Title:
METHOD FOR CALIBRATING NEURAL NETWORK QUANTIZATION, APPARATUS, DEVICE, MEDIUM, AND PROGRAM PRODUCT
Document Type and Number:
WIPO Patent Application WO/2024/082932
Kind Code:
A1
Abstract:
The present disclosure relates to the field of computers. Provided in the embodiments of the present disclosure are a method for calibrating neural network quantization, a device, an apparatus, a medium, and a program product. The method comprises: executing conversion between a first precision and a second precision on a plurality of layers of a neural network, so as to obtain a first reasoning result, the neural network being trained at the first precision, the second precision comprising a tapered floating point precision, and the first precision being higher than the second precision; if the first reasoning result does not satisfy a predetermined condition, executing calibrated conversion between the first precision and the second precision on the plurality of layers, so as to obtain a second reasoning result; and if the second reasoning result does not satisfy the predetermined condition, identifying one or more target layers among the plurality of layers so as to maintain the first precision or another precision higher than the second precision. Thus, the impact of low precisions on reasoning results of networks can be reduced by means of calibrated conversion. In addition, utilizing multiple rounds of calibration to identify target layers which are not suitable for executing quantization improves computing efficiency and also enables network reasoning results to satisfy predetermined conditions.
Inventors:
LUO YUANYONG (CN)
CHEN GUANFU (CN)
ZHENG KAI (CN)
ZHANG ZHONGXING (CN)
WU WEI HSIANG (CN)
HE ZHANYING (CN)
HUANG ZEYI (CN)
CHEN GUANFU (CN)
ZHENG KAI (CN)
ZHANG ZHONGXING (CN)
WU WEI HSIANG (CN)
HE ZHANYING (CN)
HUANG ZEYI (CN)
Application Number:
PCT/CN2023/121225
Publication Date:
April 25, 2024
Filing Date:
September 25, 2023
Export Citation:
Assignee:
HUAWEI TECH CO LTD (CN)
International Classes:
G06N3/08
Foreign References:
| CN113196305A | 2021-07-30 | |||
| CN111831355A | 2020-10-27 | |||
| CN113971457A | 2022-01-25 | |||
| US20200210840A1 | 2020-07-02 |
Attorney, Agent or Firm:
KING & WOOD MALLESONS (CN)
Download PDF: