stjordanis / pancake

Pancake is a Python package which provides a simple API to stack scikit-learn models.

README

PanCake is a Python package that allows users to stack scikit-learn models over a number of folds and train stacker models using out-of-sample predictions of input models.

The stacking tool provides the construction of a stacking module composed of in-layer (models being stacked) and out-layer (stacker models) models. The output is a list or matrix of predictions from training of the module, which can either be used as the final results, or fed into a different module.

Installation

After cloning the repository, install from the directory of the package by

pip install .

Usage

Initiating stacker

stacker = Stacker(X, y, splitter, evalMetric, family)

where X is the data matrix (numpy array), y is target vector (numpy array), splitter is a scikit-learn cross-validation generator (KFold or StratifiedKFold), evalMetric is the metric to be maximized during training, and family is the type of the problem (currently "regression" or "binary").

Adding models (in-layer):

Add a scikit-learn model modelObj to in-layer by

stacker.addModelIn(modelObj, trainable, hyperParameters)

If trainable is set to True then the model will be trained across folds using the hyperParameters which is a dictionary of hyper-parameter grid for the model (check scikit-learn's documentation for the model). If it is set to False then the model is assumed fixed and is only fitted across folds.

Adding stacker models (out-layer):

Add a scikit-learn model modelObj to out-layer by

stacker.addModelOut(modelObj, hyperParameters)

Again, hyperParameters is a dictionary containig the grid of hyper-parameters for the model.

Training and Predictions:

To train the model and get predictions on the training data, use

predsTrain = stacker.stackTrain(matrixOut)

which yields final predictions for each out-layer model as a list when matrixOut is set to False. When it is set to True, predictions for each out-model is appended as column vectors is a an array.

For predictions on the test set, use:

predsTest = stacker.stackTest(X_ts, matrixOut)

where X_ts is the test data and matrixOut is the same as above.

Summary, Saving and Loading:

To get a summary on CV scores, fit and training times for each in-layer and out-layer model, use

stacker.summary()

To save the trained stacker for later use, call

saveModel(stacker, savePath)

To load a trained model from disk, call

stacker = loadModel(savePath)

Examples

Jupyter notebooks analyzing the Boston Housing data is included in the repo:

1. Stacking linear models
2. Stacking Random Forest and Support Vector Regressors

TODO

1. Multi-class classification problems
2. Parallelization at the model and/or hyper-parameter level