Source-linked AI summary
AutoAugment: Learning Augmentation Policies from Data
Ekin D. Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, Quoc V. Le
TL;DR
Manual augmentation strategies are difficult to design and often dataset-specific, motivating automated policy discovery. AutoAugment searches operation choices, ordering, probabilities, and magnitudes using validation accuracy as the objective. It achieves state-of-the-art results across several datasets, and learned policies transfer to other datasets, although the chosen search algorithm may be improvable.
Problem
Existing image-augmentation methods are manually designed, require expert knowledge and time, and often transfer poorly across datasets.
Method
AutoAugment uses reinforcement-learning-based discrete search to select ordered image operations and their application probabilities and magnitudes for a target dataset.
Results
AutoAugment achieves state-of-the-art accuracy across CIFAR-10, CIFAR-100, SVHN, and ImageNet, while learned policies transfer across datasets and architectures.
Takeaways & Limitations
Policies can be learned directly when resources permit or transferred when direct search is resource-intensive, with transferred policies improving results across varied datasets and models.
Takeaways & Limitations
The reinforcement-learning search algorithm is one of many possible choices, and alternative algorithms might improve the results.
Abstract
from arXiv · showhide
Data augmentation is an effective technique for improving the accuracy of modern image classifiers. However, current data augmentation implementations are manually designed. In this paper, we describe a simple procedure called AutoAugment to automatically search for improved data augmentation policies. In our implementation, we have designed a search space where a policy consists of many sub-policies, one of which is randomly chosen for each image in each mini-batch. A sub-policy consists of two operations, each operation being an image processing function such as translation, rotation, or shearing, and the probabilities and magnitudes with which the functions are applied. We use a search algorithm to find the best policy such that the neural network yields the highest validation accuracy on a target dataset. Our method achieves state-of-the-art accuracy on CIFAR-10, CIFAR-100, SVHN, and ImageNet (without additional data). On ImageNet, we attain a Top-1 accuracy of 83.5% which is 0.4% better than the previous record of 83.1%. On CIFAR-10, we achieve an error rate of 1.5%, which is 0.6% better than the previous state-of-the-art. Augmentation policies we find are transferable between datasets. The policy learned on ImageNet transfers well to achieve significant improvements on other datasets, such as Oxford Flowers, Caltech-101, Oxford-IIT Pets, FGVC Aircraft, and Stanford Cars.
1. Introduction
AutoAugment addresses the manual, dataset-specific design of image augmentation by searching for policies whose operations, probabilities, and magnitudes improve validation accuracy. It achieves strong results through direct policy search and transfer across datasets.
- Data augmentation increases data amount and diversity while teaching models domain invariances such as insensitivity to translation or flipping.
- Manual augmentation strategies require expert knowledge, remain dataset-specific, and may transfer poorly because datasets contain different symmetries.
- AutoAugment searches choices and orders of image-processing operations together with their application probabilities and magnitudes to maximize validation accuracy.
- AutoAugment supports both direct policy discovery on a target dataset and transfer of learned policies to new datasets.
- 83.5% Top-1 accuracy on ImageNet surpassed the previous 83.1% record by 0.4%, while CIFAR-10 reached a 1.5% error rate.
2. Related Work
Prior augmentation methods were largely manually designed or generated data directly, motivating AutoAugment’s data-driven search for symbolic transformation policies.
- Common augmentation strategies are manually selected for particular datasets, requiring expert knowledge and time.
- Architecture-search methods discover model architectures from data, but architecture search alone had not broken the 2% CIFAR-10 error-rate barrier.
- Earlier learned-augmentation approaches generated merged samples, modeled training-set distributions, or transformed data in feature space.
- Unlike generative models that produce augmented data directly, AutoAugment generates symbolic image-transformation operations.
3. AutoAugment: Searching for best Augmentation policies Directly on the Dataset of Interest
AutoAugment formulates augmentation-policy discovery as a discrete search over ordered operations and their probabilities and magnitudes. A controller samples policies, child networks provide validation-accuracy rewards, and reinforcement learning updates the controller.
- AutoAugment treats finding the best augmentation policy as a discrete search problem with a search algorithm and a defined search space.
- Each policy contains five sub-policies, each specifying two sequential image operations with an application probability and magnitude.
- For each image, a sub-policy is selected uniformly at random, while each operation may independently be skipped according to its probability.
- The search space includes 16 operations, including geometric, color, contrast, Cutout, and Sample Pairing transformations; identity is represented by zero probability.
- A recurrent controller predicts policies, and Proximal Policy Optimization trains it using rewards based on a child model’s validation generalization.
- The authors note that alternative search algorithms, including genetic programming or random search, might improve the reported results.
4. Experiments and Results
AutoAugment improves image-classification results through direct policy search and transfer of learned policies across datasets. Experiments report state-of-the-art or significant gains on CIFAR, SVHN, ImageNet, and FGVC datasets.
- Direct application: AutoAugment achieves state-of-the-art results across CIFAR-10, CIFAR-100, SVHN, and ImageNet in direct policy-search experiments.The method is evaluated against competitive baseline models and uses no additional data for ImageNet.
- Direct application: 1.5% error on CIFAR-10 is 0.6% better than the previous state-of-the-art.This result is obtained with the ShakeDrop model.
- Direct application: The same policy achieves a 4.4% error rate on the newer CIFAR-10 test set, versus 7.0% for Shake-Shake with Cutout and 7.7% for PyramidNet+ShakeDrop.Its error increase relative to the original test set is 2.9%, smaller than the increases reported for the comparison models.
- Policy reuse: The policy found on reduced CIFAR-10 is reused for CIFAR-100, where AutoAugment beats the previous 12.19% error-rate record.The same policy is also applied to CIFAR-10 and reduced CIFAR-10.
- Transferability: AutoAugment-transfer improves generalization accuracy on all evaluated FGVC datasets, including a lowest-reported error rate on Stanford Cars with weights trained from scratch.The transferred policy is learned on ImageNet and applied to the target datasets.
- Direct application: 83.5% / 96.5% Top-1 / Top-5 accuracy on ImageNet improves AmoebaNet-C from 83.1% / 96.1%.The policy was discovered using 5,000 images, and the result is reported without multicrop or ensembling.
5. Discussion
The discussion compares AutoAugment with automated augmentation methods, examines training and policy-design choices, and finds benefits from learned policies, transfer, and selected parameters.
- AutoAugment optimizes classification accuracy directly, whereas the compared method optimizes whether augmented images resemble current training images.
- AutoAugment produced approximately 3.0% higher improvement than the compared method for both ResNet-32 and ResNet-56.
- 120 training epochs are used for child models because each sub-policy must be applied repeatedly before the model benefits from all five sub-policies.
- Policies transferred well across model architectures and datasets, but policies learned from distributions closest to the target performed best.
- 3.0% average error was obtained when probabilities and magnitudes were randomized, compared with the original policy's 2.6% ± 0.1% error rate.
- Random policies improved over baseline augmentation, but learned probabilities and magnitudes were important because randomizing them reduced effectiveness.
A. Supplementary materials for “AutoAugment: Learning Augmentation policies from Data”
The supplementary materials list the transformation search space and provide learned AutoAugment policies for reduced CIFAR-10, reduced SVHN, and reduced ImageNet.
- The controller's search space includes image transformations with operation-specific magnitude values, while some operations do not use magnitude information.
- The supplementary materials include the AutoAugment policy found on reduced CIFAR-10.
- The supplementary materials include the AutoAugment policy found on reduced SVHN.
- The supplementary materials include the AutoAugment policy found on reduced ImageNet.