MengyuanChen21 / ICLR2024-REDL

[ICLR 2024] R-EDL: Relaxing Nonessential Settings of Evidential Deep Learning

[ICLR 2024] R-EDL: Relaxing Nonessential Settings of Evidential Deep Learning

Authors: Mengyuan Chen, Junyu Gao, Changsheng Xu.

Affiliations: Institute of Automation, Chinese Academy of Sciences

Dependencies:

Here we list our used requirements and dependencies.

• GPU: GeForce RTX 3090
• Python: 3.8.5
• PyTorch: 1.12.0
• Numpy: 1.21.2
• Pandas: 1.1.3
• Scipy: 1.3.1
• Scikit-learn: 1.0.1
• Wandb: 0.12.6
• Tqdm: 4.62.3

Data preparation:

The required datasets of the classical setting (MNIST/FMNIST/KMNIST/CIFAR-10/CIFAR-100/SVHN) will be automatically downloaded if your server has an Internet connection.

The required datasets of the few-shot setting (mini-ImageNet/CUB) can be downloaded by running:

cd IEDL/code_fsl
bash ./features/download.sh

Pre-trained models:

The pre-trained models of R-EDL can be downloaded from Google Disk. They need to be unzipped and put in the directory './code_classical/saved_models/'.

Quick start for experiments of classical setting:

To test pre-trained models, run:

python main.py --configid "1_mnist/mnist-redl-test" --suffix test
python main.py --configid "2_cifar10/cifar10-redl-test" --suffix test

To train from scratch, run:

python main.py --configid "1_mnist/mnist-redl-train" --suffix test
python main.py --configid "2_cifar10/cifar10-redl-train" --suffix test

Quick start for experiments of few-shot setting:

Given that this setting involves conducting experiments across 10,000 few-shot episodes, providing pre-trained models for testing becomes nearly impossible.

To train from scratch, run:

python main.py --configid "1_mini/5w1s-redl" --suffix test
python main.py --configid "1_mini/5w5s-redl" --suffix test
python main.py --configid "1_mini/5w20s-redl" --suffix test
python main.py --configid "1_mini/10w1s-redl" --suffix test
python main.py --configid "1_mini/10w5s-redl" --suffix test
python main.py --configid "1_mini/10w20s-redl" --suffix test

Acknowledgement

This project is built upon the repository of IEDL, Posterior Network, and Firth Bias Reduction in Few-shot Distribution Calibration. We would like to thank their authors for their excellent work. If you want to use and redistribe our code, please follow this license as well.

Contact

Feel free to contact me (Mengyuan Chen: [e-mail]) if anything is unclear or you are interested in potential collaboration.