Title:
PRUNING- AND DISTILLATION-BASED CONVOLUTIONAL NEURAL NETWORK COMPRESSION METHOD
Document Type and Number:
WIPO Patent Application WO/2018/223822
Kind Code:
A1
Abstract:
A pruning- and distillation-based convolutional neural network compression method (400), comprising: pruning an original convolutional neural network model to obtain a pruned model (S401); performing fine adjustment on parameters of the pruned model (S403); using the original convolutional neural network model as a teacher network of a distillation algorithm, using the pruned model with the parameters having experienced fine adjustment as a student network of the distillation algorithm, and instructing the student network in training by the teacher network according to the distillation algorithm (S405); and using the student network trained according to the distillation algorithm as a compressed convolutional neural network model (S407). According to the method, by using two conventional network compression methods in combination, a convolutional neural network model is more effectively compressed.
Inventors:
JIANG FAN (CN)
SHAN YI (CN)
SHAN YI (CN)
Application Number:
PCT/CN2018/087063
Publication Date:
December 13, 2018
Filing Date:
May 16, 2018
Export Citation:
Assignee:
BEIJING DEEPHI INTELLIGENT TECH CO LTD (CN)
International Classes:
G06N3/04
Domestic Patent References:
WO2010042256A2 | 2010-04-15 |
Foreign References:
CN106779068A | 2017-05-31 | |||
CN105894847A | 2016-08-24 | |||
CN106355248A | 2017-01-25 |
Attorney, Agent or Firm:
IPFAITH PARTNERS (CN)
Download PDF: