Azure ML Train Core

Warnings

• deprecated The `Estimator` class, while still functional, is largely deprecated for new training scenarios in favor of `ScriptRunConfig` combined with `Environment` objects.
  Fix: Migrate training job definitions from `Estimator` to `ScriptRunConfig` and explicitly define environments using `Environment` or curated environments.
• gotcha `azureml-train-core` is primarily an internal component of the Azure ML SDK. End-users typically interact with its capabilities through higher-level abstractions imported from `azureml.core` (e.g., `Workspace`, `ScriptRunConfig`, `Environment`), rather than direct imports from `azureml.train.core`.
  Fix: Always refer to the official Azure ML SDK documentation for correct import paths and recommended API usage, focusing on modules under `azureml.core`.
• gotcha Environment management (Conda, Docker) is a common source of errors. Mismatches between local and remote environments, or incorrect `CondaDependencies` specifications, can lead to failed runs or unexpected behavior.
  Fix: Explicitly define `Environment` objects with specific Conda or Docker configurations. Use `conda_dependencies.add_pip_package()` and `conda_dependencies.add_conda_package()` to ensure all required libraries are specified. Test environments locally if possible before submitting to a remote compute target.
• gotcha The Azure ML SDK components, including `azureml-train-core`, often have strict Python version requirements. Using an unsupported Python version can lead to installation failures or runtime errors.
  Fix: Always check the `requires_python` specification on PyPI (`>=3.8, <4` for 1.62.0) and use a compatible Python version for your development and deployment environments.

Install

• pip install azureml-train-core Install core training components

Imports

• Workspace
  wrong:

  from azureml_train_core.core import Workspace

  correct:

  from azureml.core import Workspace

  `azureml-train-core` is an internal package; user-facing classes are typically imported from `azureml.core`.
• ScriptRunConfig
  wrong:

  from azureml.train.runconfig import ScriptRunConfig

  correct:

  from azureml.core import ScriptRunConfig

  The modern and recommended way to configure training runs, superseding many `Estimator` uses.
• Environment

  from azureml.core import Environment

  Used to define the software environment for training runs.
• Estimator

  from azureml.core.estimator import Estimator

  Legacy class for configuring training; largely superseded by `ScriptRunConfig` and `Environment`.

Quickstart

This quickstart demonstrates how to initialize an Azure ML Workspace and define an `Environment` using `azureml.core`, which relies on `azureml-train-core`'s underlying capabilities. It highlights the use of `ScriptRunConfig` for submitting training jobs, which has largely replaced the legacy `Estimator` for many scenarios. Authentication details are expected via environment variables or direct replacement.

import os
from azureml.core import Workspace, ScriptRunConfig, Environment
from azureml.core.conda_dependencies import CondaDependencies

# NOTE: Replace with your actual workspace details or ensure environment variables are set
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")

try:
    ws = Workspace.get(name=workspace_name, subscription_id=subscription_id, resource_group=resource_group)
    print(f"Found workspace {ws.name} at {ws.get_details()['location']}")
except Exception:
    print("Could not connect to workspace. Ensure AZURE_SUBSCRIPTION_ID, AZURE_RESOURCE_GROUP, AZURE_WORKSPACE_NAME env vars are set or replace placeholders.")
    # For a real run, you'd create a workspace if not found
    # ws = Workspace.create(name=workspace_name, subscription_id=subscription_id, resource_group=resource_group, location='eastus')

# Example: Define an environment
env = Environment('my-training-env')
c_deps = CondaDependencies()
c_deps.add_conda_package('scikit-learn')
c_deps.add_pip_package('azureml-sdk')
env.python.conda_dependencies = c_deps

# Create a dummy training script (e.g., train.py)
# with open('train.py', 'w') as f:
#     f.write("""
# import os
# print('Hello from Azure ML training!')
# print(f'Running on compute: {os.environ.get("AML_RUN_ID", "unknown")}')
#     """)

# Create a ScriptRunConfig (modern way to submit training)
# src = ScriptRunConfig(source_directory='./',
#                       script='train.py',
#                       environment=env,
#                       compute_target='cpu-cluster') # Replace with actual compute target

# Submit the run (uncomment for actual execution)
# if 'ws' in locals():
#     run = ws.submit(src)
#     run.wait_for_completion(show_output=True)
#     print(f"Run finished with status: {run.get_status()}")
else:
    print("Workspace not initialized, skipping run submission example.")