Method and apparatus for detecting face, computer device and computer-readable storage medium

Patent metadata

Field	Value
Publication number	US-12136259-B2
Application number	US-202017780840-A
Country	US
Kind code	B2
Filing date	Aug 20, 2020
Priority date	Nov 29, 2019
Publication date	Nov 5, 2024
Grant date	Nov 5, 2024

Abstract

A method for training a neural network, including: determining a neural network; training the neural network at a first learning rate according to a first optimization mode, where the first learning rate is updated each time the neural network is trained; mapping the first learning rate of the first optimization mode to a second learning rate of a second optimization mode in the same vector space; determining the second learning rate satisfies a preset update condition; and continuing to train the neural network at the second learning rate according to the second optimization mode.

First claim

What is claimed is:

1. A method for detecting a face, comprising:
receiving image data; and
identifying a region in the image data where face data is located by inputting the image data into a preset neural network for processing, wherein the preset neural network is trained by a method for training a neural network and the method for training a neural network comprises:
determining a neural network;
training the neural network at a first learning rate according to a first optimization mode, wherein the first learning rate is updated each time the neural network is trained; |
mapping the first learning rate of the first optimization mode to a second learning rate of a second optimization mode in the same vector space;
determining the second learning rate satisfies a preset update condition; and
continuing to train the neural network at the second learning rate according to the second optimization mode; and
wherein the mapping the first learning rate of the first optimization mode to the second learning rate of the second optimization mode in the same vector space comprises:
determining an update range, wherein the update range represents a range for updating a first network parameter in a case that the neural network is trained at the first learning rate according to the first optimization mode, and the first network parameter represents a parameter of the neural network in a case that the neural network is trained at the first learning rate according to the first optimization mode;
determining a parameter gradient of a second network parameter, wherein the second network parameter represents a parameter of the neural network in a case that the neural network is trained at the second learning rate according to the second optimization mode; and
determining a projection of the update range on the parameter gradient in the same vector space as the second learning rate of the second optimization mode.

2. The method according to claim 1 , wherein the determining the update range based on the first learning rate of the first optimization mode comprises:
determining a first-order momentum and a second-order momentum;
determining a ratio of a first target value to a second target value as a third target value, wherein the first target value is a product of the first learning rate of the first optimization mode and the first-order momentum, and the second target value is a root of a sum of a second momentum and a preset first value; and
determining an opposite number of the third target value as the update range.

3. The method according to claim 1 , wherein the determining the projection of the update range on the parameter gradient in the same vector space and taking the projection as the second learning rate of the second optimization mode comprises:
obtaining a target vector by transposing the update range;
determining a fourth target value and a fifth target value, wherein the fourth target value is a product of the target vector and the update range, and the fifth target value is a product of the target vector and the parameter gradient; and
calculating a ratio of the fourth target value to the fifth target value as the second learning rate of the second optimization mode.

4. The method according to claim 1 , wherein the determining the update range based on the first learning rate of the first optimization mode further comprises:
smoothing the second learning rate.

5. The method according to claim 4 , wherein the smoothing the second learning rate comprises:
determining a first weight;
determining a second weight;
determining a smoothed second learning rate at last training of the neural network; and
determining a smoothed second learning rate at this training of the neural network as a sum of a sixth target value and a seventh target value, wherein the sixth target value is a product of the first weight and the smoothed second learning rate at last training of the neural network, and the seventh target value is a product of the second weight and the second learning rate.

6. The method according to claim 5 , wherein the determining the first weight comprises:
determining a first-order momentum and a second-order momentum;
determining an eighth target value and a ninth target value, wherein the eighth target value is a difference between a preset second value and a hyperparameter of the first-order momentum, and the ninth target value is a difference between a preset third value and a hyperparameter of the second-order momentum; and
determining a ratio of a root of the eighth target value to a root of the ninth target value as the first weight.

7. The method according to claim 1 , wherein the determining that the second learning rate satisfies the preset update condition comprises:
determining a learning rate error;
determining a deviation of the second learning rate from the learning rate error as a learning rate deviation; and
determining the second learning rate satisfies the preset update condition in a case that the learning rate deviation is smaller than a preset threshold.

8. The method according to claim 7 , wherein the determining the learning rate error comprises:
determining the smoothed second learning rate;
determining a target hyperparameter, wherein the target hyperparameter is configured to control the second learning rate for training the neural network this time; and
determining a ratio of a target learning rate to a tenth target value as the learning rate error, wherein the tenth target value is a difference between a preset fourth value and the target hyperparameter.

9. The method according to claim 7 , wherein the determining the deviation of the second learning rate from the learning rate error as the learning rate deviation comprises:
determining a difference between the learning rate error and the second learning rate as an eleventh target value; and
determining an absolute value of the eleventh target value as the learning rate deviation.

10. The method according to claim 1 , wherein the neural network comprises a convolutional neural network (CNN), the first optimization mode comprises an adaptive moment estimation (Adam), and the second optimization mode comprises a stochastic gradient descent (SGD).

11. An apparatus for detecting a face, wherein the apparatus for detecting a face is configured to implement the method for detecting a face according to claim 1 .

12. A computer device, comprising:
at least one processor; and
a memory configured to store at least one program therein, wherein
the at least one program, when executed by the at least one processor, causes the at least one processor to implement the method for detecting a face according to claim 1 .

13. A computer-readable storage medium configured to store a computer program therein, wherein the computer program, when executed by a processor, causes the processor to implement the method for detecting a face according to claim 1 .

Assignees

• Bigo Tech Pte Ltd

Inventors

• Xiang Wei
• Pei Chao

Classifications

• G06V40/161

  Detection; Localisation; Normalisation · CPC title

• G06V10/82

  using neural networks · CPC title

• G06V10/7747

  Organisation of the process, e.g. bagging or boosting · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/08

  Learning methods · CPC title