NEURAL NETWORK SEGMENTATION METHOD, PREDICTION METHOD, AND RELATED APPARATUS

Title:

NEURAL NETWORK SEGMENTATION METHOD, PREDICTION METHOD, AND RELATED APPARATUS

Document Type and Number:

WIPO Patent Application WO/2021/012609

Kind Code:

A1

Abstract:

A neural network segmentation method, a prediction method, and a related apparatus, the neural network segmentation method comprising: acquiring a neural network graph, the neural network graph being used to characterize a neural network; performing segmentation of the neural network graph, so as to obtain a depth subgraph; data exchange between nodes among multiple nodes comprised by the depth subgraph being performed by means of on-chip buffer reading and writing, the depth subgraph being used to process at least two data sets obtained by splitting first input data, so as to obtain first output data, the first input data being input data of the depth subgraph. In the present method, a neural network segmentation apparatus segments a neural network graph, obtaining one or more depth subgraphs, so as to facilitate generating one or more depth sub-networks according to the one or more depth subgraphs; using these depth sub-networks to execute a neural network processing task can greatly reduce the number of times an external memory is accessed.

More Like This:

WO/2018/000639	METHOD AND SYSTEM FOR HANDLING LAG
JP6485343	Control programs, control devices, and control methods
WO/2016/045515	CLOUD TASK SCHEDULING ALGORITHM BASED ON USER SATISFACTION

Inventors:

CHU JIEYU (CN)
YE DEZHOU (CN)
HU YUZHOU (CN)

Application Number:

PCT/CN2019/128915

Publication Date:

January 28, 2021

Filing Date:

December 26, 2019

Export Citation:

Click for automatic bibliography generation Help

Assignee:

HUAWEI TECH CO LTD (CN)

International Classes:

G06F9/50

Domestic Patent References:

WO2018135515A1

2018-07-26

Foreign References:

CN108292241A	2018-07-17
CN108351805A	2018-07-31
US20190042917A1	2019-02-07

Other References: