The previous weights.pt was an epoch-1 checkpoint (step 878) and scored
macro-F1 0.674 on the PML IFCB test split. Replace it with the 20-epoch
naive-oversample run, seed 4200 (chosen by lowest validation loss, 0.2094;
W&B uofg/dinov3_PML_multi_linear_oversample, run robust-serenity-3).
Test split (14,837 images): accuracy 0.941, macro-F1 0.874, identical under
the evaluation notebook's sykepic metric and the repo's inference code.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
transformers 5 nests DINOv3 parameters under backbone.model.* where the saved
(4.x) weights have backbone.*, so load_state_dict failed with missing and
unexpected keys. Rename backbone keys only when the model expects the other
form; loading stays strict.
Verified identical preprocessing and predictions on 3,000 IFCB test images
under transformers 4.57.1 and 5.6.2.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add CLIP-LoRA classifier alongside the DINO classifier
