CN-122021863-A - System for performing estimation using hybrid precision, program for performing estimation using hybrid precision, and method for performing estimation using hybrid precision

Abstract

A system (100) for performing estimation using hybrid precision, a program for performing estimation using hybrid precision, and a method for performing estimation using hybrid precision, is provided with an information acquisition unit (121) for acquiring environmental information, which is information of the environment around an estimated object, and a setting unit (122) for setting the type of data used for each layer in the estimation, based on the acquired environmental information.

Inventors

• LIN WENCHENG
• YAMAMOTO SAKON

Assignees

• 株式会社电装
• 丰田自动车株式会社
• 未来瞻科技株式会社

Dates

Publication Date

20260512

Application Date

20251103

Priority Date

20241112

Claims (5)

1. 1. A method of performing inference using hybrid accuracy, characterized in that, The method comprises the following steps: Acquiring environmental information, which is information of the environment around the inferred object, and And setting a data type used by each layer in the estimation according to the acquired environment information.
2. 2. A computer program product comprising computer program/instructions, characterized in that, Which computer program/instructions, when executed by a processor, carries out the steps of the method according to claim 1.
3. 3. A system comprising a memory, a processor, and a computer program stored in the memory, the system characterized in that, The steps of the method of claim 1 are implemented by the processor executing the computer program.
4. 4. A system according to claim 3, The memory stores in advance a mixed precision table which is a table associating the environment information with the data type used by each layer, The processor implements the step of setting the data type used by each layer using the context information and the hybrid precision table.
5. 5. The system according to claim 3 or 4, The inference is performed to detect objects around the vehicle, The environmental information includes at least one of solar radiation information related to an amount of solar radiation of the surroundings of the vehicle, weather information related to weather of the surroundings of the vehicle, road information related to a road category of the surroundings of the vehicle, and congestion information related to a congestion condition of the surroundings of the vehicle.

Description

System for performing estimation using hybrid precision, program for performing estimation using hybrid precision, and method for performing estimation using hybrid precision Technical Field The present disclosure relates to a system for performing estimation using hybrid precision, a program for performing estimation using hybrid precision, and a method for performing estimation using hybrid precision. Background In inferences using NPUs (Neural network Processing Unit: neural network processors), various techniques for reducing computation time have been proposed. For example, patent document 1 discloses a technique using Mixed Precision (Mixed Precision). Hybrid accuracy is a method of reducing the operation time while maintaining the accuracy of inference by changing the data type per layer in inference. The type of data used in each layer is preset. Prior art literature Patent technical literature Patent document 1 Japanese patent application laid-open No. 2023-031296 Disclosure of Invention In the case where the hybrid accuracy is used for detecting the object around the vehicle, the surrounding environment changes in real time. Thus, the appropriate data type for each layer may also vary from time to time. This causes a problem that the predetermined data type is no longer an appropriate data type, and the accuracy of object detection is lowered and the calculation time is prolonged. Such a problem is not limited to the vehicle, and may be common to object detection used in an environment that changes in real time. The present disclosure can be implemented as follows. According to one aspect of the present disclosure, a system for performing inference using hybrid accuracy is provided. The system comprises an information acquisition unit for acquiring environmental information, which is information of the surrounding environment of the estimated object, and a data type setting unit for setting the data type used for each layer in the estimation based on the acquired environmental information. The system according to the above aspect includes an information acquisition unit that acquires environmental information that is information of the surrounding environment of the estimated object, and a data type setting unit that sets the data type used for each layer in the estimation based on the acquired environmental information, so that even when the surrounding environment changes at any time, the data type used for each layer can be appropriately set based on the environmental information. This can suppress a decrease in the accuracy of estimation and a longer calculation time. The present disclosure can also be implemented in various ways. For example, the present invention can be implemented in a manner of executing a method of estimating using hybrid accuracy, a computer program for implementing the method, a non-transitory recording medium storing the computer program, or the like. Drawings Fig. 1 is a diagram showing a schematic configuration of a system according to an embodiment of the present disclosure. Fig. 2 is a diagram showing an example of the mixing precision table. Fig. 3 is a diagram showing an example of Config. Fig. 4 is a flowchart showing steps of a method of performing estimation using hybrid accuracy. Fig. 5 is a flowchart showing the steps of the Config production method. Fig. 6 is a diagram for explaining a method of manufacturing Config. Detailed Description A. description of the embodiments A1. Construction of System 100 The system 100 shown in fig. 1 is used to perform inferences using hybrid accuracy. Specifically, the system 100 is used for inference of object detection in the environment surrounding the vehicle. The system 100 in the present embodiment is mounted on a vehicle. The system 100 is connected to the environment information section 110. The system 100 includes a processor 120 and a storage 130. < Construction of Environment information section 110 > The environment information unit 110 detects environment information and outputs the detected environment information to the processor 120. The environment information is information of the environment around the inferred object. The environment information unit 110 includes a camera 111, a sunlight sensor 112, a weather sensor 113, a road information sensor 114, and a traffic sensor 115. The camera 111 photographs the surroundings of the vehicle as environmental information. The camera 111 can capture not only the front of the vehicle but also the sides and the rear of the vehicle. The object of object detection is included in the angle of view of the camera 111. The camera 111 has, for example, CCD (Charge Coupled Device) or CMOS (Complementary Metal Oxide Semiconductor) sensors. The photographed image is supplied to NPU (Neural network Processing Unit) via the processor 120. The NPU200 performs inference for detecting an object contained in the photographed image. The sun exposure sensor 112 detects sun exposure i