noloox 0.1.0

is a Python library containing reference implementations of a bunch of very useful unsupervised learning algorithms that you probably won't find elsewhere.

What is:

• A collection of unsupervised machine learning algorithms
• A scikit-learn compatible library
• An educational resource containing worked examples and reference implementation

What isn't:

• The most feature-complete or efficient implementation of these algorithms
• A replacement for scikit-learn
• An all-in-one machine learning framework
• A library for complete Bayesian inference. Use a PPL like NumPyro, PyMC or Stan.

Basic usage

Install noloox from PyPI:

pip install noloox

Then you can load models from the library and use them the same way you would use scikit-learn.

from noloox.mixture import StudentsTMixture

model = StudentsTMixture(n_components=10)
cluster_labels = model.fit_predict(X)

Models

Our philosophy and goals

• Keep implementations simple and minimal, Minimal dependencies
• Everything should either be implemented in NumPy or JAX. Preferably as many in JAX as possible.
• Library structure should match sklearn standards, and all algorithms should be drop-in replacements for scikit-learn equivalents.
• Under these restrictions, algorithms should be as fast as humanly possible

The wishlist:

There are a number of algorithms that would be nice to implement in the library. Contributions are very welcome.

• ProdLDA, and amortized ProdLDA (CTMs) (without Flax)
• Parametric-TSNE, possibly also Multi-scale Parametric-TSNE
• DiRE
• Infinite NMF
• Latent Dirichlet Allocation with Gibbs Sampling
• Gaussian LDA