RandAugment is a practical and automated data augmentation method that significantly reduces the search space for augmentation policies, eliminating the need for a separate proxy task. This approach allows RandAugment to be trained directly on the target task, making it more efficient and adaptable to different models and datasets. The method is parameterized with two interpretable hyperparameters, $N$ and $M$, which control the number of transformations and the magnitude of augmentation, respectively. RandAugment achieves state-of-the-art performance on various datasets and tasks, including CIFAR-10/100, SVHN, ImageNet, and COCO, outperforming or matching previous learned augmentation methods. The results demonstrate that RandAugment can be effectively used across different tasks and datasets, providing a unified and efficient solution for data augmentation.RandAugment is a practical and automated data augmentation method that significantly reduces the search space for augmentation policies, eliminating the need for a separate proxy task. This approach allows RandAugment to be trained directly on the target task, making it more efficient and adaptable to different models and datasets. The method is parameterized with two interpretable hyperparameters, $N$ and $M$, which control the number of transformations and the magnitude of augmentation, respectively. RandAugment achieves state-of-the-art performance on various datasets and tasks, including CIFAR-10/100, SVHN, ImageNet, and COCO, outperforming or matching previous learned augmentation methods. The results demonstrate that RandAugment can be effectively used across different tasks and datasets, providing a unified and efficient solution for data augmentation.