:airplane: avion

October 2, 2023 · View on GitHub

AVION is short for A VIdeo model in ONe day. AVION (meaning plane in French and Spanish) is fast.

Training a Large Video Model on a Single Machine in a Day
Yue Zhao, Philipp Krähenbühl
UT Austin
arxiv | bibtex

Installation

See INSTALL.md to install this code.

Main results

AVION enables video-language contrastive pre-training on Ego4D (original narratives) on a single node of 8× consumer-grade GPUs within a day.

Method Backbone batch-size
per GPU GPU memory Hardware GPU×hour^ EK100 MIR
0-shot Avg. mAP
EgoVLP TSF-B 16 22 32× A100 1536 22.1
Ours ViT-B 256 19 8× A5000 130 27.4

^The reported GPU×hour is not normalized for GPU generations. The cost for EgoVLP is obtained from the original paper (Sec 6.1).
AVION speeds up LLM-augmented video-language contrastive pre-training (LaViLa) on Ego4D.

a. Pretraining cost and performance.

Method Backbone batch-size
per GPU GPU memory Hardware GPU×hour^ EK100 MIR
0-shot Avg. mAP
LaViLa TSF-B 32 25 32× V100 1824 30.9
Ours ViT-B 256 19 8× A5000 260 33.2

^The reported GPU×hour is not normalized for GPU generations.

b. Downstream performance.

Method Backbone EK100 MIR
Avg. mAP EK100 MIR
Avg. nDCG EK100 CLS
Action Top-1
LaViLa TSF-B 50.5 65.0 46.9
Ours ViT-B 51.7 66.8 49.5
LaViLa TSF-L 50.9 66.5 51.0
Ours ViT-L 54.5 69.0 54.5

:trophy: LaViLa+AVION helps us win CVPR 2023 EPIC-Kitchens Challenges in both Action Recognition and Multi-Instance Retrieval Tasks by a significant margin.
AVION speeds up VideoMAE pre-training.

Method Backbone Epochs GPU×hour^^ top-1/top-5 (w/. FT)
VideoMAE ViT-B 800 995 80.0/94.4
Ours ViT-B 800 583 80.1/94.5

^^Both GPU×hour are measured on the same hardware environment (4× A5000 GPU).

Method	Backbone	batch-size per GPU	GPU memory	Hardware	GPU×hour^	EK100 MIR 0-shot Avg. mAP
EgoVLP	TSF-B	16	22	32× A100	1536	22.1
Ours	ViT-B	256	19	8× A5000	130	27.4

Method	Backbone	batch-size per GPU	GPU memory	Hardware	GPU×hour^	EK100 MIR 0-shot Avg. mAP
LaViLa	TSF-B	32	25	32× V100	1824	30.9
Ours	ViT-B	256	19	8× A5000	260	33.2

Method	Backbone	EK100 MIR Avg. mAP	EK100 MIR Avg. nDCG	EK100 CLS Action Top-1
LaViLa	TSF-B	50.5	65.0	46.9
Ours	ViT-B	51.7	66.8	49.5
LaViLa	TSF-L	50.9	66.5	51.0
Ours	ViT-L	54.5	69.0	54.5

Method	Backbone	Epochs	GPU×hour^^	top-1/top-5 (w/. FT)
VideoMAE	ViT-B	800	995	80.0/94.4
Ours	ViT-B	800	583	80.1/94.5

For more details, please refer to MODEL_ZOO.

License

MIT License.

Acknowledgements

The vision-language contrastive pretraining part is refactored from LaViLa.
The MAE-style self-supervised pre-training part is built upon VideoMAE.

Citing AVION

@article{zhao2023training,
  title={Training a large video model on a single machine in a day},
  author={Zhao, Yue and Kr{\"a}henb{\"u}hl, Philipp},
  journal={arXiv preprint arXiv:2309.16669},
  year={2023}
}

@inproceedings{zhao2023lavila,
  title={Learning Video Representations from Large Language Models},
  author={Zhao, Yue and Misra, Ishan and Kr{\"a}henb{\"u}hl, Philipp and Girdhar, Rohit},
  booktitle={CVPR},
  year={2023}
}