- Creating conda environment
conda create -n unicorn python=3.10
conda activate unicorn
- Build from source
cd UNICORN
pip install -r requirements.txt
pip install -e .
- Unzip data & Train
download the folder /nfs/data/data/howto100m/annotations
to your server in howto100m folder.
unzip activitynet.zip -d ./activitynet
unzip youcook2.zip -d ./youcook2
unzip qvhighlights.zip -d ./qvhighlights
unzip Charades_v1_480.zip -d ./charades
cd youcook2
mv video_1fps video_1fps_new
bash run_scripts/unicorn/train/pretrain_75frames_temp_aug.sh
LAVIS from Salesforce: the codebase we built upon, and we selected InstructBLIP as our base model initialization.