AgentSkillsCN

ml-training

在训练预测模型、提取指标、配置算法，或部署模型时使用。

SKILL.md
--- frontmatter
name: ml-training
description: "Use when training prediction models, extracting metrics, configuring algorithms, or deploying models"

ML Model Training Patterns

Reference patterns for training and deploying ML models via the Dataiku Python API.

Quick Start: Create and Train a Prediction Model

python
# Create ML task
ml_task = project.create_prediction_ml_task(
    input_dataset="MY_DATASET",
    target_variable="target_column",
    prediction_type="BINARY_CLASSIFICATION",  # or REGRESSION, MULTICLASS
    ml_backend_type="PY_MEMORY",
    guess_policy="DEFAULT",
    wait_guess_complete=True
)

# Train models (waits for completion)
trained_ids = ml_task.train(session_name="My Training Session")

# Get metrics for each trained model
for model_id in trained_ids:
    details = ml_task.get_trained_model_details(model_id)
    raw = details.get_raw()
    algo = raw['modeling']['algorithm']
    metrics = raw['perf']['tiMetrics']
    print(f"Model: {algo}")
    print(f"  AUC: {metrics.get('auc')}")
    print(f"  Log Loss: {metrics.get('logLoss')}")

CRITICAL: Metrics are in raw['perf']['tiMetrics'] NOT in trainInfo.testPerf or globalMetrics.

Deploy Model to Flow

Important: Only deploy models when explicitly requested by the user. After training, show results and let the user decide. When deploying, confirm whether to create new or update existing.

First-time deployment (creates new saved model)

python
result = ml_task.deploy_to_flow(
    model_id=best_model_id,
    model_name="my_model",
    train_dataset="MY_DATASET"
)
# Returns: {"savedModelId": "...", "trainRecipeName": "..."}

Update existing saved model (new version)

python
saved_model = project.get_saved_model("my_model")
sm_id = saved_model.get_id()
ml_task.redeploy_to_flow(model_id=new_model_id, saved_model_id=sm_id)

Access Existing ML Task

python
analyses = project.list_analyses()
analysis = project.get_analysis(analyses[0]['analysisId'])
ml_tasks_info = analysis.list_ml_tasks()
ml_task = analysis.get_ml_task(ml_tasks_info['mlTasks'][0]['mlTaskId'])
model_ids = ml_task.get_trained_models_ids()

References