Skip to content

Latest commit

 

History

History
50 lines (41 loc) · 3.74 KB

TrainAndDeployViaCode.md

File metadata and controls

50 lines (41 loc) · 3.74 KB

Train & Deploy a model using Azure ML Python SDK

Step 1: Clone the Azure ML Samples

  1. Goto to new ML workspace (Called Studio now)
  2. Navigate to Notebooks in the left nav
  3. Expand the Azure ML gallery ➡️ Samples ➡️ Python ➡️ 1.x.x
  4. Hover over How-to-use-azureml ➡️ click on the three dots displayed in the right side ➡️ click Clone.

Now you will see the folder under your user files below.

Step 2: Create a notebook VM & launch it

  1. Navigate to Compute in the left nav
  2. Notebook VM Tab: Click +New -> choose a name and VM type (e.g. STANDARD_DS13_V2 which has 8 VCPUs)
  3. Click Create
  4. Once the Notebook VM is provisioned, you will see three links: JupyterLab, Jupyter and R-Studio. Open Jupyter and follow next steps.

Step 3: Train a model (Diabetes Dataset)

  1. In Jupyter, open the notebook train-on-amlcompute.ipynb, in the folder /howto-use-azureml/training/train-on-amlcompute
    It will prompt you to select a kernel: select Python 3.6 AzureML and click button Set Kernel. If you miss it accidentally, you can set it by going to top menu of the notebook: Kernel -> Change kernel

  2. Run the notebook cell by cell

    • Important:
      1. When you are running the cells, it will ask you for interactive authentication: you need to follow the steps (follow the link and enter the supplied code).
      2. In the notebook no need to run from Section Provision as a persistent compute target (Advanced) onwards

  3. After training is done, cleanup the compute cluster: In the last line uncomment this and execute cpu_cluster.delete()

  4. Explore your experiment run:

    1. In the new Azure ML workspace, navigate to the Experiments section in left nav
    2. Click on experiment name train-on-amlcompute -> Run
    3. Navigate to the Metrics tab. Check alpha and mse if not already checked
    4. Navigate to Outputs tab and you can see all the model files (pkl files) created by train.py. In the train.py you will see the model files are stored in ./outputs directory. Azure ML framework automatically moves all files in the outputs folder to the cloud workspace (including logs)

  5. Take time to understand the code in the notebook and train.py

    Optional (Hyper Param Tuning)

    Time permitting you can try this: In Jupyter landing page -> top nav How to use Azure ML -> Ml Frameworks -> Scikit Learn -> click clone on train-hyperparameter-tune-deploy-with-sklearn. This uses the diabetes dataset used by the next section (Deploy).

Step 4: Deploy and score the model (Iris Dataset)

  1. In Jupyter, open the notebook model-register-and-deploy.ipynb, in the folder howto-use-azureml/deployment/deploy-to-cloud/
  2. Run the notebook cell by cell
    Note: Before running the last cell that deletes the deployed service, goto the new ML workspace and checkout Endpoints in the left nav
  3. Take time to understand the code in the notebook and score.py

Optional Excercises

  1. Leverage AutoML via Azure ML Python SDK :Execute any automl notebook excercise from MachineLearningNotebooks/how-to-use-azureml/automated-machine-learning
  2. Leverage Databricks via Azure ML Python SDK : This is bit more involved as you need to execute this from Databricks workspace. Details are here: MachineLearningNotebooks/how-to-use-azureml/automated-machine-learning/azure-databricks/README.md

Clean up:

Wait for the discussion to be over in the workshop before cleaning up.
You can cleanup following instructions here