How to do hyperparameter tuning in pipelines

Article
09/20/2024

APPLIES TO: Azure CLI ml extension v2 (current) Python SDK azure-ai-ml v2 (current)

In this article, you learn how to automate hyperparameter tuning in Azure Machine Learning pipelines by using Azure Machine Learning CLI v2 or Azure Machine Learning SDK for Python v2.

Hyperparameters are adjustable parameters that let you control the model training process. Hyperparameter tuning is the process of finding the configuration of hyperparameters that results in the best performance. Azure Machine Learning lets you automate hyperparameter tuning and run experiments in parallel to efficiently optimize hyperparameters.

Prerequisites

Have an Azure Machine Learning account and workspace.
Understand Azure Machine Learning pipelines and hyperparameter tuning a model.

Create and run a hyperparameter tuning pipeline

Azure CLI
Python SDK

The following examples come from Run a pipeline job using sweep (hyperdrive) in pipeline in the Azure Machine Learning examples repository. For more information about creating pipelines with components, see Create and run machine learning pipelines using components with the Azure Machine Learning CLI.

Create a command component with hyperparameter inputs

The Azure Machine Learning pipeline must have a command component with hyperparameter inputs. The following train.yml file from the example projects defines a trial component that has the c_value, kernel, and coef hyperparameter inputs and runs the source code that's located in the ./train-src folder.

$schema: https://azuremlschemas.azureedge.net/latest/commandComponent.schema.json
type: command

name: train_model
display_name: train_model
version: 1

inputs: 
  data:
    type: uri_folder
  c_value:
    type: number
    default: 1.0
  kernel:
    type: string
    default: rbf
  degree:
    type: integer
    default: 3
  gamma:
    type: string
    default: scale
  coef0: 
    type: number
    default: 0
  shrinking:
    type: boolean
    default: false
  probability:
    type: boolean
    default: false
  tol:
    type: number
    default: 1e-3
  cache_size:
    type: number
    default: 1024
  verbose:
    type: boolean
    default: false
  max_iter:
    type: integer
    default: -1
  decision_function_shape:
    type: string
    default: ovr
  break_ties:
    type: boolean
    default: false
  random_state:
    type: integer
    default: 42

outputs:
  model_output:
    type: mlflow_model
  test_data:
    type: uri_folder
  
code: ./train-src

environment: azureml://registries/azureml/environments/sklearn-1.5/labels/latest

command: >-
  python train.py 
  --data ${{inputs.data}}
  --C ${{inputs.c_value}}
  --kernel ${{inputs.kernel}}
  --degree ${{inputs.degree}}
  --gamma ${{inputs.gamma}}
  --coef0 ${{inputs.coef0}}
  --shrinking ${{inputs.shrinking}}
  --probability ${{inputs.probability}}
  --tol ${{inputs.tol}}
  --cache_size ${{inputs.cache_size}}
  --verbose ${{inputs.verbose}}
  --max_iter ${{inputs.max_iter}}
  --decision_function_shape ${{inputs.decision_function_shape}}
  --break_ties ${{inputs.break_ties}}
  --random_state ${{inputs.random_state}}
  --model_output ${{outputs.model_output}}
  --test_data ${{outputs.test_data}}

Create the trial component source code

The source code for this example is a single train.py file. This code executes in every trial of the sweep job.

# imports
import os
import mlflow
import argparse

import pandas as pd
from pathlib import Path

from sklearn.svm import SVC
from sklearn.model_selection import train_test_split

# define functions
def main(args):
    # enable auto logging
    mlflow.autolog()

    # setup parameters
    params = {
        "C": args.C,
        "kernel": args.kernel,
        "degree": args.degree,
        "gamma": args.gamma,
        "coef0": args.coef0,
        "shrinking": args.shrinking,
        "probability": args.probability,
        "tol": args.tol,
        "cache_size": args.cache_size,
        "class_weight": args.class_weight,
        "verbose": args.verbose,
        "max_iter": args.max_iter,
        "decision_function_shape": args.decision_function_shape,
        "break_ties": args.break_ties,
        "random_state": args.random_state,
    }

    # read in data
    df = pd.read_csv(args.data)

    # process data
    X_train, X_test, y_train, y_test = process_data(df, args.random_state)

    # train model
    model = train_model(params, X_train, X_test, y_train, y_test)
    # Output the model and test data
    # write to local folder first, then copy to output folder

    mlflow.sklearn.save_model(model, "model")

    from distutils.dir_util import copy_tree

    # copy subdirectory example
    from_directory = "model"
    to_directory = args.model_output

    copy_tree(from_directory, to_directory)

    X_test.to_csv(Path(args.test_data) / "X_test.csv", index=False)
    y_test.to_csv(Path(args.test_data) / "y_test.csv", index=False)


def process_data(df, random_state):
    # split dataframe into X and y
    X = df.drop(["species"], axis=1)
    y = df["species"]

    # train/test split
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=random_state
    )

    # return split data
    return X_train, X_test, y_train, y_test


def train_model(params, X_train, X_test, y_train, y_test):
    # train model
    model = SVC(**params)
    model = model.fit(X_train, y_train)

    # return model
    return model


def parse_args():
    # setup arg parser
    parser = argparse.ArgumentParser()

    # add arguments
    parser.add_argument("--data", type=str)
    parser.add_argument("--C", type=float, default=1.0)
    parser.add_argument("--kernel", type=str, default="rbf")
    parser.add_argument("--degree", type=int, default=3)
    parser.add_argument("--gamma", type=str, default="scale")
    parser.add_argument("--coef0", type=float, default=0)
    parser.add_argument("--shrinking", type=bool, default=False)
    parser.add_argument("--probability", type=bool, default=False)
    parser.add_argument("--tol", type=float, default=1e-3)
    parser.add_argument("--cache_size", type=float, default=1024)
    parser.add_argument("--class_weight", type=dict, default=None)
    parser.add_argument("--verbose", type=bool, default=False)
    parser.add_argument("--max_iter", type=int, default=-1)
    parser.add_argument("--decision_function_shape", type=str, default="ovr")
    parser.add_argument("--break_ties", type=bool, default=False)
    parser.add_argument("--random_state", type=int, default=42)
    parser.add_argument("--model_output", type=str, help="Path of output model")
    parser.add_argument("--test_data", type=str, help="Path of output model")

    # parse args
    args = parser.parse_args()

    # return args
    return args


# run script
if __name__ == "__main__":
    # parse args
    args = parse_args()

    # run main function
    main(args)

Note

Make sure to log the metrics in the trial component source code with exactly the same name as the primary_metric value in the pipeline file. This example uses mlflow.autolog(), which is the recommended way to track machine learning experiments. For more information about MLflow, see Track ML experiments and models with MLflow.

Create a pipeline with a hyperparameter sweep step

Azure CLI
Python SDK

Given the command component defined in train.yml, the following code creates a two-step train and predict pipeline definition file. In the sweep_step, the required step type is sweep, and the c_value, kernel, and coef hyperparameter inputs for the trial component are added to the search_space.

The following example highlights the hyperparameter tuning sweep_step.

$schema: https://azuremlschemas.azureedge.net/latest/pipelineJob.schema.json
type: pipeline
display_name: pipeline_with_hyperparameter_sweep
description: Tune hyperparameters using TF component
settings:
    default_compute: azureml:cpu-cluster
jobs:
  sweep_step:
    type: sweep
    inputs:
      data: 
        type: uri_file
        path: wasbs://[email protected]/iris.csv
      degree: 3
      gamma: "scale"
      shrinking: False
      probability: False
      tol: 0.001
      cache_size: 1024
      verbose: False
      max_iter: -1
      decision_function_shape: "ovr"
      break_ties: False
      random_state: 42
    outputs:
      model_output:
      test_data:
    sampling_algorithm: random
    trial: ./train.yml
    search_space:
      c_value:
        type: uniform
        min_value: 0.5
        max_value: 0.9
      kernel:
        type: choice
        values: ["rbf", "linear", "poly"]
      coef0:
        type: uniform
        min_value: 0.1
        max_value: 1
    objective:
      goal: minimize
      primary_metric: training_f1_score
    limits:
      max_total_trials: 5
      max_concurrent_trials: 3
      timeout: 7200

  predict_step:
    type: command
    inputs:
      model: ${{parent.jobs.sweep_step.outputs.model_output}}
      test_data: ${{parent.jobs.sweep_step.outputs.test_data}}
    outputs:
      predict_result:
    component: ./predict.yml

In the v2 SDK, you can enable hyperparameter tuning for any command component by calling the .sweep() method. The following pipeline definition shows how to enable sweep for train_model.

The example first loads the train_component_func defined in the train.yml file. To create the train_model, the code adds the c_value, kernel, and coef0 hyperparameters into the search space. The sweep_step defines the primary_metric, sampling_algorithm, and other parameters.

train_component_func = load_component(source="./train.yml")
score_component_func = load_component(source="./predict.yml")

# define a pipeline
@pipeline()
def pipeline_with_hyperparameter_sweep():
    """Tune hyperparameters using sample components."""
    train_model = train_component_func(
        data=Input(
            type="uri_file",
            path="wasbs://[email protected]/iris.csv",
        ),
        c_value=Uniform(min_value=0.5, max_value=0.9),
        kernel=Choice(["rbf", "linear", "poly"]),
        coef0=Uniform(min_value=0.1, max_value=1),
        degree=3,
        gamma="scale",
        shrinking=False,
        probability=False,
        tol=0.001,
        cache_size=1024,
        verbose=False,
        max_iter=-1,
        decision_function_shape="ovr",
        break_ties=False,
        random_state=42,
    )
    sweep_step = train_model.sweep(
        primary_metric="training_f1_score",
        goal="minimize",
        sampling_algorithm="random",
        compute="cpu-cluster",
    )
    sweep_step.set_limits(max_total_trials=20, max_concurrent_trials=10, timeout=7200)

    score_data = score_component_func(
        model=sweep_step.outputs.model_output, test_data=sweep_step.outputs.test_data
    )


pipeline_job = pipeline_with_hyperparameter_sweep()

# set pipeline level compute
pipeline_job.settings.default_compute = "cpu-cluster"

For the full sweep job schema, see CLI (v2) sweep job YAML schema.

Submit the hyperparameter tuning pipeline job

After you submit this pipeline job, Azure Machine Learning runs the trial component multiple times to sweep over hyperparameters, based on the search space and limits you defined in the sweep_step.

View hyperparameter tuning results in studio

After you submit a pipeline job, the SDK or CLI widget gives you a web URL link to the pipeline graph in the Azure Machine Learning studio UI.

To view hyperparameter tuning results, double-click the sweep step in the pipeline graph, select the Child jobs tab in the details panel, and then select the child job.

On the child job page, select the Trials tab to see and compare metrics for all the child runs. Select any of the child runs to see the details for that run.

If a child run failed, you can select the Outputs + logs tab on the child run page to see useful debug information.

Share via

How to do hyperparameter tuning in pipelines

Prerequisites

Create and run a hyperparameter tuning pipeline

Create a command component with hyperparameter inputs

Create the trial component source code

Create a pipeline with a hyperparameter sweep step

Submit the hyperparameter tuning pipeline job

View hyperparameter tuning results in studio

Feedback

Additional resources

Share via

How to do hyperparameter tuning in pipelines

Prerequisites

Create and run a hyperparameter tuning pipeline

Create a command component with hyperparameter inputs

Create the trial component source code

Create a pipeline with a hyperparameter sweep step

Submit the hyperparameter tuning pipeline job

View hyperparameter tuning results in studio

Related content

Feedback

Additional resources