I ran IS for the birds dataset, I noticed it is non deterministic. For the same input, output scores change for different runs. I am not sure if the inception network is in test mode, maybe dropout is being applied?

Additionally, for the birds dataset, I got an IS of 5.1324 for the real data (training set) and 22.7079 for the real data (test set). Which I think is weird, I thought it would be the other way around. Any idea why is that?

Thanks