Pytorch implementation of winner from VQA Chllange Workshop

2017 VQA Challenge Winner (CVPR'17 Workshop)

pytorch implementation of Tips and Tricks for Visual Question Answering: Learnings from the 2017 Challenge by Teney et al.

Prerequisites

• python 3.6+
• numpy
• pytorch 0.4
• tqdm
• nltk
• pandas

Data

• VQA 2.0
• COCO 36 features pretrained resnet model
• GloVe pretrained Wikipedia+Gigaword word embedding

Preparation

• To download and extract vqav2, glove, and pretrained visual features:

  bash scripts/download_extract.sh
  

• To prepare data for training:

  python scripts/preproc.py
  

• The structure of data/ directory should look like this:

  - data/
    - zips/
      - v2_XXX...zip
      - ...
      - glove...zip
      - trainval_36.zip
    - glove/
      - glove...txt
      - ...
    - v2_XXX.json
    - ...
    - trainval_resnet...tsv
    (The above are files created after executing scripts/download_extract.sh)
    - tokenizers/
      - ...
    - dict_ans.pkl
    - dict_q.pkl
    - glove_pretrained_300.npy
    - train_qa.pkl
    - val_qa.pkl
    - train_vfeats.pkl
    - val_vfeats.pkl
    (The above are files created after executing scripts/preproc.py)
  

Train

Use default parameters:

bash scripts/train.sh

Notes

• Huge re-factor (especially data preprocessing), tested based on pytorch 0.4.1 and python 3.6
• Training for 20 epochs reach around 50% training accuracy. (model seems buggy in my implementation)
• After all the preprocessing, data/ directory may be up to 38G+
• Some of preproc.py and utils.py are based on this repo

Resources

• The paper.
• Their CVPR Workshop slides.

GitHub

https://github.com/markdtw/vqa-winner-cvprw-2017