US 12,326,915 B2
Method and apparatus of training model, device, medium, and program product
Ruoyu Guo, Beijing (CN); Yuning Du, Beijing (CN); Weiwei Liu, Beijing (CN); Xiaoting Yin, Beijing (CN); Qiao Zhao, Beijing (CN); Qiwen Liu, Beijing (CN); Ran Bi, Beijing (CN); Xiaoguang Hu, Beijing (CN); Dianhai Yu, Beijing (CN); and Yanjun Ma, Beijing (CN)
Assigned to Beijing Baidu Netcom Science Technology Co., Ltd., Beijing (CN)
Filed by Beijing Baidu Netcom Science Technology Co., Ltd., Beijing (CN)
Filed on Sep. 20, 2021, as Appl. No. 17/479,061.
Claims priority of application No. 202110321597.X (CN), filed on Mar. 25, 2021.
Prior Publication US 2022/0004811 A1, Jan. 6, 2022
Int. Cl. G06F 18/214 (2023.01); G06F 18/21 (2023.01); G06F 18/24 (2023.01)
CPC G06F 18/214 (2023.01) [G06F 18/217 (2023.01); G06F 18/24 (2023.01)] 17 Claims
OG exemplary drawing
 
1. A method of training a model, comprising:
determining a plurality of augmented sample sets associated with a plurality of original samples;
determining a first constraint according to a first model based on the plurality of augmented sample sets, wherein the first constraint is associated with a difference between outputs of the first model for different augmented samples in one augmented sample set of the plurality of augmented sample sets;
determining a second constraint according to the first model and a second model based on the plurality of augmented sample sets, wherein the second constraint is associated with a difference between an output of the first model and an output of the second model for one augmented sample in the plurality of augmented sample sets, and the first model has a complexity lower than that of the second model; and
training the first model based on at least the first constraint and the second constraint, so as to obtain a trained first model, wherein
the determining a first constraint according to a first model based on the plurality of augmented sample sets includes:
generating a first classification result according to the first model based on a first augmented sample in the one augmented sample set;
generating a second classification result according to the first model based on a second augmented sample in the one augmented sample set;
generating a third classification result according to the first model based on a third augmented sample in the one augmented sample set; and
calculating the first constraint as an average value of L2 loss between every two of the first classification result, the second classification result, and the third classification result.