Login| Sign Up| Help| Contact|

Patent Searching and Data


Title:
NEURAL NETWORK JOINT COMPILATION METHOD, APPARATUS AND ELECTRONIC DEVICE
Document Type and Number:
WIPO Patent Application WO/2021/098509
Kind Code:
A1
Abstract:
A neural network joint compilation method, an apparatus and an electronic device. The method comprises: acquiring multiple neural network models to be compiled and computation graphs of said multiple neural network models (S302); determining model combination schemes for said multiple neural network models, and jointly compiling computation graphs of neural network models in each model combination scheme to obtain an optimized computation graph corresponding to each model combination scheme (S304); and determining among the optimized computation graphs a target optimized computation graph, and generating a target instruction according to the target optimized computation graph (S306). When the target instruction obtained by means of the compilation of the described method is run on target hardware, it can be ensured that the performance of the hardware is fully used, and the efficient running of the target instruction can further be ensured.

Inventors:
ZHOU SHUCHANG (CN)
WANG TIAN (CN)
Application Number:
PCT/CN2020/126424
Publication Date:
May 27, 2021
Filing Date:
November 04, 2020
Export Citation:
Click for automatic bibliography generation   Help
Assignee:
MEGVII BEIJING TECH CO LTD (CN)
International Classes:
G06N3/063
Foreign References:
CN110908667A2020-03-24
CN109272109A2019-01-25
CN110321999A2019-10-11
CN109496294A2019-03-19
US20020091748A12002-07-11
Attorney, Agent or Firm:
BEIJING RUN ZEHENG INTELLECTUAL PROPERTY LAW FIRM (CN)
Download PDF: