求基于模型名与计算资源获取Azure AutoML模型解释的Python代码

First off, make sure you've got the required Azure ML SDK packages installed (if you haven't already):

pip install azureml-sdk azureml-automl-core azureml-explain-model

Below is a reusable Python function that takes your target model name and compute resource as inputs, tracks down the corresponding AutoML run, and pulls both global and summary explanations for the model:

from azureml.core import Workspace, Experiment
from azureml.explain.model._internal.explanation_client import ExplanationClient

def get_automl_model_explanations(ws_name: str, resource_group: str, subscription_id: str,
                                  experiment_name: str, model_name: str, compute_target_name: str):
    # Connect to your Azure ML workspace
    ws = Workspace(subscription_id=subscription_id,
                   resource_group=resource_group,
                   workspace_name=ws_name)
    
    # Grab the AutoML experiment you ran
    experiment = Experiment(ws, experiment_name)
    
    # Filter for completed AutoML runs (skip failed/canceled ones)
    completed_runs = list(experiment.get_runs(type='automl', status='Completed'))
    
    # Find the run that matches your model and compute target
    target_run = None
    for run in completed_runs:
        run_details = run.get_details()
        # First check if the run used the specified compute target
        if run_details['target'] == compute_target_name:
            # Pull all models generated by this run
            run_models = run.get_outputs()
            for model_output in run_models:
                # Match the model name
                if model_output[0].name == model_name:
                    target_run = run
                    break
        if target_run:
            break
    
    if not target_run:
        raise ValueError(f"Couldn't find a completed AutoML run for model '{model_name}' on compute '{compute_target_name}'")
    
    # Initialize the explanation client to fetch interpretability data
    explanation_client = ExplanationClient.from_run(target_run)
    
    # Fetch the full global explanation
    global_explanation = explanation_client.get_global_explanation()
    
    # Get the summary explanation (includes aggregated feature importance)
    summary_explanation = global_explanation.get_summary_explanation()
    
    # Return a structured result with all key info
    return {
        "run_id": target_run.id,
        "model_name": model_name,
        "global_explanation": global_explanation,
        "summary_explanation": summary_explanation,
        "feature_importance": summary_explanation.feature_importance
    }

# Example usage - fill in your own workspace/experiment details
if __name__ == "__main__":
    # Your Azure ML workspace credentials
    subscription_id = "YOUR_SUBSCRIPTION_ID"
    resource_group = "YOUR_RESOURCE_GROUP"
    ws_name = "YOUR_WORKSPACE_NAME"
    experiment_name = "YOUR_AUTOML_EXPERIMENT_NAME"
    
    # Target model and compute details
    target_model_name = "YOUR_TRAINED_MODEL_NAME"
    target_compute_name = "YOUR_COMPUTE_TARGET_NAME"
    
    try:
        explanations = get_automl_model_explanations(ws_name, resource_group, subscription_id,
                                                     experiment_name, target_model_name, target_compute_name)
        
        # Quick check: print feature importance scores
        print(f"Feature Importance for Model '{target_model_name}':")
        for feature, score in explanations['feature_importance'].items():
            print(f"- {feature}: {round(score, 4)}")
            
        # Optional: launch an interactive visualization dashboard
        # explanations['global_explanation'].visualize()
        
    except ValueError as e:
        print(e)

Quick Tips: #

• Simpler Authentication: If you have a config.json file from your Azure ML workspace in your working directory, replace the workspace connection code with ws = Workspace.from_config() to skip manual credential entry.
• Handle Duplicate Runs: If you have multiple runs with the same model name/compute pair, add a filter for run timestamp (e.g., run.created_time) to pick the most recent one.
• Visualization: Uncomment the visualize() line to open an interactive browser dashboard that lets you explore feature impacts, prediction breakdowns, and more.

内容的提问来源于stack exchange，提问作者datacore