DHA: End-to-End Joint Optimization of Data Augmentation Policy, Hyper-parameter and Architecture

Authors: kaichen zhou, Lanqing HONG, Shoukang Hu, Fengwei Zhou, Binxin Ru, Jiashi Feng, Zhenguo Li

TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that DHA achieves state-of-the-art (SOTA) results on various datasets and search spaces. Experiments show that DHA achieves a state-of-the-art (SOTA) accuracy on Image Net with both cell-based and Mobilenet-like architecture search space. DHA also provides SOTA results on various datasets, including CIFAR10, CIFAR100, SPORT8, MIT67, FLOWERS102 and Image Net with relatively low computational cost, showing the effectiveness and efficiency of joint optimization (see Fig. 2). The test accuracy and computational time of various Auto ML algorithms are summarized in Table 1 and Table 2.
Researcher Affiliation Collaboration Kaichen Zhou University of Oxford Lanqing Hong EMAIL Huawei Noah s Ark Lab Shoukang Hu The Chinese University of Hong Kong Fengwei Zhou Huawei Noah s Ark Lab Binxin Ru Sail Yond Technology & Research Institute of Tsinghua University in Shenzhen Jiashi Feng National University of Singapore Zhenguo Li Huawei Noah s Ark Lab
Pseudocode Yes Algorithm 1 DHA
Open Source Code No The text does not provide an explicit statement of code release for the described methodology nor a direct link to a source code repository.
Open Datasets Yes Datasets. Following Ru et al. (2020b), we conducted experiments on various datasets, including CIFAR10 and CIFAR100 for the object classification task (Krizhevsky et al., 2009), SPORT8 for the action classification task (Li & Fei-Fei, 2007), MIT67 for the scene classification task (Quattoni & Torralba, 2009), FLOWERS102 for the fine object classification task (Nilsback & Zisserman, 2008), and Image Net(Russakovsky et al., 2015) for the large-scale classification task.
Dataset Splits No The paper uses standard benchmark datasets but does not explicitly detail the exact percentages or counts for training, validation, and test splits used in their experiments. It mentions using a subset from the training set for hyperparameter updates without specifying its size or how the initial overall data split (train/test) is made.
Hardware Specification Yes Experiments are conducted on NVIDIA V100 under Py Torch-1.3.0 and Python 3.6.
Software Dependencies Yes Experiments are conducted on NVIDIA V100 under Py Torch-1.3.0 and Python 3.6.
Experiment Setup No The paper states that 'For hyper-parameters of DHA, we simply adopt the hyper-parameters used in previous works without many modifications, including Meta-Aug (Zhou et al., 2021), ISTA-NAS (Yang et al., 2020), and DSNAS (Hu et al., 2020).' and 'Detailed settings of baselines are provided in Appendix A.' However, specific numerical values for hyperparameters or training configurations are not explicitly provided in the main body of the paper.