WO2024046463A1 - Model construction method, apparatus and platform, electronic device and storage medium - Google Patents

Abstract

Embodiments of the present application provide a model construction method, apparatus and platform, an electronic device and a storage medium. The method comprises: in a model description stage, obtaining a source model of a neural network model on the basis of model description information, the model description data comprising information of a first model and/or a pre-stored basic model, the basic model and/or the first model being used for constructing the neural network model, and the first model being a user-defined model; and when model conversion needs to be performed, entering a model conversion stage to convert the source model into an executable model. The technical solution of the embodiments of the present application reduces the complexity of neural network model construction.

Description

Model construction method, device, platform, electronic equipment and storage medium

This application requires the priority of the Chinese patent application submitted to the China Patent Office on September 2, 2022, with the application number 202211068160.0, and the invention name is "Model Construction Method, Device, Platform, Electronic Equipment and Storage Medium", and its entire content is approved by This reference is incorporated into this application.

Technical field

The present application relates to the field of data processing technology, and in particular to a model construction method, device, platform, electronic equipment and storage medium.

Background technique

The brain-like computing system is a new high-performance computing hardware and software system that imitates the structure and working principles of the brain's biological nervous system and supports brain-like high-performance parallel computing. Brain-like intelligence and cognitive computing are based on spiking neural networks, combining a variety of neurotransmitters, neuromodulators, receptors, electrical synapses, chemical synapses, dendrites, neurons, and glial cells in the biological brain. Enrich working mechanisms for computational modeling, and construct neural circuits, neural nuclei, brain regions and whole-brain models that can simulate many cognitive mechanisms and behaviors of the biological brain.

The inventor found that the current method of building a neural network model is relatively complex and has low user experience. Therefore, a model construction method is needed to solve the above technical problems.

technical problem

The main purpose of this application is to solve the existing technical problems of relatively complex methods of building neural network models and low user experience.

Technical solutions

Embodiments of the present application provide a model construction method, device, platform, electronic equipment and storage medium, which realizes the description and conversion of neural network models and simplifies the complexity of constructing neural network models.

In the first aspect, embodiments of the present application provide a model construction method, including:

In the model description stage, the source model of the neural network model is obtained based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or the first model are For constructing a neural network model; the first model is a user-defined model;

When model conversion is required, the model conversion stage is entered to convert the source model into an executable model.

In a second aspect, embodiments of the present application also provide a model building device, which includes:

A model description module, used in the model description stage to obtain the source model of the neural network model based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or The first model is used to build a neural network model; the first model is a user-defined model;

The model conversion module is used to enter the model conversion stage and convert the source model into an executable model when model conversion is required.

In a third aspect, embodiments of the present application also provide a model construction platform, which executes the model construction method of the embodiments of the present application. The platform includes: a front-end part, a core part and a back-end part, wherein:

The front-end part is used to act as an agent outside the platform to interact with the core part and/or the back-end part;

The core part is used to respond to requests from the front-end part and/or the back-end part. In the model description phase, obtain the source model of the neural network model based on the model description information. The model description data includes the first model and /or pre-stored basic model information; the basic model and/or the first model are used to build a neural network model; the first model is a user-defined model;

The backend part is used to enter the model conversion stage when model conversion is required, and The source model is converted into an executable model.

In a fourth aspect, embodiments of the present application further provide an electronic device, where the electronic device includes:

one or more processors;

a storage device for storing one or more programs,

When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the model building method in any embodiment of the present application.

In a fifth aspect, embodiments of the present application also provide a storage medium containing computer-executable instructions, which when executed by a computer processor are used to perform the model building method in any embodiment of the present application. .

beneficial effects

Through the technical solution of the embodiment of the present application, in the model description stage, the source model of the neural network model is obtained based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or the third A model is used to build a neural network model. The first model is a user-defined model. When model conversion is required, enter the model conversion stage to convert the source model into an executable model. The technical solution of the embodiment of the present application simplifies the complexity of model construction, improves the model construction speed, and thereby improves user experience.

Description of drawings

In order to explain the embodiments of the present application or the technical solutions in the prior art more clearly, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only These are some embodiments of the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without exerting creative efforts.

in:

Figure 1 is a schematic flow chart of a model construction method in an embodiment;

Figure 2 is a schematic structural diagram of a model building device in another embodiment;

Figure 3 is a schematic structural diagram of a model building platform in another embodiment;

Figure 4 is a schematic structural diagram of an electronic device in another embodiment.

Implementation Mode of this Application

The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of this application.

Before elaborating on the technical solutions of the embodiments of the present application, first an exemplary description of the application scenarios of the embodiments of the present application is given:

The neural network model is a complex network model formed by interconnecting a large number of simple processing units (neurons), which reflects many basic characteristics of human brain functions. The neural network model has the characteristics of large-scale parallelism, distributed storage and processing, self-organization, Adaptive and self-learning abilities. The construction and training of neural network models in the existing technology need to be run on GPUs, CPUs, etc. with better performance. When users build neural network models, they need to consider the allocation of back-end hardware devices, which increases the complexity of building neural network models and reduces user experience. The embodiment of this application proposes a model construction method that can solve the above technical problems, so that users do not need to consider the back-end hardware equipment when building a model, and only need to describe the source model. After the description is completed, the model needs to be In the case of conversion, the source model is converted into an executable model. The technical solution of the embodiment of the present application reduces the complexity of model construction and improves the user experience.

In an embodiment of this application, a model construction method is provided. Figure 1 is a schematic flowchart of the model construction method provided by this embodiment of the application. The method may be performed by a model building device, which may Implemented in the form of software and/or hardware. For example, the combination of software applications and back-end hardware devices implements the device. Models include but are not limited to neuron models, synapse models, dendrite models, plasticity models, neuromodulation mechanism models, etc.

Neuron models include but are not limited to spiking neuron models, deep neuron models, and firing rate neuron models; for example, neuron models can include LIF (leaky integrate-and-fire), Hodgkin-Huxley, activation function ReLU, etc. The model can either use partial differential equations or use Euler methods or other functions to describe its dynamics.

As shown in Figure 1, the model construction method in the embodiment of this application specifically includes the following steps:

S110. In the model description stage, obtain the source model of the neural network model based on the model description information.

The model description data includes information about the first model and/or a pre-stored basic model. The basic model and/or the first model are used to construct a neural network model. The first model is a user-defined model. The basic model refers to a pre-stored model with structural information and variable information. The basic model refers to the minimum processing unit or a combination of minimum processing units that constructs a neural network model. For example, the source model can be a neuron model, a neuron group model, a synapse model, a synapse group model, a dendritic model, etc. The first model refers to the model defined by the user, for example, the programming code of the first model related to the source model written by the user. The model description information is used to describe the neural network model, that is, the source model in the embodiment of this application. The source model in the embodiment of this application refers to a neural network model with no allocated storage space.

Specifically, the source model is described based on the model description information. This step does not need to consider the memory occupied by the source model, thereby improving the efficiency of source model acquisition. Optionally, the source model may be composed of at least one basic model, or it may be composed of at least one first model, or it may be composed of at least one source model and at least one first model.

S120. When model conversion is required, enter the model conversion stage to convert the source model into an executable model.

In the embodiment of the present application, when model conversion is required, it may refer to the case where the user inputs a model conversion instruction. The executable model in the embodiment of the present application includes a source model in which storage space is allocated to the source model. That is, after the source model is allocated storage space, the executable model can be obtained.

Specifically, the source model is converted into an executable model, and the executable model is used as the completed neural network model. Of course, if the user needs a source model, they can directly obtain the source model without performing model conversion, and use the source model as the neural network model that the user needs to build.

The technical solution of the embodiment of the present application obtains the source model based on the model description information. The model description information includes the first model and/or the information of the pre-stored source model. The source model and/or the first model are used to construct the neural network model. The first model is a user-defined model. When model conversion is required, the source model is converted into an executable model. The technical solutions of the embodiments of the present application simplify the complexity of model construction, improve the model construction speed, and thereby improve the user experience.

In another embodiment of the present application, when model conversion is required, entering the model conversion stage to convert the source model into an executable model includes: if the model description information is the information of the first model , then compile the first model to obtain an executable model, and/or if the model description information is information of a basic model, obtain the executable model corresponding to the basic model.

In the embodiments of this application, there are different conversion methods for different model description information. When the model description information is the first model, the first model is compiled to obtain an executable model. When the model description information is the source model, , obtain the executable model corresponding to the source model. Since the source model is a pre-stored model, after obtaining the structure information and variable information of the source model, the corresponding executable model can be obtained. Since the first model is the programming code written by the user, the programming code is compiled to obtain an executable model.

In another embodiment of the present application, after obtaining the executable model, the executable model is run to test the constructed executable model.

In the case where the model description information is information of a pre-stored source model, before obtaining the source model of the neural network model based on the model description information, the method further includes: obtaining a model identifier of at least one basic model in the model construction information, and a variable identifier of at least one variable of the basic model; based on the model identifier, retrieve the basic model corresponding to the model identifier; based on at least one of the variable identifiers, retrieve the variable registry corresponding to the basic model Obtain a variable registration item corresponding to at least one of the variable identifiers; the variable registration table includes at least one variable registration item, and the variable registration item is used to describe the variable; the source model of the neural network model is obtained based on the model description information , including: obtaining the source model based on at least the basic model and the variable registration item corresponding to the basic model.

Among them, the model construction information refers to the information input by the user for building the model, including the model identification of at least one basic model and the variable identification of the variables of each basic model. It should be noted that in the embodiment of the present application, the basic model includes multiple variables, and at least one of the multiple variables can be obtained according to user needs. Therefore, the variables in the model building information are identified as at least one. The variable registration table includes at least one variable registration item, and the variable registration item is used to describe the variable. The variable registry refers to a table of pre-stored variables of the basic model. The variable registry includes variable identifiers and variable registration items corresponding to the variable identifiers. Variable identification can be variable name, variable number, etc. Variable registration items refer to the content describing variables, and the information of each variable in the basic model is stored in the corresponding variable registration table. Optionally, in the embodiment of this application, a basic model corresponds to a variable registry.

Specifically, the model identification of at least one basic model in the model construction information and the variable identification corresponding to each basic model are obtained. The basic model is stored in advance, and based on the model identifier, the basic model corresponding to the model identifier can be retrieved from the model storage area where the basic model is stored. Based on the at least one variable identifier, a variable registration item corresponding to the at least one variable identifier is retrieved from the variable registry corresponding to the basic model. Describe the source model based on at least one basic model and the variable registration items of the basic model. For the description of the source model, there is no need to consider the subsequent operation of the neural network model. The user only needs to express the source model that needs to be described in the form of model construction information. Yes, simplifying user operations. For example, the basic model is a neuron, including two variables that describe the input port and the output port of the neuron respectively. Then the source model of the neural network model can be described based on the neuron, the input port, and the output port. It should be noted that the source model in this step refers to the model described based on the base model and the variable registration items of the base model.

Through the technical solution of the embodiment of the present application, only when receiving the model construction information, the model identification of the basic model and the variable identification of the variables of the basic model in the model construction information are obtained, and the basic model can be quickly retrieved according to the model identification. , and quickly retrieve the variable registration items in the variable registry of the basic model based on the variable identification. As a result, the source model can be quickly described based on at least one basic model and the variable registration items of the basic model. When model conversion is required, the source model is converted into an executable model. The technical solutions of the embodiments of the present application simplify the complexity of model construction, improve the model construction speed, and thereby improve the user experience.

Optionally, when model conversion is required, it may refer to the case where a model conversion instruction input by the user is received.

In another embodiment of the present application, in the case where the model construction information also includes at least one target variable registration sub-item of a variable, the method further includes: obtaining at least one target variable registration sub-item of the variable; After retrieving the variable registration item corresponding to the at least one variable identifier from the variable registration table corresponding to the basic model based on the at least one variable identifier, the method further includes: detecting whether the variable registration item exists and matches the target variable. Initial variable registration sub-item corresponding to the registration sub-item; the variable registration item includes at least one initial variable registration sub-item, and the initial variable registration sub-item includes variable type and data class At least one item of the type; if yes, update the initial variable registration subitem based on the target variable registration subitem to obtain a new variable registration item; if not, add the at least one target variable registration subitem to the variable registration item, Get a new variable registry entry.

Among them, the variable registration items include multiple initial variable registration sub-items. The initial variable registration sub-items can be variable type, quantity type, whether it can be persisted, whether it can be shared, etc. Variable types include single variables, tensor variables, etc. The data type refers to the type of data corresponding to the variable type. For example, the data type of a single variable is a numeric value, the data type of a tensor variable is a multi-dimensional array, etc. Whether to persist includes whether the variable remains unchanged in the source model, and whether to share includes whether the variable is shared with other models.

Specifically, after obtaining the target variable registration sub-item of the model construction information, it is detected whether there is an initial variable registration sub-item corresponding to the target variable registration sub-item in the initial variable registration item of the variable. If so, based on the target variable registration sub-item Update the initial variable registration subkey and obtain a new variable registration key. If not, add the target variable registration subkey to the variable registration key to obtain a new variable registration key. Through this method, the user can adjust the variables of the existing basic model to obtain a basic model suitable for constructing a neural network model, thereby improving the flexibility of neural network model construction.

In another embodiment of the present application, in the embodiment of the present application, the basic model stored in the model storage area may be a collective name for a class of models with the same number of variables. For example, the variable registry corresponding to the neuron model includes Three variables, and when actually calling the neuron model, you can call two of the variables, or you can call one of the variables. In this case, when the two variables are called, based on the neuron model The first neural network model is described with two variables. When a variable is retrieved, the second neural network model is described based on the neuron model and one variable. It should be understood that the first neural network model and the second neural network model here are different. The basic model and the variable registry of the basic model stored in this way can reduce resource usage.

In another embodiment of the present application, obtaining the source model based on at least the basic model and the variable registration items corresponding to the basic model includes: temporarily storing the basic model and the variable registration items corresponding to the basic model. The variable registration key to get the source model.

In the embodiment of the present application, at least one basic model and at least one variable registration item corresponding to the basic model are temporarily stored to describe the neural network model, that is, to obtain the source model. So that when converting the source model, the source model can be converted based on the stored basic model and variable registration items. It should be noted that in the model description stage, only the basic model retrieved based on the model identifier and the variable registration items retrieved based on the variable identifier need to be stored, which facilitates user operation and also improves the efficiency of neural network model description. .

In another embodiment of the present application, the synchronous filling method is to fill in the variables immediately when the content to be filled is described, including: allocating storage space if the variables of the basic model and/or the first model have not yet been allocated storage space. ; Based on the content of the variable to be filled, obtain the data value to be filled, and fill in the variables of the basic model and/or the first model.

In another embodiment of the present application, the asynchronous filling method is to fill in any moment after describing the content of the variables to be filled, including: if the variables of the basic model and/or the first model have not yet allocated storage space, then Allocate storage space; based on the content of the variable to be filled, obtain the data value to be filled, and fill the variables of the basic model and/or the first model.

In another embodiment of the present application, the method further includes: using a synchronous filling method to fill in the variables of the basic model and/or the first model, or using an asynchronous filling method to fill in the basic model and/or the first model. The variables of the model are filled in.

In one embodiment, the model description information includes content to be filled in variables.

In one embodiment, the variable filler can generate the data value to be filled according to the specified content of the variable to be filled and/or the specified filling method. The content and filling method of the variables to be filled are relatively simple. The upper concept of an object, while the data value to be filled is specific.

In a possible implementation, the specified filling method includes a synchronous filling method and/or an asynchronous filling method. Among them, the synchronous filling method means that when the variable to be filled in the specified variable is determined, the system allocates storage space for the specified variable (if the storage space has been allocated, there is no need to allocate it again) and fills it; the asynchronous filling method refers to determining When a variable with a specified variable is to be filled, it will not be filled immediately, but will be filled in later (such as the model compilation stage) after the storage space is allocated to the specified variable (if the storage space has been allocated, there is no need to allocate it again) . From the perspective of storage space allocation timing, storage space is allocated immediately in synchronous filling mode, while storage space can be allocated delayed in asynchronous filling mode. Taking the asynchronous filling method as an example, when declaring a variable, the storage space may not be actually allocated for the variable, but the variable filler and/or filling method can be specified in the API where the variable is declared, that is, the content to be filled in the variable is determined. This variable is filled when the storage space is allocated, which helps decouple the source model from hardware details.

In the embodiment of the present application, variables are filled in a synchronous filling method or an asynchronous filling method, and adaptive changes can be made according to the actual situation of the system when filling variables. Filling variables through synchronous filling or asynchronous filling improves the flexibility of variable filling.

For example, if the variables in the basic model are single variables, then in the model description stage, data values to be filled can be assigned to the variables of the basic model and/or the first model. For another example, after the executable model is obtained and the initialization is completed, or after the executable model is run, the user's variable updated data value to be filled is received, and the data value to be filled is directly filled into the allocated storage space. in the variables of the base model and/or the first model.

Data values include specific values corresponding to data types. For example, the quantity type of a tensor variable is a multi-dimensional array, and the data value refers to the specific data value of the multi-dimensional array. It should be understood that the data type in the variable registration item in the embodiment of this application refers to multi-dimensional arrays, numerical values, strings, etc. Of course, the data type includes specific data values. For example, the data type is numerical value, and the specific numerical value is 2. Of course, the data value is not limited to the value corresponding to the data type. The user can add or delete the variable registration sub-items in the variable registration item according to the actual situation, or adjust the variables of the first model, etc. Therefore, the data values will also change accordingly. The storage space in the embodiment of this application may be the memory space of a GPU, CPU, etc.

It should be noted that the basic model in the embodiment of the present application includes at least one variable. If a certain basic model includes tensor variables and single variables, the basic model allocates storage space according to the tensor variable and performs data value to be filled. Filling, and calling the univariate data value to be filled, completing the initialization of the basic model. In the same way, the above operations of the first model and the basic model are the same and will not be described again here.

In another embodiment of the present application, the variable registration item includes multiple variable registration sub-items, and the data values to be filled in the variable registration sub-items of the tensor variable include multi-dimensional arrays; according to the variable The corresponding variable registration item calls the data value to be filled to fill the variables of the basic model and/or the first model that have been allocated storage space, including: when the variable type of the basic model is a tensor variable, according to The multi-dimensional array in the variable registration sub-item allocates storage space to the basic model, and fills the multi-dimensional array with the basic model allocated storage space.

In the embodiment of the present application, when the variable type of the basic model is a tensor variable, the storage space is allocated to the basic model according to the multi-dimensional array in the variable registration sub-item, and the basic model allocated the storage space is multi-dimensional. Array filling, through this filling method, tensor variables can be filled into multi-dimensional arrays to achieve preparation before running the executable model. Allocate storage space to the basic model for multi-dimensional arrays to ensure the rationality of storage space allocation.

In another embodiment of the present application, the basic model with allocated storage space is filled with a multi-dimensional array, including: using at least one of constant filling, copy filling, uniform random filling, normal random filling and local convolution filling. A variable filling method fills multi-dimensional arrays in the basic model.

In another embodiment of the present application, filling the multi-dimensional array of the basic model allocated with storage space includes: filling the multi-dimensional array with the basic model allocated with storage space through a variable filler.

In the embodiment of this application, a variable filler is preset, and the variable filler can cooperate according to a specified method or a default method to fill variables, here it is the filling of multi-dimensional arrays. The basic model with allocated storage space is filled with multi-dimensional arrays through variable fillers.

Optionally, the variable filler in the embodiment of the present application can fill in the variables of the basic model with random values according to the value interval specified by the user. That is, when the user inputs the model construction information, he not only inputs the variable identifier, but also inputs The variable registration subkey of the variable is specified, and the content in the variable registration subkey is limited or updated. When filling the data values to be filled, the data values to be filled can be filled in the basic model according to the specific data values to be filled or the numerical range in the variable registration sub-item.

In another embodiment of the present application, when the model description information is a basic model, and the model identifiers of multiple basic models are obtained from the model construction information, the method further includes: obtaining each basic model. Topological structure information between models; the source model is obtained based on at least the basic model and the variable registration items corresponding to the basic model, including: based on each basic model and the variable registration items corresponding to each basic model and topological structure information to obtain the source model.

In the embodiment of this application, the topological structure information is used to describe the connection method between basic models, the data transmission method between basic models, the data format, etc. Independent base models can be connected through topology information. Specifically, model identifiers of multiple basic models are obtained from the model construction information, and topological structure information between each basic model is obtained. Based on each basic model, the variable registration items corresponding to each basic model and the topological structure information between each basic model, the neural network model is described, that is, the source model is obtained. Through this method, source models of various frameworks and structures can be obtained.

In another embodiment of the present application, the topology information includes port information and/or model information of data transmission; the model information includes at least one item of the input model information and the output model information of the basic model; if the basic model If it is a port model, the topology information includes port information; if the basic model is a container model, the topology information includes the model information.

In the embodiment of this application, the port model includes a port and a model body, and the port model may refer to a basic model with ports. In the embodiment of this application, the model subject may refer to neurons, synapses, neuron groups, synapse groups, etc. Ports include input ports, output ports, reference ports, connection ports, etc. Port information is used to describe the connection method between basic models. For example, the output port of basic model A is connected to the input port of basic model B. The port information includes at least one of input port information, output port information, reference port information and connection port information. Reference ports provide a mechanism for port models to reference variables of other port models, and the two parties have a strict binding relationship. Reference ports are used to share information between at least two port model objects. A connection port is a model management interface that is bound to a connection model. Connection ports are used to dynamically bind other port models to another port model. The container model involves placing one or more components in a container. For example, place one or more neurons in a container to obtain container 1, and place container 1 and multiple neurons in container 3. Place container 1 and container 3 in container 4 and so on. A component can be a single neuron or a container. The model information is used to determine the input model and the output model. The input model information can be the identifier of the input model to determine the corresponding input model. The output model information can also be the identifier of the output model to determine the output model. Optionally, the model type of the input model and the model type of the output model are set in the container model in advance. For example, when the model type of the input model of container model A is the same as the model type of the output model type of container model C, the two can be connected.

Specifically, in the embodiment of the present application, when the basic model is a port model, the port information of the port model is obtained, and when the basic model is a container model, the model type is obtained. By determining the port mode The topology of type and container models enables describing the source model through multiple base models. It should be noted that when the container model sets input ports and output ports, the container model at this time can refer to the port model.

Optionally, for the port model, the source model can be described through the port model paradigm. The port model description paradigm includes at least one of a port model class and a port model template, so as to determine the port model based on at least one of the port model description paradigms of the port model class and the port model template. The port model description paradigm refers to arranging one or more components of the basic model into one or more port models. Port models can be associated through one or more ports to facilitate variable reference or sharing, forming nested, hierarchical Organizational methods such as connections and cycles. In this way, the information flow is clearer, the model organization form is more flexible, and the degree of modularization is better. It is easier for different partners to cooperate and develop, and their respective models can be formed into modules according to certain specifications, and further organized into larger models. , and users do not have to pay attention to the underlying details of the hardware. Taking the port model class to determine the port model as an example, for neurons, the neuron port model can be derived based on the port model class. The ports in the neuron port model include at least input ports and output ports, and the main body of the model is configured as The leaky integral function, a single variable as the threshold of the neuron, the single variable as the membrane potential of the neuron, and the input port is configured to accept the input of one or more neurons, and the output port is connected to one or more neurons.

Optionally, for the container model, the source model is described based on the container model, variable registration items and topology structure information. The container model paradigm can be used to describe the source model, that is, one or more components of the basic model are arranged in one or more containers. Model management and scheduling are performed through containers. The container model paradigm includes at least one of a container class and a container template. The paradigm is described based on at least one of the container class and the container template to form container models at different levels. Containers can form cascade, tree, flat, nested and other organizational forms. The container model includes node containers, that is, neuron containers. Node containers can be constructed by describing the basic model according to variables. The node container types include but are not limited to: neuron models, synapse models, dendrite models, etc. Connect the created node containers to form a network model, which is also the source model.

In this way, it is conducive to the realization of source models with arbitrary granularity, and is more conducive to the conversion of the source model at the bottom of the hardware, allowing users to design operations that are more in line with the hardware characteristics when designing the target model, so that the back-end part can be more reasonable. Arrange the call of hardware resources such as memory.

In a possible implementation, the network topology includes but is not limited to: feed-forward structure, feedback structure, cross-layer structure, lateral structure, loop structure, self-loop structure, neuron-to-neuron direct connection structure, synapse and Synaptic direct connection structure, synapse and dendrite direct connection structure.

In another embodiment of the present application, after converting the source model into an executable model, the method further includes: when receiving model update information, updating the source model based on the model update information to obtain a new source model. , and determine whether the new source model is the same as each basic model and/or topological structure information in the source model. If not, obtain the corresponding new available model based on each basic model and/or topological structure information of the new source model. Execution model.

The model update information in the embodiment of the present application may be an update to the basic model structure, for example, adding some neurons to the original basic model, or it may be a deletion of the basic model, for example, deleting At least one base model in the source model may also be an update of the topology structure between the base models. For example, if basic model A is connected to basic model B, it is updated to be connected to basic model A and basic model C. Specifically, after obtaining the executable model, the model update information is received, the source model is updated based on the model update information, and a new source model is obtained. After the source model is updated, each element in the new source model is judged. Whether the basic model and/or the topological structure between each basic model has changed? If so, it is necessary to re-allocate memory space to each basic model, fill in data values and/or call data values, etc., to obtain a new executable model. This step can improve the flexibility of model construction. If the user After obtaining the executable model, it is found that the effect is not ideal. The current source model can be updated and the converted executable model can be obtained to improve the flexibility of obtaining the executable model.

In another embodiment of the present application, after converting the source model into an executable model, the method further includes: when the update information of the variables of the received basic model/first model is detected, updating based on the update information of the variables. The variables in the basic model/first model are used to obtain an updated basic model/first model; the update information includes at least one of adding variables, deleting variables, and updating data values of variables.

Among them, for the source model, the update of the variable registration sub-item can be the adjustment of whether the variable is shared, the update of the specific data value of the variable's data type, etc. For the addition of variables, multiple first variables are added based on the variables of the original basic model. It should be understood that if the first variable is added, the update information includes the model identifier of the basic model, the variable identifier of the first variable added in the basic model, and the variable registration item corresponding to the first variable. To delete a variable, a deletion identifier can be set. When a certain variable identifier and a deletion identifier exist in the update information, the variable registration item corresponding to the variable identifier in the basic model is deleted. Of course, the modification information of the variables of the first model can also be obtained for the first model.

Specifically, after obtaining the executable model, the update information of any variable input by the user is detected, and the variables in the basic model are updated based on the update information to obtain a new basic model. For example, the update information includes the update of the variable registration sub-item of any variable, which may be to update the data value of the corresponding variable registration sub-item of the variable in the basic model. It should be understood that the data value here can refer to the specific data value corresponding to the data type, or whether the sharing is Y or F, etc., where Y here means yes, and F means no. After obtaining the executable model, when the update information of the variables is detected, the variables can be updated, and it can be determined whether the storage space needs to be re-allocated to the basic model. If so, the storage space will be reallocated and the data values will be filled in again. Convert. This step improves the flexibility of neural network model construction. Optionally, the user's variable update information can also be received during the process of converting the source model into an executable model.

In another embodiment of the present application, a model construction device is provided. Figure 2 is a schematic structural diagram of the model construction device provided by the embodiment of the present application. The model construction device provided by the embodiment of the present application can execute any embodiment of the present application. The provided model construction method has functional modules and beneficial effects corresponding to the execution method. The device includes: model description module 210 and model conversion module 220; wherein:

The model description module 210 is used to obtain the source model of the neural network model based on the model description information during the model description stage. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/ Or the first model is used to build a neural network model; the first model is a user-defined model;

The model conversion module 220 is used to enter the model conversion stage and convert the source model into an executable model when model conversion is required.

Further, in the embodiment of this application, the model conversion module is also used to:

If the model description information is the information of the first model, compile the first model to obtain an executable model, and/or if the model description information is the information of the basic model, obtain the executable model corresponding to the basic model. Execution model.

Further, in this embodiment of the application, the device further includes:

A variable registration item acquisition module is used to obtain the model identifier of at least one basic model in the model construction information and the variable identifier of at least one variable of the basic model; based on the model identifier, retrieve the model identifier corresponding to the model identifier. Basic model; based on at least one of the variable identifiers, retrieve a variable registration item corresponding to at least one of the variable identifiers from the variable registration table corresponding to the basic model; the variable registration table includes at least one variable registration item, so The variable registration item is used to describe the variable;

The model description module 210 is also configured to: based on at least the basic model and the corresponding The variable registration item obtains the source model.

Further, in this embodiment of the application, the device further includes:

A variable filling module is used to fill in the variables of the basic model and/or the first model using a synchronous filling method, or to fill the variables of the basic model and/or the first model using an asynchronous filling method.

Further, in the embodiment of the present application, the variable filling module includes: a synchronous filling sub-module, which is used to fill in the variable immediately when describing the content to be filled, and a synchronous filling sub-module, which is also used to: if the basic model and /or the variables of the first model have not been allocated storage space, then allocate storage space; based on the content to be filled in the variables, obtain the data values to be filled, and fill the variables of the basic model and/or the first model.

Further, in the embodiment of the present application, the variable filling module includes: an asynchronous filling sub-module, used to fill at any time after describing the content of the variable to be filled, an asynchronous filling sub-module, also used to if the basic If the variables of the model and/or the first model have not yet been allocated storage space, allocate storage space; based on the content to be filled in the variables, obtain the data values to be filled, and fill the variables of the basic model and/or the first model.

Further, in this embodiment of the present application, the variable registration item includes multiple variable registration sub-items, and the data values to be filled in the variable registration sub-items of the tensor variable include multi-dimensional arrays;

The variable filling module is also used to: when the variable type of the basic model is a tensor variable, allocate storage space to the basic model according to the multi-dimensional array in the variable registration sub-item, and allocate the basis of the storage space. The model performs filling of multidimensional arrays.

Further, in the embodiment of this application, the variable filling module is also used to:

Fill the multi-dimensional array of the basic model through at least one variable filling method among constant filling, copy filling, uniform random filling, normal random filling and local convolution filling.

Further, in the embodiment of this application, the variable filling module is also used to:

The basic model with allocated storage space is filled with multi-dimensional arrays through variable fillers.

Further, in the embodiment of the present application, when the model description information is a basic model, and the model identifiers of multiple basic models are obtained from the model construction information, the device further includes:

The information acquisition module is used to obtain the topological structure information between each basic model;

The model description module 210 is also used to:

The source model is obtained based on each basic model, the variable registration items corresponding to each basic model, and the topological structure information.

Further, in this embodiment of the present application, the topology information includes port information and/or model information of data transmission; the model information includes at least one of the input model information and the output model information of the basic model; if the basic If the model is a port model, the topology information includes port information; if the base model is a container model, the topology information includes the model information.

Further, in this embodiment of the present application, the port information includes at least one of input port information, output port information, reference port information and connection port information.

Further, in this embodiment of the application, the device further includes:

The model update module is used to update the source model based on the model update information when receiving the model update information, obtain a new source model, and determine whether the new source model is consistent with each basic model in the source model and/or Whether the topological structure information is the same, if not, the corresponding new executable model is obtained based on each basic model and/or topological structure information of the new source model.

Further, in this embodiment of the application, the device further includes:

Variable update module, used when detecting the update information of the variables of the received basic model/first model When, the variables in the basic model/first model are updated based on the update information of the variables to obtain an updated basic model/first model; the update information includes data values of added variables, deleted variables and updated variables. at least one of.

Further, in this embodiment of the present application, the model description module 210 is also used to:

Temporarily store the basic model and the variable registration items corresponding to the basic model to obtain the source model.

Further, in this embodiment of the application, the device further includes:

A variable registration sub-item acquisition module, used to obtain at least one target variable registration sub-item of the variable;

A variable registration item update module is used to detect whether the variable registration item has an initial variable registration sub-item corresponding to the target variable registration sub-item; the variable registration item includes at least one initial variable registration sub-item, and the initial variable The registration sub-item includes at least one of variable type and data type; if yes, update the initial variable registration sub-item based on the target variable registration sub-item to obtain a new variable registration item; if not, add the above-mentioned variable registration item to the variable registration item At least one target variable registration subkey is obtained, and a new variable registration key is obtained.

Through the technical solution of the embodiment of the present application, in the model description stage, the source model of the neural network model is obtained based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or the third A model is used to build a neural network model. The first model is a user-defined model. When model conversion is required, enter the model conversion stage to convert the source model into an executable model. The technical solution of the embodiment of the present application simplifies the complexity of model construction, improves the model construction speed, and thereby improves user experience.

It is worth noting that the various modules included in the above device are only divided according to functional logic, but are not limited to the above divisions, as long as they can achieve the corresponding functions; in addition, the specific names of each functional module are only for convenience The mutual distinction is not used to limit the scope of protection of the embodiments of the present application.

In another embodiment of the present application, a model building platform is provided. Figure 3 is a schematic structural diagram of a model building platform provided by an embodiment of the present application. The platform is used to execute the following when building a neural network model: For the model building method in any of the above embodiments, the platform includes: a front-end part 310, a core part 320 and a back-end part 330, where:

The front-end part 310 is used to interact with the core part and/or the back-end part on behalf of the outside of the platform;

The core part 320 is used to respond to requests from the front-end part and/or the back-end part. In the model description phase, obtain the source model of the neural network model based on the model description information. The model description data includes the first model. and/or pre-stored basic model information; the basic model and/or the first model are used to build a neural network model; the first model is a user-defined model;

The backend part 330 is used to enter the model conversion stage and convert the source model into an executable model when model conversion is required.

As for the front-end part 310, the outside of the system (such as users, other systems) will interact with the platform through the front-end part 310, thereby calling at least one function of the platform. For example, call the platform configuration model.

The front-end part 310 provided by the embodiment of the present application can receive requests input by the user, optionally, can display the neural network model built by the user, and can also display the running status of the neural network model. The core part 320 obtains the source model based on the model description information, and the backend part 330 converts the source model into an executable model when model conversion is required. The front-end part 310, the core part 320 and the back-end part 330 of the model building platform in the embodiment of the present application have a clear division of labor and cooperate with each other to realize the construction of the neural network model, which improves the efficiency of the construction, and because the core part 320 only describes the model , there is no need to consider issues related to model conversion. The backend part 330 converts the source model into an executable model. This application The model construction platform of the embodiment is convenient for user operations, simplifies the process of model construction, reduces the complexity of model construction, and improves user experience.

The back-end hardware devices include but are not limited to GPU, TPU, neuromorphic chips, artificial intelligence chips, etc.

In another embodiment of the present application, an electronic device is provided. FIG. 4 is a schematic structural diagram of an electronic device provided by this embodiment of the present application. 4 illustrates a block diagram of an exemplary electronic device 50 suitable for implementing embodiments of the present application. The electronic device 50 shown in FIG. 4 is only an example and should not bring any limitations to the functions and scope of use of the embodiments of the present application.

As shown in Figure 4, electronic device 50 is embodied in the form of a general computing device. The components of the electronic device 50 may include, but are not limited to: one or more processors or processing units 501, a system memory 502, and a bus 503 connecting different system components (including the system memory 502 and the processing unit 501).

Bus 503 represents one or more of several types of bus structures, including a memory bus or memory controller, a peripheral bus, a graphics accelerated port, a processor, or a local bus using any of a variety of bus structures. For example, these architectures include, but are not limited to, the Industry Standard Architecture (ISA) bus, the Micro Channel Architecture (MAC) bus, the Enhanced ISA bus, the Video Electronics Standards Association (VESA) local bus, and the Peripheral Component Interconnect ( PCI) bus.

Electronic device 50 typically includes a variety of computer system readable media. These media can be any available media that can be accessed by electronic device 50, including volatile and nonvolatile media, removable and non-removable media.

System memory 502 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) 504 and/or cache memory 505 . Electronic device 50 may further include other removable/non-removable, volatile/non-volatile computer system storage media. By way of example only, storage system 506 may be used to read and write to non-removable, non-volatile magnetic media (not shown in Figure 4, commonly referred to as a "hard drive"). Although not shown in FIG. 4, a disk drive may be provided for reading and writing to removable non-volatile disks (e.g., "floppy disks"), and for removable non-volatile optical disks (e.g., CD-ROM, DVD-ROM). or other optical media) that can read and write optical disc drives. In these cases, each drive may be connected to bus 503 through one or more data media interfaces. The memory 502 may include at least one program product having a set (eg, at least one) program module configured to perform the functions of various embodiments of the present application.

A program/utility 508 having a set of (at least one) program modules 507, including but not limited to an operating system, one or more application programs, other program modules, and program data, may be stored, for example, in memory 502 , each of these examples or some combination may include the implementation of a network environment. Program modules 507 generally perform functions and/or methods in the embodiments described herein.

Electronic device 50 may also communicate with one or more external devices 509 (e.g., keyboard, pointing device, display 510, etc.), with one or more devices that enable a user to interact with electronic device 50, and/or with Any device (eg, network card, modem, etc.) that enables the electronic device 50 to communicate with one or more other computing devices. This communication may occur through input/output (I/O) interface 511. Furthermore, the electronic device 50 may also communicate with one or more networks (eg, a local area network (LAN), a wide area network (WAN), and/or a public network, such as the Internet) through the network adapter 512. As shown, network adapter 512 communicates with other modules of electronic device 50 via bus 503 . It should be understood that, although not shown in Figure 4, other hardware and/or software modules may be used in conjunction with electronic device 50, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tapes drives and data backup storage systems, etc.

The processing unit 501 executes various functional applications and data processing by running programs stored in the system memory 502, for example, implementing the model building method provided by the embodiment of the present application.

In another embodiment of the present application, a storage medium containing computer-executable instructions is also provided. The computer-executable instructions, when executed by a computer processor, are used to perform a model construction method, and the method includes:

In the model description stage, the source model of the neural network model is obtained based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or the first model are For constructing a neural network model; the first model is a user-defined model; when model conversion is required, the model conversion stage is entered to convert the source model into an executable model.

The computer storage medium in the embodiment of the present application may be 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 is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus or device, or any combination thereof. More specific examples (non-exhaustive list) of computer readable storage media include: electrical connections having one or more conductors, portable computer disks, hard drives, random access memory (RAM), read only memory (ROM), Erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination of the above. As used herein, a computer-readable storage medium may be any tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device.

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

Program code embodied on a computer-readable medium may be transmitted using any suitable medium, including but not limited to wireless, wire, optical cable, RF, etc., or any suitable combination of the foregoing.

Computer program code for performing operations of embodiments of the present application may be written in one or more programming languages, including object-oriented programming languages such as Java, Smalltalk, C++, and combinations thereof, including Conventional procedural programming language - such as "C" or similar programming language. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In situations involving remote computers, the remote computer can be connected to the user's computer through any kind of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (such as an Internet service provider through Internet connection).

What is disclosed above is only the preferred embodiment of the present application. Of course, it cannot be used to limit the scope of rights of the present application. Therefore, equivalent changes made according to the claims of the present application still fall within the scope of the present application.

Claims (35)

1. A model building method, characterized in that the method includes:

   In the model description stage, the source model of the neural network model is obtained based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or the first model are For constructing a neural network model; the first model is a user-defined model;

   When model conversion is required, the model conversion stage is entered to convert the source model into an executable model.
2. The model construction method according to claim 1, characterized in that, when model conversion is required, entering the model conversion stage to convert the source model into an executable model includes:

   If the model description information is the information of the first model, compile the first model to obtain an executable model,

   and / or

   If the model description information is information of a basic model, the executable model corresponding to the basic model is obtained.
3. The model construction method according to claim 2, characterized in that, in the case where the model description information is information of a pre-stored source model, before obtaining the source model of the neural network model based on the model description information, it further includes: :

   Obtain the model identifier of at least one basic model in the model construction information, and the variable identifier of at least one variable of the basic model;

   Based on the model identifier, retrieve a basic model corresponding to the model identifier;

   Based on at least one of the variable identifiers, a variable registration item corresponding to at least one of the variable identifiers is retrieved from a variable registration table corresponding to the basic model; the variable registration table includes at least one variable registration item, and the variable registration item Terms are used to describe variables;

   The source model of the neural network model is obtained based on the model description information, including:

   The source model is obtained based on at least the basic model and the variable registration item corresponding to the basic model.
4. The model construction method according to claim 3, characterized in that the method further includes:

   Use a synchronous filling method to fill in the variables of the basic model and/or the first model,

   or

   An asynchronous filling method is used to fill in the variables of the basic model and/or the first model.
5. The model construction method according to claim 4, characterized in that the synchronous filling method is to fill in the variables immediately when the content to be filled is described, including:

   If the variables of the base model and/or the first model have not yet been allocated storage space, allocate storage space;

   Based on the content of the variable to be filled, the data value to be filled is obtained, and the variables of the basic model and/or the first model are filled.
6. The model construction method according to claim 4, characterized in that the asynchronous filling method is to fill in any moment after describing the content of the variable to be filled, including:

   If the variables of the base model and/or the first model have not yet been allocated storage space, allocate storage space;

   Based on the content of the variable to be filled, the data value to be filled is obtained, and the variables of the basic model and/or the first model are filled.
7. The model building method according to claim 5, wherein the variable registration item includes a plurality of variable registration sub-items, and the data values to be filled in the variable registration sub-items of the tensor variable include Multidimensional Arrays;

   Calling the data value to be filled to fill the variables of the basic model and/or the first model that has been allocated storage space includes:

   When the variable type of the basic model is a tensor variable, the storage space is allocated to the basic model according to the multi-dimensional array in the variable registration sub-item, and the multi-dimensional array is filled in the basic model to which the storage space is allocated.
8. The model construction method according to claim 7, characterized in that filling the multi-dimensional array of the basic model allocated with storage space includes:

   Fill the multi-dimensional array of the basic model through at least one variable filling method among constant filling, copy filling, uniform random filling, normal random filling and local convolution filling.
9. The model construction method according to claim 7, characterized in that filling the multi-dimensional array of the basic model allocated with storage space includes:

   The basic model with allocated storage space is filled with multi-dimensional arrays through variable fillers.
10. The model construction method according to claim 3, characterized in that when the model description information is a basic model and the model identifiers of multiple basic models are obtained from the model construction information, the method further include:

    Obtain topological structure information between each basic model;

    Obtaining the source model based on at least the basic model and the variable registration items corresponding to the basic model includes:

    The source model is obtained based on each basic model, the variable registration items corresponding to each basic model, and the topological structure information.
11. The model construction method according to claim 9, characterized in that the topological structure information includes port information and/or model information of data transmission; the model information includes at least one of the input model information and the output model information of the basic model. an item;

    If the basic model is a port model, the topology information includes port information;

    If the basic model is a container model, the topology information includes the model information.
12. The model building method according to claim 11, characterized in that the port information includes at least one of input port information, output port information, reference port information and connection port information.
13. The model construction method according to claim 10, characterized in that after converting the source model into an executable model, it further includes:

    When receiving model update information, update the source model based on the model update information, obtain a new source model, and determine whether the new source model is the same as each basic model and/or topology structure information in the source model, If not, a corresponding new executable model is obtained based on each basic model and/or topological structure information of the new source model.
14. The model construction method according to claim 1, characterized in that after converting the source model into an executable model, it further includes:

    When the update information of the received variables of the basic model/first model is detected, updating the variables in the basic model/first model based on the update information of the variables to obtain an updated basic model/first model; The update information includes at least one of adding a variable, deleting a variable, and updating a data value of a variable.
15. The model construction method according to claim 3, wherein obtaining the source model based on at least the basic model and the variable registration items corresponding to the basic model includes:

    Temporarily store the basic model and the variable registration items corresponding to the basic model to obtain the source model.
16. The model building method according to claim 3, characterized in that in the model building information If the information also includes at least one target variable registration subkey of the variable, the method further includes:

    Obtain at least one target variable registration subkey of the variable;

    After retrieving the variable registration item corresponding to the at least one variable identifier from the variable registry corresponding to the basic model based on the at least one variable identifier, the method further includes:

    Detect whether there is an initial variable registration sub-item corresponding to the target variable registration sub-item in the variable registration item; the variable registration item includes at least one initial variable registration sub-item, and the initial variable registration sub-item includes variable type and data at least one of the types;

    If so, update the initial variable registration subkey based on the target variable registration subkey to obtain a new variable registration key,

    If not, add the at least one target variable registration sub-item to the variable registration item to obtain a new variable registration item.
17. A model building device, characterized in that the device includes:

    A model description module, used in the model description stage to obtain the source model of the neural network model based on the model description information. The model description data includes the first model and/or the information of the pre-stored basic model; the basic model and/or The first model is used to build a neural network model; the first model is a user-defined model;

    The model conversion module is used to enter the model conversion stage and convert the source model into an executable model when model conversion is required.
18. The model building device according to claim 17, characterized in that the model conversion module is also used to:

    If the model description information is the information of the first model, compile the first model to obtain an executable model, and/or if the model description information is the information of the basic model, obtain the executable model corresponding to the basic model. Execution model.
19. The model building device according to claim 18, characterized in that the device further includes:

    A variable registration item acquisition module is used to obtain the model identifier of at least one basic model in the model construction information and the variable identifier of at least one variable of the basic model; based on the model identifier, retrieve the model identifier corresponding to the model identifier. Basic model; based on at least one of the variable identifiers, retrieve a variable registration item corresponding to at least one of the variable identifiers from the variable registration table corresponding to the basic model; the variable registration table includes at least one variable registration item, so The variable registration item is used to describe the variable;

    The model description module is further configured to obtain the source model based on at least the basic model and the variable registration items corresponding to the basic model.
20. The model building device according to claim 19, characterized in that the device further includes:

    A variable filling module is used to fill in the variables of the basic model and/or the first model using a synchronous filling method, or to fill the variables of the basic model and/or the first model using an asynchronous filling method.
21. The model building device according to claim 20, wherein the variable filling module includes:

    The synchronous filling sub-module is used to fill in the variables immediately when describing the content to be filled.

    The synchronous filling sub-module is also used to: allocate storage space if the variables of the basic model and/or the first model have not yet been allocated storage space;

    Based on the content of the variable to be filled, the data value to be filled is obtained, and the variables of the basic model and/or the first model are filled.
22. The model building device according to claim 20, wherein the variable filling module includes:

    The asynchronous filling sub-module is used to fill in any moment after describing the content of the variable to be filled.

    The asynchronous filling sub-module is also used to allocate storage space if the variables of the basic model and/or the first model have not yet been allocated storage space;

    Based on the content of the variable to be filled, the data value to be filled is obtained, and the variables of the basic model and/or the first model are filled.
23. The model building device according to claim 20, wherein the variable registration item includes a plurality of variable registration sub-items, and the data values to be filled in the variable registration sub-items of the tensor variable include multi-dimensional arrays;

    The variable filling module is also used to: when the variable type of the basic model is a tensor variable, allocate storage space to the basic model according to the multi-dimensional array in the variable registration sub-item, and allocate the basis of the storage space. The model performs filling of multidimensional arrays.
24. The model building device according to claim 23, characterized in that the variable filling module is also used to:

    Fill the multi-dimensional array of the basic model through at least one variable filling method among constant filling, copy filling, uniform random filling, normal random filling and local convolution filling.
25. The model building device according to claim 23, characterized in that the variable filling module is also used to:

    The basic model with allocated storage space is filled with multi-dimensional arrays through variable fillers.
26. The model construction device according to claim 19, characterized in that, when the model description information is a basic model, and the model identifiers of multiple basic models are obtained from the model construction information, the device further include:

    The information acquisition module is used to obtain the topological structure information between each basic model;

    The model description module is also used to:

    The source model is obtained based on each basic model, the variable registration items corresponding to each basic model, and the topological structure information.
27. The model building device according to claim 26, wherein the topological structure information includes port information and/or model information of data transmission; the model information includes at least one of the input model information and the output model information of the basic model. An item; if the basic model is a port model, the topology information includes port information; if the basic model is a container model, the topology information includes the model information.
28. The model building device according to claim 27, wherein the port information includes at least one of input port information, output port information, reference port information and connection port information.
29. The model building device according to claim 26, characterized in that the device further includes:

    The model update module is used to update the source model based on the model update information when receiving the model update information, obtain a new source model, and determine whether the new source model is consistent with each basic model in the source model and/or Whether the topological structure information is the same, if not, the corresponding new executable model is obtained based on each basic model and/or topological structure information of the new source model.
30. The model building device according to claim 17, characterized in that the device further includes:

    A variable update module, configured to update the variables in the basic model/first model based on the updated information of the variables when detecting the received update information of the variables of the basic model/first model, to obtain an updated basic model/ First model; the update information includes at least one of data values of added variables, deleted variables, and updated variables.
31. The model building device according to claim 19, characterized in that the model description module 210 is also used to:

    Temporarily store the basic model and the variable registration items corresponding to the basic model to obtain the source model type.
32. The model building device according to claim 19, characterized in that the device further includes:

    A variable registration sub-item acquisition module, used to obtain at least one target variable registration sub-item of the variable;

    A variable registration item update module is used to detect whether the variable registration item has an initial variable registration sub-item corresponding to the target variable registration sub-item; the variable registration item includes at least one initial variable registration sub-item, and the initial variable registration item The registration subitem includes at least one of a variable type and a data type; if yes, update the initial variable registration subitem based on the target variable registration subitem to obtain a new variable registration item; if not, add the above variable registration item to the variable registration item At least one target variable registration subkey is obtained, and a new variable registration key is obtained.
33. A model building platform, characterized in that the platform executes the model building method of claims 1-16, and the platform includes: a front-end part, a core part and a back-end part, wherein:

    The front-end part is used to interact with the core part and/or the back-end part on behalf of the outside of the platform;

    The core part is used to respond to requests from the front-end part and/or the back-end part. In the model description phase, obtain the source model of the neural network model based on the model description information. The model description data includes the first model and /or pre-stored basic model information; the basic model and/or the first model are used to build a neural network model; the first model is a user-defined model;

    The back-end part is used to enter the model conversion stage and convert the source model into an executable model when model conversion is required.
34. An electronic device, characterized in that the electronic device includes:

    one or more processors;

    a storage device for storing one or more programs,

    When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the model building method as described in any one of claims 1-16.
35. A storage medium containing computer-executable instructions, characterized in that, when executed by a computer processor, the computer-executable instructions are used to perform the model construction method according to any one of claims 1-16.