What is CML? Continuous Machine Learning (CML) is an open-source CLI tool for implementing continuous integration & delivery (CI/CD) with a focus on MLOps. Use it to automate development workflows — including machine provisioning, model training and evaluation, comparing ML experiments across project history, and monitoring changing datasets.

CML can help train and evaluate models — and then generate a visual report with results and metrics — automatically on every pull request.

An example report for a neural style transfer model.

CML principles:

• GitFlow for data science. Use GitLab or GitHub to manage ML experiments, track who trained ML models or modified data and when. Codify data and models with DVC instead of pushing to a Git repo.
• Auto reports for ML experiments. Auto-generate reports with metrics and plots in each Git pull request. Rigorous engineering practices help your team make informed, data-driven decisions.
• No additional services. Build your own ML platform using GitLab, Bitbucket, or GitHub. Optionally, use cloud storage as well as either self-hosted or cloud runners (such as AWS EC2 or Azure). No databases, services or complex setup needed.

❓ Need help? Just want to chat about continuous integration for ML? Visit our Discord channel!

⏯️ Check out our YouTube video series for hands-on MLOps tutorials using CML!

Table of Contents

1. Setup (GitLab, GitHub, Bitbucket)
2. Usage
3. Getting started (tutorial)
4. Using CML with DVC
5. Advanced Setup (Self-hosted, local package)
6. Example projects

Setup

You'll need a GitLab, GitHub, or Bitbucket account to begin. Users may wish to familiarize themselves with Github Actions or GitLab CI/CD. Here, will discuss the GitHub use case.

GitLab

Please see our docs on CML with GitLab CI/CD and in particular the personal access token requirement.

Bitbucket

Please see our docs on CML with Bitbucket Cloud.

GitHub

The key file in any CML project is .github/workflows/cml.yaml:

name: your-workflow-name
on: [push]
jobs:
  run:
    runs-on: ubuntu-latest
    # optionally use a convenient Ubuntu LTS + DVC + CML image
    # container: ghcr.io/iterative/cml:0-dvc2-base1
    steps:
      - uses: actions/checkout@v3
      # may need to setup NodeJS & Python3 on e.g. self-hosted
      # - uses: actions/setup-node@v3
      #   with:
      #     node-version: '16'
      # - uses: actions/setup-python@v4
      #   with:
      #     python-version: '3.x'
      - uses: iterative/setup-cml@v1
      - name: Train model
        run: |
          # Your ML workflow goes here
          pip install -r requirements.txt
          python train.py
      - name: Write CML report
        env:
          REPO_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          # Post reports as comments in GitHub PRs
          cat results.txt >> report.md
          cml comment create report.md

Usage

We helpfully provide CML and other useful libraries pre-installed on our custom Docker images. In the above example, uncommenting the field container: ghcr.io/iterative/cml:0-dvc2-base1) will make the runner pull the CML Docker image. The image already has NodeJS, Python 3, DVC and CML set up on an Ubuntu LTS base for convenience.

CML Functions

CML provides a number of functions to help package the outputs of ML workflows (including numeric data and visualizations about model performance) into a CML report.

Below is a table of CML functions for writing markdown reports and delivering those reports to your CI system.

Function	Description	Example Inputs
cml runner launch	Launch a runner locally or hosted by a cloud provider	See Arguments
cml comment create	Return CML report as a comment in your GitLab/GitHub workflow	<path to report> --head-sha <sha>
cml check create	Return CML report as a check in GitHub	<path to report> --head-sha <sha>
cml pr create	Commit the given files to a new branch and create a pull request	<path>...
cml tensorboard connect	Return a link to a Tensorboard.dev page	--logdir <path to logs> --title <experiment title> --md

CML Reports

The cml comment create command can be used to post reports. CML reports are written in markdown (GitHub, GitLab, or Bitbucket flavors). That means they can contain images, tables, formatted text, HTML blocks, code snippets and more — really, what you put in a CML report is up to you. Some examples:

🗒️ Text Write to your report using whatever method you prefer. For example, copy the contents of a text file containing the results of ML model training:

cat results.txt >> report.md

🖼️ Images Display images using the markdown or HTML. Note that if an image is an output of your ML workflow (i.e., it is produced by your workflow), it can be uploaded and included automaticlly to your CML report. For example, if graph.png is output by python train.py, run:

echo "![](./graph.png)" >> report.md
cml comment create report.md

Getting Started

1. Fork our example project repository.

⚠️ Note that if you are using GitLab, you will need to create a Personal Access Token for this example to work.

⚠️ The following steps can all be done in the GitHub browser interface. However, to follow along with the commands, we recommend cloning your fork to your local workstation:

git clone https://github.com/<your-username>/example_cml

2. To create a CML workflow, copy the following into a new file, .github/workflows/cml.yaml:

name: model-training
on: [push]
jobs:
  run:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v3
      - uses: actions/setup-python@v4
      - uses: iterative/setup-cml@v1
      - name: Train model
        env:
          REPO_TOKEN: ${{ secrets.GITHUB_TOKEN }}
        run: |
          pip install -r requirements.txt
          python train.py

          cat metrics.txt >> report.md
          echo "![](./plot.png)" >> report.md
          cml comment create report.md

3. In your text editor of choice, edit line 16 of train.py to depth = 5.

4. Commit and push the changes:

git checkout -b experiment
git add . && git commit -m "modify forest depth"
git push origin experiment

5. In GitHub, open up a pull request to compare the experiment branch to master.

Shortly, you should see a comment from github-actions appear in the pull request with your CML report. This is a result of the cml send-comment function in your workflow.

This is the outline of the CML workflow:

• you push changes to your GitHub repository,
• the workflow in your .github/workflows/cml.yaml file gets run, and
• a report is generated and posted to GitHub.

CML functions let you display relevant results from the workflow — such as model performance metrics and visualizations — in GitHub checks and comments. What kind of workflow you want to run, and want to put in your CML report, is up to you.

Using CML with DVC

In many ML projects, data isn't stored in a Git repository, but needs to be downloaded from external sources. DVC is a common way to bring data to your CML runner. DVC also lets you visualize how metrics differ between commits to make reports like this:

The .github/workflows/cml.yaml file used to create this report is:

name: model-training
on: [push]
jobs:
  run:
    runs-on: ubuntu-latest
    container: ghcr.io/iterative/cml:0-dvc2-base1
    steps:
      - uses: actions/checkout@v3
      - name: Train model
        env:
          REPO_TOKEN: ${{ secrets.GITHUB_TOKEN }}
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
        run: |
          # Install requirements
          pip install -r requirements.txt

          # Pull data & run-cache from S3 and reproduce pipeline
          dvc pull data --run-cache
          dvc repro

          # Report metrics
          echo "## Metrics" >> report.md
          git fetch --prune
          dvc metrics diff master --show-md >> report.md

          # Publish confusion matrix diff
          echo "## Plots" >> report.md
          echo "### Class confusions" >> report.md
          dvc plots diff --target classes.csv --template confusion -x actual -y predicted --show-vega master > vega.json
          vl2png vega.json -s 1.5 > confusion_plot.png
          echo "![](./confusion_plot.png)" >> report.md

          # Publish regularization function diff
          echo "### Effects of regularization" >> report.md
          dvc plots diff --target estimators.csv -x Regularization --show-vega master > vega.json
          vl2png vega.json -s 1.5 > plot.png
          echo "![](./plot.png)" >> report.md

          cml comment create report.md

⚠️ If you're using DVC with cloud storage, take note of environment variables for your storage format.

Configuring Cloud Storage Providers

There are many supported could storage providers. Here are a few examples for some of the most frequently used providers:

S3 and S3-compatible storage (Minio, DigitalOcean Spaces, IBM Cloud Object Storage...)

# Github
env:
  AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
  AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
  AWS_SESSION_TOKEN: ${{ secrets.AWS_SESSION_TOKEN }}

👉 AWS_SESSION_TOKEN is optional.

👉 AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY can also be used by cml runner to launch EC2 instances. See [Environment Variables].

Azure

env:
  AZURE_STORAGE_CONNECTION_STRING:
    ${{ secrets.AZURE_STORAGE_CONNECTION_STRING }}
  AZURE_STORAGE_CONTAINER_NAME: ${{ secrets.AZURE_STORAGE_CONTAINER_NAME }}

Aliyun

env:
  OSS_BUCKET: ${{ secrets.OSS_BUCKET }}
  OSS_ACCESS_KEY_ID: ${{ secrets.OSS_ACCESS_KEY_ID }}
  OSS_ACCESS_KEY_SECRET: ${{ secrets.OSS_ACCESS_KEY_SECRET }}
  OSS_ENDPOINT: ${{ secrets.OSS_ENDPOINT }}

Google Storage

⚠️ Normally, GOOGLE_APPLICATION_CREDENTIALS is the path of the json file containing the credentials. However in the action this secret variable is the contents of the file. Copy the json contents and add it as a secret.

env:
  GOOGLE_APPLICATION_CREDENTIALS: ${{ secrets.GOOGLE_APPLICATION_CREDENTIALS }}

Google Drive

⚠️ After configuring your Google Drive credentials you will find a json file at your_project_path/.dvc/tmp/gdrive-user-credentials.json. Copy its contents and add it as a secret variable.

env:
  GDRIVE_CREDENTIALS_DATA: ${{ secrets.GDRIVE_CREDENTIALS_DATA }}

Advanced Setup

Self-hosted (On-premise or Cloud) Runners

GitHub Actions are run on GitHub-hosted runners by default. However, there are many great reasons to use your own runners: to take advantage of GPUs, orchestrate your team's shared computing resources, or train in the cloud.

☝️ Tip! Check out the official GitHub documentation to get started setting up your own self-hosted runner.

Allocating Cloud Compute Resources with CML

When a workflow requires computational resources (such as GPUs), CML can automatically allocate cloud instances using cml runner. You can spin up instances on AWS, Azure, GCP, or Kubernetes.

For example, the following workflow deploys a g4dn.xlarge instance on AWS EC2 and trains a model on the instance. After the job runs, the instance automatically shuts down.

You might notice that this workflow is quite similar to the basic use case above. The only addition is cml runner and a few environment variables for passing your cloud service credentials to the workflow.

Note that cml runner will also automatically restart your jobs (whether from a GitHub Actions 35-day workflow timeout or a AWS EC2 spot instance interruption).

name: Train-in-the-cloud
on: [push]
jobs:
  deploy-runner:
    runs-on: ubuntu-latest
    steps:
      - uses: iterative/setup-cml@v1
      - uses: actions/checkout@v3
      - name: Deploy runner on EC2
        env:
          REPO_TOKEN: ${{ secrets.PERSONAL_ACCESS_TOKEN }}
          AWS_ACCESS_KEY_ID: ${{ secrets.AWS_ACCESS_KEY_ID }}
          AWS_SECRET_ACCESS_KEY: ${{ secrets.AWS_SECRET_ACCESS_KEY }}
        run: |
          cml runner launch \
            --cloud=aws \
            --cloud-region=us-west \
            --cloud-type=g4dn.xlarge \
            --labels=cml-gpu
  train-model:
    needs: deploy-runner
    runs-on: [self-hosted, cml-gpu]
    timeout-minutes: 50400 # 35 days
    container:
      image: ghcr.io/iterative/cml:0-dvc2-base1-gpu
      options: --gpus all
    steps:
      - uses: actions/checkout@v3
      - name: Train model
        env:
          REPO_TOKEN: ${{ secrets.PERSONAL_ACCESS_TOKEN }}
        run: |
          pip install -r requirements.txt
          python train.py

          cat metrics.txt > report.md
          cml comment create report.md

In the workflow above, the deploy-runner step launches an EC2 g4dn.xlarge instance in the us-west region. The model-training step then runs on the newly-launched instance. See [Environment Variables] below for details on the secrets required.

🎉 Note that jobs can use any Docker container! To use functions such as cml send-comment from a job, the only requirement is to have CML installed.

Docker Images

The CML Docker image (ghcr.io/iterative/cml or iterativeai/cml) comes loaded with Python, CUDA, git, node and other essentials for full-stack data science. Different versions of these essentials are available from different image tags. The tag convention is {CML_VER}-dvc{DVC_VER}-base{BASE_VER}{-gpu}:

{BASE_VER}	Software included (-gpu)
0	Ubuntu 18.04, Python 2.7 (CUDA 10.1, CuDNN 7)
1	Ubuntu 20.04, Python 3.8 (CUDA 11.2, CuDNN 8)

For example, iterativeai/cml:0-dvc2-base1-gpu, or ghcr.io/iterative/cml:0-dvc2-base1.

Arguments

The cml runner launch function accepts the following arguments:

  --labels                                  One or more user-defined labels for
                                            this runner (delimited with commas)
                                                       [string] [default: "cml"]
  --idle-timeout                            Time to wait for jobs before
                                            shutting down (e.g. "5min"). Use
                                            "never" to disable
                                                 [string] [default: "5 minutes"]
  --name                                    Name displayed in the repository
                                            once registered
                                                    [string] [default: cml-{ID}]
  --no-retry                                Do not restart workflow terminated
                                            due to instance disposal or GitHub
                                            Actions timeout            [boolean]
  --single                                  Exit after running a single job
                                                                       [boolean]
  --reuse                                   Don't launch a new runner if an
                                            existing one has the same name or
                                            overlapping labels         [boolean]
  --reuse-idle                              Creates a new runner only if the
                                            matching labels don't exist or are
                                            already busy               [boolean]
  --docker-volumes                          Docker volumes, only supported in
                                            GitLab         [array] [default: []]
  --cloud                                   Cloud to deploy the runner
                         [string] [choices: "aws", "azure", "gcp", "kubernetes"]
  --cloud-region                            Region where the instance is
                                            deployed. Choices: [us-east,
                                            us-west, eu-west, eu-north]. Also
                                            accepts native cloud regions
                                                   [string] [default: "us-west"]
  --cloud-type                              Instance type. Choices: [m, l, xl].
                                            Also supports native types like i.e.
                                            t2.micro                    [string]
  --cloud-permission-set                    Specifies the instance profile in
                                            AWS or instance service account in
                                            GCP           [string] [default: ""]
  --cloud-metadata                          Key Value pairs to associate
                                            cml-runner instance on the provider
                                            i.e. tags/labels "key=value"
                                                           [array] [default: []]
  --cloud-gpu                               GPU type. Choices: k80, v100, or
                                            native types e.g. nvidia-tesla-t4
                                                                        [string]
  --cloud-hdd-size                          HDD size in GB              [number]
  --cloud-ssh-private                       Custom private RSA SSH key. If not
                                            provided an automatically generated
                                            throwaway key will be used  [string]
  --cloud-spot                              Request a spot instance    [boolean]
  --cloud-spot-price                        Maximum spot instance bidding price
                                            in USD. Defaults to the current spot
                                            bidding price [number] [default: -1]
  --cloud-startup-script                    Run the provided Base64-encoded
                                            Linux shell script during the
                                            instance initialization     [string]
  --cloud-aws-security-group                Specifies the security group in AWS
                                                          [string] [default: ""]
  --cloud-aws-subnet,                       Specifies the subnet to use within
  --cloud-aws-subnet-id                     AWS           [string] [default: ""]

Environment Variables

⚠️ You will need to create a personal access token (PAT) with repository read/write access and workflow privileges. In the example workflow, this token is stored as PERSONAL_ACCESS_TOKEN.

ℹ️ If using the --cloud option, you will also need to provide access credentials of your cloud compute resources as secrets. In the above example, AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY (with privileges to create & destroy EC2 instances) are required.

For AWS, the same credentials can also be used for configuring cloud storage.

Proxy support

CML support proxy via known environment variables http_proxy and https_proxy.

On-premise (Local) Runners

This means using on-premise machines as self-hosted runners. The cml runner launch function is used to set up a local self-hosted runner. On a local machine or on-premise GPU cluster, install CML as a package and then run:

cml runner launch \
  --repo=$your_project_repository_url \
  --token=$PERSONAL_ACCESS_TOKEN \
  --labels="local,runner" \
  --idle-timeout=180

The machine will listen for workflows from your project repository.

Local Package

In the examples above, CML is installed by the setup-cml action, or comes pre-installed in a custom Docker image pulled by a CI runner. You can also install CML as a package:

npm install --location=global @dvcorg/cml

You can use cml without node by downloading the correct standalone binary for your system from the asset section of the releases.

You may need to install additional dependencies to use DVC plots and Vega-Lite CLI commands:

sudo apt-get install -y libcairo2-dev libpango1.0-dev libjpeg-dev libgif-dev \
                        librsvg2-dev libfontconfig-dev
npm install -g vega-cli vega-lite

CML and Vega-Lite package installation require the NodeJS package manager (npm) which ships with NodeJS. Installation instructions are below.

Install NodeJS

• GitHub: This is probably not necessary when using GitHub's default containers or one of CML's Docker containers. Self-hosted runners may need to use a set up action to install NodeJS:

uses: actions/setup-node@v3
  with:
    node-version: '16'

• GitLab: Requires direct installation.

curl -sL https://deb.nodesource.com/setup_16.x | bash
apt-get update
apt-get install -y nodejs

See Also

These are some example projects using CML.

• Basic CML project
• CML with DVC to pull data
• CML with Tensorboard
• CML with a small EC2 instance 🔑
• CML with EC2 GPU 🔑

🔑 needs a PAT.

⚠️ Maintenance ⚠️

• ~2023-07 Nvidia has dropped container CUDA images with 10.x/cudnn7 and 11.2.1, CML images will be updated accrodingly