WO2021259039A1

WO2021259039A1 - Neural network model customization method, system and device, and storage medium

Info

Publication number: WO2021259039A1
Application number: PCT/CN2021/098288
Authority: WO
Inventors: 熊超; 蔡权雄; 牛昕宇
Original assignee: 深圳鲲云信息科技有限公司
Priority date: 2020-06-22
Filing date: 2021-06-04
Publication date: 2021-12-30
Also published as: CN111753983A

Abstract

A neural network model customization method, system and device, and a storage medium. The neural network model customization method comprises: acquiring a preset neural network model (S110); converting the neural network model into a static computation graph model (S120); constructing a directed cyclic computation graph model according to node information of a first computation node of the static computation graph model (S130); converting the directed cyclic computation graph model into an intermediate representation computation graph by means of a preset graph parsing engine (S140); and generating a customized target neural network model according to the intermediate representation computation graph (S150).

Description

Customized method, system, equipment and storage medium of neural network model

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office with an application number of 202010575490.3 on June 22, 2020. The entire content of this application is incorporated into this application by reference.

Technical field

This application relates to the field of neural network technology, for example, to a method, system, device, and storage medium for customizing a neural network model.

Background technique

With the development of the field of deep learning, more and more companies have launched their own deep learning frameworks. Different deep learning frameworks have their own advantages, some are suitable for research, some are suitable for industrial deployment, and so on.

The deep learning neural network is realized through the data flow calculation graph. The data (tensor, etc.) flows from the input node through the collection nodes, and finally the inference result is obtained at the output node. Different frameworks implement data flow calculation graphs in different ways. Some use static calculation graphs, and some are dynamically created during operation. Either way, the structure calculation diagram and weights of a neural network model will eventually be obtained. Different frameworks have their own model deployment methods. For artificial intelligence chip manufacturers, how to analyze and deploy multiple deep learning frameworks more easily is a key issue in their development chip tool chain. For the conversion of different deep learning frameworks, it is usually necessary to analyze the neural network models of different frameworks, generate a specific intermediate representation (Intermediate Representation), customize the intermediate representation and deploy it to the hardware device.

Different framework characteristics and data structure differences of models require developers to develop multiple analytical front ends. The front-end analytical modes of different frameworks are not uniform, resulting in cumbersome and inefficient front-end development processes, which are not conducive to expansion and high maintenance costs.

Summary of the invention

This application provides a method, system, equipment, and storage medium for customizing a neural network model to realize the generation of a custom neural network model to be suitable for different deep learning frameworks.

Provides a method for customizing a neural network model, which includes:

Obtain a preset neural network model; convert the neural network model into a static calculation graph model; construct a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model; pass the preset The graph analysis engine converts the directed acyclic calculation graph model into an intermediate expression calculation graph; and generates a customized target neural network model according to the intermediate expression calculation graph.

A customized system of neural network model is also provided, which includes:

The model acquisition module is set to obtain a preset neural network model; the model conversion module is set to convert the neural network model into a static calculation graph model; the model construction module is set to perform the first calculation according to the static calculation graph model The node information of the nodes constructs a directed acyclic calculation graph model; the calculation graph conversion module is configured to convert the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine; the model generation module is set To generate a customized target neural network model according to the intermediate expression calculation graph.

A customized device for a neural network model is also provided. The device includes: one or more processors; a storage device configured to store one or more programs. Execution by each processor, so that the one or more processors implement the above-mentioned customized method of the neural network model.

A computer-readable storage medium is also provided, on which a computer program is stored, and when the program is executed by a processor, the method for customizing the above-mentioned neural network model is realized.

Description of the drawings

FIG. 1 is a schematic flowchart of a method for customizing a neural network model provided in Embodiment 1 of the present application;

FIG. 2 is a schematic flowchart of a method for customizing a neural network model provided in Embodiment 2 of the present application;

FIG. 3 is a schematic flowchart of step S230 in FIG. 2 according to the second embodiment of the present application;

FIG. 4 is a schematic flowchart of step S260 in FIG. 2 according to the second embodiment of the present application;

FIG. 5 is a schematic structural diagram of a customized system for a neural network model provided in Embodiment 3 of the present application;

Fig. 6 is a schematic structural diagram of a neural network model customized device provided in the fourth embodiment of the present application.

detailed description

The application will be described below with reference to the drawings and embodiments.

It should be mentioned before discussing the exemplary embodiments that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowchart describes multiple steps as sequential processing, many of the steps can be implemented in parallel, concurrently, or simultaneously. In addition, the order of multiple steps can be rearranged. The processing may be terminated when its operations are completed, but may also have additional steps not included in the drawings. Processing can correspond to methods, functions, procedures, subroutines, subroutines, and so on.

In addition, the terms "first", "second", etc. may be used herein to describe various directions, actions, steps or elements, etc., but these directions, actions, steps or elements are not limited by these terms. These terms are only used to distinguish a first direction, action, step or element from another direction, action, step or element. For example, without departing from the scope of the present application, the first computing node may be referred to as the second computing node, and similarly, the second computing node may be referred to as the first computing node. Both the first computing node and the second computing node are computing nodes, but they are not the same computing node. The terms "first", "second", etc. cannot be understood as indicating or implying relative importance or implicitly indicating the number of indicated technical features. Therefore, the features defined with “first” and “second” may explicitly or implicitly include one or more of the features. In the description of the embodiments of the present application, "a plurality of" means at least two, such as two, three, etc., unless otherwise defined.

Example one

As shown in Figure 1, Embodiment 1 of the present application provides a method for customizing a neural network model, which includes:

S110. Obtain a preset neural network model.

S120: Convert the neural network model into a static calculation graph model.

In this embodiment, it is first necessary to obtain a preset neural network model, that is, a neural network model of a customized deep learning framework, and then convert the neural network model into a static computational graph model. Among them, the static computational graph model is a pre-defined computational graph model that cannot be modified during the inference process. Illustratively, the static computational graph model can be a data flow-based programming that removes the training-related data and retains only the inference-related data. The symbolic mathematics system (tensorflow) static calculation graph model can also be the open source Python machine learning library (pyotrch) static calculation graph model exported using pyotrch's official deployment tool (jit).

S130. Construct a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model.

S140: Convert the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine.

S150. Generate a customized target neural network model according to the intermediate expression calculation graph.

In this embodiment, after obtaining the static calculation graph model, it is also necessary to obtain the node information of the first calculation node of the static calculation graph model. The node information includes operators, operator attributes, model parameters, and edge relations, etc., where the operator In order to complete a kind of operation node, such as the node that completes addition, subtraction, multiplication and division, the large operator can be a combination of multiple mathematical operations, such as convolution operation and pooling operation; the attribute of the operator means that some operator nodes need to be specified Some parameters of the convolution operation, such as the kernel size (kernel_size) and stride length (strides), etc.; model parameters refer to the trainable parameters of the neural network model during training; the edge relationship refers to the input of the operator node The output relationship, the output of the predecessor node is the input of the node, which is generally represented by a string. Then according to the node information of the first computing node, a preset tool is used to perform node mapping, corresponding to new nodes, adding node information, and adding edge relationships, thereby constructing a directed acyclic computing graph model. After the directed acyclic calculation graph model is obtained, multiple analysis rules are predefined in the preset graph analysis engine according to user needs, and these analysis rules are repeatedly iterated to match, write and replace the directed acyclic calculation graph model. Thereby converted into an intermediate expression calculation graph. After obtaining the intermediate expression calculation graph, you can easily generate a customized target neural network model through the preset interface.

The embodiment of the application obtains a preset neural network model; converts the neural network model into a static calculation graph model; constructs a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model Convert the directed acyclic calculation graph model into an intermediate expression calculation graph through the preset graph analysis engine; generate a customized target neural network model according to the intermediate expression calculation graph, which solves the problem that the analytical front-end is not conducive to expansion and maintenance The problem of high cost realizes the effect of generating customized neural network models to apply to different deep learning frameworks. When analyzing and deploying different deep learning frameworks, you can first customize different neural network models corresponding to different deep learning frameworks to obtain a common model representation, and then deploy different deep learning frameworks on this basis. The number of analysis front ends can be reduced, and the analysis front ends can be expanded, and the maintenance cost is low.

Example two

As shown in Figure 2, the second embodiment of the present application provides a method for customizing a neural network model. The second embodiment of the present application is described on the basis of the first embodiment of the present application. The method includes:

S210. Obtain a preset neural network model.

S220: Convert the neural network model into a static calculation graph model.

S230: Construct a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model.

S240: Visually display the directed acyclic calculation graph model through a preset interface.

S250. Receive a user's modification to the directed acyclic calculation graph model.

In this embodiment, after obtaining the directed acyclic calculation graph model, the user can visualize the directed acyclic calculation graph model in the browser through the Application Programming Interface (API), and check the directed acyclic calculation graph through visualization Whether the node mapping of the model is correct, so as to receive the user's modification of the directed acyclic computational graph model.

S260: Convert the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine.

S270: Convert the intermediate expression calculation graph into a customized intermediate expression through a preset serialization interface.

S280. Serialize the intermediate expression into a customized target neural network model.

In this embodiment, after the intermediate expression calculation graph is obtained, the intermediate expression calculation graph can be converted into a customized intermediate expression through a preset serialization interface, and the intermediate expression is serialized into a customized target neural network model, thereby Completed the conversion of the preset neural network model into a customized target neural network model. Users can simply complete the neural network model extension to different deep learning frameworks, and it is easy to maintain.

As shown in FIG. 3, step S230 in the embodiment of the present application includes:

S231. Obtain node information of the first calculation node of the static calculation graph model, where the node information includes the number of calculation nodes and the order of the calculation nodes.

In this embodiment, in the process of constructing the directed acyclic calculation graph model, first obtain the node information of the first calculation node of the static calculation graph model, and use the python dictionary class to temporarily save the node information. The node information is in addition to operators, Operator attributes, model parameters and edge relations, as well as the number of calculation nodes and the order of calculation nodes. The number of calculation nodes includes the number of operators and the number of constant nodes. The order of calculation nodes is each first calculation node in the dictionary record of python. The successor node.

S232. Initialize the same number of second computing nodes as the number of computing nodes.

S233. Correspondingly add node information to the second computing node according to the topological sorting to construct an undirected computing graph model.

In this embodiment, after obtaining the number of computing nodes, initialize the same number of second computing nodes as the number of computing nodes, that is, add the same number of second computing nodes as the number of computing nodes, and then correspond the node information one-to-one according to the topological ordering Add to the second calculation node, add the operator name, operator attributes, model parameters, input and output and other node information for the second calculation node, thereby performing a one-to-one mapping of the original static calculation graph model to construct An undirected computational graph model with the same graph structure as the original static computational graph model.

S234: Add an edge relationship to the second computing node of the undirected computing graph model according to the sequence of computing nodes to construct a directed acyclic computing graph model.

In this embodiment, after obtaining the successor nodes of each first computing node, the second computing node of the undirected computing graph model can add edge relationships one by one for each edge based on this information. For example, according to the order of computing nodes Obtain the successor nodes of node a of the first computing node including node b and node c, then correspondingly add two edges ab and ac to node a of the second computing node, and add all the edge relationships to construct a directed acyclic Computational graph model.

As shown in FIG. 4, step S260 in the embodiment of the present application includes:

S261. Obtain parsing rules defined in a preset graph parsing engine, where the parsing rules include matching rules, rewriting rules, and assignment rules.

In this embodiment, in the process of constructing the intermediate expression calculation graph, it is first necessary to obtain the parsing rules defined in the preset graph parsing engine. The parsing rules can be pre-defined and stored in the graph parsing engine according to customized requirements. The rules include matching rules, rewriting rules and assignment rules.

S262: Match the third computing node of the directed acyclic computing graph model according to the matching rule.

In this embodiment, the third calculation node of the directed acyclic calculation graph model is a calculation node obtained by adding an edge relationship to the second calculation node of the undirected calculation graph model.

S263. When the matching is successful, a node connection relationship between the fourth computing node and the fourth computing node is added according to the rewriting rule.

S264: Assign a value to the fourth computing node according to the assignment rule and the third computing node that is successfully matched, and delete the third computing node that is successfully matched to obtain an intermediate expression calculation graph.

In this embodiment, there are multiple parsing rules. One parsing rule can be iterated first, and the third computing node of the directed acyclic computing graph model is matched according to the matching rule in the parsing rule, and the special attributes of the third computing node are matched. , Or match the topological connection relationship of the subgraph. When the matching is successful, the graphical structure of the fourth computing node is determined according to the rewriting rule in the analysis rule, and then the node connection relationship between the fourth computing node and the fourth computing node is added. Finally, the fourth computing node is assigned according to the assignment rule in the parsing rule and the successfully matched third computing node. Determine the part of the third computing node that needs to be retained or replaced according to the assignment rules, determine the retention or rewrite attribute of the third computing node, and then determine the part that needs to be retained in the third computing node that is successfully matched, and the user-defined content The fourth computing node performs the assignment, and the third computing node that successfully matches is deleted. Continue to iterate the next parsing rule until the iteration of all parsing rules is completed, thereby generating an intermediate expression calculation graph according to the obtained fourth calculation node.

The embodiment of the application obtains the node information of the first computing node of the static computing graph model. The node information includes the number of computing nodes and the sequence of computing nodes; initializes the same number of second computing nodes as the number of computing nodes; and corresponds the node information according to the topological ordering Add to the second calculation node to construct an undirected calculation graph model; add edge relations to the second calculation node of the undirected calculation graph model according to the order of the calculation nodes to construct a directed acyclic calculation graph model, and obtain the definition in the preset The analysis rules of the graph analysis engine, the analysis rules include matching rules, rewriting rules and assignment rules; according to the matching rules, the third computing node of the directed acyclic computing graph model is matched; when the matching is successful, the fourth is added according to the rewriting rules The node connection relationship between the computing node and the fourth computing node; the fourth computing node is assigned according to the assignment rules and the successfully matched third computing node, and the successfully matched third computing node is deleted to obtain the intermediate expression calculation graph, which solves the analysis The front-end is not conducive to the problem of expansion and high maintenance costs, and realizes the effect of generating customized neural network models to apply to different deep learning frameworks.

Example three

As shown in FIG. 5, the third embodiment of the present application provides a neural network model customization system 100. The neural network model customization system 100 provided in the third embodiment of the present application can execute the system 100 provided by any embodiment of the present application. The customized method of the neural network model has the corresponding functional modules and effects of the execution method. The neural network model customization system 100 includes a model acquisition module 200, a model conversion module 300, a model construction module 400, a calculation graph conversion module 500, and a model generation module 600.

The model obtaining module 200 is set to obtain a preset neural network model; the model conversion module 300 is set to convert the neural network model into a static calculation graph model; the model building module 400 is set to obtain node information of the first calculation node of the static calculation graph model A directed acyclic calculation graph model is constructed; the calculation graph conversion module 500 is set to convert the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine; the model generation module 600 is set to calculate the graph based on the intermediate expression Generate customized target neural network model.

In this embodiment, the model building module 400 is configured to obtain node information, where the node information includes the number of computing nodes and the sequence of computing nodes; an undirected computing graph model is constructed according to the number of computing nodes; an undirected computing graph model is constructed according to the sequence of computing nodes The second computing node of add edge relationship to construct a directed acyclic computing graph model.

The model building module 400 is configured to construct an undirected computing graph model according to the number of computing nodes in the following manner: initialize the same number of second computing nodes as the number of computing nodes; sort the operator names and operators included in the node information according to the topological order The attributes, model parameters, inputs, and outputs are correspondingly added to the same number of second calculation nodes as the number of calculation nodes to construct an undirected calculation graph model.

The neural network model customization system 100 further includes a model display module 700, which is configured to visually display the directed acyclic calculation graph model through a preset interface; and receive the user's modification of the directed acyclic calculation graph model.

In this embodiment, the calculation graph conversion module 500 is configured to obtain the analysis rules defined in the preset graph analysis engine; replace the third calculation node of the directed acyclic calculation graph model according to the analysis rules to obtain the intermediate expression calculation graph, Wherein, the third calculation node of the directed acyclic calculation graph model is a calculation node obtained by adding the edge relationship to the second calculation node of the undirected calculation graph model.

The parsing rules include matching rules, rewriting rules, and assignment rules. The calculation graph conversion module 500 is configured to replace the third calculation node of the directed acyclic calculation graph model according to the parsing rules in the following manner to obtain the intermediate expression calculation graph: according to the matching rules Match the third computing node of the directed acyclic computing graph model; when the third computing node is successfully matched, the node connection relationship between the fourth computing node and the fourth computing node is added according to the rewriting rule; according to the assignment rule and the matching success The third computing node assigns a value to the fourth computing node, and deletes the third computing node that is successfully matched to obtain an intermediate expression calculation graph.

The model generation module 600 is configured to convert the intermediate expression calculation graph into a customized intermediate expression through a preset serialization interface; serialize the intermediate expression into a customized target neural network model.

Embodiment four

Fig. 6 is a schematic structural diagram of a neural network model customized device provided in the fourth embodiment of the present application. FIG. 6 shows a block diagram of an exemplary computer device 12 suitable for implementing the embodiments of the present application. The computer device 12 shown in FIG. 6 is only an example, and should not bring any limitation to the function and scope of use of the embodiments of the present application.

As shown in FIG. 6, the computer device 12 is represented in the form of a general-purpose computing device. The components of the computer device 12 may include, but are not limited to: one or more processors or processing units 16, a system memory 28, and a bus 18 connecting different system components (including the system memory 28 and the processing unit 16).

The bus 18 represents one or more of several types of bus structures, including a memory bus or a memory controller, a peripheral bus, a graphics acceleration port, a processor, or a local bus using any bus structure among multiple bus structures. For example, these architectures include but are not limited to Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MAC) bus, enhanced ISA bus, Video Electronics Standards Association (Video Electronics Standards) Association, VESA) local bus and Peripheral Component Interconnect (PCI) bus.

The computer device 12 includes a variety of computer system readable media. These media can be any available media that can be accessed by the computer device 12, including volatile and nonvolatile media, removable and non-removable media.

The system memory 28 may include a computer system readable medium in the form of a volatile memory, such as a random access memory (RAM) 30 and/or a cache memory 32. The computer device 12 may include other removable/non-removable, volatile/nonvolatile computer system storage media. For example only, the storage system 34 may be configured to read and write a non-removable, non-volatile magnetic medium (not shown in FIG. 6, usually referred to as a "hard drive"). Although not shown in FIG. 6, a disk drive configured to read and write to a removable non-volatile disk (such as a "floppy disk"), and a removable non-volatile optical disk (such as a compact disc read-only memory (Compact Disc)) can be provided. Read-Only Memory, CD-ROM), Digital Versatile Disc-ROM, DVD-ROM or other optical media for reading and writing optical disc drives. In these cases, each drive can pass one or Multiple data medium interfaces are connected to the bus 18. The memory 28 may include at least one program product having a set of (for example, at least one) program modules, and these program modules are configured to perform the functions of the embodiments of the present application.

A program/utility tool 40 having a set of (at least one) program module 42 may be stored in, for example, the memory 28. Such program module 42 includes, but is not limited to, an operating system, one or more application programs, and other programs Modules and program data, each of these examples or a combination may include the realization of a network environment. The program module 42 usually executes the functions and/or methods in the embodiments described in this application.

The computer device 12 can also communicate with one or more external devices 14 (such as keyboards, pointing devices, displays 24, etc.), and can also communicate with one or more devices that enable users to interact with the computer device 12, and/or communicate with Any device (such as a network card, modem, etc.) that enables the computer device 12 to communicate with one or more other computing devices. This communication can be performed through an input/output (Input/Output, I/O) interface 22. In addition, the computer device 12 may also communicate with one or more networks (for example, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet) through the network adapter 20. As shown in the figure, the network adapter 20 communicates with other modules of the computer device 12 through the bus 18. Although not shown in the figure, other hardware and/or software modules can be used in conjunction with the computer device 12, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, and disk arrays (Redundant Arrays of Independent Disks). , RAID) systems, tape drives, and data backup storage systems.

The processing unit 16 executes a variety of functional applications and data processing by running programs stored in the system memory 28, for example, to implement the methods provided in the embodiments of the present application:

Embodiment five

The fifth embodiment of the present application also provides a computer-readable storage medium on which a computer program is stored, and when the program is executed by a processor, the method as provided in all the application embodiments of the present application is implemented:

The computer storage medium of the embodiment of the present application may adopt any combination of one or more computer-readable media. The computer-readable medium may be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or device, or a combination of any of the above. Examples of computer-readable storage media (non-exhaustive list) include: electrical connections with one or more wires, portable computer disks, hard disks, RAM, ROM, Erasable Programmable Read-Only Memory, EPROM or flash memory), optical fiber, CD-ROM, optical storage device, magnetic storage device, or any suitable combination of the above. In this document, the computer-readable storage medium can be any tangible medium that contains or stores a program, and the program can be used by or in combination with an instruction execution system, apparatus, or device.

The computer-readable signal medium may include a data signal propagated in baseband or as a part of a carrier wave, and computer-readable program code is carried therein. This propagated data signal can take many forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the foregoing. The computer-readable signal medium may also be any computer-readable medium other than the computer-readable storage medium. The computer-readable medium may send, propagate or transmit the program for use by or in combination with the instruction execution system, apparatus, or device .

The program code contained on the computer-readable medium can be transmitted by any suitable medium, including but not limited to wireless, wire, optical cable, radio frequency (RF), etc., or any suitable combination of the foregoing.

The computer program code used to perform the operations of this application can be written in one or more programming languages or a combination thereof. The programming languages include object-oriented programming languages—such as Java, Smalltalk, C++, and also conventional Procedural programming language-such as "C" language or similar programming language. The program code can be executed entirely on the user's computer, partly on the user's computer, executed as an independent software package, partly on the user's computer and partly executed on a remote computer, or entirely executed on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any kind of network including LAN or WAN, or may be connected to an external computer (for example, using an Internet service provider to connect through the Internet).

Claims

A method for customizing neural network models, including:

Obtain a preset neural network model;

Converting the neural network model into a static calculation graph model;

Constructing a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model;

Converting the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine;

A customized target neural network model is generated according to the intermediate expression calculation graph.
The method according to claim 1, wherein the constructing a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model comprises:

Acquiring the node information, where the node information includes the number of computing nodes and the order of computing nodes;

Constructing an undirected calculation graph model according to the number of computing nodes;

Adding an edge relationship to the second computing node of the undirected computing graph model according to the sequence of computing nodes to construct the directed acyclic computing graph model.
The method according to claim 2, wherein said constructing an undirected calculation graph model according to the number of calculation nodes comprises:

Initialize the same number of second computing nodes as the number of said computing nodes;

According to the topological sorting, the operator names, operator attributes, model parameters, inputs, and outputs that are also included in the node information are correspondingly added to the second computing nodes of the same number as the number of computing nodes to construct the To the computational graph model.
The method according to claim 1, after constructing a directed acyclic calculation graph model based on the node information of the first calculation node of the static calculation graph model, further comprising:

Visually display the directed acyclic calculation graph model through a preset interface;

Receive a user's modification to the directed acyclic calculation graph model.
The method according to claim 1, wherein the converting the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine comprises:

Acquiring the parsing rules defined in the preset graph parsing engine;

The third calculation node of the directed acyclic calculation graph model is replaced according to the analysis rule to obtain the intermediate expression calculation graph, wherein the third calculation node of the directed acyclic calculation graph model is The calculation node obtained by adding the edge relationship to the second calculation node of the undirected calculation graph model.
The method according to claim 5, wherein the parsing rules include matching rules, rewriting rules, and assignment rules;

The replacing the third calculation node of the directed acyclic calculation graph model according to the analysis rule to obtain the intermediate expression calculation graph includes:

Matching the third computing node of the directed acyclic computing graph model according to the matching rule;

In the case that the third computing node is successfully matched, a node connection relationship between the fourth computing node and the fourth computing node is newly added according to the rewriting rule;

Assign a value to the fourth computing node according to the assignment rule and the successfully matched third computing node, and delete the successfully matched third computing node to obtain the intermediate expression computing graph.
The method according to claim 1, wherein the generating a customized target neural network model according to the intermediate expression calculation graph comprises:

Converting the intermediate expression calculation graph into a customized intermediate expression through a preset serialization interface;

The intermediate expression is serialized into the customized target neural network model.
A customized system of neural network model, including:

The model acquisition module is set to acquire the preset neural network model;

A model conversion module, configured to convert the neural network model into a static calculation graph model;

A model construction module, configured to construct a directed acyclic calculation graph model according to the node information of the first calculation node of the static calculation graph model;

The calculation graph conversion module is configured to convert the directed acyclic calculation graph model into an intermediate expression calculation graph through a preset graph analysis engine;

The model generation module is configured to generate a customized target neural network model according to the intermediate expression calculation graph.
A customized equipment of neural network model, including:

At least one processor;

The storage device is set to store at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the neural network model customization method according to any one of claims 1-7.
A computer-readable storage medium storing a computer program, wherein when the program is executed by a processor, the neural network model customization method according to any one of claims 1-7 is realized.