mask generation for training and inference with MAR

[YOU-k]

Hi author, thanks for the great work and the general concept about 'masked autoregressive'. When I try to use the model provided, I find that during training, the mask is generated randomly without any auto-regression (the forward function in MAR class), but during inference, the auto-regression is applied.
Am I correct about this or do I understand the code wrongly?

[LTH14]

Thanks for your interest! Yes, during training, we perform masked image modeling training, which simulates one step in the auto-regression.