thetaflow 0.0.20

Bridging the gap between Statistical Inference and Neural Networks

Thetaflow

Bridging the gap between Statistical Inference and Deep Learning.

Thetaflow is a Python package built on top of TensorFlow/Keras designed to fully integrate statistical modeling with neural network components. It allows researchers and data scientists to define any statistical model where parameters can be:

Dynamic: Modeled as outputs of a complex neural network. Static: Treated as independent, learnable weights (standard statistical coefficients).

It generalizes Maximum Likelihood Estimation (MLE) for a massive class of models, acting as a flexible optimizer that brings the power of backpropagation to rigorous statistical inference.

Key Features

• Flexible Parameter Definition: seamless mixing of deep learning outputs and scalar statistical parameters.
• Custom Likelihoods: Define any probability density function (PDF) or mass function (PMF) as your objective.
• TensorFlow/Keras Backend: leverages hardware acceleration (GPU/TPU) and automatic differentiation for complex optimization landscapes.
• General Optimizer: Solves for the Maximum Likelihood Estimate (MLE) across arbitrary model architectures.

Installation

The easiest and recommended way to install thetaflow is directly from PyPI using pip:

pip install thetaflow

Examples

An example application applying the standard simple linear regression model can be seen in the examples directory. I plan on adding further documentation! :)