Contents:

1. Github repo
2. What is pymc-learn?
3. Quick Install
4. Quick Start
5. Index

pymc-learn is a library for practical probabilistic machine learning in Python.

It provides probabilistic models in a syntax that mimics scikit-learn. Users can now have calibrated quantities of uncertainty in their models using powerful inference algorithms -- such as MCMC or Variational inference --provided by PyMC3. See why for a more detailed description of why pymc-learn was created.

pymc-learn mimics scikit-learn. You don't have to completely rewrite your scikit-learn ML code.

from sklearn.linear_model \                         from pmlearn.linear_model \
  import LinearRegression                             import LinearRegression
lr = LinearRegression()                             lr = LinearRegression()
lr.fit(X, y)                                        lr.fit(X, y)

The difference between the two models is that pymc-learn estimates model parameters using Bayesian inference algorithms such as MCMC or variational inference. This produces calibrated quantities of uncertainty for model parameters and predictions.

You can install pymc-learn from source as follows:

pip install git+https://github.com/pymc-learn/pymc-learn

pymc-learn is tested on Python 2.7, 3.5 & 3.6 and depends on Theano, PyMC3, NumPy, SciPy, and Matplotlib (see requirements.txt for version information).

# For regression using Bayesian Nonparametrics
>>> from sklearn.datasets import make_friedman2
>>> from pmlearn.gaussian_process import GaussianProcessRegressor
>>> from pmlearn.gaussian_process.kernels import DotProduct, WhiteKernel
>>> X, y = make_friedman2(n_samples=500, noise=0, random_state=0)
>>> kernel = DotProduct() + WhiteKernel()
>>> gpr = GaussianProcessRegressor(kernel=kernel).fit(X, y)
>>> gpr.score(X, y) # doctest: +ELLIPSIS
0.3680...
>>> gpr.predict(X[:2,:], return_std=True) # doctest: +ELLIPSIS
(array([653.0..., 592.1...]), array([316.6..., 316.6...]))

Recent research has led to the development of variational inference algorithms that are fast and almost as flexible as MCMC. For instance Automatic Differentation Variational Inference (ADVI) is illustrated in the code below.

from pmlearn.neural_network import MLPClassifier
model = MLPClassifier()
model.fit(X_train, y_train, inference_type="advi")

Instead of drawing samples from the posterior, these algorithms fit a distribution (e.g. normal) to the posterior turning a sampling problem into an optimization problem. ADVI is provided PyMC3.

To cite pymc-learn in publications, please use the following:

Pymc-learn Developers Team (2019). pymc-learn: Practical probabilistic machine
learning in Python. arXiv preprint arXiv:xxxx.xxxxx. Forthcoming.

Or using BibTex as follows:

@article{Pymc-learn,
  title={pymc-learn: Practical probabilistic machine learning in {P}ython},
  author={Pymc-learn Developers Team},
  journal={arXiv preprint arXiv:xxxx.xxxxx},
  year={2019}
}

If you want to cite pymc-learn for its API, you may also want to consider this reference:

Carlson, Nicole (2018). Custom PyMC3 models built on top of the scikit-learn
API. https://github.com/parsing-science/pymc3_models

Or using BibTex as follows:

@article{Pymc3_models,
  title={pymc3_models: Custom PyMC3 models built on top of the scikit-learn API,
  author={Carlson, Nicole},
  journal={},
  url={https://github.com/parsing-science/pymc3_models}
  year={2018}
}

New BSD-3 license

Getting Started

• install
• support
• why

User Guide

The main documentation. This contains an in-depth description of all models and how to apply them. pymc-learn leverages the Base template provided by the PyMC3 Models project: https://github.com/parsing-science/pymc3_models.

• user_guide

Examples

Pymc-learn provides probabilistic models for machine learning, in a familiar scikit-learn syntax.

• regression
• classification
• mixture
• neural_networks
• api

API Reference

pymc-learn leverages the Base template provided by the PyMC3 Models project: https://github.com/parsing-science/pymc3_models.

• api

Help & reference

• develop
• support
• changelog
• cite