lazytune 0.1.3

Fast hyperparameter optimization using screening and pruning

LazyTune 🚀

Fast, smart, and lazy hyperparameter optimization
for scikit-learn models — up to 5–10× faster than GridSearchCV
with almost the same final performance.

Live Demo → https://lazytune.vercel.app
PyPI → https://pypi.org/project/lazytune/

🔥 Why LazyTune?

• 100% compatible with any scikit-learn-style estimator
• Works for classification & regression
• Supports all scikit-learn metrics + custom scorers
• Smart pipeline: screening → ranking → pruning → full training
• Early pruning of poor configurations (prune_ratio)
• Parallel execution (n_jobs)
• Clean trial summaries & rankings in pandas DataFrame
• Returns best model, params, score + detailed report

Installation

pip install lazytune

Quick Start

from sklearn.datasets import load_breast_cancer
from sklearn.ensemble import RandomForestClassifier
from lazytune import SmartSearch

X, y = load_breast_cancer(return_X_y=True)

param_grid = {
    "n_estimators": [50, 100, 150, 200],
    "max_depth": [5, 10, 15, None],
    "min_samples_split": [2, 3, 4, 5]
}

search = SmartSearch(
    estimator=RandomForestClassifier(random_state=42),
    param_grid=param_grid,
    metric="accuracy",
    cv_folds=3,
    prune_ratio=0.5,       # keep top 50% after screening
    n_jobs=-1              # use all cores
)

search.fit(X, y)

print("Best parameters:", search.best_params_)
print("Best CV score:   ", search.best_score_)
print("\nBest model:\n", search.best_estimator_)

More Examples

SVM Classification

from sklearn.svm import SVC
from lazytune import SmartSearch

search = SmartSearch(
    estimator=SVC(random_state=42),
    param_grid={
        "C": [0.1, 1, 10, 50, 100],
        "kernel": ["linear", "rbf"],
        "gamma": ["scale", "auto", 0.001, 0.0001]
    },
    metric="f1_macro",
    cv_folds=5,
    prune_ratio=0.6
)

Regression (Random Forest)

from sklearn.ensemble import RandomForestRegressor
from lazytune import SmartSearch

search = SmartSearch(
    estimator=RandomForestRegressor(random_state=42),
    param_grid={
        "n_estimators": [100, 200, 300, 500],
        "max_depth": [8, 12, 16, None],
        "min_samples_split": [2, 4, 8]
    },
    metric="r2",
    cv_folds=4,
    n_jobs=-1
)

Supported Metrics (examples)

Classification

accuracy • f1 • f1_macro • f1_weighted • precision • recall • roc_auc • balanced_accuracy • ...

Regression

r2 • neg_mean_squared_error • neg_root_mean_squared_error • neg_mean_absolute_error • ...

Custom metrics → use sklearn.metrics.make_scorer

How LazyTune Works

1. Generate all (or sampled) hyperparameter combinations
2. Quick screening round with cross-validation
3. Rank configurations by performance
4. Prune bottom performers (prune_ratio)
5. Train remaining promising candidates thoroughly
6. Return best model + full trial summary

→ Much faster than GridSearchCV / RandomizedSearchCV
→ Usually very close (or identical) final performance

Main API – SmartSearch

Key Attributes

Attribute	Description
best_params_	Best hyperparameter dictionary
best_score_	Best cross-validated score
best_estimator_	Fully fitted estimator with best parameters
summary_	pandas DataFrame with trial results & rankings
cv_results_	Detailed CV results per candidate

Main Methods

• .fit(X, y)
• .predict(X)
• .score(X, y)
• .get_params() / .set_params()

Requirements

• Python ≥ 3.8
• numpy
• pandas
• scikit-learn

Made with ❤️ by Anik Chand
MIT License

Feedback, issues, stars, and contributions are very welcome! 🌟

Try Live Demo • Install from PyPI

Happy tuning! 🚀