To setup the docker container you first have to install docker and the nvidia toolkit. Please follow the instructions here.
To build the docker container that runs pytorch, detectron2 and the preprocessing pipeline, run
sudo docker build --build-arg UID=$(id -u) -t deep-dance:YOUR_TAG .
To build the docker container that runs tensorflow, run
sudo docker build -f Dockerfile_tensorflow --build-arg UID=$(id -u) -t tensorflow:YOUR_TAG .
To run the container do
sudo docker run --gpus all -it -p 6666:8888 -v /path_to_local_folder_/core:/home/deep-dance/core deep-dance:v0 bash
-p 6666:8888 exposes the jupyter port and is optional, 6666 is the port on the host server (you might want to adjust the port number), 8888 is the jupyter port in the docker container.
To start a jupyter notebook server in the container do
jupyter notebook --ip 0.0.0.0 --no-browser
you can then access the notebook server on your host via a browser at
localhost:8888/tree
for the correct token check the terminal output in docker container.
The -v option mounts the ../core folder to the container.
Note: You still have to download the pretrained models for VideoPose3D as described in the main readme file in point 4.