TAP: Text-Aware Pre-training

TAP: Text-Aware Pre-training for Text-VQA and Text-Caption

by Zhengyuan Yang, Yijuan Lu, Jianfeng Wang, Xi Yin, Dinei Florencio, Lijuan Wang, Cha Zhang, Lei Zhang, and Jiebo Luo

IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021, Oral

Introduction

We propose Text-Aware Pre-training (TAP) for Text-VQA and Text-Caption tasks. For more details, please refer to our paper.

Citation

@inproceedings{yang2021tap,
  title={TAP: Text-Aware Pre-training for Text-VQA and Text-Caption},
  author={Yang, Zhengyuan and Lu, Yijuan and Wang, Jianfeng and Yin, Xi and Florencio, Dinei and Wang, Lijuan and Zhang, Cha and Zhang, Lei and Luo, Jiebo},
  booktitle={CVPR},
  year={2021}
}

Prerequisites

• Python 3.6

• Pytorch 1.4.0

• Please refer to requirements.txt. Or using

  python setup.py develop
  

Installation

1. Clone the repository

   git clone https://github.com/microsoft/TAP.git
   cd TAP
   python setup.py develop
   

2. Data

• Please refer to the Readme in the data folder.

Training

3. Train the model, run the code under main folder. Using flag --pretrain to access the pre-training mode, otherwise the main QA/Captioning losses are used to optimize the model. Example yml files are in configs folder. Detailed configs are in released models.

   Pre-training:

   python -m torch.distributed.launch --nproc_per_node $num_gpu tools/run.py --pretrain --tasks vqa --datasets $dataset --model $model --seed $seed --config configs/vqa/$dataset/"$pretrain_yml".yml --save_dir save/$pretrain_savedir training_parameters.distributed True
   
   # for example
   python -m torch.distributed.launch --nproc_per_node 4 tools/run.py --pretrain --tasks vqa --datasets m4c_textvqa --model m4c_split --seed 13 --config configs/vqa/m4c_textvqa/tap_base_pretrain.yml --save_dir save/m4c_split_pretrain_test training_parameters.distributed True
   

   Fine-tuning:

   python -m torch.distributed.launch --nproc_per_node $num_gpu tools/run.py --tasks vqa --datasets $dataset --model $model --seed $seed --config configs/vqa/$dataset/"$refine_yml".yml --save_dir save/$refine_savedir --resume_file save/$pretrain_savedir/$savename/best.ckpt training_parameters.distributed True
   
   # for example
   python -m torch.distributed.launch --nproc_per_node 4 tools/run.py --tasks vqa --datasets m4c_textvqa --model m4c_split --seed 13 --config configs/vqa/m4c_textvqa/tap_refine.yml --save_dir save/m4c_split_refine_test --resume_file save/pretrained/textvqa_tap_base_pretrain.ckpt training_parameters.distributed True
   

4. Evaluate the model, run the code under main folder. Set up val or test set by --run_type.

   python -m torch.distributed.launch --nproc_per_node $num_gpu tools/run.py --tasks vqa --datasets $dataset --model $model --config configs/vqa/$dataset/"$refine_yml".yml --save_dir save/$refine_savedir --run_type val --resume_file save/$refine_savedir/$savename/best.ckpt training_parameters.distributed True
   
   # for example
   python -m torch.distributed.launch --nproc_per_node 4 tools/run.py --tasks vqa --datasets m4c_textvqa --model m4c_split --config configs/vqa/m4c_textvqa/tap_refine.yml --save_dir save/m4c_split_refine_test --run_type val --resume_file save/finetuned/textvqa_tap_base_best.ckpt training_parameters.distributed True
   

5. Captioning evaluation.

   python projects/M4C_Captioner/scripts/textcaps_eval.py --set val --pred_file YOUR_VAL_PREDICTION_FILE
   

Performance and Pre-trained Models

Please check the detailed experiment settings in our paper.

Model checkpoints (~17G).

path/to/azcopy copy https://tapvqacaption.blob.core.windows.net/data/save <local_path>/save --recursive

Please refer to the Readme in the data folder for the detailed instructions on azcopy downloading.

Credits

The project is built based on the following repository:

• MMF: A multimodal framework for vision and language research.