sbarratt/inception-score-pytorch

Unable to get the same score as the tensorflow version

xunhuang1995 opened this issue · 2 comments

xunhuang1995 commented

Thanks the authors for the contribution. Inception score for pytorch is quite useful for GAN community. However I'm not able to get the same score as the tensorflow version. The scores computed by this code are much lower. Have the authors check the correctness of the results?

sharma409 commented

It seems the popularly used tensorflow implementation of the inception score has a few bugs — we filed an issue here: openai/improved-gan#29

I believe this pytorch score is a correct implementation of the inception score as described in Section 4 of Salimans et al. (https://arxiv.org/pdf/1606.03498.pdf)

sharma409 commented

Unfortunately most results in the literature have been calculated using the tensorflow implementation. We plan on posting a note on arxiv pointing this out shortly

  • 👍2