Spleeter: Improvement: Include the model artifact into the Docker image (instead of downloading at run time)

[iMerica]

Image building happens once, running the container happens N times. Theres no reason to re-download the model every time it runs.

[pppq]

I also tried adding a volume for it at /model, mentioned in the Getting started page of the wiki, but it does not seem to pick up a pre-downloaded instance.

[alreadytaikeune]

Hello,

There is indeed a problem in the wiki. /model doesn't correspond to the path where the models are actually extracted it seems. They are in /workspace/spleeter/pretrained_models, so you can mount a volume there, e.g.

nvidia-docker run -v $(pwd)/output:/output -v $(pwd)/pretrained_models:/workspace/spleeter/pretrained_models ...

[pppq]

Thanks! I see now that the MODEL_PATH environment variable can also control the location within the container, so I could make the example work without re-download by adding -e MODEL_PATH=/model to the command.

[alreadytaikeune]

Indeed, you can also keep the command given in the wiki, and set the environment variable -e MODEL_PATH=/model when running the container.

[Faylixe]

We will probably also at some point distribute ready to use Docker image on DockerHub which will embed model directly.

[mickdekkers]

In the interim, you can add this to your Dockerfile somewhere after the WORKDIR /workspace/spleeter command:

RUN mkdir -p ./pretrained_models/2stems && \
    wget -nv https://github.com/deezer/spleeter/releases/download/v1.4.0/2stems.tar.gz && \
    tar -xvzf 2stems.tar.gz -C ./pretrained_models/2stems && \
    rm -f 2stems.tar.gz

This will download and extract the 2stems model to the right place (adjust for different models). Note that you'll need to rebuild after updating your Dockerfile.

[romi1502]

I fixed the command in the wiki with the extra -e MODEL_PATH=/model option.
As suggested by @mickdekkers, you can also build your own image with the model you need (avoiding including all models when you only need one).