Build and deploy containerized models

How to deploy a container to the cloud.

https://github.com/JDACS4C-IMPROVE/Singularity/blob/master/README.md#deploying-an-improve-containert-to-the-cloud

Requirements

  • git

  • singularity

  • IMPROVE Singularity repository

  • Path to a directory on a local file system (not shared filesystem, e.g. your home directory is usually on a shared filesystem)

  • Build permissions with singularity, e.g. fakeroot

Installation

Build singularity images for IMPROVE models and deploy them to a shared location.

  1. Clone the repository into a location of your choice:

    git clone https://github.com/JDACS4C-IMPROVE/Singularity.git

  2. Create your config file:

    cd Singularity
./setup

  3. To customize your setup modify the config in config/improve.env to change the workspace and data directory.

  4. Build and deploy:

    1. deploy image files to default location ./images:

      make configure && make deploy

    2. If you want to deploy the images at a different location invoke the make command and set PREFIX to a path of your choosing, default is the current directory. The deploy process will create an image directory at the specified location and copy the image files into it.

      make configure && make deploy PREFIX=/my/deploy/path/

  5. Every container has a standardized script for training a model called train.sh and a standard location (/candle_data_dir) for model input and output. To train a model you have to make your data directory available inside the container as /candle_data_dir.

    export IMPROVE_DATA_DIR=$HOME/improve_data_dir
export GPUID=0
export CONTAINER=images/GraphDRP.sif

singularity exec --nv --bind ${IMPROVE_DATA_DIR}:/candle_data_dir ${CONTAINER} train.sh ${GPUID} /candle_data_dir

With:

  • IMPROVE_DATA_DIR path to data directory

  • CONTAINER path/and/name.sif of image file

  • GPUID ID of GPU to use