Skip to content

This repository contains necessary infrastructure elements for automatically adding and updating docker and singularity images for models and model groups in Kipoi model zoo. These images are pre-activated with a compatible conda environment where all the model (group) specific dependencies have been installed.

License

Notifications You must be signed in to change notification settings

kipoi/kipoi-containers

Repository files navigation

kipoi-containers

Python >=3.9 Coverage Status License: MIT

alt text This repository contains necessary infrastructure elements for adding and updating docker and singularity images for models and model groups in Kipoi model zoo. These images are pre-activated with a compatible conda environment where all the model (group) specific dependencies have been installed.

Motivation

Example usage of kipoi

kipoi env create Basset
source activate kipoi-Basset

kipoi predict Basset \
--dataloader_args='{"intervals_file": "example/intervals_file", "fasta_file": "example/fasta_file"}' \
-o 'Basset.example_pred.tsv'

Main bottleneck

Kipoi uses conda for creating model specific environments.

  • It is impossible to gurantee that kipoi env create Basset resolves in every operating system since conda is not operating system agnostic.
  • It is cumbersome, labor intensive and error prone to pin all model dependencies across 31 and counting model groups in kipoi.
  • There is no gurantee that even if the dependencies are getting resolved now, they will continue to be resolved in future since the universe of python dependencies are ever changing.

Solution

Software containers were invented to handle exactly these problems by making a snapshot of a working system. We use both docker and singularity to make the containers as generalized as possible all the while remaining high performance computing cluster friendly.

Example usage of kipoi with singularity

kipoi predict Basset \
--dataloader_args='{"intervals_file": "example/intervals_file", "fasta_file": "example/fasta_file"}' \
-o 'Basset.example_pred.tsv' \
--singularity

Note: There is no need to create a separate environment as the container comes pre-installed with the model specific conda environment.

Example usage of kipoi with docker

docker run -v $PWD/app/ kipoi/kipoi-docker:sharedpy3keras2tf2 
kipoi predict Basset \
--dataloader_args='{'intervals_file': '/app/intervals.bed',
                    'fasta_file': '/app/ref.fa'}' \
-o '/app/Basset.example_pred.tsv'

Docker and singularity image hosting

  • Docker images are hosted in dockerhub.

  • Singularity images are hosted in zenodo.

  • Model specific docker and singularity image information and example usage are located under docker and singularity tab in each model's webpage at kipoi website such as here.

Installation

  • python>=3.9

  • Install docker

  • Install singularity

    • Singularity has been renamed to Apptainer. However, it is also possible to use SingularityCE from Sylabs. Current versions of kipoi containers are compatible with the latest version of Apptainer (1.0.2) and SingularityCE 3.9. Install Apptainer from here or SingularityCE from here.
  • Install kipoi_containers using pip install -e .

Environment variables

  1. DOCKER_USERNAME, DOCKER_PASSWORD

    • Only required for pushing the image to kipoi/kipoi-docker
    • Get it here
  2. ZENODO_ACCESS_TOKEN

    • Required for updating and pushing singularity images to zenodo using its rest api
    • Get it here. Make sure to check deposit:actions and deposit:write
  3. GITHUB_TOKEN

    • Required for syncing with Kipoi model zoo
    • Get it here. Make sure to add both read and write access
  4. SINGULARITY_PULL_FOLDER (Optional)

    • If specified, singularity images will be downloaded, built into and pushed from this folder. Otherwise, the current working directory is chosen as default.

Map between models (groups) and docker and singularity images

  • Docker: here

    • This maps models (groups) to a docker images. Each value here refers to a dockerhub image which can be pulled using docker cli/api.
  • Singularity: here

    • Each entry here has three keys
      • url: A globally accessible url for the image
      • name: Name of the image without any extension
      • md5: A md5 checksum used to ensure integrity during download

Sync with Kipoi model repo

As models get added and updated in the model repo, the respective docker and singularity containers should also be added and updated along with various json files in kipoi_contaners/container-info and github workflows in .github/workflows. Execute this as follows -

python kipoi_containers/updateoradd.py

If everything is succesfull kipoi_containers/kipoi-model-repo-hash will be updated to the most recent commit on the master branch of the model repo.

Tests

Testing the package

docker pull kipoi/kipoi-docker:mmsplice

pytest test-docker test-singularity test-containers/test_update_all_singularity_images.py

Testing the containers manually

Currently, there are two ways to test the docker and singularity images along with the models.

  1. Test model(s) at a time or model group(s) if it contains only one model within their respective docker and singularity containers

    pytest test-containers/test_models_from_command_line.py --model=KipoiSplice/4,Basenji
  2. Test any docker image which tests all compatible models or with a specific model group.

  • pytest test-containers/test_containers_from_command_line.py --image=kipoi/kipoi-docker:sharedpy3keras2tf1
  • pytest test-containers/test_containers_from_command_line.py --image=kipoi/kipoi-docker:sharedpy3keras2tf2 --modelgroup=HAL

Github action workflows

There are three different workflows at .github/workflow, each of which serves a different purpose. The necessary secrets and workflows are described below.

Github secrets

For a quick howto look here

  1. DOCKERUSERNAME and DOCKERPASSWORD
    • Correspond to values of env variables DOCKER_USERNAME and DOCKER_PASSWORD respectively
  2. ZENODOACCESSTOKEN
    • Corresponds to value of env variable ZENODO_ACCESS_TOKEN
  3. GITHUBPAT
    • Corresponds to value of env variable GITHUB_TOKEN

Workflows

  1. Continuous integration

    • Which
      • .github/workflows/test-images.yml
    • When
      • Push to any branch and pr to main branch in this repo
    • Why
      • kipoi_containers package is tested by this workflow
    • How
      • The package is built from scratch and tests specified in Tests section get executed. Additionally, one model from every model group gets tested within its docker and singularity containers.
  2. Sync with Kipoi model repo

    • Which

      • .github/workflows/sync-with-model-repo.yml
    • When

      • On demand and when a pull request is merged to master branch of model repo from here
    • Why

      • Keep the docker and singularity images up to date with the model definition in the model repo
    • How

      • Update existing images on dockerhub and zenodo if the model definiton has been updated
      • Add new images if new model has been added to the model repo
      • Create a new branch in model repo named target-json if it already does not exist
      • Update shared/containers/model-to-singularity.json in branch target-json of model repo if a singularity image has been updated in zenodo.
      • Update jsons in kipoi_containers/container-info/ and shared/containers/ in branch target-json of model repo in case a new model has been added
      • Update workflows in .github/workflow in case a new model has been added
      • Update kipoi_containers/kipoi-model-repo-hash
      • A pr is created automatically which then needs to be reviewed and merged.
  3. Build, test and push all docker and singularity images

    • Which

      • .github/workflows/release-workflow.yml
    • When

      • On demand and when a new package of kipoi is released to pypi from here
    • Why

      • Re-build, test and push the docker and singularity images. Some example scenarios -
        • kipoi pypi package has been updated
        • A new version has been released for continuumio/miniconda3:latest
    • How

      • Re-build, test and push the dockerhub images. Docker cli is used for this purpose.
      • A new version of the singularity image will be built based on the new docker image. A new version of the existing deposition on zenodo will be created and this modified image will be uploaded there. Finally, this new deposition will be pushed an url will be returned.
      • New url will be updated in shared/containers/model-to-singularity.json in branch target-json of model repo

About

This repository contains necessary infrastructure elements for automatically adding and updating docker and singularity images for models and model groups in Kipoi model zoo. These images are pre-activated with a compatible conda environment where all the model (group) specific dependencies have been installed.

Topics

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages