probe network need to re-train head for each task?

According your paper, I think for a specific task, the probe network should retrain, but the code in `models.py`, it seems no need to train the resnet-*.

Read the paper referenced in the repository description and inspect models.py, focusing on how the probe network and resnet-* are trained for a specific task. Determine whether the implementation matches the paper's expected retraining behavior; done means the discrepancy is resolved or clearly documented.

Written by the indexing model from the issue text.