Methods and apparatus for generating deep learning network structures

Abstract

This specification provides a method and apparatus for generating deep learning network structures. The method for generating deep learning network structures includes: acquiring an initial deep learning network structure and business requirement information corresponding to a target business, wherein the business requirement information includes deep learning network structure parameter thresholds; generating multiple augmented deep learning network structures based on the initial deep learning network structure; determining a set of deep learning network structures to be screened from the multiple augmented deep learning network structures according to the business requirement information, and determining a reference deep learning network structure from the set of deep learning network structures to be screened; and generating a target deep learning network structure corresponding to the target business based on the reference deep learning network structure and preset network generation conditions.

Description

Technical Field

[0001] The embodiments in this specification relate to the field of deep learning technology, and in particular to a method for generating deep learning network structures. Background Technology

[0002] Deep learning plays a crucial role in visual intelligence, and its effectiveness is closely related to the neural network structure. Convolutional neural networks (CNNs) are a key neural network structure for handling visual intelligence scenarios. The quality of visual intelligence based on CNNs is heavily dependent on the neural network structure. Currently, popular CNN structures are primarily configured based on the personal experience of technical personnel combined with different business scenarios. This approach to network structure design is highly dependent on individual expertise, and errors are inevitable. This can lead to neural network structures that are either overly powerful or fail to meet the requirements of the target business scenario. Overly powerful structures increase hardware costs, while those that fail to meet the requirements result in a poor user experience. Therefore, a more accurate method for designing neural network structures for specific business scenarios is urgently needed. Summary of the Invention

[0003] In view of this, embodiments of this specification provide a method for generating deep learning network structures. One or more embodiments of this specification also relate to an apparatus for generating deep learning network structures, a computing device, a computer-readable storage medium, and a computer program, to address the technical deficiencies existing in the prior art.

[0004] According to a first aspect of the embodiments of this specification, a method for generating deep learning network structures is provided, comprising:

[0005] Obtain the initial deep learning network structure and the business requirement information corresponding to the target business, wherein the business requirement information includes the threshold values ​​of the deep learning network structure parameters;

[0006] Multiple augmented deep learning network structures are generated based on the initial deep learning network structure;

[0007] Based on the business requirements information, a set of deep learning network structures to be screened is determined from the multiple augmented deep learning network structures, and a reference deep learning network structure is determined from the set of deep learning network structures to be screened.

[0008] Based on the reference deep learning network structure and the preset network generation conditions, the target deep learning network structure corresponding to the target service is generated.

[0009] According to a second aspect of the embodiments of this specification, an apparatus for generating deep learning network structures is provided, comprising:

[0010] The acquisition module is configured to acquire the initial deep learning network structure and the business requirement information corresponding to the target business, wherein the business requirement information includes the threshold values ​​of the deep learning network structure parameters.

[0011] The augmentation module is configured to generate multiple augmented deep learning network structures based on the initial deep learning network structure.

[0012] The determination module is configured to determine a set of deep learning network structures to be screened from the multiple augmented deep learning network structures based on the business requirement information, and to determine a reference deep learning network structure from the set of deep learning network structures to be screened.

[0013] The generation module is configured to generate the target deep learning network structure corresponding to the target service based on the reference deep learning network structure and preset network generation conditions.

[0014] According to a third aspect of the embodiments of this specification, a method for generating an image processing network structure is provided, comprising:

[0015] Obtain the initial convolutional network structure and the business requirement information corresponding to the target image processing business. The business requirement information includes the aspect ratio threshold of the convolutional network structure, which is determined based on the number of convolutional network layers, the size of the convolutional kernel, and the number of convolutional channels.

[0016] Multiple augmented convolutional network structures are generated based on the initial convolutional network structure;

[0017] Based on the business requirements information, a set of convolutional network structures to be screened is determined from the multiple augmented convolutional network structures, and a reference convolutional network structure is determined from the set of deep learning network structures to be screened.

[0018] Based on the reference convolutional network structure and the preset network generation conditions, the target convolutional network structure corresponding to the target image processing service is generated.

[0019] According to a fourth aspect of the embodiments of this specification, a system for generating deep learning network structures is provided, comprising:

[0020] A cloud-based device is used to acquire an initial deep learning network structure and business requirement information corresponding to a target service. The business requirement information includes deep learning network structure parameter thresholds. Based on the initial deep learning network structure, multiple augmented deep learning network structures are generated. According to the business requirement information, a set of deep learning network structures to be screened is determined from the multiple augmented deep learning network structures, and a reference deep learning network structure is determined from the set of deep learning network structures to be screened. Based on the reference deep learning network structure and preset network generation conditions, a target deep learning network structure corresponding to the target service is generated, and a target service deep learning network model is trained for the target service based on the target deep learning network structure.

[0021] An edge device is used to execute the target service according to the target service deep learning network model and obtain the processing result corresponding to the target service.

[0022] According to a fifth aspect of the embodiments of this specification, a computing device is provided, comprising:

[0023] Memory and processor;

[0024] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the above-described method for generating deep learning network structures.

[0025] According to a sixth aspect of the embodiments of this specification, a computer-readable storage medium is provided that stores computer-executable instructions, which, when executed by a processor, implement the steps of the method for generating the deep learning network structure described above.

[0026] According to a seventh aspect of the embodiments of this specification, a computer program is provided, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the method for generating the deep learning network structure described above.

[0027] One embodiment of this specification implements the acquisition of an initial deep learning network structure and business requirement information corresponding to a target service, wherein the business requirement information includes deep learning network structure parameter thresholds; generating multiple augmented deep learning network structures based on the initial deep learning network structure; determining a set of deep learning network structures to be screened from the multiple augmented deep learning network structures according to the business requirement information, and determining a reference deep learning network structure from the set of deep learning network structures 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 preset network generation conditions.

[0028] The method provided in the embodiments of this specification first mutates and amplifies the initial deep learning network structure to obtain multiple amplified deep learning network structures, improving the data foundation for selecting a better network structure. Then, based on the business requirements information set by technical personnel for the target business (specifically, the threshold values ​​for deep learning network structure parameters in the business requirements information), the multiple amplified deep learning network structures are screened, eliminating network structures that do not meet the actual needs of the target business, reducing the amount of data required for subsequent selection. A reference deep learning network structure is determined from the set of deep learning network structures to be screened. Iterative updates are performed based on the reference deep learning network structure and preset network generation conditions, ultimately generating the target deep learning network structure corresponding to the target business. This ensures that the final generated target deep learning network structure meets the business requirements of the target business and is more closely aligned with its actual application. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of a method for generating a deep learning network structure used in one embodiment of this specification;

[0030] Figure 2 This is a flowchart illustrating a method for generating a deep learning network structure according to one embodiment of this specification;

[0031] Figure 3 This is a flowchart illustrating the processing steps of a method for generating deep learning network structures for image recognition scenarios, provided in one embodiment of this specification.

[0032] Figure 4 This is a flowchart of a method for generating an image processing network structure according to an embodiment of this specification;

[0033] Figure 5 This is a schematic diagram of a deep learning network structure generation system provided in one embodiment of this specification;

[0034] Figure 6 This is a schematic diagram of a deep learning network structure generation device provided in one embodiment of this specification;

[0035] Figure 7 This is a structural block diagram of a computing device provided in one embodiment of this specification. Detailed Implementation

[0036] Many specific details are set forth in the following description to provide a full understanding of this specification. However, this specification can be implemented in many other ways than those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this specification. Therefore, this specification is not limited to the specific implementations disclosed below.

[0037] The terminology used in one or more embodiments of this specification is for the purpose of describing particular embodiments only and is not intended to be limiting of the one or more embodiments of this specification. The singular forms “a,” “described,” and “the” as used in one or more embodiments of this specification and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in one or more embodiments of this specification refers to and includes any or all possible combinations of one or more associated listed items.

[0038] It should be understood that although the terms first, second, etc., may be used to describe various information in one or more embodiments of this specification, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, first may also be referred to as second without departing from the scope of one or more embodiments of this specification, and similarly, second may also be referred to as first. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to a determination."

[0039] First, the terms and concepts used in one or more embodiments of this specification will be explained.

[0040] CNN: Convolutional neural network.

[0041] Neural network architecture design: In the embodiments provided in this specification, it refers to the optimized design of a CNN neural network architecture based on deep learning.

[0042] Aspect ratio of convolutional structure: The ratio of the depth to the width of a deep learning network structure proposed in the embodiments provided in this specification.

[0043] With the rapid development of deep learning technology, more and more deep learning-based visual intelligence services are being widely used in many scenarios, bringing diverse experiences to users. The quality of deep learning-based visual intelligence is closely related to the neural network structure, and different business scenarios require different neural network structures. Therefore, there is an urgent need to provide a method for generating neural network structures that can be applied to various business scenarios and generate corresponding neural network structures for each business scenario.

[0044] Based on this, this specification provides a method for generating deep learning network structures. This specification also relates to a device for generating deep learning network structures, a computing device, and a computer-readable storage medium, which will be described in detail in the following embodiments.

[0045] See Figure 1 , Figure 1A schematic diagram of a method for generating deep learning network structures according to an embodiment of this specification is shown, such as... Figure 1 As shown, Figure 1 As shown in the embodiments of this specification, the method for generating deep learning network structures is applied to terminal 100, which may be a laptop, smart terminal, server, cloud server, or other terminal device.

[0046] In terminal 100, an initial deep learning network structure is first determined. Then, the hyperparameters in the initial deep learning network structure are adjusted to generate the adjusted deep learning network structure. The adjusted deep learning network structure and the initial deep learning network structure are used as multiple augmented deep learning network structures. The augmented deep learning network structures serve as the data basis for determining the target network structure in the future.

[0047] After obtaining multiple augmented deep learning network structures, these structures are filtered based on business requirement information corresponding to the target business (this information includes deep learning network structure parameter thresholds, which include aspect ratio thresholds). In practical applications, the business requirement information may include not only aspect ratio thresholds but also deep learning network parameter thresholds, computational cost thresholds, etc. During the design of deep learning network structures, it was found that the aspect ratio plays a crucial role in the performance of the deep learning network. Different aspect ratios of deep learning network structures in different business scenarios will result in different network performance. In the embodiments provided in this specification, the specified aspect ratio of the deep learning network structure is used as an indicator for filtering deep learning network structures, providing a new reference metric for generating deep learning network structures.

[0048] After filtering the augmented deep learning network structures using business requirements information, multiple deep learning network structures to be screened are obtained. Then, each deep learning network structure to be screened is scored using a network structure scoring method, and the network structure with the higher score is selected as the reference deep learning network structure.

[0049] The reference deep learning network structure is used as the new initial deep learning network structure, and the above-described process of expanding, filtering, scoring, and determining the new reference deep learning network structure is repeated iteratively until the reference deep learning network structure meets the preset network generation conditions; that is, the iteration process continues until the network structure no longer changes. At this point, a target deep learning network structure corresponding to the target business can be obtained. This target deep learning network structure can then be used to train the corresponding deep learning network model for the target business.

[0050] In practical applications, the preferred deep learning network is the convolutional neural network. Accordingly, the aspect ratio threshold of the deep learning network structure is the same as the aspect ratio threshold of the convolutional structure, the parameter threshold of the deep learning network is the same as the convolutional parameter threshold, and the computational cost threshold of the deep learning network is the same as the convolutional computational cost threshold.

[0051] See Figure 2 , Figure 2 A flowchart is shown of a method for generating a deep learning network structure according to an embodiment of this specification, which specifically includes the following steps.

[0052] Step 202: Obtain the initial deep learning network structure and the business requirement information corresponding to the target business, wherein the business requirement information includes the threshold values ​​of the deep learning network structure parameters.

[0053] The initial deep learning network structure specifically refers to the foundation upon which the network structure is used for subsequent expansion. For example, if subsequent expansion is based on deep learning network structure A, then deep learning network structure A is the initial deep learning network structure. In practical applications, the deep learning network structure can be any network structure, such as convolutional network structures, residual network structures, recurrent neural network structures, fully connected network structures, etc.

[0054] The target business specifically refers to different services in practical applications, such as image recognition, image segmentation, and object tracking. The business requirement information specifically refers to the resource budget that can be provided for each target business. This resource budget includes network parameter values, computational load values, resource values, and the aspect ratio threshold of the deep learning network structure, etc. In other words, the business requirement information specifies the processing upper limit of the deep learning network structure corresponding to the target business. For example, in a practical business scenario, the memory budget that can be provided for the deep learning network structure is 2GB, the computational load is 100, and the aspect ratio threshold of the deep learning network structure is 3. Therefore, when providing the deep learning network structure for this target business, it is necessary to ensure that the memory usage of the deep learning network structure is within 2GB, the computational load is within 100, and the aspect ratio of the deep learning network structure is less than or equal to 3, etc.

[0055] In the specific embodiments provided in this specification, the aspect ratio of the specified deep learning network structure, taking a convolutional network structure as an example, is determined based on the number of convolutional network layers, the kernel size, and the number of convolutional channels. Different numbers of convolutional network layers, kernel sizes, and the number of convolutional channels will correspond to different aspect ratios of the convolutional network structure. The aspect ratio of the convolutional network structure can be used for selection.

[0056] This includes obtaining the business requirement information corresponding to the target business, including:

[0057] Receive a budget instruction for a target business, wherein the budget instruction carries deep learning network structure parameter conditions;

[0058] The deep learning network structure parameter conditions are analyzed to obtain the deep learning network structure parameter thresholds corresponding to the target service.

[0059] In practical applications, business requirement information is set by technical personnel based on the actual situation of the target business. Furthermore, technical personnel will send budget instructions for the target business to the terminal. These budget instructions will carry pre-set deep learning network structure parameter conditions. Specifically, the deep learning network structure parameter conditions include the aspect ratio threshold of the deep learning network structure. In addition, they may also include deep learning network parameter thresholds, deep learning network computation thresholds, and so on.

[0060] After receiving a budget instruction, the terminal can obtain the deep learning network structure parameter thresholds corresponding to the target service by parsing the deep learning network structure parameter conditions in the budget instruction. Taking convolutional network structure as an example again, the deep learning network structure parameter conditions include the convolutional network structure aspect ratio threshold, the convolutional network parameter threshold, and the convolutional network computational cost threshold. The aspect ratio of the convolutional network structure is determined based on the number of convolutional network layers, the convolutional kernel size, and the number of convolutional channels.

[0061] In one specific embodiment provided in this specification, taking image recognition as an example, an initial deep learning network structure a is obtained, along with the business requirement information allocated by the technicians for the image recognition business. The aspect ratio threshold of the deep learning network structure specified in the business requirement information is 2.

[0062] Step 204: Generate multiple augmented deep learning network structures based on the initial deep learning network structure.

[0063] Among them, augmented deep learning network structures can be understood as deep learning network structures generated based on the initial deep learning network structure. In practical applications, multiple augmented deep learning network structures include the initial deep learning network structure.

[0064] Furthermore, in practical applications, the initial deep learning network structure includes multiple parameters. Adjusting different parameters will generate different new deep learning network structures. Each time the network structure's parameters are adjusted, the deep learning network structure's ability to process the target business will vary. The method provided in this manual requires selecting a better network structure from multiple augmented deep learning network structures for the subsequent model training process.

[0065] Among the parameters that determine the structure of a deep learning network, the hyperparameters are particularly important. Hyperparameters are parameters pre-set in the deep learning network structure, rather than parameters adjusted during model training. Hyperparameters define higher-level concepts within the model, and different hyperparameter values ​​can yield different network structures. Based on this, multiple augmented deep learning network structures are generated from the initial deep learning network structure, including:

[0066] Adjust the hyperparameters in the initial deep learning network structure, wherein the hyperparameters include at least one of the following: number of convolution channels, kernel size, and number of convolutional network layers;

[0067] Generate the corresponding augmented deep learning network structure after each adjustment.

[0068] Specifically, since different deep learning network structures can be generated by adjusting the hyperparameters of the deep learning network structure, in order to obtain a larger number of deep learning network structures and provide a data foundation for selecting better structures in the future, the method provided in this specification can adjust the hyperparameters in the initial deep learning network structure accordingly. Specifically, the hyperparameters can include the number of deep learning network channels, the number of deep learning network layers, etc.

[0069] Taking the initial deep learning network structure 'a' as an example of a convolutional neural network structure, convolutional network structure 'a' has 3 convolutional channels, a 1*1 kernel size, and 4 layers. Based on the initial deep learning network structure 'a', adjusting the number of convolutional channels to 5 yields the adjusted convolutional network structure 'a1', with 5 convolutional channels and the same hyperparameters as 'a'. Based on the initial deep learning network structure 'a', adjusting the kernel size to 3*3 yields the adjusted convolutional network structure 'a2', with the same kernel size and hyperparameters as 'a'. Based on the initial deep learning network structure 'a', adjusting the number of convolutional network layers to 5 yields the adjusted convolutional network structure 'a3', with 5 convolutional network layers and the same hyperparameters as 'a'. Thus, four augmented convolutional network structures can be obtained: a, a1, a2, and a3. In practical applications, at least two hyperparameters can be adjusted together. For example, the number of convolutional channels and the kernel size in the convolutional network structure can be adjusted simultaneously, the number of convolutional channels and the number of convolutional network layers can be adjusted simultaneously, or the number of convolutional channels, the kernel size and the number of convolutional network layers can be adjusted simultaneously, and so on.

[0070] It is important to note that, to ensure the randomness and breadth of augmented deep learning network structures, in practical applications, the hyperparameters in the initial deep learning network structure can be randomly adjusted. This allows for the generation of a large number of augmented deep learning network structures, providing a substantial data foundation for subsequent selection.

[0071] Step 206: Based on the business requirement information, determine the set of deep learning network structures to be screened from the multiple augmented deep learning network structures, and determine the reference deep learning network structure from the set of deep learning network structures to be screened.

[0072] After determining the business requirements and multiple augmented deep learning network structures, the augmented deep learning network structures that meet the business requirements can be selected as the set of deep learning network structures to be selected. The set of deep learning network structures to be selected can be regarded as the set of deep learning network structures that satisfy the business requirements.

[0073] Specifically, based on the business requirement information, a set of deep learning network structures to be screened is determined from the multiple augmented deep learning network structures, including:

[0074] Obtain the target network structure parameters for the target augmented deep learning network structure;

[0075] If the target network structure parameters meet the business requirements, the target augmented deep learning network structure is added to the set of deep learning network structures to be screened.

[0076] In the embodiments provided in this specification, a convolutional neural network is still used as an example. The target network structure parameters for obtaining the target augmented deep learning network structure include:

[0077] Obtain the target convolutional network layer number, target convolutional kernel size, target convolutional channel number, target convolutional parameters, and target convolutional computation cost of the target augmented deep learning network structure.

[0078] In the above steps, it has been learned that the business requirements information contains the conditions for selecting deep learning network structures. Accordingly, it is necessary to obtain the network structure parameters in the augmented deep learning network structure based on the conditions required by the business requirements information. For example, if the business requirements information requires parameters such as the number of convolutional network layers, convolutional kernel size, number of convolutional channels, convolutional parameters, and convolutional computation cost, then it is necessary to obtain these parameter information in the augmented deep learning network structure.

[0079] Specifically, the target augmented deep learning network structure can be understood as a network structure selected from multiple augmented deep learning network structures, which is a network structure that is compared with the business requirement information.

[0080] After determining the target augmented deep learning network structure, the target network structure parameters can be obtained. Furthermore, the target number of convolutional network layers, kernel size, channel number, parameters, and computational cost of the target convolutional network structure can be obtained. Based on these target network structure parameters, it is determined whether the target augmented deep learning network structure meets the business requirements. If it does, the target augmented deep learning network structure is added to the set of deep learning network structures to be screened; if it does not, it is discarded.

[0081] Furthermore, after obtaining the target network structure parameters of the target augmented deep learning network structure, the method further includes:

[0082] The aspect ratio of the target convolutional structure of the target augmented deep learning network structure is determined based on the number of layers of the target convolutional network, the size of the target convolutional kernel, and the number of target convolutional channels.

[0083] If the aspect ratio of the target convolutional structure is less than or equal to the aspect ratio threshold of the convolutional structure, the target convolutional parameters are less than or equal to the convolutional parameter threshold, and the target convolutional computation is less than or equal to the convolutional computation threshold, then the target network structure parameters are determined to meet the business requirement information.

[0084] As determined in the above steps, the business requirement information contains a threshold for the aspect ratio of the convolutional structure. Based on this threshold, the target augmented deep learning network structure can be screened. The aspect ratio of the convolutional structure is determined based on the number of convolutional network layers, the size of the convolutional kernel, and the number of convolutional channels. Therefore, the target aspect ratio of the target augmented deep learning network structure can be determined from the obtained target network structure parameters based on the target number of convolutional network layers, the target size of the convolutional kernel, and the target number of convolutional channels.

[0085] Specifically, the aspect ratio of the target convolutional structure of the target augmented deep learning network structure is determined based on the number of layers in the target convolutional network, the size of the target convolutional kernel, and the number of target convolutional channels, including:

[0086] The convolutional structure length value of the target augmented deep learning network structure is determined based on the number of layers in the target convolutional network.

[0087] The convolutional structure width value of the target augmented deep learning network structure is determined based on the target convolutional kernel size and the target number of convolutional channels;

[0088] The aspect ratio of the target convolutional structure is determined based on the length and width values ​​of the convolutional structure.

[0089] In the embodiments provided in this specification, taking a convolutional network structure as an example, the convolutional network structure is abstracted into a structure with length and width, and the convolutional structure length value of the target augmented convolutional network structure is determined according to the number of layers of the target convolutional network; the convolutional structure width value of the target augmented convolutional network structure is determined according to the target convolutional kernel size and the target number of convolutional channels.

[0090] The calculation method for the length of the convolution structure is shown in Formula 1 below:

[0091]

[0092] Among them, L j This represents the depth of the j-th stage module, and n represents the number of stages in the current network structure.

[0093] The calculation method for the width of the convolution structure is shown in Formula 2 below:

[0094]

[0095] Among them, C i This represents the number of convolutional channels in the i-th convolutional network layer. This represents the kernel size of the i-th convolutional network layer.

[0096] After determining the length and width of the convolutional structure, the aspect ratio of the convolutional structure can be determined based on the ratio between the two. The aspect ratio of the convolutional structure is then compared with the aspect ratio threshold in the business requirements information. If the aspect ratio of the convolutional structure is less than or equal to the aspect ratio threshold, it means that the target augmented convolutional network structure meets the requirement of the aspect ratio threshold in the business requirements information.

[0097] Specifically, the comparison between the target network structure parameters of the target augmented convolutional network structure and the business requirement information can be found in Equations 3-1 to 3-3 below:

[0098]

[0099] size of params≤budget params Formula 3-2

[0100] size of FLOPs ≤ budget FLOPs Formula 3-3

[0101] Formulas 3-1 to 3-3 are parallel; the target network structure parameters of the target augmented convolutional network structure must simultaneously satisfy formulas 3-1 to 3-3. Specifically, L j Let C represent the depth of the j-th stage module, n represent the current network stage number, N represent the total number of network stages, and C represent the total number of stages in the network structure. i This represents the number of convolutional channels in the i-th convolutional network layer. ρ represents the kernel size of the i-th convolutional network layer. h The aspect ratio threshold of the convolutional structure is represented by `size of params`, which is the number of parameters in the convolution operation of the target augmented convolutional network structure, and `budget` is the budget. params The parameter threshold for convolution operations in the business requirements information; size of FLOPs is the computational cost of convolution operations in the target augmented convolutional network structure; budget. FLOPs As shown in Formula 3, when the aspect ratio of the target convolutional structure is less than or equal to the aspect ratio threshold, the target convolutional parameters are less than or equal to the convolutional parameter threshold, and the target convolutional computation is less than or equal to the convolutional computation threshold, it can be determined that the target network structure parameters meet the business requirement information, and the target augmented convolutional network structure is added to the set of convolutional network structures to be screened.

[0102] In one specific embodiment provided in this specification, taking the augmented convolutional network structure (a, a1, a2, a3, a4...a20) as an example, after comparing each structure with the business requirement information, the convolutional network structures that meet the business requirement information are added to the set of convolutional network structures to be screened. Based on this, the set of convolutional network structures to be screened (a, a1, a3, a4, a7, a9, a13, a15, a17, a20) can be obtained. All convolutional network structures in the set of convolutional network structures to be screened are convolutional network structures that meet the business requirement information.

[0103] After obtaining the set of deep learning network structures to be screened, it is also necessary to determine reference deep learning network structures from the set of deep learning network structures to be screened, specifically including:

[0104] Each deep learning network architecture in the set of architectures to be screened is scored;

[0105] The deep learning network architecture with the highest score among those to be screened will be used as the reference deep learning network architecture.

[0106] Since the set of network structures to be screened consists of deep learning network structures that meet the business requirements, it is necessary to select the best-performing deep learning network structures for iterative screening. Specifically, each deep learning network structure to be screened needs to be scored. In the implementation provided in this specification, the scoring is preferably based on the maximum entropy method, as shown in Formula 4 below:

[0107]

[0108] in, This represents the height of the output feature image at the nth stage. This represents the width of the output feature image in the nth stage. The numbers α1, α2, and ω represent the number of channels in the output feature image of the nth stage. n E represents the network structure hyperparameters. n This is the score for the nth stage. It is the weighted average score of N stages, which is the total score of the deep learning network structure to be screened.

[0109] After calculating the score for each deep learning network structure to be screened, the deep learning network structure with the highest score is selected as the reference deep learning network structure. The reference deep learning network structure specifically refers to the deep learning network structure used for subsequent processing.

[0110] In the implementation provided in this specification, following the example above, the set of deep learning network structures to be screened is (a, a1, a3, a4, a7, a9, a13, a15, a17, a20). After scoring each network structure to be screened, they are sorted in descending order of score, and the resulting set of network structures is (a17, a13, a9, a15, a1, a4, a3, a7, a20, a). That is, a17 has the highest score, so a17 is the reference deep learning network structure.

[0111] Step 208: Based on the reference deep learning network structure and the preset network generation conditions, generate the target deep learning network structure corresponding to the target service.

[0112] After determining the reference deep learning network structure, it can be compared with the preset network generation conditions. If the reference deep learning network structure meets the preset network generation conditions, it can be selected as the final target deep learning network structure. Otherwise, iterative optimization is performed based on the reference deep learning network structure.

[0113] In the specific embodiments provided in this specification, after obtaining the reference deep learning network structure, it is also necessary to determine whether the reference deep learning network structure meets the preset network generation conditions. Specifically, after determining the reference deep learning network structure from the set of deep learning network structures to be screened, the method further includes:

[0114] Count the number of consecutive iterations that hit the reference deep learning network structure;

[0115] If the number of consecutive iterations hits reaches a preset threshold, the reference deep learning network structure is determined to meet the preset network generation conditions.

[0116] Specifically, the consecutive iteration hit count refers to whether the reference deep learning network structure has been selected as the reference deep learning network structure in multiple iterations of optimization. For example, if the initial deep learning network structure is 'a', after the first iteration optimization, the deep learning network structure 'b' is determined. Based on 'b', after the second iteration optimization, the deep learning network structure 'c' is determined. Based on 'c', after the third iteration optimization, the deep learning network structure 'd' is determined. Based on 'd', after the fourth iteration optimization, the deep learning network structure 'd' is determined again. Then, based on 'd', after the fifth iteration optimization, the deep learning network structure 'd' is still determined. At this point, the consecutive iteration hit count of the deep learning network structure 'd' is three.

[0117] If the preset threshold is 3 times, it can be determined that the deep learning network structure d satisfies the preset network generation conditions, while the previous deep learning network structures b and c did not satisfy the preset network generation conditions.

[0118] Furthermore, based on the reference deep learning network structure and preset network generation conditions, a target deep learning network structure corresponding to the target service is generated, including:

[0119] S2082: If the reference deep learning network structure does not meet the preset network generation conditions, at least one deep learning network structure to be processed is selected from the set of deep learning network structures to be screened, and each deep learning network structure to be processed is used as the initial deep learning network structure, and the operation of generating multiple augmented deep learning network structures continues.

[0120] If the reference deep learning network structure does not meet the preset network generation conditions, iterative optimization is required again. Specifically, at least one deep learning network structure to be processed is selected from the set of deep learning network structures to be screened. For example, the top 20 deep learning network structures to be screened after ranking 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 the initial deep learning network structure, and the following operations are performed: generating multiple augmented deep learning network structures, determining the set of deep learning network structures to be screened from the multiple augmented deep learning network structures according to business requirements, determining the reference deep learning network structure from the set of deep learning network structures to be screened, and judging whether the reference convolutional network set meets the preset network generation conditions.

[0121] In a specific implementation provided in this specification, following the previous example, the initial deep learning network structure is 'a'. After the first iteration of optimization, a deep learning network structure 'b' is determined. At this point, deep learning network structure 'b' does not meet the preset network generation conditions. Therefore, deep learning network structures 'b', 'b1', 'b2', 'b3', and 'b4' are selected as the deep learning network structures to be processed, where b1, b2, b3, and b4 are the top 5 deep learning network structures in terms of score. Each of these unprocessed deep learning network structures is used as the initial deep learning network structure, and the hyperparameters in deep learning network structures b, b1, b2, b3, and b4 are randomly changed to generate multiple augmented deep learning network structures corresponding to b, b1, b2, b3, and b4. Then, based on business requirements, a new set of deep learning network structures to be screened is determined from these augmented deep learning network structures, and a reference deep learning network structure 'c' is selected from this new set. Finally, it is determined whether 'c' meets the preset network generation conditions.

[0122] S2084: If the reference deep learning network structure meets the preset network generation conditions, determine that the reference deep learning network structure is the target deep learning network structure corresponding to the target service.

[0123] If the reference deep learning network structure meets the preset network generation conditions, that is, the reference deep learning network structure tends to be stable after multiple iterations of optimization, then the reference deep learning network structure can be determined to be a relatively stable deep learning network structure that meets the preset network generation conditions. Therefore, the reference deep learning network structure can be used as the target deep learning network structure corresponding to the target business.

[0124] In a specific implementation provided in this specification, following the example above, the deep learning network structure d is used in three consecutive iterations of optimization, indicating that the deep learning network structure d has stabilized and can be used as the target deep learning network structure corresponding to the target business.

[0125] After determining the target deep learning network structure, a corresponding target service deep learning network model can be trained based on this target deep learning network structure for the target service. In the embodiments provided in this specification, the method further includes:

[0126] Based on the target deep learning network structure, a target business deep learning network model is trained for the target business.

[0127] Specifically, the training data and training labels corresponding to the target business are obtained, the training data is input into a deep learning network model based on the target deep learning network structure for processing, the prediction results output by the deep learning network model are obtained, the model loss value is calculated based on the prediction results and training labels, and the model parameters of the deep learning network model are adjusted based on the model loss value in the deep learning network model until the model training stopping condition is reached.

[0128] The method provided in the embodiments of this specification first mutates and amplifies the initial deep learning network structure to obtain multiple amplified deep learning network structures, improving the data foundation for selecting a better network structure. Then, based on the business requirements information set by technical personnel for the target business (specifically, the threshold values ​​for deep learning network structure parameters in the business requirements information), the multiple amplified deep learning network structures are screened, eliminating network structures that do not meet the actual needs of the target business, reducing the amount of data required for subsequent selection. A reference deep learning network structure is determined from the set of deep learning network structures to be screened. Iterative updates are performed based on the reference deep learning network structure and preset network generation conditions, ultimately generating the target deep learning network structure corresponding to the target business. This ensures that the final generated target deep learning network structure meets the business requirements of the target business and is more closely aligned with its actual application.

[0129] In addition, the embodiments in this specification define the length and width of the convolutional network structure and use the aspect ratio of the deep learning network structure as a constraint to select a network structure that meets the expected aspect ratio. Compared with manually designed networks, this method has a significant performance improvement and can select a better network structure based on the business requirements information corresponding to the target business.

[0130] The following is in conjunction with the appendix Figure 3 Taking the application of the deep learning network structure generation method provided in this specification in an image recognition scenario as an example, the method for generating the deep learning network structure will be further explained. Figure 3 The flowchart of a method for generating a deep learning network structure according to an embodiment of this specification is shown. The deep learning network structure used in image recognition scenarios is a convolutional network structure, and the method specifically includes the following steps.

[0131] Step 302: Obtain the initial convolutional network structure and the business requirements information corresponding to the image recognition service.

[0132] The business requirements information includes thresholds for the aspect ratio of the convolutional structure, thresholds for convolutional parameters, and thresholds for the amount of convolutional computation.

[0133] Step 304: Adjust the hyperparameters in the initial convolutional network structure to obtain the corresponding augmented convolutional network structure after each adjustment.

[0134] The hyperparameters include at least one of the following: number of convolution channels, kernel size, and number of convolutional network layers.

[0135] Step 306: Determine the target convolutional network structure, including the number of target convolutional network layers, the target convolutional kernel size, the target convolutional channel number, the target convolutional parameters, and the target convolutional computation cost.

[0136] Step 308: Determine the convolutional structure length value based on the target number of convolutional network layers, and determine the convolutional structure width value based on the target convolutional kernel size and the target number of convolutional channels.

[0137] Step 310: Determine the aspect ratio of the target convolutional structure based on the length and width values ​​of the convolutional structure.

[0138] Step 312: If the aspect ratio of the target convolutional structure is less than or equal to the aspect ratio threshold, the target convolutional parameters are less than or equal to the convolutional parameter threshold, and the target convolutional computation is less than or equal to the convolutional computation threshold, add the target augmented convolutional network structure to the set of convolutional network structures to be screened.

[0139] Step 314: Score each convolutional network structure in the set of structures to be screened.

[0140] Step 316: Select the highest-scoring convolutional network structure to be screened as the reference convolutional network structure.

[0141] Step 318: Count the number of consecutive iterations that the reference convolutional network structure hits.

[0142] Step 320: Determine whether the number of consecutive iterations hit is greater than or equal to a preset threshold. If yes, proceed to step 322. If no, select at least one initial convolutional network structure from the set of convolutional network structures to be screened, and continue to proceed to step 304.

[0143] Step 322: Determine the reference convolutional network structure as the target convolutional network structure corresponding to the image recognition service.

[0144] Step 324: Obtain the image training data corresponding to the image recognition service, and train an image recognition model based on the target convolutional network structure as the backbone network according to the image training data.

[0145] The method provided in the embodiments of this specification first performs mutation amplification on the initial convolutional network structure to obtain multiple amplified convolutional network structures, improving the data foundation for selecting a better network structure. Then, based on the business requirements information set by the technical personnel for the target business (specifically, the aspect ratio threshold of the convolutional structure in the business requirements information), the multiple amplified convolutional network structures are screened, eliminating network structures that do not meet the actual needs of the target business, reducing the amount of data during subsequent selection. A reference convolutional network structure is determined from the set of convolutional network structures to be screened, and it is determined whether the reference convolutional network structure meets the preset network generation conditions. If it does, the reference convolutional network structure is used as the target convolutional network structure. If it does not meet the conditions, at least one initial convolutional network structure is selected from the set of convolutional network structures to be screened, and the above operation is repeated.

[0146] See Figure 4 , Figure 4 A flowchart of a method for generating an image processing network structure according to an embodiment of this specification is shown, specifically including the following steps:

[0147] Step 402: Obtain the initial convolutional network structure and the business requirement information corresponding to the target image processing business. The business requirement information includes the aspect ratio threshold of the convolutional network structure, which is determined based on the number of convolutional network layers, the size of the convolutional kernel, and the number of convolutional channels.

[0148] Step 404: Generate multiple augmented convolutional network structures based on the initial convolutional network structure.

[0149] Step 406: Based on the business requirement information, determine the set of convolutional network structures to be screened from the multiple augmented convolutional network structures, and determine the reference convolutional network structure from the set of deep learning network structures to be screened.

[0150] Step 408: Based on the reference convolutional network structure and the preset network generation conditions, generate the target convolutional network structure corresponding to the target image processing service.

[0151] Specifically, based on the reference convolutional network structure and preset network generation conditions, a target convolutional network structure corresponding to the target image processing service is generated, including:

[0152] If the reference convolutional network structure does not meet the preset network generation conditions, at least one convolutional network structure to be processed is selected from the set of convolutional network structures to be screened, and each convolutional network structure to be processed is used as the initial convolutional network structure, and the operation of generating multiple expanded convolutional network structures continues.

[0153] If the reference convolutional network structure meets the preset network generation conditions, the reference convolutional network structure is determined to be the target convolutional network structure corresponding to the target image processing service.

[0154] See Figure 5 , Figure 5 This diagram illustrates a deep learning network structure generation system according to an embodiment of this specification. The system may include a cloud-side device 501 and an edge-side device 502. The cloud-side device 501 is used to select a target deep learning network structure corresponding to a target service and train a target service deep learning network model based on the target deep learning network structure. The edge-side device 502 is used to execute the target service based on the trained target service deep learning network model. Specifically, the deep learning network structure may be a convolutional network structure, a residual network structure, a recurrent neural network structure, a fully connected network structure, etc. 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, etc. In the embodiments provided in this specification, the form of the deep learning network model and the specific content of the target service are not limited.

[0155] Cloud-side device 501 is used to acquire an initial deep learning network structure and business requirement information corresponding to a target service. The business requirement information includes deep learning network structure parameter thresholds. Based on the initial deep learning network structure, multiple augmented deep learning network structures are generated. According to the business requirement information, a set of deep learning network structures to be screened is determined from the multiple augmented deep learning network structures. A reference deep learning network structure is determined from the set of deep learning network structures to be screened. Based on the reference deep learning network structure and preset network generation conditions, a target deep learning network structure corresponding to the target service is generated. Based on the target deep learning network structure, a target service deep learning network model is trained for the target service.

[0156] The edge device 502 is used to execute the target service according to the target service deep learning network model and obtain the processing result corresponding to the target service.

[0157] Cloud-side device 501 can be a central cloud device in a distributed cloud architecture, and edge-side device 502 can be an edge cloud device in a distributed cloud architecture. Both cloud-side device 501 and edge-side device 502 can be server-side devices such as conventional servers, cloud servers, or server arrays, or they can be terminal devices; this specification does not limit the specific type of device used in this embodiment. Furthermore, cloud-side device 501 provides superior computing and storage capabilities and is located relatively far from the user; while edge-side device 502 has a larger deployment range and is closer to the user. Edge-side device 502 is an extension of cloud-side device 501, allowing the computing power of cloud-side device 501 to be transferred to edge-side device 502. Through integrated and collaborative management of the cloud and edge, it addresses business needs that cannot be met under a centralized cloud computing model.

[0158] Corresponding to the above method embodiments, this specification also provides embodiments of a deep learning network structure generation apparatus. Figure 6 A schematic diagram of a deep learning network structure generation apparatus according to an embodiment of this specification is shown. Figure 6 As shown, the device includes:

[0159] The acquisition module 602 is configured to acquire the initial deep learning network structure and the business requirement information corresponding to the target business, wherein the business requirement information includes the threshold values ​​of the deep learning network structure parameters.

[0160] The augmentation module 604 is configured to generate multiple augmented deep learning network structures based on the initial deep learning network structure.

[0161] The determining module 606 is configured to determine a set of deep learning network structures to be screened from the plurality of augmented deep learning network structures according to the business requirement information, and to determine a reference deep learning network structure from the set of deep learning network structures to be screened.

[0162] The generation module 608 is configured to generate the target deep learning network structure corresponding to the target service based on the reference deep learning network structure and preset network generation conditions.

[0163] Optionally, the generation module 608 is further configured to:

[0164] If the reference deep learning network structure does not meet the preset network generation conditions, at least one deep learning network structure to be processed is selected from the set of deep learning network structures to be screened, and each deep learning network structure to be processed is used as the initial deep learning network structure, and the operation of generating multiple augmented deep learning network structures continues.

[0165] If the reference deep learning network structure meets the preset network generation conditions, the reference deep learning network structure is determined to be the target deep learning network structure corresponding to the target service.

[0166] Optionally, the acquisition module 602 is further configured to:

[0167] Receive a budget instruction for a target business, wherein the budget instruction carries deep learning network structure parameter conditions;

[0168] The deep learning network structure parameter conditions are analyzed to obtain the deep learning network structure parameter thresholds corresponding to the target service.

[0169] Optionally, the deep learning network structure parameter conditions include a convolutional network structure aspect ratio threshold, a convolutional network parameter threshold, and a convolutional network computational cost threshold, wherein the convolutional network structure aspect ratio is determined based on the number of convolutional network layers, the size of the convolutional kernel, and the number of convolutional channels.

[0170] Optionally, the determining module 606 is further configured to:

[0171] Obtain the target network structure parameters for the target augmented deep learning network structure;

[0172] If the target network structure parameters meet the business requirements, the target augmented deep learning network structure is added to the set of deep learning network structures to be screened.

[0173] Optionally, the determining module 606 is further configured to:

[0174] Obtain the target convolutional network layer number, target convolutional kernel size, target convolutional channel number, target convolutional parameters, and target convolutional computation cost of the target augmented deep learning network structure.

[0175] Optionally, the deep learning network structure parameter thresholds include the aspect ratio threshold of the convolutional network structure, the convolutional network parameter threshold, and the convolutional network computational cost threshold;

[0176] The determining module 606 is further configured to:

[0177] The aspect ratio of the target convolutional structure of the target augmented deep learning network structure is determined based on the number of layers of the target convolutional network, the size of the target convolutional kernel, and the number of target convolutional channels.

[0178] If the aspect ratio of the target convolutional structure is less than or equal to the aspect ratio threshold of the convolutional structure, the target convolutional parameters are less than or equal to the convolutional parameter threshold, and the target convolutional computation is less than or equal to the convolutional computation threshold, then the target network structure parameters are determined to meet the business requirement information.

[0179] Optionally, the determining module 606 is further configured to:

[0180] The convolutional structure length value of the target augmented deep learning network structure is determined based on the number of layers in the target convolutional network.

[0181] The convolutional structure width value of the target augmented deep learning network structure is determined based on the target convolutional kernel size and the target number of convolutional channels;

[0182] The aspect ratio of the target convolutional structure is determined based on the length and width values ​​of the convolutional structure.

[0183] Optionally, the determining module 606 is further configured to:

[0184] Each deep learning network architecture in the set of architectures to be screened is scored;

[0185] The deep learning network architecture with the highest score among those to be screened will be used as the reference deep learning network architecture.

[0186] Optionally, the device further includes:

[0187] The statistics module is configured to count the number of consecutive iterations of the reference deep learning network structure.

[0188] Accordingly, the determining module 606 is further configured to:

[0189] If the number of consecutive iterations hits reaches a preset threshold, the reference deep learning network structure is determined to meet the preset network generation conditions.

[0190] Optionally, the amplification module 604 is further configured to:

[0191] Adjust the hyperparameters in the initial deep learning network structure, wherein the hyperparameters include at least one of the following: number of convolution channels, kernel size, and number of convolutional network layers;

[0192] Generate the corresponding augmented deep learning network structure after each adjustment.

[0193] Optionally, the device further includes:

[0194] 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.

[0195] The apparatus provided in the embodiments of this specification first performs mutation amplification on the initial deep learning network structure to obtain multiple amplified deep learning network structures, improving the data foundation for selecting better network structures. Then, based on the business requirements information set by technicians for the target business (specifically, the aspect ratio threshold of the convolutional structure in the business requirements information), the multiple amplified deep learning network structures are screened, eliminating network structures that do not meet the actual needs of the target business, reducing the amount of data required for subsequent selection. A reference deep learning network structure is determined from the set of deep learning network structures to be screened, and it is determined whether the reference deep learning network structure meets the preset network generation conditions. If it does, the reference deep learning network structure is used as the target deep learning network structure; if it does not, the reference deep learning network structure is used as a new initial deep learning network structure, and the above operation is repeated.

[0196] This specification defines the length and width of the convolutional network structure in the embodiments, and uses the aspect ratio of the deep learning network structure as a constraint to select a network structure that meets the expected aspect ratio. Compared with manually designed networks, this method has a significant performance improvement and can select a better network structure based on the business requirements information corresponding to the target business.

[0197] The above is a schematic scheme of a deep learning network structure generation device according to this embodiment. It should be noted that the technical solution of this deep learning network structure generation device and the technical solution of the deep learning network structure generation method described above belong to the same concept. For details not described in detail in the technical solution of the deep learning network structure generation device, please refer to the description of the technical solution of the deep learning network structure generation method described above.

[0198] Figure 7 A structural block diagram of a computing device 700 according to one embodiment of this specification is shown. The components of the computing device 700 include, but are not limited to, a memory 710 and a processor 720. The processor 720 is connected to the memory 710 via a bus 730, and a database 750 is used to store data.

[0199] The computing device 700 also includes an access device 740, which enables the computing device 700 to communicate via one or more networks 760. Examples of these networks include Public Switched Telephone Network (PSTN), Local Area Network (LAN), Wide Area Network (WAN), Personal Area Network (PAN), or combinations of communication networks such as the Internet. The access device 740 may include one or more of any type of wired or wireless network interface (e.g., a network interface card (NIC)), such as an IEEE 802.11 Wireless Local Area Network (WLAN) wireless interface, a Wi-MAX (Worldwide Interoperability for Microwave Access) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth interface, or a Near Field Communication (NFC) interface.

[0200] In one embodiment of this specification, the above-described components of the computing device 700 and Figure 7 Other components, not shown, can also be connected to each other, for example, via a bus. It should be understood that... Figure 7 The block diagram of the computing device shown is for illustrative purposes only and is not intended to limit the scope of this specification. Those skilled in the art can add or replace other components as needed.

[0201] The computing device 700 can be any type of stationary or mobile computing device, including mobile computers or mobile computing devices (e.g., tablet computers, personal digital assistants, laptop computers, notebook computers, netbooks, etc.), mobile phones (e.g., smartphones), wearable computing devices (e.g., smartwatches, smart glasses, etc.) or other types of mobile devices, or stationary computing devices such as desktop computers or personal computers (PCs). The computing device 700 can also be a mobile or stationary server.

[0202] The processor 720 is configured to execute the following computer-executable instructions, which, when executed by the processor, implement the steps of the aforementioned data processing method. The above is an illustrative scheme of a computing device according to this embodiment. It should be noted that the technical solution of this computing device and the technical solution of the aforementioned deep learning network structure generation method belong to the same concept. Details not described in detail in the technical solution of the computing device can be found in the description of the technical solution of the aforementioned deep learning network structure generation method.

[0203] An embodiment of this specification 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 the deep learning network structure described above.

[0204] The above is an illustrative scheme of a computer-readable storage medium according to this embodiment. It should be noted that the technical solution of this storage medium belongs to the same concept as the technical solution of the deep learning network structure generation method described above. Details not described in detail in the technical solution of the storage medium can be found in the description of the technical solution of the deep learning network structure generation method described above.

[0205] An embodiment of this specification also provides a computer program, wherein when the computer program is executed in a computer, it causes the computer to perform the steps of the above-described method for generating deep learning network structures.

[0206] The above is an illustrative example of a computer program according to this embodiment. It should be noted that the technical solution of this computer program belongs to the same concept as the technical solution of the aforementioned method for generating deep learning network structures. Details not described in detail in the computer program's technical solution can be found in the description of the technical solution of the aforementioned method for generating deep learning network structures.

[0207] The foregoing description of this specification describes specific embodiments. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recited in the claims can be performed in an order different from that described in the embodiments and still achieve the desired results. Furthermore, the processes depicted in the accompanying drawings do not necessarily require the specific order shown or the sequential order to achieve the desired results. In certain embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0208] The computer instructions include computer program code, which may be in the form of source code, object code, executable file, or some intermediate form. The computer-readable medium may include: any entity or device capable of carrying the computer program code, recording media, USB flash drive, portable hard drive, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium may be appropriately added to or subtracted according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media may not include electrical carrier signals and telecommunication signals.

[0209] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the embodiments in this specification are not limited to the described order of actions, because according to the embodiments in this specification, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in this specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the embodiments in this specification.

[0210] In the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0211] The preferred embodiments disclosed above are merely illustrative of this specification. The optional embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the embodiments described herein. These embodiments are selected and specifically described in this specification to better explain the principles and practical applications of the embodiments, thereby enabling those skilled in the art to better understand and utilize this specification. This specification is limited only by the claims and their full scope and equivalents.

Claims

1. A method for generating deep learning network structures, comprising: Obtain the initial deep learning network structure and the corresponding business requirements information for image processing services, wherein the business requirements information includes deep learning network structure parameter thresholds, and the deep learning network structure parameter thresholds include the aspect ratio thresholds of the convolutional network structure. Multiple augmented deep learning network structures are generated based on the initial deep learning network structure; Obtain the aspect ratio of the convolutional structure of each augmented deep learning network. The aspect ratio of the convolutional structure of each augmented deep learning network is determined based on the number of convolutional network layers, the size of the convolutional kernel, and the number of convolutional channels of each augmented deep learning network structure. Add augmented deep learning network structures whose aspect ratio is less than or equal to the aspect ratio threshold of the convolutional network structure to the set of deep learning network structures to be screened. Determining a reference deep learning network structure from the set of deep learning network structures to be screened, wherein determining the reference deep learning network structure from the set of deep learning network structures to be screened includes: determining the reference deep learning network structure from the set of deep learning network structures to be screened based on the height, width, and number of channels of the feature image; Based on the reference deep learning network structure and the preset network generation conditions, a target deep learning network structure corresponding to the image processing service is generated, wherein the preset network generation conditions are that the number of consecutive iterations of the reference deep learning network structure hits reaches a preset threshold.

2. The method as described in claim 1, wherein generating the target deep learning network structure corresponding to the image processing service based on the reference deep learning network structure and preset network generation conditions includes: If the reference deep learning network structure does not meet the preset network generation conditions, at least one deep learning network structure to be processed is selected from the set of deep learning network structures to be screened, and each deep learning network structure to be processed is used as the initial deep learning network structure, and the operation of generating multiple augmented deep learning network structures continues. If the reference deep learning network structure meets the preset network generation conditions, the reference deep learning network structure is determined to be the target deep learning network structure corresponding to the image processing service.

3. The method as described in claim 1, wherein obtaining business requirement information corresponding to image processing services includes: Receive a budget instruction for image processing services, wherein the budget instruction carries deep learning network structure parameter conditions; The deep learning network structure parameter conditions are analyzed to obtain the deep learning network structure parameter thresholds corresponding to the image processing service.

4. The method as described in claim 1, wherein the aspect ratio of the augmented deep learning network's convolutional structure is obtained through the following steps: Obtain the target number of convolutional network layers, target convolutional kernel size, and target number of convolutional channels for the target augmented deep learning network structure; The convolutional structure length value of the target augmented deep learning network structure is determined based on the number of layers in the target convolutional network. The convolutional structure width value of the target augmented deep learning network structure is determined based on the target convolutional kernel size and the target number of convolutional channels; The aspect ratio of the convolutional structure of the target augmented deep learning network structure is determined based on the length and width values ​​of the convolutional structure.

5. The method as described in claim 1, wherein multiple augmented deep learning network structures are generated based on the initial deep learning network structure, comprising: Adjust the hyperparameters in the initial deep learning network structure, wherein the hyperparameters include at least one of the following: number of convolution channels, kernel size, and number of convolutional network layers; Generate the corresponding augmented deep learning network structure after each adjustment.

6. The method as described in claim 1, wherein the deep learning network structure parameter threshold further includes a convolutional network parameter threshold and a convolutional network computational cost threshold; The method further includes: Obtain the convolution parameters and computational cost of each augmented deep learning network; Augmented deep learning network structures whose convolutional structure aspect ratio is less than or equal to the convolutional network structure aspect ratio threshold, whose convolutional parameters are less than or equal to the convolutional parameter threshold, and whose convolutional computation is less than or equal to the convolutional computation threshold are added to the set of deep learning network structures to be screened.

7. A system for generating deep learning network structures, comprising: A cloud-based device is used to acquire initial deep learning network structures and corresponding business requirement information for image processing services. The business requirement information includes deep learning network structure parameter thresholds, including a convolutional network structure aspect ratio threshold. Multiple augmented deep learning network structures are generated based on the initial deep learning network structure. The aspect ratio of the convolutional structure of each augmented deep learning network is obtained, and this aspect ratio is determined based on the number of convolutional network layers, kernel size, and number of convolutional channels. Augmented deep learning network structures with aspect ratios less than or equal to the convolutional network structure aspect ratio threshold are added to the deep learning network to be screened. The system comprises a set of structures, which includes determining a reference deep learning network structure from the set of deep learning network structures to be screened, generating a target deep learning network structure corresponding to the image processing service based on the reference deep learning network structure and preset network generation conditions, and training a target deep learning network model for the image processing service based on the target deep learning network structure. The preset network generation conditions are defined as the number of consecutive iterations of the reference deep learning network structure hitting a preset threshold. Determining the reference deep learning network structure from the set of deep learning network structures to be screened includes: determining the reference deep learning network structure from the set of deep learning network structures to be screened based on the height, width, and number of channels of the feature image. An edge device is used to execute the image processing service according to the target service deep learning network model and obtain the processing result corresponding to the image processing service.

8. A computing device, comprising: memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the method according to any one of claims 1-6.

9. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the method according to any one of claims 1-6.

Images