how to reproduce inception / FID scores?

[jcpeterson]

Which scores in the table are from your code? Do you have the commands to reproduce them exactly?

[jcpeterson]

bump

[takerum]

Sorry for the late reply.

All scores showed in README (except for the scores in the "official" column) are calculated with the code in this repository. The scores for each method are calculated at the end of training, using the following extension functions: https://github.com/pfnet-research/chainer-gan-lib/blob/master/common/evaluation.py#L82 https://github.com/pfnet-research/chainer-gan-lib/blob/master/common/evaluation.py#L154

We cannot reproduce the scores exactly because we did not specify the random seed when we trained each model, but you should be able to get similar scores for each method by running the script in https://github.com/pfnet-research/chainer-gan-lib/blob/master/example.sh

[jcpeterson]

I'm getting varying scores that doesn't always preserve the rank in the table. Are these supposed to be averaged to say 10 runs of training a model? They don't seem reliable at all otherwise.