US 11,816,547 B2
Calibration method and apparatus, terminal device, and storage medium
Han Li, Beijing (CN); and Yaolong Zhu, Beijing (CN)
Assigned to LYNXI TECHNOLOGIES CO., LTD., Beijing (CN)
Appl. No. 18/004,021
Filed by LYNXI TECHNOLOGIES CO., LTD., Beijing (CN)
PCT Filed Jul. 23, 2021, PCT No. PCT/CN2021/108133
§ 371(c)(1), (2) Date Dec. 30, 2022,
PCT Pub. No. WO2022/022417, PCT Pub. Date Feb. 3, 2022.
Claims priority of application No. 202010747179.2 (CN), filed on Jul. 29, 2020.
Prior Publication US 2023/0196197 A1, Jun. 22, 2023
Int. Cl. G06N 20/00 (2019.01)
CPC G06N 20/00 (2019.01) 7 Claims
OG exemplary drawing
 
1. A calibration method for a neural network model, comprising:
determining layer attribute information of each to-be-calibrated layer in a model; and
determining a group to which each of the to-be-calibrated layers is assigned according to the total available resources and the layer attribute information of each of the to-be-calibrated layers;
wherein the layer attribute information of any of the to-be-calibrated layers comprises layer required resources, the layer required resources being resources needing to be occupied when the to-be-calibrated layer is calibrated; and the total available resources are the total resources used for calibration, wherein the resources are memory resources; and
wherein the determining a group to which each of the to-be-calibrated layers is assigned according to the total available resources and the layer attribute information of each of the to-be-calibrated layers comprises:
determining the to-be-calibrated layer with the greatest layer required resources as a target layer from the to-be-calibrated layers not in a group;
determining a group to which the target layer is assigned at least according to the layer attribute information of the target layer;
subtracting the layer required resources for the target layer from the group available resources of the group to which the target layer is assigned; and
if there are still to-be-calibrated layers not in a group, returning to the step of determining the to-be-calibrated layer with the greatest layer required resources as a target layer from the to-be-calibrated layers not in a group.