Feasibility Experiment and Discussion of Cyclegan Generating Data for Retraining Under Unsupervised Scenarios
Download as PDF
DOI: 10.23977/iset.2019.025
Corresponding Author
Yongjian Chen
ABSTRACT
The main contribution of this paper is a simple unsupervised pipeline that uses the training set that only has rare samples. It tries to answer this question: does it improve the accuracy of the classifier using the cyclegan generating data to expand the training set? In this work, We first selected a portion of the data from the CelebA data set as our raw data set. Using cyclegan to train some specific data transformation models on this dataset. We use these data transformation models to augment our dataset and then test it on a mobienetv2 classifier. The experimental results show that cyclegan used to augment the data set did not achieve good results. Even in some cases, it will be counterproductive. Of course, the results of this experiment are only for the conclusion of this experiment, and may be affected by insufficient experimental data and the effect of the gan model. We also discussed it later in this article.
KEYWORDS
Cyclegan, Retraining, Unsupervise