SelectAugment: Hierarchical Deterministic Sample Selection for Data Augmentation

Authors: Shiqi Lin, Zhizheng Zhang, Xin Li, Zhibo Chen

AAAI 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate that our proposed Select Augment significantly improves various off-the-shelf DA methods on image classification and fine-grained image recognition.
Researcher Affiliation Collaboration 1 University of Science and Technology of China 2 Microsoft Research Asia
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code No The paper does not provide concrete access to source code for the methodology described.
Open Datasets Yes For the target network, following (Cubuk et al. 2018; Lim et al. 2019; Lin et al. 2021; Yun et al. 2019), for CIFAR-10 and CIFAR-100, we respectively use Wide-Res Net-28-10 (WRN) (Zagoruyko and Komodakis 2016), Shake Shake(26 2 32d) and Shake Shake (26 2 96d) (Gastaldi 2017) (aliased as SS(32) and SS(96)) as target network. For Image Net (Deng et al. 2009), Res Net50 and Res Net-200 (He et al. 2016) are adopted as target models, which are trained from scratch. For fine-grained image classification, we follow the previous works (Du et al. 2020; Chen et al. 2019b) and utilize pre-trained Res Net-50 and Res Net-101 models as the target network. Unless specified otherwise, the input image size is 32 32 for CIFAR while 224 224 for Image Net, CUB-200-2011 and Stanford Dogs.
Dataset Splits Yes Validation set Top-1 and Top-5 accuracy (%) on Image Net." and "The hyperparameters for target networks, such as training epochs and the learning rate of target network are the same as previous works (Yun et al. 2019; Zhang et al. 2018; Cubuk et al. 2018) for fair comparison.
Hardware Specification No The paper mentions '1.4 GPU hours increase in the training time' but does not specify the type or model of GPU used for the experiments, nor does it provide other specific hardware details.
Software Dependencies No The paper does not provide specific software details like library names with version numbers.
Experiment Setup Yes Unless specified otherwise, the input image size is 32 32 for CIFAR while 224 224 for Image Net, CUB-200-2011 and Stanford Dogs. We set the batch size to 128 for experiments on CIFAR and 1024 for experiments on Image Net as well as two fine-grained datasets.