Vision Transformer

by Alexey Dosovitskiy*†, Lucas Beyer*, Alexander Kolesnikov*, Dirk Weissenborn*, Xiaohua Zhai*, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit and Neil Houlsby*†.

(*) equal technical contribution, (†) equal advising.

Open source release prepared by Andreas Steiner.

Note: This repository was forked and modified from google-research/big_transfer.

Introduction

In this repository we release models from the paper An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale that were pre-trained on the ImageNet-21k (imagenet21k) dataset. We provide the code for fine-tuning the released models in Jax/Flax.

Overview of the model: we split an image into fixed-size patches, linearly embed each of them, add position embeddings, and feed the resulting sequence of vectors to a standard Transformer encoder. In order to perform classification, we use the standard approach of adding an extra learnable "classification token" to the sequence.

Colab

Check out the Colab for loading the data, fine-tuning the model, evaluation, and inference. The Colab loads the code from this repository and runs by default on a TPU with 8 cores.

https://colab.research.google.com/github/google-research/vision_transformer/blob/master/vit_jax.ipynb

Note that the Colab can be run as is storing all data in the ephemeral VM, or, alternatively you can log into your personal Google Drive to persist the code and data there.

Installation

Make sure you have Python>=3.6 installed on your machine.

For installing Jax, follow the instructions provided in the corresponding repository linked here. Note that installation instructions for GPU differs slightly from the instructions for CPU.

Then, install python dependencies by running:

pip install -r vit_jax/requirements.txt

Available models

We provide models pre-trained on imagenet21k for the following architectures: ViT-B/16, ViT-B/32, ViT-L/16, ViT-L/32 and ViT-H/14. We provide the same models pre-trained on imagenet21k and fine-tuned on imagenet2012.

Update (1.12.2020): We have added the R50+ViT-B/16 hybrid model (ViT-B/16 on top of a Resnet50 backbone). When pretrained on imagenet21k, this model achieves almost the performance of the L/16 model with less than half the computational finetuning cost.

Update (9.11.2020): We have also added the ViT-L/16 and ViT-H/14 models.

Update (29.10.2020): We have added ViT-B/16 and ViT-L/16 models pretrained on ImageNet-21k and then fine-tuned on ImageNet at 224x224 resolution (instead of default 384x384). These models have the suffix "-224" in their name. They are expected to achieve 81.2% and 82.7% top-1 accuracies respectively.

You can find all these models in the following storage bucket:

https://console.cloud.google.com/storage/vit_models/

For example, if you would like to download the ViT-B/16 pre-trained on imagenet21k run the following command:

wget https://storage.googleapis.com/vit_models/imagenet21k/ViT-B_16.npz

How to fine-tune ViT

You can run fine-tuning of the downloaded model on your dataset of interest. All frameworks share the command line interface

python3 -m vit_jax.train --name ViT-B_16-cifar10_`date +%F_%H%M%S` --model ViT-B_16 --logdir /tmp/vit_logs --dataset cifar10

Currently, the code will automatically download CIFAR-10 and CIFAR-100 datasets. Other public or custom datasets can be easily integrated, using tensorflow datasets library. Note that you will also need to update vit_jax/input_pipeline.py to specify some parameters about any added dataset.

Note that our code uses all available GPUs/TPUs for fine-tuning.

To see a detailed list of all available flags, run python3 -m vit_jax.train --help.

Notes about some flags:

• --accum_steps=16 : This works well with ViT-B_16 on a machine that has 8 GPUs of type V100 with 16G memory each attached. If you have fewer accelerators or accelerators with less memory, you can use the same configuration but increase the --accum_steps. For a small model like ViT-B_32 you can even use --accum_steps=1. For a large model like ViT-L_16 you need to go in the other direction (e.g. --accum_steps=32). Note that the largest model ViT-H_14 also needs adaptation of the batch size (--accum_steps=2 --batch=16 should work on a 8x V100). tested `)
• --batch=512 : Alternatively, you can decrease the batch size, but that usually involves some tuning of the learning rate parameters.

Expected results

In this table we closely follow experiments from the paper and report results that were achieved by running this code on Google Cloud machine with eight V100 GPUs.

upstream	model	dataset	accuracy	wall_clock_time	link
imagenet21k	R50+ViT-B_16	cifar10	0.9893	10.8h	tensorboard.dev
imagenet21k	R50+ViT-B_16	cifar10	0.9885	10.9h	tensorboard.dev
imagenet21k	R50+ViT-B_16	cifar100	0.9235	10.8h	tensorboard.dev
imagenet21k	R50+ViT-B_16	cifar100	0.9239	10.8h	tensorboard.dev
imagenet21k	R50+ViT-B_16	imagenet2012	0.8505	25.9h	tensorboard.dev
imagenet21k	R50+ViT-B_16	imagenet2012	0.8492	25.9h	tensorboard.dev
imagenet21k	ViT-B_16	cifar10	0.9892	7.2h	tensorboard.dev
imagenet21k	ViT-B_16	cifar10	0.9903	7.7h	tensorboard.dev
imagenet21k	ViT-B_16	cifar100	0.9226	7.2h	tensorboard.dev
imagenet21k	ViT-B_16	cifar100	0.9264	7.5h	tensorboard.dev
imagenet21k	ViT-B_16	imagenet2012	0.8462	17.9h	tensorboard.dev
imagenet21k	ViT-B_16	imagenet2012	0.8461	17.8h	tensorboard.dev
imagenet21k	ViT-B_32	cifar10	0.9893	1.6h	tensorboard.dev
imagenet21k	ViT-B_32	cifar10	0.9889	1.6h	tensorboard.dev
imagenet21k	ViT-B_32	cifar100	0.9208	1.6h	tensorboard.dev
imagenet21k	ViT-B_32	cifar100	0.9196	1.6h	tensorboard.dev
imagenet21k	ViT-B_32	imagenet2012	0.8179	4.2h	tensorboard.dev
imagenet21k	ViT-B_32	imagenet2012	0.8179	4.1h	tensorboard.dev
imagenet21k	ViT-L_16	cifar10	0.9907	24.7h	tensorboard.dev
imagenet21k	ViT-L_16	cifar10	0.991	24.9h	tensorboard.dev
imagenet21k	ViT-L_16	cifar100	0.9304	24.8h	tensorboard.dev
imagenet21k	ViT-L_16	cifar100	0.93	24.4h	tensorboard.dev
imagenet21k	ViT-L_16	imagenet2012	0.8507	59.2h	tensorboard.dev
imagenet21k	ViT-L_16	imagenet2012	0.8505	59.2h	tensorboard.dev
imagenet21k	ViT-L_32	cifar10	0.9903	5.7h	tensorboard.dev
imagenet21k	ViT-L_32	cifar10	0.9909	5.8h	tensorboard.dev
imagenet21k	ViT-L_32	cifar100	0.9302	6.7h	tensorboard.dev
imagenet21k	ViT-L_32	cifar100	0.9306	6.7h	tensorboard.dev
imagenet21k	ViT-L_32	imagenet2012	0.8122	14.7h	tensorboard.dev
imagenet21k	ViT-L_32	imagenet2012	0.812	14.7h	tensorboard.dev

We also would like to emphasize that high-quality results can be achieved with shorter training schedules and encourage users of our code to play with hyper-parameters to trade-off accuracy and computational budget. Some examples for CIFAR-10/100 datasets are presented in the table below.

upstream	model	dataset	total_steps / warmup_steps	accuracy	wall-clock time	link
imagenet21k	ViT-B_16	cifar10	500 / 50	0.9859	17m	tensorboard.dev
imagenet21k	ViT-B_16	cifar10	1000 / 100	0.9886	39m	tensorboard.dev
imagenet21k	ViT-B_16	cifar100	500 / 50	0.8917	17m	tensorboard.dev
imagenet21k	ViT-B_16	cifar100	1000 / 100	0.9115	39m	tensorboard.dev

Bibtex

@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}
}

This is not an official Google product.