Neural network optimization method, device and processor

• Main Authors: Liu, Xiao-Long, Li, Yu-Dong
• Format: Patent
• Language: Chinese; English
• Published: 11.07.2022
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses a neural network optimization method, device and processor. The present invention can select an operand to be replaced from a plurality of operands in a network layer according to a preset condition, and replace the operand to be replaced with a plurality of equivalent operands according to a calculation function corresponding to the operand to be replaced, wherein a target operand is included in the plurality of equivalent operands. A first operand among the plurality of equivalent operands is pre-calculated, and the calculation result is input into the target operand. A second operand is identified according to the data change status of the plurality of equivalent operands, and the second operand is combined with the target operand to complete the optimization of the neural network model. The present invention can perform lossless conversion on the operands in the neural network, and can further improve the calculation performance on the basis of simplifying the network structure.