sgad-checkpoints

Trained predictor checkpoints for the SGAD paper (frozen-encoder video anomaly detection over codebook targets). Companion to the codebooks at sgad-iclr/sgad-codebooks; every checkpoint here needs its codebook from there. Checkpoints are the unmodified last.ckpt written by training (+trainer.save_checkpoint=true): model (frozen DINOv2-S/14 backbone + prototypes + predictor), cfg, epoch, global_step, optimizer, so they both evaluate and resume (+trainer.resume_from=<path>, constant LR only). Embedded cfg paths are node-local and informational.

Download to bulk storage, never into a repo tree:

uvx hf download sgad-iclr/sgad-checkpoints --repo-type dataset --local-dir /path/on/bulk/storage/checkpoints

ipad/dinov2abs_euc_k32768/ — iPAD, K=32768 euclidean codebook, 16 properties

Protocol (the paper's Ours row with K=32768 instead of 8192): DINOv2-S/14 img224, window nf6 / ctx4 / frame_step 4 / stride 1, self-attention predictor width 384 depth 4, euclidean K=32768 codebook (ipad/dinov2abs_euc/<prop>_vit_small_patch14_dinov2_lvd142m_img224_k32768.pt on sgad-codebooks), constant LR 1e-5 (optimizer.lr_base=5e-6), 120 epochs, 1 GPU. Values are final-epoch (ep119) frame AUROC at mean pooling (val/frame_auroc/mean), micro over all test frames. Files: <prop>_absabs_K32768euc_d4_120ep_constlr1e-5_<wandb_run_id>.ckpt, W&B project feature_space_video_anomaly_detection/sgad-ipad.

prop	run id	ep119 frame AUROC	sha256[:12]
R01	3pfaatuw	0.8691	168decdcb254
R02	fsh0l5zo	0.9435	078b0dfbde85
R03	17h5v2si	0.5635	9a27f1a9a69e
R04	jg6jpo8m	0.8002	6906ff8ee4e3
S01	xpkx37at	0.7877	c8ebf9722679
S02	rzaqxejc	0.6001	0cf1a5002e34
S03	tnfu4y89	0.9197	75df44ab64d1
S04	qwnbgvb4	0.6949	ee3975057531
S05	yvz4pkur	0.9818	fd74f9683026
S06	el1yswl5	0.5730	109cc27cf5a1
S07	yq1lwlen	0.7382	ab1a76414bc3
S08	gqulabpl	0.8136	efc012c0d6e0
S09	nn05gscl	0.7058	1c2bbf088642
S10	koh4oeuo	0.8369	e245a86cfd24
S11	rclhv50u	0.6950	340b37ba72a2
S12	rsb6gbh0	0.7128	2a859c495d53
mean (16)		0.765

Eval-only reproduction (research repo scripts/analysis/ipad/eval_ckpt_frame_scores.py, one GPU, ~5 min):

CUDA_VISIBLE_DEVICES=0 uv run --no-sync python scripts/analysis/ipad/eval_ckpt_frame_scores.py \
    --prop R01 --ctx 4 --nf 6 \
    --ckpt <checkpoints>/ipad/dinov2abs_euc_k32768/r01_absabs_K32768euc_d4_120ep_constlr1e-5_3pfaatuw.ckpt \
    --cb <codebooks>/ipad/dinov2abs_euc/r01_vit_small_patch14_dinov2_lvd142m_img224_k32768.pt \
    --out r01_scores.parquet

tempglitch/dinov2abs_euc_k32768/ — TempGlitch, K=32768 euclidean codebook

Protocol (the paper's TempGlitch Ours row, t057 recipe): DINOv2-S/14 img448, raw frame step 12 (5 fps from 60 fps), window nf6 / ctx4 (4 context + 2 predicted frames), stride 4, self-attention predictor width 384 depth 4, euclidean K=32768 codebook per glitch type (tempglitch/dinov2abs_euc/<type>_vit_small_patch14_dinov2_lvd142m_img448_k32768.pt on sgad-codebooks), constant effective LR 3e-5 (optimizer.lr_base=1.5e-5, global batch 16 = 4 GPUs x 4), 120 epochs (resumed in segments; constant LR makes the resume exact), 80/20 normal split (tempglitch_8020). Value = W&B val/video_auroc/mean__mean at ep119 (video AUROC, mean over tokens then over clips). W&B project feature_space_video_anomaly_detection/games-tempGlitch-ad, tag t057_tg448.

type	run id (final segment)	ep119 video AUROC	sha256[:12]
blinking	0vc8rp9y	0.7809	49efb6612c97