Unrestricted Adversarial Examples

This repository contains the Two-Class Unambiguous images required for evaluation of the warmup to the Unrestricted Adversarial Examples contest.

Installation

git clone https://github.com/google/unrestricted-adversarial-examples
pip install -e unrestricted-adversarial-examples/bird-or-bicycle

bird-or-bicycle-download

Usage

import bird_or_bicycle 
train_dataset_folder = bird_or_bicycle.get_dataset('train')

# Use a pytorch directory-based dataset loader
import torchvision.datasets
train_dataset = torchvision.datasets.ImageFolder(
    train_dataset_folder)
    
# Or a keras one
from tensorflow.keras.preprocessing.image import ImageDataGenerator
train_datagen = ImageDataGenerator(rescale=1. / 255)
train_generator = train_datagen.flow_from_directory(train_dataset_folder)

Splits

• clean train -> 1k
• clean test -> 1k
• extras -> 40k