Title:
NETWORK ARCHITECTURE SEARCH METHOD AND APPARATUS, AND STORAGE MEDIUM AND COMPUTER PROGRAM PRODUCT
Document Type and Number:
WIPO Patent Application WO/2021/081809
Kind Code:
A1
Abstract:
Provided is a network architecture search method, comprising: a whole graph training step: training a first whole graph according to a first network architecture and training data to generate a second whole graph; and a network architecture training step: determining several test sub-graphs from the second whole graph according to the first network architecture; testing the several test sub-graphs by means of test data to generate a feedback result; determining a skip constraint term according to the feedback result; and updating the first network architecture according to the feedback result and the skip constraint term. During the training process of a neural network, the requirements for skip density are different at different training stages, and the skip constraint term in the method is determined on the basis of the feedback result in the current training stage. Therefore, the skip density in the method is related to the current training stage, such that the current skip density of a controller is more suitable for the current training stage, and the training efficiency can thus be improved, while a good training result is also obtained.
More Like This:
Inventors:
JIANG YANG (CN)
PANG LEI (CN)
HU ZHAN (CN)
PANG LEI (CN)
HU ZHAN (CN)
Application Number:
PCT/CN2019/114361
Publication Date:
May 06, 2021
Filing Date:
October 30, 2019
Export Citation:
Assignee:
SZ DJI TECHNOLOGY CO LTD (CN)
International Classes:
G06N3/08
Domestic Patent References:
| WO2019084560A1 | 2019-05-02 |
Foreign References:
| CN110009048A | 2019-07-12 | |||
| CN109934336A | 2019-06-25 | |||
| US20190026639A1 | 2019-01-24 |
Attorney, Agent or Firm:
LONGSUN LEAD IP LTD. (CN)
Download PDF: