BEVT: BERT Pretraining of Video Transformers

Rui Wang¹, Dongdong Chen², Zuxuan Wu¹, Yinpeng Chen², Xiyang Dai², Mengchen Liu², Yu-Gang Jiang¹, Luowei Zhou², Lu Yuan^2
1Shanghai Key Lab of Intelligent Info. Processing, School of Computer Science, Fudan University, ²Microsoft Cloud + AI

This repository hosts the official PyTorch implementation of the paper: "BEVT: BERT Pretraining of Video Transformers".

Abstract

This paper studies the BERT pretraining of video transformers. It is a straightforward but worth-studying extension given the recent success from BERT pretraining of image transformers. We introduce BEVT which decouples video representation learning into spatial representation learning and temporal dynamics learning. In particular, BEVT first performs masked image modeling on image data, and then conducts masked image modeling jointly with masked video modeling on video data. This design is motivated by two observations: 1) transformers learned on image datasets provide decent spatial priors that can ease the learning of video transformers, which are often times computationally-intensive if trained from scratch; 2) discriminative clues, i.e., spatial and temporal information, needed to make correct predictions vary among different videos due to large intra-class and inter-class variations. We conduct extensive experiments on three challenging video benchmarks where BEVT achieves very promising results. On Kinetics 400, for which recognition mostly relies on discriminative spatial representations, BEVT achieves comparable results to strong supervised baselines. On Something-Something-V2 and Diving 48, which contain videos relying on temporal dynamics, BEVT outperforms by clear margins all alternative baselines and achieves state-of-the-art performance with a 71.4% and 87.2% Top-1 accuracy respectively.

Main Results on Downstream Tasks

Something-Something V2

Backbone	Pretrain	Tokenizer	acc@1	#params	FLOPs	Views	config	model
Swin-B	ImageNet-1K + K400	DALL-E	70.6	89M	321G	1x3	config	ToDo
Swin-B	ImageNet-1K + K400	PeCo	71.4	89M	321G	1x3	config	ToDo

Kinetics-400

Backbone	Pretrain	Tokenizer	acc@1	#params	FLOPs	Views	config	model
Swin-B	ImageNet-1K + K400	DALL-E	80.6	88M	282G	4x3	config	ToDo
Swin-B	ImageNet-1K + K400	PeCo	81.1*	88M	282G	4x3	config	ToDo

Note:

• BEVT uses the visual tokenizer of pretrained VQ-VAE from DALL-E or PeCo.
• PeCo is only pretrained on ImageNet1K and uses the same codebook size as in DALL-E.
• BEVT does not need labels during pretraining.
• ^* BEVT can achieve 81.5% Top-1 accuracy on Kinetics-400 when using PeCo tokenizer for pretraining and finetuning for 100 epochs.

Usage

Installation

Please refer to install.md for installation.

We use apex for mixed precision training by default.

Data Preparation

Please refer to data_preparation.md for a general knowledge of data preparation.

We use Kinetics-400 annotation files k400_val, k400_train from Video Swin Transformer.

BEVT Pretraining

Install DALL-E package before training:

pip install DALL-E

Download DALL-E tokenizer weight before training:

TOKENIZER_PATH=/path/to/save/dall_e_tokenizer_weight
mkdir -p $TOKENIZER_PATH
wget -O $TOKENIZER_PATH/encoder.pkl https://cdn.openai.com/dall-e/encoder.pkl
wget -O $TOKENIZER_PATH/decoder.pkl https://cdn.openai.com/dall-e/decoder.pkl

Set tokenizer_path in the config file. For example, configs/recognition/swin/swin_base_patch244_window877_bevt_in1k_k400.py:

tokenizer_path = '/path/to/save/dall_e_tokenizer_weight'

First, pretrain the image stream of BEVT (Swin-base) on ImageNet-1K (800 epochs). The pretrained model of image stream could be downloaded at google drive.

Then pretrain two stream of BEVT on ImageNet-1K and K400 (initialized from swin transformer pretrained with the image stream) with 32 GPUs (150 epochs):

bash tools/dist_train.sh configs/recognition/swin/swin_base_patch244_window877_bevt_in1k_k400.py --work-dir OUTPUT/swin_base_bevt_twostream --cfg-options total_epochs=150 model.backbone.pretrained='/path/to/save/swin_base_image_stream_pretrain.pth' --seed 0 --deterministic

The pretrained model of BEVT could be downloaded at google drive.

BEVT Finetuning

Finetune BEVT model on K400 with 8 GPUs:

bash tools/dist_train.sh configs/recognition/swin/swin_base_patch244_window877_bevt_finetune_k400.py --work-dir OUTPUT/bevt_finetune/swin_base_bevt_finetune_k400 --cfg-options model.backbone.pretrained='OUTPUT/swin_base_bevt_twostream/latest.pth' --seed 0 --deterministic --validate --test-best --test-last

Finetune BEVT model on SSv2 with 8 GPUs:

bash tools/dist_train.sh configs/recognition/swin/swin_base_patch244_window1677_bevt_finetune_ssv2.py --work-dir OUTPUT/bevt_finetune/swin_base_bevt_finetune_ssv2 --cfg-options model.backbone.pretrained='OUTPUT/swin_base_bevt_twostream/latest.pth' --seed 0 --deterministic --validate --test-best --test-last

To Do

• Release joint pretraining code
• Release fine-tuning code
• Release pretrained model
• Release finetuned model
• Release image stream pretraining code

Acknowledgements

This code is based on mmaction2 and Video Swin Transformer.

Citation

@inproceedings{wang2021bevt,
  title={BEVT: BERT Pretraining of Video Transformers},
  author={Wang, Rui and Chen, Dongdong and Wu, Zuxuan and Chen, Yinpeng and Dai, Xiyang and Liu, Mengchen and Jiang, Yu-Gang and Zhou, Luowei and Yuan, Lu},
  booktitle={CVPR},
  year={2022}
}

@article{dong2021peco,
  title={PeCo: Perceptual Codebook for BERT Pre-training of Vision Transformers},
  author={Dong, Xiaoyi and Bao, Jianmin and Zhang, Ting and Chen, Dongdong and Zhang, Weiming and Yuan, Lu and Chen, Dong and Wen, Fang and Yu, Nenghai},
  journal={arXiv preprint arXiv:2111.12710},
  year={2021}
}