CN116663614A

CN116663614A - Deep learning network structure generation method and device

Info

Publication number: CN116663614A
Application number: CN202310380674.8A
Authority: CN
Inventors: 孙修宇; 王耀华; 黄一伦
Original assignee: Alibaba China Co Ltd
Current assignee: Alibaba China Co Ltd
Priority date: 2022-12-22
Filing date: 2022-12-22
Publication date: 2023-08-29
Also published as: CN115640835A; CN115640835B

Abstract

The embodiment of the specification provides a method and a device for generating a deep learning network structure, wherein the method for generating the deep learning network structure comprises the following steps: acquiring service demand information corresponding to an initial deep learning network structure and a target service, wherein the service demand information comprises a deep learning network structure parameter threshold; generating a plurality of amplified deep learning network structures based on the initial deep learning network structure; determining a deep learning network structure set to be screened in the multiple amplified deep learning network structures according to the service demand information, and determining a reference deep learning network structure in the deep learning network structure set to be screened; and generating a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

Description

Deep learning network structure generation method and device

Technical Field

The embodiment of the specification relates to the technical field of deep learning, in particular to a method for generating a deep learning network structure.

Background

Deep learning plays an important role in visual intelligence, the effect of the deep learning is closely related to the neural network structure, and the convolutional neural network is an important neural network structure for processing visual intelligence scenes. The visual intelligent effect based on the convolutional neural network is closely related to the structure of the neural network. The current popular convolutional neural network structure is mainly set by combining personal experiences of technicians with different service scenes, the mode of setting the network structure is dependent on individuals, errors can not be avoided in personal experiences, performance of the neural network structure designed for a target service scene is over-winning or the use requirement of the target service scene can not be met, the performance is over-generating, hardware cost can be improved, and the use requirement of the target service scene can not be met, so that physical examination of a user is poor. Therefore, a more accurate way of designing neural network structures for a target traffic scenario is needed.

Disclosure of Invention

In view of this, the present embodiment provides a method of generating a deep learning network structure. One or more embodiments of the present specification relate to a deep learning network structure generating apparatus, a computing device, a computer-readable storage medium, and a computer program, which solve the technical drawbacks of the prior art.

According to a first aspect of embodiments of the present disclosure, there is provided a method for generating a deep learning network structure, including:

acquiring service demand information corresponding to an initial deep learning network structure and a target service, wherein the service demand information comprises a deep learning network structure parameter threshold;

generating a plurality of amplified deep learning network structures based on the initial deep learning network structure;

determining a deep learning network structure set to be screened in the multiple amplified deep learning network structures according to the service demand information, and determining a reference deep learning network structure in the deep learning network structure set to be screened;

and generating a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

According to a second aspect of embodiments of the present specification, there is provided a generating apparatus of a deep learning network structure, including:

the system comprises an acquisition module, a target service acquisition module and a control module, wherein the acquisition module is configured to acquire service demand information corresponding to an initial deep learning network structure and the target service, and the service demand information comprises a deep learning network structure parameter threshold value;

an amplification module configured to generate a plurality of amplified deep learning network structures based on the initial deep learning network structure;

the determining module is configured to determine a deep learning network structure set to be screened in the plurality of amplified deep learning network structures according to the service demand information, and determine a reference deep learning network structure in the deep learning network structure set to be screened;

the generation module is configured to generate a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

According to a third aspect of embodiments of the present specification, there is provided a method of generating an image processing network structure, including:

acquiring service demand information corresponding to an initial convolution network structure and a target image processing service, wherein the service demand information comprises a convolution network structure length-width ratio threshold value, and the convolution structure length-width ratio threshold value is determined according to the number of convolution network layers, the convolution kernel size and the convolution channel number;

Generating a plurality of augmented convolutional network structures based on the initial convolutional network structure;

determining a convolution network structure set to be screened in the plurality of amplification convolution network structures according to the service demand information, and determining a reference convolution network structure in the deep learning network structure set to be screened;

and generating a target convolution network structure corresponding to the target image processing service based on the reference convolution network structure and a preset network generation condition.

According to a fourth aspect of embodiments of the present specification, there is provided a generation system of a deep learning network structure, including:

the cloud side equipment is used for acquiring service demand information corresponding to an initial deep learning network structure and target service, wherein the service demand information comprises a deep learning network structure parameter threshold value, a plurality of amplified deep learning network structures are generated based on the initial deep learning network structure, a deep learning network structure set to be screened is determined in the amplified deep learning network structures according to the service demand information, a reference deep learning network structure is determined in the deep learning network structure set to be screened, a target deep learning network structure corresponding to the target service is generated based on the reference deep learning network structure and preset network generation conditions, and a target service deep learning network model is trained for the target service based on the target deep learning network structure;

And the terminal side equipment is used for executing the target service according to the target service deep learning network model to obtain a processing result corresponding to the target service.

According to a fifth aspect of embodiments of the present specification, there is provided a computing device comprising:

a memory and a processor;

the memory is configured to store computer-executable instructions that, when executed by the processor, perform the steps of the method for generating a deep learning network structure described above.

According to a sixth aspect of embodiments of the present specification, there is provided a computer-readable storage medium storing computer-executable instructions which, when executed by a processor, implement the steps of the method of generating a deep learning network structure described above.

According to a seventh aspect of the embodiments of the present specification, there is provided a computer program, wherein the computer program, when executed in a computer, causes the computer to perform the steps of the above-described deep learning network structure generation method.

One embodiment of the specification realizes obtaining service demand information corresponding to an initial deep learning network structure and a target service, wherein the service demand information comprises a deep learning network structure parameter threshold value; generating a plurality of amplified deep learning network structures based on the initial deep learning network structure; determining a deep learning network structure set to be screened in the multiple amplified deep learning network structures according to the service demand information, and determining a reference deep learning network structure in the deep learning network structure set to be screened; and generating a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

According to the method provided by the embodiment of the specification, after the initial deep learning network structure is subjected to mutation amplification, a plurality of amplified deep learning network structures are obtained, the data base of selecting a better network structure is improved, then the plurality of amplified deep learning network structures are screened according to service demand information (particularly, the deep learning network structure parameter threshold value in the service demand information) set by a technician for target service, the network structures which do not meet the actual requirements of the target service are screened, the data amount during subsequent selection is reduced, a reference deep learning network structure is determined in a deep learning network structure set to be screened, iterative updating is performed according to the reference deep learning network structure and preset network generation conditions, and finally the target deep learning network structure corresponding to the target service is generated. The finally generated target deep learning network structure meets the service requirement of the target service and is more suitable for the actual application of the target service.

Drawings

Fig. 1 is a schematic diagram of a method for generating a deep learning network structure for use in one embodiment of the present disclosure;

FIG. 2 is a flowchart of a method for generating a deep learning network structure according to one embodiment of the present disclosure;

FIG. 3 is a process flow diagram of a method for generating a deep learning network structure for application to image recognition scenes according to one embodiment of the present disclosure;

FIG. 4 is a flow chart of a method of generating an image processing network architecture provided in one embodiment of the present disclosure;

FIG. 5 is a schematic diagram of a system for generating a deep learning network architecture according to one embodiment of the present disclosure;

fig. 6 is a schematic structural diagram of a generating device of a deep learning network structure according to an embodiment of the present disclosure;

FIG. 7 is a block diagram of a computing device provided in one embodiment of the present description.

Detailed Description

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present description. This description may be embodied in many other forms than described herein and similarly generalized by those skilled in the art to whom this disclosure pertains without departing from the spirit of the disclosure and, therefore, this disclosure is not limited by the specific implementations disclosed below.

The terminology used in the one or more embodiments of the specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of the specification. As used in this specification, one or more embodiments and the appended claims, the singular forms "a," "an," and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used in one or more embodiments of the present specification refers to and encompasses any or all possible combinations of one or more of the associated listed items.

It should be understood that, although the terms first, second, etc. may be used in one or more embodiments of this specification to describe various information, these information should not be limited by these terms. These terms are only used to distinguish one type of information from another. For example, a first may also be referred to as a second, and similarly, a second may also be referred to as a first, without departing from the scope of one or more embodiments of the present description. The word "if" as used herein may be interpreted as "at … …" or "at … …" or "responsive to a determination", depending on the context.

First, terms related to one or more embodiments of the present specification will be explained.

CNN: convolutional neural network, convolutional neural network.

And (3) designing a neural network structure: in the embodiments provided in the present specification, the optimized design of the CNN neural network structure based on deep learning is referred to.

Aspect ratio of convolution structure: the embodiment provided in the specification provides a deep learning network structure ratio mode of depth to width.

Along with the rapid development of deep learning technology, more and more visual intelligent services based on deep learning are widely applied to a plurality of scenes, so that diversified experiences are brought to users, the quality of visual intelligent effects based on deep learning is closely related to the neural network structure, and the neural network structures required under different service scenes are different, therefore, based on the fact, the generation method of the neural network structure is needed to be provided, and the method can be suitable for various service scenes, and corresponding neural network structures are generated for each service scene.

Based on this, in the present specification, a method of generating a deep learning network structure is provided, and the present specification relates to a generating apparatus of a deep learning network structure, a computing device, and a computer-readable storage medium, which are described in detail in the following embodiments one by one.

Referring to fig. 1, fig. 1 is a schematic diagram illustrating a method for generating a deep learning network structure according to an embodiment of the present disclosure, as shown in fig. 1, and as shown in fig. 1, the method for generating a deep learning network structure according to the embodiment of the present disclosure is applied to a terminal 100, where the terminal 100 may be a terminal device such as a notebook computer, an intelligent terminal, a server, a cloud server, or the like.

In the terminal 100, an initial deep learning network structure is determined, then the super parameters in the initial deep learning network structure are adjusted to generate a deep learning network structure after each adjustment, the adjusted deep learning network structure and the initial deep learning network structure are used as a plurality of amplified deep learning network structures, and the amplified deep learning network structure is used as a data base for determining a target network structure subsequently.

And after obtaining the multiple amplified deep learning network structures, screening the multiple amplified deep learning network structures according to service requirement information corresponding to the target service (the service requirement information comprises a deep learning network structure parameter threshold value, the deep learning network structure parameter threshold value comprises a deep learning network structure length-width ratio threshold value), wherein in practical application, the service requirement information can comprise a deep learning network parameter threshold value, a deep learning network calculated amount threshold value and the like besides the deep learning network structure length-width ratio threshold value. In the design of the deep learning network structure, the aspect ratio of the deep learning network structure is found to play an important role in the performance of the deep learning network, and the aspect ratio of different deep learning network structures in each service scene brings different network performances to the deep learning network model. In the embodiment provided in the specification, the aspect ratio of the specified deep learning network structure is referred to as an index for screening the deep learning network structure, and a new reference index is provided for generating the deep learning network structure.

And screening the amplified deep learning network structures by using the service demand information to obtain a plurality of deep learning network structures to be screened, grading each deep learning network structure to be screened by a network structure grading method, and selecting the network structure with higher grading as a reference deep learning network structure.

And taking the reference deep learning network structure as a new initial deep learning network structure, and continuing to iterate the modes of amplifying, screening, grading and determining the new reference deep learning network structure until the reference deep learning network structure meets the preset network generation condition, namely, the iterating process is carried out until the network structure is not changed. Thus, the target deep learning network structure corresponding to the target service can be obtained. The target deep learning network structure can be used for training a corresponding deep learning network model for the target service.

In practical application, the preferred deep learning network is a convolutional neural network, and correspondingly, the deep learning network structure length-width ratio threshold is a convolutional structure length-width ratio threshold, the deep learning network parameter threshold is a convolutional parameter threshold, and the deep learning network calculated amount threshold is a convolutional calculated amount threshold.

Referring to fig. 2, fig. 2 shows a flowchart of a method for generating a deep learning network structure according to an embodiment of the present disclosure, which specifically includes the following steps.

Step 202: and acquiring service demand information corresponding to the initial deep learning network structure and the target service, wherein the service demand information comprises a deep learning network structure parameter threshold value.

The initial deep learning network structure specifically refers to a basis of a network structure for subsequent amplification, for example, when the network structure is subsequently amplified, the deep learning network structure a is based on the deep learning network structure a, and then the deep learning network structure a is the initial deep learning network structure. In practical applications, the deep learning network structure may be any network structure, such as a convolutional network structure, a residual network structure, a recurrent neural network structure, a fully connected network structure, and so on.

The target service specifically refers to different services in practical application, such as an image recognition service, an image segmentation service, a target tracking service and the like, and the service demand information specifically refers to a resource budget which can be provided for the target service by each target service, wherein the resource budget comprises a network parameter value, a calculated amount value, a resource value, a deep learning network structure length-width ratio threshold value and the like. In other words, the service requirement information specifies the upper processing limit of the deep learning network structure corresponding to the target service. For example, in an actual business scenario, the memory budget that may be provided for the deep learning network structure is 2G, the calculated amount is 100, and the aspect ratio threshold of the deep learning network structure is 3. When the deep learning network structure is provided for the target service later, the memory occupied space of the deep learning network structure is ensured to be within 2G, the calculated amount is within 100, and the aspect ratio of the deep learning network structure is less than or equal to 3, and the like.

In the specific embodiment provided in the present specification, taking a convolutional network structure as an example, the aspect ratio of the specified deep learning network structure is determined according to the number of convolutional network layers, the size of the convolutional kernel and the number of convolutional channels, and the number of layers of the convolutional network, the size of the convolutional kernel and the number of convolutional channels correspond to the aspect ratios of different convolutional structures. The convolutional network structure may be screened based on the aspect ratio of the convolutional structure.

The obtaining the service demand information corresponding to the target service includes:

receiving a budget instruction aiming at a target service, wherein the budget instruction carries a deep learning network structure parameter condition;

and analyzing the deep learning network structure parameter condition to obtain a deep learning network structure parameter threshold corresponding to the target service.

In practical application, the service requirement information is set by a technician according to the actual situation of the target service, and further, the technician sends a budget command for the target service to the terminal, wherein the budget command carries preset deep learning network structure parameter conditions, and the deep learning network structure parameter conditions include a deep learning network structure length-width ratio threshold value, a deep learning network parameter threshold value, a deep learning network calculated amount threshold value and the like.

After receiving the budget command, the terminal analyzes the deep learning network structure parameter conditions in the budget command to obtain a deep learning network structure parameter threshold corresponding to the target service. Taking a convolutional network structure as an example, the deep learning network structure parameter condition comprises a convolutional network structure length-width ratio threshold, a convolutional network parameter threshold and a convolutional network calculated amount threshold, wherein the convolutional network structure length-width ratio is determined according to the number of convolutional network layers, the convolutional kernel size and the convolutional channel number.

In a specific embodiment provided in the present specification, taking a target service as an image recognition service as an example, an initial deep learning network structure a and service requirement information allocated by a technician for the image recognition service are obtained, where an aspect ratio threshold of the deep learning network structure specified in the service requirement information is 2.

Step 204: generating a plurality of amplified deep learning network structures based on the initial deep learning network structure.

The amplified deep learning network structure may be understood as a deep learning network structure generated based on the initial deep learning network structure, and in practical application, the plurality of amplified deep learning network structures include the initial deep learning network structure.

Furthermore, in practical application, the initial deep learning network structure includes a plurality of parameters, different parameters are adjusted differently, so that different new deep learning network structures are generated, and each time the parameters of the network structures are adjusted, the capability of the deep learning network structure for processing the target service is different.

The parameters important in determining the parameters of the deep learning network structure are super parameters of the model, wherein the super parameters refer to parameters preset in the deep learning network structure, not parameters adjusted through model training, the super parameters are used for defining higher-layer concepts of the model, and different values of the super parameters can be set to obtain different network structures. Based on this, generating a plurality of augmented deep learning network structures based on the initial deep learning network structure, comprising:

adjusting super parameters in the initial deep learning network structure, wherein the super parameters comprise at least one of the number of convolution channels, the size of convolution kernels and the number of convolution network layers;

Generating an amplified deep learning network structure corresponding to each adjustment.

Specifically, because the super parameters of the deep learning network structure can be adjusted, different deep learning network structures can be generated, in order to obtain a greater number of deep learning network structures, a data base is provided for a structure with a better subsequent selection effect, in the method provided in the specification, the super parameters in the initial deep learning network structure can be correspondingly adjusted, and specifically, the super parameters can include the number of deep learning network channels, the number of deep learning network layers and the like.

Taking the initial deep learning network structure a as a convolutional neural network structure as an example, the number of convolutional channels of the convolutional network structure a is 3, the size of a convolutional kernel is 1*1, and the number of layers of the convolutional network is 4. Based on the initial deep learning network structure a, the number of convolution channels is adjusted to be 5, the adjusted convolution network structure a1, the convolution channel data of a1 is 5, and other super parameters are the same as a; based on the initial deep learning network structure a, the convolution kernel size is adjusted to 3*3, the adjusted convolution network structure a2, the convolution kernel size of a2 is 3*3, and other super parameters are the same as a; based on the initial deep learning network structure a, the number of layers of the convolution network is adjusted to 5, the adjusted convolution network structure a3 can be obtained, the number of layers of the convolution network of the a3 is 5, and other super parameters are the same as a. Up to this point, 4 amplified convolutional network structures, amplified convolutional network structures a, a1, a2 and a3, respectively, can be obtained. In practical applications, at least two super parameters may be adjusted in combination, for example, the number of convolution channels and the size of convolution kernels in the convolution network structure may be adjusted simultaneously, the number of convolution channels and the number of convolution network layers may be adjusted simultaneously, or the number of convolution channels, the size of convolution kernels and the number of convolution network layers may be adjusted simultaneously.

It should be noted that, in order to ensure randomness and universality of the amplified deep learning network structure, in practical application, the super parameters in the initial deep learning network structure can be randomly adjusted, and based on this, a large number of amplified deep learning network structures can be obtained. Thereby providing a huge data base for subsequent screening.

Step 206: and determining a deep learning network structure set to be screened in the multiple amplified deep learning network structures according to the service demand information, and determining a reference deep learning network structure in the deep learning network structure set to be screened.

After the service demand information and the multiple amplified deep learning network structures are determined, screening can be performed in the multiple amplified deep learning network structures according to the service demand information, and the amplified deep learning network structures conforming to the service demand information are used as a set of deep learning network structures to be screened, wherein the set of deep learning network structures to be screened can be regarded as a set of deep learning network structures meeting the service demand information.

Specifically, determining a deep learning network structure set to be screened from the multiple amplified deep learning network structures according to the service demand information includes:

Obtaining target network structure parameters of a target amplification deep learning network structure;

and adding the target amplified deep learning network structure to a deep learning network structure set to be screened under the condition that the target network structure parameters meet the service demand information.

In the embodiment provided in the present specification, taking a convolutional neural network as an example, obtaining a target network structure parameter of a target amplified deep learning network structure includes:

and obtaining the target convolution network layer number, the target convolution kernel size, the target convolution channel number, the target convolution parameters and the target convolution calculated amount of the target amplification deep learning network structure.

In the above step, it is known that the service requirement information is a condition for screening the deep learning network structure, and accordingly, the network structure parameters in the amplified deep learning network structure need to be obtained according to the condition required by the service requirement information, for example, the service requirement information needs to use parameter information such as the number of layers of the convolutional network, the size of the convolutional kernel, the number of convolutional channels, the convolutional parameters, the convolutional calculation amount, and the like, and then the parameter information in the amplified deep learning network structure needs to be obtained.

Specifically, the target amplified deep learning network structure may be understood as a network structure selected from a plurality of amplified deep learning network structures, where the target amplified deep learning network structure is a network structure that is compared with the service requirement information.

After determining the target amplification deep learning network structure, obtaining target network structure parameters of the target amplification deep learning network structure, further, obtaining the number of target convolution network layers, the size of target convolution kernels, the number of target convolution channels, the target convolution parameters and the target convolution calculated amount in the target amplification deep learning network structure, judging whether the target amplification deep learning network structure meets the service requirement information or not based on the target network structure parameters, if so, adding the target amplification deep learning network structure into a deep learning network structure set to be screened, and if not, discarding the target amplification deep learning network structure.

Further, after obtaining the target network structure parameter of the target amplified deep learning network structure, the method further includes:

determining a target convolution structure length-width ratio of the target amplification deep learning network structure according to the target convolution network layer number, the target convolution kernel size and the target convolution channel number;

and determining that the target network structure parameter meets the service requirement information under the conditions that the target convolution structure length-width ratio is smaller than or equal to the convolution structure length-width ratio threshold, the target convolution parameter is smaller than or equal to the convolution parameter threshold and the target convolution calculated amount is smaller than or equal to the convolution calculated amount threshold.

In the step, it has been determined that the service requirement information carries a convolution structure length-width ratio threshold, and the target amplification deep learning network structure can be screened according to the convolution structure length-width ratio threshold, and the convolution structure length-width ratio is determined according to the number of convolution network layers, the convolution kernel size and the number of convolution channels, so that in the acquired target network structure parameters, the target convolution structure length-width ratio of the target amplification deep learning network structure can be determined according to the number of target convolution network layers, the target convolution kernel size and the number of target convolution channels.

Specifically, determining the target convolution structure aspect ratio of the target amplification deep learning network structure according to the target convolution network layer number, the target convolution kernel size and the target convolution channel number comprises the following steps:

determining a convolution structure length value of the target amplification deep learning network structure according to the target convolution network layer number;

determining a convolution structure width value of the target amplification deep learning network structure according to the target convolution kernel size and the target convolution channel number;

and determining the aspect ratio of the target convolution structure according to the convolution structure length value and the convolution structure width value.

In the embodiment provided in the present specification, taking a convolutional network structure as an example, the convolutional network structure is abstracted into a structure body with a length and a width, and a convolutional structure length value of the target amplified convolutional network structure is determined according to the number of target convolutional network layers; and determining a convolution structure width value of the target amplification convolution network structure according to the target convolution kernel size and the target convolution channel number.

The calculation mode of the length value of the convolution structure is shown in the following formula 1:

wherein L is _j Representing the j-th phase module depth, n representing the current network fabric phase number.

The calculation mode of the width value of the convolution structure is shown in the following formula 2:

wherein C is _i Indicating the number of convolved channels of the ith convolved network layer,representing the convolution kernel size of the ith convolution network layer.

After the length value and the width value of the convolution structure are determined, the length-width ratio of the convolution structure can be determined according to the ratio between the length value and the width value of the convolution structure, the length-width ratio of the convolution structure is compared with the length-width ratio threshold of the convolution structure in the service requirement information, and if the length-width ratio of the convolution structure is smaller than or equal to the length-width ratio threshold of the convolution structure, the target expansion convolution network structure is indicated to meet the limitation of the length-width ratio threshold of the convolution structure in the service requirement information.

Specifically, the comparison of the target network structure parameter of the target amplification convolutional network structure and the service requirement information can be seen from the following formulas 3-1 to 3-3:

size of params≤budget _params equation 3-2

size of FLOPs≤budget _FLOPs Equation 3-3

Wherein, the formulas 3-1 to 3-3 are parallel relations, and the target network structure parameters of the target amplification convolution network structure need to satisfy the formulas 3-1 to 3-3 at the same time. Specifically, L _j Represents the module depth of the jth stage, N represents the current network structure stage number, N represents the total stage number of the network structure, and C _i Indicating the number of convolved channels of the ith convolved network layer,a convolution kernel size, ρ, representing the ith convolution network layer _h Representing the aspect ratio threshold of a convolution structure, wherein size of params is the parameter quantity of convolution operation in a target amplification convolution network structure, and the size is the widget _params For the parameter quantity threshold value of convolution operation in service demand information, size of FLPs is the calculated quantity of convolution operation in target augmentation convolution network structure, and the bandwidth is used for the calculation quantity of convolution operation in target augmentation convolution network structure _FLOPs A calculated threshold for convolution operation in the business requirement information, as shown in formula 3, when the target convolution structure length-width ratio is less than or equal to the convolution structure length-width ratio threshold, the target convolution structure length-width ratioAnd under the condition that the target convolution parameter is smaller than or equal to the convolution parameter threshold and the target convolution calculated amount is smaller than or equal to the convolution calculated amount threshold, determining that the target network structure parameter meets the service requirement information, and adding the target amplification convolution network structure to a convolution network structure set to be screened.

In one embodiment provided in the present specification, taking the example of the amplified convolutional network structure (a, a1, a2, a3, a4 … … a 20), after comparing with the service requirement information respectively, the convolutional network structure according with the service requirement information is added to the convolutional network structure set to be screened, and based on this, the convolutional network structure set to be screened (a, a1, a3, a4, a7, a9, a13, a15, a17, a 20) can be obtained. The convolution network structures in the convolution network structure set to be screened are all convolution network structures which meet the service requirement information.

After obtaining the set of deep learning network structures to be screened, determining a reference deep learning network structure in the set of deep learning network structures to be screened, which specifically comprises:

scoring each deep learning network structure to be screened in the deep learning network structure set to be screened;

and taking the deep learning network structure to be screened with the highest score as a reference deep learning network structure.

Because the network structure sets to be screened are all deep learning network structures which meet the service requirement information, the deep learning network structures to be screened with better performance are selected in the network structure sets to be screened for iterative screening again. Specifically, each deep learning network structure to be screened needs to be scored, and in the embodiment provided in this specification, scoring is preferably performed based on the maximum entropy method, specifically, see the following formula 4:

Wherein,,representing the height of the output characteristic image of the nth stage,/->Representing the width of the output characteristic image of the nth stage,/-, and>representing the number of channels of the output characteristic image of the nth stage, alpha ₁ 、α ₂ 、ω _n Representing network structure superparameter, E _n Is the score of the nth stage, +.>The weighted average of N stages is the total score of the deep learning network structure to be screened.

After calculating the score of each deep learning network structure to be screened, selecting one deep learning network structure to be screened with the highest score as a reference deep learning network structure, wherein the reference deep learning network structure specifically refers to a deep learning network structure for subsequent processing.

In the embodiment provided in the present specification, the above example is used, the set of deep learning network structures to be screened is (a, a1, a3, a4, a7, a9, a13, a15, a17, a 20), after each network structure to be screened is scored, the network structures to be screened are ranked according to the order from high score to low score, and the obtained set of network structures is (a 17, a13, a9, a15, a1, a4, a3, a7, a20, a), that is, the score of a17 is the highest, and then a17 is the reference deep learning network structure.

Step 208: and generating a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

After the reference deep learning network structure is determined, the reference deep learning network structure and the preset network generation condition can be compared, if the reference deep learning network structure meets the preset network generation condition, the reference deep learning network structure can be selected as a final target deep learning network structure, otherwise, iteration optimization processing is continuously carried out according to the reference deep learning network structure.

In a specific embodiment provided in the present disclosure, after the reference deep learning network structure is obtained, it is further required to determine whether the reference deep learning network structure meets a preset network generation condition, and specifically, after determining the reference deep learning network structure in the deep learning network structure set to be screened, the method further includes:

counting the continuous iteration hit times of the reference deep learning network structure;

and under the condition that the continuous iteration hit times reach a preset threshold value, determining that the reference deep learning network structure meets a preset network generation condition.

The number of continuous iterative hits specifically refers to counting whether the reference deep learning network structure is selected as the reference deep learning network structure in multiple iterative optimization, for example, the initial deep learning network structure is a, after the first iterative optimization, the deep learning network structure b is determined, after the second iterative optimization based on b, the deep learning network structure c is determined, after the third iterative optimization based on c, the deep learning network structure d is determined, after the fourth iterative optimization based on d, the deep learning network structure d is determined, and after the fifth iterative optimization based on d, the determined deep learning network structure d is still determined, and at this time, the number of continuous iterative hits of the deep learning network structure d is three.

If the preset threshold is 3 times, it can be determined that the deep learning network structure d meets the preset network generation condition, and the previous deep learning network structures b and c are all not meeting the preset network generation condition.

Further, generating the target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition includes:

s2082: and under the condition that the reference deep learning network structure does not meet the preset network generation condition, selecting at least one deep learning network structure to be processed from the deep learning network structure set to be screened, taking each deep learning network structure to be processed as an initial deep learning network structure, and continuously executing the operation of generating a plurality of amplified deep learning network structures.

If the reference deep learning network structure does not meet the preset network generation condition, iterative optimization processing is needed again, specifically, at least one deep learning network structure to be processed is selected from a deep learning network structure set to be screened, for example, the first 20 deep learning network structures to be screened with scores arranged from high to low are selected as the deep learning network structures to be processed, each deep learning network structure to be processed is used as an initial deep learning network structure, a plurality of amplified deep learning network structures are continuously executed, the deep learning network structure set to be screened is determined in the amplified deep learning network structures according to the service requirement information, the reference deep learning network structure is determined in the deep learning network structure set to be screened, and whether the reference convolution network set meets the preset network generation condition is judged.

In a specific embodiment provided in this specification, along the above example, after the initial deep learning network structure is a and the first iterative optimization, the deep learning network structure b is determined, and when the deep learning network structure b does not meet the preset network generation condition, the deep learning network structure b1, the deep learning network structure b2, the deep learning network structure b3, and the deep learning network structure b4 are used as the deep learning network structures to be processed, where b1, b2, b3, and b4 are the deep learning network structures with 5 bits before scoring. And taking each to-be-processed deep learning network structure as an initial deep learning network structure, randomly changing super parameters in the deep learning network structures b, b1, b2, b3 and b4, and generating a plurality of amplified deep learning network structures corresponding to the deep learning network structures b, b1, b2, b3 and b 4. And determining a new deep learning network structure set to be screened in the multiple amplified deep learning network structures according to the service demand information, and determining a reference deep learning network structure c in the new deep learning network structure set to be screened. And judging whether the c meets the preset network generation condition again.

S2084: and under the condition that the reference deep learning network structure meets the preset network generation condition, determining the reference deep learning network structure as a target deep learning network structure corresponding to the target service.

If the reference deep learning network structure meets the preset network generation condition, namely the reference deep learning network structure tends to be stable after repeated iterative optimization, the reference deep learning network structure can be determined to be a relatively stable deep learning network structure, the reference deep learning network structure meets the preset network generation condition, and the reference deep learning network structure can be further used as a target deep learning network structure corresponding to the target service.

In a specific embodiment provided in the present specification, in the above example, in three continuous iterative optimization, the deep learning network structure d is the deep learning network structure d, which indicates that the deep learning network structure d is already stable at this time, and can be used as the target deep learning network structure corresponding to the target service.

After determining the target deep learning network structure, the corresponding target service deep learning network model can be trained for the target service based on the target deep learning network structure, and in the embodiment provided in the specification, the method further includes:

And training a target service deep learning network model for the target service based on the target deep learning network structure.

Specifically, training data and training labels corresponding to target services are obtained, the training data are input into a deep learning network model based on a target deep learning network structure for processing, a prediction result output by the deep learning network model is obtained, a model loss value is calculated according to the prediction result and the training labels, and model parameters of the deep learning network model are adjusted based on the model loss value to be back propagated in the deep learning network model until model training stop conditions are reached.

In addition, the embodiment of the specification defines the length and the width corresponding to the convolution network structure, and screens the aspect ratio of the deep learning network structure as a constraint condition, so that the network structure conforming to the expected aspect ratio is selected, compared with a manually designed network, the performance is greatly improved, and a better network structure can be selected under the service requirement information corresponding to the target service.

The method for generating the deep learning network structure provided in the present specification will be further described with reference to fig. 3, by taking an application of the method for generating the deep learning network structure in an image recognition scene as an example. Fig. 3 is a flowchart illustrating a process of a method for generating a deep learning network structure according to an embodiment of the present disclosure, where the deep learning network structure applied to an image recognition scene is a convolutional network structure, and specifically includes the following steps.

Step 302: and acquiring service demand information corresponding to the initial convolution network structure and the image identification service.

The business requirement information comprises a convolution structure length-width ratio threshold value, a convolution parameter threshold value and a convolution calculated quantity threshold value.

Step 304: and adjusting the super parameters in the initial convolution network structure to obtain the corresponding amplified convolution network structure after each adjustment.

Wherein the super-parameters comprise at least one of the number of convolution channels, the size of convolution kernel and the number of convolution network layers.

Step 306: and determining the target convolution network layer number, the target convolution kernel size, the target convolution channel number, the target convolution parameters and the target convolution calculated amount of the target amplification convolution network structure.

Step 308: and determining a convolution structure length value according to the number of target convolution network layers, and determining a convolution structure width value according to the size of the target convolution kernel and the number of target convolution channels.

Step 310: and determining the aspect ratio of the target convolution structure according to the convolution structure length value and the convolution structure width value.

Step 312: and adding the target amplification convolution network structure to the convolution network structure set to be screened under the conditions that the aspect ratio of the target convolution structure is smaller than or equal to the aspect ratio threshold of the convolution structure, the target convolution parameter is smaller than or equal to the convolution parameter threshold and the target convolution calculated amount is smaller than or equal to the convolution calculated amount threshold.

Step 314: and scoring each convolution network structure to be screened in the convolution network structure set to be screened.

Step 316: and taking the convolution network structure to be screened with the highest score as a reference convolution network structure.

Step 318: and counting the number of continuous iteration hits of the reference convolution network structure.

Step 320: whether the number of continuous iteration hits is greater than or equal to a preset threshold is determined, if yes, step 322 is executed, if not, at least one initial convolutional network structure is selected from the set of convolutional network structures to be screened, and step 304 is executed continuously.

Step 322: and determining the reference convolution network structure as a target convolution network structure corresponding to the image recognition service.

Step 324: and acquiring image training data corresponding to the image recognition service, and training an image recognition model based on the target convolutional network structure as a backbone network according to the image training data.

According to the method provided by the embodiment of the specification, after the mutation amplification is carried out according to the initial convolution network structure, a plurality of amplified convolution network structures are obtained, the data base of selecting a better network structure is improved, then the plurality of amplified convolution network structures are screened according to service demand information (specifically, the aspect ratio threshold of the convolution structure in the service demand information) set by a technician, the network structure which does not meet the actual demand of the target service is screened, the data quantity during subsequent selection is reduced, one reference convolution network structure is determined in a convolution network structure set to be screened, whether the reference convolution network structure meets the preset network generation condition is judged, if yes, the reference convolution network structure is used as the target convolution network structure, if not, at least one initial convolution network structure is selected in the convolution network structure set to be screened, and the operation is repeated.

Referring to fig. 4, fig. 4 is a flowchart showing a method for generating an image processing network structure according to an embodiment of the present disclosure, and specifically includes the following steps:

step 402: and acquiring service demand information corresponding to the initial convolution network structure and the target image processing service, wherein the service demand information comprises a convolution network structure length-width ratio threshold value, and the convolution structure length-width ratio threshold value is determined according to the number of convolution network layers, the convolution kernel size and the convolution channel number.

Step 404: generating a plurality of augmented convolutional network structures based on the initial convolutional network structure.

Step 406: and determining a convolution network structure set to be screened in the plurality of amplification convolution network structures according to the service demand information, and determining a reference convolution network structure in the deep learning network structure set to be screened.

Step 408: and generating a target convolution network structure corresponding to the target image processing service based on the reference convolution network structure and a preset network generation condition.

Specifically, based on the reference convolutional network structure and a preset network generation condition, generating a target convolutional network structure corresponding to the target image processing service includes:

Under the condition that the reference convolution network structure does not meet the preset network generation condition, selecting at least one convolution network structure to be processed from the convolution network structure set to be screened, taking each convolution network structure to be processed as an initial convolution network structure, and continuously executing the operation of generating a plurality of expansion convolution network structures;

and under the condition that the reference convolution network structure meets the preset network generation condition, determining the reference convolution network structure as a target convolution network structure corresponding to the target image processing service.

Referring to fig. 5, fig. 5 shows a schematic diagram of a deep learning network structure generating system provided in an embodiment of the present disclosure, where the system may include a cloud side device 501 and an end side device 502, where the cloud side device 501 is configured to select a target deep learning network structure corresponding to a target service, and generate a target service deep learning network model corresponding to the target service based on training of the target deep learning network structure, and the end side device 502 is configured to execute the target service based on the training obtained target service deep learning network model, and specifically, the deep learning network structure may be a convolutional network structure, a residual network structure, a cyclic neural network structure, a fully connected network structure, or the like, and the target service may be a target object detection task, an image recognition task, an image segmentation task, a speech recognition service, a speech processing service, or the like.

The cloud side device 501 is configured to obtain service requirement information corresponding to an initial deep learning network structure and a target service, where the service requirement information includes a deep learning network structure parameter threshold, generate a plurality of amplified deep learning network structures based on the initial deep learning network structure, determine a set of deep learning network structures to be screened among the plurality of amplified deep learning network structures according to the service requirement information, determine a reference deep learning network structure among the set of deep learning network structures to be screened, generate a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition, and train a target service deep learning network model for the target service based on the target deep learning network structure;

and the end side device 502 is configured to execute the target service according to the target service deep learning network model, and obtain a processing result corresponding to the target service.

Cloud-side device 501 may be a central cloud device of a distributed cloud architecture, end-side device 502 may be an edge cloud device of a distributed cloud architecture, cloud-side device 501 and end-side device 502 may be service end devices such as a conventional server, a cloud server or a server array, or may be terminal devices, which is not limited in this embodiment of the present disclosure. Moreover, cloud-side device 501 provides superior computing and storage capabilities, remote from the user; while the end-side device 502 is deployed in a large range, closer to the user. The end-side device 502 is an expansion of the cloud-side device 501, and can sink the computing capability of the cloud-side device 501 to the end-side device 502, and solve the service requirements which cannot be met in the centralized cloud computing mode through the integration and collaborative management of the end cloud.

Corresponding to the method embodiment, the present disclosure further provides an embodiment of a deep learning network structure generating device, and fig. 6 shows a schematic structural diagram of a deep learning network structure generating device provided in one embodiment of the present disclosure. As shown in fig. 6, the apparatus includes:

an obtaining module 602, configured to obtain service requirement information corresponding to an initial deep learning network structure and a target service, where the service requirement information includes a deep learning network structure parameter threshold;

an augmentation module 604 configured to generate a plurality of augmented deep learning network structures based on the initial deep learning network structure;

a determining module 606 configured to determine a set of deep learning network structures to be screened among the plurality of amplified deep learning network structures according to the traffic demand information, and determine a reference deep learning network structure among the set of deep learning network structures to be screened;

the generating module 608 is configured to generate a target deep learning network structure corresponding to the target service based on the reference deep learning network structure and a preset network generation condition.

Optionally, the generating module 608 is further configured to:

Selecting at least one deep learning network structure to be processed from the deep learning network structure set to be screened under the condition that the reference deep learning network structure does not meet the preset network generation condition, taking each deep learning network structure to be processed as an initial deep learning network structure, and continuously executing the operation of generating a plurality of amplified deep learning network structures;

and under the condition that the reference deep learning network structure meets the preset network generation condition, determining the reference deep learning network structure as a target deep learning network structure corresponding to the target service.

Optionally, the obtaining module 602 is further configured to:

Optionally, the deep learning network structure parameter condition includes a convolutional network structure length-width ratio threshold, a convolutional network parameter threshold and a convolutional network calculated amount threshold, wherein the convolutional network structure length-width ratio is determined according to the convolutional network layer number, the convolutional kernel size and the convolutional channel number.

Optionally, the determining module 606 is further configured to:

Optionally, the deep learning network structure parameter threshold includes a convolutional network structure aspect ratio threshold, a convolutional network parameter threshold, and a convolutional network calculation amount threshold;

the determining module 606 is further configured to:

Optionally, the determining module 606 is further configured to:

Optionally, the apparatus further includes:

a statistics module configured to count a number of successive iterative hits of the reference deep learning network structure;

accordingly, the determining module 606 is further configured to:

Optionally, the amplification module 604 is further configured to:

Optionally, the apparatus further includes:

and the training module is configured to train a target service deep learning network model for the target service based on the target deep learning network structure.

According to the device provided by the embodiment of the specification, after the initial deep learning network structure is subjected to mutation amplification, a plurality of amplified deep learning network structures are obtained, the data base of selecting a better network structure is improved, the plurality of amplified deep learning network structures are screened according to service demand information (specifically, the aspect ratio threshold of a convolution structure in the service demand information) set by a technician for target service, the network structures which do not meet the actual demand of the target service are screened, the data amount during subsequent selection is reduced, one reference deep learning network structure is determined in a to-be-screened deep learning network structure set, whether the reference deep learning network structure meets the preset network generation condition is judged, if yes, the reference deep learning network structure is used as the target deep learning network structure, and if not, the reference deep learning network structure is used as the new initial deep learning network structure, and the operation is repeated.

According to the embodiment of the specification, the length and the width corresponding to the convolution network structure are defined, the length-width ratio of the deep learning network structure is used as a constraint condition to screen, so that the network structure conforming to the expected length-width ratio is selected, compared with a manually designed network, the performance is greatly improved, and a better network structure can be selected under the service requirement information corresponding to the target service.

The above is a schematic scheme of the generation apparatus of the deep learning network structure of the present embodiment. It should be noted that, the technical solution of the deep learning network structure generating device and the technical solution of the deep learning network structure generating method belong to the same concept, and details of the technical solution of the deep learning network structure generating device which are not described in detail can be referred to the description of the technical solution of the deep learning network structure generating method.

Fig. 7 illustrates a block diagram of a computing device 700 provided in accordance with one embodiment of the present description. The components of computing device 700 include, but are not limited to, memory 710 and processor 720. Processor 720 is coupled to memory 710 via bus 730, and database 750 is used to store data.

Computing device 700 also includes access device 740, access device 740 enabling computing device 700 to communicate via one or more networks 760. Examples of such networks include public switched telephone networks (PSTN, public Switched Telephone Network), local area networks (LAN, local Area Network), wide area networks (WAN, wide Area Network), personal area networks (PAN, personal Area Network), or combinations of communication networks such as the internet. The access device 740 may include one or more of any type of network interface, wired or wireless, such as a network interface card (NIC, network interface controller), such as an IEEE802.11 wireless local area network (WLAN, wireless Local Area Network) wireless interface, a worldwide interoperability for microwave access (Wi-MAX, worldwide Interoperability for Microwave Access) interface, an ethernet interface, a universal serial bus (USB, universal Serial Bus) interface, a cellular network interface, a bluetooth interface, near field communication (NFC, near Field Communication).

In one embodiment of the present description, the above-described components of computing device 700, as well as other components not shown in FIG. 7, may also be connected to each other, such as by a bus. It should be understood that the block diagram of the computing device illustrated in FIG. 7 is for exemplary purposes only and is not intended to limit the scope of the present description. Those skilled in the art may add or replace other components as desired.

Computing device 700 may be any type of stationary or mobile computing device, including a mobile computer or mobile computing device (e.g., tablet, personal digital assistant, laptop, notebook, netbook, etc.), mobile phone (e.g., smart phone), wearable computing device (e.g., smart watch, smart glasses, etc.), or other type of mobile device, or a stationary computing device such as a desktop computer or personal computer (PC, personal Computer). Computing device 700 may also be a mobile or stationary server.

Wherein the processor 720 is configured to execute computer-executable instructions that, when executed by the processor, perform the steps of the data processing method described above. The foregoing is a schematic illustration of a computing device of this embodiment. It should be noted that, the technical solution of the computing device and the technical solution of the generating method of the deep learning network structure belong to the same concept, and details of the technical solution of the computing device, which are not described in detail, can be referred to the description of the technical solution of the generating method of the deep learning network structure.

An embodiment of the present disclosure also provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method for generating a deep learning network structure described above.

The above is an exemplary version of a computer-readable storage medium of the present embodiment. It should be noted that, the technical solution of the storage medium and the technical solution of the generating method of the deep learning network structure belong to the same concept, and details of the technical solution of the storage medium which are not described in detail can be referred to the description of the technical solution of the generating method of the deep learning network structure.

An embodiment of the present disclosure further provides a computer program, where the computer program when executed in a computer causes the computer to perform the steps of the method for generating a deep learning network structure described above.

The above is an exemplary version of a computer program of the present embodiment. It should be noted that, the technical solution of the computer program and the technical solution of the method for generating a deep learning network structure belong to the same concept, and details of the technical solution of the computer program, which are not described in detail, can be referred to the description of the technical solution of the method for generating a deep learning network structure.

The foregoing describes specific embodiments of the present disclosure. Other embodiments are within the scope of the following claims. In some cases, the actions or steps recited in the claims can be performed in a different order than in the embodiments and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

The computer instructions include computer program code that may be in source code form, object code form, executable file or some intermediate form, etc. The computer readable medium may include: any entity or device capable of carrying the computer program code, a recording medium, a U disk, a removable hard disk, a magnetic disk, an optical disk, a computer Memory, a Read-Only Memory (ROM), a random access Memory (RAM, random Access Memory), an electrical carrier signal, a telecommunications signal, a software distribution medium, and so forth. It should be noted that the computer readable medium contains content that can be appropriately scaled according to the requirements of jurisdictions in which such content is subject to legislation and patent practice, such as in certain jurisdictions in which such content is subject to legislation and patent practice, the computer readable medium does not include electrical carrier signals and telecommunication signals.

It should be noted that, for simplicity of description, the foregoing method embodiments are all expressed as a series of combinations of actions, but it should be understood by those skilled in the art that the embodiments are not limited by the order of actions described, as some steps may be performed in other order or simultaneously according to the embodiments of the present disclosure. Further, those skilled in the art will appreciate that the embodiments described in the specification are all preferred embodiments, and that the acts and modules referred to are not necessarily all required for the embodiments described in the specification.

In the foregoing embodiments, the descriptions of the embodiments are emphasized, and for parts of one embodiment that are not described in detail, reference may be made to the related descriptions of other embodiments.

The preferred embodiments of the present specification disclosed above are merely used to help clarify the present specification. Alternative embodiments are not intended to be exhaustive or to limit the invention to the precise form disclosed. Obviously, many modifications and variations are possible in light of the teaching of the embodiments. The embodiments were chosen and described in order to best explain the principles of the embodiments and the practical application, to thereby enable others skilled in the art to best understand and utilize the invention. This specification is to be limited only by the claims and the full scope and equivalents thereof.

Claims

1. A method for generating a deep learning network structure includes:

acquiring service demand information corresponding to an initial deep learning network structure and an image processing service, wherein the service demand information comprises a deep learning network structure parameter threshold, and the deep learning network structure parameter threshold comprises a convolution network structure length-width ratio threshold;

acquiring the length-width ratio of the convolution structure of each amplified deep learning network, wherein the length-width ratio of the convolution structure of each amplified deep learning network is determined according to the number of layers of the convolution network, the size of convolution kernel and the number of convolution channels of each amplified deep learning network structure;

adding the amplified deep learning network structure with the aspect ratio of the convolution structure smaller than or equal to the aspect ratio threshold value of the convolution network structure to a deep learning network structure set to be screened;

determining a reference deep learning network structure in the deep learning network structure set to be screened, wherein determining the reference deep learning network structure in the deep learning network structure set to be screened comprises: determining a reference deep learning network structure from the deep learning network structure set to be screened according to the height, the width and the channel number of the feature image;

and generating a target deep learning network structure corresponding to the image processing service based on the reference deep learning network structure and a preset network generation condition.

2. The method of claim 1, generating a target deep learning network structure corresponding to the image processing service based on the reference deep learning network structure and a preset network generation condition, comprising:

and under the condition that the reference deep learning network structure meets the preset network generation condition, determining the reference deep learning network structure as a target deep learning network structure corresponding to the image processing service.

3. The method of claim 1, obtaining service requirement information corresponding to an image processing service, comprising:

receiving a budget instruction aiming at an image processing service, wherein the budget instruction carries a deep learning network structure parameter condition;

and analyzing the deep learning network structure parameter condition to obtain a deep learning network structure parameter threshold corresponding to the image processing service.

4. The method of claim 1, wherein the aspect ratio of the convolutions of the augmented deep learning network is obtained by:

acquiring the number of target convolution network layers, the size of a target convolution kernel and the number of target convolution channels of a target amplification deep learning network structure;

and determining the length-width ratio of the convolution structure of the target amplification deep learning network structure according to the length value of the convolution structure and the width value of the convolution structure.

5. The method of claim 1, after determining a reference deep learning network structure in the set of deep learning network structures to be screened, the method further comprising:

6. The method of claim 1, generating a plurality of augmented deep learning network structures based on the initial deep learning network structure, comprising:

7. The method of claim 1, the deep learning network structure parameter threshold further comprising a convolutional network parameter threshold and a convolutional network computational load threshold;

the method further comprises the steps of:

acquiring convolution parameters and convolution calculated quantity of each amplified deep learning network;

and adding the amplified deep learning network structure with the aspect ratio of the convolution structure smaller than or equal to the aspect ratio threshold of the convolution network structure, with the convolution parameter smaller than or equal to the convolution parameter threshold and with the convolution calculation amount smaller than or equal to the convolution calculation amount threshold to a deep learning network structure set to be screened.

8. The method of claim 7, determining a reference deep learning network structure from the set of deep learning network structures to be screened according to the height, width, and number of channels of the feature image, comprising:

scoring each deep learning network structure to be screened in the deep learning network structure set to be screened according to the height, the width and the channel number of the feature image;

9. A deep learning network architecture generation system, comprising:

The cloud side device is configured to obtain service requirement information corresponding to an initial deep learning network structure and an image processing service, where the service requirement information includes a deep learning network structure parameter threshold, where the deep learning network structure parameter threshold includes a convolutional network structure aspect ratio threshold, generate a plurality of amplified deep learning network structures based on the initial deep learning network structure, obtain a convolutional structure aspect ratio of each amplified deep learning network, determine the convolutional structure aspect ratio of each amplified deep learning network according to a convolutional network layer number, a convolutional kernel size, and a convolutional channel number of each amplified deep learning network, add the amplified deep learning network structure with the convolutional structure aspect ratio less than or equal to the convolutional network structure aspect ratio threshold to a deep learning network structure set to be screened, determine a reference deep learning network structure in the deep learning network structure set to be screened, generate a target deep learning network structure corresponding to the image processing service based on the reference deep learning network structure and a preset network generation condition, and train the target service deep learning network model for the image processing service based on the target deep learning network structure, where determining the reference deep learning network structure includes: determining a reference deep learning network structure from the deep learning network structure set to be screened according to the height, the width and the channel number of the feature image;

And the terminal side equipment is used for executing the image processing service according to the target service deep learning network model to obtain a processing result corresponding to the image processing service.

10. A computing device, comprising:

a memory and a processor;

the memory is configured to store computer executable instructions, the processor being configured to execute the computer executable instructions, which when executed by the processor, implement the steps of the method of any one of claims 1-8.

11. A computer readable storage medium storing computer executable instructions which when executed by a processor implement the steps of the method of any one of claims 1 to 8.