takuya-andou / stacking

for generalizing stacking

practical_stacking

practical_stacking is python package made in consideration of the practical use, not contests such as Kaggle. All trained model and out-of-fold predictions is automatically saved for next time. And if you try to train again a saved model, it is loaded automatically. In addition, you can only predict newly geted test data without training again a saved model. Of course you can choose scikit-learn, XGboost, Keras and other for stacking.

practical_stackingはKaggleなどのコンテストではなく実用することを考えて作られたpythonのパッケージです。全ての学習済モデルとOOF予測は、自動的に保存されます。そしてもし、学習済のモデルを再び学習しようとした場合、そのモデルは自動的にロードされます。加えて、保存してあるモデルを再学習することなく、新しく取得したテストデータの予測だけをすることもできます。当然、スタッキングのモデルにはscikit-learnやxgboostやkeras等を自由に選択することができます。

Description

Stacking (sometimes called stacked generalization) learns several models and combines their predictions. This blog is very helpful to understand stacking and ensemble learning.

スタッキングはいくつかのモデルを学習して、それらの予測を組み合わせます。このブログがスタッキングやアンサンブル学習の参考にとてもなります。

Requirement

• python 3.5+(don't know about version < 3.5, sorry)
• os
• pickle
• functools.reduce
• scipy.stats
• numpy
• pandas
• sklearn

Installation

You can install stacking from PyPI:

pip install practical_stacking

Usage

See working example:

• Regression
• Classification (with class labels)
• Classification (with class probabilities)

See basic usage:

# Import practical_stacking
from pracstack import StackModel, StackMaster

# Get your data

# Initialize 1st level models
models = [
    StackModel(model_name='Li',model=LinearRegression),
    StackModel(model_name='Ridge',model=Ridge)]

# Initialize StackMaster
master = StackMaster(models=models)

# Fit 1st level models
master.fit(X_train, y_train)

# Predict 1st level models
master.predict(X_test)

# Get your stacked features
S_train = master.S_train
S_test = master.S_test

# Use 2nd level models with stacked features