Performance Evaluation and Analysis of Deep Learning Frameworks
Xiaoyan Xie, Wenjuan He, Yun Zhu, Hao Xu
Abstract
The rapid development of deep learning has contributed to the increasing number of open-source deep learning frameworks, and in practice, benchmarking deep learning frameworks to effectively understand the performance characteristics of these frameworks and make choices becomes a challenge. Based on this, this paper uses three types of neural networks (convolutional neural networks, recurrent neural networks, and vision transformer models) to conduct extensive experimental evaluation and analysis of three popular deep learning frameworks, TensorFlow, PyTorch, and PaddlePaddle. Experiments are mainly conducted in CPU and GPU environments using different datasets, and performance parameters such as accuracy, training time, inference time, hardware utilization and other non-performance factors are considered. Finally, the performance characteristics, advantages and disadvantages of different frameworks are analyzed based on the above indexes, which provides theoretical guidance for users to choose.