AutoTrain: An Efficient Auto-training System for Small-scale Image Classification
Shaoxiao Yi, Junwei Zou, Wei Zheng Ren, Hong Yan Luo · 2020
Machine learning has become the most promising research field. However, the involved models usually require complex, tedious and expensive manual intervention. The automated machine learning technology plays a significant role in mitigating this issue. However, the current studies ignore the importance of automation in data preprocessing. In this paper, we propose an efficient automatic training system, AutoTrain, to solve small-scale image classification problems. First, we design sample equalization in data augmentation to improve the performance of training on uneven data. Then, a Bayesian optimization-based strategy controller is introduced to rapidly find the strategy applied in data augmentation. Additionally, we present a dynamic adjustment model to fit tasks with different scales and complexity. Finally, experimental results show that the AutoTrain's training speed is about 3 times faster on average than the conventional methods. And the avergae accuracy of AutoTrain has 2% improved to the conventional methods.