Previous

Next

Create a pipeline

Completed

In Azure Machine Learning, a pipeline is a workflow of machine learning tasks in which each task is defined as a component.

Components can be arranged sequentially or in parallel, enabling you to build sophisticated flow logic to orchestrate machine learning operations. Each component can be run on a specific compute target, making it possible to combine different types of processing as required to achieve an overall goal.

A pipeline can be executed as a process by running the pipeline as a pipeline job. Each component is executed as a child job as part of the overall pipeline job.

Build a pipeline

An Azure Machine Learning pipeline is defined in a YAML file. The YAML file includes the pipeline job name, inputs, outputs, and settings.

You can create the YAML file, or use the @pipeline() function to create the YAML file.

Tip

Review the reference documentation for the @pipeline() function.

For example, if you want to build a pipeline that first prepares the data, and then trains the model, you can use the following code:

from azure.ai.ml.dsl import pipeline

@pipeline()
def pipeline_function_name(pipeline_job_input):
    prep_data = loaded_component_prep(input_data=pipeline_job_input)
    train_model = loaded_component_train(training_data=prep_data.outputs.output_data)

    return {
        "pipeline_job_transformed_data": prep_data.outputs.output_data,
        "pipeline_job_trained_model": train_model.outputs.model_output,
    }

To pass a registered data asset as the pipeline job input, you can call the function you created with the data asset as input:

from azure.ai.ml import Input
from azure.ai.ml.constants import AssetTypes

pipeline_job = pipeline_function_name(
    Input(type=AssetTypes.URI_FILE, 
    path="azureml:data:1"
))

The @pipeline() function builds a pipeline consisting of two sequential steps, represented by the two loaded components.

To understand the pipeline built in the example, let's explore it step by step:

1. The pipeline is built by defining the function pipeline_function_name.
2. The pipeline function expects pipeline_job_input as the overall pipeline input.
3. The first pipeline step requires a value for the input parameter input_data. The value for the input will be the value of pipeline_job_input.
4. The first pipeline step is defined by the loaded component for prep_data.
5. The value of the output_data of the first pipeline step is used for the expected input training_data of the second pipeline step.
6. The second pipeline step is defined by the loaded component for train_model and results in a trained model referred to by model_output.
7. Pipeline outputs are defined by returning variables from the pipeline function. There are two outputs:
   • pipeline_job_transformed_data with the value of prep_data.outputs.output_data
   • pipeline_job_trained_model with the value of train_model.outputs.model_output

The result of running the @pipeline() function is a YAML file that you can review by printing the pipeline_job object you created when calling the function:

print(pipeline_job)

The output will be formatted as a YAML file, which includes the configuration of the pipeline and its components. Some parameters included in the YAML file are shown in the following example.

display_name: pipeline_function_name
type: pipeline
inputs:
  pipeline_job_input:
    type: uri_file
    path: azureml:data:1
outputs:
  pipeline_job_transformed_data: null
  pipeline_job_trained_model: null
jobs:
  prep_data:
    type: command
    inputs:
      input_data:
        path: ${{parent.inputs.pipeline_job_input}}
    outputs:
      output_data: ${{parent.outputs.pipeline_job_transformed_data}}
  train_model:
    type: command
    inputs:
      input_data:
        path: ${{parent.outputs.pipeline_job_transformed_data}}
    outputs:
      output_model: ${{parent.outputs.pipeline_job_trained_model}}
tags: {}
properties: {}
settings: {}

Tip

Learn more about the pipeline job YAML schema to explore which parameters are included when building a component-based pipeline.

Continue