stickbugml

stick-bug-ml

A framework to ease the burden of organizing code of a supervised machine learning system.

It provides decorators that manage data & pass it between common steps in building a machine learning system, such as: - loading the dataset - preprocessing - feature generation - model definition

While doing this, it keeps the global namespace free of clutter such as that from an endless chain of features and models.

In addition, it makes it easy to put new, real life, data through the exact same process that training data goes through.

Installation

Install simply via pip (Python 3):

.. code:: bash

$ pip install stickbugml

Dependencies: - Python 3 - sklearn - pandas - numpy

Example

Note: there is also a great example for use in Jupyter Notebooks <demo.ipynb>__

First, import this library:

.. code:: python

import stickbugml
from stickbugml.decorators import dataset, feature, model

Load your dataset:

.. code:: python

import seaborn.apionly as sns
import pandas as pd

@dataset(train_valid_test=(0.6, 0.2, 0.2)) # define your train/test/validation data splits
def raw_dataset():
    titanic_dataset = sns.load_dataset('titanic')

    # Drop NaN rows for simplicity
    titanic_dataset.dropna(inplace=True)

    # Extract X and y
    X = titanic_dataset.drop('survived', axis=1)
    y = titanic_dataset['survived']
    return X, y

print(raw_dataset.head()) # yes, this does work! raw_dataset is now a pandas DataFrame

(Optionally) do some pre-processing:

.. code:: python

@preprocess
def preprocessed_dataset(X):
    # Encode categorical columns
    categorical_column_names = [
            'sex', 'embarked', 'class',
            'who', 'adult_male', 'deck',
            'embark_town', 'alive', 'alone']

    X = pd.get_dummies(X,
                       columns=categorical_column_names,
                       prefix=categorical_column_names)

    return X

print(preprocessed_dataset.head()) # See the first code block for explaination

Generate some features:

.. code:: python

from sklearn import decomposition
import numpy as np

@feature('pca')
def pca_feature(X):
    pca = decomposition.PCA(n_components=3)
    pca.fit(X)
    pca_out = pca.transform(X)

    pca_out = np.transpose(pca_out, (1, 0))
    return pd.DataFrame(pca_out)

# let's preview
print(pca_feature.head()) # See the first code block for explaination

# you can add more features, btw

And define your (machine learning) model(s):

.. code:: python

import xgboost as xgb

@model('xgboost')
def xgboost_model():
    def define(num_columns):
        return None # xgboost models aren't pre-defined


    def train(model, params, train, validation):
        params['objective'] = 'binary:logistic' # Static parameters can be defined here
        params['eval_metric'] = 'logloss'

        d_train = xgb.DMatrix(train['X'], label=train['y'])
        d_valid = xgb.DMatrix(validation['X'], label=validation['y'])

        watchlist = [(d_train, 'train'), (d_valid, 'valid')]

        trained_model = xgb.train(params, d_train, 2000, watchlist, early_stopping_rounds=50, verbose_eval=10)

        return trained_model

    def predict(model, X):
        return model.predict(xgb.DMatrix(X))

    return define, train, predict

Now you can train your model, trying out different parameters if you want:

.. code:: python

stickbugml.train('xgboost', {
    'max_depth': 7,
    'eta': 0.01
})

The library keeps the test data's ground truth values locked away so your models won't train on it. After you train your model, have the framework evaluate it for you:

.. code:: python

logloss_score = stickbugml.evaluate('xgboost')
print(logloss_score)

You can add more models and features if so desired.

Since this library is built with reality in mind, you can easily get predictions for new/real-life data:

.. code:: python

raw_X = pd.read_csv('2018_titanic_manifesto.csv') # It will probably sink, but we don't know who will survive
processed_X = stickbugml.process(raw_X) # Process the data
del raw_X # Gotta keep that namespace clean, right?

y = stickbugml.predict('xgboost', processed_X) # Make predictions

print(y)

Contributing & Feedback

If you have any problems, or would like a new feature, submit an Issue.

If you want to help out, feel free to submit a Pull Request.

License

This project uses the Apache 2.0 License