Back

Randaugment: Practical automated data augmentation with a reduced search space

Ekin D. CubukBarret ZophJonathon ShlensQuoc V. Le

article2020NeurIPS4,400 citations

Proposes RandAugment, an automated data augmentation strategy that eliminates computationally expensive proxy searches by reducing the search space to two tunable parameters, matching or exceeding prior state-of-the-art methods across standard computer vision benchmarks.

Listen

Recent automated data augmentation methods have delivered large gains in image classification and object detection accuracy, yet they require a costly separate search phase on small proxy tasks. This approach raises training complexity, inflates compute cost, and produces policies that cannot adapt to different model or dataset sizes.

The document introduces RandAugment, a method that removes the separate search entirely by collapsing the augmentation policy into two simple, interpretable hyperparameters: the number of transformations applied per image (N) and a single global distortion magnitude (M). The approach was tested through direct grid search on the target datasets and models, covering CIFAR-10/100, SVHN, ImageNet classification, and COCO object detection with architectures ranging from Wide-ResNet to EfficientNet-B7 and RetinaNet.

Experiments show that optimal augmentation strength increases systematically with both model size and training-set size, confirming that policies learned on proxy tasks are often suboptimal. RandAugment matches or exceeds the accuracy of prior methods such as AutoAugment, Fast AutoAugment, and Population Based Augmentation while reducing the search space by roughly ten orders of magnitude. On ImageNet it reaches 85.0 % top-1 accuracy, a 0.6 % gain over the previous state of the art and a 1.0 % gain over standard baseline augmentation. On COCO it improves mean average precision by 1.0–1.3 % over baseline and stays within 0.3 % of AutoAugment despite using a smaller transformation library and no separate search.

These results indicate that data-augmentation strength can be tuned directly to the model and data at hand, eliminating the computational overhead and transfer assumptions of earlier techniques. The gains incur no extra inference cost and appear consistently across classification and detection tasks.

The method can be applied immediately to new image-classification and detection pipelines by performing a small grid search over N and M. Further gains may come from expanding the transformation set for specialized tasks such as bounding-box augmentation, and from testing the approach in domains such as segmentation, speech, and audio. The primary limitation is that the current transformation list was not optimized per task; results on very different data modalities therefore remain to be verified. Overall the empirical evidence across multiple benchmarks supports high confidence in the reported accuracy improvements.

Googlecomputational efficiencydata augmentationdeep learningRandAugmentData AugmentationSearch Space
arXiv: 1909.13719tensorflow/tpu
Cover for Randaugment: Practical automated data augmentation with a reduced search space

Abstract

Recent work has shown that data augmentation has the potential to significantly improve the generalization of deep learning models. Recently, automated augmentation strategies have led to state-of-the-art results in image classification and object detection. While these strategies were optimized for improving validation accuracy, they also led to state-of-the-art results in semi-supervised learning and improved robustness to common corruptions of images. An obstacle to a large-scale adoption of these methods is a separate search phase which increases the training complexity and may substantially increase the computational cost. Additionally, due to the separate search phase, these approaches are unable to adjust the regularization strength based on model or dataset size. Automated augmentation policies are often found by training small models on small datasets and subsequently applied to train larger models. In this work, we remove both of these obstacles. RandAugment has a significantly reduced search space which allows it to be trained on the target task with no need for a separate proxy task. Furthermore, due to the parameterization, the regularization strength may be tailored to different model and dataset sizes. RandAugment can be used uniformly across different tasks and datasets and works out of the box, matching or surpassing all previous automated augmentation approaches on CIFAR-10/100, SVHN, and ImageNet. On the ImageNet dataset we achieve 85.0% accuracy, a 0.6% increase over the previous state-of-the-art and 1.0% increase over baseline augmentation. On object detection, RandAugment leads to 1.0-1.3% improvement over baseline augmentation, and is within 0.3% mAP of AutoAugment on COCO. Finally, due to its interpretable hyperparameter, RandAugment may be used to investigate the role of data augmentation with varying model and dataset size. Code is available online.

Topics

computational efficiencydata augmentationdeep learningobject detectionsemi-supervised learningsupervised machine learningtransfer learningvalidationsearch processsupervised classification

Keywords

RandAugmentData AugmentationSearch SpaceMagnitudeNumber of TransformationsCIFAR‑10ImageNetAutoAugmentProxy TaskRegularization Strength

Table of Contents

  • 1 Introduction
  • 2 Related Work
  • 3 Methods
  • 4 Results
  • 4.1 Systematic failures of a separate proxy task
  • 4.2 CIFAR
  • 4.3 SVHN
  • 4.4 ImageNet
  • 4.5 COCO
  • 4.6 Investigating the dependence on the included transformations
  • 4.7 Learning the probabilities for selecting image transformations
  • 5 Discussion
  • 6 Acknowledgements
  • References
  • A Appendix
  • A.1 Second order term from bilevel optimization
  • A.1.1 Magnitude methods
  • A.1.2 Optimizing individual transformation magnitudes
  • A.2 Experimental Details
  • A.2.1 CIFAR
  • A.2.2 SVHN
  • A.2.3 ImageNet
  • A.3 COCO

Knowls

  1. Knowl 1 — RandAugment Automated Data Augmentation Algorithm

    algorithm

    RandAugment is an automated data augmentation method designed to eliminate the separate search phase required by prior learned augmentation techniques (such as AutoAugment) by drastically reducing the search space to two hyper-parameters: NN (the number of sequential transformation operations to apply) and MM (the global distortion magnitude applied across all selected transformations).

    During training, for each image, NN transformations are sampled uniformly at random with replacement from a candidate pool of K=14K = 14 standard image transformations, each having equal selection probability 1/K1/K. The candidate transformations are:

    • Identity
    • AutoContrast
    • Equalize
    • Rotate
    • Solarize
    • Color
    • Posterize
    • Contrast
    • Brightness
    • Sharpness
    • ShearX
    • ShearY
    • TranslateX
    • TranslateY

    Each transformation's distortion strength is mapped to a shared linear integer scale M∈[0,10]M \in [0, 10] (or extended to [0,30][0, 30]), where 00 corresponds to minimum distortion and the maximum value corresponds to the largest transformation strength defined for that operation. Because all operations share the single global magnitude MM and are selected uniformly at random, the policy search space is reduced to the order of 10210^2 possible hyperparameter pairs (N,M)(N, M), which can be optimized directly on the target training task via a minimal grid search.

    import numpy as np

transforms = [
    'Identity', 'AutoContrast', 'Equalize',
    'Rotate', 'Solarize', 'Color', 'Posterize',
    'Contrast', 'Brightness', 'Sharpness',
    'ShearX', 'ShearY', 'TranslateX', 'TranslateY'
]

def randaugment(N, M):
    """Generate a set of distortions.
    Args:
        N: Number of augmentation transformations to apply sequentially.
        M: Magnitude applied to all sampled transformations.
    Returns:
        A list of (operation_name, magnitude) pairs of length N.
    """
    sampled_ops = np.random.choice(transforms, N)
    return [(op, M) for op in sampled_ops]

  2. Knowl 2 — Dependence of Optimal Augmentation Distortion on Model Capacity and Dataset Size

    empirical result

    Experiments on CIFAR-10 using Wide-ResNet-28 architectures demonstrate that the optimal data augmentation strength depends systematically on both model capacity and dataset size:

    1. Model Capacity Dependence: When varying the network widening factor kk in Wide-ResNet-28-kk (from k=2k=2 to k=10k=10, with N=1N=1), the optimal distortion magnitude M∗M^* increases monotonically with model capacity. For example, the optimal magnitude M∗M^* for Wide-ResNet-28-2 is approximately 99, whereas for Wide-ResNet-28-10 it increases to approximately 1717. Larger neural networks demand substantially stronger data distortions for effective regularization.

    2. Training Dataset Size Dependence: For a fixed network architecture (Wide-ResNet-28-10, N=1N=1), the optimal distortion magnitude M∗M^* increases monotonically with the number of training examples, rising from M∗≈5M^* \approx 5 for a 1K training subset to M∗≈17M^* \approx 17 for the full 45K training dataset.

    These findings challenge the standard assumption of automated data augmentation methods that a policy optimized on a small proxy task (i.e., a smaller network or a small data subset) transfers effectively to larger target models and datasets. Because small proxy tasks systematically bias the optimal distortion magnitude toward lower strengths, proxy-searched policies are sub-optimal for large-scale training.

  3. Knowl 3 — Image Classification Accuracy of RandAugment on CIFAR-10, CIFAR-100, and SVHN

    data/table

    RandAugment matches or exceeds the predictive accuracy of complex learned augmentation methods across standard image classification benchmarks without requiring a separate proxy search phase. Evaluated methods include default baseline augmentation (flips and crops), Population Based Augmentation (PBA), Fast AutoAugment (Fast AA), AutoAugment (AA), and RandAugment (RA). Reported results represent test accuracy (%) averaged over 10 independent runs:

    Dataset / Architecture Baseline PBA Fast AA AA RA (Ours)
    CIFAR-10
    Wide-ResNet-28-2 94.9 – – 95.9 95.8
    Wide-ResNet-28-10 96.1 97.4 97.3 97.4 97.3
    Shake-Shake (26 2×\times96d) 97.1 98.0 98.0 98.0 98.0
    PyramidNet + ShakeDrop 97.3 98.5 98.3 98.5 98.5
    CIFAR-100
    Wide-ResNet-28-2 75.4 – – 78.5 78.3
    Wide-ResNet-28-10 81.2 83.3 82.7 82.9 83.3
    SVHN (core set, 73K)
    Wide-ResNet-28-2 96.7 – – 98.0 98.3
    Wide-ResNet-28-10 96.9 – – 98.1 98.3
    SVHN (full)
    Wide-ResNet-28-2 98.2 – – 98.7 98.7
    Wide-ResNet-28-10 98.5 98.9 98.8 98.9 99.0

    On Wide-ResNet-28-2 trained on the SVHN core set (73K examples), applying RandAugment achieves 98.3%98.3\% accuracy, outperforming the baseline model trained with the additional 531K extra training images (98.2%98.2\%).

  4. Knowl 4 — ImageNet Classification Performance Across ResNet-50 and EfficientNet Architectures

    data/table

    On the ImageNet classification task, RandAugment matches previous learned augmentation strategies on smaller models and provides larger accuracy gains on high-capacity architectures, where proxy-searched policies tend to yield diminishing returns.

    Model Baseline Fast AA AutoAugment RandAugment (Ours)
    ResNet-50 76.3 / 93.1 77.6 / 93.7 77.6 / 93.8 77.6 / 93.8
    EfficientNet-B5 83.2 / 96.7 – 83.3 / 96.7 83.9 / 96.8
    EfficientNet-B7 84.0 / 96.9 – 84.4 / 97.1 85.0 / 97.2

    Accuracies are reported as Top-1 / Top-5 percentages (%). On EfficientNet-B7, RandAugment improves Top-1 accuracy by 1.0%1.0\% over standard baseline augmentation (from 84.0%84.0\% to 85.0%85.0\%) and by 0.6%0.6\% over AutoAugment (84.4%84.4\%), demonstrating that scaling augmentation magnitude on the full target architecture yields improvements comparable to neural architecture search without inference overhead.

  5. Knowl 5 — Object Detection Performance of RandAugment on COCO Benchmark

    data/table

    RandAugment generalizes directly to object detection on the COCO dataset using RetinaNet with ResNet-101 and ResNet-200 backbones trained for 300 epochs from random initialization using focal loss (α=0.25,γ=1.5\alpha = 0.25, \gamma = 1.5):

    Model Backbone Augmentation mAP (%) Search Space Size
    ResNet-101 Baseline 38.8 0
    ResNet-101 AutoAugment 40.4 103410^{34}
    ResNet-101 RandAugment 40.1 10210^{2}
    ResNet-200 Baseline 39.9 0
    ResNet-200 AutoAugment 42.1 103410^{34}
    ResNet-200 RandAugment 41.9 10210^{2}

    RandAugment outperforms the baseline augmentation by 1.3%1.3\% mAP on ResNet-101 and 2.0%2.0\% mAP on ResNet-200, achieving performance within 0.3%0.3\% mAP of AutoAugment. While AutoAugment required approximately 15,000 GPU hours of search on proxy tasks and utilized specialized localized bounding-box transformations, RandAugment was tuned over only 6 hyperparameter combinations (N=1N=1, distortion magnitudes between 4 and 9) using only image-level transformations applied to bounding boxes geometrically.

  6. Knowl 6 — Effect of Candidate Transformation Diversity and Individual Operations on Validation Accuracy

    empirical result

    Ablation experiments on CIFAR-10 using Wide-ResNet-28-2 (N=3,M=4N=3, M=4, without standard flips, pad-and-crop, or cutout) evaluate the effect of candidate transformation pool size and individual operations:

    1. Transformation Pool Size: Median validation accuracy improves monotonically as the number of randomly sampled candidate transformations in the pool increases from 1 to 14 (rising from an unaugmented baseline of ≈86%\approx 86\% to >94%> 94\%). Even a pool of only 2 transformations yields over 1%1\% average accuracy improvement over the unaugmented baseline.

    2. Individual Transformation Impact: The average marginal contribution Δ(%)\Delta (\%) to validation accuracy when adding a specific transformation to a randomly sampled subset of transformations is:

    Transformation Δ\Delta (%) Transformation Δ\Delta (%)
    Rotate +1.3 ShearX +0.9
    ShearY +0.9 TranslateY +0.4
    TranslateX +0.4 AutoContrast +0.1
    Sharpness +0.1 Identity +0.1
    Contrast 0.0 Color 0.0
    Brightness 0.0 Equalize -0.0
    Solarize -0.1 Posterize -0.3

    Geometric transformations (Rotate, ShearX, ShearY, TranslateX, TranslateY) provide the largest individual accuracy gains. Although certain color transformations (such as Posterize and Solarize) slightly decrease average accuracy when evaluated in isolation, including the full set of 14 transformations jointly achieves the best overall performance without needing dataset-specific pruning.

  7. Knowl 7 — Invariance to Distortion Magnitude Schedules and Efficacy of Shared Magnitude Parameter

    empirical result

    Evaluating strategies for scheduling the global distortion magnitude MM on CIFAR-10 with Wide-ResNet-28-10 (trained for 200 epochs across 48 hyperparameter configurations per method) reveals that scheduling complexity does not improve validation accuracy over a fixed constant magnitude:

    • Random Magnitude (uniformly sampled between two values): 97.3%97.3\%
    • Constant Magnitude (fixed single value throughout training): 97.2%97.2\%
    • Linearly Increasing Magnitude (linearly interpolated over training): 97.2%97.2\%
    • Random Magnitude with Increasing Upper Bound (upper bound increases linearly): 97.3%97.3\%

    Because all four schedules achieve statistically indistinguishable performance, a single constant magnitude parameter MM is sufficient.

    Furthermore, fixing a shared magnitude across all transformations and varying only a single transformation's individual magnitude across [0,19][0, 19] results in minimal accuracy variation compared to the optimal individual magnitude: a maximum difference of 0.18%−0.19%0.18\% - 0.19\% for Rotate and 0.05%−0.07%0.05\% - 0.07\% for TranslateX on Wide-ResNet-28-2 and Wide-ResNet-28-10. Tying all transformation magnitudes to a single global hyperparameter MM thus drastically reduces the search space with negligible loss in peak accuracy.

  8. Knowl 8 — First-Order Differentiable Density Matching for Augmentation Selection Probabilities

    model/method

    To evaluate whether learning non-uniform transformation selection probabilities improves upon RandAugment's uniform 1/K1/K sampling, selection probabilities αij\alpha_{ij} are parameterized and learned via gradient descent, where αij\alpha_{ij} is the probability of selecting transformation ii (i∈{1,…,K}i \in \{1, \dots, K\}, with K=14K=14) at sequential step jj (j∈{1,…,N}j \in \{1, \dots, N\}, with N=2N=2), totaling K×N=28K \times N = 28 parameters.

    Because 11 of the 14 transformations are differentiable (excluding Posterize, Equalize, and AutoContrast), the selection weights are initialized to uniform probability (1/K1/K) and optimized via a first-order density matching objective on held-out validation images.

    Performance Comparison (Test Accuracy %):

    • Reduced CIFAR-10 (4K examples, 500 epochs):
      • Wide-ResNet-28-2: Baseline 82.0%82.0\%, AutoAugment 85.6%85.6\%, RandAugment 85.3%85.3\%, 1st-Order Density Matching 85.5%85.5\%
      • Wide-ResNet-28-10: Baseline 83.5%83.5\%, AutoAugment 87.7%87.7\%, RandAugment 86.8%86.8\%, 1st-Order Density Matching 87.4%87.4\%
    • Full CIFAR-10:
      • Wide-ResNet-28-2: Baseline 94.9%94.9\%, AutoAugment 95.9%95.9\%, RandAugment 95.8%95.8\%, 1st-Order Density Matching 96.1%96.1\%
      • Wide-ResNet-28-10: Baseline 96.1%96.1\%, AutoAugment 97.4%97.4\%, RandAugment 97.3%97.3\%, 1st-Order Density Matching 97.4%97.4\%

    While first-order density matching slightly improves accuracy over uniform RandAugment on small datasets, it requires evaluating all KK transformations NN times independently per image (K×NK \times N operations per image), making it computationally prohibitive for large-scale datasets such as ImageNet.

Coverage note — None was omitted; all primary contributions, empirical analyses, benchmark comparisons, ablations on schedules and transformation subsets, and differentiable optimization extensions from the paper and appendix are covered.

References

  1. 1.Antreas Antoniou, Amos Storkey, and Harrison Edwards. Data augmentation generative adversarial networks. arXiv preprint arXiv:1711.04340, 2017. 2
  2. 2.Liang-Chieh Chen, Maxwell Collins, Yukun Zhu, George Papandreou, Barret Zoph, Florian Schroff, Hartwig Adam, and Jon Shlens. Searching for efficient multi-scale architectures for dense image prediction. In Advances in Neural Information Processing Systems, pages 8699–8710, 2018. 2
  3. 3.Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. IEEE transactions on pattern analysis and machine intelligence, 40(4):834–848, 2017. 8
  4. 4.Dan Ciregan, Ueli Meier, and Jurgen Schmidhuber. Multi-column deep neural networks for image classification. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, pages 3642–3649. IEEE, 2012. 2
  5. 5.Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation policies from data. arXiv preprint arXiv:1805.09501, 2018. 1, 2, 3, 4, 5, 6, 7, 8
  6. 6.Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2009. 1, 2
  7. 7.Terrance DeVries and Graham W Taylor. Dataset augmentation in feature space. arXiv preprint arXiv:1702.05538, 2017. 1, 2, 8
  8. 8.Terrance DeVries and Graham W Taylor. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017. 2, 6
  9. 9.Debidatta Dwibedi, Ishan Misra, and Martial Hebert. Cut, paste and learn: Surprisingly easy synthesis for instance detection. In Proceedings of the IEEE International Conference on Computer Vision, pages 1301–1310, 2017. 2
  10. 10.Hao-Shu Fang, Jianhua Sun, Runzhong Wang, Minghao Gou, Yong-Lu Li, and Cewu Lu. Instaboost: Boosting instance segmentation via probability map guided copypasting. arXiv preprint arXiv:1908.07801, 2019. 1
  11. 11.Nic Ford, Justin Gilmer, Nicolas Carlini, and Dogus Cubuk. Adversarial examples are a natural consequence of test error in noise. arXiv preprint arXiv:1901.10513, 2019. 2
  12. 12.Xavier Gastaldi. Shake-shake regularization. arXiv preprint arXiv:1705.07485, 2017. 4, 13
  13. 13.Ross Girshick, Ilija Radosavovic, Georgia Gkioxari, Piotr Dollar, and Kaiming He. Detectron, 2018. 1, 2, 8
  14. 14.Daniel Golovin, Benjamin Solnik, Subhodeep Moitra, Greg Kochanski, John Karro, and D Sculley. Google vizier: A service for black-box optimization. In Proceedings of the 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1487–1495. ACM, 2017. 4
  15. 15.Dongyoon Han, Jiwhan Kim, and Junmo Kim. Deep pyramidal residual networks. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 6307–6315. IEEE, 2017. 1
  16. 16.Awni Hannun, Carl Case, Jared Casper, Bryan Catanzaro, Greg Diamos, Erich Elsen, Ryan Prenger, Sanjeev Satheesh, Shubho Sengupta, Adam Coates, et al. Deep speech: Scaling up end-to-end speech recognition. arXiv preprint arXiv:1412.5567, 2014. 1
  17. 17.Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2016. 1, 6
  18. 18.Shawn Hershey, Sourish Chaudhuri, Daniel PW Ellis, Jort F Gemmeke, Aren Jansen, R Channing Moore, Manoj Plakal, Devin Platt, Rif A Saurous, Bryan Seybold, et al. Cnn architectures for large-scale audio classification. In 2017 ieee international conference on acoustics, speech and signal processing (icassp), pages 131–135. IEEE, 2017. 8
  19. 19.Geoffrey Hinton, Li Deng, Dong Yu, George Dahl, Abdelrahman Mohamed, Navdeep Jaitly, Andrew Senior, Vincent Vanhoucke, Patrick Nguyen, Brian Kingsbury, et al. Deep neural networks for acoustic modeling in speech recognition. IEEE Signal processing magazine, 29, 2012. 8
  20. 20.Daniel Ho, Eric Liang, Ion Stoica, Pieter Abbeel, and Xi Chen. Population based augmentation: Efficient learning of augmentation policy schedules. arXiv preprint arXiv:1905.05393, 2019. 1, 2, 3, 4, 7, 8
  21. 21.Naoyuki Kanda, Ryu Takeda, and Yasunari Obuchi. Elastic spectral distortion for low resource speech recognition with deep neural networks. In 2013 IEEE Workshop on Automatic Speech Recognition and Understanding, pages 309–314. IEEE, 2013. 1
  22. 22.Alex Krizhevsky and Geoffrey Hinton. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. 1, 2
  23. 23.Alex Krizhevsky, Ilya Sutskever, and Geoffrey E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems, 2012. 1, 2, 8
  24. 24.Joseph Lemley, Shabab Bazrafkan, and Peter Corcoran. Smart augmentation learning an optimal data augmentation strategy. IEEE Access, 5:5858–5869, 2017. 2
  25. 25.Sungbin Lim, Ildoo Kim, Taesup Kim, Chiheon Kim, and Sungwoong Kim. Fast autoaugment. arXiv preprint arXiv:1905.00397, 2019. 1, 2, 3, 4, 7, 8
  26. 26.Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. Focal loss for dense object detection. In Proceedings of the IEEE international conference on computer vision, pages 2980–2988, 2017. 6
  27. 27.Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 2, 6
  28. 28.Chenxi Liu, Barret Zoph, Jonathon Shlens, Wei Hua, Li-Jia Li, Li Fei-Fei, Alan Yuille, Jonathan Huang, and Kevin Murphy. Progressive neural architecture search. arXiv preprint arXiv:1712.00559, 2017. 2, 6
  29. 29.Hanxiao Liu, Karen Simonyan, Oriol Vinyals, Chrisantha Fernando, and Koray Kavukcuoglu. Hierarchical representations for efficient architecture search. In International Conference on Learning Representations, 2018. 2
  30. 30.Hanxiao Liu, Karen Simonyan, and Yiming Yang. Darts: Differentiable architecture search. arXiv preprint arXiv:1806.09055, 2018. 2, 7, 8, 12
  31. 31.Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21–37. Springer, 2016. 2
  32. 32.Raphael Gontijo Lopes, Dong Yin, Ben Poole, Justin Gilmer, and Ekin D Cubuk. Improving robustness without sacrificing accuracy with patch gaussian augmentation. arXiv preprint arXiv:1906.02611, 2019. 1, 2, 6, 8
  33. 33.Seongkyu Mun, Sangwook Park, David K Han, and Hanseok Ko. Generative adversarial network based acoustic scene training set augmentation and selection using svm hyperplane. In Detection and Classification of Acoustic Scenes and Events Workshop, 2017. 2
  34. 34.Yuval Netzer, Tao Wang, Adam Coates, Alessandro Bissacco, Bo Wu, and Andrew Y Ng. Reading digits in natural images with unsupervised feature learning. In NIPS Workshop on Deep Learning and Unsupervised Feature Learning, 2011. 1, 2, 6
  35. 35.Jiquan Ngiam, Benjamin Caine, Wei Han, Brandon Yang, Yuning Chai, Pei Sun, Yin Zhou, Xi Yi, Ouais Alsharif, Patrick Nguyen, et al. Starnet: Targeted computation for object detection in point clouds. arXiv preprint arXiv:1908.11069, 2019. 8
  36. 36.Daniel S Park, William Chan, Yu Zhang, Chung-Cheng Chiu, Barret Zoph, Ekin D Cubuk, and Quoc V Le. Specaugment: A simple data augmentation method for automatic speech recognition. arXiv preprint arXiv:1904.08779, 2019. 1, 4, 8
  37. 37.Luis Perez and Jason Wang. The effectiveness of data augmentation in image classification using deep learning. arXiv preprint arXiv:1712.04621, 2017. 2
  38. 38.Hieu Pham, Melody Y Guan, Barret Zoph, Quoc V Le, and Jeff Dean. Efficient neural architecture search via parameter sharing. In International Conference on Machine Learning, 2018. 2
  39. 39.Alexander J Ratner, Henry Ehrenberg, Zeshan Hussain, Jared Dunnmon, and Christopher Re. Learning to compose domain-specific transformations for data augmentation. In Advances in Neural Information Processing Systems, pages 3239–3249, 2017. 2
  40. 40.Suman Ravuri and Oriol Vinyals. Classification accuracy score for conditional generative models. arXiv preprint arXiv:1905.10887, 2019. 2
  41. 41.Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to imagenet? arXiv preprint arXiv:1902.10811, 2019. 1, 8
  42. 42.Ikuro Sato, Hiroki Nishimura, and Kensuke Yokoi. Apac: Augmented pattern classification with neural networks. arXiv preprint arXiv:1505.03229, 2015. 2
  43. 43.Patrice Y Simard, David Steinkraus, John C Platt, et al. Best practices for convolutional neural networks applied to visual document analysis. In Proceedings of International Conference on Document Analysis and Recognition, 2003. 1, 2, 8
  44. 44.Leon Sixt, Benjamin Wild, and Tim Landgraf. Rendergan: Generating realistic labeled data. arXiv preprint arXiv:1611.01331, 2016. 2
  45. 45.Jasper Snoek, Hugo Larochelle, and Ryan P Adams. Practical bayesian optimization of machine learning algorithms. In Advances in neural information processing systems, pages 2951–2959, 2012. 4
  46. 46.Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 2
  47. 47.Mingxing Tan and Quoc V Le. Efficientnet: Rethinking model scaling for convolutional neural networks. arXiv preprint arXiv:1905.11946, 2019. 1, 6, 7, 13
  48. 48.Toan Tran, Trung Pham, Gustavo Carneiro, Lyle Palmer, and Ian Reid. A bayesian data augmentation approach for learning deep models. In Advances in Neural Information Processing Systems, pages 2794–2803, 2017. 2
  49. 49.Li Wan, Matthew Zeiler, Sixin Zhang, Yann Le Cun, and Rob Fergus. Regularization of neural networks using dropconnect. In International Conference on Machine Learning, pages 1058–1066, 2013. 2
  50. 50.Qizhe Xie, Zihang Dai, Eduard Hovy, Minh-Thang Luong, and Quoc V Le. Unsupervised data augmentation. arXiv preprint arXiv:1904.12848, 2019. 1, 6, 8
  51. 51.Yoshihiro Yamada, Masakazu Iwamura, and Koichi Kise. Shakedrop regularization. arXiv preprint arXiv:1802.02375, 2018. 4, 13
  52. 52.Dong Yin, Raphael Gontijo Lopes, Jonathon Shlens, Ekin D Cubuk, and Justin Gilmer. A fourier perspective on model robustness in computer vision. arXiv preprint arXiv:1906.08988, 2019. 1, 2, 8
  53. 53.Sergey Zagoruyko and Nikos Komodakis. Wide residual networks. In British Machine Vision Conference, 2016. 1, 2, 4, 5, 7
  54. 54.Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412, 2017. 1, 2, 8
  55. 55.Zhun Zhong, Liang Zheng, Guoliang Kang, Shaozi Li, and Yi Yang. Random erasing data augmentation. arXiv preprint arXiv:1708.04896, 2017. 2
  56. 56.Xinyue Zhu, Yifan Liu, Zengchang Qin, and Jiahong Li. Data augmentation in emotion classification using generative adversarial networks. arXiv preprint arXiv:1711.00648, 2017. 2
  57. 57.Barret Zoph, Ekin D Cubuk, Golnaz Ghiasi, Tsung-Yi Lin, Jonathon Shlens, and Quoc V Le. Learning data augmentation strategies for object detection. arXiv preprint arXiv:1906.11172, 2019. 1, 4, 6, 7, 8, 13
  58. 58.Barret Zoph and Quoc V Le. Neural architecture search with reinforcement learning. In International Conference on Learning Representations, 2017. 2, 4
  59. 59.Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V Le. Learning transferable architectures for scalable image recognition. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2, 4, 6

Citation

MLA
Cubuk, E. D., et al. “Randaugment: Practical Automated Data Augmentation with a Reduced Search Space”. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2020, pp. 3008–17, https://doi.org/10.1109/CVPRW50498.2020.00359.
APA
Cubuk, E. D., Zoph, B., Shlens, J., & Le, Q. V. (2020). Randaugment: Practical automated data augmentation with a reduced search space. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 3008–3017. https://doi.org/10.1109/CVPRW50498.2020.00359
Chicago
Cubuk, E. D., B. Zoph, J. Shlens, and Q. V. Le. 2020. “Randaugment: Practical Automated Data Augmentation with a Reduced Search Space”. 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 3008–17. https://doi.org/10.1109/CVPRW50498.2020.00359.
Harvard
Cubuk, E.D. et al. (2020) “Randaugment: Practical automated data augmentation with a reduced search space”, 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, pp. 3008–3017. Available at: https://doi.org/10.1109/CVPRW50498.2020.00359.
Vancouver
1. Cubuk ED, Zoph B, Shlens J, Le QV (2020) Randaugment: Practical automated data augmentation with a reduced search space. In: 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). IEEE, pp 3008–3017

BibTeX

@inproceedings{Cubuk_2020, title={Randaugment: Practical automated data augmentation with a reduced search space}, url={http://dx.doi.org/10.1109/CVPRW50498.2020.00359}, DOI={10.1109/cvprw50498.2020.00359}, booktitle={2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)}, publisher={IEEE}, author={Cubuk, Ekin D. and Zoph, Barret and Shlens, Jonathon and Le, Quoc V.}, year={2020}, month=June, pages={3008–3017} }
Metadata:Crossref

Source Code

This paper has an official code repository available. Click below to access the source code.

View Repository

Access the Paper

This paper is available from its original source. Click below to access the PDF.

Open PDF

License: Published with permission