A genetic algorithm-based convolutional neural network and an optimization method thereof

Abstract

The present application relates to the technical field of neural network model structure optimization, and particularly discloses a hollow convolutional neural network based on a genetic algorithm and an optimization method thereof, which comprises the following steps: first, constructing an architecture and hyperparameter hierarchical collaborative evolution population and pairing; then, calculating a task performance and structure efficiency fitness vector through hierarchical collaborative fitness evaluation; after that, performing sub-population competition interaction, architecture-guided collaborative mutation, and elite reservation screening to obtain an optimal individual pair; finally, decoding and fully training the optimal individual pair to obtain a final model; and the optimized hollow convolutional neural network realizes dynamic adaptation of the architecture and hyperparameters, balances high task performance and optimal structure efficiency, has high automation, reduces the design threshold, improves training stability and scene adaptability, and is suitable for computer vision inference tasks such as image segmentation and target detection. The present application realizes collaborative optimization of the architecture and hyperparameters, balances task performance and running efficiency, has high automation, avoids falling into a local optimum, and is suitable for computer vision tasks such as image segmentation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of structure optimization of neural network models, in particular to a cavity convolutional neural network based on a genetic algorithm and an optimization method thereof. BACKGROUND

[0002] With the rapid development of computer vision technology, the cavity convolutional neural network has been widely used in image segmentation, target detection, semantic understanding and other fields due to its core advantage of expanding the receptive field without increasing the number of parameters. By introducing a cavity rate parameter into the convolution kernel, this type of network can flexibly balance the needs of local detail capture and global context information integration, becoming one of the key model architectures for processing complex scene data.

[0003] However, there are still many problems to be solved in the design and optimization of existing cavity convolutional neural networks. First, the network architecture and hyperparameter configuration rely on manual adjustment of artificial experience, which not only requires researchers to have deep professional knowledge, but also has the problems of long optimization period and low efficiency. Moreover, manual design is difficult to achieve precise adaptation of structure and parameters, which easily leads to performance bottlenecks or computational redundancy of the model. Second, most automatic optimization methods use single-dimensional optimization strategies, or search for network architecture alone, or adjust hyperparameters alone, ignoring the synergistic relationship between architecture topology and hyperparameters, so that the optimized model is often difficult to balance task performance and running efficiency. Third, existing neural network optimization schemes based on genetic algorithms mostly use single population evolution mode, lack of hierarchical and collaborative design of architecture and hyperparameters, and are prone to problems of insufficient diversity or slow convergence speed in the population evolution process. Moreover, the fitness evaluation mostly only focuses on task performance indicators, without fully considering the structure efficiency factors such as the amount of network computation consumption, resulting in that the optimized model can meet the accuracy requirement, but is difficult to adapt to hardware environments with limited computing power. Fourth, the crossover and mutation operations of genetic algorithm lack targeted guidance, and architecture changes and parameter adjustments are often independent of each other, which easily produces invalid candidate models that do not match structure and parameters, further reducing the optimization efficiency and reliability of the final model.

[0004] These problems seriously restrict the deployment and application of cavity convolutional neural networks in actual scenarios, so there is an urgent need for an optimization method that can realize collaborative optimization of architecture and hyperparameters, balance task performance and structure efficiency, and has high automation, to solve the deficiencies of existing technologies. SUMMARY

[0005] The present application relates to the technical field of structure optimization of neural network models, in particular to a cavity convolutional neural network based on a genetic algorithm and an optimization method thereof.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A method for optimizing a dilated convolutional neural network based on a genetic algorithm, comprising the following steps:

[0008] Step S1: constructing a hierarchical coevolutionary population, including an architecture subpopulation and a hyperparameter subpopulation using different encoding methods respectively, randomly initializing the two subpopulations and pairing them to form initial co-individual pairs;

[0009] Step S2: hierarchical co-adaptive fitness evaluation: decoding and quickly training the network for each co-individual pair, and calculating a hierarchical fitness vector containing task performance fitness and structural efficiency fitness, wherein the structural efficiency fitness is determined by the average receptive field growth rate and the computational cost consumption of the network;

[0010] Step S3: intra-subpopulation competition and interaction: performing fitness-based competitive selection in the two subpopulations respectively, and then performing architecture-individual-guided directional crossover operation to ensure the coordination of structural changes and parameter changes;

[0011] Step S4: architecture-guided co-variation: performing probabilistic variation on the offspring individuals, and when the architecture individual is varied, forcing the hyperparameter individual paired with it to have a correlated adaptive variation at the corresponding position;

[0012] Step S5: elite preservation and environmental selection: merging all co-individual pairs of the parent and offspring, and based on the hierarchical fitness vector, using the mechanism of first sorting by task performance and then screening by structural efficiency to retain a predetermined number of optimal individual pairs;

[0013] Step S6: iteration and model generation: repeating steps S2 to S5 until a predetermined number of evolution generations is reached, selecting the co-individual pair with the highest task performance fitness, decoding and fully training to obtain the final dilated convolutional neural network model.

[0014] As a preferred scheme, the step S1 of constructing a hierarchical coevolutionary population specifically comprises:

[0015] initializing the architecture subpopulation and the hyperparameter subpopulation in parallel, the architecture subpopulation containing a first predetermined number of architecture individuals, and the hyperparameter subpopulation containing the same number of hyperparameter individuals;

[0016] for each architecture individual in the architecture subpopulation, using a first encoding sequence and a second encoding sequence to represent the network structure together; the first encoding sequence is a binary encoding sequence, each bit of which represents whether the skip connection between the current network layer and the previous specific layer is valid; the second encoding sequence is an integer encoding sequence, each bit of which represents the type of dilated convolution kernel used by the corresponding network layer, which is selected from a predefined candidate kernel set containing standard convolution kernels and multiple expansion patterns;

[0017] For each hyper-parameter individual in the hyper-parameter sub-population, its coding structure is guided to be generated by the architecture individual paired therewith; first, according to the total number of network layers of the paired architecture individual, the third coding sequence length of the hyper-parameter individual is determined, the third coding sequence being a real number coding sequence, each bit of which representing a hole rate parameter of a corresponding network layer in the paired architecture individual; then, an independent real number coding is generated to represent the network filter base parameter corresponding to the paired architecture individual;

[0018] After the above encoding is completed, each architecture individual in the architecture sub-population is randomly paired with one hyper-parameter individual in the hyper-parameter sub-population, so that the structure information of each architecture individual and the parameter information of one hyper-parameter individual are combined to form a complete initial collaborative individual pair to represent a candidate model of a dilated convolutional neural network.

[0019] As a preferred scheme, the hierarchical collaborative fitness evaluation in step S2 specifically includes:

[0020] The collaborative individual pair is decoded, and a corresponding dilated convolutional neural network model is constructed according to the first coding sequence and the second coding sequence of the architecture individual in the collaborative individual pair and the third coding sequence and the filter base parameter of the hyper-parameter individual;

[0021] The dilated convolutional neural network model is quickly trained using the training data subset to obtain a preliminary training model;

[0022] The preliminary training model is forward-propagated on the validation set to obtain a classification accuracy of the validation set;

[0023] According to the classification accuracy of the validation set, the task performance fitness is calculated through linear normalization mapping;

[0024] Based on the hole rate parameter sequence represented by the third coding sequence of the hyper-parameter individual, the receptive field growth step of each network layer is dynamically calculated through recursive multiplication, and the average receptive field growth rate of all network layers is counted;

[0025] According to the first coding sequence and the second coding sequence of the architecture individual and the filter base parameter of the hyper-parameter individual, the computational complexity consumption of the dilated convolutional neural network model in single forward propagation is evaluated;

[0026] The average receptive field growth rate and the computational complexity consumption are combined to calculate the structure efficiency fitness through weighted reciprocal fusion of a preset weight coefficient;

[0027] The task performance fitness and the structure efficiency fitness are combined in a preset order to form a hierarchical fitness vector, which is used for subsequent competition and interaction within the sub-population.

[0028] As a preferred scheme, the competition and interaction within the sub-population in step S3 specifically includes:

[0029] Based on the hierarchical fitness vectors of each pair of collaborative individuals, the entire architecture individuals are first sorted according to the task performance fitness, and then a first number of superior architecture individuals are selected from the architecture sub-population according to a preset selection ratio by using a ranking-based roulette wheel selection method;

[0030] Based on the paired hyperparameter individuals corresponding to the selected superior architecture individuals, the structure efficiency fitness in the hierarchical fitness vectors corresponding to the hyperparameter individuals is sorted, and the same number of superior hyperparameter individuals are selected from the hyperparameter sub-population by using the same ranking-based roulette wheel selection method;

[0031] For each pair of selected superior architecture individuals and superior hyperparameter individuals, the skip connection mode represented by the first encoding sequence of the superior architecture individual is analyzed; the network layer positions where the skip connection changes are identified, and these positions are set as forced crossover points for the third encoding sequence of the hyperparameter individual in subsequent crossover operations;

[0032] Single-point crossover operations are performed between the superior architecture individuals to generate offspring architecture individuals; the skip connection change layers inherited by the offspring architecture individuals from their parents are determined based on the skip connection mode of the offspring architecture individuals; multi-point crossover operations are performed between the corresponding superior hyperparameter individuals, wherein the crossover points are forced to be set at the forced crossover points, to generate offspring hyperparameter individuals that are semantically collaborative with the offspring architecture individuals in structural changes.

[0033] As a preferred scheme, the architecture-guided collaborative variation in step S4 specifically includes:

[0034] The offspring collaborative individual pairs composed of the offspring architecture individuals generated in step S3 and their paired offspring hyperparameter individuals are respectively set with independent architecture variation probabilities and hyperparameter variation probabilities;

[0035] For the offspring architecture individuals, the architecture variation probability triggers variation; the variation operation includes a first type of variation and a second type of variation; the first type of variation is to flip the binary bits at randomly selected positions in the first encoding sequence to achieve the addition or deletion of skip connections; the second type of variation is to replace the integer values at randomly selected positions in the second encoding sequence to reselect the types of the dilated convolution kernels of the current network layer from a predefined candidate kernel set;

[0036] When the offspring architecture individuals undergo the second type of variation, the system records the specific network layer positions where the dilated convolution kernel type switching occurs, and marks these positions as associated variation guide points; at the same time, according to the switched dilated convolution kernel type, the allowed domain of the candidate values of the hole rate parameter of the corresponding network layer is obtained from the preset kernel type-hole rate mapping relationship;

[0037] Based on the correlation variation guide point, the correlation variation of the paired offspring hyperparameter individual is forcibly triggered; in the third encoding sequence of the offspring hyperparameter individual, the position corresponding to the correlation variation guide point is located, and the original void rate parameter value is replaced by a value selected from the void rate parameter candidate value domain according to a preset adaptive strategy, so as to adapt to the change of the void convolution kernel type in the offspring architecture individual;

[0038] After the above mutation operation is completed, the offspring architecture individual and the offspring hyperparameter individual which has experienced the correlation adaptive variation jointly constitute a post-mutation collaborative individual pair.

[0039] As a preferred scheme, the elite reservation and environment selection in step S5 specifically includes:

[0040] All offspring collaborative individual pairs generated after the competition and interaction in the subpopulation in step S3 and the collaborative variation based on the architecture guide in step S4 are combined with the parent collaborative individual pairs of the current generation to form a total set of collaborative individual pairs containing all candidate schemes;

[0041] According to the hierarchical fitness vector, all collaborative individual pairs in the total set of collaborative individual pairs are first sorted, and the sorting basis of the first sorting is the task performance fitness corresponding to the collaborative individual pairs, and the arrangement is in the order from high to low;

[0042] Based on the result of the first sorting, the total set of collaborative individual pairs after arrangement is dynamically divided into performance levels; when dividing, the absolute difference of the task performance fitness between adjacent collaborative individual pairs is calculated, and when the absolute difference is less than a preset performance difference threshold, the adjacent collaborative individual pairs are classified into the same performance level;

[0043] Within each performance level, all collaborative individual pairs belonging to the level are second sorted, and the basis of the second sorting is the structure efficiency fitness corresponding to the collaborative individual pairs, and the arrangement is in the order from high to low;

[0044] According to the number of the new generation population, the collaborative individual pairs in front of each performance level in the second sorting are selected in order from high to low performance level, until the total number of selected collaborative individual pairs reaches the number of the new generation population;

[0045] The finally selected collaborative individual pairs form the architecture subpopulation and the hyperparameter subpopulation of the new generation, which are used for the next round of iteration in step S6 or as the input for generating the final model.

[0046] As a preferred scheme, step S6 includes the following specific steps:

[0047] Extracting the collaborative individual pair with the highest task performance fitness in the new generation population generated from step S5, decoding and reconstructing it into a complete dilated convolutional neural network model, and archiving it as the optimal candidate model of the current generation;

[0048] Using the optimal candidate model of the current generation and the historical archived optimal candidate model, a dynamically growing experience model library is constructed; all models in the experience model library are analyzed for structural similarity, and the preset evolution number is dynamically adjusted according to the analysis result;

[0049] Judging whether the current evolution number has reached the dynamically adjusted preset evolution number or meets other preset convergence conditions; if not, taking the new generation population as the parent population of the next round of iteration, returning to step S2 to perform the next round of iteration;

[0050] When the evolution reaches the preset number of generations or meets the convergence condition, terminate the evolution process; select the candidate model with the highest task performance fitness from the final experience model library, and decode to obtain the optimal network architecture and hyperparameter configuration corresponding thereto;

[0051] Based on the optimal network architecture and hyperparameter configuration, a complete dilated convolutional neural network is re-instantiated; using the complete training data set and adopting the preset optimizer and dynamically adjusted learning rate scheduling strategy, the network is fully trained;

[0052] Evaluating the performance of the fully trained network on an independent test set to obtain the final performance indicator; encapsulating the structural parameters of the network, the trained weight parameters, and the final performance indicator to generate a final dilated convolutional neural network model that can be directly used for inference tasks.

[0053] A dilated convolutional neural network based on a genetic algorithm, wherein the dilated convolutional neural network is optimized by an optimization method.

[0054] As can be seen from the technical solutions provided by the above-mentioned application, the dilated convolutional neural network based on a genetic algorithm and the optimization method thereof provided by the application have the beneficial effects that:

[0055] The dynamic collaborative adaptation of the network architecture and the hyperparameters is realized, the performance bottleneck caused by the disconnection between the structure and the parameters in the traditional manual design is solved, the model has reasonable structural parameter matching in the initial stage, and the training stability and the final task performance are improved;

[0056] Through the hierarchical collaborative fitness evaluation mechanism, the task performance and the structural efficiency of the model are taken into account, the average receptive field growth rate and the computational consumption are optimized while ensuring the core performance indicators such as high classification accuracy, the balance between performance and computing power is realized, and the inference requirements of different hardware environments are adapted;

[0057] With the mechanism of hierarchical co-evolution genetic algorithm, through competitive selection, directional crossover, cooperative mutation and elite preservation, better network configuration is continuously mined, effectively avoiding the evolution process from falling into local optimal solution, while maintaining population diversity, so that the model has wider scene adaptation ability;

[0058] The optimization process is fully automated, from population construction, fitness evaluation to iterative optimization and model generation, without manual adjustment of network structure or hyperparameters, reducing the model design threshold, improving the research and development efficiency, and suitable for users lacking professional parameter tuning experience or large-scale model deployment scenarios;

[0059] Flexible adaptation of the types of cavity convolution kernels and the cavity rate parameters, combined with dynamic configuration of the jump connection, enhances the model's ability to capture local details and global context information, and shows stronger feature expression ability in image segmentation, target detection and other computer vision tasks, improving the inference accuracy in complex scenarios. BRIEF DESCRIPTION OF DRAWINGS

[0060] Figure 1 A flowchart of the steps of the genetic algorithm-based cavity convolutional neural network optimization method of the application. DETAILED DESCRIPTION

[0061] In order to make the purpose, technical scheme and advantages of the application clearer, the application will be further described in detail below in combination with the drawings and examples. It should be understood that the specific examples described herein are only used to explain the application and do not limit the application.

[0062] In order to better understand the above technical scheme, the above technical scheme will be described in detail below in combination with the drawings and specific embodiments of the specification.

[0063] As shown in Figure 1 The application embodiment provides a genetic algorithm-based cavity convolutional neural network optimization method, which includes the following steps:

[0064] Step S1: Construct a hierarchical co-evolution population, including an architecture sub-population and a hyperparameter sub-population using different encoding methods respectively, randomly initialize the two sub-populations and pair them to form an initial co-evolution individual pair;

[0065] Step S2: Hierarchical co-evolution fitness evaluation: for each co-evolution individual pair, decode and quickly train the network, calculate a hierarchical fitness vector containing task performance fitness and structure efficiency fitness, wherein the structure efficiency fitness is determined by the average receptive field growth rate and the computational load consumption of the network;

[0066] Step S3: Intra-subpopulation competition and interaction: Perform fitness-based competitive selection in the two subpopulations respectively, followed by performing architecture individual-guided directional crossover operation, ensuring that structural changes and parameter changes are coordinated;

[0067] Step S4: Architecture-guided coordinated mutation: Perform probabilistic mutation on the offspring individuals, and when the architecture individuals are mutated, force the associated hyperparameter individuals to undergo correlated adaptive mutation at the corresponding positions;

[0068] Step S5: Elite preservation and environmental selection: Merge all coordinated individual pairs of the parent and offspring, and based on the hierarchical fitness vector, use the mechanism of first sorting by task performance and then filtering by structural efficiency to retain a preset number of optimal individual pairs;

[0069] Step S6: Iteration and model generation: Repeat steps S2 to S5 until a preset number of evolutionary generations is reached, select the coordinated individual pair with the highest task performance fitness, decode and fully train to obtain the final hollow convolutional neural network model.

[0070] In this embodiment, the role of step S1 is to construct a hierarchical coordinated evolutionary population. By initializing the architecture subpopulation and the hyperparameter subpopulation in parallel, the network structure and parameter configuration are represented using differentiated encoding methods, and then randomly paired to form complete initial coordinated individual pairs. This provides a structure-specified and data-complete candidate model set for subsequent hierarchical coordinated fitness evaluation, population interaction, and mutation. The detailed steps are as follows:

[0071] Step S1-1: Population initialization parameter configuration:

[0072] First, the core parameter preset rules for population construction are defined, and the first preset number is determined as the individual number benchmark for the architecture subpopulation and the hyperparameter subpopulation. The number of individuals in the two subpopulations remains the same to meet the subsequent pairing requirements. At the same time, three types of key configuration information are predefined: the first type is the encoding rule of the architecture subpopulation, which specifies the binary bit value range of the first encoding sequence and the integer mapping rule of the second encoding sequence; the second type is the encoding constraint of the hyperparameter subpopulation, which sets the real number value interval of the third encoding sequence and the reasonable range of the filter basis parameter; the third type is the predefined candidate kernel set, which includes standard convolution kernels and various dilated mode hollow convolution kernels, each convolution kernel corresponds to a unique integer identifier for the second encoding sequence mapping. After parameter configuration is completed, the system automatically checks the rationality of the configuration information, such as confirming whether the first preset number meets the computing resource bearing capacity and whether the types of the candidate kernel set cover the network structure design requirements. If there is a configuration exception, the parameter re-adjustment process is triggered to ensure that the population initialization can be started based on compliant parameters;

[0073] Step S1-2: Architecture subpopulation encoding generation:

[0074] The architecture sub-population parallel initialization process is started, and for each architecture individual, a first encoding sequence and a second encoding sequence are generated, which are associated with each other and together represent the network structure; the first encoding sequence uses binary encoding, and the sequence length is consistent with the preset maximum number of network layers, and each binary number corresponds to a network layer, where a value of 1 represents that the current network layer is connected to the previous specific layer by a jump connection, and a value of 0 represents that the jump connection is invalid, and the binary sequence is initialized by randomly generating 0 or 1; the second encoding sequence uses integer encoding, and the sequence length is the same as that of the first encoding sequence, and each integer corresponds to a network layer, and the value range of the integer is the index interval of the predefined candidate kernel set, and the type of the convolution kernel of the corresponding network layer is determined by randomly selecting an integer in the interval, ensuring that the type of the convolution kernel of each network layer is selected from the preset candidate set; after the two encoding sequences of all architecture individuals are generated, the encoding data of each individual is automatically stored by the system to form a complete architecture sub-population;

[0075] Step S1-3: Hyperparameter sub-population encoding generation:

[0076] Based on the individual data of the architecture sub-population, the hyperparameter sub-population parallel initialization process is started, and the encoding structure of each hyperparameter individual is directly generated by the paired architecture individual; first, the total number of network layers of the paired architecture individual is extracted, and the number of layers is directly used as the length of the third encoding sequence of the hyperparameter individual, the third encoding sequence uses real number encoding, and each real number corresponds to the hole rate parameter of the network layer at the same position in the paired architecture individual, and the initialization of the third encoding sequence is completed by randomly selecting a value in the preset real number interval; then, an independent real number encoding is generated as the filter base parameter, which is used to represent the network filter number benchmark corresponding to the paired architecture individual, and the value needs to meet the constraint requirements of network computing efficiency and performance balance; after the encoding of all hyperparameter individuals is generated, the system automatically associates the identification of the paired architecture individual to ensure that the encoding data of each hyperparameter individual and the structure information of the corresponding architecture individual are logically associated to complete the construction of the hyperparameter sub-population;

[0077] Step S1-4: Coordination individual pairing and effectiveness verification:

[0078] The one-to-one random pairing mechanism is used to pair each architecture individual in the architecture sub-population with a hyperparameter individual in the hyperparameter sub-population, to ensure that each architecture individual corresponds to only one hyperparameter individual, and that all individuals are involved in pairing without omission. After pairing is completed, an initial set of collaborative individual pairs is formed, each of which includes the first and second encoding sequences of the architecture individual and the third encoding sequence and filter base parameter of the hyperparameter individual, collectively representing a complete candidate model of the dilated convolutional neural network. Subsequently, the system performs effectiveness verification on all initial collaborative individual pairs: checks whether the length of the encoding sequence of the architecture individual is consistent with the length of the third encoding sequence of the hyperparameter individual; checks whether the integers in the second encoding sequence all correspond to valid indexes of the predefined candidate kernel set; checks whether the real numbers of the third encoding sequence and the filter base parameter are within the preset reasonable range; if there is a collaborative individual pair that fails the verification, the pairing and encoding generation process of the individual pair is re-executed until all collaborative individual pairs meet the effectiveness requirements, and finally an initial set of collaborative individual pairs that are structurally standardized and data complete is output.

[0079] In this embodiment, the role of step S2 is to perform hierarchical collaborative fitness evaluation on each collaborative individual pair in the initial or iterative process, to build a network model by decoding the individual pair, perform rapid training and verification, quantify the task performance fitness and structure efficiency fitness, and finally form a hierarchical fitness vector, which provides a core evaluation basis for subsequent operations such as competition and selection within the sub-population and directional crossover. The following are the detailed steps:

[0080] Step S2-1: decoding of collaborative individual pairs and network model construction

[0081] First, decoding is performed on a single collaborative individual pair to extract the first and second encoding sequences of the architecture individual and the third encoding sequence and filter base parameter of the hyperparameter individual in the individual pair. According to the decoded information, a dilated convolutional neural network model is generated according to the preset network construction rules: according to the binary bit distribution of the first encoding sequence, establish or cancel the jump connection between the corresponding network layers, and if the binary bit is 1, keep the jump connection between the current layer and the previous specific layer, and if it is 0, remove the connection; according to the integer value of the second encoding sequence, call the corresponding type of dilated convolution kernel from the predefined candidate kernel set and assign it to each network layer as the convolution operation core; according to the real number value of the third encoding sequence, set the dilated rate parameter of each network layer to control the dilatation degree of the convolution kernel; according to the filter base parameter, determine the filter number reference of the entire network, and allocate the input and output channel numbers of each layer according to the inter-layer proportion; after the model construction is completed, the system automatically checks the integrity of the network layer connection and the rationality of the parameter configuration, such as confirming that there is no closed loop in the jump connection and the dilated rate parameter does not exceed the device calculation limit, and if there is an abnormality, the model structure is re-decoded and corrected to ensure that the network can be normally started and trained;

[0082] Step S2-2: Network model fast training and preliminary verification:

[0083] The constructed hole convolutional neural network model is subjected to fast training by calling a training data subset. The training process adopts a simplified training strategy: a short training round is set, 10% to 30% of the samples in the training data set are selected as the training subset, a stochastic gradient descent optimizer is adopted, and a large initial learning rate is set to accelerate parameter updating. During the training process, the system monitors the loss function change trend of the model in real time, and terminates the fast training in advance when the loss function does not decrease significantly for consecutive rounds to avoid invalid iteration. After the fast training is completed, a preliminary training model is obtained, the validation set data is input into the model to perform forward propagation verification, the matching situation of the predicted class and the real class of each sample is calculated, the number of samples that are correctly predicted in all validation samples is counted, and finally the classification accuracy of the validation set is calculated, which is used as the basis data for subsequent task performance fitness calculation;

[0084] Step S2-3: Task performance fitness calculation:

[0085] Based on the classification accuracy of the validation set obtained in step S2-2, the task performance fitness is calculated by using a linear normalization mapping method. First, the classification accuracy of the validation set of the model corresponding to the current all collaborative individual pairs is counted to determine the maximum value and the minimum value ; then the task performance fitness of a single collaborative individual pair is calculated by the following formula:

[0086] , wherein is the task performance fitness, is the classification accuracy of the validation set of the model corresponding to the current collaborative individual pair, is the highest classification accuracy of the validation set of the model of all individual pairs, is the lowest classification accuracy of the validation set of the model of all individual pairs; the accuracy of different models is mapped to the interval of 0 to 1 by the formula, and the closer the mapping result is to 1, the better the task performance of the model, and vice versa;

[0087] Step S2-4: Average receptive field growth rate calculation:

[0088] All hole rates in the third encoding sequence of the hyperparameter individual are extracted to form a hole rate parameter sequence in the order of network layers. Based on the sequence, the receptive field growth step of each network layer is calculated by using recursive multiplication: the initial receptive field growth step of the first layer network is , which is equal to the basic size of the convolution kernel of the layer; from the second layer, the receptive field growth step of the layer is , which is equal to the receptive field growth step of the -1 layer receptive field growth step With the first Layer cavity rate parameters The product of ; after calculating the receptive field growth step of all network layers in turn, the arithmetic mean of all step values is obtained, which is the average receptive field growth rate of the model , the calculation formula is as follows:

[0089] , wherein is the average receptive field growth rate, to are the receptive field growth steps of the first layer to the nth layer, and n is the total number of layers of the network; the greater the average receptive field growth rate, the stronger the model's ability to capture global information of the input image;

[0090] Step S2-5: network model computation consumption evaluation:

[0091] According to the first encoding sequence, the second encoding sequence of the architecture individual and the filter base parameter of the hyperparameter individual, the computation consumption of the dilated convolutional neural network model in single forward propagation is evaluated; the calculation process is disassembled by layer: for each layer of dilated convolution operation, the computation is calculated by the following formula:

[0092] , wherein is the computation of a single layer, is the base size of the dilated convolution kernel of the layer, is the number of input channels of the layer, is the number of output channels of the layer (obtained by proportional distribution according to the filter base parameter), is the height of the input feature map of the layer, is the width of the input feature map of the layer; for the network layer with a jump connection, the computation increment of the jump connection is additionally calculated, that is, the element addition operation amount of the jump connection input feature map and the current layer output feature map, which is equal to the number of elements of the current layer output feature map; the computation of all network layers and the computation increment of the jump connection are summed up, and the total computation consumption of the model in single forward propagation is obtained.

[0093] Step S2-6: structure efficiency fitness calculation:

[0094] The average receptive field growth rate obtained in step S2-4 is combined with the total computation consumption obtained in step S2-5, and the structure efficiency fitness is calculated by using the weighted reciprocal fusion method of the preset weight coefficient; first, two weight coefficients and , and satisfies , wherein is the weight of the average receptive field growth rate, is the weight of the total computational cost; then the structure efficiency fitness is calculated by the following formula:

[0095] , wherein is the structure efficiency fitness, and is a preset weight coefficient, is the average receptive field growth rate, is the total computational cost of a single forward propagation; the formula realizes the evaluation logic that "the greater the receptive field growth rate, the smaller the computational cost, and the higher the structure efficiency fitness" in the form of reciprocal, and the weighting operation can adjust the importance of the two indicators according to actual needs, for example, increasing the value of in the limited computing power scene, to preferentially select the model with smaller computational cost;

[0096] Step S2-7: hierarchical fitness vector formation:

[0097] The task performance fitness calculated in step S2-3 is combined with the structure efficiency fitness calculated in step S2-6 in a preset order to form a hierarchical fitness vector; the default combination order of the vector is , that is, the first element is the task performance fitness, and the second element is the structure efficiency fitness; after the hierarchical fitness vector of a single cooperative individual pair is generated, the system automatically stores it in association with the identification information of the individual pair, and simultaneously aggregates the hierarchical fitness vectors of all cooperative individual pairs to form a global evaluation dataset, which provides data support for the competition selection in the sub-population of step S3.

[0098] In this embodiment, the role of step S3 is to filter superior individuals through competition selection based on fitness in the sub-population, and then perform directed crossover operation guided by the architecture individual, so that the changes of the architecture sub-population and the hyperparameter sub-population remain semantically coordinated, providing a basis for the offspring individuals with both performance advantages and coordination for subsequent mutation operation; the following are the detailed steps:

[0099] Step S3-1: architecture sub-population superior individual screening:

[0100] Based on the hierarchical fitness vectors of all cooperative individual pairs, first, sort all architecture individuals in the architecture sub-population: take the task performance fitness in the hierarchical fitness vector as the only sorting basis, determine the ranking of each architecture individual in order from high to low according to the numerical value, and the earlier the ranking, the better the task performance of the model corresponding to the architecture individual.

[0101] Set a preset selection ratio (e.g., 30% to 50%), and calculate the number of dominant architecture individuals to be screened based on this ratio, denoted as . ( It is a positive integer, and Total number of individuals in the architecture subpopulation The ranking-based roulette wheel selection method is used to select dominant individuals. First, the selection probability of each architecture individual is calculated, and the calculation formula is as follows:

[0102] ,in, For the first The probability of choosing an individual with a named structure. The total number of individuals in the subpopulation. The ranking of the individuals in this architecture ( Values ​​range from 1 to 1 corresponds to the individual with the highest task performance fitness;

[0103] Repeated selections are made using a roulette wheel selection mechanism. (each extraction is performed without replacement) to ultimately select the best individuals. Each dominant architectural individual forms a dominant subset of the architectural subpopulation;

[0104] Step S3-2: Screening for dominant individuals in hyperparameter subpopulations:

[0105] The selected in step S3-1 Based on a subset of dominant architecture individuals, extract the hyperparameter individuals paired with each dominant architecture individual in the initial pair of collaborating individuals, forming a candidate set of hyperparameters associated with the subset of dominant architectures (the number of individuals in the candidate set is...). );

[0106] All hyperparameter individuals in the candidate hyperparameter set are sorted: the structural efficiency fitness in the hierarchical fitness vector is used as the sole sorting criterion, and the ranking of each hyperparameter individual is determined in descending order of value. The higher the ranking, the better the structural efficiency of its corresponding model.

[0107] The same ranking-based roulette wheel selection method as in step S3-1 is used to screen dominant individuals: Let the total number of individuals in the hyperparameter candidate set be... , No. The formula for calculating the selection probability of an individual with a number of hyperparameters is as follows:

[0108] ,in, For the first The selection probability of a single hyperparameter. The total number of individuals in the hyperparameter candidate set. The ranking of this hyperparameter individual ( Values ​​range from 1 to 1 corresponds to the individual with the highest structural efficiency fitness; repeated sampling is performed using a roulette wheel selection mechanism. (each extraction is performed without replacement) to ultimately select the best individuals. Each dominant hyperparameter individual forms a dominant subset of the hyperparameter subpopulation, and each dominant hyperparameter individual maintains an initial pairing association with one dominant architecture individual from step S3-1, thus forming... A temporary collaborative pair;

[0109] Step S3-3: Forced Intersection Point Determination:

[0110] For each dominant architecture individual in a temporary cooperative pair, the skip connection pattern represented by its first encoding sequence (binary encoding sequence) is analyzed: the first encoding sequence of the dominant architecture individual is compared bit by bit with the dominant architecture individuals in other temporary pairs to identify the network layer positions where the binary bit values ​​of the two encoding sequences are different; these positions are the network layers where skip connections change (for example, if the third bit of the first encoding sequence of individual A is 1 and the third bit of individual B is 0, then the third layer is the skip connection change layer).

[0111] All identified skip connection change layer locations are uniformly marked as "forced crossover points". Each forced crossover point corresponds to a position in the third encoding sequence of the hyperparameter individual (i.e., the position of the hole rate parameter corresponding to this network layer). Subsequent crossover operations of hyperparameter individuals must be forced to be performed at these points to ensure that hyperparameter changes are synchronized with skip connection changes.

[0112] Step S3-4: Generate offspring through single-point crossover of individual architectures:

[0113] From the dominant subset of the architectural subpopulation, two pairs of dominant architectural individuals are randomly selected as parent architectural pairs (denoted as parent A and parent B). A single-point crossover operation is performed on each parent pair:

[0114] A random intersection point is generated, which is an integer within the common length range of the first and second encoding sequences of parent A and parent B (let the encoding sequence length be 1). The intersection point value is 1 to... );

[0115] For the first encoding sequence: concatenate the binary bits to the left of the cross point of parent A with the binary bits to the right of the cross point of parent B to form the first encoding sequence of child architecture individual 1; at the same time, concatenate the binary bits to the left of the cross point of parent B with the binary bits to the right of the cross point of parent A to form the first encoding sequence of child architecture individual 2.

[0116] For the second encoding sequence: adopt the same cross point position as the first encoding sequence, splice the integer sequence on the left side of the cross point position of parent A with the integer sequence on the right side of the cross point position of parent B to form the second encoding sequence of the offspring architecture individual 1; similarly, splice the integer sequence on the left side of the cross point position of parent B with the integer sequence on the right side of the cross point position of parent A to form the second encoding sequence of the offspring architecture individual 2;

[0117] Repeat the above operation until offspring architecture individuals (consistent with the number of dominant architecture individuals) are generated, and all offspring architecture individuals collectively form an offspring architecture subset;

[0118] Step S3-5: generating offspring through multi-point crossover of hyperparameter individuals:

[0119] For each pair of parent dominant hyperparameter individuals (parent hyperparameter A, parent hyperparameter B) corresponding to each pair of offspring architecture individuals (offspring 1, offspring 2) in step S3-4, perform a multi-point crossover operation:

[0120] Extract the forced cross point positions determined in step S3-3, and let the forced cross point position set be ( is the index of the third encoding sequence for the network layer position);

[0121] For the third encoding sequences of parent hyperparameter A and parent hyperparameter B (let the third encoding sequence of parent hyperparameter A be , and the third encoding sequence of parent hyperparameter B be ): at each forced cross point position , exchange the encoding values at the corresponding positions of the two parent individuals, i.e., the third encoding sequence of the offspring hyperparameter 1 after crossover satisfies (the non-cross point position remains , , and the third encoding sequence of the offspring hyperparameter 2 satisfies (the non-cross point position remains ;

[0122] For the filter basis parameters of the hyperparameter individuals (let the filter basis of parent hyperparameter A be , and the filter basis of parent hyperparameter B be ): through a random selection mechanism (such as generating a 0-1 random number, if the number is ≤0.5, then inherit , otherwise inherit );

[0123] Through the above operation, one offspring hyperparameter individual is matched for each offspring architecture individual to form The third encoding sequence of the offspring parameter individual is changed in full correspondence with the change of the jump connection of the offspring architecture individual, so that the semantics of the structure and the parameters are coordinated.

[0124] In this embodiment, the step S4 is to perform a probabilistic mutation operation on the offspring synergistic individual pair, guide the associated adaptive mutation of the hyperparameter individual through the architecture individual mutation, ensure that the network structure change and the parameter configuration adjustment remain adaptive, avoid the decline of the model performance caused by the mismatch between the structure and the parameters, and provide candidate individuals with diversity and rationality for subsequent elite reservation and environmental selection; the following are the detailed steps:

[0125] Step S4-1: mutation probability initialization and parameter setting:

[0126] For the offspring synergistic individual pair (including the offspring architecture individual and the paired offspring hyperparameter individual) generated in step S3, two independent mutation probability parameters are preset: the architecture mutation probability and the hyperparameter mutation probability ; wherein the architecture mutation probability is used to control the probability of triggering mutation of the offspring architecture individual, and the hyperparameter mutation probability is used to control the probability of triggering mutation of the offspring hyperparameter individual in the non-associated mutation scenario, both of which have a value range of 0.01 to 0.05 (dynamically adjusted according to the population diversity demand, and the value can be appropriately increased when the initial population diversity is low);

[0127] At the same time, a predefined kernel type-dilated rate mapping relationship table is loaded, which records the allowed dilated rate parameter range corresponding to each dilated convolution kernel type, for example, the allowed dilated rate range corresponding to the standard 3x3 convolution kernel is 1 to 5, and the allowed dilated rate range corresponding to the 3x3 convolution kernel with expansion mode 1 is 2 to 6, so that the hyperparameter individual can quickly match the appropriate dilated rate parameter when the architecture individual occurs convolution kernel type mutation;

[0128] Step S4-2: mutation triggering and execution of the offspring architecture individual:

[0129] For each offspring architecture individual, a random number between 0 and 1 is generated , and is compared with the architecture mutation probability .

[0130] If , the mutation operation of the offspring architecture individual is triggered, and the mutation type is divided into two categories, one of which is randomly selected for execution.

[0131] The first type of mutation (jump connection mutation): a position in the first encoding sequence (binary encoding sequence) of the offspring architecture individual is randomly selected flip the binary value of the position (0 becomes 1, 1 becomes 0); after flipping, if the value of the position is 1, it indicates that a skip connection is added between the corresponding network layer and the previous specific layer; if it is 0, it indicates that the original skip connection is deleted;

[0132] The second type of variation (convolution kernel type variation): randomly select a position in the second encoding sequence (integer encoding sequence) of the offspring architecture individual , replace the original integer in the position with a randomly selected integer different from the current value; after replacement, the type of the dilated convolution kernel of the corresponding network layer is changed, for example, from a standard convolution kernel to a convolution kernel with a certain expansion mode;

[0133] If , the offspring architecture individual does not trigger variation, and the original encoding sequence remains unchanged;

[0134] The above judgment and operation are performed on all offspring architecture individuals one by one, and the relevant information of each offspring architecture individual that has undergone the second type of variation is recorded, including the variation position , the network layer number corresponding to the position (denoted as ), and the type of the dilated convolution kernel after variation (denoted as ) ;

[0135] Step S4-3: association of variation guide point marker and candidate value domain acquisition:

[0136] For the offspring architecture individual that has undergone the second type of variation in step S4-2, the network layer number corresponding to the variation position is marked as an associated variation guide point, which is used to locate the dilated rate parameter position in the hyperparameter individual that needs to be synchronously varied;

[0137] According to the type of the dilated convolution kernel after variation , the pre-defined kernel type-dilated rate mapping table is queried to extract the allowed dilated rate parameter range corresponding to the type of the convolution kernel, denoted as the candidate value domain (where is the minimum allowed dilated rate, is the maximum allowed dilated rate, and ); for example , when the 3x3 convolution kernel with expansion mode 2 is used, the query result is , indicating that the dilated rate parameter of the corresponding network layer needs to be selected between 3 and 7;

[0138] If the offspring architecture individual has not undergone the second type of variation (only the first type of variation or no variation), no associated variation guide point is marked, and the subsequent hyperparameter individual only determines whether to trigger independent variation according to the hyperparameter variation probability ;

[0139] Step S4-4: Execution of associated mutations in offspring hyperparameter individuals:

[0140] For each offspring hyperparameter individual, first determine whether its paired offspring architecture individuals are marked with associated mutation guide points:

[0141] If associated mutation guide points exist :

[0142] In the third coding sequence (real number coding sequence) of the offspring hyperparameter individual, the relationship between... Corresponding position This location stores the hole rate parameter of the network layer corresponding to the associated mutation guide point. ;

[0143] Adaptive strategy is used to select candidate value ranges Selecting new void ratio parameters The adaptive strategy is as follows: if the average hole rate of this type of convolutional kernel in the historical best model is... ,and ,but ;like Then in A random integer is generated within the range as... (The void ratio parameter is usually a positive integer);

[0144] use Replace the third coding sequence position place This completes the association mutation of hyperparameter individuals in the offspring;

[0145] If no associated mutation leader point exists:

[0146] Generate a random number between 0 and 1 ,Will With hyperparameter mutation probability Compare:

[0147] like Then, a position in the third encoded sequence is randomly selected. Add or subtract 1 from the original void ratio parameter at that location (ensuring that the value after the addition or subtraction is still within the predefined global void ratio range, usually 1 to 10);

[0148] like If the offspring hyperparameter individuals do not trigger mutations, the original coding sequence and filter cardinality parameters remain unchanged;

[0149] Step S4-5: Integration and verification of post-mutation cooperative individual pairs:

[0150] Recombine the offspring architecture individual after completing the variation with its counterpart, the offspring hyperparameter individual after completing the associated variation or independent variation, to form the cooperative individual pair after variation;

[0151] Perform a check on each cooperative individual pair after variation:

[0152] Check whether the first encoding sequence of the offspring architecture individual has a logical conflict, such as whether a calculation redundancy caused by the same network layer connecting multiple previous layers occurs, and if so, cancel the variation operation;

[0153] Check whether all hole rates in the third encoding sequence of the offspring hyperparameter individual are within the allowed range of the corresponding convolution kernel type (in combination with the kernel type-hole rate mapping relationship table), and if a parameter exceeds the range, re-execute the parameter selection at the position;

[0154] After the check passes, the cooperative individual pair after variation is retained; if the check fails, the state before the variation is restored, and a variation operation is triggered again (up to 3 retries, and if it still fails, the variation of the individual pair is abandoned); finally, all cooperative individual pairs after variation that pass the check form the offspring cooperative individual pair set, which is used for elite retention and environmental selection in the subsequent step S5.

[0155] In this embodiment, the role of step S5 is to perform a screening logic of “retaining performance first and then efficiency” according to the hierarchical fitness vector by merging the parent and offspring cooperative individual pairs, and retain the optimal individual pairs to form a new generation population, so as to ensure that the task performance does not degrade and the structural efficiency continuously improves in the population evolution process; the following are the detailed steps:

[0156] Step S5-1: Total set of cooperative individual pairs:

[0157] Merge the two types of cooperative individual pairs: the first type is the parent cooperative individual pairs of the current generation (i.e., the initial cooperative individual pair set before step S3 is performed); the second type is all offspring cooperative individual pairs generated after the variation based on the architecture guidance in step S4 through the competition and interaction within the subpopulation in step S3; after merging, a total set of cooperative individual pairs containing all candidate schemes is formed, denoted as ;

[0158] During the merging process, the system automatically checks the integrity of the individual pairs: ensures that each individual pair contains a complete architecture individual (first encoding sequence, second encoding sequence) and a hyperparameter individual (third encoding sequence, filter basis parameter), and there is no duplicate individual pair (judged by comparing the hash values of the encoding sequences); if there are duplicate or incomplete individual pairs, automatically remove the duplicates and supplement the missing items (if the offspring individual pairs are missing, re-call step S4 to generate supplements), and finally ensure that the total set contains all valid candidate schemes;

[0159] Step S5-2: First-level sorting (task performance priority):

[0160] Based on task performance fitness in the hierarchical fitness vector For the total set All collaborating individuals are sorted using the first-level sorting rule: by... Arranging the values ​​in descending order results in an ordered set. ;

[0161] During the sorting process, if there are two or more pairs of cooperating individuals... If the values ​​are completely equal, then the structural efficiency fitness factor will be used. The values ​​are sorted from high to low as an auxiliary sorting method (to avoid randomness caused by identical performance); after sorting, Individuals with higher rankings in the model have better performance on the corresponding task and can be given priority to enter the subsequent screening stage.

[0162] Step S5-3: Dynamic performance level classification:

[0163] based on The ranking results are used to dynamically classify the performance of all collaborative pairs, ensuring that the performance difference between pairs within the same level is less than a preset threshold; the specific process is as follows:

[0164] Preset performance difference threshold (The value ranges from 0.01 to 0.03, and can be adjusted according to the performance and accuracy requirements of the task. For example, for image classification tasks, a value of 0.01 is appropriate.) );

[0165] from Starting with the first pair of individuals, the absolute difference in task performance fitness between adjacent pairs of individuals is calculated sequentially using the following formula:

[0166] ,in, For the first The and the first The absolute difference in task performance fitness between collaborative individuals for The Middle Task performance fitness of each collaborative individual pair for The Middle Task performance fitness of each collaborative individual pair;

[0167] like Then the first Individuals are included in the first Individuals with the same performance level; if Then the first Each individual pair serves as the first individual pair for the new performance level;

[0168] Repeat steps 2 and 3 until... All individual pairs in the process are classified into different levels, resulting in several performance levels, denoted as follows: ( The total number of levels, (The highest performance level corresponding to the task);

[0169] Step S5-4: Second-level sorting within the hierarchy (structural efficiency optimization):

[0170] For each performance level For all cooperative pairs within the hierarchy, perform a second-level sort: based on the structural efficiency fitness in the hierarchical fitness vector. As the sole sorting criterion, by Arrange the values ​​in descending order to obtain ordered subsets of each level. ;

[0171] The purpose of the second-level sorting is to prioritize individuals with higher structural efficiency among pairs of individuals with similar task performance. That is, under the same task performance, the model with a larger average receptive field growth rate and less computational consumption ensures that the new generation of population has better operating efficiency while meeting the performance requirements.

[0172] Step S5-5: Selection of individuals in the next generation of the population:

[0173] Preset the number of individuals in the next generation population (Usually consistent with the initial population size, for example) To ensure population stability, the following rules apply to ordered subsets at each level. Individuals selected from the middle:

[0174] In order of performance level from highest to lowest (first) ,Again (and so on) Prioritize individual pairs;

[0175] In each level From The selection process continues sequentially, starting with the first individual pair, until all individual pairs within that level have been selected or the total number of selected individual pairs reaches [a certain threshold]. ;

[0176] If a certain level After the individual selections were completed, the total number of selected individuals still did not reach [the target]. Then continue from the next level. The same rules are applied to select individuals until the total number of selected individuals is exactly equal to the total number of selected individuals. ;

[0177] In the selection process, the system real-time statistics of the selected individual pairs, if the final selected number and the target number of the selected individual pairs have a deviation (such as less selection due to insufficient number of individuals in the grade), the system automatically selects the remaining individual pairs from the last grade of the current selection to ensure that the number of the new generation population is accurate and meets the requirements.

[0178] Step S5-6: New generation sub-population construction:

[0179] Step S5-5: Split the selected cooperative individual pairs in step S5-4:

[0180] Extract the architecture individuals in all individual pairs to form a new generation architecture sub-population (the number of individuals is );

[0181] Extract the hyperparameter individuals paired with the architecture individuals in all individual pairs to form a new generation hyperparameter sub-population (the number of individuals is also );

[0182] After splitting, the system automatically associates the pairing relationship between the individuals in the new generation architecture sub-population and the hyperparameter sub-population (maintaining the pairing logic of the original cooperative individual pairs), forming a complete new generation hierarchical cooperative evolution population; this population will serve as the parent population for step S6 iteration (if the number of evolution generations has not been reached), or as the input population for the final model generation (if the number of evolution generations has been reached).

[0183] In this embodiment, the role of step S6 is to optimize the population through multiple iterations, dynamically construct an experience model library and judge the evolution convergence, finally decode the network architecture and hyperparameter configuration from the optimal candidate model, after sufficient training and performance evaluation, encapsulate and generate the final hollow convolutional neural network model which can be directly used for inference tasks; the following are the detailed steps:

[0184] Step S6-1: Archive the optimal candidate model of the current generation:

[0185] From the new generation population generated in step S5, extract the hierarchical fitness vectors of all cooperative individual pairs, select the cooperative individual pairs with the highest task performance fitness value (if there are multiple individual pairs with the same fitness value, select the individual pair with the highest structure efficiency fitness );

[0186] ​Performing decoding operation on the optimal synergy individual pair screened out: determining network layer hop connection mode according to the first encoding sequence (binary sequence) of the architecture individual, determining the type of each layer's hole convolution kernel according to the second encoding sequence (integer sequence), determining each layer's hole rate parameter according to the third encoding sequence (real number sequence) of the hyperparameter individual, and determining the number of input and output channels of each layer according to the filter base parameter; reconstructing a complete hole convolutional neural network model according to the decoding result, marking the model as the optimal candidate model of the current generation, and storing it in the temporary archive area, while recording the corresponding hierarchical fitness vector and the number of generation iterations;

[0187] Step S6-2: empirical model library construction and evolution generation adjustment:

[0188] Combine the current generation optimal candidate model with all the optimal candidate models stored in the historical iteration process to construct a dynamically growing empirical model library; the empirical model library needs to save the complete information of each candidate model: including the first and second encoding sequences of the architecture individual, the third encoding sequence of the hyperparameter individual and the filter base parameter, the structure configuration required for model reconstruction, and the corresponding task performance fitness and structural efficiency fitness ;

[0189] Perform structural similarity analysis on all candidate models in the empirical model library: calculate the similarity of the encoding sequence of the architecture individual of any two models, and the similarity calculation formula is:

[0190] , wherein the structural similarity of the two models (the value range is 0 to 1, the closer to 1, the more similar the structure), the same number of bits is the total number of bits with the same value at the corresponding positions in the first and second encoding sequences of the two models, and the total length of the encoding sequence of model A and B is the sum of the lengths of the first and second encoding sequences;

[0191] Dynamically adjust the preset evolution generation according to the structural similarity analysis result: if the highest similarity of the newly added optimal candidate model in the last 3 generations and the existing model in the empirical model library , it indicates that the population evolution has tended to converge, and the preset evolution generation can be reduced by 20%; if the highest similarity of the newly added optimal candidate model in the last 3 generations and the existing model is , it indicates that the population diversity is still high, and the preset evolution generation can be increased by 10%; otherwise, the preset evolution generation remains unchanged;

[0192] Step S6-3: iteration termination judgment:

[0193] Statistically count the total number of the current evolution iteration, denoted as , and compare it with the preset evolution generation adjusted in step S6-2 The comparison is performed while simultaneously determining whether other preset convergence conditions are met.

[0194] Algebraic condition: if If , then the algebraic termination condition is satisfied;

[0195] Performance convergence condition: Calculate the change in task performance fitness of the best candidate models in the empirical model library for three consecutive generations. The calculation formula is as follows:

[0196] ,in, This represents the average change in performance fitness across three consecutive generations of tasks. For the first Task performance fitness of the optimal candidate model. For the first Task performance fitness of the optimal candidate model. For the first The task performance fitness of the optimal candidate model; if (This threshold is set according to the accuracy requirements of the task), then the performance convergence condition is met;

[0197] If both the algebraic condition and the performance convergence condition are met, the evolution process is terminated; if either condition is not met, the new generation population generated in step S5 is used as the parent population for the next iteration, and the process returns to step S2 to re-execute the hierarchical cooperative fitness evaluation and subsequent operations until the iteration termination condition is met.

[0198] Step S6-4: Determining the optimal network architecture and hyperparameter configuration:

[0199] Once the evolutionary process terminates, the task performance fitness of all candidate models is extracted from the empirical model library. Filter out The candidate model with the highest numerical value (if multiple candidates exist) The same model, selecting structural efficiency fitness The model with the highest score is marked as the global optimal candidate model.

[0200] Perform decoding operations on the globally optimal candidate model:

[0201] Analysis of architecture configuration: Extract the effective states of skip connections of each network layer from the first encoding sequence, extract the dilated convolution kernel type of each layer from the second encoding sequence, and form a complete network architecture topology;

[0202] Analysis of hyperparameter configuration: The hole rate parameters of each layer are extracted from the third coding sequence, and the baseline number of channels of the network as a whole is determined from the filter cardinality parameters to form a complete hyperparameter configuration scheme; The network architecture topology obtained by analysis is associated and stored with the hyperparameter configuration scheme as the basis for the final model construction;

[0203] Step S6-5: Full training of the final model:

[0204] Based on the network architecture topology and hyperparameter configuration scheme determined in step S6-4, a complete convolutional neural network model is re-instantiated; the full training dataset (different from the training data subset in step S2) is used to perform full training on the model, and the training process follows the following rules:

[0205] Optimizer selection: a preset adaptive optimizer (such as Adam optimizer) is used, with an initial learning rate of (0.001 to 0.01); learning rate scheduling: a dynamic learning rate scheduling strategy (such as cosine annealing scheduling) is used, and the learning rate update formula is:

[0206] wherein, is the learning rate of the th training round, is the minimum learning rate (usually of is the total number of training rounds;

[0207] Training termination condition: when the validation set loss function value does not decrease (or the decrease amplitude is less than 1e-5) for 10 consecutive rounds, the training is terminated to ensure that the model is fully converged and overfitting is avoided;

[0208] Step S6-6: Performance evaluation of the final model:

[0209] The fully trained model is loaded into an independent test set (test set data does not participate in the training and validation process), forward propagation inference is performed, and the final performance indicators of the model are calculated:

[0210] Task performance indicators: calculate the test set classification accuracy , precision , and recall , wherein the classification accuracy calculation formula is:

[0211] wherein, is the test set classification accuracy, is the number of samples correctly predicted by the model in the test set, is the total number of test set samples;

[0212] Structural efficiency indicators: recalculate the total computational amount consumed by the model in a single forward propagation and the average receptive field growth rate to verify whether the model structural efficiency meets the expectations;

[0213] All performance indicators are summarized to form a final performance report. If the performance indicators do not meet the preset target (such as ), return to step S6-4 to reselect a suboptimal candidate model in the experience model library, repeat steps S6-5 to S6-6 until the performance meets the standard;

[0214] Step S6-7: final model packaging

[0215] Perform a packaging operation on the performance-meeting hollow convolutional neural network model:

[0216] Store the core parameters: save the structure parameters of the model (including the number of network layers, the mode of skip connection, the type of convolution kernel and the parameters of the hollow rate of each layer, the filter base), and the trained weight parameters (the weight of each layer of convolution kernel, the bias term);

[0217] Additional performance information: store the final performance report (including classification accuracy, computational load consumption, etc.) generated in step S6-6 in association with the model parameters;

[0218] Generate an inference file: package the above data in a general deep learning model format (such as ONNX format) to ensure that the model can be directly loaded and run in different inference frameworks (such as TensorRT, ONNX Runtime) without the need for retraining or parameter configuration;

[0219] After packaging is completed, the final hollow convolutional neural network model file is output, which can be directly used for inference calculation of the target task (such as image segmentation, target detection, etc.).

[0220] A hollow convolutional neural network based on a genetic algorithm, the hollow convolutional neural network is optimized by an optimization method, the hollow convolutional neural network is a deep learning model optimized by a hierarchical collaborative evolution genetic algorithm, the core feature is the collaborative adaptation of architecture and hyperparameters, and the hollow convolutional neural network has high task performance and efficient structure and can be directly applied to computer vision inference tasks such as image segmentation and target detection; the detailed description is as follows:

[0221] I. Optimization association and core design logic of network

[0222] All structural configurations and parameter settings of the hollow convolutional neural network are determined by the optimization method of steps S1 to S6; the optimization process is based on a hierarchical collaborative evolution population, through the competitive interaction and collaborative variation of the architecture sub-population and the hyperparameter sub-population, and the elite reservation, the collaborative individual pair with the optimal task performance fitness and structural efficiency fitness is screened out; the final architecture topology and hyperparameter configuration of the network are obtained after decoding the individual pair, and then the network model directly available is formed through full training data set sufficient training and independent test set performance verification; the core design logic is to realize the dynamic collaboration of network structure and parameters, and to avoid the performance bottleneck or computational redundancy caused by the mismatch between structure and parameters in traditional manual design;

[0223] II. Core structure composition:

[0224] Architecture configuration (based on architecture sub-population encoding and decoding):

[0225] The network architecture is defined by the first encoding sequence and the second encoding sequence to form a fixed and efficient topology:

[0226] Jump connection mode: the first encoding sequence is a binary sequence, each bit corresponds to a network layer, and the value 1 indicates that there is an effective jump connection between the current network layer and the previous specific layer, and the value 0 indicates no jump connection; the existence of jump connection enables the direct transmission and fusion of shallow features and deep features, reduces the risk of gradient disappearance, and improves the feature expression ability;

[0227] Hollow convolution kernel type configuration: the second encoding sequence is an integer sequence, each bit corresponds to the convolution kernel type of a network layer, and all types are selected from a predefined candidate kernel set; the candidate kernel set contains standard convolution kernels and various dilated mode hollow convolution kernels, and different types of convolution kernels adapt to the feature extraction needs of different scenarios, such as standard convolution kernels for fine-grained feature capture and dilated mode convolution kernels for expanding the receptive field;

[0228] Hyperparameter configuration (based on hyperparameter sub-population encoding and decoding):

[0229] Hyperparameters and architecture configurations are one-to-one, ensuring that the parameters of each layer are adapted to its structural design:

[0230] Hollow rate parameter sequence: the third encoding sequence is a real number sequence, the length is consistent with the total number of network layers, and each bit value is the hollow rate parameter of the corresponding network layer; the hollow rate parameter determines the dilated degree of the convolution kernel, and by controlling this parameter, the network's ability to capture local details and global information can be adjusted, and the larger the value, the wider the receptive field coverage;

[0231] Filter base parameter: an independent real number coding corresponding to the filter base parameter, providing a benchmark for the input and output channel number of each layer of the network; the channel number of each layer is derived from the base parameter according to a pre-set ratio, ensuring the continuity of feature transmission and avoiding the redundancy of calculation caused by excessive channel number;

[0232] III. Hierarchy and functional division:

[0233] The network is divided into five functional levels according to the feature processing flow, and the structure and parameters of each level are determined by the optimized configuration:

[0234] Input layer: receives the original input data (such as image data) and converts the data into a feature map format that the network can process; the channel number of the feature map is consistent with the input channel number derived from the filter base parameter; the input layer does not perform convolution operation, but only responsible for data format conversion and dimension adaptation, ensuring that the subsequent levels can be processed normally;

[0235] Cave convolution layer: the core feature extraction level of the network, the number of which is consistent with the total number of network layers in the architecture configuration; each layer is equipped with corresponding cave convolution kernel type and cave rate parameter, which extracts features and converts dimensions from the input feature map through convolution operation; during the convolution process, the cave convolution kernel expands the receptive field without increasing the number of parameters, while preserving the feature map resolution, realizing the synchronous capture of local details and global context information;

[0236] Skip connection layer: alternately cooperates with the cave convolution layer, establishes direct connection between specified network layers according to the binary configuration of the first coding sequence; this layer directly transmits the fine-grained features extracted by the shallow cave convolution layer to the deep feature fusion layer, supplements the missing details in the deep features, and accelerates the gradient backpropagation, improving the stability and convergence speed of network training;

[0237] Feature fusion layer: receives the feature maps output by each cave convolution layer and the feature maps transmitted by the skip connection, realizes multi-scale feature fusion through element addition or channel splicing; during the fusion process, the features of different levels complement each other, preserving the position information and detail features of the shallow layer, integrating the semantic information and global features of the deep layer, and improving the adaptability of the network to complex scenes;

[0238] Output layer: designed according to specific task requirements (such as pixel classification for segmentation task or target box regression for detection task), receives the fusion feature map output by the feature fusion layer, and outputs the task result through the fully connected layer or convolution layer; the number of neurons or channels of the output layer adapts to the dimension of the task target, ensuring that the final output format meets the requirements of inference application;

[0239] IV. Feature extraction and inference mechanism:

[0240] Feature extraction process:

[0241] After the input data is converted into a feature map by the input layer, it enters each hollow convolutional layer in turn to perform feature extraction:

[0242] Each hollow convolutional layer operates according to the configured convolution kernel type and the hollow rate parameter. When the convolution kernel slides on the feature map, it only performs weighted summation with the pixels at the corresponding position to generate the feature response map of the layer. The hollow rate parameter determines the stride interval of the convolution kernel sliding. The larger the value, the larger the interval, and the wider the receptive field coverage;

[0243] The receptive field growth follows the recursive multiplication rule. The initial receptive field growth step of the first layer network is the basic size of the convolution kernel of the layer. Starting from the second layer, the receptive field growth step of the i-th layer is equal to the product of the receptive field growth step of the i-1-th layer and the hollow rate parameter of the i-th layer. The average value of the receptive field growth steps of all network layers is the average receptive field growth rate, which directly affects the network's ability to capture global information;

[0244] The skip connection layer directly transmits the feature map of the specified level to the feature fusion layer according to the configuration, synchronously enters the fusion process with the feature map of the subsequent level, and forms multi-scale fusion features;

[0245] Inference process:

[0246] The network performs single forward propagation in the inference stage, and the process is efficient and the computational complexity is controllable:

[0247] The fusion feature map enters the output layer, generates the task result after targeted operation, and the whole process does not need back propagation and parameter update;

[0248] The computational complexity consumption of forward propagation is composed of the convolution operation amount of each layer and the computational complexity increment of the skip connection. The single layer convolution operation amount is related to the square of the basic size of the convolution kernel, the input channel number, the output channel number, and the feature map size. The computational complexity increment of the skip connection is equal to the number of elements of the output feature map of the current layer. The optimized network is screened by the structural efficiency fitness to ensure that the computational complexity consumption in the inference process is within a reasonable range, taking into account the running speed and hardware adaptability.

[0249] Although embodiments of the present application have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A dilated convolutional neural network optimization method based on genetic algorithm, characterized in that: Includes the following steps: Step S1: Construct a hierarchical co-evolutionary population, including an architecture subpopulation and a hyperparameter subpopulation using different encoding methods. Randomly initialize the two subpopulations and pair them to form initial co-evolutionary pairs. Specifically, this includes: The architecture subpopulation and hyperparameter subpopulation are initialized in parallel. The architecture subpopulation contains a first preset number of architecture individuals, and the hyperparameter subpopulation contains the same number of hyperparameter individuals. For each individual architecture in the architecture subpopulation, the network structure is represented by a first encoding sequence and a second encoding sequence. The first encoding sequence is a binary encoding sequence, where each bit represents whether the skip connection between the current network layer and the previous specific layer is valid. The second encoding sequence is an integer encoding sequence, where each bit represents the type of dilated convolution kernel used by the corresponding network layer. The dilated convolution kernel type is selected from a predefined candidate kernel set that includes standard convolution kernels and multiple dilation modes. For each hyperparameter individual in the hyperparameter subpopulation, its coding structure is generated by the architecture individual paired with it. First, the length of the third coding sequence of the hyperparameter individual is determined according to the total number of network layers of the paired architecture individual. The third coding sequence is a real number coding sequence, and each bit of it represents the hole rate parameter of the corresponding network layer in the paired architecture individual. Then, an independent real number code is generated to represent the cardinality parameter of the network filter corresponding to the paired architecture individual. After completing the above encoding, each architecture individual in the architecture subpopulation is randomly paired with a hyperparameter individual in the hyperparameter subpopulation, so that the structural information of each architecture individual is combined with the parameter information of a hyperparameter individual to form a complete initial cooperative individual pair to represent a dilated convolutional neural network candidate model. Step S2: Hierarchical Cooperative Fitness Evaluation: For each cooperative individual pair, decode and quickly train the network, and calculate a hierarchical fitness vector that includes task performance fitness and structural efficiency fitness, where structural efficiency fitness is determined by the average receptive field growth rate and computational cost of the network. Step S3: Competition and interaction within subpopulations: Perform fitness-based competitive selection within each of the two subpopulations, followed by architecture-guided directional crossover to ensure that structural changes and parameter changes are coordinated; Step S4: Architecture-guided cooperative mutation: Perform probabilistic mutation on offspring individuals. When an architecture individual mutates, force its paired hyperparameter individuals to undergo associative adaptive mutation at the corresponding position. Step S5: Elite Preservation and Environment Selection: Merge all cooperative individual pairs between the parent and offspring generations, and based on the hierarchical fitness vector, adopt a mechanism of first sorting by task performance and then filtering by structural efficiency to retain a preset number of optimal individual pairs; Step S6: Iteration and Model Generation: Repeat steps S2 to S5 until the preset number of generations is reached. Select the cooperative individual pair with the highest task performance fitness, decode and fully train to obtain the final dilated convolutional neural network model.

2. The dilated convolutional neural network optimization method based on genetic algorithm according to claim 1, characterized in that: The hierarchical collaborative fitness assessment in step S2 specifically includes: Decode the cooperative individual pairs, and construct the corresponding dilated convolutional neural network model based on the first and second encoding sequences of the architecture individuals in the cooperative individual pairs, the third encoding sequence of the hyperparameter individuals, and the filter cardinality parameters; The dilated convolutional neural network model is quickly trained using a subset of training data to obtain a preliminary trained model. The initially trained model is then validated by forward propagation on the validation set to obtain the classification accuracy of the validation set. Based on the classification accuracy of the validation set, the task performance fitness is calculated through linear normalization mapping; Based on the hole rate parameter sequence represented by the third encoding sequence of the hyperparameter individual, the receptive field growth step of each network layer is dynamically calculated by recursive multiplication, and the average receptive field growth rate of all network layers is statistically analyzed. Based on the first and second encoding sequences of the architecture individuals and the filter cardinality parameters of the hyperparameter individuals, the computational cost of the dilated convolutional neural network model in a single forward propagation is evaluated. By combining the average receptive field growth rate and computational cost, a weighted inverse fusion is performed using preset weighting coefficients to calculate the structural efficiency fitness. The task performance fitness and structural efficiency fitness are combined in a preset order to form a hierarchical fitness vector, which is used for subsequent competition and interaction within subpopulations.

3. The dilated convolutional neural network optimization method based on genetic algorithm according to claim 1, characterized in that: The competition and interaction within the subpopulation in step S3 specifically includes: Based on the hierarchical fitness vector of each collaborative individual pair, all architecture individuals are first sorted according to task performance fitness. Then, according to the preset selection ratio, the ranking-based roulette wheel selection method is used to select the first number of dominant architecture individuals from the architecture subpopulation. Based on the paired hyperparameter individuals corresponding to the selected dominant architecture individuals, the individuals are sorted according to the structural efficiency fitness in the hierarchical fitness vector corresponding to the hyperparameter individuals. The same ranking-based roulette wheel selection method is used to select the same number of dominant hyperparameter individuals from the hyperparameter subpopulation. For each temporary pairing formed by the selected dominant architecture individual and the dominant hyperparameter individual, the skip connection pattern represented by the first encoding sequence of the dominant architecture individual is analyzed; the network layer positions where the skip connections change are identified, and these positions are set as the forced crossover points for the third encoding sequence of the hyperparameter individual in subsequent crossover operations; Perform a single-point crossover operation between dominant architecture individuals to generate child architecture individuals; determine the jump connection change layer inherited from the parent based on the jump connection pattern of the child architecture individuals; perform a multi-point crossover operation between the corresponding dominant hyperparameter individuals, wherein the crossover point is forcibly set at the forced crossover point to generate child hyperparameter individuals that are semantically coordinated with the child architecture individuals in terms of structural changes.

4. The dilated convolutional neural network optimization method based on genetic algorithm according to claim 1, characterized in that: The architecture-guided collaborative mutation in step S4 specifically includes: For the offspring architecture individuals generated in step S3 and the offspring hyperparameter individuals they are paired with, separate architecture mutation probabilities and hyperparameter mutation probabilities are set for each pair of offspring cooperative individuals. For individual child architectures, mutations are triggered by the probability of architectural mutation. The mutation operations include a first type of mutation and a second type of mutation. The first type of mutation involves flipping the binary bits at a randomly selected position in the first encoding sequence to add or delete skip connections. The second type of mutation involves replacing the integer value at a randomly selected position in the second encoding sequence and reselecting the dilated convolution kernel type of the current network layer from a predefined set of candidate kernels. When a second type of mutation occurs in a child architecture individual, the system records the specific network layer location where the dilated convolution kernel type switch occurs and marks this location as the associated mutation initiation point; at the same time, based on the dilated convolution kernel type after the switch, the system obtains the candidate value range of the dilation rate parameter allowed by the corresponding network layer from the preset kernel type-dilation rate mapping relationship. Based on the associated mutation guide point, the associated mutation of the paired offspring hyperparameter individuals is forcibly triggered; in the third encoding sequence of the offspring hyperparameter individuals, the position corresponding to the associated mutation guide point is located, and its original hole rate parameter value is replaced with a value reselected from the candidate value range of the hole rate parameter according to the preset adaptive strategy, so as to adapt to the change of hole convolution kernel type in the offspring architecture individuals. After the above mutation operations are completed, the offspring architecture individuals and the offspring hyperparameter individuals that have undergone association adaptive mutation together constitute a mutated cooperative individual pair.

5. The dilated convolutional neural network optimization method based on genetic algorithm according to claim 1, characterized in that: The elite retention and environment selection in step S5 specifically include: All the offspring cooperative individual pairs generated after the competition and interaction within the subpopulation in step S3 and the architecture-guided cooperative mutation in step S4 are merged with the parent cooperative individual pairs of the current generation to form a total set of cooperative individual pairs containing all candidate schemes. Based on the hierarchical fitness vector, all collaborative individual pairs in the total set are sorted in the first level. The sorting in the first level is based on the task performance fitness corresponding to the collaborative individual pair, and they are arranged in descending order. Based on the results of the first-level sorting, the total set of the sorted collaborative individuals is dynamically divided into performance levels. During the division, the absolute difference in task performance fitness between adjacent collaborative individuals is calculated. When the absolute difference is less than the preset performance difference threshold, the adjacent collaborative individuals are classified into the same performance level. Within each performance level, all cooperative pairs belonging to that level are sorted in a second-level order based on the structural efficiency fitness of the cooperative pairs, arranged in descending order. Based on the preset number of the new generation population, in descending order of performance level, select the top-ranked cooperative individual pairs in the second-level sorting from each performance level until the total number of selected cooperative individual pairs reaches the number of the new generation population. The selected collaborative pairs will form a new generation of architecture subpopulations and hyperparameter subpopulations, which will be used in the next iteration of step S6 or as input for the final model generation.

6. The dilated convolutional neural network optimization method based on genetic algorithm according to claim 1, characterized in that: Step S6 includes the following specific steps: Extract the cooperative individual pairs with the highest task performance fitness in the current iteration from the new generation population generated in step S5, decode and reconstruct them into a complete dilated convolutional neural network model, and archive them as the best candidate model for the current generation. By utilizing the current generation's best candidate models and the historical archive's best candidate models, a dynamically growing empirical model library is constructed. Structural similarity analysis is performed on all models in the empirical model library, and the preset number of evolutionary generations is dynamically adjusted based on the analysis results; Determine whether the current evolutionary generation has reached the dynamically adjusted preset evolutionary generation or meets other preset convergence conditions; If the target is not reached and convergence is not achieved, the new generation population will be used as the parent population for the next iteration, and the process will return to step S2 for the next iteration. The evolution process terminates when the evolution reaches the preset number of generations or meets the convergence condition. From the final empirical model library, select the candidate model with the highest task performance fitness, and decode it to obtain its corresponding optimal network architecture and hyperparameter configuration; Based on the optimal network architecture and hyperparameter configuration, a complete dilated convolutional neural network is re-instantiated. The network is fully trained using the complete training dataset and a pre-defined optimizer and a dynamically adjusted learning rate scheduling strategy. The performance of the fully trained network is evaluated on an independent test set to obtain the final performance metric. The network's structural parameters, trained weight parameters, and final performance metric are then encapsulated to generate the final dilated convolutional neural network model that can be directly used for inference tasks.

7. A dilated convolutional neural network based on a genetic algorithm, characterized in that: The dilated convolutional neural network is optimized by the optimization method described in any one of claims 1-6.

Images