WO2024234986A1 - Method for customizing sensor function and smart device comprising sensor - Google Patents

Abstract

The present invention relates to a method for customizing a sensor function based on Wasm technology. The method comprises: providing a mathematical model for training a sensor in a smart device, and a post-processing program customized for the mathematical model, the post-processing program being used to execute post-processing of output data of the mathematical model; deploying the mathematical model in the smart device; compiling the post-processing program into a Wasm program and deploying the Wasm program in the smart device; providing an interface defined on the basis of a WASI specification in the smart device, the Wasm program being capable of accessing and operating the sensor by means of the interface. In addition, the present invention relates to a smart device comprising a sensor, the function of the sensor being customized using the method according to the present invention.

Description

Method for customizing sensor functions and smart device including sensor

Citation of Related Applications

This application claims the benefit of Chinese Patent Application No. 202310552597.X filed with the State Intellectual Property Office of the People's Republic of China on May 16, 2023, the entire contents of which are hereby incorporated by reference into this document in their entirety.

Technical Field

The present invention relates to the field of computers, and in particular to a technology for customizing sensor functions based on Wasm technology.

Background Art

In the field of intelligent technology, sensors can be used in autonomous driving, equipment detection, environmental detection, logistics detection, security detection, health detection, smart home appliances, retail industry, health care, agriculture, smart cities, etc.

The sensors in smart devices contain computing models or AI models that are developed, trained, or customized for specific purposes. These models are collectively referred to as mathematical models in this article. The sensors can use the corresponding mathematical models to provide computing or reasoning results based on the sensor's detection. Here, the format of the computing or reasoning results may vary depending on the model.

For these different formats, smart devices need to post-process the post-processing results on site to obtain the expected information. A typical post-processing method is to convert the calculation or reasoning results into a readable, visible and structured metadata character stream.

As mentioned above, different models may output calculation or reasoning results in different formats. As a result, it becomes difficult to customize the post-processing function for the model. The post-processing function needs to be provided by the model developer. However, model developers usually do not understand the software and hardware architecture of smart devices, and developers pay more attention to the logic of post-processing and ensure its adaptation to the model.

In order to customize the post-processing function, you should know how to identify the data in a specific model and how to communicate with the sensor. However, developers usually do not know the working environment of smart devices, and therefore it is difficult to deploy matching post-processing functions that can run on any smart device. There are currently various ways to customize post-processing functions using Wasm technology, but they all have some problems to a greater or lesser extent.

For example, the Camera Remote SDK, a software development kit provided by Sony, does not provide a general API for sensor control. Waft (WebAssembly Framework for Things), a new development framework proposed by Alibaba engineers, does not provide an API for accessing AI sensors or general sensors to read data. In US20220198071A1, Wasm technology is suitable for task sharing to reduce the burden on the main computing node, but no method is provided for task simulation on child nodes. CN202210944973, CN202210096203, etc. lack sandbox control, have security issues, or do not comply with wasi specifications.

In addition, current model development is highly dependent on several mainstream software frameworks, such as tensorflow, pytorch, onnx, etc. On the one hand, these frameworks do accelerate the model development process to some extent, but on the other hand, mathematical models and possible post-processing programs often have some problems during the deployment phase and are difficult to debug, especially for some low-level, cost-effective industrial equipment.

Summary of the invention

The purpose of the present invention is to propose a method and a module for customizing sensor functions based on Wasm technology, which do not have the disadvantages of the prior art and have additional advantages. The present invention allows developers to focus only on the logic of model post-processing without having to pay attention to the specific implementation and deployment of the model.

One aspect of the present invention relates to a method for customizing sensor functions based on Wasm technology, which includes: providing a mathematical model trained for a sensor in a smart device and a post-processing program customized for the mathematical model, wherein the post-processing program is used to perform post-processing on output data of the mathematical model; deploying the mathematical model in the smart device; compiling the post-processing program into a Wasm program and deploying it in the smart device; providing an interface defined based on the WASI specification in the smart device, and the Wasm program can access and operate the sensor through the interface.

Another aspect of the present invention relates to a smart device comprising a sensor, wherein the functionality of the sensor is customized according to the method of the present invention.

Another aspect of the present invention relates to a debugging tool for debugging a mathematical model, which includes a post-processing module and a compiler stack, wherein the post-processing module generates a post-processing template file for post-processing the mathematical model, provides the post-processing template file to the compiler stack, and receives output data of the mathematical model from a runtime running the mathematical model and provides it to a developer, wherein the compiler stack compiles the received post-processing template file into an executable file that can be executed by the runtime, and provides the executable file to the runtime, so that the runtime can generate the output data using the mathematical model.

According to the present invention, Wasm technology is used to enable the post-processing function to be deployed across platforms and in a modular manner, thereby enabling it to be deployed on smart devices. Customize sensor specific functions in

In addition, according to the present invention, developers can easily debug post-processing, and have the following advantages: no need to compulsorily learn low-level languages or employ specific hardware engineers for debugging; guaranteed execution efficiency; easy to transplant.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 shows the creation and training of the mathematical model and the development of the post-processing procedure.

FIG. 2 shows the development process of a mathematical model according to the prior art.

FIG. 3 shows a debugging tool for debugging a mathematical model according to an embodiment of the present invention.

FIG. 4 shows a flow chart of a process for debugging a mathematical model according to an embodiment of the present invention.

DETAILED DESCRIPTION

The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it. However, the embodiments are not intended to limit the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments may be combined with each other.

In various smart devices of IoT there are various sensors like image sensors, sound sensors, light sensors etc. These sensors can be used to achieve different purposes depending on the smart device.

In order to use the sensor for a specific purpose, the developer can create and train the corresponding mathematical model in, for example, a host. The developer can use a model framework tool to create and train the mathematical model. For example, development platforms such as tensorflow, pytorch, and onnx can be used as model framework tools. The sensor can use the mathematical model to calculate or infer the sensor's detection results to obtain a calculation result or an inference result. The sensor then outputs the obtained calculation result or inference result as output data in a specific format.

For the output data, developers also need to develop a post-processing program for post-processing in, for example, a host computer. Developers can develop post-processing programs in their favorite programming languages, such as Python, C/C++, Java, Go, or Rust. Through the post-processing program, the smart device can post-process the output data of the sensor (such as format conversion, etc.) to provide a post-processing result containing the desired information.

As an example, FIG1 shows a case where an IoT device including a camera is used as a smart device. Here, the camera is a sensor in the sense of the present invention. Note that the smart device of the present invention is not limited to IoT devices. According to the technical inspiration of the present invention, those skilled in the art can also apply the present invention to the fields of autonomous driving, equipment detection, environmental detection, logistics detection, security detection, health detection, smart home appliances, retail industry, health care, agriculture, smart cities, etc. Here, they are not listed one by one.

The left side of Figure 1 shows the creation and training of the mathematical model of the camera and the development of the post-processing program. Here, the mathematical model of the camera can be an object detection model, such as a pedestrian detection model, a face recognition model, etc.

In the prior art, the completed mathematical model and post-processing program are usually directly deployed on the hardware of the smart device, for example, in the example of FIG1, they are respectively deployed in the digital signal processor (DSP) of the camera and the processor of the IoT device. When deploying the mathematical model, the power consumption, performance, and resources of the camera are taken into consideration, and the mathematical model can be lightweighted, for example, by using methods such as deep separable convolution, grouped convolution, adjustable hyperparameters to reduce spatial resolution, knowledge distillation, low-rank pruning, and model compression.

The camera uses the mathematical model to calculate or infer the camera detection results and outputs the calculation or inference results as output data. Then, the IoT device including the camera can use the post-processing program to post-process the camera output data to obtain the desired information. The programming environment of the mathematical model can be Python, C/C++, Java, Rust, Go, etc.

However, unlike the prior art, the present invention proposes a technical solution for customizing sensor functions based on Wasm technology. Wasm, or WebAssembly, is a new bytecode format that is gradually being widely recognized due to its portability, small size, and high security. Wasm allows users to write in a language they are familiar with (currently supports programming languages such as python, C/C++, Rust, Java, or Go), and then use a virtual machine engine to run on the web. Wasm technology supports running Wasm programs in a sandbox environment.

As shown in Figure 1, in the present invention, the post-processing program is compiled into a Wasm program and deployed in an IoT device (e.g., a processor) to run in a virtual machine provided by the IoT device. Depending on the programming language of the post-processing program and the embedded environment of the Wasm program, various compilers, such as the LLVM backend, can be used.

In order to enable the Wasm program to implement functions consistent with the post-processing program, the present invention uses the WASI (WebAssembly System Interface) specification to define a unified sensor system interface for the Wasm program, which is referred to as WASI sensor. WASI is a new API system, which is a set of engine-independent, non-Web system-oriented API standards designed for WASM programs, that is, the standard for how WASM programs interact with the host environment, for example, including standard system calls such as file system, network stack, time, and random number generator. These system calls are provided to the Wasm program through predefined API interfaces, allowing it to access data resources in the sensor in a sandbox environment. Based on the support of WASI, Wasm programs can also run in non-web environments. As long as the WASI standard is supported, any runtime that runs Wasm programs can Implement system calls in the host environment of the IoT device. The runtime library provides the execution environment of the program, such as memory allocation, variable assignment, function call, etc. It can also be used to manage threads or processes, execute interpreted code, dynamic link libraries and shared libraries, etc.

Through the sensor system interface, the Wasm program can correctly parse the output data generated by the sensor's mathematical model, and can access and operate the sensor (for example, turn it on, turn it off, perform reasoning, retrieve results, etc.). As a result, the Wasm program can access and control the various functions provided by the sensor through the Wasi interface through the sandbox boundary, such as obtaining reasoning data, controlling the life cycle of the sensor, etc., which can not only ensure the security and reliability of the program, but also protect the security of various types of information and equipment. Specifically, when the Wasm program needs to access sensor resources, it can pass the access request to the sensor through the interface. The sensor responds to the access request and returns the reasoning data to the Wasm program.

In the present invention, there may be multiple sensors in the smart device, and each sensor may be provided with at least one mathematical model to achieve a specific purpose. For example, for the example of FIG1 , multiple mathematical models may be provided for the camera, including but not limited to a pedestrian detection model, a face recognition model, a text recognition model, a license plate information recognition model, etc.

A post-processing program may be provided for each mathematical model to perform post-processing on the output data of the mathematical model. Of course, a post-processing program may also be provided for multiple mathematical models to perform post-processing on the output data of these mathematical models in combination.

Since the present invention customizes sensor functions based on Wasm technology, developers can also develop post-processing programs in any language they like, without having to consider what platform it runs on and how to adjust the program in different devices. In addition, Wasm technology restricts Wasm programs to run in a sandbox environment, so it has high security.

In addition, the present invention provides a sensor system interface based on the WASI specification. It runs in a sandbox environment to control the entire life cycle, which becomes safer. The present invention also implements modular sensor control functions, and the function of the sensor can be expanded or changed by updating a single Wasm program instead of updating the entire platform.

In summary, the present invention proposes a set of cross-platform sensor control interfaces, which developers can use to connect mathematical model development locally and implement specific post-processing functions. The Wasm technology enables the post-processing function to be deployed cross-platform and modularly, so that specific functions of sensors can be customized in smart devices.

Hereinafter, the present invention will explain the deployment and debugging of Wasm programs in detail.

FIG. 2 shows a process of developing a mathematical model using, for example, the Python programming language according to the prior art.

As shown in Figure 2, first, create and train a mathematical model. Model developers can use model framework tools to train mathematical models. For example, here, development platforms such as tensorflow, pytorch, and onnx can be used as model framework tools in the Python programming environment. Programming tools are not limited to Python, but can also be C/C++, Java, Rust, and Go.

Specifically, the model framework tool is first used to generate an initial model, and then the generated initial model is trained through machine learning. After the initial model is trained, the trained initial model is exported.

Next, depending on the type of model and the operating environment, developers can use the corresponding model deployment tools to deploy the completed mathematical model in a specific environment, such as an edge device such as a personal computer or a smartphone. Here, the edge device can be the camera in Figure 1.

Finally, the deployed mathematical model can be run in the edge device to obtain calculation results or inference results.

However, at present, most developers are only familiar with programming languages based on software framework requirements, such as Python, but are not familiar with hardware-level programming languages. This phenomenon makes it difficult for algorithm developers to directly debug Wasm programs, which seriously limits the richness of mathematical models.

Furthermore, even if mathematical models can be deployed using a hardware platform-specific software stack, they are usually implemented through a series of complex manual operations and are difficult to port to other devices.

Furthermore, developers mainly focus on model design creation without ensuring the efficiency requirements of industrial software.

To solve the above problems, the present invention proposes a debugging tool that can be created in a Docker image to provide developers with out-of-the-box services. As is known to all, using Docker images, development platforms for various programming languages can be created, such as Python, Java, Go, and other programming languages. This article uses the Python programming language as an example for illustration.

In addition, for the operation of mathematical models, a runtime (e.g., WAMR Runtime) in edge devices (e.g., the smart devices mentioned above) is also necessary so that the services can be applied to a wide range of hardware platforms and the entire solution can be smoothly ported when necessary.

Specifically, FIG3 shows a debugging tool according to an embodiment of the present invention, which is used to debug a mathematical model to obtain a higher task index, a faster running speed, and a smaller model volume.

According to the present invention, developers can, for example, create the entire debugging tool in a Docker image on their own host to provide a Docker container service from the Docker image, so developers can provide this service at any time on almost all PC or server operating system platforms, such as in x86 or arm hardware structures. Docker container is an open source application container engine that allows developers to package their applications and dependent packages into portable containers in a unified manner. Specifically, in order to create a debugging tool in a Docker image, for example, first write a Dockerfile file, which is used to define the environment and application in the Docker image; then, in the Dockerfile file Add the required debugging tools. For example, you can choose to add gdb, strace, tcpdump and other tools as needed. In addition, depending on the type of debugging tools used and the requirements of the application, you may also need to set the corresponding environment variables in the Dockerfile file or expose the port of the container. Then, build the Docker image. For example, execute the "docker build -t <image name>" command in the directory containing the Dockerfile file. This command will start building the Docker image. Next, start the Docker container. For example, run the following command to start the Docker container: "docker run -it –name <container name><imagename>"; Finally, test the debugging tool, run the required application in the started container, and then use the added debugging tool for debugging.

The debugging tool according to the present invention may include a post-processing module, a compiler stack and an optional optimizer. The post-processing module may include a post-processing code module, a decoder and a raw data viewer. The compiler stack may include a C++ converter, a WASM compiler and a publisher. The optimizer may include a standard analyzer, a comparator and a scheduler.

According to the present invention, in order to debug the mathematical model, a runtime (e.g., WAMR runtime) can be used to run the mathematical model in the edge device.

According to the present invention, the Docker runtime environment of the debugging tool can be connected to the runtime of the edge device via a connector interface, and the connector interface can be used to handle communication protocols, data format conversion, security authentication and other issues to achieve communication and interaction. If necessary, data encryption, identity authentication and other issues need to be considered to ensure the security of communication and the integrity of data. Here, the connection can be automatically performed using various communication protocols, such as wired or wireless communication protocols.

In FIG. 3 , as shown in mark ①, the developer can fill the Python post-processing code for post-processing into the template Python file through the post-processing code module. The Python post-processing code may include instructions for post-processing the output data of the mathematical model. The post-processing code can be customized by the developer according to the desired post-processing function.

As shown in mark ②, the C++ converter in the compiler stack converts the Python post-processing code from the post-processing code module into C++ code. Then, the WASM compiler in the compiler stack converts the C++ code into a Wasm executable file. Finally, as shown in marks ③ and ④, the compiler stack uses the publisher to publish the converted Wasm executable file to the Runtime that runs the mathematical model via the connector interface.

As shown in marks ⑤ and ⑥, the mathematical model outputs the calculation or reasoning result (i.e., output data), and outputs the calculation or reasoning result to the decoder in the post-processing module via the connector interface. The decoder decodes the calculation or reasoning result. As shown in mark ⑦, the raw data viewer in the post-processing module displays the decoded calculation or reasoning data to the developer in the form of, for example, images, sounds, text, etc. If the calculation or reasoning result is incorrect or unsatisfactory, the developer can adjust the mathematical model and then repeat the process as shown in marks ① to ⑦ until the correct or ideal calculation or reasoning result is obtained.

Optionally, after obtaining the correct or ideal calculation or inference results, the developer can also optimize the mathematical model.

As shown in mark ⑧, developers can submit performance criteria, trial time cost constraints, and current results to the optimizer. In the optimizer, the standard analyzer parses the performance standard configuration to extract parameters and time costs, and the comparator compares the standard request and the current result. If the result meets the request, then as shown in mark As shown, the current result and time cost are directly returned to the developer. Otherwise, as shown in mark ⑨, the scheduler is started to test various optimization hyperparameters and scheduling strategies, and the publisher is controlled to run the optimized test code on the edge device in the process shown in marks ③ to ⑤, and the result of each test case is sent to the comparator in the manner shown in mark ⑩ until a satisfactory value is reached or the best result within the set time is found, and then it is sent back to the developer together with the time cost.

FIG. 4 shows a flow chart of a process for debugging a mathematical model according to an embodiment of the present invention.

Determine whether to start debugging. When debugging is started, fill the post-processing code into the template file through the post-processing code module.

Next, the template file is converted into an execution file through the compiler stack.

Next, it is determined whether the connector interface is ready for connection. If this is not the case, the device is restarted to connect until the connection is successful. In the case of a successful connection, the executable file is published to the runtime via the connector interface through the compiler stack.

Next, the mathematical model outputs the calculation or reasoning result (ie, output data) to the post-processing module via the connector interface. Then, the post-processing module decodes the calculation or reasoning result and displays the decoded calculation or reasoning result to the developer.

If the inference result is incorrect or unsatisfactory, the developer can adjust the mathematical model and repeat the above process until the correct or ideal calculation or inference result is obtained.

After obtaining the correct or ideal calculation or reasoning results, determine whether to further optimize the mathematical model.

If optimization is not required, the debugging process ends directly.

If optimization is required, the performance standard configuration, experimental time cost limit, and current result can be submitted to the optimizer. In the optimizer, the parameters and time limits are extracted, and the standard request and initial results are compared. If the result meets the request, the current result and time cost are directly returned to the developer. Otherwise, the scheduler is started to test various optimization hyperparameters and scheduling strategies, the publisher is controlled to run the optimized test code on the edge device, and the results of each test case are sent to the comparator until a satisfactory value is reached or the best result within the set time is found, and then it is sent back to the developer together with the time cost. At this point, the debugging process ends.

According to the present invention, the following configuration may be adopted.

(1) A method for customizing sensor functions based on Wasm technology, comprising:

Providing a mathematical model trained for a sensor in an intelligent device and a post-processing program customized for the mathematical model, wherein the post-processing program is used to perform post-processing on output data of the mathematical model;

Deploying the mathematical model in the smart device;

Compiling the post-processing program into a Wasm program and deploying it in the smart device;

An interface defined based on the WASI specification is provided in the smart device, and the Wasm program can access and operate the sensor through the interface.

(2) The method according to (1) above, wherein the Wasm program and the post-processing program have the same post-processing function.

(3) The method according to (1) above, wherein the mathematical model is deployed in the sensor.

(4) The method according to (1) above, wherein the Wasm program runs in a virtual machine provided by the smart device.

(5) The method according to (1) above, wherein the post-processing is a format conversion of the output data.

(6) According to the method described in (1) above, the post-processing program and the mathematical model are developed using any one of the programming languages selected from Python, C/C++, Java, Rust and Go.

(7) According to the method described in (1) above, the method further comprises the step of debugging the mathematical model in the step of deploying the mathematical model in the smart device, wherein the debugging comprises:

executing the mathematical model in a suitable runtime;

Creating a debugging tool in a Docker image and running the debugging tool in a Docker runtime environment;

The debugging tool fills the post-processing code for the post-processing into the template file;

The debugging tool compiles the template file into an execution file that can be executed by the runtime;

The runtime receives the execution file from the Docker runtime environment and executes the execution file to generate the output data using the mathematical model;

The debugging tool receives the output data from the runtime, decodes the output data, and provides the decoded output data to the developer;

In case the output data is incorrect, the developer adjusts the mathematical model;

Repeat the above steps until the correct output data is obtained.

(8) The method according to (7) above, wherein when compiling the execution file,

The debugging tool converts the template file into C++ code, converts the C++ code into the execution file, and publishes the execution file to the runtime.

(9) According to the method described in (7) above, after obtaining the correct output data, the debugging tool analyzes the standard request set by the developer, compares the standard request and the output data, and when the result of the comparison is unsatisfactory, tests multiple scheduling strategies and finds the best scheduling strategy.

(10) The method according to (7) above, wherein the runtime is provided by the smart device.

(11) The method according to (7) above, wherein the Docker operating environment is provided by the developer's host.

(12) The method according to (7) above, wherein the Docker image creates a development environment for any one of the programming languages of Python, C/C++, Java, Rust and Go, and the mathematical model, the debugging tool and the post-processing program are developed or created using the same programming language.

(13) A smart device including a sensor,

The function of the sensor is customized according to any one of the methods described in (1) to (12) above.

(14) A debugging tool for debugging a mathematical model, comprising a post-processing module and a compiler stack,

The post-processing module generates a post-processing template file for post-processing the mathematical model, provides the post-processing template file to the compiler stack, and receives output data of the mathematical model from a runtime running the mathematical model and provides it to a developer.

The compiler stack compiles the received post-processing template file into an execution file executable by the runtime, and provides the execution file to the runtime, so that the runtime can generate the output data using the mathematical model.

(15) The debugging tool according to (14) above, wherein the debugging tool and the mathematical model are developed or created using the same programming language.

(16) The debugging tool according to (14) above, wherein the post-processing module comprises:

A post-processing code module, which is used to generate the post-processing template file and provide it to the compiler stack;

a decoder for receiving the output data of the mathematical model from the runtime and decoding the output data; and

A raw result viewer is used to provide the decoded output data to the developer.

(17) The debugging tool according to (14) above, wherein the compiler stack comprises:

A C++ converter, which is used to convert the post-processing template file into C++ code;

A WASM compiler, which is used to convert the C++ code into the execution file; and

A publisher is used to publish the execution file to the runtime.

(18) The debugging tool according to (14) above further includes an optimizer, wherein the optimizer includes:

A standard analyzer for analyzing standard requests set by the developer;

a comparator for comparing the standard request with the output data of the mathematical model; and

The scheduler is used to test multiple scheduling strategies and find the best scheduling strategy when the result of the comparison is not satisfactory.

(19) The debugging tool according to (14) above, wherein:

The debugging tool is created in a Docker image.

It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims (19)

A method for customizing sensor functions based on Wasm technology, comprising: Providing a mathematical model trained for a sensor in an intelligent device and a post-processing program customized for the mathematical model, wherein the post-processing program is used to perform post-processing on output data of the mathematical model; Deploying the mathematical model in the smart device; Compiling the post-processing program into a Wasm program and deploying it in the smart device; An interface defined based on the WASI specification is provided in the smart device, and the Wasm program can access and operate the sensor through the interface. The method according to claim 1, wherein the Wasm program and the post-processing program have consistent post-processing functions. The method of claim 1, wherein the mathematical model is deployed in the sensor. The method according to claim 1, wherein the Wasm program runs in a virtual machine provided by the smart device. The method according to claim 1, wherein the post-processing is format conversion of the output data. According to the method of claim 1, the post-processing program and the mathematical model are developed by any one of the programming languages selected from Python, C/C++, Java, Rust, and Go. The method according to claim 1, wherein the method further comprises the step of debugging the mathematical model in the step of deploying the mathematical model in the smart device, wherein the debugging comprises: executing the mathematical model in a suitable runtime; Creating a debugging tool in a Docker image and running the debugging tool in a Docker runtime environment; The debugging tool fills the post-processing code for the post-processing into the template file; The debugging tool compiles the template file into an execution file that can be executed by the runtime; The runtime receives the execution file from the Docker runtime environment and executes the execution file to generate the output data using the mathematical model; The debugging tool receives the output data from the runtime, decodes the output data, and provides the decoded output data to the developer; In case the output data is incorrect, the developer adjusts the mathematical model; Repeat the above steps until the correct output data is obtained. The method according to claim 7, wherein, when compiling the executable file, The debugging tool converts the template file into C++ code, converts the C++ code into the execution file, and publishes the execution file to the runtime. The method according to claim 7, wherein, after obtaining the correct output data, the debugging tool analyzes the standard request set by the developer, compares the standard request and the output data, and when the result of the comparison is not satisfactory, tests multiple scheduling strategies and finds the best scheduling strategy therefrom. The method of claim 7, wherein the runtime is provided by the smart device. The method according to claim 7, wherein the Docker runtime environment is provided by the developer's host. The method according to claim 7, wherein the Docker image creates a development environment for any one of the programming languages of Python, C/C++, Java, Rust and Go, and the mathematical model, the debugging tool and the post-processing program are developed or created using the same programming language. A smart device including a sensor, Wherein, the function of the sensor is customized using the method according to any one of claims 1 to 12. A debugging tool for debugging a mathematical model, comprising a post-processing module and a compiler stack, wherein the post-processing module generates a post-processing template file for post-processing the mathematical model, provides the post-processing template file to the compiler stack, and receives output data of the mathematical model from a runtime running the mathematical model to provide to a developer, The compiler stack compiles the received post-processing template file into an execution file executable by the runtime, and provides the execution file to the runtime, so that the runtime can generate the output data using the mathematical model. The debugging tool according to claim 14, wherein the debugging tool and the mathematical model are developed or created using the same programming language. The debugging tool according to claim 14, wherein the post-processing module comprises: A post-processing code module, which is used to generate the post-processing template file and provide it to the compiler stack; a decoder for receiving the output data of the mathematical model from the runtime and decoding the output data; and A raw result viewer is used to provide the decoded output data to the developer. The debugging tool according to claim 14, wherein the compiler stack comprises: A C++ converter, which is used to convert the post-processing template file into C++ code; A WASM compiler, which is used to convert the C++ code into the execution file; and A publisher is used to publish the execution file to the runtime. The debugging tool according to claim 14, further comprising an optimizer, wherein the optimizer comprises: A standard analyzer for analyzing standard requests set by the developer; a comparator for comparing the standard request with the output data of the mathematical model; and The scheduler is used for testing a plurality of scheduling strategies and finding the best scheduling strategy therefrom when the result of the comparison is not satisfactory. The debugging tool according to claim 14, wherein: The debugging tool is created in a Docker image.