Patch Gaussian

This is an unofficial pytorch implementation of Patch Gaussian augmentation which is proposed in the paper, Improving Robustness Without Sacrificing Accuracy with Patch Gaussian Augmentation [Lopes+, ArXiv 2019].

Patch Gaussian is a simple augmentation scheme that adds noise to randomly selected patches in an input image. This simple augmentation help to overcome the trade-off between robustness and accuracy.

Requirements

You will need the following to run the codes:

• Python 3.7+
• PyTorch 1.4+
• TorchVision
• numpy
• tqdm

Note that I run the code with Ubuntu 18, Pytorch 1.4.0, CUDA 10.1.

Usage

Generating sample image

By running patch_gaussian.py, you will be able to obtain a sample image like the teaser image of this repo. If you want to apply Patch Gaussian to an other image, please specify options as follows.

Example code:

python patch_gaussian.py -i [IMAGE_PATH] -o [OUTPUT_PATH]

Adding Patch Gaussian to your augmentation

If you want to add Patch Gaussin augmentation to your training, you can to AddPatchGaussian class in patch_gaussian.py. This class is able to use as same as classes in torchvision.transforms (e.g. torchvision.transforms.CenterCrop or torchvision.transforms.ToTensor , etc.).