we run our code in the following environment:
- CUDA 11.7
- Python 3.9
- Pytorch 2.1.0
First, clone the repository.
git clone https://github.com/FoundationVision/UniRef.git
Second, install the detectron2.
pip3 install -e . --user
Third, install the necessary packages.
pip3 install -r requirements.txt --user
pip3 install git+https://github.com/youtubevos/cocoapi.git#"egg=pycocotools&subdirectory=PythonAPI" --user
# flash-attn
python3 -m pip install ninja
python3 -m pip install flash-attn==2.0.6
Finally, compile deformable attention CUDA operator.
cd projects/UniRef/uniref/models/deformable_detr/ops;
bash make.sh