DADA: Differentiable Automatic Data Augmentation

TL;DR

DADA introduces a differentiable approach to automatic data augmentation, significantly reducing computational costs while maintaining high accuracy, thereby enabling efficient training of deep networks across multiple datasets.

Contribution

DADA proposes a novel differentiable optimization method for data augmentation policies using Gumbel-Softmax and RELAX, improving efficiency over prior approaches.

Findings

DADA is at least ten times faster than AutoAugment.

It achieves comparable accuracy on CIFAR-10, CIFAR-100, SVHN, and ImageNet.

Effective in pre-training for downstream detection tasks.

Abstract

Data augmentation (DA) techniques aim to increase data variability, and thus train deep networks with better generalisation. The pioneering AutoAugment automated the search for optimal DA policies with reinforcement learning. However, AutoAugment is extremely computationally expensive, limiting its wide applicability. Followup works such as Population Based Augmentation (PBA) and Fast AutoAugment improved efficiency, but their optimization speed remains a bottleneck. In this paper, we propose Differentiable Automatic Data Augmentation (DADA) which dramatically reduces the cost. DADA relaxes the discrete DA policy selection to a differentiable optimization problem via Gumbel-Softmax. In addition, we introduce an unbiased gradient estimator, RELAX, leading to an efficient and effective one-pass optimization strategy to learn an efficient and accurate DA policy. We conduct extensive experiments on CIFAR-10, CIFAR-100, SVHN, and ImageNet datasets. Furthermore, we demonstrate the value of Auto DA in pre-training for downstream detection problems. Results show our DADA is at least one order of magnitude faster than the state-of-the-art while achieving very comparable accuracy. The code is available at https://github.com/VDIGPKU/DADA.