CN112836794A - Method, device and equipment for determining image neural architecture and storage medium - Google Patents

Method, device and equipment for determining image neural architecture and storage medium Download PDF

Info

Publication number
CN112836794A
CN112836794A CN202110106031.5A CN202110106031A CN112836794A CN 112836794 A CN112836794 A CN 112836794A CN 202110106031 A CN202110106031 A CN 202110106031A CN 112836794 A CN112836794 A CN 112836794A
Authority
CN
China
Prior art keywords
preset
individuals
current
target
image neural
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202110106031.5A
Other languages
Chinese (zh)
Other versions
CN112836794B (en
Inventor
骆剑平
蔡榕鸿
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Shenzhen University
Original Assignee
Shenzhen University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Shenzhen University filed Critical Shenzhen University
Priority to CN202110106031.5A priority Critical patent/CN112836794B/en
Publication of CN112836794A publication Critical patent/CN112836794A/en
Application granted granted Critical
Publication of CN112836794B publication Critical patent/CN112836794B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/086—Learning methods using evolutionary algorithms, e.g. genetic algorithms or genetic programming

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Evolutionary Biology (AREA)
  • Physiology (AREA)
  • Image Analysis (AREA)

Abstract

The embodiment of the invention discloses a method, a device, equipment and a storage medium for determining an image neural architecture. Wherein, the method comprises the following steps: performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; inputting the current child set into a preset multitask learning agent model to generate an updated target child set; adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement; and judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set. The method and the device realize efficient determination of the image neural framework and effectively save determination time.

Description

Method, device and equipment for determining image neural architecture and storage medium
Technical Field
The embodiment of the invention relates to computer technology, in particular to a method, a device, equipment and a storage medium for determining an image neural architecture.
Background
The deep convolutional neural network in deep learning has a wide prospect in solving various practical problems, but the network can achieve the best performance only when the architecture of the deep convolutional neural network is optimally constructed. If the data to be processed changes, a new neural network architecture must be redesigned.
Deep convolutional neural networks are typically designed manually, requiring designers to have a wealth of expertise in both neural networks and data research. However, the process of manual design is a continuous trial and error and time-consuming process, most of common users with insufficient experience are difficult to individually construct the neural network required by themselves in practical application, and the determination efficiency of the image neural architecture is low.
Disclosure of Invention
The embodiment of the invention provides a method, a device and equipment for determining an image neural architecture and a storage medium, which are used for improving the determination efficiency of the image neural architecture.
In a first aspect, an embodiment of the present invention provides a method for determining an image neural architecture, where the method includes:
performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement;
and judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
In a second aspect, an embodiment of the present invention further provides an apparatus for determining an image neural architecture, where the apparatus includes:
the current offspring set generation module is used for carrying out mutation operation on the parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
the target child set generation module is used for inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
the target population generation module is used for adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set and generating a target population based on the current population and a preset fitness requirement;
and the neural framework set determining module is used for judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
In a third aspect, an embodiment of the present invention further provides an apparatus for determining an image neural architecture, including a memory, a processor, and a computer program stored in the memory and executable on the processor, where the processor implements a method for determining an image neural architecture according to any embodiment of the present invention when executing the program.
In a fourth aspect, embodiments of the present invention further provide a storage medium containing computer-executable instructions, which when executed by a computer processor, are configured to perform the method for determining an image neural architecture according to any of the embodiments of the present invention.
The embodiment of the invention obtains the varied current offspring set by varying the parent individuals in the population, and the offspring individuals of the current offspring set only have the structural information and the network scale of the image neural architecture but have no network performance. Through mutation operation, the possibility of the image neural architecture is increased, and the determination precision of the image neural architecture is improved. And inputting the data of the current offspring set into a pre-trained multi-task learning agent model, and predicting the network performance of the offspring individuals so as to obtain the information of the complete offspring individuals and generate a target offspring set. The problem of among the prior art, through the image classification task data set that predetermines, consume the long time to calculate network performance is solved, effective saving image neural framework confirms time. Through the fitness ranking of the individuals in the target population, the individual with the lowest fitness can be found, and the individual is combined with the current population to generate the target population. If the iteration is finished, the pareto optimal set of the image neural framework can be searched from the target population, so that the target neural framework can be determined from the pareto optimal set conveniently, and the determination efficiency of the image neural framework is effectively improved.
Drawings
Fig. 1 is a flowchart illustrating a method for determining an image neural architecture according to a first embodiment of the present invention;
FIG. 2 is a schematic structural diagram of a multi-task learning agent model according to one embodiment of the present invention;
FIG. 3 is a flowchart illustrating a method for determining an image neural architecture according to a second embodiment of the present invention;
FIG. 4 is a block diagram of an apparatus for determining neural architecture of an image according to a third embodiment of the present invention;
fig. 5 is a schematic structural diagram of an image neural architecture determining apparatus in the fourth embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.
Example one
Fig. 1 is a flowchart illustrating a method for determining an image neural architecture according to an embodiment of the present invention, where the method is applicable to automatically determining an image neural architecture, and the method can be executed by a device for determining an image neural architecture. As shown in fig. 1, the method specifically includes the following steps:
110, performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; and the child individuals of the current child set comprise the structural information and the network scale of the image neural architecture.
The current population is a population of the image neural architecture, and the current population may include a plurality of image neural architecture individuals. And during the first iteration, the current population is the initial population, and for the number of iterations after the first iteration, the current population is the target population generated by the last iteration. The genetic operator may include selection, intersection, mutation, and the like, and in this embodiment, the predetermined genetic operator is a mutation operator, and may be used for performing mutation operation. The parent individuals are image neural architecture individuals with better performance in the current population, and one individual can be selected from the current population as a parent individual. And carrying out mutation operation on the parent individuals according to a preset genetic operator to generate a plurality of filial individuals with different genotypes, wherein the generated filial individuals form a current filial collection, and each filial individual represents an image neural architecture.
The single individual in the current population comprises three parts of information, namely structural information of the image neural architecture, network performance of the image neural architecture and network scale of the image neural architecture. The structural information of the image neural architecture represents the genotype of the image neural architecture, and is described by a character string. The genotype composition comprises functional layers commonly used by the neural framework of the convolution image and parameter configuration corresponding to the functional layers. Common functional layers may include two-dimensional convolutional layers, dropout layers, two-dimensional max pooling layers, full connectivity layers, and the like, and dropout layers may be used to reduce and improve generalization. The parameter configuration of the functional layer is specified as follows: for a two-dimensional convolution layer, the character c may be used to represent the size of the convolution kernel3 × 3, the number of convolution kernels is an integer, and the activation function of the convolution layer is fixed as a ReLU (Rectified Linear Unit); for dropout layers, which may be represented by the character d, dropout ratio ranges from 0 to 1; for the two-dimensional maximum pooling layer, which may be represented by the character p, the size of the pooling kernel is fixed to 2 x 2; for the fully connected layer, the activation function is fixed to softmax (normalized exponential function), which can be represented by the character fc. One individual I in the current populationiGenotype G ofiThe following may be exemplified:
Gi＝['c',87,'c',326,'p','c',...,'p','d',0.14,'c',...,'p',...,'fc',10]；
wherein, IiRepresenting the i-th image of the individual with neural architecture, genotype GiCan be uniquely mapped into corresponding individuals IiI.e. the corresponding convolutional image neural architecture, the first layer c of the image neural architecture is the convolutional layer, the number of convolutional kernels is 87, the default convolutional kernel size is 3 × 3, and the default activation function is ReLU; the next layer c is convolution layer, the number of convolution kernels is 326, the default convolution kernel size is 3 x 3, and the default activation function is ReLU; the next layer p is the largest pooling layer, with the default pooling core size of 2 x 2; by analogy, d represents the dropout layer, the floating point number 0.14 after d represents the ratio of dropout; the last layer fc is the fully connected layer, and the integer after fc represents the output dimension of the image neural architecture.
The network scale of the image neural architecture can be rapidly calculated according to the structural information of the image neural architecture, belongs to an extremely cheap optimization target, and can be obtained by calculating the parameter configuration of a common functional layer. For the convolutional image neural architecture, the parameter configuration amount of each convolutional layer is calculated as follows:
convparameters＝C0×(kw×kh×Ci+1)；
wherein, convparametersRepresenting the quantity of configuration of the convolutional layer, C0Representing the number of output channels of the convolutional layer, i.e. the number of convolutional kernels, CiRepresenting the number of input channels, k, of the convolutional layerwRepresenting the width, k, of the convolution kernelhHigh, 1 table representing convolution kernelThe convolution kernel is shown as biased. Usually, there is kw＝khK is a constant. Therefore, the parameter configuration amount of the convolutional layer can be calculated as follows:
convparameters＝C0×(k2×Ci+1)；
for the parameter configuration quantity of the full connection layer, the parameter configuration quantity can be calculated by the following formula:
denseparameters＝(I+1)×O＝I×O+O；
wherein denseparametersThe parameter configuration quantity of the full connection layer is obtained, I is the size of the column vector input by the full connection layer, O is the size of the column vector output, and 1 is offset. According to the genotype G of the individuals in the populationiAnd calculating and summing the parameter configuration quantity of each functional layer, so that the network scale of the image neural architecture can be quickly calculated.
The network performance of the image neural architecture can be obtained only after training on a CIFAR data set, which belongs to a very expensive optimization target, and the CIFAR data set is an existing picture data set. In the prior art, when the image neural architecture is determined, the network performance of each image neural architecture needs to be determined through the CIFAR data set, and the determination efficiency of the image neural architecture is influenced. In order to improve the determination efficiency of the image neural architecture, the network performance of the child individuals in the current child set does not need to be determined. Therefore, the offspring individuals of the current offspring set include the structural information and the network scale of the image neural architecture.
In this embodiment, optionally, before performing mutation operation on the parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals, the method further includes: adopting a preset multi-objective optimization algorithm to carry out fitness ranking on the image neural framework individuals in the current population to obtain a fitness ranking result; obtaining at least one parent individual according to the fitness sorting result and preset selection operation; the image neural architecture individuals of the current population comprise structure information, network performance and network scale of the image neural architecture.
Specifically, before generating the current child set, the parent individuals need to be determined. A multi-objective optimization algorithm is preset, for example, NSGA-II (multi-objective genetic algorithm), namely a fast non-dominated sorting genetic algorithm with an elite strategy. And calculating the fitness value of each image neural framework individual in the current population by adopting a preset multi-objective optimization algorithm, and sequencing the fitness values to obtain a fitness sequencing result. The smaller the fitness value, the more excellent the image neural architecture of the individual is. And selecting at least one parent individual from the individuals according to the fitness ranking result and a preset selection operation, wherein the selection operation can be a tournament method and the like. And in each iteration, the structure information, the network performance and the network scale of the image neural architecture individuals in the current population can be obtained. The structural information and the network scale of the image neural framework can be quickly obtained, the network performance of the image neural framework only needs to be determined by the initial population during initial iteration, and can be directly determined by the result of the last iteration in the iteration times after the first iteration, so that the network performance is prevented from being calculated in each iteration, the iteration time of the image neural framework is greatly shortened, and the determination efficiency of the image neural framework is improved. In addition, by determining the parent individuals, a better image neural architecture can be obtained, and the determination accuracy of the image neural architecture is favorably improved.
In this embodiment, optionally, the current population is an initial population of the first iteration; correspondingly, before the fitness ranking of the image neural architecture individuals in the current population is performed by adopting a preset multi-objective optimization algorithm and a fitness ranking result is obtained, the method further comprises the following steps: training the image neural architecture individuals in the initial population according to a preset image classification task data set to obtain the network performance of the image neural architecture individuals; and determining the network scale of the image neural architecture individuals according to the structure information of the image neural architecture individuals in the initial population.
Specifically, if the current population is the initial population, training the image neural architecture individuals in the initial population according to a preset image classification task data set to obtain the network performance of the image neural architecture individuals, for example, the image classification task data set may be a CIFAR data set. For the current population after the first iteration, a large amount of time is not needed to be spent on training to obtain the network performance of the image neural architecture individual. And calculating to obtain the network scale of the image neural architecture individuals according to the structure information of the image neural architecture individuals in the initial population, namely according to the configuration parameters of each functional layer and each functional layer of the image neural architecture. In the iteration times after the first iteration, the network scale of each filial generation individual can be quickly calculated according to the structure information of the image neural architecture individual. By determining the network performance and the network scale of the individuals in the initial population, the subsequent iteration is facilitated, and the subsequent iteration efficiency is improved.
In this embodiment, optionally, performing mutation operation on the parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals, including: and performing mutation operations of increasing genes, reducing genes and/or replacing genes on the parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals.
Specifically, a plurality of mutation operations may be performed on the parent individuals through a predetermined genetic operator, for example, the mutation operations may include adding genes, reducing genes, and/or replacing genes. The specific operation of adding genes is to add genes at a random position in the genotype of a parent, namely adding a common functional layer, and performing specific parameter configuration on the newly added functional layer. When mutation manipulation for increasing genes is employed, it is noted that genes are not added at the ends of the parent genotypes. The specific operation of reducing the gene is to reduce the gene at a random position of the genotype of the parent, namely deleting the common functional layer and deleting the parameter configuration corresponding to the functional layer. When mutation reduction is used, care should be taken not to delete the first and last genes of the parent genotype. For the image classification task, the first layer of the image neural framework must be a convolutional layer, and other common layers are not used as the first layer, so that the first end and the last end of the genotype are not operated when the genes are reduced. The specific operation of replacing the gene is to replace the gene at a random position in the genotype of the parent with a new gene generated randomly, wherein the replacement comprises the replacement of a common functional layer and the replacement of the parameter configuration of the corresponding layer. Similar to the addition of genes, genes were not replaced at the end of the parent genotype. When the gene to be replaced is the first gene of the genotype, only the parameter configuration thereof can be replaced, since the first layer must be fixed as a convolutional layer. The method has the advantages of increasing the diversity of genotypes, reducing the possibility of local convergence and improving the determination precision of the image neural architecture.
Step 120, inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of the image neural architecture, a network performance prediction result and a network scale.
The multi-task learning agent model may be a pre-trained model based on RBFNN (Radial Basis Function Neural Network), and may predict Network performance of a descendant individual. And inputting the current offspring set serving as input data into the multitask learning agent model, combining the output result with information of the offspring in the current offspring set, and updating the current offspring set to obtain a target offspring set. The child information in the current child set comprises the structural information of the image neural architecture and the network scale, and the child individual information of the target child set comprises the structural information of the image neural architecture, the network performance prediction result and the network scale.
In this embodiment, optionally, the inputting the current child set into a preset multitask learning agent model to generate an updated target child set includes: inputting the current offspring set into a preset multitask learning agent model, and outputting the image neural architecture network performance prediction result of the offspring individuals in the current offspring set; and obtaining an updated target child set according to the structure information, the network scale and the network performance prediction result of the child individuals in the current child set.
Specifically, after the training of the multi-task learning agent model is completed, the multi-task learning agent model can be used for predicting the network performance of the image neural architecture of the offspring individuals. The filial individuals generated by the mutation of the parent individuals only have genotype information, namely only the structural information of the image neural architecture is known, and the network scale of the image neural architecture can be quickly calculated according to the structural information of the image neural architecture. For the network performance of the image neural architecture of the offspring individuals, the trained multi-task learning agent model directly uses the genotype G of the individualiThe prediction of the network performance is made to quickly obtain the network performance of the image neural framework, and the real training process of the image neural framework on a CIFAR data set is replaced, so that the offspring individuals IiIs updated to (I)i:Gi,MTSM(Gi),Oj2)。MTSM(Gi) Represents an individual IiNetwork performance prediction result of, Oj2Represents an individual IiThe network size of (a). The multi-task learning agent model can be composed of a plurality of radial basis function neural networks, one radial basis function neural network can solve one picture classification task, namely the multi-task learning agent model can simultaneously solve a plurality of picture classification tasks, for example, input data of two groups of tasks can be simultaneously input into the multi-task learning agent model, and output data of the two groups of tasks can be obtained. The multi-task learning agent model may include an input layer, a hidden layer, an interaction layer, and an output layer, and each radial basis function neural network may include an input layer, a hidden layer, and an output layer. Through the joint learning of a plurality of tasks, the problem that image neural architecture search cannot be directly carried out due to the fact that the difficulty of a data set is too large and the data volume of the data set is insufficient in an actual scene can be effectively solved. And when facing a plurality of similar tasks, the similarity between the tasks can be processed and utilized simultaneously. The similarity task can be that two tasks are to classify pictures, wherein the task is a simple ten-class classification task, and the task is a complex hundred-class classification task. Joint learning of multi-task knowledge is more favorable for searching optimal image neural frameAnd multi-task joint search is superior to single-task search.
FIG. 2 is a schematic diagram of a multi-task learning agent model. The multi-task learning agent model shown in fig. 2 can simultaneously solve two image classification tasks, wherein the task 1 and the task 2 correspond to a radial basis network respectively, the radial basis network comprises an input layer, a hidden layer and an output layer, and the two radial basis networks are connected through an interaction layer to realize the interaction of knowledge between the two tasks. The multi-task learning agent model comprises 4 trained parameters which are respectively the central point c of the hidden layer Gaussian kernel functioniThe method comprises the following steps of obtaining a width parameter sigma of a Gaussian kernel function of a hidden layer, a weight omega of each central point in the hidden layer and a correlation parameter L of an interaction layer. x is input data and y is output data. Before inputting the input data of each task, the genotype of the offspring individuals is encoded again, and a set X of data to be predicted is constructedpredictAs a collection of input data. The image neural architecture information of the individual is represented by genotypes, different image neural architectures have different genotypes, and the genotypes are represented in a character string mode and are composed of characters. The length of the character string is determined by the number of functional layers of the image neural architecture, and the content of the character string is determined by the sequence of the functional layers and the parameter configuration of the functional layers. Thus, the genotypes of different individuals within a population vary in length and content. The genotype needs to be re-encoded before it can be input as data to be predicted. In this embodiment, a 12-dimensional array is defined to re-encode each individual genotype, and genotypes with different lengths and contents are mapped onto the 12-dimensional array with a fixed length, where the 12-dimensional array includes information such as the total number of functional layers, the number of convolutional layers, the number of pooling layers, and the number of dropout layers of the image neural architecture, which are represented by the 1 st to 4 th dimensions; the 5 th to 7 th dimensions represent the maximum convolution kernel number, the minimum convolution kernel number and the average convolution kernel number in the parameter configuration of the convolution layer; dimensions 8 to 10 represent the maximum dropout ratio, the minimum dropout ratio and the average dropout ratio in the parameter configuration of the dropout layer; the 11 th dimension represents the network scale of the image neural architecture; the 12 th dimension represents the network performance of the image neural architecture. Due to different orders of magnitude of parameter configuration of different functional layers, the method can be used forThe data normalization of the statistical data of each dimension is performed, and the effect is to map the value of each dimension between 0 and 1. For the descendant individuals to be predicted, there is no 12 th dimension data, so the input data is 11 dimensions, and the data X to be predictedpredictAnd (4) showing. After determining the input data, the RBFNN input layer to hidden layer mapping may be calculated according to the following formula:
Figure BDA0002917691620000121
wherein m is the number of the center points of the hidden layers,
Figure BDA0002917691620000124
is a Gaussian kernel function and represents the mapping of the input layer of the jth central point of the ith task to the hidden layer.
According to the trained weight parameter omega, the output of the hidden layer in the RBFNN can be calculated, and the output of the hidden layer can be calculated through the following formula:
Figure BDA0002917691620000122
wherein h isiHidden layer output, ω, representing the ith taskijAnd representing the weight parameter of the jth central point of the ith task.
According to the trained correlation parameter L, calculating output data corresponding to each task through the following formula:
Figure BDA0002917691620000123
wherein, yiOutput data representing the ith task, LiqAnd a relevance parameter representing data interaction of the ith task and the qth task. Using the activation function for the full connection layer of the interaction layer as the final output value of the radial basis function neural network, and finishing the prediction of the offspring individuals to obtain (I)i:Gi,MTSM(Gi),Oj2)。
The radial basis function neural network has good local approximation characteristics, and has the advantages of simplicity in training and high learning speed. Therefore, the embodiment of the invention constructs a multi-task learning agent model based on the radial basis function neural network. Macroscopically, the multitask agent model is formed by combining 2 radial basis function neural networks, and the multitask agent model has a structure with 4 layers including an input layer, a hidden layer, an interaction layer and an output layer. In the embodiment of the invention, the number of tasks in a multitask scene is set to be 2, so that the number of radial basis function neural networks forming the multitask agent model is 2, and each RBFNN corresponds to a single task. It is noted that the input to each RBFNN is the same, taking as input the training data of all tasks, but the output of each RBFNN is different, with its output corresponding to a single task. Dimension n of input data of the input layer is data X to be predictedpredictI.e. dimension n is 11; the number of the central points of the hidden layers of the single RBFNN is m, and the central points are predefined by a user; and the parameter omega behind the hidden layer corresponds to a weight vector from the hidden layer to the output layer in a single RBFNN, and is obtained after the multi-task agent model is learned and trained. For the multi-task learning agent model, besides sharing data of all tasks at an input layer, in order to enable knowledge learned by a plurality of RBFNNs to be interacted fully, an interaction layer is added between a hidden layer and an output layer, and the interaction layer enables outputs of the RBFNNs to be combined in a full connection mode. The parameter L of the interaction layer is a correlation weight parameter and represents the knowledge contribution rate of a single task, and a vector formed by the parameter L reflects the similarity among multiple tasks. After the fully-connected layer of the interaction layer, a sigmod (S-shaped growth curve function) activation function can be adopted, the output of the sigmod function is mapped to a (0,1) interval, the sigmod function is nonlinear change, the nonlinear mapping capability of the multi-task learning agent model is further improved, and the generalization of the model is enhanced. The multi-task learning agent model enables a multi-task network to more effectively utilize the correlation among a plurality of tasks by sharing the data of all the tasks and the knowledge of an interaction layer, can simultaneously model a plurality of similar tasks, and is also beneficial to the joint learning of multi-task knowledgeThe method improves the capability of the multi-task learning agent model for predicting the network performance of the unknown neural architecture. Through the multi-task learning agent model, the determination time of network performance is effectively saved, and multi-task data interaction is realized. The performance of the image neural framework is regressed by utilizing the radial basis network, so that the network performance of the image neural framework is directly predicted, a large amount of network training time is saved, a larger search space is searched under the condition of limited resources and cost, and the determination efficiency of the image neural framework is improved.
And step 130, performing fitness sequencing on the offspring individuals in the target offspring set by adopting a preset multi-objective optimization algorithm, and generating a target population based on the current population and a preset fitness requirement.
After the target offspring set is obtained, the offspring individuals in the target offspring set are subjected to fitness ranking by using a preset multi-objective optimization algorithm, for example, an NSGA-II algorithm may be adopted. A fitness requirement is preset, for example, the fitness requirement may be to select a child individual with the lowest fitness value. And selecting the offspring individuals meeting the requirements according to the preset fitness requirement. And combining the offspring individuals with the current population to generate a target population.
In this embodiment, optionally, the method for generating the target population based on the current population and the preset fitness requirement by using a preset multi-objective optimization algorithm to rank the fitness of the child individuals in the target child set includes: adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and determining the target offspring individuals meeting the preset fitness requirement; determining the network performance of the target child individual according to the image neural architecture structure information of the target child individual and a preset image classification task data set; adding the target offspring individuals into the current population to generate a target population; the image neural architecture individuals of the target population comprise structure information, network performance and network scale of the image neural architecture.
Specifically, the fitness value of the descendant individuals in the target descendant set can be calculated based on a preset NSGA-II algorithm according to the network performance prediction result and the network scale of the descendant individuals in the target descendant set, and the fitness is ranked. And selecting the filial generation individuals meeting the preset fitness requirement from the target filial generation set as target filial generation individuals according to the sorting result. For example, the preset fitness requirement may be to select the descendant with the lowest fitness value as the target descendant. Training the image neural architecture of the target offspring individual according to a preset image classification task data set, and obtaining the image neural architecture network performance of the target offspring individual according to the structural information of the image neural architecture of the target offspring individual. The preset image classification task dataset may be a CIFAR dataset. The network structure prediction result is obtained through the multi-task learning agent model, the target child individual can be conveniently and rapidly obtained according to the network structure prediction result, the better individual can be found in the iteration process, and the determination precision of the image neural framework is improved. And performing expensive training of network performance through a CIFAR data set to obtain image neural architecture structure information, network performance and network scale of the target offspring individuals. By adopting the multi-task learning agent model, expensive training of network performance on all filial generations in the target filial generation set is avoided, and time is effectively saved, for example, 100 filial generations exist in the target filial generation set, a network performance prediction result is determined by adopting the multi-task learning agent model, one target filial generation is selected from the target filial generation set, only the network performance training on the target filial generation is needed, and the training on the 100 filial generations is not needed. And each individual in the current population also comprises image neural architecture structure information, network performance and network scale, target child individuals are added into the current population to generate a target population, and the image neural architecture individuals of the target population all comprise the image neural architecture structure information, the network performance and the network scale image neural architecture information. The method has the advantages that the complete image neural architecture information of the target population can be obtained, the final target image neural architecture can be obtained from the target population, the image neural architecture information of the image neural architecture individuals in the target population can be used in the next iteration, the network performance of the current population of each iteration is prevented from being determined, and the determination efficiency of the image neural architecture is improved.
And 140, judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, so that a user can determine the target neural framework from the pareto optimal set.
An iteration end condition is preset, for example, the iteration end condition may be that the iteration ends when the iteration number reaches the maximum iteration number. And after the target architecture is obtained, judging whether the current iteration frequency is the last iteration, namely the maximum iteration frequency, if so, determining that the iteration is finished, and determining the target neural architecture from the target population.
After the iteration is determined, a pareto optimal set of the image neural framework can be determined from the target population according to a preset multi-objective optimization algorithm, for example, the pareto optimal set can be determined according to an NSGA-II algorithm. After determining the pareto optimal set, the user may select a target neural architecture from the pareto optimal set according to actual needs.
In this embodiment, optionally, determining whether the current iteration number meets a preset iteration end condition, and if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, where the determining includes: if the current iteration times meet the preset iteration ending conditions, dividing the image neural framework individuals in the target population into non-dominant leading edge sets of at least two levels according to a preset multi-objective optimization algorithm, and determining the non-dominant leading edge sets meeting the preset level requirements as pareto optimal sets.
Specifically, after determining that the iteration is over, the NSGA-II can be adopted to rapidly sort the population, and after the NSGA-II is completed, the population is divided into non-dominated leading edge sets with different grades. A rank requirement is preset, and for example, a rank-optimal non-dominant leading edge set may be determined as a pareto-optimal set. The method is convenient to obtain a good image neural framework, and the determination precision of the image neural framework is improved.
In this embodiment, optionally, after determining whether the current iteration number meets a preset iteration end condition, the method further includes: and if the current iteration times do not meet the preset iteration ending condition, determining the target population as the current population of the next iteration, and adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the image neural architecture individuals in the current population of the next iteration to obtain new parent individuals.
Specifically, if it is determined that the current iteration number is not the maximum iteration number, the target population generated by the current iteration is used as the current population of the next iteration, and the determination of the image neural architecture is continued. The method comprises the steps of adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on image neural architecture individuals in a target population to obtain new parent individuals, and continuing to finish the next iteration process. The method has the advantages of realizing iterative automatic circulation, reducing user operation and improving the determination precision and the determination efficiency of the image neural framework.
According to the technical scheme, the parent individuals in the population are mutated to obtain the mutated current offspring set, and the offspring individuals of the current offspring set only have the structural information and the network scale of the image neural architecture but have no network performance. Through mutation operation, the possibility of the image neural architecture is increased, and the determination precision of the image neural architecture is improved. And inputting the data of the current offspring set into a pre-trained multi-task learning agent model, and predicting the network performance of the offspring individuals so as to obtain the information of the complete offspring individuals and generate a target offspring set. The problem of among the prior art, through the image classification task data set that predetermines, consume the long time to calculate network performance is solved, effective saving image neural framework confirms time. Through the fitness ranking of the individuals in the target population, the individual with the lowest fitness can be found, and the individual is combined with the current population to generate the target population. If the iteration is finished, the pareto optimal set of the image neural framework can be searched from the target population, so that the target image neural framework can be determined from the pareto optimal set conveniently, and the determination efficiency of the image neural framework is effectively improved.
Example two
Fig. 3 is a flowchart illustrating a method for determining an image neural architecture according to a second embodiment of the present invention, which is further optimized based on the second embodiment. As shown in fig. 3, the method specifically includes the following steps:
step 310, performing mutation operation on the parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; and the child individuals of the current child set comprise the structural information and the network scale of the image neural architecture.
And 320, generating training data of the image neural architecture individuals in the current population based on a preset data format according to the structure information, the network scale and the network performance of the image neural architecture individuals in the current population.
Wherein the multi-task learning agent model may be trained prior to each iteration using the multi-task learning agent model. The aim of training the multi-task learning agent model is to obtain four parameters to be trained of the model, namely a central point c of a hidden layer Gaussian kernel functioniThe method comprises the following steps of obtaining a width parameter sigma of a Gaussian kernel function of a hidden layer, a weight omega of each central point in the hidden layer and a correlation parameter L of an interaction layer. The multi-task learning agent model comprises an interaction layer and at least two radial basis function neural networks, wherein each radial basis function neural network comprises an input layer, a hidden layer and an output layer, and the hidden layer can comprise a plurality of central points.
A standard data format is set in advance, for example, a 12-dimensional data format is set. The image neural architecture information of individuals in the current population is represented by genotypes, different image neural architectures have different genotypes, and the genotypes are represented in a character string mode and are composed of characters. The length of the character string is determined by the number of functional layers of the image neural architecture, and the content of the character string is determined by the sequence of the functional layers and the parameter configuration of the functional layers. Thus, the genotypes of different individuals within a population vary in length and content. The genotype needs to be re-encoded before it can be input as training data. In this embodiment, a 12-dimensional array is defined to re-encode each individual genotype, and genotypes with different lengths and contents are mapped onto the 12-dimensional array with a fixed length, where the 12-dimensional array includes information such as the total number of functional layers, the number of convolutional layers, the number of pooling layers, and the number of dropout layers of the image neural architecture, which are represented by the 1 st to 4 th dimensions; the 5 th to 7 th dimensions represent the maximum convolution kernel number, the minimum convolution kernel number and the average convolution kernel number in the parameter configuration of the convolution layer; dimensions 8 to 10 represent the maximum dropout ratio, the minimum dropout ratio and the average dropout ratio in the parameter configuration of the dropout layer; the 11 th dimension represents the network scale of the image neural architecture; the 12 th dimension represents the network performance of the image neural architecture.
Recoding the genotype G of each individualiAnd mapping to data for training of the multi-task learning agent model. The requirement for genotype G statistics during recodingiIn the case of the number of various common functional layers and the parameter configuration of the functional layers, because the magnitude of the parameter configuration of different functional layers is different, data normalization needs to be performed on the statistical data of each dimension, and the value of each dimension is mapped to a range from 0 to 1, so that a training set of the multi-task agent model, namely training data, is formed. The first 11-dimensional data of the training set is XtrainAnd the last 1-dimensional data is Ytrain. The training data comprises network structure data and training labels, the network structure data comprises structure information and network scale, and the training labels comprise network performance, namely XtrainRepresenting network structure data, YtrainRepresenting a training label.
And 330, inputting the training data into the multi-task learning agent model to be trained, and determining the output value and the output layer error of the training label according to the output value of the output layer of the radial basis function neural network.
And inputting the training data into a multi-task learning agent model to be trained to obtain an output value of the radial basis function neural network output layer. If the training data of a plurality of image classification tasks exist, the training data of the plurality of tasks can be input into the multi-task learning agent model together, and the corresponding radial basis function neural network output layer outputs the output values of the tasks. And obtaining the output layer error of the task according to the output value of the task and the training label in the task training data. For example, the output value of one task and the corresponding training label may be subtracted to obtain the radial basis function neural network output layer error. If only one task exists, the error of the output layer of the radial basis function neural network of the task is the error of the output layer of the multi-task learning agent model; and if a plurality of tasks exist, adding the output layer errors of the radial basis function neural networks of the tasks to obtain the output layer errors of the multi-task learning agent model.
Step 340, judging whether the error of the output layer meets the requirement of a preset error threshold value; and if so, determining that the multi-task learning agent model is trained, inputting the current offspring set into a preset multi-task learning agent model, and generating an updated target offspring set.
After the error of the output layer of the multi-task learning agent model is obtained, the error of the output layer is compared with the error threshold, and whether the error of the output layer meets the requirement of the preset error threshold is judged. If the error of the output layer is smaller than the error threshold, the error of the output layer is determined to meet the requirement of the preset error threshold, the training of the multi-task learning agent model is completed, and the current offspring set can be input into the preset multi-task learning agent model for calculation.
In this embodiment, optionally, after determining whether the output layer error meets the preset error threshold requirement, the method further includes: if the error of the output layer does not meet the requirement of a preset error threshold, updating the multi-task learning agent model by adopting a preset gradient descent method; and inputting the training data into the updated multi-task learning agent model, and determining whether the multi-task learning agent model completes training according to the preset error threshold requirement.
If the error of the output layer is equal to or larger than the error threshold, the error of the output layer is determined not to meet the requirement of the preset error threshold, and the multi-task learning agent model needs to be trained continuously. And during the initial training of each iteration, the adopted four parameters to be trained are initialization parameters, the four parameters to be trained are updated after each training, and if the training is not finished, the updated parameters to be trained are adopted as the parameter values to be trained used in the next training. For example, four parameters to be trained can be updated by adopting a gradient descent method, and the updated parameters to be trained are substituted into the multi-task learning agent model to obtain the updated multi-task learning agent model. Inputting the training data into the updated multi-task learning agent model, then calculating the output value of the output layer, determining the error between the output value and the output layer of the training label until the error of the output layer meets the requirement of a preset error threshold value, and finishing the training. The method has the advantages that the multitask learning agent model can be repeatedly trained, the precision of the multitask learning agent model is improved, and the determining efficiency and precision of the image neural framework are further improved.
And 350, adopting a preset multi-objective optimization algorithm to perform fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement.
And 360, judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
The embodiment of the invention obtains the varied current offspring set by varying the parent individuals in the population, and the offspring individuals of the current offspring set only have the structural information and the network scale of the image neural architecture but have no network performance. Through mutation operation, the possibility of the image neural architecture is increased, and the determination precision of the image neural architecture is improved. And inputting the data of the current offspring set into a pre-trained multi-task learning agent model, and predicting the network performance of the offspring individuals so as to obtain the information of the complete offspring individuals and generate a target offspring set. The multi-task learning agent model can be trained for multiple times during each iteration, and the precision of the multi-task learning agent model is improved. The problem of among the prior art, through the image classification task data set that predetermines, consume the long time to calculate network performance is solved, effective saving image neural framework confirms time. Through the fitness ranking of the individuals in the target population, the individual with the lowest fitness can be found, and the individual is combined with the current population to generate the target population. If the iteration is finished, the pareto optimal set of the image neural framework can be searched from the target population, so that the image neural framework can be determined from the pareto optimal set conveniently, and the determination efficiency of the image neural framework is effectively improved.
EXAMPLE III
Fig. 4 is a block diagram of a device for determining an image neural architecture according to a third embodiment of the present invention, which is capable of performing a method for determining an image neural architecture according to any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the method for performing the method. As shown in fig. 4, the apparatus specifically includes:
a current offspring set generating module 401, configured to perform mutation operation on a parent individual in a current population of the image neural framework according to a preset genetic operator, so as to generate a current offspring set of the parent individual; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
a target child set generating module 402, configured to input the current child set into a preset multitask learning agent model, and generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
a target population generating module 403, configured to perform fitness ranking on the child individuals in the target child set by using a preset multi-objective optimization algorithm, and generate a target population based on the current population and a preset fitness requirement;
and a neural architecture set determining module 404, configured to determine whether the current iteration number meets a preset iteration end condition, and if so, determine a pareto optimal set of the image neural architecture from the target population according to a preset multi-objective optimization algorithm, so that the user determines the target neural architecture from the pareto optimal set.
Optionally, the apparatus further comprises:
the fitness sorting module is used for carrying out fitness sorting on the image neural framework individuals in the current population by adopting a preset multi-objective optimization algorithm before carrying out mutation operation on the parent individuals in the current population of the image neural framework according to a preset genetic operator and generating a current offspring set of the parent individuals so as to obtain a fitness sorting result;
the parent individual determining module is used for obtaining at least one parent individual according to the fitness sorting result and preset selecting operation; the image neural architecture individuals of the current population comprise structure information, network performance and network scale of the image neural architecture.
Optionally, the current population is an initial population of the first iteration;
correspondingly, the device also comprises:
the network performance determining module is used for training the image neural architecture individuals in the initial population according to a preset image classification task data set before fitness ranking is carried out on the image neural architecture individuals in the current population by adopting a preset multi-objective optimization algorithm to obtain a fitness ranking result, so that the network performance of the image neural architecture individuals is obtained;
and the network scale determining module is used for determining the network scale of the image neural architecture individuals according to the structural information of the image neural architecture individuals in the initial population.
Optionally, the current child set generating module 401 is specifically configured to:
and performing mutation operations of increasing genes, reducing genes and/or replacing genes on parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals.
Optionally, the target child set generating module 402 is specifically configured to:
inputting the current offspring set into a preset multitask learning agent model, and outputting the image neural architecture network performance prediction result of the offspring individuals in the current offspring set;
and obtaining an updated target child set according to the structure information, the network scale and the network performance prediction result of the child individuals in the current child set.
Optionally, the target population generating module 403 is specifically configured to:
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and determining the target offspring individuals meeting the preset fitness requirement;
determining the network performance of the target child individual according to the image neural architecture structure information of the target child individual and a preset image classification task data set;
adding the target progeny individual to the current population to generate the target population; the image neural architecture individuals of the target population comprise structure information, network performance and network scale of the image neural architecture.
Optionally, the neural architecture set determining module 404 is specifically configured to:
if the current iteration times meet the preset iteration ending conditions, dividing the image neural framework individuals in the target population into non-dominant leading edge sets of at least two levels according to a preset multi-objective optimization algorithm, and determining the non-dominant leading edge sets meeting the preset level requirements as pareto optimal sets.
Optionally, the apparatus further comprises:
and the repeated iteration module is used for determining the target population as the current population of the next iteration after judging whether the current iteration number meets a preset iteration ending condition or not, and performing fitness sequencing on the image neural architecture individuals in the current population of the next iteration by adopting a preset multi-objective optimization algorithm to obtain new parent individuals if the current iteration number does not meet the preset iteration ending condition.
Optionally, the multi-task learning agent model includes an interaction layer and at least two radial basis function neural networks, and each radial basis function neural network includes an input layer, a hidden layer and an output layer;
correspondingly, the device also comprises:
the training data determining module is used for generating training data of the image neural architecture individuals in the current population based on a preset data format according to the structural information, the network scale and the network performance of the image neural architecture individuals in the current population before inputting the current offspring set into a preset multitask learning agent model and generating an updated target offspring set; the training data comprises network structure data and training labels, the network structure data comprises structure information and network scale, and the training labels comprise network performance;
the output layer error determining module is used for inputting the training data into a multi-task learning agent model to be trained and determining the output value and the output layer error of the training label according to the output value of the output layer of the radial basis function neural network;
the error judgment module is used for judging whether the error of the output layer meets the requirement of a preset error threshold value;
and the model training completion module is used for determining that the multi-task learning agent model is trained and inputting the current offspring set into a preset multi-task learning agent model if the multi-task learning agent model is trained.
Optionally, the apparatus further comprises:
the model updating module is used for updating the multitask learning agent model by adopting a preset gradient descent method if the output layer error does not meet the preset error threshold requirement after judging whether the output layer error meets the preset error threshold requirement;
and the model repeated training module is used for inputting training data into the updated multi-task learning agent model and determining whether the multi-task learning agent model completes training or not according to the preset error threshold requirement.
In the embodiment, the parent individuals in the population are mutated to obtain the mutated current offspring set, and the offspring individuals of the current offspring set only have the structural information of the image neural architecture and the network scale, but do not have the network performance. Through mutation operation, the possibility of the image neural architecture is increased, and the determination precision of the image neural architecture is improved. And inputting the data of the current offspring set into a pre-trained multi-task learning agent model, and predicting the network performance of the offspring individuals so as to obtain the information of the complete offspring individuals and generate a target offspring set. The problem of among the prior art, through the image classification task data set that predetermines, consume the long time to calculate network performance is solved, effective saving image neural framework confirms time. Through the fitness ranking of the individuals in the target population, the individual with the lowest fitness can be found, and the individual is combined with the current population to generate the target population. If the iteration is finished, the pareto optimal set of the image neural framework can be searched from the target population, so that the image neural framework can be determined from the pareto optimal set conveniently, and the determination efficiency of the image neural framework is effectively improved.
Example four
Fig. 5 is a schematic structural diagram of an image neural architecture determining apparatus according to a fourth embodiment of the present invention. The image neural architecture determining device may be a computer device, and FIG. 5 illustrates a block diagram of an exemplary computer device 500 suitable for use in implementing embodiments of the present invention. The computer device 500 shown in fig. 5 is only an example and should not bring any limitations to the functionality or scope of use of the embodiments of the present invention.
As shown in fig. 5, computer device 500 is in the form of a general purpose computing device. The components of computer device 500 may include, but are not limited to: one or more processors or processing units 501, a system memory 502, and a bus 503 that couples the various system components (including the system memory 502 and the processing unit 501).
Bus 503 represents one or more of any of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, a processor, or a local bus using any of a variety of bus architectures. By way of example, such architectures include, but are not limited to, Industry Standard Architecture (ISA) bus, micro-channel architecture (MAC) bus, enhanced ISA bus, Video Electronics Standards Association (VESA) local bus, and Peripheral Component Interconnect (PCI) bus.
Computer device 500 typically includes a variety of computer system readable media. Such media can be any available media that is accessible by computer device 500 and includes both volatile and nonvolatile media, removable and non-removable media.
The system memory 502 may include computer system readable media in the form of volatile memory, such as Random Access Memory (RAM)504 and/or cache memory 505. The computer device 500 may further include other removable/non-removable, volatile/nonvolatile computer system storage media. By way of example only, storage system 506 may be used to read from and write to non-removable, nonvolatile magnetic media (not shown in FIG. 5, commonly referred to as a "hard drive"). Although not shown in FIG. 5, a magnetic disk drive for reading from and writing to a removable, nonvolatile magnetic disk (e.g., a "floppy disk") and an optical disk drive for reading from or writing to a removable, nonvolatile optical disk (e.g., a CD-ROM, DVD-ROM, or other optical media) may be provided. In these cases, each drive may be connected to the bus 503 by one or more data media interfaces. Memory 502 may include at least one program product having a set (e.g., at least one) of program modules that are configured to carry out the functions of embodiments of the invention.
A program/utility 508 having a set (at least one) of program modules 507 may be stored, for instance, in memory 502, such program modules 507 including, but not limited to, an operating system, one or more application programs, other program modules, and program data, each of which examples or some combination thereof may comprise an implementation of a network environment. Program modules 507 generally perform the functions and/or methodologies of embodiments of the invention as described herein.
The computer device 500 may also communicate with one or more external devices 509 (e.g., keyboard, pointing device, display 510, etc.), with one or more devices that enable a user to interact with the computer device 500, and/or with any devices (e.g., network card, modem, etc.) that enable the computer device 500 to communicate with one or more other computing devices. Such communication may occur via input/output (I/O) interfaces 511. Moreover, computer device 500 may also communicate with one or more networks (e.g., a Local Area Network (LAN), a Wide Area Network (WAN) and/or a public network such as the Internet) via network adapter 512. As shown in FIG. 5, network adapter 512 communicates with the other modules of computer device 500 via bus 503. It should be appreciated that although not shown in FIG. 5, other hardware and/or software modules may be used in conjunction with computer device 500, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems, among others.
The processing unit 501 executes various functional applications and data processing by running a program stored in the system memory 502, for example, to implement a method for determining an image neural architecture provided by an embodiment of the present invention, including:
performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement;
and judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
EXAMPLE five
The fifth embodiment of the present invention further provides a storage medium containing computer executable instructions, on which a computer program is stored, where the computer program, when executed by a processor, implements a method for determining an image neural architecture, according to the fifth embodiment of the present invention, where the method includes:
performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement;
and judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
Computer storage media for embodiments of the invention may employ any combination of one or more computer-readable media. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. The computer-readable storage medium may be, for example, but is not limited to: an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of the computer readable storage medium would include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a Random Access Memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.
A computer readable signal medium may include a propagated data signal with computer readable program code embodied therein, for example, in baseband or as part of a carrier wave. Such a propagated data signal may take many forms, including, but not limited to, electro-magnetic, optical, or any suitable combination thereof. A computer readable signal medium may also be any computer readable medium that is not a computer readable storage medium and that can communicate, propagate, or transport a program for use by or in connection with an instruction execution system, apparatus, or device.
Program code embodied on a computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.
Computer program code for carrying out operations for aspects of the present invention may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C + + or the like and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network, including a Local Area Network (LAN) or a Wide Area Network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet service provider).
It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.

Claims (13)

1. A method for determining image neural architecture, comprising:
performing mutation operation on parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and generating a target population based on the current population and a preset fitness requirement;
and judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
2. The method of claim 1, further comprising, before performing mutation operations on parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals:
adopting a preset multi-objective optimization algorithm to carry out fitness ranking on the image neural framework individuals in the current population to obtain a fitness ranking result;
obtaining at least one parent individual according to the fitness sorting result and preset selection operation; the image neural architecture individuals of the current population comprise structure information, network performance and network scale of the image neural architecture.
3. The method of claim 2, wherein the current population is an initial population of a first iteration;
correspondingly, before the fitness ranking of the image neural architecture individuals in the current population is performed by adopting a preset multi-objective optimization algorithm and a fitness ranking result is obtained, the method further comprises the following steps:
training the image neural architecture individuals in the initial population according to a preset image classification task data set to obtain the network performance of the image neural architecture individuals;
and determining the network scale of the image neural architecture individuals according to the structure information of the image neural architecture individuals in the initial population.
4. The method of claim 1, wherein performing mutation operations on parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals comprises:
and performing mutation operations of increasing genes, reducing genes and/or replacing genes on parent individuals in the current population of the image neural architecture according to a preset genetic operator to generate a current offspring set of the parent individuals.
5. The method of claim 1, wherein inputting the current child set into a preset multitask learning agent model, and generating an updated target child set comprises:
inputting the current offspring set into a preset multitask learning agent model, and outputting the image neural architecture network performance prediction result of the offspring individuals in the current offspring set;
and obtaining an updated target child set according to the structure information, the network scale and the network performance prediction result of the child individuals in the current child set.
6. The method of claim 1, wherein fitness ranking of the offspring individuals in the target offspring set is performed by using a preset multi-objective optimization algorithm, and generating a target population based on the current population and a preset fitness requirement comprises:
adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set, and determining the target offspring individuals meeting the preset fitness requirement;
determining the network performance of the target child individual according to the image neural architecture structure information of the target child individual and a preset image classification task data set;
adding the target progeny individual to the current population to generate the target population; the image neural architecture individuals of the target population comprise structure information, network performance and network scale of the image neural architecture.
7. The method of claim 1, wherein determining whether a current iteration number meets a preset iteration end condition, and if so, determining a pareto optimal set of image neural frameworks from the target population according to a preset multi-objective optimization algorithm comprises:
if the current iteration times meet the preset iteration ending conditions, dividing the image neural framework individuals in the target population into non-dominant leading edge sets of at least two levels according to a preset multi-objective optimization algorithm, and determining the non-dominant leading edge sets meeting the preset level requirements as pareto optimal sets.
8. The method according to claim 1, further comprising, after determining whether the current iteration number satisfies a preset iteration end condition:
and if the current iteration times do not meet the preset iteration ending condition, determining the target population as the current population of the next iteration, and adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the image neural architecture individuals in the current population of the next iteration to obtain new parent individuals.
9. The method of claim 1, wherein the multitasking learning agent model comprises an interaction layer and at least two radial basis function neural networks, and wherein the radial basis function neural networks comprise an input layer, a hidden layer and an output layer;
correspondingly, before the current child set is input into a preset multitask learning agent model and an updated target child set is generated, the method further includes:
generating training data of the image neural architecture individuals in the current population based on a preset data format according to the structure information, the network scale and the network performance of the image neural architecture individuals in the current population; the training data comprises network structure data and training labels, the network structure data comprises structure information and network scale, and the training labels comprise network performance;
inputting the training data into a multi-task learning agent model to be trained, and determining an output layer error between an output value and the training label according to the output value of a radial basis function neural network output layer;
judging whether the output layer error meets the preset error threshold value requirement or not;
and if so, determining that the multi-task learning agent model is trained, and inputting the current filial generation set into a preset multi-task learning agent model.
10. The method of claim 9, after determining whether the output layer error meets a preset error threshold requirement, further comprising:
if the error of the output layer does not meet the requirement of a preset error threshold, updating the multi-task learning agent model by adopting a preset gradient descent method;
and inputting the training data into the updated multi-task learning agent model, and determining whether the multi-task learning agent model completes training according to the preset error threshold requirement.
11. An apparatus for determining neural architecture of an image, comprising:
the current offspring set generation module is used for carrying out mutation operation on the parent individuals in the current population of the image neural framework according to a preset genetic operator to generate a current offspring set of the parent individuals; wherein, the offspring individuals of the current offspring set comprise the structure information and the network scale of the image neural architecture;
the target child set generation module is used for inputting the current child set into a preset multitask learning agent model to generate an updated target child set; the descendant individuals of the target descendant set comprise structural information of an image neural architecture, a network performance prediction result and a network scale;
the target population generation module is used for adopting a preset multi-objective optimization algorithm to carry out fitness sequencing on the offspring individuals in the target offspring set and generating a target population based on the current population and a preset fitness requirement;
and the neural framework set determining module is used for judging whether the current iteration times meet a preset iteration ending condition, if so, determining a pareto optimal set of the image neural framework from the target population according to a preset multi-objective optimization algorithm, and allowing a user to determine the target neural framework from the pareto optimal set.
12. An apparatus for determining an image neural architecture, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method for determining an image neural architecture according to any one of claims 1-10 when executing the program.
13. A storage medium containing computer-executable instructions for performing the method for image neural architecture determination according to any one of claims 1-10 when executed by a computer processor.
CN202110106031.5A 2021-01-26 2021-01-26 Method, device, equipment and storage medium for determining image neural architecture Active CN112836794B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110106031.5A CN112836794B (en) 2021-01-26 2021-01-26 Method, device, equipment and storage medium for determining image neural architecture

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110106031.5A CN112836794B (en) 2021-01-26 2021-01-26 Method, device, equipment and storage medium for determining image neural architecture

Publications (2)

Publication Number Publication Date
CN112836794A true CN112836794A (en) 2021-05-25
CN112836794B CN112836794B (en) 2023-09-29

Family

ID=75931687

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110106031.5A Active CN112836794B (en) 2021-01-26 2021-01-26 Method, device, equipment and storage medium for determining image neural architecture

Country Status (1)

Country Link
CN (1) CN112836794B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780518A (en) * 2021-08-10 2021-12-10 深圳大学 Network architecture optimization method, terminal device and computer-readable storage medium
CN116108764A (en) * 2023-04-14 2023-05-12 中国科学院长春光学精密机械与物理研究所 Optical intelligent optimization method, device, equipment and medium

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109598092A (en) * 2018-12-28 2019-04-09 浙江工业大学 Merge the air source heat pump multi-objective optimization design of power method of BP neural network and more parent genetic algorithms
CN109932903A (en) * 2019-02-25 2019-06-25 北京妙微科技有限公司 The air-blower control Multipurpose Optimal Method of more parent optimization networks and genetic algorithm
CN110135498A (en) * 2019-05-17 2019-08-16 电子科技大学 A kind of image-recognizing method based on depth Evolutionary Neural Network
CN110443364A (en) * 2019-06-21 2019-11-12 深圳大学 A kind of deep neural network multitask hyperparameter optimization method and device
CN111325284A (en) * 2020-03-09 2020-06-23 武汉大学 Self-adaptive learning method and device based on multi-target dynamic distribution
CN111898689A (en) * 2020-08-05 2020-11-06 中南大学 Image classification method based on neural network architecture search

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109598092A (en) * 2018-12-28 2019-04-09 浙江工业大学 Merge the air source heat pump multi-objective optimization design of power method of BP neural network and more parent genetic algorithms
CN109932903A (en) * 2019-02-25 2019-06-25 北京妙微科技有限公司 The air-blower control Multipurpose Optimal Method of more parent optimization networks and genetic algorithm
CN110135498A (en) * 2019-05-17 2019-08-16 电子科技大学 A kind of image-recognizing method based on depth Evolutionary Neural Network
CN110443364A (en) * 2019-06-21 2019-11-12 深圳大学 A kind of deep neural network multitask hyperparameter optimization method and device
CN111325284A (en) * 2020-03-09 2020-06-23 武汉大学 Self-adaptive learning method and device based on multi-target dynamic distribution
CN111898689A (en) * 2020-08-05 2020-11-06 中南大学 Image classification method based on neural network architecture search

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113780518A (en) * 2021-08-10 2021-12-10 深圳大学 Network architecture optimization method, terminal device and computer-readable storage medium
CN113780518B (en) * 2021-08-10 2024-03-08 深圳大学 Network architecture optimization method, terminal equipment and computer readable storage medium
CN116108764A (en) * 2023-04-14 2023-05-12 中国科学院长春光学精密机械与物理研究所 Optical intelligent optimization method, device, equipment and medium
CN116108764B (en) * 2023-04-14 2023-07-07 中国科学院长春光学精密机械与物理研究所 Optical intelligent optimization method, device, equipment and medium

Also Published As

Publication number Publication date
CN112836794B (en) 2023-09-29

Similar Documents

Publication Publication Date Title
Blondel et al. A ranking approach to genomic selection
CN114048331A (en) Knowledge graph recommendation method and system based on improved KGAT model
CN111881359B (en) Ordering method, ordering system, ordering equipment and ordering storage medium in internet information retrieval
US20120123980A1 (en) Optimization technique using evolutionary algorithms
CN112836794B (en) Method, device, equipment and storage medium for determining image neural architecture
CN109086900B (en) Electric power material guarantee and allocation platform based on multi-target particle swarm optimization algorithm
Nadimi-Shahraki et al. MMKE: Multi-trial vector-based monkey king evolution algorithm and its applications for engineering optimization problems
CN112906865B (en) Neural network architecture searching method and device, electronic equipment and storage medium
US20110173145A1 (en) Classification of a document according to a weighted search tree created by genetic algorithms
Wang et al. Multiple surrogates and offspring-assisted differential evolution for high-dimensional expensive problems
CN113627471A (en) Data classification method, system, equipment and information data processing terminal
CN105808582A (en) Parallel generation method and device of decision tree on the basis of layered strategy
Bi et al. A genetic algorithm-assisted deep learning approach for crop yield prediction
CN114417058A (en) Video material screening method and device, computer equipment and storage medium
CN118036649A (en) Diversifying recommendations by improving embedded generation of a graph neural network model
CN116992151A (en) Online course recommendation method based on double-tower graph convolution neural network
CN115203631A (en) Multi-modal data analysis method and system based on improved genetic algorithm
CN113469354B (en) Memory-constrained neural network training
CN114358257A (en) Neural network pruning method and device, readable medium and electronic equipment
CN107516020B (en) Method, device, equipment and storage medium for determining importance of sequence sites
Rasekh et al. EDNC: Evolving differentiable neural computers
CN108229572B (en) Parameter optimization method and computing equipment
Gao et al. Multi-objective pointer network for combinatorial optimization
Ghorbel et al. Smart adaptive run parameterization (SArP): enhancement of user manual selection of running parameters in fluid dynamic simulations using bio-inspired and machine-learning techniques
CN115544699A (en) Submarine cable layout optimization method, device, equipment and medium based on genetic algorithm

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant