US 12,393,823 B2
Data processing method for neural network accelerator, device and storage medium
Chao Tian, Beijing (CN); Lei Jia, Beijing (CN); Junhui Wen, Beijing (CN); and Qiang Li, Beijing (CN)
Assigned to BEIJING BAIDU NETCOM SCIENCE TECHNOLOGY CO., LTD., Beijing (CN)
Filed by BEIJING BAIDU NETCOM SCIENCE TECHNOLOGY CO., LTD., Beijing (CN)
Filed on Nov. 15, 2021, as Appl. No. 17/526,755.
Claims priority of application No. 202011566189.2 (CN), filed on Dec. 25, 2020.
Prior Publication US 2022/0138528 A1, May 5, 2022
Int. Cl. G06N 3/04 (2023.01); G06F 9/50 (2006.01); G06F 17/15 (2006.01); G06N 3/045 (2023.01); G06N 3/0464 (2023.01); G06N 3/049 (2023.01); G06N 3/063 (2023.01); G06N 3/08 (2023.01)
CPC G06N 3/04 (2013.01) [G06F 9/5027 (2013.01); G06F 17/15 (2013.01); G06N 3/045 (2023.01); G06N 3/0464 (2023.01); G06N 3/049 (2013.01); G06N 3/063 (2013.01); G06N 3/08 (2013.01)] 17 Claims
OG exemplary drawing
 
1. A data processing method for a neural network accelerator, comprising:
obtaining data to be processed and a first operation to be executed, wherein the data to be processed is voice frequency domain data, in which the voice frequency domain data is data obtained by frequency domain conversion of voice data;
obtaining a real-number full-connection operation corresponding to the first operation; and
performing the real-number full-connection operation on the data based on a real-number full-connection unit of the neural network accelerator to obtain a result of the first operation for the data, so that hardware logics of the neural network accelerator is minimized;
wherein the first operation is a complex convolution operation; and
wherein obtaining the real-number full-connection operation corresponding to the first operation comprises:
obtaining a complex weight matrix corresponding to the complex convolution operation;
splitting the complex weight matrix to obtain a real part weight matrix and an imaginary part weight matrix;
determining a real-part full-connection operation based on the real part weight matrix, and determining an imaginary-part full-connection operation based on the imaginary part weight matrix; and
combining the real-part full-connection operation and the imaginary-part full-connection operation to obtain the real-number full-connection operation.