Title:
CLASSIFICATION NETWORK, AND METHOD AND APPARATUS FOR IMPLEMENTING SAME
Document Type and Number:
WIPO Patent Application WO/2022/160856
Kind Code:
A1
Abstract:
A classification network, and a method and apparatus for implementing same. The method comprises: constructing, on the basis of an IRv2 algorithm, a classification network comprising a Stem module, multiple Inception modules and multiple IR modules, a batch normalization (BN) layer of the classification network using a γ parameter; initializing the classification network, and performing sparsification training on an initialized classification network to obtain a sparsified γ parameter; and on the basis of the sparsified γ parameter, performing network slimming pruning on the Stem module, a pruned Inception module and the multiple IR modules, and not performing network slimming pruning on a non-pruned Inception module. The classification effect of the classification network can be ensured, and the structure of the network is simplified, so that the training efficiency is improved, the inference time and the memory occupation are reduced, and the labor, the working time, and costs are reduced, and thus the classification network can be deployed on hardware devices having different hashrates.
Inventors:
FENG PENGBO (CN)
ZHANG YIFAN (CN)
LIU JIE (CN)
ZHANG YIFAN (CN)
LIU JIE (CN)
Application Number:
PCT/CN2021/129578
Publication Date:
August 04, 2022
Filing Date:
November 09, 2021
Export Citation:
Assignee:
GOERTEK INC (CN)
International Classes:
G06K9/62
Foreign References:
| CN112766397A | 2021-05-07 | |||
| CN111612144A | 2020-09-01 | |||
| CN111832705A | 2020-10-27 | |||
| CN111401523A | 2020-07-10 |
Download PDF: