[train_vqvae] generated images look blank

[kimdn]

trafficstars

I downloaded some FFHQ images from https://drive.google.com/drive/folders/1tZUcXDBeOibC6jcMCtgRRz67pzrAHeHL

Then, I ran python /people/kimd999/script/python/cryoEM/vq-vae-2-pytorch/train_vqvae.py /people/kimd999/MARScryo/dn/data/devel/FFHQ/data

Since it is a default running, --size should be 256 (line 142 of train_vqvae.py). I think that @rosinality used 256 size as well.

However, trained images look blank https://drive.google.com/file/d/1reILjem27LlcNsptUFEFoy-VCDLz9Poq/view?usp=sharing

Can anyone guess why?

[rosinality]

If your directory structure is something like FFHQ/data/, then you should set path to FFHQ, not FFHQ/data.

(Also you need to set permission of google drive to allow public access.)

[kimdn]

If your directory structure is something like FFHQ/data/, then you should set path to FFHQ, not FFHQ/data.

-> My folder (.../FFHQ/data) has 'checkpoint', 'sample' and 'drive-download-20200823T020941Z-001' folders. 'drive-download-20200823T020941Z-001' folder has 22 images

(Also you need to set permission of google drive to allow public access.)

-> Now I changed to "Anyone on the internet with this link can view". Thank you for pointing out.

I ran again this time with 3,000 images (instead of 22 images only). This time, I could see regenerated images as in https://drive.google.com/file/d/1yHEDhJpKm2Aa9GZi4xCre0Dl0_A2Loqf/view?usp=sharing

Therefore, I think that at least 3,000 > x > 22 number of images are needed for proper running.