Improving Auto-Augment via Augmentation-Wise Weight Sharing

TL;DR

This paper introduces an efficient proxy evaluation method for auto-augmentation policies using augmentation-wise weight sharing, significantly improving search speed and accuracy on CIFAR-10 and ImageNet.

Contribution

The paper proposes a novel augmentation-wise weight sharing approach for fast and reliable evaluation of augmentation policies in auto-augmentation search.

Findings

Achieves 1.24% top-1 error on CIFAR-10, the best for single models without extra data.

Reduces top-1 error on ImageNet ResNet-50 to 20.36%.

Outperforms existing auto-augmentation methods in both effectiveness and efficiency.

Abstract

The recent progress on automatically searching augmentation policies has boosted the performance substantially for various tasks. A key component of automatic augmentation search is the evaluation process for a particular augmentation policy, which is utilized to return reward and usually runs thousands of times. A plain evaluation process, which includes full model training and validation, would be time-consuming. To achieve efficiency, many choose to sacrifice evaluation reliability for speed. In this paper, we dive into the dynamics of augmented training of the model. This inspires us to design a powerful and efficient proxy task based on the Augmentation-Wise Weight Sharing (AWS) to form a fast yet accurate evaluation process in an elegant way. Comprehensive analysis verifies the superiority of this approach in terms of effectiveness and efficiency. The augmentation policies found by our method achieve superior accuracies compared with existing auto-augmentation search methods. On CIFAR-10, we achieve a top-1 error rate of 1.24%, which is currently the best performing single model without extra training data. On ImageNet, we get a top-1 error rate of 20.36% for ResNet-50, which leads to 3.34% absolute error rate reduction over the baseline augmentation.