A dynamic graph execution method and device for neural network calculation

Abstract

The invention discloses a dynamic graph execution method and device for neural network calculation, comprising the following steps: S1: constructing and distributing operators and tensors; S2: operator interpreter deriving the operator execution process; S3: operator The interpreter constructs the instructions of the runtime virtual machine; S4: the operator interpreter sends instructions to the runtime virtual machine; S5: the virtual machine schedules instructions; S6: the virtual machine releases the executed instructions. The present invention provides a dynamic graph execution method and device for neural network computing. By abstracting the runtime into a virtual machine, the virtual machine obtains the subgraph scheduling and execution of each step built by the user through the interpreter in real time. Each subgraph not only meets the needs of users for real-time debugging, but also can be locally tuned to obtain the optimal local model. It meets the needs of algorithm researchers to instantly verify the correctness of the algorithm and the local performance of the model in the process of developing the model.

Description

technical field [0001] The present invention relates to the field of computer systems based on specific computing models, in particular to a dynamic graph execution method and device for neural network computing. Background technique [0002] With the rapid development of the industrial application of artificial intelligence, the dynamic graph computing method for deep neural networks is still a hot issue explored by deep learning framework researchers in the process of developing algorithm models for researchers and engineering users. Most of the existing graph computing technologies perform runtime computing tasks based on an entire static graph, so the algorithm model must be completely constructed before debugging and tuning can be performed. The disadvantage of this is that in the process of building the model, researchers cannot meet the needs of real-time verification of the local subgraphs of the model, nor can they meet the needs of real-time adjustment of the model...

AI Technical Summary

Problems solved by technology

The disadvantage of this is that in the process of building the model, researchers cannot meet the needs of real-time verification of the local subgraph of the model, and can not meet the needs of real-time adjustment of the model structure. In order to solve the real-time debugging of the dynamic graph calculated by the neural network The present invention discloses a dynamic graph execution method and device for neural network calculation, and provides a dynamic graph execution method and device for neural network model calculation in a deep learning training system

Images