Global Context Vision Transformer (GC ViT)

This repository is the official PyTorch implementation of Global Context Vision Transformers.

Global Context Vision Transformers
Ali Hatamizadeh, Hongxu (Danny) Yin, Jan Kautz, and Pavlo Molchanov.

GC ViT achieves state-of-the-art results across image classification, object detection and semantic segmentation tasks. On ImageNet-1K dataset for classification, the tiny, small, base and large variants of GC ViT with 28M, 51M, 90M and 201M parameters achieve 83.4, 83.9, 84.5 and 84.8 Top-1 accuracy, respectively, surpassing comparably-sized prior art such as CNN-based ConvNeXt and ViT-based Swin Transformer by a large margin. Pre-trained GC ViT backbones in downstream tasks of object detection, instance segmentation, and semantic segmentation using MS COCO and ADE20K datasets outperform prior work consistently, sometimes by large margins.

The architecture of GC ViT is demonstrated in the following:

Updates

01/12/2023

1. Updated pre-trained weights for GC ViT Large model. Please see below for download link.

2. ImageNet-v2 benchmarks released for all GC ViT models.

11/15/2022

1. Pre-trained weights for GC ViT Large model are released. Please see below for download link.

08/28/2022

1. GC ViT and its pre-trained weights are now available as part of timm library.

08/11/2022

1. New Pre-trained model weights with improved performance have been released. Please see below for download link.
2. GC ViT model has been updated with enhanced global query generator.

06/23/2022

1. Pre-trained model weights released. Please see below for download link.

06/17/2022

1. GC ViT model, training and validation scripts released for ImageNet-1K classification.
2. Pre-trained model checkpoints will be released soon.

Introduction

GC ViT leverages global context self-attention modules, joint with local self-attention, to effectively yet efficiently model both long and short-range spatial interactions, without the need for expensive operations such as computing attention masks or shifting local windows.

ImageNet Benchmarks

ImageNet-1K Pretrained Models

Name	Acc@1	Acc@5	Resolution	#Params	FLOPs	Summary	Download
GC ViT-XXT	79.8	95.1	224x224	12	2.1	summary	model
GC ViT-XT	82.0	96.0	224x224	20	2.6	summary	model
GC ViT-T	83.4	96.4	224x224	28	4.7	summary	model
GC ViT-S	83.9	96.6	224x224	51	8.5	summary	model
GC ViT-B	84.5	96.8	224x224	90	14.8	summary	model
GC ViT-L	84.8	97.1	224x224	201	32.6	summary	model

ImageNet-v2 Benchmarks

Name	Acc@1-ImageNet-v2	Acc@1-ImageNet-1K	Resolution
GC ViT-XXT	69.3	79.8	224x224
GC ViT-XT	71.3	82.0	224x224
GC ViT-T	73.1	83.4	224x224
GC ViT-S	73.8	83.9	224x224
GC ViT-B	74.4	84.5	224x224
GC ViT-L	74.9	84.8	224x224

Installation

This repository is compatible with NVIDIA PyTorch docker nvcr>=21.06 which can be obtained in this link.

The dependencies can be installed by running:

pip install -r requirements.txt

Data Preparation

Please download the ImageNet dataset from its official website. The training and validation images need to have sub-folders for each class with the following structure:

  imagenet
  ├── train
  │   ├── class1
  │   │   ├── img1.jpeg
  │   │   ├── img2.jpeg
  │   │   └── ...
  │   ├── class2
  │   │   ├── img3.jpeg
  │   │   └── ...
  │   └── ...
  └── val
      ├── class1
      │   ├── img4.jpeg
      │   ├── img5.jpeg
      │   └── ...
      ├── class2
      │   ├── img6.jpeg
      │   └── ...
      └── ...
 

Commands

Training on ImageNet-1K From Scratch (Multi-GPU)

The GC ViT model can be trained from scratch on ImageNet-1K dataset by running:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus> --master_port 11223  train.py \ 
--config <config-file> --data_dir <imagenet-path> --batch-size --amp <batch-size-per-gpu> --tag <run-tag> --model-ema

To resume training from a pre-trained checkpoint:

python -m torch.distributed.launch --nproc_per_node <num-of-gpus> --master_port 11223  train.py \ 
--resume <checkpoint-path> --config <config-file> --amp --data_dir <imagenet-path> --batch-size <batch-size-per-gpu> --tag <run-tag> --model-ema

Evaluation

To evaluate a pre-trained checkpoint using ImageNet-1K validation set on a single GPU:

python validate.py --model <model-name> --checkpoint <checkpoint-path> --data_dir <imagenet-path> --batch-size <batch-size-per-gpu>

Citation

Please consider citing GC ViT paper if it is useful for your work:

@article{hatamizadeh2022global,
  title={Global Context Vision Transformers},
  author={Hatamizadeh, Ali and Yin, Hongxu and Kautz, Jan and Molchanov, Pavlo},
  journal={arXiv preprint arXiv:2206.09959},
  year={2022}
}

Third-party Implementations and Resources

In this section, we list third-party contributions by other users. If you would like to have your work included here, please raise an issue in this repository.

Name	Link	Contributor	Note
timm	Link	@rwightman	PyTorch
tfgcvit	Link	@shkarupa-alex	Tensorflow 2.0 (Keras)
gcvit-tf	Link	@awsaf49	Tensorflow 2.0 (Keras)
GCViT-TensorFlow	Link	@EMalagoli92	Tensorflow 2.0 (Keras)
keras_cv_attention_models	Link	@leondgarse	Keras
Paper Explanation	Link	@awsaf49	Annotated GC ViT
Colab Notebook	Link	@awsaf49	Flower classification
Kaggle Notebook	Link	@awsaf49	Flower classification
Live Demo	Link	@awsaf49	Hugging Face demo

Acknowledgement

• This repository is built upon the timm library.

• We would like to sincerely thank the community especially Github users @rwightman, @shkarupa-alex, @awsaf49, @leondgarse, who have provided insightful feedback, which has helped us to further improve GC ViT and achieve even better benchmarks.

Licenses

Copyright © 2022, NVIDIA Corporation. All rights reserved.

This work is made available under the Nvidia Source Code License-NC. Click here to view a copy of this license.

The pre-trained models are shared under CC-BY-NC-SA-4.0. If you remix, transform, or build upon the material, you must distribute your contributions under the same license as the original.

For license information regarding the timm repository, please refer to its official website.

For license information regarding the ImageNet dataset, please refer to its official website.