Vision Transformer(ViT) in Tensorflow2

Tensorflow2 implementation of the Vision Transformer(ViT).

This repository is for An image is worth 16x16 words: Transformers for image recognition at scale and How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers.

Limitations.

• Due to memory limitations, only the ti/16, s/16, and b/16 models were tested.
• Due to memory limitations, batch_size 2048 in s16 and 1024 in b/16 (in paper, 4096).
• Due to computational resource limitations, only reproduce using imagenet1k.

All experimental results and graphs are opend in Wandb.

• https://docs.google.com/spreadsheets/d/1j0lFlaMuqccFiHj3eQVpZYIbSoXY6Pz6oEW76x7g25M/edit?usp=sharing
• upstream: https://wandb.ai/justhungryman/vit
• downstream: https://wandb.ai/justhungryman/vit-downstream/
• In case of an experiment in which the tpu is stopped, it is resumed (duplicated experiment name but different start epoch).

Model weights

Since this is personal project, it is hard to train with large datasets like imagenet21k. For a pretrain model with good performance, see the official repo. But if you really need it, contact to me.

Install dependencies

pip install -r requirements

All experiments were done on tpu_v3-8 with the support of TRC. But you can experiment on GPU. Check conf/config.yaml and conf/downstream.yaml

  # TPU options
  env:
    mode: tpu
    gcp_project: {your_project}
    tpu_name: node-1
    tpu_zone: europe-west4-a
    mixed_precision: True
  # GPU options
  # env:
  #   mode: gpu
  #   mixed_precision: True

Train from scratch

# example
python run.py experiment=vit-s16-aug_light1-bs_2048-wd_0.1-do_0.1-dp_0.1-lr_1e-3 \
base.project_name=vit-s16-aug_light1-bs_2048-wd_0.1-do_0.1-dp_0.1-lr_1e-3 \
base.save_dir={your_save_dir} base.env.gcp_project={your_gcp_project} \
base.env.tpu_name={your_tpu_name} base.debug=False

Downstream

# example
python run.py --config-name=downstream experiment=downstream-imagenet-ti16_384 \
base.pretrained={your_checkpoint} base.project_name={your_project_name} \
base.save_dir={your_save_dir} base.env.gcp_project={your_gcp_project} \
base.env.tpu_name={your_tpu_name} base.debug=False

Board

To track metics, you can use wandb or tensorboard (default: wandb). You can change in conf/callbacks/{filename.yaml}.

modules:
  - type: MonitorCallback
  - type: TerminateOnNaN
  - type: ProgbarLogger
    params:
      count_mode: steps
  - type: ModelCheckpoint
    params:
      filepath: ???
      save_weights_only: True
  - type: Wandb
    project: vit
    nested_dict: False
    hide_config: True
    params: 
      monitor: val_loss
      save_model: False
  # - type: TensorBoard
  #   params:
  #     log_dir: ???
  #     histogram_freq: 1

TFC

This open source was supported by TPU Research Cloud (TRC) program

Thank you for providing the TPU.

Citations

@article{dosovitskiy2020image,
  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 others},
  journal={arXiv preprint arXiv:2010.11929},
  year={2020}
}

@article{steiner2021train,
  title={How to train your ViT? Data, Augmentation, and Regularization in Vision Transformers},
  author={Steiner, Andreas and Kolesnikov, Alexander and Zhai, Xiaohua and Wightman, Ross and Uszkoreit, Jakob and Beyer, Lucas},
  journal={arXiv preprint arXiv:2106.10270},
  year={2021}
}