Welcome to TF-MELT’s documentation!

TF-MELT (TensorFlow Machine Learning Toolbox) is a collection of architectures, processing, and utilities that are transferable over a range of ML applications.

TF-MELT is a toolbox for researchers to use for machine learning applications in the TensorFlow language. The goal of this software is to enable fast start-up of machine learning tasks and to provide a reliable and flexible framework for development and deployment. The toolbox contains generalized methods for every aspect of the machine learning workflow while simultaneously providing routines that can be tailored to specific application spaces.

The toolbox is structured with the following modules further described in the TF-MELT Package section:

  • TF-MELT Models Module - Contains a collection of pre-built models that can be used for a variety of machine learning tasks.

    The models currently available are:

    • Artificial Neural Network (ANN) - A simple feedforward neural network with customizable layers and activation functions.

    • Residual Neural Network (ResNet) - A neural network architecture with customizable residual blocks.

    • Bayesian Neural Network (BNN) - A neural network architecture with Bayesian layers for probabilistic modeling.

  • TF-MELT Blocks Module - Contains a collection of pre-built blocks that can be used to build custom models. These blocks are designed to be easily imported and used in custom models. Refer to the Models module for examples of how to use these effectively.

    The blocks currently available are:

    • DenseBlock - A dense block for fully-connected models.

    • ResidualBlock - A residual block with skip connections.

    • BayesianBlock - A Bayesian block using flipout layers for probabilistic modeling.

    • DefualtOutput - A single dense layer for output.

    • SingleMixtureOutput - A single dense layer with mean and variance output.

    • MultipleMixturesOutput - A dense layer with mixture model output for multiple means and variances with learnable mixture coefficients.

    • BayesianAleatoricOutput - A Bayesian output layer with aleatoric uncertainty.

  • TF-MELT Losses Module - Contains a collection of pre-built loss functions that can be used for a variety of machine learning tasks. These loss functions are designed to be easily imported and used in custom models. Refer to the Models module for examples of how to use these effectively.

    The loss functions currently available are:

    • SingleMixtureLoss - A negative log likelihood loss function for single mixture models.

    • MultipleMixturesLoss - A negative log likelihood loss function for multiple mixture models.

The toolbox also includes a Utilities Subpackage, which contains a collection of functions useful for data preprocessing, model evaluation, visualization, and other tasks.

The utility modules currently available are:

  • Evaluation Module - Contains a collection of functions for evaluating machine learning models. Useful for evaluating TF-MELT model performance and extracting uncertainty quantification metrics.

  • Preprocessing Module - Contains a collection of functions for preprocessing data for machine learning tasks. Leverages Scikit-learn preprocessing functions and implements additional helper functions.

  • Statistics Module - Contains a collection of functions for calculating statistics and metrics for machine learning tasks. Designed to be utilized by the other utility functions.

  • Visualization Module - Contains a collection of functions for visualizing data and model performance. Designed to easily generate plots of model performance, but can also be customized for user preferences.

Also included in the TF-MELT repo is an Examples directory, which contains a set of jupyter notebooks that demonstrate how to use the different modules in the toolbox for the full machine learning workflow (e.g., data preprocessing, model creation, training, evaluation, and visualization).

There is also a Tests Module that contains a set of unit tests for the different modules in the toolbox.

Finally, these docs are contained in the Docs directory, which can be built using Sphinx.

Contact

If you have any questions, issues, or feedback regarding TF-MELT, please feel free to contact the authors:

We look forward to hearing from you!

Contents

Indices and tables