Skip to content


Switch branches/tags

Name already in use

A tag already exists with the provided branch name. Many Git commands accept both tag and branch names, so creating this branch may cause unexpected behavior. Are you sure you want to create this branch?


Failed to load latest commit information.

Vision Transformer and MLP-Mixer Architectures

In this repository we release models from the papers

The models were pre-trained on the ImageNet and ImageNet-21k datasets. We provide the code for fine-tuning the released models in JAX/Flax.

The models from this codebase were originally trained in where you can find more advanced code (e.g. multi-host training), as well as some of the original training scripts (e.g. configs/ for pre-training a ViT, or configs/ for transfering a model).

Table of contents:


Below Colabs run both with GPUs, and TPUs (8 cores, data parallelism).

The first Colab demonstrates the JAX code of Vision Transformers and MLP Mixers. This Colab allows you to edit the files from the repository directly in the Colab UI and has annotated Colab cells that walk you through the code step by step, and lets you interact with the data.

The second Colab allows you to explore the >50k Vision Transformer and hybrid checkpoints that were used to generate the data of the third paper "How to train your ViT? ...". The Colab includes code to explore and select checkpoints, and to do inference both using the JAX code from this repo, and also using the popular timm PyTorch library that can directly load these checkpoints as well. Note that a handful of models are also available directly from TF-Hub: sayakpaul/collections/vision_transformer (external contribution by Sayak Paul).

The second Colab also lets you fine-tune the checkpoints on any tfds dataset and your own dataset with examples in individual JPEG files (optionally directly reading from Google Drive).

Note: As for now (6/20/21) Google Colab only supports a single GPU (Nvidia Tesla T4), and TPUs (currently TPUv2-8) are attached indirectly to the Colab VM and communicate over slow network, which leads to pretty bad training speed. You would usually want to set up a dedicated machine if you have a non-trivial amount of data to fine-tune on. For details see the Running on cloud section.


Make sure you have Python>=3.10 installed on your machine.

Install JAX and python dependencies by running:

# If using GPU:
pip install -r vit_jax/requirements.txt

# If using TPU:
pip install -r vit_jax/requirements-tpu.txt

For newer versions of JAX, follow the instructions provided in the corresponding repository linked here. Note that installation instructions for CPU, GPU and TPU differs slightly.

Install Flaxformer, follow the instructions provided in the corresponding repository linked here.

For more details refer to the section Running on cloud below.

Fine-tuning a model

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

For example for fine-tuning a ViT-B/16 (pre-trained on imagenet21k) on CIFAR10 (note how we specify b16,cifar10 as arguments to the config, and how we instruct the code to access the models directly from a GCS bucket instead of first downloading them into the local directory):

python -m vit_jax.main --workdir=/tmp/vit-$(date +%s) \
    --config=$(pwd)/vit_jax/configs/,cifar10 \

In order to fine-tune a Mixer-B/16 (pre-trained on imagenet21k) on CIFAR10:

python -m vit_jax.main --workdir=/tmp/vit-$(date +%s) \
    --config=$(pwd)/vit_jax/configs/ \

The "How to train your ViT? ..." paper added >50k checkpoints that you can fine-tune with the configs/ config. When you only specify the model name (the value from configs/, then the best i21k checkpoint by upstream validation accuracy ("recommended" checkpoint, see section 4.5 of the paper) is chosen. To make up your mind which model you want to use, have a look at Figure 3 in the paper. It's also possible to choose a different checkpoint (see Colab vit_jax_augreg.ipynb) and then specify the value from the filename or adapt_filename column, which correspond to the filenames without .npz from the gs://vit_models/augreg directory.

python -m vit_jax.main --workdir=/tmp/vit-$(date +%s) \
    --config=$(pwd)/vit_jax/configs/ \
    --config.dataset=oxford_iiit_pet \

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/ 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 on memory:

  • Different models require different amount of memory. Available memory also depends on the accelerator configuration (both type and count). If you encounter an out-of-memory error you can increase the value of --config.accum_steps=8 -- alternatively, you could also decrease the --config.batch=512 (and decrease --config.base_lr accordingly).
  • The host keeps a shuffle buffer in memory. If you encounter a host OOM (as opposed to an accelerator OOM), you can decrease the default --config.shuffle_buffer=50000.

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.

Figure 1 from paper

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.

Available ViT models

We provide a variety of ViT models in different GCS buckets. The models can be downloaded with e.g.:


The model filenames (without the .npz extension) correspond to the config.model_name in vit_jax/configs/

We recommend using the following checkpoints, trained with AugReg that have the best pre-training metrics:

Model Pre-trained checkpoint Size Fine-tuned checkpoint Resolution Img/sec Imagenet accuracy
L/16 gs://vit_models/augreg/L_16-i21k-300ep-lr_0.001-aug_strong1-wd_0.1-do_0.0-sd_0.0.npz 1243 MiB gs://vit_models/augreg/L_16-i21k-300ep-lr_0.001-aug_strong1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz 384 50 85.59%
B/16 gs://vit_models/augreg/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0.npz 391 MiB gs://vit_models/augreg/B_16-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz 384 138 85.49%
S/16 gs://vit_models/augreg/S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0.npz 115 MiB gs://vit_models/augreg/S_16-i21k-300ep-lr_0.001-aug_light1-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz 384 300 83.73%
R50+L/32 gs://vit_models/augreg/R50_L_32-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1.npz 1337 MiB gs://vit_models/augreg/R50_L_32-i21k-300ep-lr_0.001-aug_medium1-wd_0.1-do_0.1-sd_0.1--imagenet2012-steps_20k-lr_0.01-res_384.npz 384 327 85.99%
R26+S/32 gs://vit_models/augreg/R26_S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0.npz 170 MiB gs://vit_models/augreg/R26_S_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz 384 560 83.85%
Ti/16 gs://vit_models/augreg/Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0.npz 37 MiB gs://vit_models/augreg/Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz 384 610 78.22%
B/32 gs://vit_models/augreg/B_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0.npz 398 MiB gs://vit_models/augreg/B_32-i21k-300ep-lr_0.001-aug_light1-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz 384 955 83.59%
S/32 gs://vit_models/augreg/S_32-i21k-300ep-lr_0.001-aug_none-wd_0.1-do_0.0-sd_0.0.npz 118 MiB gs://vit_models/augreg/S_32-i21k-300ep-lr_0.001-aug_none-wd_0.1-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.01-res_384.npz 384 2154 79.58%
R+Ti/16 gs://vit_models/augreg/R_Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0.npz 40 MiB gs://vit_models/augreg/R_Ti_16-i21k-300ep-lr_0.001-aug_none-wd_0.03-do_0.0-sd_0.0--imagenet2012-steps_20k-lr_0.03-res_384.npz 384 2426 75.40%

The results from the original ViT paper ( have been replicated using the models from gs://vit_models/imagenet21k:

model dataset dropout=0.0 dropout=0.1
R50+ViT-B_16 cifar10 98.72%, 3.9h (A100), 98.94%, 10.1h (V100),
R50+ViT-B_16 cifar100 90.88%, 4.1h (A100), 92.30%, 10.1h (V100),
R50+ViT-B_16 imagenet2012 83.72%, 9.9h (A100), 85.08%, 24.2h (V100),
ViT-B_16 cifar10 99.02%, 2.2h (A100), 98.76%, 7.8h (V100),
ViT-B_16 cifar100 92.06%, 2.2h (A100), 91.92%, 7.8h (V100),
ViT-B_16 imagenet2012 84.53%, 6.5h (A100), 84.12%, 19.3h (V100),
ViT-B_32 cifar10 98.88%, 0.8h (A100), 98.75%, 1.8h (V100),
ViT-B_32 cifar100 92.31%, 0.8h (A100), 92.05%, 1.8h (V100),
ViT-B_32 imagenet2012 81.66%, 3.3h (A100), 81.31%, 4.9h (V100),
ViT-L_16 cifar10 99.13%, 6.9h (A100), 99.14%, 24.7h (V100),
ViT-L_16 cifar100 92.91%, 7.1h (A100), 93.22%, 24.4h (V100),
ViT-L_16 imagenet2012 84.47%, 16.8h (A100), 85.05%, 59.7h (V100),
ViT-L_32 cifar10 99.06%, 1.9h (A100), 99.09%, 6.1h (V100),
ViT-L_32 cifar100 93.29%, 1.9h (A100), 93.34%, 6.2h (V100),
ViT-L_32 imagenet2012 81.89%, 7.5h (A100), 81.13%, 15.0h (V100),

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 98.59% 17m
imagenet21k ViT-B_16 cifar10 1000 / 100 98.86% 39m
imagenet21k ViT-B_16 cifar100 500 / 50 89.17% 17m
imagenet21k ViT-B_16 cifar100 1000 / 100 91.15% 39m


by Ilya Tolstikhin*, Neil Houlsby*, Alexander Kolesnikov*, Lucas Beyer*, Xiaohua Zhai, Thomas Unterthiner, Jessica Yung, Andreas Steiner, Daniel Keysers, Jakob Uszkoreit, Mario Lucic, Alexey Dosovitskiy.

(*) equal contribution.

Figure 1 from paper

MLP-Mixer (Mixer for short) consists of per-patch linear embeddings, Mixer layers, and a classifier head. Mixer layers contain one token-mixing MLP and one channel-mixing MLP, each consisting of two fully-connected layers and a GELU nonlinearity. Other components include: skip-connections, dropout, and linear classifier head.

For installation follow the same steps as above.

Available Mixer models

We provide the Mixer-B/16 and Mixer-L/16 models pre-trained on the ImageNet and ImageNet-21k datasets. Details can be found in Table 3 of the Mixer paper. All the models can be found at:

Note that these models are also available directly from TF-Hub: sayakpaul/collections/mlp-mixer (external contribution by Sayak Paul).

Expected Mixer results

We ran the fine-tuning code on Google Cloud machine with four V100 GPUs with the default adaption parameters from this repository. Here are the results:

upstream model dataset accuracy wall_clock_time link
ImageNet Mixer-B/16 cifar10 96.72% 3.0h
ImageNet Mixer-L/16 cifar10 96.59% 3.0h
ImageNet-21k Mixer-B/16 cifar10 96.82% 9.6h
ImageNet-21k Mixer-L/16 cifar10 98.34% 10.0h

LiT models

For details, refer to the Google AI blog post LiT: adding language understanding to image models, or read the CVPR paper "LiT: Zero-Shot Transfer with Locked-image text Tuning" (

We published a Transformer B/16-base model with an ImageNet zeroshot accuracy of 72.1%, and a L/16-large model with an ImageNet zeroshot accuracy of 75.7%. For more details about these models, please refer to the LiT model card.

We provide a in-browser demo with small text encoders for interactive use (the smallest models should even run on a modern cell phone):

And finally a Colab to use the JAX models with both image and text encoders:

Note that none of above models support multi-lingual inputs yet, but we're working on publishing such models and will update this repository once they become available.

This repository only contains evaluation code for LiT models. You can find the training code in the big_vision repository:

Expected zeroshot results from model_cards/ (note that the zeroshot evaluation is slightly different from the simplified evaluation in the Colab):

Model B16B_2 L16L
ImageNet zero-shot 73.9% 75.7%
ImageNet v2 zero-shot 65.1% 66.6%
CIFAR100 zero-shot 79.0% 80.5%
Pets37 zero-shot 83.3% 83.3%
Resisc45 zero-shot 25.3% 25.6%
MS-COCO Captions image-to-text retrieval 51.6% 48.5%
MS-COCO Captions text-to-image retrieval 31.8% 31.1%

Running on cloud

While above colabs are pretty useful to get started, you would usually want to train on a larger machine with more powerful accelerators.

Create a VM

You can use the following commands to setup a VM with GPUs on Google Cloud:

# Set variables used by all commands below.
# Note that project must have accounting set up.
# For a list of zones with GPUs refer to
PROJECT=my-awesome-gcp-project  # Project must have billing enabled.

# Below settings have been tested with this repository. You can choose other
# combinations of images & machines (e.g.), refer to the corresponding gcloud commands:
# gcloud compute images list --project ml-images
# gcloud compute machine-types list
# etc.
gcloud compute instances create $VM_NAME \
    --project=$PROJECT --zone=$ZONE \
    --image=c1-deeplearning-tf-2-5-cu110-v20210527-debian-10 \
    --image-project=ml-images --machine-type=n1-standard-96 \
    --scopes=cloud-platform,storage-full --boot-disk-size=256GB \
    --boot-disk-type=pd-ssd --metadata=install-nvidia-driver=True \
    --maintenance-policy=TERMINATE \

# Connect to VM (after some minutes needed to setup & start the machine).
gcloud compute ssh --project $PROJECT --zone $ZONE $VM_NAME

# Stop the VM after use (only storage is billed for a stopped VM).
gcloud compute instances stop --project $PROJECT --zone $ZONE $VM_NAME

# Delete VM after use (this will also remove all data stored on VM).
gcloud compute instances delete --project $PROJECT --zone $ZONE $VM_NAME

Alternatively, you can use the following similar commands to set up a Cloud VM with TPUs attached to them (below commands copied from the TPU tutorial):

PROJECT=my-awesome-gcp-project  # Project must have billing enabled.

# Required to set up service identity initially.
gcloud beta services identity create --service

# Create a VM with TPUs directly attached to it.
gcloud alpha compute tpus tpu-vm create $VM_NAME \
    --project=$PROJECT --zone=$ZONE \
    --accelerator-type v3-8 \
    --version tpu-vm-base

# Connect to VM (after some minutes needed to setup & start the machine).
gcloud alpha compute tpus tpu-vm ssh --project $PROJECT --zone $ZONE $VM_NAME

# Stop the VM after use (only storage is billed for a stopped VM).
gcloud alpha compute tpus tpu-vm stop --project $PROJECT --zone $ZONE $VM_NAME

# Delete VM after use (this will also remove all data stored on VM).
gcloud alpha compute tpus tpu-vm delete --project $PROJECT --zone $ZONE $VM_NAME

Setup VM

And then fetch the repository and the install dependencies (including jaxlib with TPU support) as usual:

git clone --depth=1 --branch=master
cd vision_transformer

# optional: install virtualenv
pip3 install virtualenv
python3 -m virtualenv env
. env/bin/activate

If you're connected to a VM with GPUs attached, install JAX and other dependencies with the following command:

pip install -r vit_jax/requirements.txt

If you're connected to a VM with TPUs attached, install JAX and other dependencies with the following command:

pip install -r vit_jax/requirements-tpu.txt

Install Flaxformer, follow the instructions provided in the corresponding repository linked here.

For both GPUs and TPUs, Check that JAX can connect to attached accelerators with the command:

python -c 'import jax; print(jax.devices())'

And finally execute one of the commands mentioned in the section fine-tuning a model.


  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},

  title={MLP-Mixer: An all-MLP Architecture for Vision},
  author={Tolstikhin, Ilya and Houlsby, Neil and Kolesnikov, Alexander and Beyer, Lucas and Zhai, Xiaohua and Unterthiner, Thomas and Yung, Jessica and Steiner, Andreas and Keysers, Daniel and Uszkoreit, Jakob and Lucic, Mario and Dosovitskiy, Alexey},
  journal={arXiv preprint arXiv:2105.01601},

  title={How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers},
  author={Steiner, Andreas and Kolesnikov, Alexander and and Zhai, Xiaohua and Wightman, Ross and Uszkoreit, Jakob and Beyer, Lucas},
  journal={arXiv preprint arXiv:2106.10270},

  title={When Vision Transformers Outperform ResNets without Pretraining or Strong Data Augmentations},
  author={Chen, Xiangning and Hsieh, Cho-Jui and Gong, Boqing},
  journal={arXiv preprint arXiv:2106.01548},

  title={Surrogate Gap Minimization Improves Sharpness-Aware Training},
  author={Zhuang, Juntang and Gong, Boqing and Yuan, Liangzhe and Cui, Yin and Adam, Hartwig and Dvornek, Nicha and Tatikonda, Sekhar and Duncan, James and Liu, Ting},

  title={LiT: Zero-Shot Transfer with Locked-image Text Tuning},
  author={Zhai, Xiaohua and Wang, Xiao and Mustafa, Basil and Steiner, Andreas and Keysers, Daniel and Kolesnikov, Alexander and Beyer, Lucas},


In reverse chronological order:

  • 2022-08-18: Added LiT-B16B_2 model that was trained for 60k steps (LiT_B16B: 30k) without linear head on the image side (LiT_B16B: 768) and has better performance.

  • 2022-06-09: Added the ViT and Mixer models trained from scratch using GSAM on ImageNet without strong data augmentations. The resultant ViTs outperform those of similar sizes trained using AdamW optimizer or the original SAM algorithm, or with strong data augmentations.

  • 2022-04-14: Added models and Colab for LiT models.

  • 2021-07-29: Added ViT-B/8 AugReg models (3 upstream checkpoints and adaptations with resolution=224).

  • 2021-07-02: Added the "When Vision Transformers Outperform ResNets..." paper

  • 2021-07-02: Added SAM (Sharpness-Aware Minimization) optimized ViT and MLP-Mixer checkpoints.

  • 2021-06-20: Added the "How to train your ViT? ..." paper, and a new Colab to explore the >50k pre-trained and fine-tuned checkpoints mentioned in the paper.

  • 2021-06-18: This repository was rewritten to use Flax Linen API and ml_collections.ConfigDict for configuration.

  • 2021-05-19: With publication of the "How to train your ViT? ..." paper, we added more than 50k ViT and hybrid models pre-trained on ImageNet and ImageNet-21k with various degrees of data augmentation and model regularization, and fine-tuned on ImageNet, Pets37, Kitti-distance, CIFAR-100, and Resisc45. Check out vit_jax_augreg.ipynb to navigate this treasure trove of models! For example, you can use that Colab to fetch the filenames of recommended pre-trained and fine-tuned checkpoints from the i21k_300 column of Table 3 in the paper.

  • 2020-12-01: Added the R50+ViT-B/16 hybrid model (ViT-B/16 on top of a Resnet-50 backbone). When pretrained on imagenet21k, this model achieves almost the performance of the L/16 model with less than half the computational finetuning cost. Note that "R50" is somewhat modified for the B/16 variant: The original ResNet-50 has [3,4,6,3] blocks, each reducing the resolution of the image by a factor of two. In combination with the ResNet stem this would result in a reduction of 32x so even with a patch size of (1,1) the ViT-B/16 variant cannot be realized anymore. For this reason we instead use [3,4,9] blocks for the R50+B/16 variant.

  • 2020-11-09: Added the ViT-L/16 model.

  • 2020-10-29: 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.


Open source release prepared by Andreas Steiner.

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

This is not an official Google product.