billpsomas/efficient-probing-heads

Efficient Probing heads

Trained EP (efficient probing) heads for all 37 frozen encoders of the standing ImageNet-1k benchmark at billpsomas/efficient-probing (ICLR 2026). Each head is a multi-query cross-attention pooling + BatchNorm + linear classifier, trained for 90 epochs with LARS on frozen features. No backbone weights are included -- each config.json records the exact --model / loader flags to rebuild the encoder from its original source.

Why two kinds of head: the training code initially kept only a rolling per-epoch checkpoint, so a finished run left its final epoch behind rather than its best. 12 early-peaking models were later re-run with best-epoch checkpointing, so their heads are the peak and reproduce the leaderboard number exactly. Every file's metadata records both its own accuracy at the saved epoch and the table's best-epoch figure, so nothing has to be taken on trust.

Loading

from huggingface_hub import hf_hub_download
import torch

path = hf_hub_download("billpsomas/efficient-probing-heads",
                       "dinov3_vit7b/ep_head.pth")
ck = torch.load(path, map_location="cpu", weights_only=False)
head_state, meta = ck["state_dict"], ck["meta"]
# with the benchmark repo on PYTHONPATH:
#   model = backbones.build_backbone(args, device)   # args from meta
#   probe_heads.build_probe_head(model, args)
#   model.head.load_state_dict(head_state, strict=True)

Or evaluate directly with the benchmark's tool:

python tools/eval_reimagenet.py predict <backbone flags from config.json> \
    --head_ckpt ep_head.pth --pred_out preds.json

Heads

encoder	EP variant	top-1 @ saved epoch	epoch	checkpoint is
DINOv3 ViT-7B/16	ep_all	88.36	6	peak
MetaCLIP2 ViT-bigG/14-378	ep	88.12	6	peak
EVA02-CLIP E-14-plus	ep	87.98	6	peak
EVA02-CLIP E-14	ep	87.70	6	peak
SigLIP2 SO400M/14	ep	85.64	29	final epoch
PE-Core L-14/336	ep	87.25	12	peak
MetaCLIP2 ViT-bigG/14	ep	87.11	6	peak
SigLIP2 ViT-L/16	ep	87.06	6	peak
DINOv3 ViT-L/16	ep_all	86.73	19	final epoch
AIMv2 ViT-L/14	ep	85.62	19	final epoch
SigLIP ViT-L/16	ep	85.93	6	peak
DINOv2 ViT-L/14	ep_all	85.56	15	peak
Franca ViT-L/14	ep_all	84.28	14	peak
DINOv3 ViT-B/16	ep_all	83.77	20	final epoch
DINOv2 ViT-B/14	ep	83.61	25	final epoch
RADIO ViT-L/16	ep	83.40	89	final epoch
EVA02 ViT-L/14	ep	83.22	89	final epoch
CLIP ViT-L/14	ep	83.22	11	peak
CAPI ViT-L/14	ep	82.43	89	final epoch
BEiTv2 ViT-B/16	ep	81.32	89	final epoch
RADIO ViT-B/16	ep	80.26	89	final epoch
iBOT ViT-L/16	ep_all	79.43	89	final epoch
Hiera ViT-H/16	ep	79.82	89	final epoch
MAE ViT-L/16	ep	79.43	89	final epoch
I-JEPA ViT-H/14	ep	78.80	89	final epoch
iBOT ViT-B/16	ep_all	78.62	89	final epoch
Hiera ViT-L/16	ep	78.51	83	final epoch
CLIP ViT-B/16	ep_all	77.85	11	peak
DINO ViT-B/16	ep_all	77.08	89	final epoch
MoCov3 ViT-B/16	ep_all	76.21	89	final epoch
Hiera ViT-B/16	ep	75.63	88	final epoch
MAE ViT-B/16	ep	75.35	86	final epoch
MaskFeat ViT-B/16	ep	71.68	89	final epoch
MaskFeat ViT-L/16	ep	69.56	89	final epoch
SimMIM ViT-B/16	ep	64.81	89	final epoch
MAE ViT-S/16	ep	64.56	89	final epoch
DiT DiT-XL/2	ep	56.94	86	final epoch

Full provenance (training logs, exact commands, the leaderboard itself) lives in the GitHub repo. Heads were trained on ImageNet-1k; use accordingly.