README.md

Multimodal Product Manual Question Answering

This is the official repository of the paper MPMQA: Multimodal Question Answering on Product Manuals (AAAI 2023). It provides the PM209 dataset and code of the URA model.
The metadata (products and brands) of PM209 can be found in meta_data/

Setup

Create virtual environment with conda

conda create -n mpmqa python==3.7.6
conda activate mpmqa

Install environments

# Install pytorch
pip install torch torchvision torchaudio 
# Please make sure that the versions of torch and CUDA is compatible.

# Install detectron2-0.2
# Clone detectron2 from github
cd detectron2
git checkout be792b959bca9a
python python -m pip install -e detectron2
cd ..

# Install apex 
# Clone apex from github
cd apex
python setup.py build develop

# Install bua detector
cd ./detector
python setup.py build develop
cd ..

# Install java>=8
sudo apt-get install openjdk-8-jdk

# Install other requirements
pip install -r requirements.txt

PM209 dataset

mkdir -p data && cd data

# Download PM209.zip from: https://drive.google.com/file/d/1K6BPBYdTwKgA1OkNt_BUqVJAn3RDy59B/view?usp=sharing
unzip PM209.zip
cd ..

URA config file and checkpoint

# Download expr.zip from: https://drive.google.com/file/d/1CgY5pg2Z1DtfZVcvLtSIsxYjMiBSBQyx/view?usp=sharing
unzip expr.zip

Pre-trained T5 and BUA models

# Download pretrained.zip from: https://drive.google.com/file/d/1-o1LMbZKCZQOBJtX1C1T_MkZzOG8tsxu/view?usp=share_link
unzip pretrained.zip

# Make symbolic link to the ./detector dir 
cd detector
ln -s ../pretrained .
cd ..

Config file

Hyper-parameters can be found in config.json, and their default values are in parser.py. The config file of the URA model can be found at expr/URA/config.json.

Train, validate, and evaluate

config_json_path="expr/URA/config.json"
gpu_id=0
bash ds_train.sh ${config_json_path} ${gpu_id}

Load checkpoint and evaluate

# Change the 'checkpoint' field to the model path.

config_json_path="expr/URA/eval_opt/config.json"
gpu_id=0
bash ds_eval.sh ${config_json_path} ${gpu_id}

Acknowledge & License

The code is partly based on transformers and detectron2, and is released under the Apache 2.0 license.