3bcf441c7a3229985e3e18c7c03ea3984ff74836, which sets transform_input=False. That is required to get the same results.

bknyaz commented 6 years ago

My inception score for CIFAR-10 using this implementation is 9.86. Using normalization values as suggested in pytorch docs helped me to increase the score up to 10.56, but is still below the tensorflow results, which must be around 11.24 according to https://arxiv.org/pdf/1606.03498.pdf. I think the reason for this gap is that the model in pytorch makes slightly different predictions for the same images.