Inception Model Snapshot Version

Hi!, I’ve noticed that here (in Chainer GAN lib) they use 1008 out classes while
in implementation provided by pytorch team used here, you use 1000 output classes. Messing around with comparing Inception scores I found that numbers do not match. I want to make some comparative study between models in their repo and mine (in pytorch). What would you suggest to do? By now I see the only way in somehow adopting chainer model, but I’d like to aviod it.