Target detection model construction method based on evolutionary neural network architecture search

Abstract

The present application relates to a kind of target detection model construction methods based on evolutionary neural network architecture search, comprising the following steps: step S1, given target detection dataset with label, the target detection dataset is divided into training dataset, validation dataset and test dataset, and training dataset is divided into three sub-training datasets with certain proportion;Step S2, define super network as search space and coding scheme;Step S3, construct multi-objective evolutionary neural network architecture search algorithm;Step S4, the optimal integer sequence activated in super network corresponding node connection output by multi-objective evolutionary neural network architecture search algorithm is decoded into corresponding trunk network structure and is integrated into RetinaNet network architecture to form a complete target detection model;Using training dataset and using SGD optimizer trains target detection model until convergence;After model training is completed, the performance of the model is tested using test dataset.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of target detection model design and optimization, and particularly relates to a target detection model construction method based on evolutionary neural network architecture search. BACKGROUND

[0002] Target detection technology has been widely applied in various fields of life, such as intelligent monitoring systems, automatic driving systems and industrial detection systems. The core of the target detection technology based on deep learning is the design of the target detection neural network architecture. The traditional manual design method needs to rely heavily on human expert knowledge and the data distribution of the corresponding data set, and careful design and tuning are needed to obtain good performance. However, this method needs to consume a large amount of manpower and material resources, and is not conducive to large-scale promotion. In particular, when the data set is updated online or the data distribution changes due to changes in natural conditions such as light and humidity, the old model often fails, and the model needs to be redesigned or tuned. In a target detection model, the task of the backbone network is to extract image features, and the feature extraction capability and network parameter quantity have a very important influence on the performance and deployment of the target detection model. A backbone network with excellent performance can extract more effective features from the data, thereby improving the detection and classification capabilities of the target detection model.

[0003] In recent years, automated machine learning (AutoML) technology has developed rapidly, providing the possibility of automated model deployment. Among them, neural network architecture search (NAS) as a subfield of AutoML has attracted widespread attention from researchers due to its ability to automatically design neural network architectures. However, the NAS-based target detection model design algorithm still has problems to be solved, and the computational bottleneck is the pre-training weight acquisition of the backbone network architecture and the evaluation process of a large number of candidate network architectures. SUMMARY

[0004] The present application is carried out in order to solve the above-mentioned problems, and aims at providing a target detection model construction method based on evolutionary neural network architecture search. On the one hand, the present application first defines a super network as a search space; then uses a multi-layer perceptron (MLP) model as a proxy model, i.e. a performance predictor; a certain number of neural network models are trained to convergence by using the MLP proxy model as a performance predictor; then, the fitness value of the candidate neural network model can be obtained directly by the performance predictor without any training, which avoids the time-consuming and laborious model training process and improves the search efficiency of the algorithm. On the other hand, the present application uses a multi-objective evolutionary algorithm as a search algorithm, takes the accuracy of the model and the size of the model as the optimization target, and designs different customized models for different deployment scenarios, which balances the size of the model and the performance of the model.

[0005] In order to achieve the above-mentioned purposes, the present application adopts the following scheme:

[0006] The present application provides a target detection model construction method based on evolutionary neural network architecture search, which has the following characteristics, including the following steps:

[0007] Step S1, a target detection data set with labels is given, and the target detection data set is divided into a training data set D train , a validation data set D val and a test data set D test , and the training data set D train is divided into three sub-training data sets D train1 , D train2 and D train3 in a certain proportion;

[0008] Step S2, define a super network as a search space and an encoding scheme, specifically:

[0009] Step S2-1, define the super network as a search space, and the search space is composed of normal cells and reduction cells, and the normal cells and the reduction cells have the same number of calculation nodes, but the connection modes of the internal nodes are different; the step length of all nodes in the normal cell is 1, and the step length of all nodes in the reduction cell is 2; each calculation node can be connected to any previous node as the input of the current node; each calculation node is a binary tree structure, including two inputs and one output, and the feature maps obtained by the two inputs are output after addition operation; the normal cell and the reduction cell each include two input nodes for receiving the input of the current cell;

[0010] Step S2-2, encoding the computing node of step S2-1;

[0011] Step S3, constructing a multi-objective evolutionary neural network architecture search algorithm, specifically:

[0012] Step S3-1, connecting normalcell and reductioncell in sequence to form a complete neural network;

[0013] Step S3-2, pre-training the super network using the ImageNet dataset until convergence, and retaining the network model weight of the super network;

[0014] Step S3-3, taking the MAP value of the validation set D val as the evaluation index of the candidate target detection model;

[0015] Step S3-4, integrating the super network as the backbone network architecture into the RetinaNet target detection framework to form a complete target detection model, and further training to convergence using the sub-training dataset D tain1 ;

[0016] Step S3-5, randomly generating M integer gene sequences, M being an integer greater than 0; each integer gene sequence is a backbone network architecture of a target detection model, and the FPN structure and RetinaNetHead structure are fixed and unchanged; the fitness value of each integer gene sequence is evaluated through weight sharing technology;

[0017] Step S3-6, arranging the M integer gene sequences and their corresponding fitness values into data pairs, using the data pairs as input to train the MLP model, and using the trained MLP model as a performance predictor;

[0018] Step S3-7, initializing the multi-objective evolutionary algorithm, initializing the population size M, the crossover probability Pc, the mutation probability Pm, and the maximum evolution generation G, and taking the loss value of each individual based on the validation set D val and the size of the model as the optimization target;

[0019] Step S3-8, randomly initializing the population P, that is, randomly generating N integer gene sequences, N being an integer greater than 0;

[0020] Step S3-9, decoding the individuals in the population P into corresponding backbone network architectures and integrating them into the RetinaNet framework to become target detection models;

[0021] Step S3-10, evaluating the fitness value of each individual using the performance predictor, and the fitness value is the loss value of each individual based on the validation set D val ;

[0022] Step S3-11, based on the current population P, two individuals are selected as parent individuals by using the tournament selection method;

[0023] Step S3-12, according to the crossover rate P c and using the single-point crossover method to perform a crossover operation on the sequence of the two parent individuals to obtain two offspring individuals;

[0024] Step S3-13, according to the mutation rate Pm, a certain calculation module in the sequence of the individual in the offspring population is replaced or deleted;

[0025] Step S3-14, steps S3-11 and S3-12 are repeated until the number of offspring individuals reaches N, forming the offspring population Q;

[0026] Step S3-15, the performance predictor is used to evaluate the fitness value of each individual in the offspring population, and the fitness value is the loss value of each individual based on the validation set D val ;

[0027] Step S3-16, the current population P and the offspring population Q are combined into a population containing 2N individuals as an intermediate population, and the environment selection method is used to select the top N individuals in the intermediate population to form the next generation population P; and feedback to step S3-9 until the preset number of evolution generations is reached;

[0028] Step S4, the optimal integer sequence output by the multi-objective evolutionary neural network architecture search algorithm is used to activate the corresponding node connection in the super network, decoded into the corresponding backbone network structure and integrated into the RetinaNet network architecture to form a complete target detection model; the training data set D train is used to train the target detection model using the SGD optimizer until convergence; after the model training is completed, the test data set D test is used to test the performance of the model.

[0029] Effects of the application

[0030] Compared with the prior art, the application has the following beneficial effects: first, the application uses a multi-objective evolutionary algorithm as a search algorithm, and takes the accuracy of the model and the size of the model as optimization targets, and designs different customized models for different deployment scenarios, balancing the size of the model and the performance of the model. Second, the application uses a multi-layer perceptron MLP model as a proxy model and integrates it into the multi-objective evolutionary algorithm framework for individual fitness evaluation. The MLP model can be used as a proxy model and can directly output the fitness value of the individual according to the encoding of the individual. Therefore, each individual can obtain its fitness without any training, thereby saving a lot of training time and improving the search efficiency of the algorithm. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 is an action flow chart of a target detection model construction method based on evolutionary neural network architecture search in an embodiment of the present application; and

[0032] Figure 2 is a structural block diagram of a multi-objective evolutionary neural network architecture search algorithm in an embodiment of the present application. DETAILED DESCRIPTION

[0033] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application is specifically described below in combination with embodiments and drawings.

[0034] In the neural network algorithm framework, all candidate network architectures are contained in the search space, and a reasonable search space is very critical to the performance of the finally generated model, and also has a very important influence on the search efficiency of the algorithm. The search algorithm defines how to quickly find the neural network architecture for the task target in the search space. Therefore, the efficiency of the search algorithm is also very critical, and it is very important for the promotion of the algorithm to quickly and accurately find the appropriate neural network architecture in the search space. The size of the model also directly affects the deployment of the model. The present application starts from the above two parts to improve the efficiency of the search algorithm and the size and performance of the final model.

[0035] To this end, on the one hand, the present application first defines a super network as a search space; then uses a multi-layer perceptron (MLP) model as a proxy model, i.e. a performance predictor; a certain number of neural network models are trained to convergence by using the MLP proxy model as a performance predictor; then, the candidate neural network model can directly obtain its fitness value through the performance predictor without any training, avoiding the time-consuming and laborious model training process and improving the search efficiency of the algorithm. On the other hand, the present application uses a multi-objective evolutionary algorithm as a search algorithm, taking the accuracy of the model and the size of the model as optimization objectives, and designs different customized models for different deployment scenarios, balancing the size of the model and the performance of the model.

[0036] <EMBODIMENT>

[0037] REFERENCE Figure 1 The target detection model construction method based on evolutionary neural network architecture search of the present application specifically includes the following steps S1 to S4.

[0038] Step S1, a target detection data set with labels is given, and the target detection data set is divided into a training data set D train , a validation data set D val and a test data set D testand the training data set D train is divided into three sub-training data sets D train1 , D train2 , and D train3 .

[0039] Step S2, define a super network as a search space and an encoding scheme.

[0040] Specifically, this step S2 includes the following sub-steps S2-1 to S2-2.

[0041] Step S2-1, define a super network as a search space, the search space is composed of normal cells and reduction cells, the normal cells and the reduction cells have the same number of computing nodes, but the connection modes of the internal nodes are different; the step size of all nodes in the normal cell is 1, and the step size of all nodes in the reduction cell is 2; each computing node can be connected to any previous node as the input of the current node; each computing node is a binary tree structure, including two inputs and one output, and the feature maps obtained by the two inputs are output after addition operation; the normal cell and the reduction cell each include two input nodes for receiving the input of the current cell.

[0042] Step S2-2, encode the computing nodes of step S2-1.

[0043] Step S3, construct a multi-objective evolutionary neural network architecture search algorithm.

[0044] Specifically, referring to Figure 2 , this step S3 includes the following sub-steps S3-1 to S3-16.

[0045] Step S3-1, connect the normal cells and the reduction cells in sequence to form a complete neural network.

[0046] Step S3-2, pre-train the super network using the ImageNet data set until convergence, and retain the network model weights of the super network.

[0047] Step S3-3, take the MAP value of the validation set D val as an evaluation index of the candidate target detection model.

[0048] Step S3-4, integrate the super network as a backbone network architecture into the RetinaNet target detection framework to form a complete target detection model, and further train using the sub-training data set D train1 until convergence.

[0049] Step S3-5, randomly generate M integer gene sequences, M is an integer greater than 0; each integer gene sequence is a backbone network architecture of a target detection model, the FPN structure and the RetinaNetHead structure are fixed and unchanged; the fitness value of each integer gene sequence is evaluated by the weight sharing technology.

[0050] Step S3-6, arrange the M integer gene sequences and their corresponding fitness values into data pairs, use the data pairs as input to train the MLP model, and use the trained MLP model as a performance predictor.

[0051] Step S3-7, multi-objective evolutionary algorithm initialization, initialize population size M, crossover probability Pc, mutation probability Pm, and maximum evolution generation G, and take the loss value of each individual based on the validation set D val and the size of the model as the optimization target.

[0052] Step S3-8, randomly initialize the population P, that is, randomly generate N integer gene sequences, N is an integer greater than 0.

[0053] Step S3-9, decode the individuals in the population P into corresponding backbone network architectures and integrate them into the RetinaNet framework, thereby becoming a target detection model.

[0054] Step S3-10, use the performance predictor to evaluate the fitness value of each individual, and the fitness value is the loss value of each individual based on the validation set D val .

[0055] Step S3-11, based on the current population P, use the tournament selection method to select two individuals as parent individuals.

[0056] Step S3-12, according to the crossover rate P c and using the single-point crossover method to perform crossover operation on the sequences of the two parent individuals to obtain two offspring individuals.

[0057] Step S3-13, according to the mutation rate Pm, replace or delete a certain calculation module in the sequence of the individual in the offspring population.

[0058] Step S3-14, repeat steps S3-11 and S3-12 until the number of offspring individuals reaches N, forming an offspring population Q.

[0059] Step S3-15, use the performance predictor to evaluate the fitness value of each individual in the offspring population, and the fitness value is the loss value of each individual based on the validation set D val .

[0060] Step S3-16, merge the current population P and the offspring population Q into a population containing 2N individuals as an intermediate population, select the top N individuals in terms of fitness value from the intermediate population using the environmental selection method to form the next generation population P; and feedback to step S3-9 until the preset evolution number is reached.

[0061] Step S4, use the optimal integer sequence output by the multi-objective evolutionary neural network architecture search algorithm to activate the corresponding node connection in the super network, decode into the corresponding backbone network structure and integrate into the RetinaNet network architecture to form a complete target detection model; use the training data set D train and train the target detection model using the SGD optimizer until convergence; after the model training is completed, use the test data set D test Test the performance of the model.

[0062] Effects of the embodiment

[0063] Compared with the prior art, the embodiment has the following beneficial effects: first, the present application uses a multi-objective evolutionary algorithm as a search algorithm, and takes the accuracy of the model and the size of the model as optimization targets, designs different customized models for different deployment scenarios, and balances the size of the model and the performance of the model. Second, the present application uses a multi-layer perceptron MLP model as a proxy model and integrates it into a multi-objective evolutionary algorithm framework for individual fitness evaluation. The MLP model can be used as a proxy model and can directly output the fitness value of the individual according to the encoding of the individual. Therefore, each individual can obtain its fitness without any training, thereby saving a lot of training time and improving the search efficiency of the algorithm.

[0064] The above embodiments are preferred cases of the present application and do not limit the protection scope of the present application.

Claims

1.A method for constructing a target detection model based on evolutionary neural network architecture search, characterized in that, The method comprises the following steps: Step S1, given a labeled target detection dataset, divide the target detection dataset into a training dataset D train , a validation dataset D val , and a test dataset D test , and divide the training dataset D train into three sub-training datasets D train1 , D train2 , D train3 in a certain ratio; Step S2, defining a hypernetwork as a search space and an encoding scheme, specifically: Step S2-1, defining a hypernetwork as a search space, the search space being composed of a normal cell and a reduction cell, the normal cell and the reduction cell having the same number of computing nodes but different connection modes of internal nodes; the step length of all nodes in the normal cell is 1, and the step length of all nodes in the reduction cell is 2; each computing node can be connected to any previous node as the input of the current node; each computing node is a binary tree structure, comprising two inputs and one output, and the feature maps obtained by the two inputs are output after addition operation; The normal cell and the reduction cell each comprise two input nodes for receiving the input of the current cell; Step S2-2, encoding the computing nodes in step S2-1; Step S3, constructing a multi-objective evolutionary neural network architecture search algorithm, specifically: Step S3-1, connecting the normal cell and the reduction cell in sequence to form a complete neural network; Step S3-2, pre-training the hypernetwork using an ImageNet dataset until convergence, and retaining the network model weight of the hypernetwork; Step S3-3, taking the MAP value of the verification set D val as the evaluation index of the candidate target detection model; Step S3-4, incorporate the super network as the backbone network architecture into the RetinaNet target detection framework to form a complete target detection model, and use the sub-training data set D train1 further training to convergence; Step S3-5, randomly generating M integer gene sequences, M being an integer greater than 0; each integer gene sequence is a backbone network architecture of a target detection model, and the FPN structure and the RetinaNet Head structure are fixed and unchanged; the fitness value of each integer gene sequence is evaluated through weight sharing technology; Step S3-6, arranging the M integer gene sequences and the corresponding fitness values into data pairs, using the data pairs as input to train the MLP model, and using the trained MLP model as a performance predictor; Step S3-7, multi-objective evolutionary algorithm initialization, initialize population size M, crossover probability Pc, mutation probability Pm, maximum evolution generation G, and each individual is based on the validation set D val loss value and the size of the model as the optimization target; Step S3-8, randomly initializing a population P, that is, randomly generating N integer gene sequences, N being an integer greater than 0; Step S3-9, decoding the individual solutions in the population P into corresponding backbone network architectures and integrating them into the RetinaNet framework, thereby becoming a target detection model; Step S3-10, evaluation of the fitness value of each individual with the performance predictor, the fitness value being the loss value of each individual based on the validation set D val ; Step S3-11, based on the current population P, using the tournament selection method to select two individuals as parent individuals; Step S3-12, performing crossover operation in the sequences of the two parent individuals according to the crossover rate P c and performing crossover operation in the sequences of the two parent individuals by using the single-point crossover method to obtain two offspring individuals; Step S3-13, replacing or deleting a certain computing module in the individual sequence in the child population according to the mutation rate Pm; Step S3-14, repeating steps S3-11 and S3-12 until the number of child individuals reaches N, forming a child population Q; Step S3-15, evaluating each individual in the offspring population with the performance predictor for a fitness value, the fitness value being the loss value of each individual based on the validation set D val ; Step S3-16, combining the current population P and the child population Q into a population containing 2N individuals as an intermediate population, and using the environmental selection method to select the top N individuals in terms of fitness value from the intermediate population to form the next generation population P; And feedback to step S3-9 until the preset evolution number is reached; Step S4, using the optimal integer sequence output by the multi-object evolutionary neural network architecture search algorithm to activate the corresponding node connection in the super network, decoding into the corresponding backbone network structure and integrating into the RetinaNet network architecture to form a complete target detection model; using the training data set D train and using the SGD optimizer to train the target detection model until convergence; after the model training is completed, the test data set D test test the performance of the model.

Images