Method, device, equipment and medium for determining performance parameter value of network

By using parameter sharing mechanism and cumulative gradient value evaluation in neural network architecture search, the problem of insufficient accuracy of sub-models on verification data sets is solved, the search accuracy of neural network architecture search method is improved, and the sub-network is inferred stage of high performance.

CN112686321BActive Publication Date: 2024-12-03FORCE POLYMERIZATION (CHONGQING) INFORMATION TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202011639067.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2020-12-31
Publication Date
2024-12-03
Estimated Expiration
2040-12-31

AI Technical Summary

Technical Problem

The existing neural network architecture search method has insufficient accuracy on the verification data set, resulting in the searched sub-model performance in the inference stage. How to improve the search accuracy of neural network architecture search method is an urgent problem to be solved. question.

Method used

By combining multiple network modules with the original network, multiple subnets are alternately trained, and their performance parameter values ​​are determined based on the accumulated gradient values ​​of the subnet during training, the parameter sharing mechanism is used to make the parameters of the subnet be updated by multiple subnets. This truly reflects its performance and improves evaluation accuracy.

Benefits of technology

By accumulating gradient values ​​as evaluation indicators, the performance of the subnet can be truly reflected, the search accuracy of the neural network architecture search method can be improved, and the high performance of the searched subnets in the inference stage can be ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN112686321B_ABST
    Figure CN112686321B_ABST
Patent Text Reader

Abstract

The embodiments of the present application provide a method, apparatus, device and medium for determining the performance parameter value of a network, belonging to the field of machine learning technology. The method comprises: combining a plurality of network modules with an original network respectively to obtain a plurality of sub-networks of the original network, wherein a network module is used to implement a type of data processing operation; training the plurality of sub-networks alternately for multiple times to complete the training of the original network, wherein the parameters of the network module are shared among the sub-networks including the same network module; and determining the performance parameter value of any sub-network among the plurality of sub-networks according to the accumulated gradient values ​​of the respective network modules included in the sub-network during the training process of the original network.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a method, device, equipment and medium for determining a performance parameter value of a network. Background Art

[0002] With the development of deep learning, the performance of deep neural networks has achieved significant performance improvements in various visual and speech tasks. However, existing neural network structures, such as ResNet and DenseNet, only propose a general model topology paradigm. Although these paradigms can achieve relatively good results in some general computer tasks, with the complexity of application scenarios, it is necessary to adapt the model structure for different tasks and different conditional constraints (such as computing power constraints and storage constraints in embedded devices) to achieve the best balance between performance and conditional constraints. However, this process requires high labor costs, so the neural network architecture search method came into being.

[0003] Among them, the neural network architecture search method mainly defines the search space, samples multiple sub-models from the search space for training from scratch, uses the accuracy of the sub-model on the validation data set as an evaluation indicator of model performance, and then combines reinforcement learning or evolutionary algorithms to search the model structure. However, since the accuracy of the sub-model on the validation data set is not sufficient to reflect the actual performance of the sub-model, the performance of the sub-model searched out subsequently in the inference stage is not high. Therefore, how to improve the accuracy of the sub-model searched out by the neural network architecture search method is a technical problem that needs to be solved urgently. Summary of the invention

[0004] In view of the above problems, a method, apparatus, device and medium for determining a performance parameter value of a network are proposed in an embodiment of the present application to overcome the above problems or at least partially solve the above problems.

[0005] In order to solve the above problem, the first aspect of the present application discloses a method for determining a performance parameter value of a network, the method comprising:

[0006] Combining a plurality of network modules with the original network respectively to obtain a plurality of sub-networks of the original network, wherein one network module is used to implement one type of data processing operation;

[0007] The plurality of sub-networks are alternately trained multiple times to complete the training of the original network, wherein the parameters of the network module are shared among the sub-networks including the same network module;

[0008] For any sub-network among the multiple sub-networks, a performance parameter value of the sub-network is determined according to the accumulated gradient values ​​of each network module included in the sub-network during the training process of the original network.

[0009] Optionally, the method further comprises:

[0010] During the training process of the original network, the accumulated gradient values ​​of each of the plurality of network modules during the training process of the original network are counted to obtain accumulated gradient value sequences corresponding to each of the plurality of network modules;

[0011] For any sub-network among the multiple sub-networks, determining a performance parameter value of the sub-network includes:

[0012] Taking the accumulated gradient value sequence corresponding to each network module included in the sub-network as a row, so as to obtain the accumulated gradient value matrix of the sub-network;

[0013] The accumulated gradient value matrix of the sub-network is processed to obtain the performance parameter value of the sub-network.

[0014] Optionally, during the training process of the original network, the subnetwork targeted at each training is any subnetwork among the multiple subnetworks; the method further includes:

[0015] Take the subnetwork targeted for each training as the target subnetwork and perform the following steps:

[0016] Using the training samples, the target sub-network is trained to obtain the gradient values ​​of the respective network modules included in the target sub-network;

[0017] Using the respective gradient values ​​of the respective network modules included in the target sub-network, updating the respective accumulated gradient values ​​of the respective network modules belonging to the target sub-network among the multiple network modules;

[0018] In the training process of the original network, a cumulative gradient value sequence corresponding to any network module in the plurality of network modules is obtained, including:

[0019] During the training process of the original network, the accumulated gradient values ​​after each update of the network module are arranged in sequence according to the order of update time to obtain the accumulated gradient value sequences corresponding to the multiple network modules.

[0020] Optionally, for at least one group of edges connecting adjacent original network modules in the original network, perform the following steps:

[0021] Configuring candidate operands corresponding to the plurality of network modules respectively for the edge;

[0022] Sampling a target operand of the edge from the candidate operands respectively corresponding to the plurality of network modules;

[0023] A network module corresponding to the target candidate operand is added between a group of adjacent original network modules connected by the edge.

[0024] Optionally, after determining the performance parameter values ​​of each of the multiple sub-networks, the method further includes:

[0025] Determining at least some of the multiple subnetworks as search spaces according to respective performance parameter values ​​of the multiple subnetworks;

[0026] Taking the maximum performance parameter value as the search target, searching the search space to obtain the search space for the next search;

[0027] When a preset search end condition is met, the searched network is determined as the target network.

[0028] Optionally, determining at least some of the multiple subnetworks as search spaces according to respective performance parameter values ​​of the multiple subnetworks includes:

[0029] Determine a first preset number of subnetworks with top performance parameter values ​​among the multiple subnetworks as a search space;

[0030] The search space is searched with the maximum performance parameter value as the search target to obtain the search space for the next search, including:

[0031] Selecting a second preset number of sub-networks with top ranking performance parameter values ​​from the search space;

[0032] The second preset number of sub-networks is determined as the search space for the next search.

[0033] Optionally, after determining the performance parameter values ​​of each of the multiple sub-networks, the method further includes:

[0034] Using at least some of the multiple subnetworks as an initialization population, and determining a performance parameter value of each subnetwork in the initialization population;

[0035] Select K sub-networks with top ranking performance parameter values ​​from the initialization population;

[0036] Perform crossover mutation on the K sub-networks to obtain a new population, and determine the performance parameter value of each new network in the new population;

[0037] The following steps are performed multiple times: K new networks with top ranking performance parameter values ​​are selected from the new population, cross-mutation is performed on the K new networks to obtain a new population, and the performance parameter values ​​of each new network in the new population are determined to obtain a target network.

[0038] Optionally, after obtaining the target network, the method further includes:

[0039] The target network is trained using a plurality of sample images to obtain an image processing network, wherein the sample images carry classification labels or object recognition labels.

[0040] Optionally, the method further comprises:

[0041] Obtaining an image to be processed;

[0042] Inputting the image to be processed into the image processing network to obtain a processing result output by the image processing network;

[0043] The processing result is a classification result of the image to be processed or a recognition result of an object contained in the image to be processed.

[0044] According to a second aspect of an embodiment of the present application, a device for determining a performance parameter value of a network is provided, the device comprising:

[0045] A sub-network determination module, used to combine multiple network modules with the original network respectively to obtain multiple sub-networks of the original network, wherein one network module is used to implement one type of data processing operation;

[0046] A training module, used for alternately training the multiple sub-networks for multiple times to complete the training of the original network, wherein the sub-networks of the same network module share the parameters of the network module;

[0047] The parameter value determination module is used to determine the performance parameter value of any sub-network among the multiple sub-networks according to the accumulated gradient values ​​of each network module included in the sub-network during the training process of the original network.

[0048] The third aspect of the embodiments of the present application further discloses an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes, the method for determining the performance parameter value of a network as described in the first aspect of the embodiments is implemented.

[0049] The fourth aspect of the embodiments of the present application further discloses a computer-readable storage medium, which stores a computer program that enables a processor to execute the method for determining a performance parameter value of a network as described in the embodiments of the first aspect of the present application.

[0050] The embodiments of the present application include the following advantages:

[0051] In an embodiment of the present application, multiple network modules can be combined with the original network respectively to obtain multiple sub-networks of the original network, wherein one network module is used to implement one type of data processing operation; then, the multiple sub-networks are alternately trained multiple times to complete the training of the original network, wherein the parameters of the network module are shared among the sub-networks including the same network module; thereafter, for any sub-network among the multiple sub-networks, the performance parameter value of the sub-network is determined according to the accumulated gradient values ​​of each network module included in the sub-network during the training process of the original network.

[0052] In an embodiment of the present application, since the parameters of the network modules of each sub-network can be shared during the process of alternating training of multiple sub-networks, the parameters of a network module can be updated by multiple sub-networks through parameter sharing. Therefore, the cumulative gradient value of the network module included in a sub-network can truly reflect the optimization goal and optimization method of the sub-network itself. In this way, the performance parameter value of the sub-network determined according to the cumulative gradient value can truly reflect the performance of the sub-network, and then using the performance parameter value as an evaluation indicator can reflect the true performance of the sub-network, so that using the performance parameter value as a search factor can improve the accuracy of the sub-network searched by the neural network architecture search method. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments of the present application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0054] Figure 1 It is a schematic diagram of the overall process of a method for determining a performance parameter value of a network according to an embodiment of the present application;

[0055] Figure 2 is a schematic diagram of the structure of an original network to be processed in an embodiment of the present application;

[0056] Figure 3 It is a flowchart of the steps of obtaining multiple network modules in the implementation of this application;

[0057] Figure 4 It is a schematic diagram of the steps of a method for determining a performance parameter value of a network in the implementation of the present application;

[0058] Figure 5 is a flow chart of the steps for determining the performance parameter value of a subnetwork in the implementation of the present application;

[0059] Figure 6 It is a flowchart of the steps of searching for a target network in the implementation of this application;

[0060] Figure 7 This is another flowchart of steps for searching for a target network in the implementation of this application;

[0061] Figure 8 It is a structural block diagram of a device for determining performance parameter values ​​of a network in the implementation of the present application. DETAILED DESCRIPTION

[0062] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and understandable, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0063] In the related art, in neural network search, a neural network structure search method based on parameter sharing is proposed to accelerate the model search efficiency. Specifically, the neural network structure search method based on parameter sharing is: the entire search space is defined as a super network, which is composed of an exponential number of sub-networks, and there is parameter sharing between each sub-network and alternating training. In this way, a model parameter in the super network may be updated by multiple sub-networks, thereby reaching a convergence state.

[0064] However, under the parameter sharing mechanism, a certain parameter of the super network exists in multiple sub-network structures, and the multiple sub-network structures correspond to different optimization objectives and optimization methods, which will lead to a gap between the prediction performance of the obtained sub-network and the prediction performance obtained by training the sub-network from scratch. For example, sub-network 1 and sub-network 2 have parameters, and one update of the parameter may be triggered by the training of sub-network 1, and another update may be triggered by the training of sub-network 2. The optimization objectives and optimization methods of sub-network 1 and sub-network 2 may be different. Therefore, the final parameter value obtained after multiple updates of the parameter integrates the respective optimization objectives and optimization methods of network 1 and network 2, resulting in that when evaluating sub-network 1 and sub-network 2 through the final value of the parameter, it cannot truly reflect the optimization objectives and optimization methods of sub-network 1 itself, nor can it truly reflect the optimization objectives and optimization methods of sub-network 2 itself.

[0065] This results in a poor correlation between the ranking relationship of the prediction performance of multiple sub-networks and the ranking relationship of their actual performance. Therefore, under the parameter sharing mechanism, directly using the prediction accuracy on the validation set as the evaluation indicator of the model prediction performance cannot truly reflect the performance of the searched sub-networks.

[0066] In view of this, the applicant has proposed the following technical concept: recording the training process of the super network, specifically, statistically accumulating the gradient values ​​of the network modules included in each sub-network in the super network during the alternating training process, and then statistically analyzing the dynamic changes of the gradient values ​​of the corresponding network modules of each sub-network in the super network during the training process. The dynamic changes can faithfully reflect the optimization objectives and optimization methods of each sub-network itself, so that using the accumulated gradient value in the training process as the performance evaluation indicator of the sub-network can faithfully reflect the actual performance of each sub-network and improve the evaluation accuracy.

[0067] Reference Figure 1 As shown, a schematic diagram of the overall process of a method for determining a performance parameter value of a network according to an embodiment of the present application is shown, such as Figure 1 As shown, the overall process includes three processes: hypernetwork definition, network search and network evaluation.

[0068] The process defined by the super network is mainly used to construct multiple sub-networks of the original network, wherein the integration of multiple sub-networks of the original network can be understood as a super network.

[0069] Reference Figure 2 As shown, a schematic diagram of the structure of an original network to be processed in the present application is shown, combined with Figure 2 As shown, multiple sub-networks of the original network constructed in this application are introduced. Figure 2 As shown, the original network includes the following network layers:

[0070] The convolution layer is used to perform convolution operations on the image input to the original network;

[0071] There are multiple original modules, each of which is composed of N identical original network modules. The multiple original network modules in each original module are connected in pairs.

[0072] The staggered layer is connected between every two adjacent original modules and is used to downsample the feature maps output by the original modules;

[0073] The global pooling layer is used to pool the feature map output by the last original module.

[0074] Among them, among the multiple original network modules in the original module, the edges connecting the two connected original network modules can preset multiple candidate operands. A single sampling of the candidate operands of the edges of the original network modules can obtain a target candidate operand connected between adjacent original modules, and thus obtain the network module corresponding to the target candidate operand. In this way, different sampling means obtaining different network modules connected between adjacent original modules, so as to combine the original network with different network modules, thereby obtaining different sub-networks of the original network. The collection of multiple different sub-networks constitutes a super network. This process is Figure 1 The hypernetwork definition shown.

[0075] In one sampling, the edges connecting each adjacent original network modules will be sampled. In this way, a sampling result can be obtained, which can include the network modules sampled between each adjacent original network modules. One sampling result can be considered to correspond to a sub-network, and multiple samplings can obtain multiple sub-networks.

[0076] For example, Figure 2 As shown in , each original module includes 4 original network modules, and every two original network modules can be connected to each other, thus obtaining 6 edges, as shown in Figure 2 As shown in the figure, a curve between the original network modules is regarded as an edge. Each of the six edges can be preset with two operands, for example, it can be preset as a conv1×1 operand and a conv3×3 operand, where conv1×1 can be understood as a point convolution process, and conv3×3 operand can be understood as a convolution process with a convolution kernel size of 3×3. During one sampling, the candidate operands on all six edges can be uniformly sampled. After the sampling is completed, a sampling result is obtained, and then a sub-network is obtained.

[0077] Among them, two original network modules connected by an edge are called adjacent original network modules. For example, original network module 1 and original network module 3 are connected, so original network module 1 and original network module 3 can be called adjacent original network modules.

[0078] Reference Figure 3 As shown, a flowchart of the steps of processing the edges connected between adjacent original network modules to obtain multiple network modules is shown, such as Figure 3 As shown, the following steps may be specifically included:

[0079] Step S301: configuring candidate operands corresponding to the plurality of network modules respectively for the edge.

[0080] In an embodiment of the present application, different candidate operands correspond to a different network module. For example, when the candidate operand is conv1×1, it corresponds to the convolution module of conv1×1, and when the candidate operand is conv3×3, it corresponds to the convolution module of conv3×3. In this way, the edges connected between each two adjacent original network modules carry multiple candidate operands. Since multiple original network modules are connected in pairs, multiple edges can be obtained. In an embodiment of the present application, each of the multiple edges carries the same multiple candidate operands, that is, different edges carry the same multiple candidate operands. In this way, the complexity of the constructed super network can be reduced.

[0081] For example, Figure 2 As shown, the edge connecting the original network module 1 and the original network module 2 has two candidate operands, namely conv1×1 and conv3×3, and the edge connecting the original network module 1 and the original network module 3 also has the above two candidate operands, namely conv1×1 and conv3×3.

[0082] Step S302: sampling a target operand of the edge from the candidate operands respectively corresponding to the plurality of network modules.

[0083] In this embodiment, since the edge connecting adjacent original network modules carries multiple candidate operands, different candidate operands correspond to different network modules. Therefore, sampling of the multiple candidate operands carried by the edge can be understood as sampling of multiple network modules.

[0084] In specific implementation, multiple candidate operands carried by each edge may be uniformly sampled to sample the target candidate operand. Uniform sampling means that the sampling probabilities of different candidate operands among multiple candidate operands are the same, such as Figure 2 For example, the candidate operands corresponding to the edge include conv1×1 and conv3×3, then the sampling probability of conv1×1 is 0.5, and the sampling probability of conv3×3 is 0.5. In this way, the probability of each candidate operand being sampled as the target candidate operand is the same.

[0085] Among them, uniform sampling of multiple candidate operands can ensure that the probability of each candidate operand being sampled is the same, so that the diversity of the final sub-network can be improved to construct a broader super network search space.

[0086] Of course, according to actual needs, when sampling multiple candidate operands carried by each edge, the sampling may not be uniform. For example, sampling weights may be preset for different candidate operands, and sampling is performed according to the sampling weights.

[0087] Step S303: adding a network module corresponding to the target candidate operand between a group of adjacent original network modules connected by the edge.

[0088] In the embodiment of the present application, since different candidate operands correspond to different network modules, after the target candidate operand is sampled from the edge connecting two adjacent original network modules, the target candidate operand corresponds to a network module. In this way, the network module corresponding to the target candidate operand can be connected between adjacent original network modules, thereby realizing the combination between the network module and the original network. In this way, in one sampling, each edge is sampled in the above manner, thereby obtaining a network module connected between each two adjacent original network modules, and after one sampling is completed, a sub-network is obtained.

[0089] For example, Figure 2 As shown, the edge connecting the original network module 1 and the original network module 2 has two candidate operands, namely conv1×1 and conv3×3. However, after conv1×1 is sampled, the convolution module of conv1×1 can be connected between the original network module 1 and the original network module 2. Similarly, the candidate operands of the edges connecting the four original network modules in pairs can be sampled to obtain the target candidate operands corresponding to the six edges. Then, the network modules corresponding to the sampled target candidate operands are connected between the original network modules connected in pairs, thereby realizing the combination of the sampled six network modules with the original network to obtain a sub-network.

[0090] Then, if Figure 1 As shown, the network search process can be entered, that is, multiple sub-networks can be trained alternately to determine the performance parameter values ​​of each of the multiple sub-networks, and network search can be performed based on the performance parameter values.

[0091] Reference Figure 4 As shown, a flowchart of the steps of the method for determining the performance parameter value of the network in an embodiment of the present application is shown. The method for determining the performance parameter value of the network in an embodiment of the present application can be applied to a terminal device, such as Figure 4 As shown, the following steps may be specifically included:

[0092] Step S401: Combine multiple network modules with the original network respectively to obtain multiple sub-networks of the original network, wherein one network module is used to implement one type of data processing operation.

[0093] In the embodiment of the present application, the process of combining multiple network modules with the original network can refer to the process described in the above steps S301 to S303. Specifically, for the multiple original network modules in each original module in the original network, by sampling the multiple candidate operands of the edges connected to the two original network modules, the network modules corresponding to the sampled target candidate operands are added between the two original network modules to realize the combination of the network module and the original network, and obtain a sub-network. Through continuous sampling, multiple different sub-networks are obtained, so that multiple different sub-networks constitute the search space to be searched, that is, constitute the super network to be trained.

[0094] Among them, as described in the above embodiment, different network modules correspond to different candidate operands, that is, corresponding to a type of data processing operation. In the embodiment of the present application, due to the different candidate operands, the data processing operation performed by the network module can at least include a conv1×1 convolution operation and a conv3×3 convolution operation. Of course, it can also include convolution operations of other sizes, or other types of data processing operations. The present application does not limit the types of data processing operations.

[0095] In the subnetwork construction process implemented by the present application, since each subnetwork is obtained by combining the network module corresponding to the sampled target candidate operand with the original network, the same network module can be included in different subnetworks, that is, different subnetworks can share the same candidate operand.

[0096] Step S402: training the multiple sub-networks alternately for multiple times to complete the training of the original network.

[0097] Among them, the sub-networks of the same network module share the parameters of the network module.

[0098] In this embodiment, performing alternating training on multiple subnetworks may mean that the subnetwork trained last time and the subnetwork trained this time are different subnetworks. For example, subnetwork 1 is trained for the nth time, and subnetwork 2 is trained for the n+1th time. When performing alternating training on multiple subnetworks, since it involves updating the parameters of multiple subnetworks, the alternating training on multiple subnetworks can be understood as a training process for a super network. When the super network reaches convergence, the training process can be considered to be completed.

[0099] Among them, when training multiple sub-networks, for each sub-network trained, a set of training data can be sampled from the training set as the input of the sub-network, the label of the training data is used to update the cross-entropy loss function, and the back-propagation algorithm is used on the cross-entropy function to obtain the gradient of each network module in the sub-network.

[0100] Among them, in the process of alternating training of multiple sub-networks, different sub-networks can correspond to different optimization objectives and optimization methods. Since the same network module can be included in different sub-networks, in the process of training multiple sub-networks, for the same network module, these multiple sub-networks including the network module can share the parameters of the network module.

[0101] Specifically, for each sub-network among multiple sub-networks including the same network module, when training the sub-network, the basic parameters of the network module in the current training can be used as a benchmark for training. At the end of the current training, the basic parameters of the network module are updated to obtain the updated parameters of the network module after the end of the current training. Then, when training the next sub-network, the updated parameters of the network module can be used as the benchmark parameters for training. After the training is completed, the updated parameters of the network module are updated again, and this cycle is repeated so that the parameters of the network module are updated together by multiple sub-networks.

[0102] For example, assuming that the obtained sub-network 1 has a network module conv1×1, and sub-network 2 also has a network module conv1×1, then the sub-network 1 and sub-network 2 share the parameters of the network module conv1×1. That is, when sub-network 1 is trained for the nth time, the training is based on the parameters of the network module conv1×1 at the nth time, and at the end of the nth training, the updated parameters of the network module conv1×1 are obtained. Then, when sub-network 2 is trained for the n+mth time, the training is based on the updated parameters of the network module conv1×1 obtained after the nth training, and at the end of the n+mth training, the updated parameters of the network module conv1×1 are updated again, and so on.

[0103] In this embodiment, the parameters of the network module are shared between the various sub-networks of the same network module, which is a neural network structure search based on parameter sharing. Through parameter sharing, the parameters of a network module can be updated by multiple sub-networks, so as to quickly reach a convergence state, thereby accelerating the training efficiency and thus improving the network search efficiency. At the same time, the update process of the parameters of a network module can reflect the optimization objectives and optimization methods of multiple sub-networks.

[0104] Step S403: for any sub-network among the multiple sub-networks, determine the performance parameter value of the sub-network according to the accumulated gradient values ​​of each network module included in the sub-network during the training process of the original network.

[0105] In an embodiment of the present application, since the parameters of the same network module are shared among the various sub-networks of the same network module, the parameters of one network module are updated by multiple sub-networks. Therefore, in the process of alternating training of multiple sub-networks, the cumulative gradient values ​​of each network module in multiple updates can be recorded, that is, the cumulative gradient values ​​of each update can be recorded, and the cumulative gradient values ​​in the training process can characterize the dynamic changes in the gradient values ​​of the parameters of the network module during the update process of each sub-network.

[0106] Since each sub-network of the present application is obtained by combining the network module corresponding to the sampled target candidate operand with the original network, any of the multiple sub-networks includes at least one corresponding network module, and since the cumulative gradient value of each network module can characterize the dynamic change of the gradient value of the parameters of the network module during the update process of the multiple sub-networks. Therefore, the cumulative gradient values ​​of the multiple network modules included in the sub-network during the training process can faithfully reflect the optimization target and optimization method of the sub-network during the training process, so that the performance parameter value of the sub-network can be determined according to the cumulative gradient values ​​of the network modules included in the sub-network during the training process, so that the determined performance parameter value of the sub-network can faithfully reflect the performance of the sub-network.

[0107] By adopting the technical solution of the embodiment of the present application, since the parameters of the same network module are shared among the various sub-networks of the network module, the parameters of one network module can be updated by multiple sub-networks through parameter sharing, and then the current cumulative gradient value of the network module included in the sub-network can faithfully reflect the optimization objectives and optimization methods of each sub-network itself, so that the performance parameter value of the sub-network calculated by the cumulative gradient value can faithfully reflect the performance of the sub-network, and using the performance parameter value as the search factor can improve the accuracy of the sub-network searched by the neural network architecture search method.

[0108] In one embodiment, referring to Figure 5 As shown, a flowchart of the steps of determining the performance parameter value of the sub-network is shown, specifically, the following steps are included:

[0109] Step S501: During the training process of the original network, the accumulated gradient values ​​of each of the multiple network modules during the training process of the original network are counted to obtain accumulated gradient value sequences corresponding to each of the multiple network modules.

[0110] In an embodiment of the present application, since the parameters of the network module can be shared by multiple sub-networks, that is, multiple sub-networks contribute to the parameter update of the common network module during the training process, the cumulative gradient value of the network module can truthfully reflect the optimization objectives and optimization methods of the multiple sub-networks during the training process.

[0111] Among them, for each network module, the process of multiple updates of the parameters of the network module in the process of training multiple sub-networks can be recorded, that is, the cumulative gradient value obtained after each update of the gradient value of the network module is recorded, so as to obtain the cumulative gradient value sequence corresponding to each network module. Correspondingly, the cumulative gradient value sequence corresponding to the network module can include the cumulative gradient value of each update of the parameters of the network module, and one cumulative gradient value can truthfully reflect a training of a corresponding sub-network.

[0112] In one example, during the training process of the original network, the accumulated gradient values ​​after each update of the network module are arranged in sequence according to the update time to obtain the accumulated gradient value sequences corresponding to the multiple network modules.

[0113] In this embodiment, the sequence of update time can be arranged to truly reflect the actual process of the gradient value of the network module being updated by multiple sub-networks. In this way, the cumulative gradient value sequence of the network module can truly reflect the optimization process of each of the multiple sub-networks.

[0114] In one example, for the same network module, multiple accumulated gradient values ​​corresponding to multiple different sub-networks including the network module can be screened out from the accumulated gradient value sequence corresponding to the network module, thereby obtaining multiple accumulated gradient values ​​obtained by each sub-network when the network module is updated during multiple trainings. Similarly, multiple accumulated gradient values ​​obtained when multiple network modules included in a sub-network are updated multiple times by the sub-network can be obtained to reflect the performance of the sub-network during the training process.

[0115] In this way, the accumulated gradient value sequence of a network module can reflect the update process of the corresponding sub-network on the network module each time the corresponding sub-network is trained. In this way, the optimization objectives and optimization methods of different sub-networks can be known through the accumulated gradient value sequence.

[0116] Example #, for network module A, sub-network 1 and sub-network 2 share the parameters of network module A, where sub-network 1 is trained for the 1st, 3rd and 5th times, and sub-network 2 is trained for the 2nd, 6th and 7th times. Accordingly, sorted by the number of training times, the cumulative gradient value sequence of network module A can be expressed as {g, g+g1, g+g1+g2, g+g1+g2+g3, g+g1+g2+g3+g4, g+g1g2+g3+g4+g5}. Among them, the multiple cumulative gradient values ​​of sub-network 1 are {g, g+g1+g2, g+g1+g2+g3}, and the multiple cumulative gradient values ​​of sub-network 2 are {g+g1, g+g1+g2+g3+g4, g+g1g2+g3+g4+g5}. Thus, through {g, g+g1+g2, g+g1+g2+g3}, we can get the optimization target and optimization method of sub-network 1 during training. Through {g+g1, g+g1+g2+g3+g4, g+g1g2+g3+g4+g5}, we can get the optimization target and optimization method of sub-network 2 during training.

[0117] In one example, during the training process of the original network, the subnetwork targeted at each training is any subnetwork among the multiple subnetworks, and the subnetwork trained each time can be called a target subnetwork. Then, the process of updating the gradient value of the network module can be as follows:

[0118] Step S5011: using the training samples, the target sub-network is trained to obtain the gradient values ​​of the respective network modules included in the target sub-network.

[0119] In an embodiment of the present application, when training the target sub-network, training samples can be input into the target sub-network to obtain the prediction results of the target sub-network, and the prediction results and corresponding labels are used to update the cross-entropy loss function. The back propagation algorithm is used on the cross-entropy function to obtain the gradient values ​​of each network module in the sub-network.

[0120] As shown in the above example #, taking sub-network 1 as the target sub-network as an example, sub-network 1 includes network module A. Sub-network 1 is trained for the first time, and the gradient value of network module A obtained by sub-network 1 during the third training is g2.

[0121] Step S5012: using the gradient values ​​of the respective network modules included in the target sub-network, updating the accumulated gradient values ​​of the respective network modules in the plurality of network modules belonging to the target sub-network.

[0122] In the embodiment of the present application, since the parameters of the same network module can be shared by multiple sub-networks, that is, the parameters of one network module are updated by multiple sub-networks. In this case, when obtaining the gradient values ​​of each network module included in the target sub-network during the current training, it is necessary to add the gradient values ​​of each network module to the accumulated gradient values ​​of each network module to obtain the accumulated gradient values ​​of each network module. By adding the gradient values ​​of each network module to the accumulated gradient values ​​of each network module, the update of the accumulated gradient values ​​of each network module is realized.

[0123] As shown in Example # above, the gradient value of network module A obtained by sub-network 1 during the third training is g2, and the accumulated gradient value of network module A is g+g1. In this way, after updating the gradient value of network module A, the updated accumulated gradient value of network module A is g+g1+g2.

[0124] Step S502: taking the accumulated gradient value sequence corresponding to each network module included in the sub-network as a row to obtain the accumulated gradient value matrix of the sub-network.

[0125] Since for each network module, the cumulative gradient value after each update of the network module can be arranged in sequence according to the order of the update time, the cumulative gradient value sequence corresponding to the network module is obtained. In an embodiment of the present application, in order to associate the cumulative gradient value sequence of the network module with the training process of the subnetwork including the network module, so as to more intuitively reflect the network optimization process of each of the multiple subnetworks during the training process, for each subnetwork, the cumulative gradient value sequence corresponding to each network module included in the subnetwork can be taken as a row, thereby obtaining the cumulative gradient value matrix of the subnetwork. Each row in the cumulative gradient value matrix corresponds to a network module included in a subnetwork, and each column represents the cumulative gradient value of each network module included when the subnetwork is trained once.

[0126] In the embodiment of the present application, the cumulative gradient value matrices corresponding to the multiple sub-networks can be obtained. Through the cumulative gradient value matrices corresponding to the multiple sub-networks, the optimization objectives and optimization methods of the multiple sub-networks during the training process can be reflected.

[0127] In this embodiment, since multiple sub-networks are obtained by combining multiple network modules with the original network in different ways, different network modules correspond to different data processing operations, and therefore, different sub-networks may correspond to different optimization objectives and optimization methods. Among them, the optimization objective may refer to the accuracy that the sub-network needs to achieve. For example, taking the training task as an object classification task as an example, the optimization objective may be set to the degree of classification accuracy of the object classification. The optimization method may refer to an optimization method for accelerating neural network training, for example, it may include gradient descent method (BGD and SGD), Adadelta method, Adagrad method (Adagrad and AdagradDAO), Momentum method (Momentum and Nesterov Momentum), Adam, Ftrl method and RMSProp method.

[0128] Therefore, during the training process, the cumulative gradient values ​​of the statistical network module updated by multiple sub-networks can intuitively reflect the contribution of each sub-network to the parameter update of the network module under its corresponding optimization objectives and optimization methods, so as to truly reflect the performance changes of each sub-network during the training process.

[0129] Referring to Table 1, Table 1 is a statistical table of the cumulative gradient value matrix obtained by training sub-network 1 multiple times in an embodiment of the present application.

[0130] As shown in Table 1, a column represents a training session, and a row represents a network module. As shown in Table 1, subnetwork 1 includes network modules OP1, OP3, and OP6, wherein the subnetwork is trained for the W1th, W3th, and W6th times, and the gradient value of each training session is △g, and the cumulative gradient value obtained for each training session is g+△g. Of course, the △g obtained for different training sessions is different, which is not shown in Table 1.

[0131] Table 1: Statistics of the cumulative gradient value matrix of subnetwork 1

[0132] W1 W2 W3 W4 W5 W6 OP1 g+△g g+0 g+△g g+0 g+0 g+△g OP2 g g g g g g OP3 g+△g g+0 g+△g g+0 g+0 g+△g OP4 g g g g g g OP5 g g g g g g OP6 g+△g g+0 g+△g g+0 g+0 g+△g

[0133] As shown in Table 1, the row where OP1 is located can intuitively reflect the contribution of sub-network 1 to the parameter update of OP1 in each of the six trainings from W1 to W6. In this way, the combination of OP1, OP3 and OP6 can reflect the optimization goal and optimization method of sub-network 1 in the training process. The performance parameter value obtained by processing the cumulative gradient values ​​of OP1, OP3 and OP6 in the training process can accurately describe the performance of sub-network 1.

[0134] Step S503: Process the accumulated gradient value matrix of the sub-network to obtain the performance parameter value of the sub-network.

[0135] After obtaining the cumulative gradient value matrix corresponding to each sub-network, the cumulative gradient value matrix of the sub-network can be processed to obtain the performance parameter value of the sub-network.

[0136] Specifically, for each sub-network, each network module constituting the sub-network can be defined as a random variable, such as OP1, OP3 and OP6 in Table 1 are random variables, and in the cumulative gradient matrix, each cumulative gradient value related to each network module can be defined as a sample of the random variable, such as each cumulative gradient value related to OP1 in Table 1 is used as a sample, and the performance of the sub-network is defined by counting the correlation coefficients of X random variables. The higher the correlation coefficient, the better the sub-network performance, and the correlation coefficient of the X random variables can be obtained by solving the determinant of the correlation coefficient matrix corresponding to the X random variables.

[0137] Thus, the determinant of the cumulative gradient value matrix of the sub-network can be solved to obtain the correlation coefficient between the multiple network modules included in the sub-network, and the correlation coefficient can be used as the performance parameter value of the sub-network.

[0138] In this embodiment, after each training, the cumulative gradient value matrix of each sub-network is obtained, and the performance parameter values ​​of the multiple sub-networks trained can be determined according to the cumulative gradient value matrix obtained by each update. Of course, in an example, the performance parameter values ​​of the trained sub-networks can be determined according to the cumulative gradient value matrix obtained by updating each sub-network after each N training, and then according to the performance parameter values ​​of the multiple sub-networks, some sub-networks are screened out to continue training until the training end condition is reached, that is, until the super network converges, thereby searching for the target network.

[0139] Specifically, the target network can be searched out according to the new energy parameter value of the sub-network in the following two ways. Figure 6 and Figure 7 As shown, Figure 6 A flowchart showing the steps of searching for a target network by way of method A is shown. Figure 7 A flowchart of the steps of searching for a destination network through method B is shown.

[0140] like Figure 6 As shown, the target network can be searched out by way A, which can specifically include the following steps:

[0141] Step S601: according to the performance parameter values ​​of the multiple sub-networks, at least some of the multiple sub-networks are determined as search spaces.

[0142] In the present method A, since the correlation coefficients between the multiple network modules included in the sub-network obtained by solving the determinant of the cumulative gradient value matrix of the sub-network can be used as the performance parameter value of the sub-network, the higher the performance parameter value, the better the performance of the sub-network is guaranteed. In this way, at least part of the sub-networks can be screened out from the multiple sub-networks according to their respective performance parameter values, and the screened part of the sub-networks are used as new search spaces.

[0143] In an example, when at least some of the subnetworks among the multiple subnetworks are determined as search spaces according to the performance parameter values ​​of the multiple subnetworks, a first preset number of subnetworks with top-ranked performance parameter values ​​among the multiple subnetworks may be determined as search spaces.

[0144] Among them, ranking the performance parameter values ​​at the top may refer to sorting multiple sub-networks in order of performance parameter values ​​from high to low, the performance parameter values ​​of the sub-networks arranged in the front are higher than the performance parameter values ​​of the sub-networks arranged in the back, and the first preset number of sub-networks ranked at the top of the performance parameter values ​​may refer to: sub-networks with higher performance parameter values, that is, sub-networks with better performance.

[0145] Specifically, for each N trainings, the performance parameter values ​​of the trained multiple sub-networks can be obtained according to their respective accumulated gradient value matrices, and each sub-network can be evaluated according to the performance parameter values. Then, according to the evaluation results, some sub-networks can be selected from the trained multiple sub-networks as the search space to be searched. Then, the search space to be searched is searched with the maximum performance parameter value as the search target to obtain the search space for the next search.

[0146] Step S602: Taking the maximum performance parameter value as the search target, searching the search space to obtain the search space for the next search.

[0147] In one example, when searching the search space with the maximum performance parameter value as the search target to obtain the search space for the next search, a second preset number of sub-networks with top-ranked performance parameter values ​​can be screened from the search space to be searched; and the second preset number of sub-networks are determined as the search space for the next search.

[0148] Among them, taking the maximum performance parameter value as the search target may mean searching until the performance parameter value of the sub-network meets the preset maximum performance parameter value.

[0149] Among them, the second preset number is smaller than the first preset number, so that the search space can be further narrowed. Among them, when the search space for the next search is obtained, the above steps S402 to S403 can be repeated, that is, some sub-networks in the search space are alternately trained, and the cumulative gradient values ​​of each network module included in each sub-network in the partial sub-network during multiple trainings are recorded, and then the performance parameter value of each sub-network in the partial sub-network is determined according to the cumulative gradient value, and then the sub-network for the next search is screened out according to the respective performance parameter values ​​of the partial sub-networks, thereby forming the search space for the next search, and repeating in this way can continuously narrow the search space, that is, the number of retained sub-networks becomes smaller and smaller, until the preset search end condition is met, and the sub-network retained when the preset search end condition is met can be determined as the target network.

[0150] Step S603: When a preset search end condition is met, the searched network is determined as a target network.

[0151] In the embodiment of the present application, the preset search end condition may refer to that the search space determination is repeated for a preset number of times, or that the performance parameter value of the sub-network reaches a preset maximum performance parameter value.

[0152] By using this method A to search for the target network, the constructed multiple sub-networks can be used as the original search space, and the performance parameter values ​​of the multiple sub-networks can be determined according to the cumulative gradient values ​​of the multiple network modules during the alternating training of the multiple sub-networks. Based on the performance parameter values, some sub-networks are screened out to continuously shrink the search space to obtain the target network. The network structure of the searched target network can be consistent with the network structure of the target network before training, which can facilitate the evaluation of the target network.

[0153] like Figure 7 As shown, the target network can be searched out by way B, which can specifically include the following steps:

[0154] Step S701: using at least some of the multiple sub-networks as an initialization population, and determining a performance parameter value of each sub-network in the initialization population.

[0155] In this method B, after alternately training multiple subnetworks, some subnetworks can be selected from the multiple subnetworks as the initialization population, or all subnetworks can be used as the initialization population. Then, the performance parameter value of each subnetwork in the initialization population can be determined according to the cumulative gradient value matrix of each subnetwork in the initialization population, and the subnetworks in the initialization population can be sorted in descending order of the performance parameter value.

[0156] Step S702: Select K sub-networks with top ranking performance parameter values ​​from the initialization population.

[0157] In this method B, since the subnetworks in the initialization population are sorted in descending order of performance parameter values, the top K subnetworks can be screened from the initialization population, and the top K subnetworks are subnetworks with higher performance.

[0158] Step S703: performing crossover mutation on the K sub-networks to obtain a new population, and determining the performance parameter value of each new network in the new population.

[0159] In this way, the following steps are performed multiple times: K new networks with top performance parameter values ​​are selected from the new population, cross-mutation is performed on the K new networks to obtain a new population, and the performance parameter values ​​of each new network in the new population are determined to obtain the target network. Each time the above steps are repeated, the value of K decreases as the number of executions increases, thereby continuously reducing the new population.

[0160] Among them, a genetic algorithm can be used to realize crossover mutation of K sub-networks, so that K new networks are obtained, and thus the K new networks form a new population, and then the K new networks in the new population are alternately trained, and the cumulative gradient values ​​of each network module included in each of the K new networks during multiple trainings are recorded, and then the performance parameter value of each new network in the K new networks is determined according to the cumulative gradient value, and then some sub-networks are screened out according to the respective performance parameter values ​​of the K new networks, and then the newly screened sub-networks are cross-mutated to form the next search population, and this is repeated. When the preset search end condition is met, the retained sub-network is used as the target network.

[0161] By using this method B to search for the target network, the constructed multiple sub-networks can be used as the initialization population, and the performance parameter value of each sub-network is determined according to the cumulative gradient values ​​of multiple network modules in each sub-network in the initialization population during the alternating training process, and the performance parameter value is used as an indicator to screen out some sub-networks for cross-mutation to obtain a new population, and repeat this process to continuously obtain new populations, and the size of the new population is continuously reduced to obtain the target network. Since the new population is obtained by continuous cross-mutation, the network structure of the target network searched in this way can be different from the network structure before training, thereby improving the novelty of the target network of the network search.

[0162] In an embodiment of the present application, after searching for the target network, the target network can be retrained on the ImageNet dataset, and its prediction accuracy on the test set after training can be used as an evaluation indicator of the quality of the neural network algorithm, that is, it can be used as an evaluation indicator of the target network search process performed by the above-mentioned method A or method B.

[0163] In the embodiment of the present application, after searching for the target network, the target network can be trained again to obtain an image processing network, so as to perform image processing using the image processing network. Specifically, the target network can be trained using multiple sample images to obtain an image processing network, and the sample images carry classification labels or object recognition labels.

[0164] In this embodiment, the prediction accuracy of the image processing network on the test set can be determined, and then the prediction accuracy is used as an evaluation index for the target network search process performed by the above-mentioned method A or method B. Specifically, if the prediction accuracy is high, it means that a target network with a high accuracy can be searched through method A or method B, so the efficiency of the above-mentioned target network search can be fed back through the evaluation index.

[0165] Among them, multiple sample images can be used for the same image processing task. According to actual training requirements, each sample image can carry a classification label or an object recognition label. The classification label can be a category that represents the object, and the recognition label can represent the attributes of the object to be recognized.

[0166] Image processing tasks can be face recognition tasks, image classification tasks, attribute recognition tasks, fingerprint recognition tasks, iris recognition tasks, etc. Accordingly, for face recognition tasks, sample images can be face images; for attribute recognition tasks, sample images can be sample images with different attributes; for fingerprint recognition tasks, sample images can be fingerprint images; for iris recognition tasks, sample images can be iris images of eyes.

[0167] In this embodiment, for different image processing tasks, the target network can be trained according to corresponding related technologies to obtain an image processing network, wherein the structure of the obtained image processing network is consistent with the structure of the target network.

[0168] In a specific implementation, when the target network is trained using sample images as training samples, the target network at the end of the training can be determined as the image processing network. In practice, the training can be considered to be completed when the accuracy of image recognition reaches a preset accuracy, or when the target network converges, and then the target network at the end of the training can be determined as the image processing network.

[0169] In this way, an image processing network is obtained, and then the image processing network can be used to perform corresponding image processing. Specifically, an image to be processed can be obtained, the image to be processed is input into the image processing network, and a processing result output by the image processing network is obtained, and then the processing result is a classification result of the image to be processed or a recognition result of an object contained in the image to be processed.

[0170] The image to be processed may be a face image, a fingerprint image, or an image taken of a specific object. Specifically, the image to be processed may be associated with an image processing task to be performed, for example, if face recognition is performed, the image to be processed is a face image.

[0171] In this embodiment, the image to be processed can be input into the input end of the image processing network to obtain the processing result output by the image processing network, wherein the processing result is associated with the image processing task.

[0172] Among them, if the image processing task is to identify the object in the image to determine whether the object is a specified object, the processing result is the recognition result of the object contained in the image to be processed. If the image recognition task is to classify the object in the image to determine the category to which the object belongs, the processing result is the classification result of the object contained in the image to be processed.

[0173] Of course, actual image processing tasks are not limited to the above-mentioned recognition tasks, and may also be other image processing tasks, such as object classification tasks.

[0174] The method for determining the new energy parameter value of the network by using the embodiment of the present application has at least the following advantages:

[0175] 1. While ensuring the efficiency of network search, the evaluation accuracy of the searched target network can be improved.

[0176] In this embodiment, since the parameters of the same network module can be shared by multiple sub-networks, that is, the parameters of a network module are updated by multiple sub-networks, each sub-network including the network module can quickly reach a convergence state, thereby improving the network search efficiency. On this basis, since the cumulative gradient values ​​of the same network module updated by multiple sub-networks are recorded, the process of each sub-network updating the parameters of the network module can be obtained, and thus, the performance parameter value of the sub-network can be determined according to the change of the cumulative gradient values ​​of each network module included in each sub-network, which can reflect the optimization goal and optimization method of the sub-network during the training process, so that the performance parameter value can be used as an evaluation indicator to accurately evaluate the true performance of the sub-network, and the performance parameter value can be used as a network search factor to improve the accuracy of the search for the sub-network.

[0177] 2. Flexibly use two methods to perform network searches to meet the needs of searching the target network.

[0178] This embodiment uses two methods to search the target network. One is to directly use the original multiple sub-networks as the search space, and continuously narrow the search space, that is, continuously reduce the number of sub-networks screened out, so as to obtain the target network. In this search method, the network structure of the searched target network is still the network structure before training, which can meet the user's demand for the stability of the network structure of the target network. The other is to use part of the sub-network as the initialization population. During the search process, the screened sub-networks are continuously cross-mutated to obtain a new population. In this way, the variability of the network structure of the sub-network can be improved, making the sub-network flexible and changeable, so as to meet the user's demand for automatic variation of the network structure during the network search process.

[0179] It should be noted that the method for determining the performance parameter value of the network described in the above embodiment of the present application can be executed by a computer, and the above network module and the original network can be considered as neural network data to be processed. Then, the computer can process the neural network data according to the performance parameter value of the network determined as described above to determine the performance parameter value of the sub-network.

[0180] It should be noted that, for the method embodiments, for the sake of simplicity, they are all expressed as a series of action combinations, but those skilled in the art should be aware that the embodiments of the present application are not limited by the described order of actions, because according to the embodiments of the present application, certain steps can be performed in other orders or simultaneously. Secondly, those skilled in the art should also be aware that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily required by the embodiments of the present application.

[0181] Reference Figure 8 As shown, a structural block diagram of a device for determining a performance parameter value of a network according to an embodiment of the present application is shown, such as Figure 8 As shown, the device may specifically include the following modules:

[0182] A sub-network determination module 801 is used to combine multiple network modules with the original network respectively to obtain multiple sub-networks of the original network, wherein one network module is used to implement one type of data processing operation;

[0183] A first training module 802 is used to alternately train the multiple sub-networks for multiple times to complete the training of the original network, wherein the sub-networks of the same network module share the parameters of the network module;

[0184] The parameter value determination module 803 is used to determine the performance parameter value of any sub-network among the multiple sub-networks according to the accumulated gradient values ​​of each network module included in the sub-network during the training process of the original network.

[0185] Optionally, the device may further include the following modules:

[0186] A statistical module, used for, during the training process of the original network, performing statistics on the accumulated gradient values ​​of each of the plurality of network modules during the training process of the original network, to obtain accumulated gradient value sequences corresponding to each of the plurality of network modules;

[0187] The parameter value determination module 803 may specifically include the following units:

[0188] A matrix obtaining unit, used to take the cumulative gradient value sequence corresponding to each network module included in the sub-network as a row to obtain the cumulative gradient value matrix of the sub-network;

[0189] The processing unit is used to process the accumulated gradient value matrix of the sub-network to obtain the performance parameter value of the sub-network.

[0190] Optionally, during the training process of the original network, the subnetwork targeted at each training is any subnetwork among the multiple subnetworks; the device may further include the following modules:

[0191] The second training module is used to use the sub-network targeted by each training as the target sub-network, and train the target sub-network using the training samples to obtain the gradient values ​​of each network module included in the target sub-network;

[0192] A gradient updating module, configured to update the accumulated gradient values ​​of the respective network modules in the target sub-network by using the respective gradient values ​​of the respective network modules included in the target sub-network;

[0193] The statistical module is specifically used to arrange the cumulative gradient values ​​after each update of the network module in sequence according to the update time during the training process of the original network, so as to obtain the cumulative gradient value sequences corresponding to each of the multiple network modules.

[0194] Optionally, the device further includes a combining module, which is specifically configured to perform the following steps for at least one group of edges connected to adjacent original network modules in the original network:

[0195] Configuring candidate operands corresponding to the plurality of network modules respectively for the edge;

[0196] Sampling a target operand of the edge from the candidate operands respectively corresponding to the plurality of network modules;

[0197] A network module corresponding to the target candidate operand is added between a group of adjacent original network modules connected by the edge.

[0198] Optionally, the device further includes the following modules:

[0199] A search space determination module, configured to determine at least some of the multiple subnetworks as search spaces according to the performance parameter values ​​of the multiple subnetworks;

[0200] A search module, used to search the search space with the maximum performance parameter value as the search target, to obtain the search space for the next search;

[0201] The target network determination module is used to determine the searched network as the target network when a preset search end condition is met.

[0202] Optionally, the search space determination module may be specifically configured to determine a first preset number of subnetworks with top ranking performance parameter values ​​among the multiple subnetworks as the search space;

[0203] The search module may specifically include the following units:

[0204] A first determining unit, configured to select a second preset number of sub-networks with top ranking performance parameter values ​​from the search space;

[0205] The second determining unit is configured to determine the second preset number of sub-networks as a search space for a next search.

[0206] Optionally, the device further comprises the following modules:

[0207] A first population determination module is used to use at least some of the multiple subnetworks as an initialization population and determine a performance parameter value of each subnetwork in the initialization population;

[0208] A screening module is used to screen K sub-networks with top ranking performance parameter values ​​from the initialization population;

[0209] A second population determination module is used to perform crossover mutation on the K sub-networks to obtain a new population, and determine the performance parameter value of each new network in the new population;

[0210] The repeated execution module is used to perform the following steps multiple times: selecting K new networks with top performance parameter values ​​from the new population, performing crossover mutation on the K new networks to obtain a new population, and determining the performance parameter values ​​of each new network in the new population to obtain a target network.

[0211] Optionally, the device further comprises the following modules:

[0212] The third training module is used to train the target network using multiple sample images to obtain an image processing network, wherein the sample images carry classification labels or object recognition labels.

[0213] Optionally, an embodiment of the present application further provides an image processing device, the image processing device comprising the following modules:

[0214] An image acquisition module, used for acquiring an image to be processed;

[0215] The input module is used to input the image to be processed into the image processing network to obtain the processing result output by the image processing network; the processing result is the classification result of the image to be processed or the recognition result of the object contained in the image to be processed.

[0216] It should be noted that the device embodiment is similar to the method embodiment, so the description is relatively simple, and the relevant parts can be referred to the method embodiment.

[0217] An embodiment of the present application also provides an electronic device that can be used to execute a method for determining a performance parameter value of a network. The electronic device may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the method for determining a performance parameter value of a network.

[0218] The embodiment of the present application also provides a computer-readable storage medium, which stores a computer program that enables a processor to execute the method for determining a performance parameter value of a network as described in the embodiment of the present application.

[0219] The various embodiments in this specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be referenced to each other.

[0220] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, devices, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0221] The embodiments of the present application are described with reference to the flowcharts and / or block diagrams of the methods, terminal devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing terminal device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing terminal device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0222] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing terminal device to operate in a specific manner, so that the instructions stored in the computer readable memory produce a manufactured product including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0223] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal device so that a series of operating steps are executed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions for executing on the computer or other programmable terminal device to implement the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0224] Although the preferred embodiments of the present application have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the embodiments of the present application.

[0225] Finally, it should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or terminal device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or terminal device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or terminal device including the elements.

[0226] The above is a detailed introduction to a method, device, equipment and storage medium for determining the performance parameter value of a network provided by the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the method of the present application and its core idea; at the same time, for general technical personnel in this field, according to the idea of ​​the present application, there will be changes in the specific implementation method and application scope. In summary, the content of this specification should not be understood as a limitation on the present application.

Claims

1. A method for determining a performance parameter value of a network, It is characterized in that The method comprises: Combining a plurality of network modules with the original network respectively to obtain a plurality of sub-networks of the original network, wherein one network module is used to implement one type of data processing operation; The plurality of sub-networks are alternately trained multiple times to complete the training of the original network, wherein the parameters of the network module are shared among the sub-networks including the same network module; For each of the network modules, the accumulated gradient value obtained after each updating of the gradient value of the network module during the alternating training process is recorded, and the accumulated gradient values ​​after each update are arranged in sequence according to the order of the updating time, so as to obtain the accumulated gradient value sequences corresponding to the plurality of network modules respectively; For any sub-network among the multiple sub-networks, taking the accumulated gradient value sequence corresponding to each network module included in the sub-network as a row to obtain an accumulated gradient value matrix of the sub-network; Processing the accumulated gradient value matrix of the sub-network to determine the performance parameter value of the sub-network; The method further comprises: Searching for a target network according to the performance parameter values ​​of the multiple sub-networks; Training the target network using a plurality of sample images to obtain an image processing network, wherein the sample images carry classification labels or object recognition labels; After obtaining the image processing network, the method further includes: Obtaining an image to be processed; Inputting the image to be processed into the image processing network to obtain a processing result output by the image processing network; The processing result is a classification result of the image to be processed or a recognition result of an object contained in the image to be processed.

2. The method according to claim 1, It is characterized in that During the training process of the original network, the subnetwork targeted at each training is any subnetwork among the multiple subnetworks; the method further includes: Take the subnetwork targeted for each training as the target subnetwork and perform the following steps: Using the training samples, the target sub-network is trained to obtain the gradient values ​​of the respective network modules included in the target sub-network; Using the gradient values ​​of the respective network modules included in the target sub-network, the accumulated gradient values ​​of the respective network modules belonging to the target sub-network among the multiple network modules are updated.

3. The method according to any one of claims 1-2, It is characterized in that For at least one group of edges connected to adjacent original network modules in the original network, perform the following steps: Configuring candidate operands corresponding to the plurality of network modules respectively for the edge; Sampling a target candidate operand of the edge from the candidate operands respectively corresponding to the plurality of network modules; A network module corresponding to the target candidate operand is added between a group of adjacent original network modules connected by the edge.

4. The method according to any one of claims 1-2, It is characterized in that After determining the performance parameter values ​​of each of the plurality of sub-networks, the method further includes: Determining at least some of the multiple subnetworks as search spaces according to respective performance parameter values ​​of the multiple subnetworks; Taking the maximum performance parameter value as the search target, searching the search space to obtain the search space for the next search; When a preset search end condition is met, the searched network is determined as the target network.

5. The method according to claim 4, It is characterized in that Determining at least some of the multiple subnetworks as search spaces according to the respective performance parameter values ​​of the multiple subnetworks includes: Determine a first preset number of subnetworks with top performance parameter values ​​among the multiple subnetworks as a search space; The search space is searched with the maximum performance parameter value as the search target to obtain the search space for the next search, including: Selecting a second preset number of sub-networks with top ranking performance parameter values ​​from the search space; The second preset number of sub-networks is determined as the search space for the next search.

6. The method according to any one of claims 1-2, It is characterized in that After determining the performance parameter values ​​of each of the plurality of sub-networks, the method further includes: Using at least some of the multiple subnetworks as an initialization population, and determining a performance parameter value of each subnetwork in the initialization population; Select K sub-networks with top ranking performance parameter values ​​from the initialization population; Perform crossover mutation on the K sub-networks to obtain a new population, and determine the performance parameter value of each new network in the new population; The following steps are performed multiple times: K new networks with top ranking performance parameter values ​​are selected from the new population, cross-mutation is performed on the K new networks to obtain a new population, and the performance parameter values ​​of each new network in the new population are determined to obtain a target network.

7. A device for determining a value of a performance parameter of a network, It is characterized in that The device comprises: A sub-network determination module, used to combine multiple network modules with the original network respectively to obtain multiple sub-networks of the original network, wherein one network module is used to implement one type of data processing operation; A training module, used for alternately training the multiple sub-networks for multiple times to complete the training of the original network, wherein the sub-networks of the same network module share the parameters of the network module; A statistical module is used to record, for each of the network modules, the cumulative gradient value obtained after the gradient value of the parameters of the network module is updated during the alternating training process, and arrange the cumulative gradient values ​​after each update in order of the update time to obtain the cumulative gradient value sequences corresponding to the multiple network modules; A parameter value determination module is used to, for any sub-network among the multiple sub-networks, take the cumulative gradient value sequence corresponding to each network module included in the sub-network as a row to obtain a cumulative gradient value matrix of the sub-network; process the cumulative gradient value matrix of the sub-network to determine the performance parameter value of the sub-network; A target network determination module, configured to search for a target network according to the performance parameter values ​​of the plurality of sub-networks; A network training module, used to train the target network using a plurality of sample images to obtain an image processing network, wherein the sample images carry classification labels or object recognition labels; The device also includes: An image acquisition module, used for acquiring an image to be processed; An input module is used to input the image to be processed into the image processing network to obtain a processing result output by the image processing network; the processing result is a classification result of the image to be processed or a recognition result of an object contained in the image to be processed.

8. An electronic device, It is characterized in that The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor, wherein when the processor executes the method, the method for determining the performance parameter value of a network as claimed in any one of claims 1 to 6 is implemented.

9. A computer-readable storage medium, It is characterized in that The computer program stored therein enables the processor to execute the method for determining the performance parameter value of a network as described in any one of claims 1-6.