Docker support

For your convenience, there is also a Docker image with the pre-installed environments that supports all models. The Docker image does not have the video_features library inside which allows you to tweak the code locally, mount the new version, and just use the container as an environment. It is assumed that you have Docker and nvidia-container-runtime installed.

Setup

Start by cloning the repo locally if you haven't done it already

git clone https://github.com/v-iashin/video_features.git

Download the docker image or build it yourself:

docker pull iashin/video_features
# preventing newer versions of the image to be downloaded unexpectedly
docker tag iashin/video_features video_features
# or
# cd video_features
# docker build --tag video_features .

Once it is done, mount (--mount) the cloned repository folder, and initialize a container with the 0th GPU but remember: just like with any mount, a change from inside of the container will be reflected in the mounted folder (/absolute/path/to/video_features/).:

docker run -it \
    --mount type=bind,source="/absolute/path/to/video_features/",destination="/home/ubuntu/video_features/" \
    --shm-size 8G \
    -it --gpus '"device=0"' \
    video_features:latest \
    bash
# and you should get the bash shell:
# ubuntu@56b1bf77a20c:~$

Check if a GPU is available to PyTorch:

# ubuntu@56b1bf77a20c:~$
python -c "import torch; print(torch.cuda.is_available())"
# True

Finally, try to extract video features:

# cd to `./video_features`
# ubuntu@56b1bf77a20c:~/video_features $
python main.py \
    feature_type=r21d \
    device="cuda:0" \
    video_paths="[./sample/v_ZNVhz7ctTq0.mp4, ./sample/v_GGSY1Qvo990.mp4]"

Extract features from custom videos

You need to mount the folders with video files before you start the container.

If the folder with custom videos is already in ./video_features, you don't have to do anything. Any changes from inside of the container will be reflected in your original dataset (use a backup!). Here is an example of mounting a folder from somewhere else (mounts /absolute/path/somewhere/else/ to /home/ubuntu/video_features/dataset):

docker run -it \
    --mount type=bind,source="/absolute/path/to/video_features/",destination="/home/ubuntu/video_features/" \
    --mount type=bind,source="/absolute/path/somewhere/else/",destination="/home/ubuntu/video_features/dataset/" \
    --shm-size 8G \
    -it --gpus '"device=0"' \
    video_features:latest \
    bash
# ubuntu@56b1bf77a20c:~$
ls ./video_features
# ... dataset ...

If you want to save outputs to another folder on your local machine, you may want to mount it as well: e.g. by adding

...
    --mount type=bind,source="/absolute/path/somewhere/else/",destination="/home/ubuntu/video_features/output/" \
...

Then, run your command. For instance:

# cd to `./video_features`
# ubuntu@56b1bf77a20c:~/video_features $
python main.py \
    feature_type=r21d \
    device="cuda:0" \
    video_paths="[./dataset/vid_1.mp4, ./dataset/vid_2.mp4]" \
    on_extraction="save_numpy"
# you should have features in `./output`
# (and in the source location if you mount to it)

Switching conda environments

By default, the video_features environment is activated once you attach the shell. The image supports both conda environments and you can switch it simply as follows:

# ubuntu@56b1bf77a20c:~$
conda activate video_features
# which python