About Imagenet256 Latent

[weixiang-zhang]

Dear Author,

Great Work! I am working on reproducing this paper but I met some problems. I noticed that the input in cofig.py is imagenet256_latents_wds. I download the imagenet256_latent_lmdb from MaskDIT official repo. After processing it with Imdb2wds.py (MaskDIT repo), I got imagenet latent webset data. I used --resolution 32, --num_channels' 4, but the dimension I got is still (8,32,32) rather than (4,32,32) and then I failed to run this repo.

Is there any problems for my operation? Thank you

[weixiang-zhang]

Or, would you mind sharing /export/group/datasets/ILSVRC/imagenet256_latents_wds in config.yaml to help us reproduce your paper? Thank you!

[flixmk]

Thank you for trying out TREAD. We are not allowed to reupload the imagenet dataset unfortunately. But you can also just use the normal ImageNet and preprocess the information or directly use the produced VAE latents to train the model.