深度学习计算图优化综述

With the development of deep learning, the scale of deep learning models has become larger and more complex, and how to speed up the computation of deep learning models has attracted the attention of researchers. In this thesis, we first introduce the deep learning computational graph, and then introduce the optimization methods based on the computational graph from the research results in recent years, including device placement optimization, optimization based on graph replacement and related joint optimization, and introduce the results of related research according to the algorithms, including traditional algorithms such as backtracking and dynamic programming, as well as algorithms based on deep reinforcement learning.