Results Reproducibility

[eslambakr]

Hello everyone,

Thanks for sharing your great work, however I couldn't manage to reproduce the results mentioned in the paper despite training from scratch using the config files provided in the repo and following the implementation details mentioned in the paper.

I tried different Auto-Encoders and also different datasets, e.g. LSUN-Churches and Bedrooms, and none of them converged to the same number mentioned in the paper.

For instance for LSUN-Churches I only got 13 FID score, despite it is mentioned it should be 4. Thus, your help is much appreciated as it seems there is another tricks or hyper-parameters such as LR should be set or tuned correctly to get these numbers mentioned in the paper.

Thanks in advance!