Cascade Transformer for Hierarchical Semantic Reasoning in Text-based VQA

Repository Setup

Create a fresh conda environment, and install all dependencies.

conda create -n CT-HSR_env python=3.8
conda activate CT-HSR
git clone https://github.com/FTFWO/CT-HSR
cd CT-HSR
pip install -r requirements.txt

Data

data
├── detectron
├── dict
├── feat_resx
├── imdb
├── m4c_vocabs
├── ocr_feat_resx
├── original_dl
├── 1600-400-20

We recommend using the following AzCopy command to download. AzCopy executable tools can be downloaded here. TextVQA/Caps/STVQA Data (~62G).

path/to/azcopy copy <folder-link> <target-address> --resursive

# for example, downloading TextVQA/Caps/STVQA Data
path/to/azcopy copy https://tapvqacaption.blob.core.windows.net/data/data <local_path>/CT-HSR/data --recursive

Training and Evaluation