- train=optim
- dataset.dataset=KMNIST
- model=PF
- train.batch_size=1024