Benchmarks¶
Monitizer currently supports 9 ID-datasets: MNIST, CIFAR-10, CIFAR-100, German Traffic Sign Recognition Benchmark (GTSRB), SVHN, Describable Textures Dataset (DTD), Fashion-MNIST, Kuzushiji-MNIST, ImageNet.
Note that it is fairly simple to include any other dataset, especially if it already exists in Torchvision. Check out Implementation of a dataset to see how you can include another dataset. There you will see how you can include your own custom dataset.
CIFAR-10¶
The CIFAR-10 dataset consists of 60,000 32x32 color images in 10 different classes. It is widely used for object recognition tasks.
Monitizer provides four specific OOD classes for CIFAR-10:
NewWorld/GTSRB: (handselected) GTSRB images share no similarity with any CIFAR-100 image
NewWorld/DTD: DTD images share no similarity with any CIFAR-10 image
UnseenObject/Cifar100: the general type of pictures is similar, but the (handselected) CIFAR-100 images contain objects that do not appear in CIFAR-10
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
CIFAR-100¶
Similar to CIFAR-10 but with 100 classes containing 600 images each, CIFAR-100 offers a more challenging classification problem.
Monitizer provides three specific OOD classes for CIFAR-100:
NewWorld/GTSRB: (handselected) GTSRB images share no similarity with any CIFAR-100 image
NewWorld/DTD: DTD images share no similarity with any CIFAR-100 image
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
Fashion-MNIST¶
Fashion-MNIST contains 70,000 grayscale images of 28x28 pixels representing 10 different fashion categories.
Monitizer provides one specific OOD classes for FashionMNIST
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
GTSRB (German Traffic Sign Recognition Benchmark)¶
GTSRB is a multi-class traffic sign recognition dataset with more than 50,000 images of traffic signs in different conditions.
Monitizer provides four specific OOD classes for GTSRB:
NewWorld/CIFAR10: (handselected) CIFAR-10 images share no similarity with any GTSRB image
NewWorld/DTD: DTD images share no similarity with any GTSRB image
NewWorld/SVHN: (handselected) SVHN images that share no similarity with any GTSRB image
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
ImageNet¶
ImageNet is a large-scale dataset with over 14 million images across more than 20,000 categories, commonly used for deep learning.
KMNIST (Kuzushiji-MNIST)¶
KMNIST is a dataset of 70,000 grayscale images of Japanese Kuzushiji characters, designed as a drop-in replacement for MNIST.
Monitizer provides one specific OOD classes for FashionMNIST
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
MNIST¶
MNIST contains 70,000 grayscale images of handwritten digits (0-9) and is a classic benchmark for image classification.
Monitizer provides six specific OOD classes for MNIST:
NewWorld/CIFAR10: CIFAR-10 images share no similarity with any MNIST image
NewWorld/DTD: DTD images share no similarity with any MNIST image
UnseenEnvironment/SVHN: SVHN also shows numbers but in a different style than MNIST
UnseenObject/FashionMNIST: the type of images is the same, but the objects differ
UnseenObject/KMNIST: the type of images is the same, but the objects differ
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
SVHN (Street View House Numbers)¶
SVHN consists of real-world house number images obtained from Google Street View, used for digit recognition tasks.
Monitizer provides one specific OOD classes for FashionMNIST
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples
DTD (Describing Textures in the Wild)¶
DTD contains 5,640 images categorized into 47 texture classes and is used for texture recognition and classification.
Monitizer provides one specific OOD classes for FashionMNIST
WrongPrediction/FGSM: automatically generated OOD images by using a provided NN to generate adversarial examples