Fast computation of a convolutional neural network

Patent metadata

Field	Value
Publication number	US-10635951-B1
Application number	US-201916390042-A
Country	US
Kind code	B1
Filing date	Apr 22, 2019
Priority date	Oct 24, 2018
Publication date	Apr 28, 2020
Grant date	Apr 28, 2020

Abstract

A computer-implemented method includes obtaining a trained convolutional neural network comprising one or more convolutional layers, each of the one or more convolutional layers comprising a plurality of filters with known filter parameters; pre-computing a reusable factor for each of the one or more convolutional layers based on the known filter parameters of the trained convolutional neural network; receiving input data to the trained convolutional neural network; computing an output of the each of the one or more convolutional layers using a Winograd convolutional operator based on the pre-computed reusable factor and the input data; and determining output data of the trained convolutional network based on the output of the each of the one or more convolutional layers.

First claim

What is claimed is:

1. A computer-implemented method, comprising:
obtaining, by a data processing apparatus, a trained convolutional neural network comprising at least a first convolutional layer and a second convolutional layer, each of the at least the first convolutional layer and the second convolutional layer comprising a corresponding plurality of filters with known filter parameters;
pre-computing, by the data processing apparatus, a first reusable factor for the first convolutional layer based on the known filter parameters regardless of any input data to the trained convolutional neural network;
pre-computing, by the data processing apparatus, a second reusable factor for the second convolutional layer based on the known filter parameters regardless of any input data to the trained convolutional neural network;
identifying, by the data processing apparatus, a first input data to the first convolutional layer of the trained convolutional neural network;
computing, by the data processing apparatus, a first output of the first convolutional layer using a Winograd convolutional operator based on the first reusable factor and the first input data;
using, by the data processing apparatus, the first output as a second input data to the second convolutional layer of the trained convolutional neural network; and
computing, by the data processing apparatus, a second output of the second convolutional layer using the Winograd convolutional operator based on the second reusable factor and the second input data.

2. The computer-implemented method of claim 1 , further comprising saving the first reusable factor for the first convolutional layer.

3. The computer-implemented method of claim 1 , wherein computing a first output of the first convolutional layer using a Winograd convolutional operator based on the first reusable factor and the first input data comprises computing the first output of the first convolutional layer according to a Winograd minimal filtering algorithm with the first reusable factor and the first input data.

4. The computer-implemented method of claim 1 , wherein the first input data to the trained convolutional neural network comprises one or more images.

5. The computer-implemented method of claim 1 , further comprising:
determining, by the data processing apparatus, output data of the trained convolutional neural network based on the second output of the second convolutional layer.

6. The computer-implemented method of claim 1 , further comprising outputting output data of the trained convolutional neural network via a user interface.

7. A non-transitory, computer-readable medium storing one or more instructions executable by a computer system to perform operations comprising:
obtaining a trained convolutional neural network comprising at least a first convolutional layer and a second convolutional layer, each of the at least the first convolutional layer and the second convolutional layer comprising a corresponding plurality of filters with known filter parameters;
pre-computing a first reusable factor for the first convolutional layers based on the known filter parameters regardless of any input data to the trained convolutional neural network;
pre-computing a second reusable factor for the second convolutional layer based on the known filter parameters regardless of any input data to the trained convolutional neural network;
identifying a first input data to the first convolutional layer of the trained convolutional neural network;
computing a first output of the first convolutional layer using a Winograd convolutional operator based on the first reusable factor and the first input data;
using the first output as a second input data to the second convolutional layer of the trained convolutional neural network; and
computing a second output of the second convolutional layer using the Winograd convolutional operator based on the second reusable factor and the second input data.

8. The non-transitory, computer-readable medium of claim 7 , the operations further comprising saving the first reusable factor for the first convolutional layer.

9. The non-transitory, computer-readable medium of claim 7 , wherein computing a first output of the first convolutional layer using a Winograd convolutional operator based on the first reusable factor and the first input data comprises computing the first output of the first convolutional layer according to a Winograd minimal filtering algorithm with the first reusable factor and the first input data.

10. The non-transitory, computer-readable medium of claim 7 , wherein
the first input data to the trained convolutional neural network comprises one or more images.

11. The non-transitory, computer-readable medium of claim 7 , the operations further comprising:
determining output data of the trained convolutional neural network based on the second output of the second convolutional layer.

12. The non-transitory, computer-readable medium of claim 7 , the operations further comprising outputting output data of the trained convolutional neural network via a user interface.

13. A computer-implemented system, comprising:
one or more computers; and
one or more computer memory devices interoperably coupled with the one or more computers and having tangible, non-transitory, machine-readable media storing instructions, that when executed by the one or more computers, perform operations comprising:
obtaining a trained convolutional neural network comprising at least a first convolutional layer and a second convolutional layer, each of the at least the first convolutional layer and the second convolutional layer comprising a corresponding plurality of filters with known filter parameters;
pre-computing a first reusable factor for the first convolutional layer based on the known filter parameters regardless of any input data to the trained convolutional neural network;
pre-computing a second reusable factor for the second convolutional layer based on the known filter parameters regardless of any input data to the trained convolutional neural network;
identifying a first input data to the first convolutional layer of the trained convolutional neural network;
computing a first output of the first convolutional layer using a Winograd convolutional operator based on the first reusable factor and the first input data;
using the first output as a second input data to the second convolutional layer of the trained convolutional neural network; and
computing a second output of the second convolutional layer using the Winograd convolutional operator based on the second reusable factor and the second input data.

14. The computer-implemented system of claim 13 , the operations further comprising saving the first reusable factor for the first convolutional layer.

15. The computer-implemented system of claim 13 , wherein computing a first output of the first convolutional layers using a Winograd convolutional operator based on the first reusable factor and the first input data comprises computing the first output of the first convolutional layer according to a Winograd minimal filtering algorithm with the first reusable factor and the first input data.

16. The computer-implemented system of claim 13 , the operations further comprising:
determining output data of the trained convolutional neural network based on the second output of the second convolutional layer.

17. The computer-implemented system of claim 13 , further comprising a user interface; and the operations further comprising outputting output data of the trai

Assignees

• Alibaba Group Holding Ltd

Inventors

• Liu Yongchao
• Huang Qiyin
• Pan Guozhen
• Li Sizhong
• Xu Jianguo
• Zhang Haitao
• Wang Lin

Classifications

• G06N3/063 Primary

  using electronic means · CPC title

• G06N3/08

  Learning methods · CPC title

• G06N3/04

  Architecture, e.g. interconnection topology · CPC title

• G06K9/66 Primary

  Physics · mapped topic

• G06V40/172 Primary

  Classification, e.g. identification · CPC title