IJEPAForImageClassification

Reference: Assran et al., arXiv:2301.08243, Appendix A.2 — the paper evaluates by linear probing the frozen target encoder, which is what this is. The target encoder cannot be trained through this class: its parameters are frozen where they are built, so an optimiser handed everything still only moves the head.

It carries the whole pretraining model — both encoders and the predictor — because that is what a checkpoint holds and what continuing to pretrain would need. Inference uses one of the three, so an exported package is about three times the size it has to be.