US 12,444,179 B2
Model training method and model training system
Shih-Jen Chu, Taipei (TW)
Assigned to PEGATRON CORPORATION, Taipei (TW)
Filed by PEGATRON CORPORATION, Taipei (TW)
Filed on May 5, 2023, as Appl. No. 18/313,303.
Claims priority of application No. 111125804 (TW), filed on Jul. 8, 2022.
Prior Publication US 2024/0013523 A1, Jan. 11, 2024
Int. Cl. G06V 10/776 (2022.01); G06V 10/764 (2022.01); G06V 10/774 (2022.01)
CPC G06V 10/776 (2022.01) [G06V 10/774 (2022.01)] 14 Claims
OG exemplary drawing
 
1. A model training method, using a task filter in a model training system to improve training efficiency thereof, wherein the model training method comprises:
in a first iteration training, inputting first training data corresponding to a first sub-task and second training data corresponding to a second sub-task to a target model comprising a neural network;
evaluating and executing, by a processor using the task filter, a first image identification rate and a second image identification rate of the target model with respect to the first sub-task and the second sub-task respectively according to a first output corresponding to the first sub-task and a second output corresponding to the second sub-task from the target model; and
adjusting and executing, by the processor using the task filter, a first sampling rate and a second sampling rate respectively corresponding to the first training data and the second training data in a second iteration training according to the first image identification rate and the second image identification rate, wherein the first sampling rate is negatively correlated to the first image identification rate, and the second sampling rate is negatively correlated to the second image identification rate, performing, by the target model, repeated learning with respect to one of the first sub-task and the second sub-task to reduce a reconstruction error of the target model, thus improving the training efficiency of the model training system.