Closed zx1301 closed 4 years ago
Inception scores are extremely sensitive to the exact model you use to compute them.
It looks like the code you link is using the torchvisvion inception_v3 model; however most people who report inception scores (including my paper) typically use the Inception v1 model from Google. I used the original OpenAI code to compute inception scores:
https://github.com/openai/improved-gan/blob/master/inception_score/model.py
Hello,
I'm trying to replicate the Inception Scores shown in the sg2im paper and I have been using a version of this: https://github.com/sbarratt/inception-score-pytorch/blob/master/inception_score.py by giving the predicted and ground-truth image tensors as input. When I do this, my results are significantly lower than the inception scores in the paper. Can someone provide any guidance for how I should try to calculate the inception scores so that I can replicate the numbers in the Sg2im paper?