The following Python packages are required: numpy, torch, torchvision, tqdm.

An example job is

python main.py --seed 0 --optim fromage --initial_lr 0.01

See inside batch.sh for the settings used in the paper.

• The self attention block implementation is originally by https://github.com/zhaoyuzhi.
• The FID score implementation is by https://github.com/mseitzer/pytorch-fid.
• Thanks also go to Jiahui Yu.

This repository (exluding the fid/ subdirectory) is made available under a CC BY-NC-SA 4.0 license.