Abstract:
The optimization of computation graphs for neural network models is a crucial task in the field of deep learning. In this paper, we propose a method based on TVM to optimize the computation graph of neural network models. Our approach involves analyzing and optimizing the computation graph by adjusting the node order and merging operations.
We first introduce the background and purpose of our research, highlighting the importance and significance of optimizing computation graphs for neural network models. We then describe the methods and techniques used in our study, including TVM and neural network model optimization.
Our experimental results demonstrate that our method can significantly improve the efficiency of neural network models. Specifically, we achieved a 2x speedup on average across various neural network models. Furthermore, our method is highly practical and scalable, making it suitable for use in real-world applications.
In conclusion, our research provides a promising solution for optimizing computation graphs of neural network models. Future work will focus on further exploring the potential of TVM-based optimization methods and applying them to more complex neural network models.