Local Quickstart

In this tutorial, we will walk through the training, packaging and serving of a machine learning model locally by leveraging ODAHUFlow’s main components.

Prerequisites

  • Docker engine (at least version 17.0) with access from current user (docker ps should executes without errors)
  • Odahu-flow CLI
  • git
  • wget

Tutorial

We will consider the wine model from Cluster Quickstart. But now, we will train, package and deploy the model locally.

Note

Code for the tutorial is available on GitHub.

odahuflowctl has commands for local training and packaging.

$ odahuflowctl local --help

To train a model locally, you have to provide an ODAHU model training manifest and training toolchain. odahuflowctl tries to find them on your local filesystem. If it can not do it, then the CLI requests to ODAHU API.

Local training arguments:
  --train-id, --id TEXT        Model training ID  [required]
  -f, --manifest-file PATH     Path to a ODAHU-flow manifest file
  -d, --manifest-dir PATH      Path to a directory with ODAHU-flow manifests

The mlflow/sklearn/wine/odahuflow directory already contains training manifest file for wine model. If we don’t have a running ODAHUFlow API server, we should create toolchain manifest manually.

Paste the toolchain manifest into the mlflow/sklearn/wine/odahuflow/toolchain.yaml file:

kind: ToolchainIntegration
id: mlflow
spec:
  defaultImage: "odahu/odahu-flow-mlflow-toolchain:1.1.0-rc11"
  entrypoint: /opt/conda/bin/odahu-flow-mlflow-runner

We are ready to launch the local training. Copy, past and execute the following command.

$ odahuflowctl local train run -d mlflow/sklearn/wine/odahuflow --id wine

Warning

MLFlow metrics does not propagate to the tracking server during training. This will be implemented in the near future.

odahuflowctl trains the model, verify that it satisfy the GPPI spec and save GPPI binary in the host filesystem. Execute the following command to take a look all trained models.

$ odahuflowctl local train list

Our next step is to package the trained model to a REST service. Like for local training, local packaging requires a model packaging and packaging integration manifests.

Local packaging arguments:
  --pack-id, --id TEXT            Model packaging ID  [required]
  -f, --manifest-file PATH        Path to a ODAHU-flow manifest file
  -d, --manifest-dir PATH         Path to a directory with ODAHU-flow manifest files
  --artifact-path PATH            Path to a training artifact
  -a, --artifact-name TEXT        Override artifact name from file

Paste the packaging integration manifest into the mlflow/sklearn/wine/odahuflow/packager.yaml file:

kind: PackagingIntegration
id: docker-rest
spec:
  entrypoint: "/usr/local/bin/odahu-flow-pack-to-rest"
  defaultImage: "odahu/odahu-flow-packagers:1.1.0-rc11"
  privileged: true
  schema:
    targets:
      - name: docker-push
        connectionTypes: ["docker", "ecr"]
        required: true
      - name: docker-pull
        connectionTypes: ["docker", "ecr"]
        required: false
    arguments:
      properties:
        - name: dockerfileAddCondaInstallation
          parameters:
            - name: description
              value: Add conda installation code to training.Dockerfile
            - name: type
              value: boolean
            - name: default
              value: true
        - name: dockerfileBaseImage
          parameters:
            - name: description
              value: Base image for training.Dockerfile.
            - name: type
              value: string
            - name: default
              value: 'odahu/odahu-flow-docker-packager-base:1.1.0-rc11'
        - name: dockerfileCondaEnvsLocation
          parameters:
            - name: description
              value: Conda env location in training.Dockerfile.
            - name: type
              value: string
            - name: default
              value: /opt/conda/envs/
        - name: host
          parameters:
            - name: description
              value: Host to bind.
            - name: type
              value: string
            - name: default
              value: 0.0.0.0
        - name: port
          parameters:
            - name: description
              value: Port to bind.
            - name: type
              value: integer
            - name: default
              value: 5000
        - name: timeout
          parameters:
            - name: description
              value: Serving timeout in seconds.
            - name: type
              value: integer
            - name: default
              value: 60
        - name: workers
          parameters:
            - name: description
              value: Count of serving workers.
            - name: type
              value: integer
            - name: default
              value: 1
        - name: threads
          parameters:
            - name: description
              value: Count of serving threads.
            - name: type
              value: integer
            - name: default
              value: 4
        - name: imageName
          parameters:
            - name: description
              value: |
                This option provides a way to specify the Docker image name. You can hardcode the full name or specify a template. Available template values:
                  - Name (Model Name)
                  - Version (Model Version)
                  - RandomUUID
                The default value is '{{ Name }}/{{ Version }}:{{ RandomUUID }}'.
                Image name examples:
                  - myservice:123
                  - {{ Name }}:{{ Version }}
            - name: type
              value: string
            - name: default
              value: "{{ Name }}-{{ Version }}:{{ RandomUUID }}"

Choose the name of trained artifact and execute the following command:

$ odahuflowctl --verbose local pack run -d mlflow/sklearn/wine/odahuflow --id wine -a wine-1.0-wine-1.0-01-Mar-2020-18-33-35

The last lines of output must contains a name of model REST service.

At the last step, we run our REST service and make a predict.

$ docker run -it --rm -p 5000:5000 wine-1.0:cbf184d0-4b08-45c4-8efb-17e28a3b537e
$ odahuflowctl model invoke --url http://0:5000 --json-file mlflow/sklearn/wine/odahuflow/request.json