TensorFlow架构与实现机制的研究

Research on Tensor Flow Architecture and Mechanism

在大数据时代,云计算和大规模并行处理基础架构的共同发展不仅使得机器学习和深度人工智能有了更为广阔的应用空间,也激发了人工智能框架的快速迭代和部署。TensorFlow是Google发布的开放源代码的深度学习平台,已经在工业界有了广泛的应用。文中从TensorFlow平台的设计理念出发,分析了平台的框架和基本结构,对每个模块的功能和应用做了详尽阐述。在此基础上,通过建立一个多层深度学习神经网络,分析了输入层、隐藏层、输出层及激励函数的构建方法。最后在对TensorFlow实例运行和调试的基础上,演示了通过TensorBoard跟踪程序运行状态和参数调制的方法,给出了一维数据和多维数据的可视化结果。研究表明,相比较其他学术界的人工智能平台,TensorFlow有着更好的生态系统,支持更多的硬件架构,具备了一定的实用基础。

In the era of big data,the common development of cloud computing and large-scale parallel processing infrastructure not only makes machine learning and deep artificial intelligence have broader application space,but also stimulates the rapid iteration and deployment of artificial intelligence framework. TensorFlow is an open source deep learning platform released by Google that has been widely used in the industry. Based on the design concept and mechanism of TensorFlow,we analyze the framework and basic structure of the platform and elaborate the functions and applications of each module. By building a multiple layer deep learning program,we investigate the way to construct input layer,hidden layer,output layer,and activation function. Finally,on the basis of running and debugging the TensorFlow instance,the method of tracking the running state and parameter modulation by TensorBoard is demonstrated,and the visualization of one-dimensional data and multidimensional data are presented. Research shows that compared with other academic artificial intelligence platforms,TensorFlow has a better ecosystem,supports more hardware architectures and has a certain practical basis.