search

VinAIResearch / CREPS

Efficient Scale-Invariant Generator with Column-Row Entangled Pixel Synthesis (CVPR 2023)
https://vinairesearch.github.io/CREPS/
GNU Affero General Public License v3.0
29 stars 1 forks source link

How to cal metrics on FFHQ1024x1024? #2

Closed FerryHuang closed 1 year ago

FerryHuang commented 1 year ago

Thanks for providing all these clear guides. However I'm stucked with the metrics calculation on FFHQ1024x1024. Following the data preparation I run dataset_tool.py to generate a ffhq-1024x1024.zip using downloaded FFHQ1024x1024(around 90G so I confirm it's right 1024 ver.). Then I just call cal_metrics.py with ckpt creps-ffhq-512x512.pkl. I just got a error report like "OSError: Image files do not match the specified resolution". So where didi git it wrong?

thuanz123 commented 1 year ago

Hi @FerryHuang. With the latest update, you now can calculate metrics with images at resolution different to the resolution of the generator. For example:

python calc_metrics.py --metrics=fid50k_full --data=~/datasets/ffhq-1024x1024.zip --mirror=1 \
    --network=creps-ffhq-512x512.pkl --scale=2

This command will compute FID score between dataset at resolution 1024x1024 and images at resolution 1024x1024 generated by CREPS trained at resolution 512x512

thuanz123 commented 1 year ago

More generally, you can use this command

python calc_metrics.py --metrics=fid50k_full --data=~/datasets/ffhq-MxM.zip --mirror=1 \
    --network=creps-ffhq-NxN.pkl --scale=M/N

in case you need to compute metrics between a dataset at resolution MxM against CREPS trained at resolution NxN.