Global Context Vision Transformer (GC ViT)

October 14, 2023 · View on GitHub

This repository presents the official PyTorch implementation of Global Context Vision Transformers (ICML2023)

Global Context Vision Transformers
Ali Hatamizadeh, Hongxu (Danny) Yin, Greg Heinrich, 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, GC ViT variants with 51M, 90M and 201M parameters achieve 84.3, 85.9 and 85.7 Top-1 accuracy, respectively, surpassing comparably-sized prior art such as CNN-based ConvNeXt and ViT-based Swin Transformer.

The architecture of GC ViT is demonstrated in the following:

gc_vit

💥 News 💥

[10.14.2023] 🔥 We have released the object detection code !
[07.27.2023] We will present GC ViT in the (1:30-3:30 HDT) ICML23 session in exhibit hall#1, poster #516.
[07.22.2023] 🔥🔥 We have released pretrained 21K GC ViT-L checkpoint for 512 x 512 resolution !
[07.22.2023] Pretrained checkpoints are now available in official NVIDIA GCViT HuggingFace page !
[07.21.2023] 🔥 We have released the object detection/instance segmentation code !
[05.21.2023] 🔥 We have released ImageNet-21K fine-tuned GC ViT model weights for 224x224 and 384x384.
[05.21.2023] 🔥🔥 We have released new ImageNet-1K GC ViT model weights with better performance !
[04.24.2023] 🔥🔥🔥 GC ViT has been accepted to ICML 2023 !

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

Model Variant	Acc@1	#Params(M)	FLOPs(G)	Download
GC ViT-XXT	79.9	12	2.1	model
GC ViT-XT	82.0	20	2.6	model
GC ViT-T	83.5	28	4.7	model
GC ViT-T2	83.7	34	5.5	model
GC ViT-S	84.3	51	8.5	model
GC ViT-S2	84.8	68	10.7	model
GC ViT-B	85.0	90	14.8	model
GC ViT-L	85.7	201	32.6	model

ImageNet-21K Pretrained Models

Model Variant	Resolution	Acc@1	#Params(M)	FLOPs(G)	Download
GC ViT-L	224 x 224	86.6	201	32.6	model
GC ViT-L	384 x 384	87.4	201	120.4	model
GC ViT-L	512 x 512	87.6	201	245.0	model

Installation

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 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:

@inproceedings{hatamizadeh2023global,
  title={Global context vision transformers},
  author={Hatamizadeh, Ali and Yin, Hongxu and Heinrich, Greg and Kautz, Jan and Molchanov, Pavlo},
  booktitle={International Conference on Machine Learning},
  pages={12633--12646},
  year={2023},
  organization={PMLR}
}

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	Framework
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
flaim	Link	@BobMcDear	JAX/Flax

Additional Resources

We list additional GC ViT resources such as notebooks, demos, paper explanations in this section. If you have created similar items and would like to be included, please raise an issue in this repository.

Name	Link	Contributor	Note
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

Licenses

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, please refer to its repository.

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

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.