Project Overview

The muben package consists of several sub-packages, each tailored for specific functionalities within the backbond and/or UQ method trainig workflow. Below is a detailed overview of each sub-package and its purpose:

module muben.args

Manages command-line arguments and configuration settings for the muben package.

• Command-line argument/configuration file parsing
• Parameter validation and defaults

module muben.dataset

Provides functionality for loading, processing, and handling datasets.

• Dataset loading
• Feature generation
• Data preparation for dataloader
• Batching functions
• Collating functions

module muben.layers

Defines the output layer for compatibility with various objects (classification/regression), number of tasks, and UQ method (especially for Bayes-by-Backprop).

• Custom the output layer

module muben.model

Focuses on model definition and implementation.

• Model architecture and forward functions
• Loading functions for pre-trained model weights

module muben.uncertainty

Specializes in uncertainty estimation model architecture and training schemes. Notice that some UQ methods without special training steps or the need for modifying backbone layers are directly incorporated in the trainer.

• Uncertainty estimation functions
• Uncertainty estimation training schemes

module muben.train

Dedicated to the training process, including batch processing, epoch management, and callback functionalities. This module ensures efficient and effective model training, providing a robust framework for different training regimes.

• Training loops and batch processing
• Callbacks and training hooks
• Training metrics and evaluation

module muben.utils

Offers a collection of utility functions and helper tools that support the broader functionality of the muben package. This module includes miscellaneous functionalities such as logging, data manipulation, and performance metrics.

• Logging and debugging tools
• Data manipulation utilities
• Performance and evaluation metrics