Differentiable Automatic Data Augmentation

Yonggang Li, Guosheng Hu, yongtao wang, Timothy Hospedales, Neil Robertson, Yongxin Yang

Research output: Chapter in Book/Report/Conference proceedingConference contribution

11 Downloads (Pure)


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-theart while achieving very comparable accuracy. The code is available at https://github.com/VDIGPKU/DADA.
Original languageEnglish
Title of host publicationEuropean Conference on Computer Vision 2020: Proceedings
Number of pages16
ISBN (Electronic)978-3-030-58542-6
ISBN (Print)978-3-030-58541-9
Publication statusPublished - 17 Nov 2020
EventEuropean Conference on Computer Vision 2020 - Glasgow, United Kingdom
Duration: 23 Aug 202028 Aug 2020

Publication series

NameLecture Notes in Computer Science
ISSN (Print)0302-9743


ConferenceEuropean Conference on Computer Vision 2020
Country/TerritoryUnited Kingdom
Internet address


Dive into the research topics of 'Differentiable Automatic Data Augmentation'. Together they form a unique fingerprint.

Cite this