DhlinV
commented
2 weeks ago @isanztorres Sorry for the delay in responding due to my schedule with graduation and NeurIPS submission, the checkpoint we released is not the one reported in our paper but trained for longer, which is with better visual quality. If you want to reproduce the FID result, we can provide you the checkpoint and our training log. Here is the link: https://drive.google.com/file/d/1FSbG8Zvf8XLMUEZxxBBc7Puu-g5YhQc3/view?usp=drive_link [checkpoint] and https://drive.google.com/file/d/14fld9h8I1DLIjte1JDOQvS_URCiocHkG/view?usp=drive_link [log]
Hello! I've been trying to reproduce your result in the FID metric. To do so I've called your function calc_metrics.py like this:
python op3d/test_scripts/calc_metrics.py --network=op3d-ffhq-512-128-withc-12plane.pkl --gpus=1 --metrics=fid50k_full --data=FFHQ_512The network was downloaded from https://drive.google.com/file/d/1gR-NyfDLZGcPD1lmM_W1ptxatP4f1OSv/view. I downloaded the data FFHQ_512 from images1024x1024 https://drive.google.com/open?id=1u2xu7bSrWxrbUxk-dT-UvEJq8IjdmNTP. To prepare the dataset I did the last three steps of EG3D runme.py for preparing the dataset (mirroring dataset, downloading dataset and creating dataset zip). I evaluated the model three times, two with mirrored images and one without them and I've obtained an FID of ~7.95 in each of them. What should I change to obtain the same FID as you did in your paper?Thanks!