This document provides instructions for running the code used in the paper "Uncertainty Estimation by Flexible Evidential Deep Learning", submitted to NeurIPS 2025.
The provided code enables reviewers to reproduce the experimental results and explore various settings.

[Description]
The main code for running F-EDL experiments is located in main_fedl.py. The script sequentially performs the following steps:
* Training
* Testing
* Misclassification detection
* OOD detection
* Distribution shift detection (applicable for CIFAR-10)

To ensure a comprehensive evaluation, we also include implementations of baseline methods in main_edl.py and main_softmax.py. 

[How to Use]
To run F-EDL experiments, execute the main_fedl.py script with the desired arguments. Below are example commands for different experimental setups:

1. Classical setting: python main_fedl.py --spect_norm
2. Long-tailed setting (mild imbalance): python main_fedl.py --imbalance_factor 0.1 --spect_norm 
3. Long-tailed setting (heavy imbalance): python main_fedl.py --imbalance_factor 0.01 --spect_norm
4. Noisy setting: python main_fedl.py --ID_dataset "MNIST" --noise --spect_norm 

You can customize hyperparameters and options, such as dataset, batch size, learning rate, dropout rate, and weight decay, based on your experimental preferences.

[Environment Details] 
* GPU: NVIDIA GeForce RTX 4060
* Python: 3.10.12
* PyTorch: 2.0.0
* Torchvision: 0.15.1
* Numpy: 1.26.4
* Pandas: 2.2.3
* Scipy: 1.11.1
* Scikit-learn: 1.3.0
* Tqdm: 4.65.0

[Notes] 
1. Dataset availability: Some dataset files (e.g., CIFAR-10-C) have been intentionally excluded due to size constraints. To fully reproduce the experiments, please download the required datasets and store them in the designated directory specified in the code. 
2. Error handling: While we did not encounter any issues during experimentation, unintended errors might occur depending on the environment or settings. Please verify the dependencies and settings to ensure smooth execution.

[Conclusion] 
We hope this code allows reviewers to seamlessly reproduce our experiments and evaluate the potential of F-EDL. We appreciate your time, effort, and valuable feedback in reviewing our work. Thank you!