Training and Running CoreML Models using GitHub Actions

Apple's CoreML provides a way to run model predictions on iOS and OSX. Some parts of the libarary, such as the model conversion tools, can be run on non-Apple platforms, but the prediction methods will fail unless run on an iPhone (iOS) or Mac (OSX). This can present a problem for datascientists and developers without access to an Apple device.

Fortunately, GitHub Actions provides macOS as a virtual machine type for running jobs. This makes it possible for developers to run CoreML predictions from inside a GitHub Actions workflow.

Training the model

First, we'll train an extremely simple sklearn model and export it to CoreML: see train.py for details, or clone this repository and run:

pip install -r requirements.txt
python train.py 

This ingests known square footage and listing prices of some Seattle-area homes from prices.csv, fits a linear regression, and exports pricing.mlmodel.

Running a prediction

On iOS or OSX only, we can run predict.py, which ingests a list of square-footage values from input.csv and uses pricing.mlmodel to generate and save a list of predicted prices as output.csv.

pip install coremltools
python predict.py

Setting up GitHub Actions

We can now configure a sequence of actions to be triggered when specific files are pushed to our repository. We'll set up two separate workflows, both stored in /.github/workflows:

• train.yml will be executed whenever prices.csv or train.py changes, and will run train.py to generate a new pricing.mlmodel

• predict.yml will be executed whenever input.csv, predict.py, or pricing.mlmodel changes. It will run predict.py to generate a new output.csv

In addition, either job will be run if the job configuration file itself is changed, e.g.:

on: 
  push:
    paths:
    - .github/workflows/predict.yml
    - input.csv
    - predict.py
    - pricing.mlmodel

Let's examine each section of the jobs/build section of predict.yml to better understand what's going on:

runs-on: macOS-latest
strategy:
  max-parallel: 1
  matrix:
    python-version: [3.6]

Here, we specify that this job will run in the latest OSX container, with Python 3.6 installed.

steps:
- uses: actions/checkout@v1
- name: Set up Python ${{ matrix.python-version }}
  uses: actions/setup-python@v1
  with:
    python-version: ${{ matrix.python-version }}

Our first steps check out the current repo and set up the specified Python.

- name: Install dependencies
  run: |
    python -m pip install --upgrade pip
    pip install coremltools
- name: Run predictions
  run: python predict.py

Next, we install our dependencies and run the prediction script.

- name: Commit changed output
  env:
    GITHUB_ACTOR: ${{ secrets.GITHUB_ACTOR }}
    GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
    GITHUB_REPOSITORY: ${{ secrets.GITHUB_REPOSITORY }}
  run: |
    git remote set-url origin "https://$GITHUB_ACTOR:$GITHUB_TOKEN@github.com/$GITHUB_REPOSITORY.git"
    git commit -m "updated predictions" output.csv
    git push origin HEAD:master

Lastly, we commit our changed file(s) back to the repo. This requires some backflips using runtime variables, as our code inside the workflow has been automatically checked out on a detached branch.

Try it yourself

Read more about GitHub actions, then clone this repository and commit a change to the input.csv file. Click the "Actions" tab at the top of your repo to examine the run, then check to see if output.csv has changed.

Enjoy!