magnusax/AutoML

The project aims to develop a customized ML framework on top of existing libraries

Gazer

"Standing on the shoulders of giants, gazing at the stars"

The project aims to develop a customized "automl" framework, leveraging the power of existing popular libraries. As time passes the intention of the author is to add non-standard capabilities such as advanced forms of ensemble building and Markov Chain Monte Carlo (MCMC) like optimization algorithms (see further below for a brief overview of ongoing efforts). The project is still in its infancy, and fairly basic functionality has been implemented; but it is currently under active development (conditional on time and energy).

Install is now available through the python package index. Run pip install gazer==0.1.dev1 to install development version. Currently tested for python versions >=3.4 only.

Using the library

The entrance to functionality is the GazerMetaLearner object;

from gazer import GazerMetaLearner
learner = GazerMetaLearner()

Set parameter method='all' to initialize all available algorithms. Default is to initialize 3 algorithms randomly.

Post initialization, (meta) algorithms are available in the .clf variable of the initialized object. To be more specific, the variable returns a dictionary of (name, MetaClassifier) tuples, where the MetaClassifier object is a wrapper around a "sklearn-like" classifier. The current version of the library also implements xgboost and keras classifiers.

learner.clf = {name1: MetaClassifier1, name2: MetaClassifier2, ... , nameN: MetaClassifierN}

To inspect loaded algorithms, simply call learner.names.

Assuming a standard supervised learning scenario with data available in the form of X, y, you can easily fit available algorithms using default (i.e. reasonble) hyperparameter settings:

learner.fit(X, y)

This method trains all initialized algorithms. Moreover, Random, Grid search, and Bayesian optimization methods are implemented and may be directly called from the learner instance. To see parameters, simply call:

learner.clf['name'].cv_params

This returns a list containing one or more parameter dictionaries that may be edited in place.

Roadmap

Some ideas are currently being implemented, e.g.

• Markov Chain Monte Carlo scheme for optimization of (very-) large ensemble models for large data sets where training times are relatively long.
• Model perturbation and greedy adaptation
• Advanced ensembling techniques
  • See these two papers from Caruana et al.

ML libraries which will be heavily relied upon are first and foremost:

• (Done) Scikit-learn
• (Done) Xgboost
• (Done) Keras
• (Coming) Pytorch (\ge 0.4)

TODO: add direct links to each github project mentioned on this page, to ensure that credit and recognition goes to developers

Comments

As this project is in its initial development phase, the code has not been tested properly yet. It thus follows that it has not yet been uploaded to the PyPi repository yet. Project developer's contact info is available in the setup.py file.