Evidential Deep Learning with PyTorch

This repository contains a demontstration of how to build, train and evaluate a neural network capable of measuring epistemic uncertainty as proposed by the authors of Evidential Deep Learning to Quantify Classification Uncertainty, as well as some experiments performed on such architectures.

Resources

• Dataset: MNIST - sourced from TorchVision
• Working environment: Ubuntu18.04 LTS / Python 3.6.9 / virtualenv
• Use the Makefile commands to:
  • create the project virtual environment
  • print the source terminal command to activate environment in terminal
  • run tensorboard to view training progress & results

Project structure

├── data                                # Data used in the project
├── environment                         # Definition and contents of the project virtualenv
├── output                              # Default location for model training results
│   └── lightning_logs                  # Generated automatically by pytorch-lightning during training
└── src                                 # Source files of the project
    ├── dataset                         # Classes used in building and managing the project dataset
    ├── model                           # Classes and functions used in building and running the models
    ├── settings                        # Setting variables used across the repo 
    └── utilities                       # Utility functions