ZitongYu/PhysFormer ? reverse-engineered prompt

Build me a working version of PhysFormer for facial video based heart rate estimation.

I want to be able to train it on the VIPL HR dataset, then run inference on a single test video to predict the rPPG signal and estimate heart rate. It should include the data loading, training script, test or inference script, and a simple way to evaluate the heart rate error from the predicted signal.

Please make it runnable with the usual Python setup, and keep the commands simple so I can train and test without having to figure out the whole pipeline myself. If you need to check current package details or best practices, look them up online first. Also make sure the code is organized so I can easily plug in my own checkpoint and test data later.