mcnets 2.0.7

Lightweight Machine Learning package with models that train using simple Monte Carlo-like methods.

Monte-Carlo-Neural-Nets

Overview

A very lightweight machine learning package.

Has models that use a unique and simple Monte Carlo approach to training. This method used is very generalizable and can therefore be extended to a variety of models both known and new.

The primary model, the NeuralNetwork class, is on par with other similar models such as the MLPRegressor/MLPClassifier featured in SciKit-Learn, but has more customizability.

The current list of models avaliable and some of their features includes:

• NeuralNetwork
  • Complete hidden layer size and activations customization
  • Supports externally defined activation functions
  • Allows customizing the input and output activations
  • Easy-access hyperparam ranges for Optuna (via .get_param_ranges_for_optuna)
• SoupRegressor
  • A unique combination of many various functions with trainable coefficients
  • Typically on-par with the NeuralNetwork; slightly more interpretable
  • Many hyperparams to adjust, with more to come

Some QoL functions and features included are:

• TTSplit: Included train-test splitter
• cross_val: Simple cross validation system
• Built-in scorer functions with support for external functions
• Ability to save and load models at any point (.save & load_model)
• Ability to copy a model via .copy

GitHub and QuickStart

More explanations, examples, and technicals can be found on the GitHub page: https://github.com/SciCapt/Monte-Carlo-Neural-Nets