Skip to content

Latest commit

 

History

History
65 lines (43 loc) · 2.54 KB

README.md

File metadata and controls

65 lines (43 loc) · 2.54 KB

Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It

arXiv Models on Hugging Face

These experiments build on top of TorchUncertainty.

Requirements

Make sure that TorchUncertainty >= 0.4.0 is installed.

First, install your desired PyTorch version for instance with

pip3 install torch torchvision

Then install the local torch-uncertainty-ls on your computer with:

pip install -e .

The latest release of TorchUncertainty will be installed automatically.

LaTeX

You will need LaTeX installed on your computer to plot the figures that include LaTeX symbols.

How to use this software

Experiments

The experiments folder provides the configuration files and the commands to run and reproduce the results of the paper. The tabular data experiments are small scale and therefore included in separate notebooks.

Please find the commands in the readme.md files contained in each of the subfolders.

If you do not wish to perform the experiments on your machine, you may download the models directly on HuggingFace.

Notebooks

The notebooks folder provides jupyter notebooks to reproduce the plots made in the paper. Just update the paths to the models that you trained to create your own figures.

We include:

  • CIFAR-100 with DenseNet
  • ImageNet-1k with VIT-S-16 and with ResNet-50 (the latter with logit normalization and hyperparameter optimization)
  • Cityscapes with DeepLabV3+ (ResNet-101 backbone)
  • IMDB with an LSTM-based model

Citation

If you find this work helpful for your research, consider citing:

@inproceedings{xia2024understanding,
    title={Towards Understanding Why Label Smoothing Degrades Selective Classification and How to Fix It},
    author={Xia, Guoxuan and Laurent, Olivier and Franchi, Gianni and Bouganis, Christos-Savvas},
    booktitle={ICLR},
    year={2025}       
}

You will find 3 other notebooks directly in the experiments folder:

  • Bank Marteking with multilayer perceptron in experiments/tabular
  • Online Shoppers with multilayer perceptron in experiments/tabular
  • CIFAR-10 with ResNet-20 in experiments/cifar10