maple-research-lab/AdPE ? reverse-engineered prompt

Build me the full training code for this AdPE project, so I can pretrain and fine tune a Vision Transformer with masked image modeling on ImageNet. I want one main script for pretraining and one for finetuning, with support for both absolute and relative positional embeddings, plus the adversarial versions that perturb either the embedding positions or the coordinates during training.

It should be able to run the MAE+ baseline too, not just the adversarial setup, and I want the usual training stuff handled cleanly, like distributed training, gradient accumulation, masking ratio, learning rate, warmup, and saving checkpoints and logs. Please make it work with the ImageNet folder structure the README describes, and include the pieces needed for the RPE acceleration if that is part of the project.

If anything is missing from the docs, look up the current usage online and make the scripts runnable end to end with sensible defaults.