US 11,853,760 B2
Model conversion method, device, computer equipment, and storage medium
Shaoli Liu, Beijing (CN); Jun Liang, Beijing (CN); and Qi Guo, Beijing (CN)
Assigned to CAMBRICON TECHNOLOGIES CORPORATION LIMITED, Beijing (CN)
Filed by Cambricon Technologies Corporation Limited, Beijing (CN)
Filed on Mar. 24, 2022, as Appl. No. 17/703,757.
Application 17/703,757 is a continuation of application No. 16/667,593, filed on Oct. 29, 2019, granted, now 11,314,507.
Application 16/667,593 is a continuation of application No. PCT/CN2019/080510, filed on Mar. 29, 2019.
Claims priority of application No. 201810913895.6 (CN), filed on Aug. 10, 2018.
Prior Publication US 2022/0214875 A1, Jul. 7, 2022
This patent is subject to a terminal disclaimer.
Int. Cl. G06F 9/30 (2018.01); G06N 20/00 (2019.01); G06F 9/445 (2018.01); G06N 3/08 (2023.01); G06F 8/35 (2018.01)
CPC G06F 9/3004 (2013.01) [G06F 8/35 (2013.01); G06F 9/44505 (2013.01); G06N 3/08 (2013.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A model conversion method, the method comprising:
obtaining, by an input/output (I/O) interface of a computer system, an initial offline model, wherein the initial offline model is obtained by compiling an original neural network, and the initial offline model includes network weights and instructions of converted from respective compute nodes in the original neural network;
determining, by a processor of the computer system, whether model attributes of the initial offline model match a plurality of hardware attributes of the computer system; and
converting, by the processor, the network weights and the instructions included in the initial offline model to generate a target offline model based on the determination that the model attributes of the initial offline model do not match the plurality of hardware attributes of the computer system; wherein the processor is capable of executing the target offline model to implement a corresponding artificial intelligence application.