Azure Machine Learning SDK (v1)

Warnings

• breaking Azure Machine Learning SDK v1 is officially deprecated as of March 31, 2025, with end of support on June 30, 2026. After this date, existing workflows may still run but will not receive technical support or updates, potentially exposing them to security risks or breaking changes.
  Fix: Migrate your workflows to the Azure Machine Learning Python SDK v2 (`azure-ai-ml`). This involves significant API changes; refer to the official migration guides.
• gotcha SDK v1 (`azureml-sdk`) and SDK v2 (`azure-ai-ml`) are incompatible and should generally not be installed in the same Python environment to avoid package clashes and confusion.
  Fix: Use separate Python environments for SDK v1 and SDK v2 projects. If mixed interaction with a single workspace is needed, ensure distinct environments for each SDK version.
• gotcha Authentication to an Azure ML Workspace in v1 often relies on a `config.json` file (containing subscription ID, resource group, and workspace name) placed in a `.azureml` subdirectory or explicitly passed parameters. Without proper configuration, connection attempts will fail. Interactive authentication is also common for initial setup.
  Fix: Ensure a valid `config.json` file is present in the default search path (`.azureml/`) or specify the workspace details directly. For interactive authentication, follow the browser prompts. For automated scripts, consider service principal authentication.
• deprecated `Estimator` classes, a common way to define training jobs in earlier v1 versions, are effectively superseded by `ScriptRunConfig` and later `Command` (in v2). While still functional in v1, `ScriptRunConfig` offers more flexibility.
  Fix: Prefer `ScriptRunConfig` for submitting training jobs in SDK v1. When migrating to SDK v2, `Command` jobs are the direct equivalent.

Install

• pip install azureml-sdk Install Azure ML SDK v1

Imports

• Workspace

  from azureml.core import Workspace

  Used to connect to your Azure ML workspace.
• Experiment

  from azureml.core import Experiment

  Used to create and manage ML experiments.
• Environment

  from azureml.core import Environment

  Used to define the reproducible Python environment for runs and deployments.
• ScriptRunConfig

  from azureml.core import ScriptRunConfig

  Encapsulates the script, compute target, and environment for a training run.
• ComputeTarget
  wrong:

  from azureml.core import ComputeTarget

  correct:

  from azureml.core.compute import ComputeTarget

  ComputeTarget is typically imported from the `azureml.core.compute` submodule, not directly from `azureml.core`.

Quickstart

This quickstart demonstrates how to connect to an Azure ML Workspace, define a custom environment, create a compute target (or use 'local'), and submit a simple Python script as an experiment using `ScriptRunConfig` in Azure ML SDK v1. A `config.json` file in a `.azureml` subdirectory is the recommended way to connect to a workspace.

import os
from azureml.core import Workspace, Experiment, Environment, ScriptRunConfig
from azureml.core.compute import ComputeTarget, AmlCompute
from azureml.core.compute_target import ComputeTargetException
from azureml.core.conda_dependencies import CondaDependencies

# Create a dummy script file
with open('train_script.py', 'w') as f:
    f.write("""
import argparse
import os
import time
print("Hello from Azure ML v1 training script!")
parser = argparse.ArgumentParser()
parser.add_argument('--arg1', type=str, default='default_value')
args = parser.parse_args()
print(f"Argument 1: {args.arg1}")
time.sleep(5) # Simulate work
print("Script finished.")
""")

# Create a dummy conda environment file
with open('conda_env.yml', 'w') as f:
    f.write("""
name: my_env
dependencies:
  - python=3.8
  - pip:
    - azureml-defaults
""")

# NOTE: For a real scenario, replace placeholder values and ensure a config.json is available
# Authenticate and connect to your workspace
try:
    ws = Workspace.from_config(path='./.azureml', _file_name='config.json') # Reads from a local config.json
    print(f"Connected to workspace {ws.name}")
except Exception as e:
    print(f"Could not load workspace from config. Ensure .azureml/config.json exists or provide details manually: {e}")
    # Fallback to manual connection (replace with your actual details)
    subscription_id = os.environ.get('AZURE_SUBSCRIPTION_ID', 'YOUR_SUBSCRIPTION_ID')
    resource_group = os.environ.get('AZURE_RESOURCE_GROUP', 'YOUR_RESOURCE_GROUP')
    workspace_name = os.environ.get('AZURE_WORKSPACE_NAME', 'YOUR_WORKSPACE_NAME')
    ws = Workspace(subscription_id, resource_group, workspace_name)
    print(f"Connected to workspace {ws.name} via manual details.")

experiment_name = "my-first-v1-experiment"
experiment = Experiment(workspace=ws, name=experiment_name)

# Choose a name for your CPU cluster (or use 'local')
compute_name = "cpu-cluster"
compute_target = None

try:
    compute_target = ComputeTarget(workspace=ws, name=compute_name)
    print(f"Found existing compute target: {compute_name}")
except ComputeTargetException:
    print(f"Creating a new compute target: {compute_name}")
    compute_config = AmlCompute.provisioning_configuration(vm_size='STANDARD_DS1_V2', max_nodes=1)
    compute_target = ComputeTarget.create(ws, compute_name, compute_config)
    compute_target.wait_for_completion(show_output=True)

# Define the environment
env = Environment.from_conda_specification(name='my-custom-env', file_path='conda_env.yml')

# Create a ScriptRunConfig
src = ScriptRunConfig(
    source_directory='.',
    script='train_script.py',
    compute_target=compute_target,
    environment=env,
    arguments=['--arg1', 'hello_from_run']
)

# Submit the run
run = experiment.submit(src)
print(f"Submitted run: {run.get_portal_url()}")
run.wait_for_completion(show_output=True)
print(f"Run completed with status: {run.status}")