US 12,314,863 B1
Determining quantization scale factors for layers of a machine learning model
Shaoxia Fang, Beijing (CN); Jiangsha Ma, Beijing (CN); Xi Wang, Beijing (CN); Junbin Wang, Beijing (CN); Cheng Chen, Beijing (CN); and Taobo Wang, Beijing (CN)
Assigned to XILINX, INC., San Jose, CA (US)
Filed by Xilinx, Inc., San Jose, CA (US)
Filed on Apr. 12, 2021, as Appl. No. 17/227,492.
Int. Cl. G06N 3/086 (2023.01); G06N 3/126 (2023.01); G06N 20/00 (2019.01)
CPC G06N 3/086 (2013.01) [G06N 3/126 (2013.01); G06N 20/00 (2019.01)] 20 Claims
OG exemplary drawing
 
1. A method, comprising:
generating a population of chromosomes in a memory arrangement by a calibrator circuit, each chromosome having a plurality of genes, wherein each gene specifies a scale factor associated with a layer of a machine learning model;
evaluating the population of chromosomes by the calibrator circuit, the evaluating including for each chromosome in the population:
quantizing floating point weights and floating point values of a representative dataset using the scale factors of the plurality of genes of each chromosome to produce quantized weights and a quantized representative dataset in the memory arrangement,
initiating processing of the quantized representative dataset using the quantized weights by an accelerator circuit configured to perform operations of the machine learning model, and
gauging a level of accuracy of results produced by the processing of the quantized representative dataset and associating the level of accuracy with each chromosome; and
deploying the accelerator circuit with the machine learning model at an application runtime to process application data quantized according to the scale factors of a selected chromosome of the population having a greatest associated level of accuracy in response to satisfaction of a termination criteria.