Expressions in Azure Machine Learning SDK and CLI v2

  • Article

With Azure Machine Learning SDK and CLI v2, you can use expressions when a value may not be known when you're authoring a job or component. When you submit a job or call a component, the expression is evaluated and the value is substituted.

The format for an expression is ${{ <expression> }}. Some expressions are evaluated on the client, when submitting the job or component. Other expressions are evaluated on the server (the compute where the job or component is running.)

Client expressions

Note

The "client" that evaluates the expression is where the job is submitted or component is ran. For example, your local machine or a compute instance.

Expression Description Scope
${{inputs.<input_name>}} References to an input data asset or model. Works for all jobs.
${{outputs.<output_name>}} References to an output data asset or model. Works for all jobs.
${{search_space.<hyperparameter>}} References the hyperparameters to use in a sweep job. The hyperparameter values for each trial are selected based on the search_space. Sweep jobs only.
${{parent.inputs.<input_name>}} Binds the inputs of a child job (pipeline step) in a pipeline to the inputs of the top-level parent pipeline job. Pipeline jobs only.
${{parent.outputs.<output_name>}} Binds the outputs of a child job (pipeline step) in a pipeline to the outputs of the top-level parent pipeline job. Pipeline jobs only.
${{parent.jobs.<step-name>.inputs.<input-name>}} Binds to the inputs of another step in the pipeline. Pipeline jobs only.
${{parent.jobs.<step-name>.outputs.<output-name>}} Binds to the outputs of another step in the pipeline. Pipeline jobs only.

Server expressions

Important

The following expressions are resolved on the server side, not the client side. For scheduled jobs where the job creation time and job submission time are different, the expressions are resolved when the job is submitted. Since these expressions are resolved on the server side, they use the current state of the workspace, not the state of the workspace when the scheduled job was created. For example, if you change the default datastore of the workspace after you create a scheduled job, the expression ${{default_datastore}} is resolved to the new default datastore, not the default datastore when the scheduled job was created.

Expression Description Scope
${{default_datastore}} If pipeline default datastore is configured, is resolved as pipeline default datastore name; otherwise is resolved as workspace default datastore name.

Pipeline default datastore can be controlled using pipeline_job.settings.default_datastore.		 Works for all jobs.

Pipeline jobs have a configurable pipeline default datastore.
${{name}} The job name. For pipelines, it's the step job name, not the pipeline job name. Works for all jobs
${{output_name}} The job output name Works for all jobs

For example, if azureml://datastores/${{default_datastore}}/paths/${{name}}/${{output_name}} is used as the output path, at runtime it's resolved as a path of azureml://datastores/workspaceblobstore/paths/<job-name>/model_path.

Next steps

For more information on these expressions, see the following articles and examples: