Granted Patent B2

US 12,608,593 · App. 17/558,327 · Granted Apr 21, 2026

Inventors: Jian Ren (Marina Del Ray, CA); Oliver Woodford (Santa Monica, CA); Sergey Tulyakov (Marina del Rey, CA); Jiazhuo Wang (Santa Monica, CA); Qing Jin (Palo Alto, CA)

Assignee: Snap Inc.

G06N3/045G06N3/088

Abstract

Claims (51)

1 . A method comprising:

generating a first generative adversarial network (GAN) comprising a first type of convolutional layer, a second type of convolutional layer and a plurality of kernel sizes;

identifying a threshold, the threshold representing a measure of computational costs of the GAN;

based on the threshold, generating a second GAN by pruning channels of the first GAN;

training the second GAN using similarity-based knowledge distillation from the first GAN, the training comprising:

computing a global similarity metric between features of the first GAN and features of the second GAN on channels of the first GAN and channels of the second GAN, the global similarity metric characterizing batch-wise and spatial-wise similarity; and

storing the trained second GAN.

2 . The method of claim 1 , wherein the first type of convolutional layer is a conventional convolutional layer and the second type of convolutional layer is a depth-wise convolutional layer.

3 . The method of claim 1 , wherein the threshold is a number of multiply-accumulate operations.

4 . The method of claim 1 , wherein training the second GAN further comprises:

minimizing a distillation loss on a feature space comprising the first GAN and the second GAN using the computed global similarity metric.

5 . The method of claim 1 , wherein the first GAN comprises of a plurality of channels and wherein generating the second GAN further comprises:

for each channel of the first GAN:

identifying a weight of the channel;

determining that the weight of the channel is less than the threshold; and

based on the determination, pruning the channel.

6 . The method of claim 1 , wherein the second GAN is stored on a mobile computing device.

7 . The method of claim 1 , wherein the second GAN is a compressed form of the first GAN.

8 . The method of claim 1 , wherein the first GAN is a pre-trained GAN.

9 . A system comprising:

a processor; and

a memory storing instructions that, when executed by the processor, configure the system to perform operations comprising:

generating a first generative adversarial network (GAN) comprising a first type of convolutional layer, a second type of convolutional layer and a plurality of kernel sizes;

identifying a threshold, the threshold representing a measure of computational costs of the GAN;

based on the threshold, generating a second GAN by pruning channels of the first GAN;

training the second GAN using similarity-based knowledge distillation from the first GAN, the training comprising:

computing a global similarity metric between features of the first GAN and features of the second GAN on channels of the first GAN and channels of the second GAN, the global similarity metric characterizing batch-wise and spatial-wise similarity; and

storing the trained second GAN.

10 . The system of claim 9 , wherein the first type of convolutional layer is a conventional convolutional layer and the second type of convolutional layer is a depth-wise convolutional layer.

11 . The system of claim 9 , wherein the threshold is a number of multiply-accumulate operations.

12 . The system of claim 9 , wherein training the second GAN further comprises:

minimizing a distillation loss on a feature space comprising the first GAN and the second GAN using the computed global similarity metric.

13 . The system of claim 9 , wherein the first GAN comprises of a plurality of channels and wherein generating the second GAN further comprises:

for each channel of the first GAN:

identifying a weight of the channel;

determining that the weight of the channel is less than the threshold; and

based on the determination, pruning the channel.

14 . The system of claim 9 , wherein the second GAN is stored on a mobile computing device.

15 . The system of claim 9 , wherein the second GAN is a compressed form of the first GAN.

16 . The system of claim 9 , wherein the first GAN is a pre-trained GAN.

17 . A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that when executed by a computer, cause the computer to perform operations comprising:

generating a first generative adversarial network (GAN) comprising a first type of convolutional layer, a second type of convolutional layer and a plurality of kernel sizes;

identifying a threshold, the threshold representing a measure of computational costs of the GAN;

based on the threshold, generating a second GAN by pruning channels of the first GAN;

training the second GAN using similarity-based knowledge distillation from the first GAN, the training comprising:

computing a global similarity metric between features of the first GAN and features of the second GAN on channels of the first GAN and channels of the second GAN, the global similarity metric characterizing batch-wise and spatial-wise similarity; and

storing the trained second GAN.

18 . The computer-readable storage medium of claim 17 , wherein the first type of convolutional layer is a conventional convolutional layer and the second type of convolutional layer is a depth-wise convolutional layer.

19 . The computer-readable storage medium of claim 17 , wherein the threshold is a number of multiply-accumulate operations.

20 . The computer-readable storage medium of claim 17 , wherein training the second GAN further comprises:

minimizing a distillation loss on a feature space comprising the first GAN and the second GAN using the computed global similarity metric.

Assignments (2)

Continuity (2)

References Cited (106)