yanhuacheng / vilt

Code for the ICML 2021 (long talk) paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"

ViLT

Code for the ICML 2021 (long talk) paper: "ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision"

---

Install

pip install -r requirements.txt
pip install -e .

Download Pretrained Weights

We provide five pretrained weights

1. ViLT-B/32 Pretrained with MLM+ITM for 200k steps on GCC+SBU+COCO+VG (ViLT-B/32 200k) link
2. ViLT-B/32 200k finetuned on VQAv2 link
3. ViLT-B/32 200k finetuned on NLVR2 link
4. ViLT-B/32 200k finetuned on COCO IR/TR link
5. ViLT-B/32 200k finetuned on F30K IR/TR link

Out-of-the-box MLM + Visualization Demo

pip install gradio==1.6.4
python demo.py with num_gpus=<0 if you have no gpus else 1> load_path="<YOUR_WEIGHT_ROOT>/vilt_200k_mlm_itm.ckpt"

ex)
python demo.py with num_gpus=0 load_path="weights/vilt_200k_mlm_itm.ckpt"

Out-of-the-box VQA Demo

pip install gradio==1.6.4
python demo_vqa.py with num_gpus=<0 if you have no gpus else 1> load_path="<YOUR_WEIGHT_ROOT>/vilt_vqa.ckpt" test_only=True

ex)
python demo_vqa.py with num_gpus=0 load_path="weights/vilt_vqa.ckpt" test_only=True

Dataset Preparation

See DATA.md

Train New Models

See TRAIN.md

Evaluation

See EVAL.md

Citation

If you use any part of this code and pretrained weights for your own purpose, please cite our paper.

@article{kim2021vilt,
  title={ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision},
  author={Kim, Wonjae and Son, Bokyung and Kim, Ildoo},
  journal={arXiv preprint arXiv:2102.03334},
  year={2021}
}

Contact for Issues

• Wonjae Kim
• Bokyung Son
• Ildoo Kim