AutoMLパイプラインの構築

See Extended Examples for complete configuration files and templates.

Automate hyperparameter tuning and model selection using Optuna or Ray Tune with efficient search strategies.

使用タイミング

• Starting new ML project and need to quickly find good model configurations
• Retraining existing model with new data and want to re-optimize hyperparameters
• Comparing multiple algorithms and their optimal configurations
• Limited time for manual tuning but need near-optimal performance
• Team lacks deep expertise in specific algorithm hyperparameters
• Need reproducible and documented optimization process

入力

• 必須: Training dataset with features and labels
• 必須: Validation dataset for objective evaluation
• 必須: Model type(s) to optimize (e.g., XGBoost, LightGBM, neural network)
• 必須: Optimization objective (metric to maximize/minimize)
• 必須: Compute budget (time or number of trials)
• 任意: Search space constraints (min/max values for hyperparameters)
• 任意: Prior knowledge of good hyperparameter ranges

手順

ステップ1: Install Dependencies and Set Up Environment

Install Optuna or Ray Tune with appropriate backends.

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Option 1: Optuna (simpler, good for single-machine)
pip install optuna optuna-dashboard
pip install scikit-learn xgboost lightgbm

# Option 2: Ray Tune (distributed, good for multi-machine/GPU)
pip install "ray[tune]" optuna hyperopt bayesian-optimization
pip install torch torchvision  # if optimizing neural networks

# Visualization and tracking
pip install mlflow tensorboard plotly

Create project structure:

mkdir -p automl/{configs,experiments,models,results}

期待結果: Clean environment with required packages installed, no dependency conflicts.

失敗時: Use Python 3.8-3.11 (compatibility issues with 3.12+), if CUDA errors occur install CPU-only versions first, on M1/M2 Mac use conda instead of pip for scikit-learn.

ステップ2: Define Search Space and Objective (Optuna)

Create configuration for hyperparameter search.

# automl/optuna_config.py
import optuna
from optuna.pruners import HyperbandPruner
from optuna.samplers import TPESampler
import xgboost as xgb
from sklearn.metrics import roc_auc_score, mean_squared_error
import numpy as np

# ... (see EXAMPLES.md for complete implementation)

期待結果: Search space covers reasonable hyperparameter ranges, objective function runs without errors, pruning stops unpromising trials early.

失敗時: If trials crash, reduce search space (e.g., lower max n_estimators), verify data has no NaN/inf values, check memory usage (reduce batch size if OOM), ensure eval_metric matches task type.

ステップ3: Run Optimization with Advanced Samplers

Execute hyperparameter search with efficient sampling strategies.

# automl/run_optimization.py
import optuna
from optuna.samplers import TPESampler, CmaEsSampler, NSGAIISampler
from optuna.pruners import HyperbandPruner, MedianPruner, SuccessiveHalvingPruner
import joblib
import pandas as pd
from pathlib import Path

# ... (see EXAMPLES.md for complete implementation)

期待結果: Optimization completes with 50-70% of trials pruned early, best parameters found, visualization plots generated showing convergence.

失敗時: If no pruning happens, verify objective reports intermediate values correctly, if optimization doesn't improve try different sampler (TPE → CmaES), if crashes with n_jobs>1 use n_jobs=1 for debugging.

ステップ4: Set Up Ray Tune for Distributed Optimization (Alternative)

Use Ray Tune for multi-GPU or multi-node optimization.

# automl/ray_tune_config.py
from ray import tune
from ray.tune.schedulers import ASHAScheduler, PopulationBasedTraining
from ray.tune.search.optuna import OptunaSearch
from ray.tune.search import ConcurrencyLimiter
import xgboost as xgb
from sklearn.metrics import roc_auc_score
import os
# ... (see EXAMPLES.md for complete implementation)

期待結果: Ray Tune runs trials in parallel across CPUs/GPUs, ASHA scheduler stops bad trials early, best configuration found and logged.

失敗時: If Ray crashes, start with ray.init(num_cpus=2, num_gpus=0) for debugging, reduce concurrent trials if OOM, check that train function doesn't modify shared data, use tune.report() not return for metrics.

ステップ5: Track Experiments with MLflow

Integrate with MLflow for experiment tracking and model registry.

# automl/mlflow_tracking.py
import mlflow
import mlflow.xgboost
from mlflow.tracking import MlflowClient
import optuna
from pathlib import Path


# ... (see EXAMPLES.md for complete implementation)

期待結果: All trials logged to MLflow with parameters and metrics, best model registered in MLflow registry, experiments viewable in MLflow UI.

失敗時: Start MLflow UI with mlflow ui --backend-store-uri file:./automl/mlruns, check write permissions to mlruns directory, if registration fails verify model registry is configured, ensure model artifact size < 2GB.

ステップ6: Deploy Best Model and Monitor Performance

Save optimized model and set up monitoring.

# automl/deploy_model.py
import joblib
import json
from pathlib import Path
import optuna
import xgboost as xgb


# ... (see EXAMPLES.md for complete implementation)

期待結果: Model saved in production-ready format, configuration documented, inference script created for deployment.

失敗時: If model file too large (>100MB), consider model compression or feature selection, verify model loads correctly in fresh Python session, test inference script with sample data before deployment.

バリデーション

• Optuna/Ray Tune installs without dependency conflicts
• Search space includes reasonable hyperparameter ranges
• Objective function runs successfully for single trial
• Optimization completes 50+ trials within time budget
• Pruning stops 40-70% of unpromising trials early
• Best parameters improve over default configuration by >5%
• Visualizations show convergence (optimization history flattens)
• MLflow logs all trials with parameters and metrics
• Final model saved and loads correctly
• Deployment package includes all necessary files

よくある落とし穴

• Overfitting to validation set: Running 1000s of trials implicitly optimizes for validation set; use holdout test set or time-based split for final evaluation
• Ignoring feature engineering: AutoML finds best hyperparameters but doesn't create features; invest in feature engineering first
• Search space too wide: Unbounded or very wide ranges waste trials on unrealistic values; use domain knowledge to constrain
• Not using early stopping: Training full epochs for every trial is wasteful; enable early stopping in objective function
• Ignoring compute costs: 100 trials × 10 minutes = 16 hours; consider compute budget when setting n_trials
• Categorical features not encoded: Most algorithms need numeric features; encode categoricals before optimization
• Imbalanced data: Default metrics may mislead with class imbalance; use F1, AUC, or custom metrics
• Not saving intermediate results: Crashes lose all progress; use persistent storage (Optuna SQLite, MLflow) to resume

関連スキル

• track-ml-experiments - MLflow experiment tracking and versioning
• orchestrate-ml-pipeline - Airflow/Kubeflow for production AutoML pipelines