Drag and Drop Question You use a training pipeline in the Azure Machine Learning designer. You register a datastore named ds1. The datastore contains multiple training data files. You use the Import…

Question

Exhibit

Answer Area

Explanation

To efficiently retrain an Azure ML model with different data, the pipeline's data input path must be parameterized, then the pipeline published, and finally run with the new parameter value.

Approach. The correct approach involves parameterizing the Import Data module's path to allow flexible data input for retraining, followed by publishing and running the pipeline. The sequence is as follows:

1. Specify a new path to the training file as a parameter value.

   • Before you can create a parameter for a path, you first decide or identify what that new path (value) will be. This step conceptually establishes the target data location for the retraining process that will be driven by a parameter.

2. Add a new parameter in the module indicating the path to the training file.

   • Once the new path is determined, you modify the Import Data module in the Azure Machine Learning designer. You designate the path input field as a pipeline parameter. This action creates the parameter placeholder that can be configured with values at runtime or publication.

3. Publish a training pipeline.

   • After the pipeline has been configured with the new parameter, it needs to be published. Publishing makes the pipeline executable as a REST endpoint or a reusable component within the Azure ML workspace, allowing it to be invoked with specific parameter values.

4. Run the training pipeline by using the studio portal.

   • Finally, with the pipeline published and parameterized, you can initiate a run through the Azure ML studio portal. During this run, you will provide the 'new path to the training file' as the specific value for the parameter defined earlier, thereby retraining the model on the different dataset.

Common mistakes.

• common_mistake. A common mistake would be attempting to 'Register each training file as a new datastore'. Datastores typically point to storage containers or folders, not individual files, especially when dealing with 'multiple training data files'. Registering a new datastore for each file is inefficient, unnecessary, and not the standard method for simply changing the input data path for an existing pipeline. Another mistake would be to try to re-edit the Import Data module's hardcoded path every time, which lacks the flexibility and automation that pipeline parameters provide. Incorrectly ordering the 'Add parameter' and 'Specify parameter value' steps (e.g., adding a parameter after trying to specify its value at runtime without prior definition) or placing 'Publish' after 'Run' would also be incorrect, as publishing is a prerequisite for running a parameterized pipeline in a reusable manner.

Concept tested. Azure Machine Learning pipeline parameterization, dynamic data input for model retraining, and the lifecycle of an Azure ML pipeline (design, parameterization, publishing, and execution).

No community discussion yet for this question.