Inference Time Problem and Model Storage Problem

[Fyzjym]

Thank you for your contribution to HTG and sharing the code.

Q1:

When calculating the FID, is it generating a dataset that is the same as the test set and then calculating the distance between the two. For example, fid (TestSet_iam & TestSet_WordStylist) The reason for this problem is that WordStylist takes a lot of time to generate images, it takes 30s to reason 1000 steps for one image, and the complete test set is about 25k images. This means that testing FID will take about 8 days.

Q2:

I trained WordStylist using 3 V100s, however the size of the stored model A does not match the size of the model B you published. When inferencing, I got an error (some layers are missing) using Model A. I can inference normally using Model B that you have made public. Is this because I am training with multiple cards?

Thank you again and look forward to your reply. :D