Skip to content

Latest commit

 

History

History
186 lines (159 loc) · 7.13 KB

README.md

File metadata and controls

186 lines (159 loc) · 7.13 KB

VILLA: Vision-and-Language Adversarial Training

This is the official repository of VILLA (NeurIPS 2020 Spotlight). This repository currently supports adversarial finetuning of UNITER on VQA, VCR, NLVR2, and SNLI-VE. Adversarial pre-training with in-domain data will be available soon. Both VILLA-base and VILLA-large pre-trained checkpoints are released.

Overview of VILLA

Most of the code in this repo are copied/modified from UNITER.

Requirements

We provide Docker image for easier reproduction. Please install the following:

Our scripts require the user to have the docker group membership so that docker commands can be run without sudo. We only support Linux with NVIDIA GPUs. We test on Ubuntu 18.04 and V100 cards. We use mixed-precision training hence GPUs with Tensor Cores are recommended.

Quick Start

NOTE: Please run bash scripts/download_pretrained.sh $PATH_TO_STORAGE to get our latest pretrained VILLA checkpoints. This will download both the base and large models.

We use VQA as an end-to-end example for using this code base.

  1. Download processed data and pretrained models with the following command.

    bash scripts/download_vqa.sh $PATH_TO_STORAGE

    After downloading you should see the following folder structure:

    ├── finetune 
    ├── img_db
    │   ├── coco_test2015
    │   ├── coco_test2015.tar
    │   ├── coco_train2014
    │   ├── coco_train2014.tar
    │   ├── coco_val2014
    │   ├── coco_val2014.tar
    │   ├── vg
    │   └── vg.tar
    ├── pretrained
        ├── uniter-base.pt
    │   └── villa-base.pt
    └── txt_db
        ├── vqa_devval.db
        ├── vqa_devval.db.tar
        ├── vqa_test.db
        ├── vqa_test.db.tar
        ├── vqa_train.db
        ├── vqa_train.db.tar
        ├── vqa_trainval.db
        ├── vqa_trainval.db.tar
        ├── vqa_vg.db
        └── vqa_vg.db.tar
    
    

    You can put different pre-trained checkpoints inside the /pretrained folder based on your need.

  2. Launch the Docker container for running the experiments.

    # docker image should be automatically pulled
    source launch_container.sh $PATH_TO_STORAGE/txt_db $PATH_TO_STORAGE/img_db \
        $PATH_TO_STORAGE/finetune $PATH_TO_STORAGE/pretrained

    The launch script respects $CUDA_VISIBLE_DEVICES environment variable. Note that the source code is mounted into the container under /src instead of built into the image so that user modification will be reflected without re-building the image. (Data folders are mounted into the container separately for flexibility on folder structures.)

  3. Run finetuning for the VQA task.

    # inside the container
    horovodrun -np $N_GPU python train_vqa_adv.py --config $YOUR_CONFIG_JSON
    
    # specific example
    horovodrun -np 4 python train_vqa_adv.py --config config/train-vqa-base-4gpu-adv.json
  4. Run inference for the VQA task and then evaluate.

    # inference
    python inf_vqa.py --txt_db /txt/vqa_test.db --img_db /img/coco_test2015 \
    --output_dir $VQA_EXP --checkpoint 6000 --pin_mem --fp16

    The result file will be written at $VQA_EXP/results_test/results_6000_all.json, which can be submitted to the evaluation server

  5. Customization

    # training options
    python train_vqa_adv.py --help
    • command-line argument overwrites JSON config files
    • JSON config overwrites argparse default value.
    • use horovodrun to run multi-GPU training
    • --gradient_accumulation_steps emulates multi-gpu training
    • --checkpoint selects UNITER or VILLA pre-trained checkpoints
    • --adv_training decides using adv. training or not
    • --adv_modality takes values from ['text'], ['image'], ['text','image'], and ['text','image','alter'], the last two correspond to adding perturbations on two modalities simultaneously or alternatively

Downstream Tasks Finetuning

VCR

NOTE: train and inference should be ran inside the docker container

  1. download data
    bash scripts/download_vcr.sh $PATH_TO_STORAGE
    
  2. train
    horovodrun -np 4 python train_vcr_adv.py --config config/train-vcr-base-4gpu-adv.json \
        --output_dir $VCR_EXP
    
  3. inference
    horovodrun -np 4 python inf_vcr.py --txt_db /txt/vcr_test.db \
        --img_db "/img/vcr_gt_test/;/img/vcr_test/" \
        --split test --output_dir $VCR_EXP --checkpoint 8000 \
        --pin_mem --fp16
    
    The result file will be written at $VCR_EXP/results_test/results_8000_all.csv, which can be submitted to VCR leaderboard for evaluation.

NLVR2

NOTE: train and inference should be ran inside the docker container

  1. download data
    bash scripts/download_nlvr2.sh $PATH_TO_STORAGE
    
  2. train
    horovodrun -np 4 python train_nlvr2_adv.py --config config/train-nlvr2-base-1gpu-adv.json \
        --output_dir $NLVR2_EXP
    
  3. inference
    python inf_nlvr2.py --txt_db /txt/nlvr2_test1.db/ --img_db /img/nlvr2_test/ \
    --train_dir /storage/nlvr-base/ --ckpt 6500 --output_dir . --fp16
    

Visual Entailment (SNLI-VE)

NOTE: train should be ran inside the docker container

  1. download data
    bash scripts/download_ve.sh $PATH_TO_STORAGE
    
  2. train
    horovodrun -np 2 python train_ve_adv.py --config config/train-ve-base-2gpu-adv.json \
        --output_dir $VE_EXP
    

Adversarial Training of LXMERT

To keep things simple, we provide another separate repo that can be used to reproduce our results on adversarial finetuning of LXMERT on VQA, GQA, and NLVR2.

Citation

If you find this code useful for your research, please consider citing:

@inproceedings{gan2020large,
  title={Large-Scale Adversarial Training for Vision-and-Language Representation Learning},
  author={Gan, Zhe and Chen, Yen-Chun and Li, Linjie and Zhu, Chen and Cheng, Yu and Liu, Jingjing},
  booktitle={NeurIPS},
  year={2020}
}

@inproceedings{chen2020uniter,
  title={Uniter: Universal image-text representation learning},
  author={Chen, Yen-Chun and Li, Linjie and Yu, Licheng and Kholy, Ahmed El and Ahmed, Faisal and Gan, Zhe and Cheng, Yu and Liu, Jingjing},
  booktitle={ECCV},
  year={2020}
}

License

MIT