LA3: Efficient Label-Aware AutoAugment

Automated augmentation is an emerging and effective technique to search for data policies improve generalizability of deep neural network training. Most existing work focuses on constructing a unified policy applicable all samples in given dataset, without considering sample or class variations. In this paper, we propose novel two-stage algorithm, named Label-Aware AutoAugment (LA3), which takes advantage the label information, learns separately different labels. LA3 consists two learning stages, where first stage, individual methods are evaluated ranked each via Bayesian Optimization aided by predictor, allows us identify techniques under low cost. And second composite constructed out selection as well complementary augmentations, produces significant performance boost can be easily deployed typical model Extensive experiments demonstrate that achieves excellent matching surpassing CIFAR-10 CIFAR-100, new state-of-the-art ImageNet accuracy 79.97% ResNet-50 among auto-augmentation methods, while maintaining computational