vit

Vision Transformer

An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

Introduction

Official Repo

Code Snippet

Abstract

While the Transformer architecture has become the de-facto standard for natural language processing tasks, its applications to computer vision remain limited. In vision, attention is either applied in conjunction with convolutional networks, or used to replace certain components of convolutional networks while keeping their overall structure in place. We show that this reliance on CNNs is not necessary and a pure transformer applied directly to sequences of image patches can perform very well on image classification tasks. When pre-trained on large amounts of data and transferred to multiple mid-sized or small image recognition benchmarks (ImageNet, CIFAR-100, VTAB, etc.), Vision Transformer (ViT) attains excellent results compared to state-of-the-art convolutional networks while requiring substantially fewer computational resources to train.

Usage

To use other repositories' pre-trained models, it is necessary to convert keys.

We provide a script vit2mmseg.py in the tools directory to convert the key of models from timm to MMSegmentation style.

python tools/model_converters/vit2mmseg.py ${PRETRAIN_PATH} ${STORE_PATH}

E.g.

python tools/model_converters/vit2mmseg.py https://github.com/rwightman/pytorch-image-models/releases/download/v0.1-vitjx/jx_vit_base_p16_224-80ecf9dd.pth pretrain/jx_vit_base_p16_224-80ecf9dd.pth

This script convert model from PRETRAIN_PATH and store the converted model in STORE_PATH.

Results and models

ADE20K

Method	Backbone	Crop Size	Lr schd	Mem (GB)	Inf time (fps)	Device	mIoU	mIoU(ms+flip)	config	download
UPerNet	ViT-B + MLN	512x512	80000	9.20	6.94	V100	47.71	49.51	config	model | log
UPerNet	ViT-B + MLN	512x512	160000	9.20	7.58	V100	46.75	48.46	config	model | log
UPerNet	ViT-B + LN + MLN	512x512	160000	9.21	6.82	V100	47.73	49.95	config	model | log
UPerNet	DeiT-S	512x512	80000	4.68	29.85	V100	42.96	43.79	config	model | log
UPerNet	DeiT-S	512x512	160000	4.68	29.19	V100	42.87	43.79	config	model | log
UPerNet	DeiT-S + MLN	512x512	160000	5.69	11.18	V100	43.82	45.07	config	model | log
UPerNet	DeiT-S + LN + MLN	512x512	160000	5.69	12.39	V100	43.52	45.01	config	model | log
UPerNet	DeiT-B	512x512	80000	7.75	9.69	V100	45.24	46.73	config	model | log
UPerNet	DeiT-B	512x512	160000	7.75	10.39	V100	45.36	47.16	config	model | log
UPerNet	DeiT-B + MLN	512x512	160000	9.21	7.78	V100	45.46	47.16	config	model | log
UPerNet	DeiT-B + LN + MLN	512x512	160000	9.21	7.75	V100	45.37	47.23	config	model | log

Citation

@article{dosoViTskiy2020,
  title={An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale},
  author={DosoViTskiy, Alexey and Beyer, Lucas and Kolesnikov, Alexander and Weissenborn, Dirk and Zhai, Xiaohua and Unterthiner, Thomas and  Dehghani, Mostafa and Minderer, Matthias and Heigold, Georg and Gelly, Sylvain and Uszkoreit, Jakob and Houlsby, Neil},
  journal={arXiv preprint arXiv:2010.11929},
  year={2020}
}