VAD-JEPA Model Checkpoints

Pretrained temporal model weights for the VAD-JEPA project — online video anomaly detection on traffic dashcam footage.

Variants

Model	VCL	NF	Type	Best Epoch
SG-SlotSSM (Sparse Gated)	64	4	frozen	50
SG-SlotSSM (Sparse Gated)	64	4	finetuned	14
SG-SlotSSM (Sparse Gated)	28	4	frozen	50
SG-SlotSSM (Sparse Gated)	28	4	finetuned	50
SG-SlotSSM (Sparse Gated)	16	4	finetuned	100
SG-SlotSSM (Sparse Gated)	8	4	frozen	50
SG-SlotSSM (Sparse Gated)	8	4	finetuned	100
SlotSSM (Dense)	64	4	frozen	50
SlotSSM (Dense)	64	4	finetuned	10
SlotSSM (Dense)	28	4	frozen	50
SlotSSM (Dense)	28	4	finetuned	60
SlotSSM (Dense)	16	4	finetuned	100
SlotSSM (Dense)	8	4	frozen	50
SlotSSM (Dense)	8	4	finetuned	100
Mamba	28	4	finetuned	30
Mamba	16	4	finetuned	100
Mamba	8	4	finetuned	100
Linear Probe	12	8	finetuned	150
Linear Probe	16	12	finetuned	150
Linear Probe	8	4	finetuned	150

Usage

Download a checkpoint and use it with the VAD-JEPA repo:

python main.py --config cfgs/vjepa_sparse_slotssm.yaml --phase test --epoch 50

Each folder contains checkpoints/model-{epoch}.pt.