Paper Link: https://arxiv.org/abs/2201.04850
We clone the repo first,
# clone the repository then change the directory/branch
git clone https://github.com/ilkerkesen/MCQ.git
cd MCQ
git checkout vl-bench origin/vl-bench
Then we create the conda environment,
# setup the environment and directories
conda create -n mcq python=3.8 # I suggest micromamba (it's way faster)
conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 cudatoolkit=11.0 -c pytorch
pip install -r requirements.txt
Finally, setup the directories/symlinks and download the model weights,
mkdir checkpoints
mkdir data
mkdir exps
# download checkpoint and create symlinks for the dataset
wget -c http://balina.ku.edu.tr/download/MCQ.pth -P ./checkpoints/
ln -s /path/to/vl-bench/annotations ./data/vl-bench
Just run the following command,
python mcq_test_vlbench.py \
--config configs/vlbench.json
--metadata_filename json_filename.json \
--quva_dir /path/to/quva \
--something_something_dir /path/to/dataset-videos \
--output_file /path/to/output.json
This command will produce a results annotation file /path/to/output.json
. To generate scores for the proficiency task, pass the --proficiency
flag.