robustness.data_augmentation module

Module responsible for data augmentation constants and configuration.

class robustness.data_augmentation.Lighting(alphastd, eigval, eigvec)

Bases: object

Lighting noise (see https://git.io/fhBOc)

robustness.data_augmentation.TRAIN_TRANSFORMS_IMAGENET

Random crop, Random flip, Color Jitter, and Lighting Transform (see https://git.io/fhBOc)

Type:Standard training data augmentation for ImageNet-scale datasets

robustness.data_augmentation.TEST_TRANSFORMS_IMAGENET

Standard test data processing (no augmentation) for ImageNet-scale datasets, Resized to 256x256 then center cropped to 224x224.

robustness.data_augmentation.TRAIN_TRANSFORMS_DEFAULT(size)

Generic training data transform, given image side length does random cropping, flipping, color jitter, and rotation. Called as, for example, robustness.data_augmentation.TRAIN_TRANSFORMS_DEFAULT(32)() for CIFAR-10.

robustness.data_augmentation.TEST_TRANSFORMS_DEFAULT(size)

Generic test data transform (no augmentation) to complement robustness.data_augmentation.TEST_TRANSFORMS_DEFAULT(), takes in an image side length.