MetaAugment: Sample-Aware Data Augmentation Policy Learning

Authors: Fengwei Zhou, Jiawei Li, Chuanlong Xie, Fei Chen, Lanqing Hong, Rui Sun, Zhenguo Li11097-11105

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Superior performance is achieved on widely-used benchmarks including CIFAR-10/100, Omniglot, and Image Net. Experimental Results In this section, we evaluate Meta Augment for image recognition tasks on CIFAR-10/100 (Krizhevsky and Hinton 2009), Omniglot (Lake et al. 2011), and Image Net (Deng et al. 2009).
Researcher Affiliation Industry Huawei Noah s Ark Lab {zhoufengwei, li.jiawei, xie.chuanlong, chen.f, honglanqing, sun.rui3, li.zhenguo}@huawei.com
Pseudocode Yes Algorithm 1 Meta Augment: Sample-Aware Data Augmentation Policy Learning
Open Source Code No The paper states "For FAA and PBA, we do experiments with their open-source codes." but does not explicitly state that the code for Meta Augment is open-source or provide a link to it.
Open Datasets Yes Experimental Results In this section, we evaluate Meta Augment for image recognition tasks on CIFAR-10/100 (Krizhevsky and Hinton 2009), Omniglot (Lake et al. 2011), and Image Net (Deng et al. 2009).
Dataset Splits Yes CIFAR. CIFAR-10 and CIFAR-100 consist of 50,000 images for training and 10,000 images for testing. For our method, we hold out 1,000 training images as the validation data.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU models, CPU types, memory) used for running the experiments.
Software Dependencies No The paper does not provide specific software dependencies with version numbers (e.g., Python, PyTorch, CUDA versions).
Experiment Setup Yes The augmentation policy network is an MLP that takes the embedding of the transformation and the corresponding augmented image feature as inputs, each followed by a fully-connected layer of size 100 with Re LU nonlinearities. The two intermediate features are then concatenated together, followed by a fully-connected output layer of size 1. The Sigmoid function is applied to the output. We also normalize the output weights of training samples in each mini-batch, i.e., each weight is divided by the sum of all weights in the mini-batch. We use K = 14 image processing functions: Auto Contrast, Equalize, Rotate, Posterize, Solarize, Color, Contrast, Brightness, Sharpness, Shear X/Y, Translate X/Y, Identity. For Meta Augment, the transformation is applied after horizontal flipping, and then Cutout (De Vries and Taylor 2017) with 16 16 pixels is applied. For each iteration, a mini-batch of training data Dtr mi = {(xi, yi)}ntr i=1 with batch size ntr is sampled and for each xi, a transformation T m1,m2 ji,ki is sampled to augment xi. The outer loop updates of θ and α are formulated by (θ(t+1), α(t+1)) = (θ(t), α(t)) β i =1 (θ,α)Lval i ( ˆw(t)(θ(t), α(t))), where β is a learning rate.