beit

Introduction

Official Repo

Code Snippet

BEiT (ICLR'2022)

@article{bao2021beit,
    title={Beit: Bert pre-training of image transformers},
    author={Bao, Hangbo and Dong, Li and Wei, Furu},
    journal={arXiv preprint arXiv:2106.08254},
    year={2021}
}

Results

ADE20k

Segmentor	Pretrain	Backbone	Crop Size	Schedule	Train/Eval Set	mIoU	Download
UperNet	ImageNet-22k-224x224	BEiT-B	640x640	LR/POLICY/BS/EPOCH: 3e-5/poly/16/130	train/val	53.12%	cfg | model | log
UperNet	ImageNet-22k-224x224	BEiT-L	640x640	LR/POLICY/BS/EPOCH: 3e-5/poly/16/130	train/val	56.52%	cfg | model | log

More

You can also download the model weights from following sources:

• BaiduNetdisk: https://pan.baidu.com/s/1gD-NJJWOtaHCtB0qHE79rA with access code s757