Local-Mixup

This repo contains the code (notebooks, scripts) for Local Mixup, as presented in our paper "Preventing Manifold Intrusion with Locality: Local Mixup" (https://arxiv.org/abs/2201.04368). We also provide the dataset of the two coiling spirals used in our paper. Note that this dataset can also be generated by the user with different values of noise.

Requirements and Installation:

1. A Pytorch installation https://pytorch.org
2. Python version 3.8.1 (lower versions might not work)
3. (optional) CIFAR10 dataset,CIAFAR100 dataset, SVHN, Fashion MNIST from https://pytorch.org/vision/stable/datasets.html.

Experiments

Low dimension (spirals dataset)

We provide:

• Generate_Spiral_dataset.ipynb used to generate the spirals dataset.
• Spiral_Experiment.ipynb used to carried out the spirals dataset.
• data\spirals_datasets the folder that contains the spirals dataset.

High Dimension

Lipschitz Lower Bound

We provide the script cifar10_lipschitz.py used to compute $Q(D)$ for Mixup, Local Mixup and Vanilla.

CIFAR10 Resnet18

We provide the notebook cifar.ipynb used to compute the error rates on Cifar10.

Fashion-MNIST DenseNet

We provide the script fashionmnist.py used to compute the error rates on Fashion MNIST.

SVHN Lenet

We provide the script svhn.py used to compute the error rates on MNIST.