Update (2.7.2021): Added the "When Vision Transformers Outperform ResNets..." paper, and SAM (Sharpness-Aware Minimization) optimized ViT and MLP-Mixer checkpoints.
Update (20.6.2021): 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.
Update (18.6.2021): This repository was rewritten to use Flax Linen API and
ml_collections.ConfigDict for configuration.
In this repository we release models from the papers
- An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- MLP-Mixer: An all-MLP Architecture for Vision
- How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers
- When Vision Transformers Outperform ResNets without Pretraining or Strong Data Augmentations
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.
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>=3.6 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
For more details refer to the section Running on cloud below.
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/vit.py:b16,cifar10 \
--config.pretrained_dir='gs://vit_models/imagenet21k'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/mixer_base16_cifar10.py \
--config.pretrained_dir='gs://mixer_models/imagenet21k'The "How to train your ViT? ..." paper added >50k checkpoints that you can
fine-tune with the configs/augreg.py config. When you only specify the model
name (the config.name value from configs/model.py), 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/augreg.py:R_Ti_16 \
--config.dataset=oxford_iiit_pet \
--config.base_lr=0.01Currently, 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 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_lraccordingly). - 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.
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.
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.
We provide models pre-trained on ImageNet-21k for the following architectures: ViT-B/16, ViT-B/32, ViT-L/16 and ViT-L/32. We provide the same models pre-trained on ImageNet-21k and fine-tuned on ImageNet.
Update (29.7.2021): Added ViT-B/8 AugReg models (3 upstream checkpoints and adaptations with resolution=224).
Update (2.7.2021): We added the ViT models trained from scratch with
SAM optimizer on ImageNet
(with basic Inception-style preprocessing). The resultant ViTs outperform
ResNets of similar size and throughput without large-scale pre-training or
strong data augmentations. They also possess more perceptive attention maps.
To use those models, you can simply replace the model path
in vit_jax.ipynb with gs://vit_models/sam.
Update (19.5.2021): 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:
| 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% |
Update (1.12.2020): We have 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.
Update (9.11.2020): We have also added the ViT-L/16 model.
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
Table below runs experiments both with transformer.dropout_rate=0.1 (as in the
ViT paper), and with transformer.dropout_rate=0.0, which improves results
somewhat for models B=16, B/32, and L/32. The better setting was chosen for the
default config of the models in this repository. Note also that all these models
have representation_size=None, i.e. the last layer before the classification
layer is dropped for fine-tuning.
| model | dataset | dropout=0.0 | dropout=0.1 |
|---|---|---|---|
| R50+ViT-B_16 | cifar10 | 98.72%, 3.9h (A100), tb.dev | 98.94%, 10.1h (V100), tb.dev |
| R50+ViT-B_16 | cifar100 | 90.88%, 4.1h (A100), tb.dev | 92.30%, 10.1h (V100), tb.dev |
| R50+ViT-B_16 | imagenet2012 | 83.72%, 9.9h (A100), tb.dev | 85.08%, 24.2h (V100), tb.dev |
| ViT-B_16 | cifar10 | 99.02%, 2.2h (A100), tb.dev | 98.76%, 7.8h (V100), tb.dev |
| ViT-B_16 | cifar100 | 92.06%, 2.2h (A100), tb.dev | 91.92%, 7.8h (V100), tb.dev |
| ViT-B_16 | imagenet2012 | 84.53%, 6.5h (A100), tb.dev | 84.12%, 19.3h (V100), tb.dev |
| ViT-B_32 | cifar10 | 98.88%, 0.8h (A100), tb.dev | 98.75%, 1.8h (V100), tb.dev |
| ViT-B_32 | cifar100 | 92.31%, 0.8h (A100), tb.dev | 92.05%, 1.8h (V100), tb.dev |
| ViT-B_32 | imagenet2012 | 81.66%, 3.3h (A100), tb.dev | 81.31%, 4.9h (V100), tb.dev |
| ViT-L_16 | cifar10 | 99.13%, 6.9h (A100), tb.dev | 99.14%, 24.7h (V100), tb.dev |
| ViT-L_16 | cifar100 | 92.91%, 7.1h (A100), tb.dev | 93.22%, 24.4h (V100), tb.dev |
| ViT-L_16 | imagenet2012 | 84.47%, 16.8h (A100), tb.dev | 85.05%, 59.7h (V100), tb.dev |
| ViT-L_32 | cifar10 | 99.06%, 1.9h (A100), tb.dev | 99.09%, 6.1h (V100), tb.dev |
| ViT-L_32 | cifar100 | 93.29%, 1.9h (A100), tb.dev | 93.34%, 6.2h (V100), tb.dev |
| ViT-L_32 | imagenet2012 | 81.89%, 7.5h (A100), tb.dev | 81.13%, 15.0h (V100), tb.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 | 98.59% | 17m | tensorboard.dev |
| imagenet21k | ViT-B_16 | cifar10 | 1000 / 100 | 98.86% | 39m | tensorboard.dev |
| imagenet21k | ViT-B_16 | cifar100 | 500 / 50 | 89.17% | 17m | tensorboard.dev |
| imagenet21k | ViT-B_16 | cifar100 | 1000 / 100 | 91.15% | 39m | tensorboard.dev |
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.
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.
Update (2.7.2021): We added the MLP-Mixer models trained with
SAM on ImageNet without strong
augmentations (gs://mixer_models/sam). The loss landscapes become
much smoother, and we found that the activated neurons for the
first few layers decrease dramatically after SAM,
indicating the potential redundency of image patches.
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:
https://console.cloud.google.com/storage/mixer_models/
Note that these models are also available directly from TF-Hub: sayakpaul/collections/mlp-mixer (external contribution by Sayak Paul).
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 | tensorboard.dev |
| ImageNet | Mixer-L/16 | cifar10 | 96.59% | 3.0h | tensorboard.dev |
| ImageNet-21k | Mixer-B/16 | cifar10 | 96.82% | 9.6h | tensorboard.dev |
| ImageNet-21k | Mixer-L/16 | cifar10 | 98.34% | 10.0h | tensorboard.dev |
While above colabs are pretty useful to get started, you would usually want to train on a larger machine with more powerful accelerators.
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
# https://cloud.google.com/compute/docs/gpus/gpu-regions-zones
PROJECT=my-awesome-gcp-project # Project must have billing enabled.
VM_NAME=vit-jax-vm-gpu
ZONE=europe-west4-b
# 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 \
--accelerator=type=nvidia-tesla-v100,count=8
# 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_NAMEAlternatively, 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.
VM_NAME=vit-jax-vm-tpu
ZONE=europe-west4-a
# Required to set up service identity initially.
gcloud beta services identity create --service tpu.googleapis.com
# 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 v2-alpha
# 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_NAMEAnd then fetch the repository and the install dependencies (including jaxlib
with TPU support) as usual:
git clone --depth=1 --branch=master https://github.com/google-research/vision_transformer
cd vision_transformer
pip3 install virtualenv
python3 -m virtualenv env
. env/bin/activateIf you're connected to a VM with GPUs attached, install JAX with the following command:
pip3 install --upgrade jax jaxlib \
-f https://storage.googleapis.com/jax-releases/jax_releases.htmlIf you're connected to a VM with TPUs attached, install JAX with the following command:
pip3 install --upgrade jax jaxlibFor both GPUs and TPUs, then proceed to install the remaining dependencies and check that accelerators can indeed show up in JAX:
pip install -r vit_jax/requirements.txt
# Check that JAX can connect to attached accelerators:
python -c 'import jax; print(jax.devices())'And finally execute one of the commands mentioned in the section fine-tuning a model.
@article{dosovitskiy2020vit,
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={ICLR},
year={2021}
}
@article{tolstikhin2021mixer,
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},
year={2021}
}
@article{steiner2021augreg,
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},
year={2021}
}
@article{chen2021outperform,
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},
year={2021},
}
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.