Title:
DATA FLOW METHOD AND APPARATUS FOR NEURAL NETWORK COMPUTING
Document Type and Number:
WIPO Patent Application WO/2023/093185
Kind Code:
A1
Abstract:
A data flow method and apparatus for neural network computing. The method comprising the following steps: step 1, initializing a life cycle of a variable in a computational graph, and initializing a time period starting from defining the variable to the end of use as the life cycle of the variable in the computational graph; step 2, defining a propagation rule for a variable flowing pass a node during use, defining a variable at a certain node of the calculation graph to be used, and generating a definition of the variable at a precursor node of said node, such that an input set of effective variables flowing pass the node contains the variable; the method can analyze, from the perspective of a global action domain, memory unit blocks required to be allocated by the nodes of each calculation sub-graph in a local circulation. By means of the present data flow method for neural network computing, the computing efficiency of the whole computational graph is improved, and the time cost is saved.
Inventors:
WANG HONGSHENG (CN)
CHEN GUANG (CN)
CHEN GUANG (CN)
Application Number:
PCT/CN2022/116171
Publication Date:
June 01, 2023
Filing Date:
August 31, 2022
Export Citation:
Assignee:
ZHEJIANG LAB (CN)
International Classes:
G06F9/50; G06N3/08
Foreign References:
| US20190279086A1 | 2019-09-12 | |||
| US20200151580A1 | 2020-05-14 | |||
| US20210174190A1 | 2021-06-10 | |||
| CN114746871A | 2022-07-12 | |||
| CN114461351A | 2022-05-10 | |||
| CN114186687A | 2022-03-15 |
Attorney, Agent or Firm:
BEIJING ZHILIN HENGYUAN INTELLECTUAL PROPERTY AGENCY CO., LTD. (CN)
Download PDF: