Regularized Evidential Model Code (RED)

Paper Title: Learn to Accumulate Evidence from All Training Samples: Theory and Practice
Paper link: https://openreview.net/forum?id=2MaUpKBSju

This repository contains source code for the Regularized Evidential Model Code (RED) model.

Requirements

Runnning Experiments in this repository requires following :

• Python
• pytorch
• Torchvision

Note: Additional details of the requirements can be found in the file "requirements.txt" and corresponding code folders

Running Experiments

There are 5 sets of experiments

1. Toy MNIST Experiments

   For The ToyMnist Experiments and results, Consider the folder ToyMnistExperiment. To run all experiments, run ``` bash submit_demo.sh ```

2. All Remaining MNIST Experiments

   For all remaining MNIST experiments and results, consider the folder MainMnistExperiment

   Run ``` bash submit_baseline.sh ``` to obtain Cross-Entropy based standard classifier model's results
   Run ``` bash submit_job.sh ``` to obtain all Evidential model results (All hyperparameters and settings)
   
3. All Cifar10 Experiments

   For all Cifar10 experiments and results, consider the folder MainCifar10Experiments

   Run ``` bash submit_baseline_ce.sh ``` to obtain Cross-Entropy based standard classifier model's results
   Run ``` bash submit_evidential.sh ``` to obtain all Evidential model results (All hyperparameters and settings)
   
4. All Cifar100 Experiments

   For all Cifar100 experiments and results, consider the folder Cifar100Experiments

   Run ``` bash 1_submit_baseline_ce.sh ``` to obtain Cross-Entropy based standard classifier model's results
   Run ``` bash 2_submit_evidential.sh ``` to obtain all Evidential model results (All hyperparameters and settings)
   
5. All Few Shot Classification Experiments

   For all Few-Shot Classification experiments with mini-ImageNet, consider the folder FewShotExperiments

   Run ``` bash 1_submit_baseline_ce.sh ``` to obtain Cross-Entropy based standard classifier model's results
   Run ``` bash 2_submit_evidential.sh ``` to obtain all Evidential model results (All hyperparameters and settings)
   

Datasets:

MNIST dataset, Cifar10 dataset, and Cifar100 dataset are automatically downloaded from the script.
mini-Imagenet dataset needs to be downloaded and placed in the materials folder. Follow instructions in 0_mini_imagenet_instructions.txt for mini-ImageNet.

References:

• MNIST experiments are based on the github repository: https://github.com/dougbrion/pytorch-classification-uncertainty
• Cifar10 experiments are based on the github repostiory: https://github.com/kuangliu/pytorch-cifar.git
• Cifar100 experiments are based on the github repository: https://github.com/weiaicunzai/pytorch-cifar100.git
• mini-ImageNet experiments are based on the github repository: https://github.com/yinboc/few-shot-meta-baseline.git