A general neural network creation method, device, platform, medium and equipment

Abstract

The application discloses a general neural network creation method, device, platform, medium and equipment, and belongs to the field of deep learning. The method comprises the following steps: processing parameters according to custom functions of a user, creating an algorithm configuration file, and integrating the algorithm configuration file into a model package; setting data processing logic of the algorithm configuration file by using the processing parameters in one or more algorithm configuration files, obtaining an algorithm inference library matched with the model package; creating a model package loading component according to a path of the model package; running a specified model in the model package through a model loading interface by using the model package loading component, and setting a default running device; running the specified model in the model package through the algorithm inference library in the default running device, processing user-defined data to obtain network processing data, and creating a general neural network. The efficiency of the general neural network in running the algorithm inference library to obtain a conversion result is higher, the effect is better, and the time for actual deployment of a vehicle is saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of deep learning, and in particular to a general neural network creation method, device, platform, medium and equipment. BACKGROUND

[0002] In some applications, the algorithm related to automatic driving is based on deep learning, and by collecting a large amount of data, the algorithm model is constantly iterated, so that the algorithm model is more accurate and the application scenarios covered are more extensive. From algorithm research and development to engineering and productization, once the algorithm model is iterated in the middle, it will lead to higher cost of intermediate data interaction and switching, and more time consumption.

[0003] The algorithm inference library containing multiple algorithm models runs better in the neural network, and takes less time, so how to develop a neural network that directly runs the algorithm inference library to get a more efficient conversion result is a problem that needs to be solved urgently. SUMMARY

[0004] In view of the above technical problems, the present application mainly provides a general neural network creation method, device, platform, medium and equipment.

[0005] In a first aspect, the embodiments of the present application provide a general neural network creation method, which comprises:

[0006] Processing parameters are set according to user's customized function, an algorithm configuration file is created, and the algorithm configuration file is integrated into a model package;

[0007] The data processing logic of the algorithm configuration file is set by using the processing parameters in one or more algorithm configuration files, and an algorithm inference library matching the model package is obtained;

[0008] A model package loading component is created according to the path of the model package, wherein the model package loading component acquires the model package;

[0009] The specified model in the model package is run through the model loading interface by using the model package loading component, and a default running device is set;

[0010] In the default running device, the specified model in the model package is run through the algorithm inference library, the user-defined data is processed to obtain network processing data, and a general neural network is created.

[0011] It can be understood that the general neural network creation method described above can be applied to devices including embedded devices. However, the embodiments of the present application do not limit the specific application devices thereof.

[0012] In a second aspect, the embodiments of the present application provide a general neural network creation device, which comprises:

[0013] a model package generation module configured to process parameters according to custom functions of a user, create an algorithm configuration file, and integrate the algorithm configuration file into a model package;

[0014] an algorithm inference library generation module configured to set data processing logic of the algorithm configuration file by using the processing parameters in the one or more algorithm configuration files, and obtain an algorithm inference library matched with the model package;

[0015] a model package acquisition module configured to create a model package loading component according to a path of the model package, wherein the model package loading component acquires the model package;

[0016] a model running module configured to run a specified model in the model package by using the model package loading component through a model loading interface, and set a default running device;

[0017] a network generation module configured to run the specified model in the model package through the algorithm inference library in the default running device, process user-defined data to obtain network processing data, and create a general neural network.

[0018] In a third aspect, an embodiment of the present application provides a general neural network creation platform, which comprises:

[0019] a model package loading component configured to load algorithm configuration files in a model package, and acquire a single model loading component to create a neural network, wherein the model package comprises an encrypted model package and / or a compressed model package;

[0020] a single model loading component configured to load a specified model in the model package, and encapsulate a loading process of an algorithm inference library;

[0021] an algorithm inference library component configured to load and run the algorithm configuration files in the model package;

[0022] a network object component configured to input user-defined data and output network processing data, and provide an interface for data interaction between a user and the general neural network;

[0023] a network component configured to run the general neural network, and provide various types of input interfaces and / or output interfaces.

[0024] In a fourth aspect, an embodiment of the present application provides a computer readable storage medium, which stores computer instructions, and the computer instructions are executed to cause a computer to perform the general neural network creation method in the first aspect.

[0025] In a fifth aspect, an embodiment of the present application provides a computer device, comprising a processor and a memory, the memory storing computer instructions, and the computer instructions are executed by the processor to implement the general neural network creation method in the first aspect.

[0026] The technical scheme of the embodiment of the present application integrates each algorithm model and its configuration file into a model package, and runs the algorithm model by using an algorithm inference library matched with the model package, inputs and outputs data required for network operation by calling various interfaces, and processes data by using a neural network. The general neural network proposed in the present scheme makes the algorithm inference library directly run, and the efficiency of obtaining a conversion result is higher, the effect is better, and the time for actual deployment of a vehicle is saved. BRIEF DESCRIPTION OF DRAWINGS

[0027] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings required in the embodiment description will be briefly introduced. Obviously, the drawings in the following description exemplarily show some embodiments of the present application.

[0028] Figure 1 is a schematic diagram of one specific embodiment of the general neural network creation method of the present application;

[0029] Figure 2 is a schematic diagram of one specific embodiment of the general neural network creation device of the present application.

[0030] The above drawings have shown the specific embodiments of the present application, and more detailed description will be given in the following. These drawings and textual descriptions are not intended to limit the scope of the concept of the present application by any means, but to illustrate the concept of the present application to those skilled in the art by referring to specific embodiments. DETAILED DESCRIPTION

[0031] The preferred embodiments of the present application will be described in detail below with reference to the drawings, so that the advantages and features of the present application can be more easily understood by those skilled in the art, and the protection scope of the present application can be more clearly and definitely defined.

[0032] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0033] The technical solutions of this application and how they solve the aforementioned technical problems will be described in detail below with specific embodiments. The specific embodiments described below can be combined with each other to form new embodiments. The same or similar ideas or processes described in one embodiment may not be repeated in other embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0034] Figure 1 This paper illustrates a specific implementation of a general neural network creation method according to this application.

[0035] exist Figure 1 In the specific implementation shown, a general neural network creation method mainly includes:

[0036] Step S101: Based on the user's customized function settings, the processing parameters are set, an algorithm configuration file is created, and the algorithm configuration file is integrated into the model package.

[0037] In this implementation, based on the user's request for data customization, product requirements are determined. Algorithm developers configure various algorithm models into corresponding algorithm configuration files, and integrate these models and their corresponding configuration files into a model package. This model package can be encrypted, compressed, and watermarked to prevent leakage for commercial purposes. The algorithm configuration file contains the parameters for configuring the algorithm; these parameters define the algorithm model.

[0038] In a specific instance of this application, there are many types of embedded devices, and the file formats of algorithm configuration files vary among different types of embedded devices. However, regardless of the file format, the algorithm configuration file essentially contains parameters; that is, the parameters determine the model. The model is equivalent to a function, which outputs the processing result after execution. Once the parameters are determined, the variables are also determined, ultimately defining the function. The models in the model package are single models trained for a specific scenario. Different models may have different accuracies, but their functions are all defined.

[0039] It should be noted that the general neural network creation method in the present application can be applied to devices including embedded devices. However, the embodiments of the present application do not limit the specific application devices thereof.

[0040] In a specific example of the present application, for the related items of traffic lights, multiple algorithm models may be configured for different detection functions, such as multiple algorithm configuration files for detecting the image path of traffic lights, image processing operations of traffic lights, etc., to form multiple corresponding algorithm models.

[0041] In Figure 1 In the specific embodiment shown, a general neural network creation method further includes:

[0042] In step S102, the data processing logic of the algorithm configuration file is set by using the processing parameters in one or more algorithm configuration files, and an algorithm inference library matched with the model package is obtained.

[0043] In the present embodiment, the algorithm inference library is essentially a program compiled from algorithm code, and the program contains the logic of the algorithm and can run each algorithm configuration file. The algorithm inference library is matched with the model package, and only the algorithm inference library in the present solution can run the model.

[0044] It should be noted that the algorithm inference library can not only contain a program compiled from algorithm code, but also contain a deep learning model matched with the program, i.e., a model matched in the model package.

[0045] In a specific example of the present application, the image captured by the camera is transmitted to the algorithm inference library. After the algorithm inference library obtains the image, the pre-processing unit is called through the function processing interface. After the pre-processing unit processes the image, the algorithm inference library reads the parameters in the algorithm configuration file, generates a specific function processing unit from the parameters, inputs the pre-processed image into the function processing unit through the corresponding interface, and runs to obtain the result.

[0046] In Figure 1 In the specific embodiment shown, a general neural network creation method further includes:

[0047] In step S103, a model package loading component is created according to the path of the model package, wherein the model package loading component obtains the model package.

[0048] In the present embodiment, the model package loading component is used to load the model package, supports loading of encrypted models, loading of compressed models, and also supports obtaining the operation pointer of a specified model in the model package.

[0049] In one specific embodiment of the present application, according to the path of the model package, the model package loading component is created, including: under the condition that the model contains a specific model name and an algorithm configuration file name corresponding to the specific model, the model package is parsed to obtain file data stream of the algorithm configuration file corresponding to the specific model; and according to the path of the model package and the file data stream, the model package loading component is created.

[0050] In the embodiment, whether the model package contains a model with a specific name is determined, and if so, an operation pointer of the specific model is obtained, which can be used to load a single model; whether the model package contains an algorithm configuration file with a specific name is determined, and if so, file data stream of the algorithm configuration file is obtained; and according to the file data stream and the path of the model package, a model package loading component object instance is created to read the model package.

[0051] In Figure 1 In the specific embodiment shown in the specific embodiment, a general neural network creation method further includes:

[0052] In step S104, the specified model in the model package is run by using the model package loading component through the model loading interface, and a default running device is set.

[0053] In the embodiment, through the model loading interface, the specified model is loaded according to the operation pointer of the single model obtained by the model package loading component, and is run in the default running device.

[0054] In one specific embodiment of the present application, the specified model in the model package is run by using the model package loading component through the model loading interface, and a default running device is set, including: a single model loading component of the specified model in the model package is obtained by using the model package loading component; file data stream of the algorithm configuration file corresponding to the specified model is obtained through the model loading interface to read the user-defined configuration file; the restriction information of the specified model is obtained by setting the image size of the network input and the limit batch size; the specified model is loaded by using the single model loading component to obtain model data information, and data preprocessing information is set; the single model loading component calls multiple device types of a list of executable devices to set the default running device.

[0055] In this embodiment, the single model loading component is a whole package, obtains a configuration reading object of an algorithm configuration file, resets the configuration key-value pairs read in the algorithm configuration file at runtime, sets the size of the image allowed to be input by the neural network and the limit batch size allowed to be input by the neural network, obtains limit information of the algorithm model, wherein the limit information includes the image size and the limit batch size, obtains an algorithm configuration file data stream for reading a user-defined configuration file, and the like. The single model loading component loads a specified model to obtain model-related information including a network object of an operating network, sets image preprocessing information, sets a plurality of executable device types, and creates the single model loading component according to the read model package.

[0056] In Figure 1 In the specific embodiment shown in the specific embodiment, a general neural network creation method further includes:

[0057] In step S105, the specified model in the model package is run in the default running device through the algorithm inference library, user-defined data is processed to obtain network processing data, and a general neural network is created.

[0058] In this embodiment, the specified model in the model package is run through the algorithm inference library, the neural network is run, the serialized model data is obtained, the input or output network object is obtained through various types of input or output interfaces, the neural network model is trained through data, and finally the general neural network is established, which can meet most of the needs of users.

[0059] In one specific embodiment of the present application, the specified model in the model package is run through the algorithm inference library, user-defined data is processed to obtain network processing data, and a general neural network is created, including: obtaining user-defined data; the algorithm inference library calls a preprocessing unit to preprocess the user-defined data to obtain preprocessed data; the specified model in the model package is run through the algorithm inference library to perform network configuration and network management, and a neural network is obtained; the neural network is trained through the preprocessed data, and network processing data is obtained, and a general neural network is obtained.

[0060] In this embodiment, by calling various interfaces, the single model loading component loads the specified model in the model package into the algorithm inference library, the algorithm inference library loads the algorithm configuration file in the specified model, network configuration and network management can be performed, after the network configuration and network management are created, the neural network is trained through the input interface to input user-defined data, thereby obtaining a general neural network, which can meet most of the needs of users.

[0061] In one specific embodiment of the present application, running a specified model in the model package through the algorithm inference library includes: the algorithm inference library calling processing parameters in the algorithm configuration file corresponding to the specified model in the model package to generate a data processing function unit, and running the data processing function unit to process data, wherein the algorithm configuration file contains processing parameters and inter-model relationships.

[0062] In this embodiment, the algorithm inference library reads and runs the parameters in the algorithm configuration file corresponding to the specified model, inputs the data required by the user through the input interface, and runs the parameter configuration of each layer to run the specified model.

[0063] In one specific embodiment of the present application, the neural network is trained by processing preprocessed data, and further includes: inputting preprocessed preprocessed data in different devices into a network object of the neural network; obtaining the shape of the network object and returning it through a shape interface provided by the algorithm inference library; connecting different network objects to form a neural network topology structure; wherein, for image data in the preprocessed data, the image data is synchronized from the central processing unit to the graphics processing unit, and the image data buffered in the central processing unit is cleaned up; for point cloud data in the preprocessed data, the point cloud data is synchronized to the object network.

[0064] In this embodiment, the neural network is pre-trained by preprocessed data to obtain a layer of neural network layer specially for data preprocessing; the neural network is trained by preprocessed data for other layers to obtain three layers of neural network for data processing. The image data and the point cloud data are input into the network object for synchronous training, and finally various network objects are connected to form a network topology graph to obtain the neural network.

[0065] The present application sets different processing parameters according to different customization functions of users, creates corresponding algorithm configuration files, and integrates them into a model package; according to different customization needs of different users, product needs are determined, and algorithm developers configure corresponding algorithm configuration files for algorithm models with customized functions, and integrate each algorithm model and its corresponding algorithm configuration file into a model package. The model package can be encrypted, compressed, and watermarked to prevent commercial behavior from being leaked. One or more algorithm configuration files are used to set the data processing logic of the algorithm configuration file, and an algorithm inference library matching the model package is obtained; the algorithm inference library matches the model package, and the essence of the algorithm inference library is a program compiled from algorithm code, which contains the logic of the algorithm and can run each algorithm configuration file. According to the path of the model package, a model package loading component is created, which reads the model package; whether the model package contains a specific named model, if so, the operation pointer of the specific model is obtained, which can be used to load a single model; whether the model package contains an algorithm configuration file with a specific name, if so, the file data stream of the algorithm configuration file is obtained; according to the file data stream and the path of the model package, a model package loading component object instance is created to read the model package. Using the model package loading component, the specified model in the model package is run through the model loading interface, and the default running device is set; the single model loading component is a whole encapsulation, the configuration reading object of the algorithm configuration file is obtained, and the configuration key-value pairs read in the algorithm configuration file are re-set during running; the size of the image allowed to be input by the neural network and the limit batch size allowed to be input by the neural network are set; the limit information of the algorithm model is obtained, which includes the image size and the limit batch size; the algorithm configuration file data stream is obtained, which is used to read the user-defined configuration file, etc.; the single model loading component loads the specified model to obtain model-related information, which includes a network object for operating the network; image preprocessing information is set, and multiple executable device types are set; a single model loading component is created according to the read model package. In the default running device, the specified model in the model package is run through the algorithm inference library, user-defined data and network processing data are obtained through the input or output interface, and a general neural network is created; through the model loading interface, the single model loading component loads the specified model in the model package into the algorithm inference library, the algorithm inference library loads the algorithm configuration file in the specified model, and network configuration and network management can be performed. After network configuration and network management are created, the neural network is trained through the input interface to obtain a general neural network that can meet most of the user's needs.

[0066] Figure 2 A specific embodiment of a general neural network creation device of the present application is shown.

[0067] In Figure 2In the specific embodiment shown, a general neural network creation device mainly comprises:

[0068] a model package generation module 201 for processing parameters according to a user's customized function setting, creating an algorithm configuration file, and integrating the algorithm configuration file into a model package;

[0069] an algorithm inference library generation module 202 for setting data processing logic of the algorithm configuration file by using the processing parameters in one or more algorithm configuration files, and obtaining an algorithm inference library matched with the model package;

[0070] a model package acquisition module 203 for creating a model package loading component according to a path of the model package, wherein the model package loading component acquires the model package;

[0071] a model running module 204 for running a specified model in the model package by a model loading interface by using the model package loading component, and setting a default running device;

[0072] a network generation module 205 for running the specified model in the model package by the algorithm inference library in the default running device, processing user-defined data to obtain network processing data, and creating a general neural network.

[0073] In the embodiment, the model package generation module 201 determines product requirements according to a user's customized function on data, algorithm development personnel configure corresponding algorithm configuration files for various algorithm models, and integrate the various algorithm models and their corresponding algorithm configuration files into the model package. The model package can be encrypted, compressed, and watermarked to prevent commercial behavior from being leaked. The algorithm inference library generation module 202 is essentially a program compiled from algorithm code, which contains the logic of the algorithm and can run various algorithm configuration files. The model package acquisition module 203 is a model package loading component for loading the model package, supporting loading of encrypted models, loading of compressed models, and supporting operation pointers of specified models in the model package. Through a model loading interface, a specific specified model is loaded according to the operation pointer of the single model acquired by the model package loading component, and is run in a default running device. Through the model loading interface, the single model loading component loads the specified model in the model package into the algorithm inference library, the algorithm inference library loads the algorithm configuration file in the specified model, and can perform network configuration and network management. After the network configuration and network management are created, the neural network is trained by inputting user-defined data through an input interface, thereby obtaining a general neural network.

[0074] The general neural network creation device provided in the present application can be used to execute the general neural network creation method described in any of the embodiments, and has similar implementation principles and technical effects, which will not be described here again.

[0075] In one embodiment of the present application, the functional modules of the general neural network creation apparatus of the present application can be directly in hardware, in software modules executed by a processor, or a combination of both.

[0076] The software modules can reside in RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium.

[0077] The processor can be a central processing unit (CPU), a general purpose processor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any conventional processor, controller, microcontroller, or state machine. The processor can also be implemented as a combination of computing devices, such as a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0078] In another embodiment of the present application, a general neural network creation platform comprises:

[0079] a model package loading component for loading an algorithm configuration file in a model package, and obtaining a single model loading component to create a neural network, wherein the model package comprises an encrypted model package and / or a compressed model package;

[0080] a single model loading component for loading a specified model in the model package, and encapsulating a loading process of an algorithm inference library;

[0081] an algorithm inference library component for loading and running the algorithm configuration file in the model package;

[0082] a network object component for inputting user-defined data and outputting network processing data, and providing an interface for the user to interact with the general neural network;

[0083] a network component for running the general neural network, and providing various types of input interfaces and / or output interfaces.

[0084] In this embodiment, the model package loading component supports loading of encrypted models and / or compressed models, reads an algorithm configuration file of the model package, obtains an operation pointer of a single model loading component so that the single model loading component loads a specific model in the model package; the single model loading component obtains a configuration object of the algorithm configuration file through encapsulation of a loading process of an algorithm inference library, obtains restriction information and file data flow of the algorithm model, and finally successfully loads the specified model; the algorithm inference library component is used in conjunction with the model package, and is specially used for running the single model in the model package; the network object component trains the neural network through interactive data, and runs the neural network through the network component, and finally trains a platform that can run the general neural network.

[0085] In another specific embodiment of the present application, a computer readable storage medium stores computer instructions, which, when executed, cause a computer to perform the general neural network creation method in any embodiment.

[0086] In another specific embodiment of the present application, a computer device includes a processor and a memory, and the memory stores computer instructions, which, when executed by the processor, implement the general neural network creation method in any embodiment.

[0087] In the several embodiments of the present application, it should be understood that the disclosed device and method can be implemented in other ways. For example, the device embodiments described above are only schematic. The division of the units is only a logical function division. There can be another division manner for actual implementation, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units shown or discussed can be indirect coupling or communication connection through some interface, device or unit, and can be electrical, mechanical or in other forms.

[0088] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., they can be located in one place, or can be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0089] The above merely describes the embodiments of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation or direct or indirect application in other related technical fields based on the content of the present application specification and drawings is also included in the patent protection scope of the present application.

Claims

1. A general neural network creation method, applied to embedded devices, characterized in that, include: The algorithm model parameters are set according to the user's customized function settings, an algorithm configuration file is created, and the algorithm configuration file is integrated into the model package; By using the parameters of the algorithm model in one or more of the algorithm configuration files, the data processing logic of the algorithm configuration files is set to obtain an algorithm inference library that matches the model package, wherein the algorithm inference library contains programs and deep learning models that reflect the algorithm logic; Based on the path of the model package, a model package loading component is created, wherein the model package loading component obtains the model package; Using the model package loading component, a specified model in the model package is run through the model loading interface, and a default running device is set; In the default running device, the specified model in the model package is run through the algorithm inference library to process user-defined data to obtain network processing data and create a general neural network.

2. The general neural network creation method as described in claim 1, characterized in that, The process of running the specified model in the model package through the algorithm inference library, processing user-defined data to obtain network processing data, and creating a general neural network includes: Obtain the user-defined data; The algorithm inference library calls the preprocessing unit to preprocess the user-defined data to obtain preprocessed data; The specified model in the model package is run using the algorithm inference library to perform network configuration and network management, thereby obtaining a neural network; The neural network is trained by processing the preprocessed data, and the network processing data is obtained to obtain the general neural network.

3. The general neural network creation method as described in claim 1, characterized in that, The step of creating a model package loading component based on the path of the model package includes: Given that the model contains a specific model name and the name of the corresponding algorithm configuration file, the model package is parsed to obtain the file data stream of the algorithm configuration file corresponding to the specific model. Based on the path of the model package and the file data stream, create the model package loading component.

4. The general neural network creation method as described in claim 1, characterized in that, The step of using the model package loading component to run a specified model in the model package through the model loading interface and setting the default running device includes: Use the model package loading component to obtain a single model loading component for the specified model in the model package; The file data stream of the algorithm configuration file corresponding to the specified model is obtained through the model loading interface in order to read the user-defined configuration file; By setting the image size and batch size limit for network input, the limitation information of the specified model can be obtained; The specified model is loaded through the single model loading component to obtain model data information, and data preprocessing information is set. The single model loading component calls multiple device types from the list of runnable devices and sets the default running device.

5. The general neural network creation method as described in claim 2, characterized in that, The step of training the neural network by processing the preprocessed data further includes: The preprocessed data from different devices is input into the network object of the neural network; Obtain the shape of the network object and return it using the shape interface provided by the algorithm inference library; Connect the different network objects to form a neural network topology; Specifically, for the image data in the preprocessed data, the image data is synchronized from the central processing unit to the graphics processing unit, and the image data cached in the central processing unit is cleared. The point cloud data in the preprocessed data is synchronized to the network object.

6. The general neural network creation method as described in claim 1, characterized in that, Running the specified model in the model package through the algorithm inference library includes: The algorithm inference library calls the parameters of the algorithm model in the algorithm configuration file corresponding to the specified model in the model package, generates a data processing function unit, and runs the data processing function unit to process the data. The algorithm configuration file contains the parameters of the algorithm model and the relationships between models.

7. A universal neural network creation device, characterized in that, include: The model package generation module is used to set the parameters of the algorithm model according to the user's customized function settings, create an algorithm configuration file, and integrate the algorithm configuration file into the model package; An algorithm inference library generation module is used to set the data processing logic of the algorithm configuration file by utilizing the parameters of the algorithm model in one or more algorithm configuration files to obtain an algorithm inference library that matches the model package, wherein the algorithm inference library contains a program and a deep learning model that reflect the algorithm logic; A model package acquisition module is used to create a model package loading component based on the path of the model package, wherein the model package loading component acquires the model package; The model execution module is used to utilize the model package loading component to run a specified model in the model package through the model loading interface, and to set the default running device; The network generation module is used to run the specified model in the model package through the algorithm inference library in the default running device, process user-defined data to obtain network processing data, and create a general neural network.

8. A general neural network creation system, characterized in that, include: A model package loading component is used to load algorithm configuration files from a model package and obtain a single model loading component to create a neural network, wherein the model package includes encrypted model packages and / or compressed model packages, and the model package contains different algorithm models and the algorithm configuration files; The single model loading component is used to load a specified model in the model package and encapsulate the loading process of the algorithm inference library. The single model loading component obtains the configuration object of the algorithm configuration file, obtains the limitation information and file data stream of the algorithm model, and loads it into the specified model by encapsulating the loading process of the algorithm inference library. The algorithm inference library is a component used to load and run the algorithm configuration files in the model package; The network object component is used to input user-defined data and output network-processed data, and provides an interface for users to interact with general neural networks. A network component that runs the general neural network and provides various types of input and / or output interfaces.

9. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed, the computer performs the general neural network creation method as described in any one of claims 1-6.

10. A computer device comprising a processor and a memory, the memory storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the general neural network creation method as described in any one of claims 1-6.

Images