Title:
METHOD, APPARATUS, ELECTRONIC DEVICE, AND COMPUTER STORAGE MEDIUM FOR OPTIMIZING QUANTIZATION MODEL
Document Type and Number:
WIPO Patent Application WO/2020/228655
Kind Code:
A1
Abstract:
Provided are a method, apparatus, electronic device, and computer storage medium for optimizing a quantization model, relating to the technical field of deep learning. The method comprises: determining the jump ratios of parameters of the embedding layer of a trained quantization model within a predetermined time range, the quantization model being a neural network model in which the parameters of the embedding layer have been quantized (S110); then, according to each jump ratio, determining a jump curve within a predetermined time range, and fitting the jump curve to obtain a corresponding temporal scaling parameter (S120); then, optimizing an original optimization algorithm of the quantization model on the basis of the temporal scaling parameter to obtain an optimized target optimization algorithm, and training the quantization model on the basis of the target optimization algorithm (S130). When optimizing the original optimization algorithm, the time parameters are fully considered, enabling rapid learning of new data features, improving online prediction ability and recommendation accuracy.
Inventors:
YUAN YI (CN)
MAO ZHICHENG (CN)
WANG YONGZHUANG (CN)
XU YUHUI (CN)
MAO ZHICHENG (CN)
WANG YONGZHUANG (CN)
XU YUHUI (CN)
Application Number:
PCT/CN2020/089543
Publication Date:
November 19, 2020
Filing Date:
May 11, 2020
Export Citation:
Assignee:
TENCENT TECH SHENZHEN CO LTD (CN)
International Classes:
G06F16/9535; G06N3/04
Foreign References:
| CN110096647A | 2019-08-06 | |||
| CN105786830A | 2016-07-20 | |||
| CN103605909A | 2014-02-26 | |||
| US20130198729A1 | 2013-08-01 | |||
| CN201910390616A | 2019-05-10 |
Other References:
See also references of EP 3968184A4
Attorney, Agent or Firm:
DEQI INTELLECTUAL PROPERTY LAW CORPORATION (CN)
Download PDF: