Evolutionary Approach for AutoAugment Using the Thermodynamical Genetic Algorithm

Authors: Akira Terauchi, Naoki Mori9851-9858

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

Reproducibility Variable Result LLM Response
Research Type Experimental To confirm the effectiveness of the proposed method, computational experiments were conducted using two benchmark datasets, CIFAR-10 and SVHN, as examples. The experimental results show that the proposed method can obtain various useful augmentation sub-policies for the problems while reducing the computational cost.
Researcher Affiliation Academia Akira Terauchi, Naoki Mori Osaka Prefecture University, 1-1 Gakuen-cho, Naka-ku, Sakai, Osaka, Japan terauchi@ss.cs.osakafu-u.ac.jp, mori@cs.osakafu-u.ac.jp
Pseudocode Yes The pseudocode of the TDGA algorithm is provided in Algorithm 1. [...] The overall procedure is summarized in Algorithm 2.
Open Source Code No The paper does not include an unambiguous statement that the source code for the described methodology is released, nor does it provide a direct link to a code repository for it.
Open Datasets Yes To investigate the performance of TDGA Auto Augment, we applied the method to CIFAR-10 and SVHN (Netzer et al. 2011) datasets and compared the results with those of the baseline and previous methods. [...] CIFAR-10 is a benchmark dataset for image classification that has been studied for a long time, and includes 50,000 training images and 10,000 test images.
Dataset Splits Yes In TDGA Auto Augment, the training sample is divided into a training dataset Dtrain and an evaluation dataset Dvalid by stratified shuffling (Shahrokh Esfahani and Dougherty 2013). [...] In the search process, 90% of the training data were used for Dtrain, and the remaining 10% were used for Dvalid.
Hardware Specification Yes Table 3: Comparison of the number of GPU hours required by TDGA Auto Augment with those required by other methods. GPU hours GPU AA 5000 Tesla P100 PBA 5 Titan XP Fast AA 3.5 Tesla V100 TDGA AA 2.5 RTX 2080Ti
Software Dependencies No The paper does not provide specific software dependencies with version numbers.
Experiment Setup Yes Pre-training was conducted for 200 epochs. [...] In the GA, the final generation was set to 30 (Ng = 30), the mutation rate of each locus was set to 0.06, and uniform crossover was used [...] For the models, we used a learning rate of 0.1, weight decay of 5e-4, batch size of 128, and cosine learning rate decay for 200 epochs. We set m = 2 for the experiments. [...] We set Np = 64, M = 5, m = 2, and T = 0.02.