Skip to content

StartBaseScenario.md

Latest commit

 

History

History
111 lines (101 loc) · 5.83 KB

StartBaseScenario.md

File metadata and controls

111 lines (101 loc) · 5.83 KB

Run an end-to-end MLOps pipeline

Section 1: Setup Github & Integration with Devops

  1. Perform the initial setup if not already done
  2. Fork the MLOps repository
    1. Visit the MLOps github project and click Fork on top right.
    2. Give a name & let it remain "Public" for course of this workshop
    3. From your github repository, directly edit the file mlops/recipes/common/Variables.yml: Change the BASE_NAME as per the instructions given there. Note: If you wish to reuse your existing ML workspace, then update the Variables.yml with the details
    4. In your project page copy your clone https url by clicking on the Clone or Download button - you will need this for next step
  3. Setup integration between Devops & Github
    1. In the Azure Devops project you created earlier goto: Pipelines -> Pipelines
    2. Create new pipeline by clicking "New pipeline" button on top right
    3. To respond to question "Where is your code", click Github (YAML)
      1. Authorize the two requests: One for Azure devops to connect to github to load your pipelines & another for integration of triggers from github to Devops
      2. Click Existing Azure Pipelines YAML file
      3. Now you will see a list of YML (yaml) files. Follow next step below.

Section 2: Create the environment

Note: Skip this step if you wish to reuse your existing Azure ML workspace
This step will create the cloud environment and provision all the required services including Resource group & Azure ML Workspace

  1. Continuing from the last section: select mlops/recipes/IaC/ProvisionMLWorkspace.yml -> Click Continue -> Click Run
  2. You can monitor the status by clicking on the stages. Check if your pipeline ran successfully.
  3. [Optional] Rename the pipeline run in Azure Devops to give a meaningful name. You can do this by visiting Pipeline in the left nav bar of Azure Devops. You can then From the recent runs, you can click the three dots in the right and click rename

Section 3: Create a Dataset in Azure ML

  1. Login into the Azure ML portal and switch to the new workspace that was created using the build pipeline:
    1. In the top nav bar, click the Switch Directory icon
    2. Click on the drop down Machine learning workspace and select the one that was just created (BASE_NAME-aml-ws)
  2. In the left panel click Dataset -> Create Dataset -> from web files
    1. Paste this url https://raw.githubusercontent.com/rsethur/MLOps/master/models/risk-model/dataset/german_credit_data.csv
    2. Important Change dataset name to credit_dataset
    3. If a CORS error shows up, open the page and check all boxes in the Allowed methods drop down.
    4. Follow the onscreen instructions by clicking Next till completion

Section 4: Run the Build Release pipeline

Now we can run training & automate the deployment by running the build release pipeline.

  1. Similar to first pipeline we ran, in the Azure Devops project goto: Pipelines -> Pipelines
  2. Create new pipeline by clicking "New pipeline" button on top right
  3. To respond to question "Where is your code", click Github (YAML)
    1. Click Existing Azure Pipelines YAML file
    2. Now you will see a list of YML (yaml) files: select mlops/model_pipelines/risk-model/BuildReleasePipeline.yml -> Click Continue -> Click Run

You can monitor the status by clicking on the stages. Check if your pipeline ran successfully.

[Optional] Section 5: Try out other pipelines

You can try various other pipelines from mlops/model_pipelines/risk-model. You will see pipelines fo

  1. Batch Scoring
  2. Release Pipeline for Pretrained model
  3. No code deploy
  4. In the snippets folder in mlops/model_pipelines/risk-model you can find few smaller atomic pipelines to play with

Section 5: Test the deployed service manually

  1. Get the service endpoint:
    1. Navigate to the Azure ML portal
    2. Go to Endpoints in the left nav bar -> click on the deployed service
    3. copy the url of the REST endpoint from the details
  2. Two options for testing the end point
    1. Easy option:
      1. Go to the prefilled request details in this page
      2. Replace the existing URL with the new endpoint url -> click the Test button
      3. You should see predictions in the response body.
    2. Another option:
      1. Use your favourite REST endpoint testing tool like Postman or API Tester (online)
      2. Fill the details:
        1. Your rest endpoint URL
        2. Set content-type to application/json
        3. Request type to POST
        4. Copy the following post data -> press submit/test
        {
  'data': {
    "Age": [
      20
    ],
    "Sex": [
      "male"
    ],
    "Job": [
      0
    ],
    "Housing": [
      "own"
    ],
    "Saving accounts": [
      "little"
    ],
    "Checking account": [
      "little"
    ],
    "Credit amount": [
      100
    ],
    "Duration": [
      48
    ],
    "Purpose": [
      "radio/TV"
    ]
  }
}

Section 6: Clean up

Once you understand the concepts & ready to delete the resources you can fo the the following:

From Azure Devops run the pipeline mlops/EnvTearDownPipeline.yml. Alternatively you can login to Azure Portal and delete the resource group that we created (i.e. your-unique-name-aml-rg).

Do ⭐ the repo if you like it.