WO2023272987A1

WO2023272987A1 - Model recommendation method and apparatus, and device and computer storage medium

Info

Publication number: WO2023272987A1
Application number: PCT/CN2021/121263
Authority: WO
Inventors: 袁坤; 余锋伟
Original assignee: 深圳市商汤科技有限公司
Priority date: 2021-06-29
Filing date: 2021-09-28
Publication date: 2023-01-05
Also published as: CN113408634A; CN113408634B

Abstract

Provided in the present disclosure are a model recommendation method and apparatus, and a device and a computer storage medium. The model recommendation method comprises: acquiring a target attribute parameter of a neural network model, which runs in first hardware, wherein the target attribute parameter comprises a desired speed value and/or a desired precision value; and on the basis of the first hardware and the target attribute parameter, screening each neural network model in a preset neural network model library, so as to obtain a neural network model which matches the target attribute parameter, wherein an attribute parameter of each neural network model in the preset neural network model library is obtained by performing a test in second hardware, and the second hardware comprises the first hardware. Therefore, automatic model recommendation is realized.

Description

Model recommendation method and device, equipment, computer storage medium

Cross References to Related Applications

This disclosure is based on the Chinese patent application with the application number 202110730001.1, the application date is June 29, 2021, and the application name is "model recommendation method and device, equipment, computer storage medium", and claims the priority of the Chinese patent application. The entire content of this Chinese patent application is hereby incorporated by reference into this disclosure.

technical field

The present disclosure relates to the field of artificial intelligence, and in particular to a model recommendation method, device, device, and computer storage medium.

Background technique

With the rapid development of artificial intelligence, deep learning technology has been successfully applied to the field of computer vision. It makes the feature extraction of the image change from traditional manual design to automatic extraction based on data, which greatly improves the robustness of image features and the accuracy of recognition. Among them, the design of the model plays a crucial role.

Since different models can achieve different task processing effects for the same task, it is of great significance to select an appropriate model for a specific task. However, in related technologies, when selecting a model, engineers often choose based on work experience, which is difficult to select and has poor accuracy, which makes the trial and error cost of the model higher, which in turn leads to a longer cycle of model selection. defect.

Contents of the invention

Embodiments of the present disclosure provide a model recommendation method, device, device, and computer storage medium.

The disclosed technical solution is achieved in this way:

An embodiment of the present disclosure provides a model recommendation method, the method comprising:

Obtaining the target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed values and/or expected accuracy values; based on the first hardware and the target attribute parameters, in the preset neural network model Each neural network model is screened in the library to obtain a neural network model that matches the target attribute parameter; the attribute parameter of each neural network model in the preset neural network model library is obtained in the second hardware test, The second hardware includes the first hardware.

In this way, after the target attribute parameters such as the expected speed value and the expected accuracy value are given under the expected hardware platform environment, the neural network that matches the target attribute parameters can be quickly and accurately determined from the pre-set model library. The network model realizes the automatic recommendation of the model.

An embodiment of the present disclosure provides a model recommendation device, including:

The acquiring part is configured to acquire target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed values and/or expected accuracy values;

The screening part is configured to screen each neural network model in a preset neural network model library based on the first hardware and the target attribute parameter to obtain a neural network model that matches the target attribute parameter; The attribute parameters of each neural network model in the preset neural network model library are obtained through testing on the second hardware, and the second hardware includes the first hardware.

An embodiment of the present disclosure provides a model recommendation device. The model recommendation device includes a processor and a memory storing executable instructions of the processor. When the instructions are executed by the processor, the above-mentioned model is realized. recommended method.

An embodiment of the present disclosure provides a computer-readable storage medium, on which a program is stored and applied to a model recommendation device. When the program is executed by a processor, the above-mentioned model recommendation method is implemented.

An embodiment of the present disclosure provides a computer program, including computer readable codes. When the computer readable codes run in an electronic device and are executed by a processor in the electronic device, the above-mentioned model recommendation is implemented. method.

An embodiment of the present disclosure provides a computer program product, which, when run on a computer, enables the computer to execute the above-mentioned model recommendation method.

According to the technical solution proposed by the embodiments of the present disclosure, the model recommendation device can obtain the target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed value and/or expected accuracy value; based on the first hardware and target attribute parameters, Each neural network model is screened in the preset neural network model library to obtain a neural network model that matches the target attribute parameter; the attribute parameters of each neural network model in the preset neural network model library are obtained in the second hardware The test shows that the second hardware includes the first hardware. In this way, after the target attribute parameters such as the expected speed value and/or the expected accuracy value included in the expected hardware platform environment, it can be quickly and accurately determined from the preset model library that matches the target attribute parameters. The neural network model realizes the automatic recommendation of models, improves the accuracy of model selection, reduces the cost of model trial and error, and further overcomes the defect of long model selection cycle.

Description of drawings

FIG. 1 is a first schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 2 is a second schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 3 is a schematic diagram of the third implementation process of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 4 is a schematic diagram 4 of the implementation flow of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 5 is a schematic diagram of the fifth implementation flow of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 6 is a sixth schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure;

FIG. 7 is a schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure VII;

FIG. 8 is a schematic diagram of an application scenario of a model recommendation method proposed by an embodiment of the present disclosure;

FIG. 9 is a schematic diagram of the composition and structure of a model recommendation device proposed by an embodiment of the present disclosure;

FIG. 10 is a schematic diagram of the composition and structure of a model recommendation device proposed by an embodiment of the present disclosure.

detailed description

In order to make the purpose, technical solutions and advantages of the present disclosure clearer, the present disclosure will be further described in detail below in conjunction with the accompanying drawings. All other embodiments obtained under the premise of creative labor belong to the protection scope of the present disclosure.

In the following description, references to "some embodiments" describe a subset of all possible embodiments, but it is understood that "some embodiments" may be the same subset or a different subset of all possible embodiments, and Can be combined with each other without conflict.

In the following description, the term "first\second\third" is only used to distinguish similar objects, and does not represent a specific ordering of objects. Understandably, "first\second\third" Where permitted, the specific order or sequencing may be interchanged such that the embodiments of the disclosure described herein can be practiced in sequences other than those illustrated or described herein.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. The terms used herein are only for the purpose of describing the embodiments of the present disclosure, and are not intended to limit the present disclosure.

Before the embodiments of the present disclosure are further described in detail, the nouns and terms involved in the embodiments of the present disclosure will be described, and the nouns and terms involved in the embodiments of the present invention are applicable to the following explanations.

1) ImageNet-1k Val standard test set: a computer vision data set, which is a large-scale image data set established to promote the development of computer image recognition technology, used for training and testing neural network models, and can be used to evaluate the performance of image classification algorithms benchmark.

2) Test accuracy: refers to the accuracy index obtained by evaluating the trained network on the standard test set.

3) Parameter quantity: refers to the number of learnable parameters contained in the model.

4) Calculation amount: refers to the number of multiplication and addition calculations required by the model for a given input size image.

5) Running time: refers to the running speed of the model on a specific hardware platform.

6) Pareto algorithm: a multi-objective optimization algorithm. Many problems in real life are composed of multiple goals that conflict and influence each other. These goals cannot reach the optimal state at the same time. Objective optimization means that there are two or more optimization objectives under constraint conditions, and these objectives are contradictory. One objective is often at the expense of the other objective, and the Pareto algorithm can be used to solve the Pareto optimal solution.

With the rapid development of artificial intelligence, deep learning technology has been successfully applied to the field of computer vision. It makes the feature extraction of the image change from traditional manual design to automatic extraction based on data, which greatly improves the robustness of image features and the accuracy of recognition. Among them, the design of the structure/model of the neural network plays a crucial role.

Since different models can achieve different task processing effects for the same task, it is of great significance to select an appropriate model for a specific task. And because the selection of the model is related to many parameters, it is necessary to consider these factors comprehensively when selecting the model for recommendation. However, when selecting a model in related technologies, not only the model structure is relatively limited, but also out of touch with reality, and only limited software attribute parameters such as parameter amount and calculation amount are considered. At the same time, the model structure is generally supported by a single task, which has poor flexibility. Even the selection of models is often made by engineers based on work experience, which makes model selection difficult, poor accuracy, and high trial and error costs for models, which in turn leads to the defect of a long model selection cycle.

In view of this, how to realize efficient model selection is an urgent problem to be solved, which is the content to be discussed in the embodiments of the present disclosure, and will be described below in conjunction with the following specific embodiments.

Embodiments of the present disclosure provide a model recommendation method, device, device, and computer storage medium. After the target attribute parameters such as the expected speed value and/or the expected accuracy value are given under the expected hardware platform environment, it can be obtained from Quickly and accurately determine the neural network model that matches the target attribute parameters in the pre-set model library, realize the automatic recommendation of the model, improve the accuracy of model selection, reduce the cost of model trial and error, and further overcome the model selection cycle longer flaws.

The model recommendation method proposed by the embodiment of the present disclosure is applied to a model recommendation device. The following describes the exemplary application of the model recommendation device proposed by the embodiment of the present disclosure. The model recommendation device proposed by the embodiment of the present disclosure can be implemented as a mobile phone, a notebook computer, a tablet computer, a desktop computer, a smart TV, a vehicle-mounted device, a wearable device, an industrial equipment etc.

In the following, the technical solutions in the embodiments of the present disclosure will be clearly and completely described with reference to the drawings in the embodiments of the present disclosure.

An embodiment of the present disclosure provides a model recommendation method. FIG. 1 is a schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure. As shown in FIG. 1 , in the embodiment of the present disclosure, the model recommendation device executes the model A recommended approach could include the following steps:

S100. Obtain a target attribute parameter of the neural network model running on the first hardware; the target attribute parameter includes an expected speed value and/or an expected accuracy value.

In some embodiments, the model recommendation device is configured with a search engine, and the front end of the search engine corresponds to the first interface. Wherein, the user can configure the performance requirement parameters of the neural network model in the first interface, so as to obtain the performance requirement parameters used to describe the model to be recommended in response to the user's configuration operation in the first interface.

In some embodiments, the performance requirement parameters used to describe the neural network model matching the target attribute parameters may include target application scenarios and target attribute parameters.

Wherein, the target application scenario at least includes the expected hardware platform environment, that is, the first hardware that supports model deployment and operation.

Wherein, the target attribute parameter includes at least one of an expected speed value and an expected accuracy value, and the expected accuracy value is the accuracy index obtained when the neural network model matching the target attribute parameter is deployed and tested under the first hardware, The expected speed value is the expected computing speed when the model to be recommended is deployed and run under the first hardware, in other words, the running time.

That is to say, in the embodiment of the present disclosure, in response to the user's configuration operation on the first interface, at least the target attribute parameters of the neural network model specified by the user to run on the first hardware deployment can be obtained, including the expected speed value and expected At least one of the precision values.

S110. Based on the first hardware and the target attribute parameters, screen each neural network model in the preset neural network model library to obtain a neural network model that matches the target attribute parameter; each of the preset neural network model libraries The attribute parameters of the neural network model are obtained through testing on the second hardware, and the second hardware includes the first hardware.

In an embodiment of the present disclosure, after obtaining the target attribute parameter under the first hardware specified by the user, that is, at least one of the expected speed value and the expected accuracy value, the model recommendation device may select the target attribute parameter from the preset The neural network model is screened in the established model library.

It should be understood that the preset model library is a model library containing a large number of neural network models. The neural network models in this model library cover a wide range, and can be neural network models deployed on mobile terminals, that is, small models on the terminal; The neural network model deployed on the cloud, that is, the large model on the cloud, covers a wide range from small models on the end to large models on the cloud. In addition, various models different in at least one of type, depth, width, and resolution are covered in the model library.

Wherein, each neural network model in the preset model library is obtained by training through a preset training data set, and each neural network model is tested on a preset testing data set.

In some embodiments, each neural network model in the preset model library corresponds to a piece of associated information, and a corresponding relationship between the identification of the neural network model and the associated information can be established; wherein, the associated information can represent the neural network model The test results obtained by testing on the preset test data set can associate each neural network model in the model library with its corresponding test results.

Among them, each neural network model in the model library can be deployed and tested in any preset hardware platform environment, so as to obtain the corresponding test results of each neural network model. The calculation speed value and calculation accuracy value corresponding to the model, and then the test result can be associated with the corresponding neural network model.

Here, any preset hardware platform environment may refer to second hardware capable of model deployment, operation and testing, and the first hardware may be one of the second hardware.

In some embodiments, the model recommendation device is provided with a search engine, and the back-end access of the search engine contains a preset model library with a large number of model structures, and the model recommendation device can use the search engine based on the first hardware including The target attribute parameters of the expected speed value and/or the expected accuracy value are screened for the neural network model in the preset model library connected to the backend.

Wherein, when screening the model to be recommended based on the target attribute parameters under the first hardware, the screening of the neural network model can be performed based on at least one set of associated information; wherein, each set of associated information can be represented under the first hardware, and the model Each neural network model in the library takes a specific batch size (batch size) as input to obtain the calculation speed value and calculation accuracy value after testing.

In an implementation manner of the embodiment of the present disclosure, the model library may be searched based on the target attribute parameters under the first platform, and then the neural network model matching the target attribute parameters may be obtained.

For example, the first hardware can be a mobile phone terminal. The expected speed value is the upper limit of the running time (in other words, running speed) when the model is running on the mobile phone terminal, and the expected accuracy value is the lower limit of the model’s accuracy when the model is running on the mobile phone terminal. The model recommendation device retrieves and matches the neural network models that can support the deployment and operation of the first hardware from the back-end model library based on the above parameters through the model search engine, so as to determine the neural network model that matches the target attribute parameters.

In some embodiments, after the neural network model matching the target attribute parameter is determined, the neural network model can be presented on a second interface; wherein, the second interface can be the same interface as the first interface, or can be is a different interface.

The embodiment of the present disclosure proposes a model recommendation method, by obtaining the target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include the expected speed value and/or the expected accuracy value; based on the first hardware and the target attribute parameters, Each neural network model is screened in the preset neural network model library to obtain a neural network model that matches the target attribute parameter; the attribute parameters of each neural network model in the preset neural network model library are obtained in the second hardware The test shows that the second hardware includes the first hardware. In this way, after the target attribute parameters such as the expected speed value and/or the expected accuracy value included in the expected hardware platform environment, it can be quickly and accurately determined from the preset model library that matches the target attribute parameters. The neural network model realizes the automatic recommendation of models, improves the accuracy of model selection, reduces the cost of model trial and error, and further overcomes the defect of long model selection cycle.

Fig. 2 is a schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure. The method for screening each neural network model in the model library to obtain a neural network model that matches the target attribute parameters may include the following steps:

S200. Based on the first hardware, batch size, and target attribute parameters, screen each neural network model in the preset neural network model library to obtain a Pareto model; wherein, the Pareto model is to satisfy the expected speed value and /or the expected accuracy value, and the neural network model with optimal calculation speed and calculation accuracy.

S210. Determine the Pareto model as a neural network model that matches the target attribute parameter.

In some embodiments, the target attribute parameter also includes a batch size, that is, the batch size processed by the neural network model when the neural network model is deployed and run under the expected hardware platform environment, ie, the first hardware.

In an embodiment of the present disclosure, after acquiring the expected batch size under the first hardware specified by the user, and at least one of the expected speed value and expected precision value, the model recommendation device may select from the preset Screening of neural network models in the model library.

In some embodiments, at least one batch size can be preset, and each neural network model can be tested with any batch size as input for each neural network model in the model library under any hardware platform environment , so as to obtain the test results corresponding to each neural network model. The test results are the calculation speed values and calculation accuracy values corresponding to each neural network model under any batch size under any hardware platform environment, and then the test results can be The results are correlated with the corresponding neural network model.

In some embodiments, when the neural network model is screened based on the target attribute parameters including batch size, expected speed value and/or expected precision value under the first hardware, the neural network model can be performed based on at least one set of associated information. Screening of network models; wherein, each group of associated information can be characterized under the first hardware, and each neural network model in the model library uses the same batch size (batch size) as an input to obtain the calculation speed value and calculation speed obtained after testing precision value.

For example, the first hardware can be a mobile phone terminal, the batch size of the neural network model run by the first hardware can be 256, and the expected speed value is the upper limit of the running time of the model when running on the mobile phone terminal (in other words, the running speed) , the expected accuracy value is the lower limit of the model’s accuracy when the model is running on the mobile phone terminal. Based on the above parameters, the model recommendation device performs search and matching in the neural network model that can support the deployment and operation of the first hardware and has a processing batch size of 256. All candidate models that meet the target attribute parameters, that is, the running time is less than the expected speed value and the accuracy is greater than the expected accuracy value, can be determined, and then based on the optimal solution algorithm, such as the Pareto algorithm, it can be determined from these candidate models that can be deployed On the mobile phone terminal, the Pareto optimal neural network model with the least time-consuming movement and the highest precision is used as the neural network model matching the target attribute parameters.

In this way, given the expected hardware platform environment, expected batch size, expected speed value, and expected accuracy value, it can be quickly and accurately determined from the pre-set model library that meets the target index value and the calculation speed And the Pareto model with optimal calculation accuracy realizes the automatic recommendation of models, improves the accuracy of model selection, reduces the cost of model trial and error, and further overcomes the defect of long model selection cycle.

Fig. 3 is a schematic diagram of the third implementation process of the model recommendation method proposed by the embodiment of the present disclosure. As shown in Fig. 3, in the embodiment of the present disclosure, the method for the model recommendation device to perform model recommendation may include the following steps:

S300. Acquire a first network structure library, where the first network structure library includes different types of initial neural network structures.

S310. Extend the acquired first network structure library to obtain a second network structure library.

In this embodiment of the present disclosure, the model recommendation device may pre-build a model library. Among them, the construction of the model library can be realized in the following ways: a large number of neural network structures can be defined first, and these neural network structures are trained and processed to obtain neural network models, and each neural network model is tested and processed to obtain the representation of each neural network. Each test result of the model attribute, and then associate each test result with the corresponding neural network structure, so as to build a model library based on each neural network structure and the association relationship between each neural network structure and the corresponding test result.

In some embodiments, defining a large number of neural network structures can be implemented in the following manner: a first network structure library including different types of initial neural network structures can be obtained, and by expanding the dimension of the initial neural network structure, the first network The structure library is expanded to obtain a second network structure library; here, the second network structure library contains a large number of neural network structures.

In an implementation manner of an embodiment of the present disclosure, the model recommendation device may first obtain the first network structure library, which includes the initial neural network structure, respectively residual neural network (ResNet), dense neural network ( DenseNet), efficient neural network (EfficientNet), mobile terminal neural network (MobileNet), normative neural network (RegNet), etc., these neural network structures can be expanded and transformed in dimension, so as to achieve the first network structure library contained in The neural network structure is expanded to obtain the second network structure library.

S320. Perform training processing on each neural network structure in the second network structure library based on the first data set, to obtain corresponding neural network models.

In the embodiment of the present disclosure, the model recommendation device may perform model training processing on each neural network structure in the expanded second network structure library, so as to obtain a neural network model corresponding to each neural network structure.

Wherein, each neural network structure in the second network structure library can be trained and processed based on a preset training data set, ie, the first data set, according to a unified standard, and then a corresponding neural network model can be obtained. Here, the unified standard may be that each neural network structure follows a unified target loss function and a unified learning rate, which is not specifically limited in this application.

In the embodiment of the present disclosure, each neural network structure in the second network structure library can be trained on the first task type based on the preset training data set, wherein the first task type is not limited to any task type , such as classification tasks, or object detection tasks or image segmentation tasks.

Wherein, the training process of classifying each neural network structure in the second network structure library may be performed based on the preset training data set, so as to obtain the neural network model corresponding to each neural network structure. Or, it is also possible to perform target detection task training processing on each neural network structure in the second network structure library based on the preset training data set, so as to obtain the neural network model corresponding to each neural network structure; or, it can also be based on the preset The training data set performs image segmentation task training processing on each neural network structure in the second network structure library, so as to obtain the neural network model corresponding to each neural network structure.

For example, the model recommendation device respectively trains ResNet, DenseNet, and EfficientNet based on the preset training data set to obtain the trained neural network model corresponding to ResNet, the trained neural network model corresponding to DenseNet, and the trained neural network model corresponding to EfficientNet. Model.

S330. Under each second hardware, use each type of batch size as an input to test each neural network model in the model library, and obtain a calculation speed value and a calculation accuracy value of each neural network model.

S340. Associating the second hardware, the batch size, the calculation speed value, and the calculation accuracy value with the corresponding neural network model to obtain the attribute parameters of each neural network model in the preset neural network model library.

In the embodiment of the present disclosure, the model recommendation device can perform a performance test on each neural network model obtained after training; wherein, model testing can be performed based on a preset test data set such as the ImageNet-1k Val standard test set, so as to obtain each The test results corresponding to the neural network model.

In some embodiments, the model recommendation device can be implemented under a variety of hardware platform environments, that is, a variety of second hardware, including a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) or a mobile phone chip The deployment test of the model is carried out in the mobile phone chip and so on. Wherein, under each second hardware, for each neural network model, different batch sizes may be used as input to test each neural network model to obtain a test result.

The test results here may include the running speed of the model obtained under each second hardware and each batch size as input, that is, the calculation speed value, and the test accuracy of the model, that is, the calculation accuracy value. On the other hand, the test results may also include the amount of parameters of the model and the amount of computation of the model.

In the embodiment of the present disclosure, the relationship between the neural network structure model and the corresponding test results can be established; here, under each hardware platform environment, that is, under each second hardware, each batch size can be used as an input. The test results of each neural network model obtained from the test are associated with the corresponding neural network model.

Here, the calculation speed and calculation accuracy of each neural network model obtained by testing each batch size as input under each hardware platform environment can be associated with the corresponding neural network identification, that is, the neural network structure is established The relationship between models and model attribute parameters.

It can be seen that in the embodiments of the present disclosure, by defining a large number of neural network structures different in at least one of type, depth, width, and resolution, and training each neural network structure, a corresponding neural network model is obtained, and a A model library with a rich and extensive range of neural network models. Each neural network model is tested to obtain test results, and the identification of each neural network model is further associated with the corresponding test results, so as to quickly detect the neural network model that matches the target attribute parameters based on the association relationship.

Fig. 4 is a schematic diagram of the fourth implementation process of the model recommendation method proposed by the embodiment of the present disclosure. As shown in Fig. 4, in the embodiment of the present disclosure, the model recommendation device expands the acquired first network structure library to obtain the second network A method of structuring a library may include the steps of:

S401. Extend each initial neural network structure in the first network structure library in at least one dimension to obtain an expanded neural network structure set corresponding to each initial neural network structure; the dimension includes at least one of the following: neural network structure width, depth and resolution.

S402. Construct a second network structure library based on each initial neural network structure and the corresponding expanded neural network structure set.

In the embodiment of the present disclosure, when the first network structure library containing the initial neural network structure is extended by the second network structure library, the expansion of the network structure library can be realized based on the expansion of each initial neural network structure in different dimensions. expand.

Wherein, each initial neural network structure may be expanded in at least one dimension of width, depth, and resolution, so as to obtain an expanded neural network structure set corresponding to each initial neural network structure.

For example, the ResNet in the initial neural network structure is expanded and changed in one dimension, that is, the expanded transformation in depth, width, and resolution respectively, to obtain the first ResNet after depth expansion, and the second ResNet after width expansion. The third ResNet after resolution expansion; or the expansion transformation in two dimensions, that is, the expansion transformation in depth and width, or depth and resolution, or width and resolution respectively, to obtain the depth and width expansion of the first Four ResNets, the fifth ResNet after depth and resolution expansion, the sixth ResNet after width and resolution expansion; or the seventh ResNet after expansion and change in three dimensions and resolution expansion, that is In other words, after expanding and changing each initial neural network structure in at least one dimension of depth, width, and resolution, the expanded neural network structure set corresponding to each initial neural network structure can be obtained, which can be based on the expanded The set of neural network structures expands the first network structure library to obtain the second network structure library.

It can be seen that, in the embodiments of the present disclosure, the initially defined initial neural network structure is extended and transformed in at least one dimension of type, depth, width and resolution to further expand the neural network structure.

FIG. 5 is a schematic diagram of the implementation process of the model recommendation method proposed in the embodiment of the present disclosure. As shown in FIG. 5, in the embodiment of the present disclosure, the model recommendation device tests each neural network model in the model library, and obtains the results of the test The method may include the steps of:

S500. Extract at least one batch size from the preset test data set.

S510. Determine at least one preset second hardware.

S511. Under each second hardware, for each neural network model in the model library, each neural network model is tested with each batch size as input, and each neural network model is tested in each batch size Below is the corresponding calculation speed value and calculation precision value.

In the embodiment of the present disclosure, the model recommendation device can test each neural network model based on different batch sizes under various hardware platform environments based on a preset standard test set.

Among them, the model recommendation device can extract at least one batch size from a preset test data set, such as a standard test set, and determine at least one second hardware that supports model deployment testing, and then under each second hardware, for each A neural network model is tested with various batch sizes as input, and the calculation speed and calculation accuracy of each neural network model under each second hardware with each batch size as input can be obtained.

It can be seen that in each hardware platform environment, for each neural network model in the model library, each neural network model is tested with each batch size as input, and the running time of each neural network model on each hardware platform is obtained. and precision.

Fig. 6 is a schematic diagram 6 of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure. As shown in Fig. 6, in the embodiment of the present disclosure, after the model recommendation device obtains the neural network model that matches the target attribute parameters The method may include the steps of:

S601. Acquire a second task type to be processed by the neural network model on the first hardware.

In some embodiments, after determining the neural network model that matches the target attribute parameters and presenting the recommended neural network model on the second interface, it may respond to the model creation operation performed by the user on the second interface, based on the The task requirement parameters of the model and the obtained neural network model matching the target attribute parameters are used to create a target neural network model that meets the task requirement parameters.

Wherein, the task requirement parameter may refer to the second task type to be processed by the neural network model on the first hardware.

In an implementation manner of the embodiments of the present disclosure, the model recommendation device is configured with a creation interface for model creation, and the front end of the creation interface corresponds to the second interface. Among them, the user can perform the creation operation of the neural network model on the creation interface, such as specifying the second task type to be processed by the neural network model on the first hardware, and then the model recommendation device can respond to the user's creation operation on the creation interface, and obtain the user Used to describe the second task type to be processed.

Here, the second task type may be a classification task; or may also be an object detection task; or may also be an image segmentation task, which is not specifically limited in the present application.

In other embodiments, the task requirement parameter may also include the number of categories of the neural network model.

It can be understood that the output data of each layer of the neural network model is different, and the output data of the middle layer of the neural network model can be obtained, and the output data of the last layer of the neural network model can also be obtained. In the embodiment of the present disclosure, when creating a model, the user can also specify which layer of the model needs to obtain the output data, that is, specify the data output layer of the model, in other words, the depth of the model or the number of categories.

In another embodiment of the present disclosure, the user can create a neural network model on the creation interface, such as specifying the type of task to be processed and the number of categories of the neural network model, and then the model recommendation device can respond to The creation operation of the creation interface by the user obtains the second task type and the corresponding number of categories used to describe the neural network model to be processed on the first hardware.

S602. When the second task type does not match the first task type, retrain the neural network model based on the second data set corresponding to the second task type, so as to fine-tune the parameters of the neural network model.

In some embodiments, after determining the neural network model that matches the target attribute parameters, and presenting the neural network model on the second interface, and obtaining a description for processing in response to the model creation operation performed by the user on the second interface After the second task type, if the second task type does not match the preset first task type, that is, the task type when training the neural network model in the model library is different from the pending task when the current model is created different types, then the model recommendation device can retrain the neural network model based on the second data set corresponding to the second task type to fine-tune the parameters of the neural network model, for example, some hyperparameters of the model, learning rate, optimization Adjustment of the device, number of iterations, etc.

It can be seen that in the embodiment of the present disclosure, when the model is created based on the user's task requirements, according to the specified task type and the neural network structure model to be recommended, if the task type to be processed is different from the preset task type during model training, The neural network model can be further retrained based on new task types and new data sets to achieve fine-tuning of model parameters.

Fig. 7 is a schematic diagram of the implementation process of the model recommendation method proposed by the embodiment of the present disclosure VII. As shown in Fig. 7, in the embodiment of the present disclosure, after the model recommendation device obtains the neural network model that matches the target attribute parameters, the method further Can include the following steps:

S701. Acquire a second task type to be processed by the neural network model on the first hardware.

S702. In the case that the second task type matches the first task type, create a corresponding target neural network model based on the second task type, at least one set of preset task specification information, and a neural network model that matches the target attribute parameters ; wherein: each set of task specification information is used to represent at least one preset task type, and the corresponding input format and output format under each neural network model in the model library.

In the embodiment of the present disclosure, in response to the creation operation of the creation interface, after obtaining the second task type used to describe the neural network model matching the target attribute parameter on the first hardware to be processed, if the second task type is consistent with If the preset first task type matches, the model can be created based on the second task type and the neural network model matching the target attribute parameters.

In some embodiments, the model recommendation device can predefine the multiple task types supported by each neural network model in the specification model library, and the corresponding input and output formats of each neural network model under each task type, that is, at least one set of tasks Specification information. Wherein, the at least one set of specification information is at least one task type, and the corresponding input format and output format under each neural network model in the model library.

In some embodiments, after obtaining the second task type to be processed on the first hardware for describing the neural network model matching the target attribute parameter, based on the second task type and at least one set of task specification information, Determine the input and output formats of the model under the second task type, and then standardize the input and output formats of the neural network model that match the target attribute parameters, so as to further construct the target neural network model that supports the second task type.

In the embodiment of the present disclosure, in order to enable each neural network model in the model library to support different tasks, such as classification tasks, target detection tasks, image classification tasks, etc., the task type and input and output formats of each neural network model can be adjusted. specification definition; where at least one set of task specification information can be implemented in the following ways:

Determine at least one preset task type; for each neural network model in the model library, perform specification definition processing based on each task type and corresponding input format and output format, and obtain a corresponding set of task specification information.

For example, classification tasks, object detection tasks, image classification tasks, etc.

Among them, the classification task is standardized, given the specified input, and the returned output format is a fixed-length two-dimensional vector, which can support the use of classifiers for category determination.

Among them, the target detection task or image segmentation task is standardized and defined, given the specified input, the returned output format is a set of feature matrices of different scales, and the feature extraction of the task is supported.

It can be seen that various task types are defined and standardized for each model in the model library, so that each neural network model can be called by different downstream tasks.

In some other embodiments, after obtaining the second task type and the number of categories to be processed on the first hardware for describing the neural network model matching with the target attribute parameters, based on the second task type and at least one set of Task specification information, determine the input and output formats of the model under the second task type, determine the data output layer of the model based on the number of categories, and then standardize the input and output formats of the neural network model that match the target attribute parameters, and The data output layer of the neural network model matching the target attribute parameters, thereby further constructing the target neural network model supporting the second task type.

It can be seen that in the embodiments of the present disclosure, based on user task requirements, a target neural network model capable of supporting a specific task can be constructed according to the specified task type and the neural network structure model to be recommended.

Exemplarily, FIG. 8 is a schematic diagram of the application scenario of the model recommendation method proposed by the embodiment of the present disclosure. As shown in FIG. 8, the performance (running time and accuracy) distribution of each neural network model in the model library on the GPU hardware platform, the model The library contains 11 types of neural network models, including resnet, regnet, bignas, bignas, dmcp, shufflenet_v2, mobilenet_v2, oneshot_supcell, crnas_resnet, efficient, and netmobilenet_v3. The structure of each neural network model can be adjusted in width, By expanding at least one of the dimensions of depth and resolution, a set of neural network models corresponding to each type can be obtained. For example, after the neural network structure corresponding to resnet is expanded in at least one dimension, neural network models of the same type but different dimensional structures such as resnet18c_×0_25, resnet18c_×0.5, resnet18c_×0_125, and dmcp_resnet18_47M can be obtained. The extensions of corresponding structures of other types of neural network models are similar, and will not be repeated here.

Further, the target attribute parameters based on the GPU hardware platform, for example, the running time is 1ms. The accuracy is 60%. When screening the model library to determine the neural network model that matches the target attribute parameters, you can first determine all candidate models that take less than 1ms to run and have an accuracy greater than 60%, that is, the upper left corner corresponding to the dotted line intersection The multiple neural network models are all candidate models that satisfy the requirement that the running time is less than 1ms and the accuracy is greater than 60%. Further, the fastest and best accuracy can be determined from these candidate models based on the Pareto optimal solution method. The Pareto model, that is, the neural network model bignas_resnet18_492M corresponding to the points on the Pareto curve.

Based on the above-mentioned embodiment, in an embodiment of the present disclosure, FIG. 9 is a schematic diagram of the composition and structure of the model recommendation device proposed by the embodiment of the present disclosure. As shown in FIG. 9, the model recommendation device 10 includes an acquisition part 11, a screening Part 12, extension part 13, training part 14, testing part 15, association part 16, determination part 17.

The acquiring part 11 is configured to acquire target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed values and/or expected accuracy values;

The screening part 12 is configured to screen each neural network model in a preset neural network model library based on the first hardware and the target attribute parameter to obtain a neural network model that matches the target attribute parameter; The attribute parameters of each neural network model in the preset neural network model library are obtained through testing on the second hardware, and the second hardware includes the first hardware.

In some embodiments, the target attribute parameter further includes a batch amount processed by the neural network model based on the first hardware, and the screening part 12 is further configured to be based on the first hardware, the batch quantity, the target attribute parameter, and screen each neural network model in the preset neural network model library to obtain a Pareto model; wherein, the Pareto model is to satisfy the desired speed value and/or Or a neural network model with an expected accuracy value and optimal calculation speed and calculation accuracy; and determining the Pareto model as a neural network model that matches the target attribute parameter.

In some embodiments, the obtaining part 11 is configured to obtain a first network structure library, and the first network structure library includes different types of initial neural network structures.

In some embodiments, the extension part 13 is configured to expand the obtained first network structure library to obtain the second network structure library.

In some embodiments, the training part 14 is configured to perform training processing on each neural network structure in the second network structure library based on the first data set to obtain corresponding neural network models.

In some embodiments, the testing part 15 is configured to test each of the neural network models by using each of the batch sizes as input under each of the second hardware to obtain the neural network models. The calculation speed value and calculation accuracy value of the network model.

In some embodiments, the association part 16 is configured to associate the second hardware, the batch size, the calculation speed value and the calculation accuracy value with the corresponding neural network model, The attribute parameters of each neural network model in the preset neural network model library are obtained.

In some embodiments, the expansion part 13 is configured to perform expansion processing on each initial neural network structure in the first network structure library in at least one dimension, to obtain the expanded neural network structure corresponding to each initial neural network structure. The set of neural network structures; the dimension includes at least one of the following: the width, depth and resolution of the neural network structure; and based on each initial neural network structure and the corresponding expanded neural network structure set, construct the Describe the second network structure library.

In some embodiments, the training part 14 is configured to use the first data set to train each neural network structure in the second network structure library based on the preset first task type to obtain the corresponding Each neural network model of .

In some embodiments, the testing part 15 is configured to extract at least one batch size from a preset test data set; and determine at least one preset second hardware; and in each of the second Under the hardware, for each neural network model in the model library, each neural network model is tested with each of the batch sizes as input, and each neural network model is tested under each of the batch sizes. Corresponding calculation speed value and calculation accuracy value.

In some embodiments, the acquisition part 11 is configured to acquire the second task type to be processed by the neural network model on the first hardware after obtaining the neural network model that matches the target attribute parameter .

In some embodiments, the training part 14 is configured to, in the case that the second task type does not match the first task type, based on the second data set corresponding to the second task type, pair the The neural network model is retrained to fine-tune the parameters of the neural network model.

In some embodiments, the determining part 17 is configured to determine a preset task type; wherein, the preset task type includes at least the first task type and the second task type; and based on each one of the task types and its corresponding input format and output format, and confirm the input format and output format of each neural network model in the preset neural network model library.

In the embodiment of the present disclosure, further, FIG. 10 is a schematic diagram of the composition and structure of the model recommendation device proposed in the embodiment of the present disclosure. As shown in FIG. 10 , the model recommendation device 20 proposed in the embodiment of the present disclosure may also include a processor 21 , a memory 22 storing instructions executable by the processor 21 , further, the living body detection device 20 may further include a communication interface 23 , and a bus 24 for connecting the processor 21 , the memory 22 and the communication interface 23 .

In an embodiment of the present disclosure, the above-mentioned processor 21 may be an application specific integrated circuit (Application Specific Integrated Circuit, ASIC), a digital signal processor (Digital Signal Processor, DSP), a digital signal processing device (Digital Signal Processing Device, DSPD ), Programmable Logic Device (ProgRAMmable Logic Device, PLD), Field Programmable Gate Array (Field Prog RAMmable Gate Array, FPGA), Central Processing Unit (Central Processing Unit, CPU), controller, microcontroller, microprocessor at least one of the It can be understood that, for different devices, the electronic device used to implement the above processor function may also be other, which is not specifically limited in this embodiment of the present disclosure. The living body detection device 20 may also include a memory 22, which may be connected to the processor 21, wherein the memory 22 is used to store executable program codes, the program codes include computer operation instructions, and the memory 22 may include a high-speed RAM memory, or may Also included is non-volatile memory, eg, at least two disk memories.

In the embodiment of the present disclosure, the bus 24 is used to connect the communication interface 23 , the processor 21 and the memory 22 and communicate with each other among these devices.

In an embodiment of the present disclosure, the memory 22 is used to store instructions and data.

Further, in the embodiment of the present disclosure, the above-mentioned processor 21 is configured to acquire a target application scenario and a target index value used to describe a neural network model that matches the target attribute parameters, and the target application scenario includes at least: expected hardware platform environment; the target index value includes at least an expected speed value and/or an expected accuracy value; based on the target application scenario and the target index value, each neural network model in the preset model library is screened, Obtain the neural network model that matches the target attribute parameter; the neural network model that matches the target attribute parameter is that the test result obtained by testing under the expected hardware platform environment meets the target index value, and the calculation speed and Calculate the neural network model with optimal accuracy.

In practical applications, the above-mentioned memory 22 can be a volatile memory (volatile memory), such as a random access memory (Random-Access Memory, RAM); or a non-volatile memory (non-volatile memory), such as a read-only memory (Read-Only Memory, ROM), flash memory (flash memory), hard disk (Hard Disk Drive, HDD) or solid-state drive (Solid-State Drive, SSD); Provide instructions and data.

In addition, each functional module in this embodiment may be integrated into one recommendation unit, each unit may exist separately physically, or two or more units may be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software function modules.

If the integrated unit is implemented in the form of a software function module and is not sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this embodiment is essentially or The part contributed by the prior art or the whole or part of the technical solution can be embodied in the form of software products, the computer software products are stored in a storage medium, and include several instructions to make a computer device (which can be a personal A computer, a server, or a network device, etc.) or a processor (processor) executes all or part of the steps of the method of this embodiment. The aforementioned storage medium includes: U disk, mobile hard disk, read only memory (Read Only Memory, ROM), random access memory (Random Access Memory, RAM), magnetic disk or optical disk and other various media that can store program codes.

An embodiment of the present disclosure provides a model recommendation device, which can pre-build a target model library; where the target model library is used to characterize the correspondence between candidate models, software attribute parameters, and hardware attribute parameters; and then receive In the case of a recommendation request to a model; wherein, the recommendation request carries recommended software attribute parameters and recommended hardware attribute parameters; the target model library is searched and processed according to the recommended software attribute parameters and recommended hardware attribute parameters, and then the target recommendation model is obtained. In this way, by building a target model library that contains rich model structures and rich model attributes, it is possible to automatically search for a suitable recommended model in the target model library according to the specified model recommendation requirements, realizing automatic model recommendation and improving the efficiency of model selection. Accuracy reduces the cost of model trial and error, and further overcomes the defect of long model selection cycle. .

An embodiment of the present disclosure provides a computer-readable storage medium on which a program is stored, and when the program is executed by a processor, the above-mentioned model recommendation method is implemented.

Specifically, the program instructions corresponding to a model recommendation method in this embodiment can be stored on a storage medium such as an optical disk, a hard disk, or a USB flash drive. When the program instructions corresponding to a model recommendation method in the storage medium are stored by a When an electronic device is read or executed, the following steps are included:

Obtaining a target application scenario and a target index value used to describe a neural network model that matches the target attribute parameter, the target application scenario at least includes: an expected hardware platform environment; the target index value includes at least an expected speed value and/or expected precision value;

Based on the target application scenario and the target index value, filter each neural network model in the preset model library to obtain the neural network model matching the target attribute parameters;

The neural network model matched with the target attribute parameter is a neural network model whose test results obtained by testing under the expected hardware platform environment meet the target index value, and whose calculation speed and calculation accuracy are optimal.

Correspondingly, an embodiment of the present disclosure further provides a computer program product, where the computer program product includes computer-executable instructions, and the computer-executable instructions are used to implement the steps in the model recommendation method proposed by the embodiments of the present disclosure.

Those skilled in the art should understand that the embodiments of the present disclosure may be provided as methods, systems, or computer program products. Accordingly, the present disclosure may take the form of a hardware embodiment, a software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present disclosure may take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) having computer-usable program code embodied therein.

The present disclosure is described with reference to the implementation flow diagrams and/or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present disclosure. It should be understood that each process and/or block in the schematic flowchart and/or block diagram, and a combination of processes and/or blocks in the schematic flowchart and/or block diagram can be realized by computer program instructions. These computer program instructions may be provided to a general purpose computer, special purpose computer, embedded processor, or processor of other programmable data processing equipment to produce a machine such that the instructions executed by the processor of the computer or other programmable data processing equipment produce a Means for realizing the functions specified in one or more steps of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing apparatus to operate in a specific manner, such that the instructions stored in the computer-readable memory produce an article of manufacture comprising instruction means, the instructions The device realizes the function specified in implementing one or more procedures of the flowchart and/or one or more blocks of the block diagram.

These computer program instructions can also be loaded onto a computer or other programmable data processing device, causing a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process, thereby The instructions provide steps for implementing the functions specified in implementing the process flow or processes of the flowchart diagrams and/or the block or blocks of the block diagrams.

The above descriptions are only preferred embodiments of the present disclosure, and are not intended to limit the protection scope of the present disclosure.

Industrial Applicability

In the embodiment of the present disclosure, by obtaining the target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include the expected speed value and/or the expected accuracy value; based on the first hardware and the target attribute parameters, in the preset neural network Each neural network model is screened in the model library to obtain a neural network model that matches the target attribute parameters; the attribute parameters of each neural network model in the preset neural network model library are obtained from the second hardware test, and the second hardware Includes first hardware. Realized the automatic recommendation of the model.

Claims

A model recommendation method, the method comprising:

Obtain target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed values and/or expected accuracy values;

Based on the first hardware and the target attribute parameters, each neural network model is screened in a preset neural network model library to obtain a neural network model that matches the target attribute parameters; the preset neural network model The attribute parameters of each neural network model in the network model library are obtained through testing on the second hardware, and the second hardware includes the first hardware.
The method according to claim 1, wherein the target property parameter further comprises a batch size processed by a neural network model running on the first hardware; the target property parameter based on the first hardware, Each neural network model is screened in the preset neural network model library to obtain a neural network model that matches the target attribute parameters, including:

Based on the first hardware, the batch size, and the target attribute parameters, each neural network model is screened in the preset neural network model library to obtain a Pareto model; wherein, the Pareto The support model is a neural network model that satisfies the desired speed value and/or desired precision value, and has optimal calculation speed and calculation accuracy;

The Pareto model is determined as a neural network model matching the target attribute parameter.
The method according to claim 1 or 2, wherein the preset neural network model library is constructed based on the following method, comprising:

Obtaining a first network structure library, the first network structure library including at least one type of initial neural network structure;

expanding the acquired first network structure library to obtain a second network structure library;

performing training processing on each neural network structure in the second network structure library based on the first data set to obtain corresponding neural network models;

Under each of the second hardware, each of the neural network models is tested by using each of the batch sizes as an input, and the calculation speed value and the calculation accuracy value of the various neural network models are obtained;

associating the second hardware, the batch size, the calculation speed value, and the calculation accuracy value with the corresponding neural network model to obtain the The attribute parameters of each neural network model.
The method according to claim 3, wherein said expanding the obtained first network structure library to obtain a second network structure library includes:

Each initial neural network structure in the first network structure library is extended in at least one dimension to obtain a set of expanded neural network structures corresponding to each initial neural network structure; the dimension includes at least one of the following : The width, depth and resolution of the neural network structure;

The second network structure library is constructed based on each initial neural network structure and the corresponding expanded neural network structure set.
The method according to claim 3, wherein, the training process is performed on each neural network structure in the first neural network structure set based on the preset training data set to obtain corresponding neural network models, including:

Based on the preset first task type, the first data set is used to train each neural network structure in the second network structure library to obtain corresponding neural network models.
The method according to any one of claims 3 to 5, wherein the testing of each neural network model in the model library to obtain test results includes:

Extract at least one batch size from a preset test data set;

determining at least one preset second hardware;

Under each of the second hardware, for each neural network model in the model library, each neural network model is tested with each batch size as input, and each neural network model is obtained in each neural network model. The corresponding calculation speed value and calculation accuracy value under the batch size.
The method according to any one of claims 1 to 6, wherein, after obtaining the neural network model matched with the target attribute parameters, the method further comprises:

Acquiring a second task type to be processed by the neural network model on the first hardware;

When the second task type does not match the first task type, retrain the neural network model based on the second data set corresponding to the second task type, so as to retrain the neural network model fine-tuning of the parameters.
The method according to claim 7, wherein the method further comprises:

determining a preset task type; wherein, the preset task type includes at least the first task type and the second task type;

Based on each task type and its corresponding input format and output format, confirm the input format and output format of each neural network model in the preset neural network model library.
A model recommendation device, the model recommendation device comprising:

The acquiring part is configured to acquire target attribute parameters of the neural network model running on the first hardware; the target attribute parameters include expected speed values and/or expected accuracy values;

The screening part is configured to screen each neural network model in a preset neural network model library based on the first hardware and the target attribute parameter to obtain a neural network model that matches the target attribute parameter; The attribute parameters of each neural network model in the preset neural network model library are obtained through testing on the second hardware, and the second hardware includes the first hardware.
The model recommendation device according to claim 9, wherein the target attribute parameter further includes the batch size of the neural network model run based on the first hardware,

The screening part is further configured to screen each neural network model in the preset neural network model library based on the first hardware, the batch size, and the target attribute parameter to obtain a Pareto model; wherein, the Pareto model is a neural network model that satisfies the desired speed value and/or desired precision value, and has optimal calculation speed and calculation accuracy; and the Pareto model is determined to be compatible with the described A neural network model that matches the target attribute parameters.
The model recommendation device according to claim 9 or 10, wherein,

The obtaining part is configured to obtain a first network structure library, and the first network structure library includes different types of initial neural network structures;

The extension part is configured to expand the acquired first network structure library to obtain a second network structure library;

The training part is configured to perform training processing on each neural network structure in the second network structure library based on the first data set, to obtain corresponding neural network models;

The testing part is configured to test each of the neural network models with each of the batch quantities as input under each of the second hardware, and obtain the calculation speed value and the value of each neural network model. Calculation precision value;

The associating part is configured to associate the second hardware, the batch size, the calculation speed value, and the calculation accuracy value with the corresponding neural network model to obtain the preset neural network model. The attribute parameters of each neural network model in the network model library.
The model recommendation device according to claim 11, wherein,

The extension part is configured to perform extension processing on at least one dimension of each initial neural network structure in the first network structure library, to obtain an expanded neural network structure set corresponding to each initial neural network structure; The dimensions include at least one of the following: width, depth and resolution of the neural network structure; and building the second network structure library based on the initial neural network structures and the corresponding expanded neural network structure set.
The model recommendation device according to claim 11, wherein,

The training part is configured to use the first data set to train each neural network structure in the second network structure library based on a preset first task type to obtain corresponding neural network models.
The model recommendation device according to any one of claims 11 to 13, wherein,

The test part is configured to extract at least one batch size from a preset test data set; and determine at least one preset second hardware; and under each of the second hardware, for the model For each neural network model in the library, each neural network model is tested with each of the batch sizes as input, and the corresponding calculation speed value and calculation speed value of each neural network model under each of the batch sizes are obtained. precision value.
The model recommendation device according to any one of claims 9 to 14, wherein,

The obtaining part is configured to obtain a second task type to be processed by the neural network model on the first hardware after obtaining the neural network model matching the target attribute parameter;

The training part is configured to retrain the neural network model based on a second data set corresponding to the second task type when the second task type does not match the first task type, to fine-tune the parameters of the neural network model.
The model recommendation device according to claim 15, wherein,

The determining part is configured to determine a preset task type; wherein, the preset task type includes at least the first task type and the second task type; and based on each of the task types and The corresponding input format and output format confirm the input format and output format of each neural network model in the preset neural network model library.
A model recommendation device, the model recommendation device includes a processor and a memory storing instructions executable by the processor, and when the instructions are executed by the processor, the model described in any one of claims 1-8 is implemented. described method.
A computer-readable storage medium, on which a program is stored and applied to a model recommendation device, and when the program is executed by a processor, the method according to any one of claims 1-8 is implemented.
A computer program, comprising computer readable code, when the computer readable code runs in an electronic device and is executed by a processor in the electronic device, any one of claims 1-8 can be realized the method described.
A computer program product, which, when run on a computer, causes the computer to execute the method according to any one of claims 1-8.