A multi-objective adaptive evolutionary neural architecture search method for generative adversarial networks

Abstract

The application belongs to the technical field of automatic design of generative adversarial network, and discloses a multi-target adaptive evolutionary neural architecture search method for generative adversarial network. The evolutionary process is processed in stages according to the neural architecture search, and the evolutionary process is divided into three stages of exploration, development and convergence. A controller is designed, so that the neural architecture search can be adjusted according to the evolutionary process, and the performance of the neural architecture search is effectively improved. Through the newly designed probability function, the crossover probability and mutation probability of the evolutionary process are controlled in different stages, and the convergence is accelerated in the later evolutionary stage, and a large number of ablation experiments are carried out. The evolutionary process based on the double-factor cooperative variation mechanism designed by the application dynamically adjusts and controls the neural architecture search process.

Description

Technical Field

[0001] This invention relates to the field of automated design technology for generative adversarial networks (GANs), and more particularly to a multi-objective adaptive evolutionary neural architecture search method for GANs. Background Technology

[0002] Current deep learning network architectures enable computers to understand images better than humans, perceiving depth information within them. However, designing these networks is extremely difficult. In current experiments, most networks are designed based on researchers' experience and intuition. Inappropriate or redundant network structures can lead to a large number of model parameters or slow training, and more seriously, problems such as vanishing gradients, exploding gradients, pattern collapse, and checkerboarding during training. Therefore, network design requires a significant investment of researchers' time and effort, and even well-designed networks may not necessarily perform well. Since 2016, neural architecture search methods have emerged, automating the design of neural network structures and allowing researchers of deep neural networks to obtain a high-performing network architecture without relying on intuition and experience. The neural architecture search process involves researchers or humans defining and encoding a search space, then sampling a candidate neural network architecture from it, and evaluating it through downstream tasks. This evaluation strategy is also pre-defined, such as comparison or non-dominated ranking. After obtaining the performance scores and feedback of the candidate network architectures, the network is resampled and evaluated according to a certain search strategy, and this process is repeated until the optimal neural network architecture is obtained.

[0003] Generative Adversarial Networks (GANs) have repeatedly broken records in tasks such as image recognition, image generation, super-resolution, motion deblurring, and style transfer. A GAN is trained collaboratively by two different network architectures: a generator (G) that learns the distribution of real data and fits the deep semantic information within it; and a discriminator (D) that distinguishes between real data and fake data generated by the G network, feeding the result back to the generator for its training. This data generation process is a fusion of self-supervised learning and zero-sum game theory. Because the optimization objectives of the G and D networks are inconsistent, manually designing a balanced G and D network is extremely difficult. In this unbalanced design, the generator may become too powerful, unable to be updated by the discriminator's feedback, thus failing to learn the real data distribution effectively, leading to pattern collapse and convergence difficulties.

[0004] However, existing neural architecture search methods for Generative Adversarial Networks (GANs) include gradient-based, reinforcement learning, and evolutionary optimization methods. Most of these methods select network architectures based on a single objective, leading to a situation where the target network performs well on one evaluation metric but only moderately or poorly on another. Therefore, utilizing multi-objective neural architecture search algorithms is necessary. However, evolutionary algorithms heavily rely on random operations, or rather, randomizing crossover and mutation processes. While this aligns with the heuristic process of evolutionary algorithms, it is fatal for GANs. Due to the vast neural architecture search space and numerous local optima, and the fact that the D network can only influence the G network by distinguishing between true and false states, current evolutionary NAS is prone to entering local optima. Furthermore, GANs with poor global search capabilities or those trapped in local optima suffer from pattern collapse. However, exploring the entire search space is impossible. This means that we must not only explore the global space as much as possible within a limited number of search rounds, but also prevent excessive exploration from causing convergence difficulties later on. By designing an evolutionary algorithm based on a two-factor cooperative mutation mechanism, the crossover and mutation probabilities of the evolutionary process are dynamically adjusted at different stages, enabling it to adaptively adjust its global exploration capabilities while simultaneously developing local information. In the later stages of evolution, this accelerates population convergence and helps the population escape local optima. Furthermore, the designed controller integrates the neural architecture search process for generative adversarial networks (GANs) with multi-objective optimization methods, resulting in a GAN that performs well across multiple objectives. Summary of the Invention

[0005] The objective of this invention is to improve evolutionary algorithms. Based on neural architecture search, this invention divides the evolutionary process into three stages: exploration, development, and convergence. A controller is designed to allow the neural architecture search to adjust according to the evolutionary process, effectively improving its performance. This invention designs an evolutionary algorithm based on a two-factor cooperative mutation mechanism. Through a newly designed probability function, the crossover and mutation probabilities of the evolutionary process are controlled at different stages, effectively improving global exploration and local development capabilities while accelerating the algorithm's optimization speed. This allows the algorithm to escape local optima and accelerate convergence in the later stages of evolution. Extensive ablation experiments have been conducted. This invention also involves multi-objective evolutionary neural architecture search and generative adversarial networks (GANs), combining the two to automate the design of GANs. The neural architecture search process is dynamically adjusted and controlled through the two-factor cooperative mutation mechanism-based evolutionary process designed in this invention.

[0006] The technical solution of the present invention is as follows: a multi-objective adaptive evolutionary neural architecture search method for generative adversarial networks, the specific steps of which are as follows;

[0007] Step 1: Construct a two-factor cooperative mutation mechanism;

[0008] First, the scores of candidate individuals in the image generation task are used as information such as fitness and dispersion to construct a population fitness evaluation factor. Second, the current generation and total number of iterations are used to construct a population variation adjustment factor. Finally, the two factors are combined to construct a crossover and mutation probability control function based on a two-factor collaborative variation mechanism.

[0009] S1.1 Population suitability assessment factor;

[0010] The population wellness assessment factor influences the evolutionary direction of a population by evaluating its wellness. Each individual in the population participates in controlling changes to the population wellness assessment factor. The specific calculation formula is as follows:

[0011]

[0012] Among them, F fit Used to represent the designed population adequateity evaluation factor; f i Indicates the appropriate level of individual i in the population, f avg Indicates the average fitness of the population, f max Indicates the maximum adequate population, f min Indicates the minimum adequate level of a population; This represents the total number of new individuals generated in each generation of a given population; population dispersion and fitness index are obtained through population fitness evaluation factors.

[0013] S1.2 Population variation adjustment factor;

[0014] The population variation adjustment factor is a factor that controls the direction of population evolution based on the current iteration round. The specific calculation formula is as follows:

[0015]

[0016] Where n is determined by the current evolutionary cycle; s represents the given overall evolutionary cycle; C is a hyperparameter used to adjust the population variation adjustment factor;

[0017] S1.3 Establish the evolutionary process based on the two-factor collaborative mutation mechanism;

[0018] By combining population fitness evaluation factors with population variation adjustment factors, an evolutionary process based on a two-factor collaborative variation mechanism is constructed. This evolutionary process is influenced by population fitness, population dispersion, and population generation index, and exhibits adaptability. The following probability function is established based on the sigmoid function:

[0019]

[0020] Among them, w i Let j be a constant, and j be a, b, c; based on the different values, we obtain a family of curves about the probability function P, where -20x+10 represents the normalization process;

[0021] Based on the process and expectations of neural architecture search, the evolutionary process is divided into three stages: exploration, development, and convergence.

[0022]

[0023] threshold1, threshold2, and threshold3 represent the evolutionary thresholds for the exploration, development, and convergence stages, respectively; m1 and m2 are set values ​​used to segment the entire evolutionary process; w a and w c Less than 1, w b Greater than 1; n1 and n2 are hyperparameters, and k is a constant;

[0024] Step 2: Construct a multi-stage adaptive evolutionary process based on a two-factor collaborative mutation mechanism;

[0025] For population P t Calculate the fitness of each population and substitute it into formula (1) to calculate F. fit Substitute the current generation number n into formula (2) to calculate F. acc According to the obtained F fit With F acc The current evolutionary stage is determined using formula (4), and the current threshold is calculated. Each individual in the current population is assigned a random number between 0 and 1. If the random number exceeds the current evolutionary threshold, the individual is considered to be undergoing a crossover operation; if the random number is below the current threshold, a mutation operation is considered. After all individuals in the population have completed their evolutionary operations, they are merged with their parent populations to form a new population. Then, a fast non-dominated sort is performed on the population containing all individuals before and after evolution, and the number of individuals generated in each generation is used as the basis for the sorting. Select the next generation population P t+1 ;

[0026] Step 3: Construct a multi-objective evolutionary neural architecture search framework for generative adversarial networks;

[0027] S3.1 Determine multiple optimization objectives for generative adversarial networks: Establish multiple optimization objectives based on image generation tasks;

[0028] Simultaneously considering the IS evaluation metric, FID evaluation metric, and model parameter quantity (Params) evaluation metric, a multi-objective evolutionary neural architecture search framework based on IS, FID, and model parameter quantity is proposed; the specific formula is as follows:

[0029]

[0030] A higher IS score indicates a better performance of the generative adversarial network, while a lower FID and model parameter count indicate better performance of the generative adversarial network.

[0031] Image Classification (IS) evaluates the quality of generated images using an image category classifier. The specific formula for IS is as follows:

[0032] IS(G) = exp(E x～Pg D KL (p(y∣x)||p(y))) (6)

[0033] Among them, E x～Pg D represents the average value obtained by iterating through all generated samples. KL This represents the KL divergence, which measures the distance between two distributions.

[0034] The specific formula for FID is as follows:

[0035] FID = ||μ r -μ g || 2 +T r (∑r+∑g)-2(ΣrΣg) 1 / 2 (7)

[0036] Where, μ r and μ g Let be the feature mean of the real image and the feature mean of the generated image, respectively; ∑r and ∑g are the covariance matrices of the real image and the generated image, respectively; and T is the feature mean of the generated image. r It is a trace;

[0037] S3.2 Construct the search space;

[0038] Based on AdversarialNAS, the search space is fixed at four nodes for both SuperNet-G and SuperNet-D. Noise samples on the SuperNet-G network pass through a fully connected layer and then through three upsampling layers to obtain generated samples, which are then passed to the SuperNet-D network. On the SuperNet-D network, the generated samples and the real images are compared through three downsampling layers and then through a fully connected layer.

[0039] S3.3 Constructing a multi-objective adaptive evolutionary neural architecture search process based on dual-factor collaborative control;

[0040] By employing a weight-sharing strategy, the generator and discriminator inherit weights from the SuperNet-G and SuperNet-D networks, respectively, and the search process is divided into steps. A controller is proposed to control the search process of the generator and discriminator, both of which follow a multi-stage evolutionary process based on a two-factor cooperative mutation mechanism. Fast non-dominated sorting is used to obtain a set of Pareto front solutions under multiple set objectives.

[0041] The first stage uses a fixed discriminator to search for a generator and adopts a many-to-one training strategy; the second stage uses the optimal generator found in the first stage to search for a discriminator and adopts a one-to-one training weight reset strategy to enhance the stability of the generative adversarial network training.

[0042] The fixed discriminator, during the many-to-one training strategy, generates [a certain value] in each round. Several candidate generators are selected and their weights are shared using a weight-sharing strategy. The sampled generators are then subjected to adversarial training against an identical discriminator to obtain the optimal generator network. After obtaining the optimal generator, its weights are copied and labeled as G. * Then, it is trained one-to-one with different discriminators, and the weight reset strategy is started before the start of each round with the optimal generator network weight as the benchmark.

[0043] The beneficial effects of this invention are as follows: This invention proposes a population fitness-based population suitability evaluation factor and a population variation adjustment factor based on the current iteration number, and organically combines them through a newly designed probability function to form a two-factor collaborative control mechanism. Based on the neural architecture search process, this invention divides the evolutionary process into three stages: exploration, development, and convergence. At different stages, the neural architecture search possesses search capabilities based on the current search state, enabling it to adjust according to the evolutionary process and exhibit a certain degree of adaptability. This invention designs a controller to apply the above evolutionary process to the automated design of generative adversarial networks (GANs) in image generation tasks. STL-10 and CIFAR-10 achieved excellent performance, surpassing most similar methods, and numerous ablation experiments have demonstrated the effectiveness of this invention. Attached Figure Description

[0044] Figure 1 This is a technical roadmap for the present invention.

[0045] Figures 2-5 This is a schematic diagram of population distribution at different evolutionary stages. Figure 2 This indicates the first stage of the exploration phase. Figure 3 This indicates the second stage of the exploration phase; Figure 4 This indicates the first stage of development. Figure 5 This indicates the second stage of development.

[0046] Figure 6 The search space is described; the upsampling and downsampling refer to the network structure in AdversarialNAS.

[0047] Figure 7 This is a search framework for multi-objective neural architectures based on two-factor collaborative control.

[0048] Figure 8 This refers to the G network search process.

[0049] Figure 9 The results are from experiments with no factor control and single factor control. Detailed Implementation

[0050] A multi-objective adaptive evolutionary neural architecture search method for generative adversarial networks (GANs) is proposed. By constructing a two-factor co-mutation mechanism, the evolutionary process is endowed with adaptive capabilities, allowing it to dynamically adjust crossover and mutation probabilities based on factors such as population state, iteration rounds, and artificially imposed development expectations. The evolutionary process is divided into three stages—exploration, development, and convergence—based on a modular neural architecture search process. The two-factor co-mutation mechanism enhances the global exploration and local development capabilities of the neural architecture search process, as well as accelerating convergence in the later stages of evolution. This enables the method to adaptively adjust the global exploration capability of the evolutionary neural architecture search while simultaneously developing local information. In the later stages of evolution, it can accelerate population convergence and escape locally optimal states. Finally, by constructing a multi-objective optimization task for image generation and a search space for GANs, a multi-objective evolutionary neural architecture search framework based on the two-factor co-mutation mechanism is established. This enables the automated design of GANs and their completion of image generation tasks. This achieves the automated design of GAN architectures and enables them to perform computer vision tasks such as image generation.

[0051] See Figure 1 The present invention provides a multi-objective neural architecture search based on a two-factor collaborative mutation mechanism, comprising the following steps:

[0052] S1 constructs a two-factor co-variation mechanism

[0053] For the modular neural architecture search process, this invention constructs a two-factor collaborative mutation mechanism to enable the evolutionary process to dynamically adjust mutation and crossover probabilities through two factors reflecting the population state, thereby controlling the population's evolutionary process. In this method, the population suitability evaluation factor and the population mutation adjustment factor are coordinated to ensure that the evolutionary process has a certain degree of self-adaptability while also allowing the population to develop according to expectations.

[0054] S1.1 Population suitability assessment factor;

[0055] The population fitness evaluation factor influences the evolutionary direction of a population by assessing its fitness level. Each individual in the population participates in controlling changes to the population fitness evaluation factor. The specific calculation formula is as follows:

[0056]

[0057] Among them, F fit f represents the population adequateness evaluation factor. i f avg f max f min Let i represent the fitness of individual i, the average fitness of the population, the maximum fitness, and the minimum fitness, respectively. This indicates the number of individuals generated in each generation. By evaluating the population's fitness level, we can obtain the population's dispersion and fitness index. This factor can carry some information about the population's fitness level, rather than simply making the population more dispersed.

[0058] S1.2 Population variation adjustment factor;

[0059] The population mutation adjustment factor is a factor that controls the direction of population evolution based on the current iteration round. This factor can help the algorithm escape local optima by increasing the number of individuals participating in mutation in the current population. The specific calculation formula is as follows:

[0060]

[0061] Where n is determined by the current evolutionary cycle; s represents the given overall evolutionary cycle; and C is a hyperparameter used to regulate the population variation adjustment factor.

[0062] S1.3 Establish the evolutionary process based on the two-factor collaborative mutation mechanism;

[0063] By combining a population fitness evaluation factor with a population variation adjustment factor, an evolutionary process based on a two-factor collaborative variation mechanism is constructed. This process is influenced by multiple indicators such as population fitness, population dispersion, and population generation number, and exhibits a certain degree of adaptability. The following variation control function is designed, influenced by the sigmoid function:

[0064]

[0065] Where w i It is a constant, and a family of curves with respect to P is obtained based on different values. -20x+10 indicates that the factors involved in the control have been normalized so that w i The index ranges from 0 to 1.

[0066] Based on the neural architecture search process, the population evolution process is redesigned and decomposed into three stages, with different tasks and abilities assigned to the population at each stage to adapt to the neural architecture search process. It is well known that a higher crossover probability indicates a population is more inclined to engage in local searches, while a higher mutation probability indicates a population is more inclined to explore. In the early stages of evolution, more exploration occurs, i.e., more mutation operations; in the later stages, more local searches and convergence occur, i.e., more crossover operations. The evolutionary process is decomposed into an exploration stage, a development stage, and a convergence stage, such as... Figure 2-5 As shown.

[0067] (1) Exploration phase

[0068] like Figure 2 As shown, at this point, the population is in a relatively discrete state, with strong global search capabilities. No additional operations are needed; simply let it develop naturally and gradually transition to the development phase. However, if the population exhibits "precocious" behavior during the exploration phase, i.e., excessive concentration, such as... Figure 3 As shown, the mutation probability of the population is set to 0.9, which means that 90% of the individuals will be mutated in order to restore the global search capability of the population.

[0069] (2) Development Stage

[0070] After the exploration phase, the population enters the development phase. At this stage, the crossover probability and mutation probability are almost equal. Furthermore, as evolution progresses, the crossover probability gradually increases, while the mutation probability gradually decreases. Additionally, if the population is relatively concentrated at this stage... Figure 4 and Figure 5 As shown, the algorithm can automatically increase the mutation probability to a certain extent, giving more individuals the opportunity to undergo mutation, thereby making the population more discrete and restoring the global search capability of the population.

[0071] (3) Convergence phase

[0072] After the development phase, the algorithm will enter the convergence phase. At this point, some individuals carry operation modules that perform well. We need to pass these models to the next generation through crossover. At this point, global exploration is no longer necessary. In this phase, we accelerate model convergence, which means that increasing the mutation probability will promote model convergence.

[0073] The above three stages can be represented in the following functional form:

[0074]

[0075] threshold1 represents the threshold of the exploration stage, threshold2 represents the threshold of the development stage, and threshold3 represents the threshold of the convergence stage; m1 and m2 are set values used for segmenting the entire evolution process, w a and w c are less than 1, w b is greater than 1; n1 and n2 are hyperparameters, and k is a constant, indicating the capability of restoring the global search ability of the population when the population gathers prematurely during the development stage.

[0076] According to formula (4), in the exploration stage, the population mutation probability increases with the evolution process of the population, that is, the dispersion degree of the population is continuously increased, and the mutation probability also increases with the increase of the dispersion degree of the population. At this time, the two factors continuously promote each other to make the population more discrete, so that a more global exploration is performed. If the population is too aggregated and has a "premature" behavior, it will cause F fit <n1 condition to trigger, and the population mutation probability is forcibly set to 0.9, that is, 90% of all individuals undergo mutation, forcing the population to diverge. In the development stage, the population mutation probability decreases with the increase of the current evolution generation n, that is, the dispersion degree of the population decreases, and at this time the mutation probability also increases with the increase of the dispersion degree of the population. The two factors can be in an antagonistic relationship: the population fitness evaluation factor makes the population discrete, and the increase of the current iteration round n makes the population converge, that is, it gradually converges while maintaining a certain global exploration capability. The two factors can also be in a mutually promoting relationship: they promote each other to make the population more convergent, but if the population is too aggregated and the "premature" behavior occurs, the condition F fit <n2 is triggered, according to formula (4), it can be known that the mutation probability increases with the decrease of the population dispersion degree and the mutation probability is greater than 0.5, that is, more than half of the individuals undergo mutation, forcing the population to diverge. In the convergence stage, the mutation probability decreases with the increase of the current iteration round n, that is, the population starts to accelerate convergence.

[0077] For population P t calculate the fitness of each individual in the population, substitute it into formula (1) to calculate F fit , substitute the current evolution generation n into formula (2) to calculate F acc ; according to the obtained F fit and F accThe current evolutionary stage is determined using formula (4), and the current threshold is calculated. Each individual in the current population is assigned a random number between 0 and 1. If the random number exceeds the current evolutionary threshold, the individual is considered to be undergoing a crossover operation; if the random number is below the current threshold, a mutation operation is considered. After all individuals in the population have completed their evolutionary operations, they are merged with their parent populations to form a new population. Then, a fast non-dominated sort is performed on the population containing all individuals before and after evolution, and the number of individuals generated in each generation is used as the basis for the sorting. Select the next generation population P t+1 The specific algorithm process is shown in Table 1.

[0078] Table 1. Evolutionary process algorithm based on two-factor cooperative mutation mechanism

[0079]

[0080] S3 constructs a multi-objective evolutionary neural architecture search framework for generative adversarial networks;

[0081] S3.1 Define multiple optimization objectives for generative adversarial networks;

[0082] This invention addresses the image generation task using generative adversarial networks (GANs). The goal is to generate high-quality images, which can be reflected by IS (Image Quality) and FID (Fixed Integrity Index). However, while considering IS and FID as evaluation metrics, this invention aims to maximize model efficiency. Therefore, it also incorporates the number of model parameters as an evaluation metric, designing a multi-objective neural architecture search based on IS, FID, and the number of model parameters. The specific formula is as follows:

[0083]

[0084] Since a higher IS score indicates a better model performance, this invention uses the reciprocal of the IS score as one of the targets for multi-objective neural architecture search. The smaller the FID and the number of model parameters, the better the network performance.

[0085] Image Filtering (IS) uses an image category classifier, specifically Inception Net-V3, to evaluate the quality of the generated images. IS is a metric for evaluating the sharpness and diversity of the generated images; a higher IS value indicates better image quality. The specific formula for IS is as follows:

[0086] IS(G) = exp(E x～Pg D KL (p(y∣x)||p(y))) (6)

[0087] Where E x～PgD represents the average value obtained by iterating through all generated samples. KL KL divergence measures the distance between two distributions, where p(y|x) represents the probability distribution of an image x belonging to all classes. IS score, which relies heavily on the classifier, is an indirect method for evaluating image quality and does not consider the specific differences between real and generated data.

[0088] FID (Fixed Identifier) ​​primarily considers the difference between the generated image and the real image data distribution. A smaller FID indicates that the generated data is closer to the real data, meaning the generated image has better clarity and diversity. The specific formula for FID is as follows:

[0089] FID = ||μ r -μ g || 2 +T r (∑r+∑g)-2(∑r∑g) 1 / 2 (7)

[0090] Where, μ r and μ g Let be the feature means of the real image and the generated image, respectively; let ∑r and ∑g be the covariance matrices of the real image and the generated image, respectively; and let T be the feature means of the generated image and the feature means of the generated image, respectively. r It is a trace.

[0091] S3.2 Construct the search space;

[0092] Based on AdversarialNAS, the search space is fixed at four nodes in both SuperNet-G and SuperNet-D. On the SuperNet-G network, noisy samples pass through a fully connected layer, then three upsampling layers to obtain generated samples, which are then passed to the SuperNet-D network. On the SuperNet-D network, the generated samples and the real image are compared through three downsampling layers and then a fully connected layer. Specific node operations are as follows: Figure 6 As shown.

[0093] S3.3 Constructing a multi-objective neural architecture search based on dual-factor collaborative control;

[0094] To reduce redundant training processes and improve search efficiency, this invention employs a weight-sharing strategy, allowing the generator and discriminator to inherit weights from the SuperNet-G network and SuperNet-D network, respectively. The search process is divided into steps: in the first stage, G is searched with a fixed D, using a many-to-one training strategy; in the second stage, D is searched with the optimal G found in the first stage, using a one-to-one training and weight resetting strategy to enhance the stability of generative adversarial network training.

[0095] This invention designs a controller to manage the search process of G and D networks, enabling the neural architecture search to adjust according to the evolutionary process. Furthermore, it utilizes a non-dominated ranking method to generate Pareto front architectures under multiple objectives. The specific algorithm is as follows: Figure 7 As shown.

[0096] like Figure 7 The left side shows the process of searching a many-to-one G network after fixing the D network. The controller in this process controls the neural architecture search by using two factors: population adaptation and the current iteration number, to guide the search into different stages. In each round, a... Several candidate generators are selected and their weights are shared using a weight-sharing strategy. The same discriminator is used to train different generators to find the optimal one. Furthermore, generators that experience mode collapse will not interfere with other generators, as these generators are removed during the selection module.

[0097] After obtaining the optimal generator, it is copied and then trained one-to-one with different discriminators. This invention uses the generator's evaluation index to indirectly evaluate the discriminator. Before the start of each round, the best G weight is assigned to each G* and a weight reset strategy is used to mitigate mode collapse.

[0098] Example:

[0099] This invention first searches for generative adversarial network architectures on the CIFAR-10 dataset, and then trains and tests the obtained neural network architectures on both the supervised CIFAR-10 dataset and the unsupervised STL-10 dataset. Extensive ablation experiments are conducted to verify the effectiveness of this method.

[0100] The data in Table 2 can be obtained by comparing the experimental results of this method with the models that have better performance at present.

[0101] Table 2 Comparison of Experimental Results

[0102]

[0103] In the experimental results shown in Table 1, the present invention achieved an IS score of 8.64 (±0.09) and an FID score of 10.87 on the CIFAR-10 dataset, with a G network parameter count of 7.4 MB. On the STL-10 dataset, the present invention achieved an IS score of 9.93 (±0.13) and an FID score of 21.64, with a G network parameter count of 11.55 MB. On the CIFAR-10 dataset, the manually designed Big Generative Adversarial Network (GAN) achieved an IS score of 9.22. However, since the Big GAN is designed with a single objective, its FID score is lower than that of most GANs. The gradient-based neural architecture search algorithm AdversarialNAS performed well in both IS and FID scores (IS = 8.74 ± 0.07, FID = 10.87), but its performance on the STL-10 dataset was not as good as the network searched by the present invention.

[0104] like Figure 8 The standard deviation of the population evaluation index is statistically analyzed during the search of the G network. This can reflect the global search capability of the population to a certain extent. It can be seen that the standard deviation of the population tends to increase during the exploration phase, meaning the population becomes more dispersed and has a stronger global search capability. During the development phase, the standard deviation of the population gradually decreases, meaning it begins to converge while maintaining a certain degree of dispersion. It is evident that the standard deviations of the IS and FID indices are very low during the convergence phase, indicating that the population has converged considerably. This demonstrates that the evolutionary process designed in this invention maintains a high global search capability during the exploration phase, gradually converges while maintaining a certain global search capability during the development phase, and accelerates convergence during the convergence phase, proving the effectiveness of the method of this invention.

[0105] Ablation experiments were conducted on this invention, and the method of this invention was tested on a single target. The results showed good performance on one target, but the results on other targets were far inferior to the experimental results of this invention.

[0106] Table 3 Comparison of experimental results between multi-objective optimization and single-objective optimization

[0107]

[0108] Experiments were conducted with the population fitness factor removed, transforming the evolutionary process into one strictly controlled iterative workflow. Specific experimental results are as follows: Figure 9 As shown, the horizontal axis represents the number of generations, and the vertical axis represents the standard deviation of the population evaluation index; the triangle indicates that no control factors for crossover and mutation probabilities are added, and the probabilities are all set to 0.5; the box indicates that the control is strictly based on the current iteration number; the dot indicates the two-factor collaborative mutation mechanism.

[0109] As can be seen, convergence is very slow, far slower than with factor control without it. Furthermore, when strictly following the process control, convergence clearly begins around round 110, because the crossover probability is already greater than the mutation probability at this point. Moreover, the population dispersion at its highest point is far less than that under the two-factor cooperative mutation mechanism. In other words, under single-factor control, the population is in a gradually converging state, but its global search capability is far inferior to that of the two-factor cooperative mutation mechanism, demonstrating the effectiveness of the method presented in this invention.

Claims

1. A multi-objective adaptive evolutionary neural architecture search method for generative adversarial networks, characterized in that, The specific steps are as follows; Step 1: Construct a two-factor cooperative mutation mechanism; First, the scores of candidate individuals in the image generation task are used as information such as fitness and dispersion to construct a population fitness evaluation factor. Second, the current generation and total number of iterations are used to construct a population variation adjustment factor. Finally, the two factors are combined to construct a crossover and mutation probability control function based on a two-factor collaborative variation mechanism. S1.1 Population suitability assessment factor; The population wellness assessment factor influences the evolutionary direction of a population by evaluating its wellness. Each individual in the population participates in controlling changes to the population wellness assessment factor. The specific calculation formula is as follows: in, Used to represent the designed population adequateity evaluation factor; Individuals in a population Moderation Indicates the average fitness of the population; This represents the total number of new individuals generated in each generation of a given population; population dispersion and fitness index are obtained through population fitness evaluation factors. S1.2 Population variation adjustment factor; The population variation adjustment factor is a factor that controls the direction of population evolution based on the current iteration round. The specific calculation formula is as follows: in, Determined by the current evolutionary stage; Represents a given overall evolutionary cycle; These are hyperparameters used to regulate population variation adjustment factors; S1.3 Establish the evolutionary process based on the two-factor collaborative mutation mechanism; By combining population fitness evaluation factors with population variation adjustment factors, an evolutionary process based on a two-factor collaborative variation mechanism is constructed. This evolutionary process is influenced by population fitness, population dispersion, and population generation index, and exhibits adaptability. The following probability function is established based on the sigmoid function: in, It is a constant. j for a , b , c Based on different values, we can obtain the probability function. family of curves, This indicates that normalization is being performed. Based on the process and expectations of neural architecture search, the evolutionary process is divided into three stages: exploration, development, and convergence. in, These represent the evolutionary thresholds for the three stages of exploration, development, and convergence, respectively. and This is a set value used to segment the entire evolutionary process. and Less than 1, Greater than 1; and It's a hyperparameter. It is a constant; Step 2: Construct a multi-stage adaptive evolutionary process based on a two-factor cooperative mutation mechanism. For population Calculate the fitness of each population and substitute it into formula (1) to calculate... The current generation Substitute into formula (2) to calculate According to the obtained and The current evolutionary stage is determined by formula (4), and the current evolutionary stage is calculated. Each individual in the current population is assigned a random number between 0 and 1. If the random number exceeds the current evolutionary threshold, the individual undergoes a crossover operation; if the random number is below the threshold, a mutation operation is performed. After all individuals in the population have completed their evolutionary operations, they are merged with their parent populations to form a new population. Then, a fast non-dominated sort is performed on this population containing all individuals before and after evolution, based on the number of individuals generated in each generation. Selecting the next generation of population ; Step 3: Construct a multi-objective evolutionary neural architecture search framework for generative adversarial networks; S3.1 Establish multiple optimization objectives based on the image generation task; Simultaneously considering IS evaluation index, FID evaluation index, and model parameter quantity. Evaluation metrics are proposed, including a multi-objective evolutionary neural architecture search framework based on IS, FID, and model parameter count; the specific formula is as follows: A higher IS score indicates a better performance of the generative adversarial network, while a lower FID and model parameter count indicate better performance of the generative adversarial network. Image Classification (IS) evaluates the quality of generated images using an image category classifier. The specific formula for IS is as follows: in, This represents the average value obtained by iterating through all generated samples. This represents the KL divergence, which measures the distance between two distributions. Indicates for images The probability distribution of belonging to all categories; The specific formula for FID is as follows: in, and These are the feature mean of the real image and the feature mean of the generated image, respectively. These are the covariance matrices of the real image and the generated image, respectively. It is a trace; S3.2 Construct the search space; Based on AdversarialNAS, the search space is fixed at four nodes for both SuperNet-G and SuperNet-D. Noise samples on the SuperNet-G network pass through a fully connected layer and then through three upsampling layers to obtain generated samples, which are then passed to the SuperNet-D network. On the SuperNet-D network, the generated samples and the real images are compared through three downsampling layers and then through a fully connected layer. S3.3 Constructing a multi-objective neural architecture search based on dual-factor collaborative control; By employing a weight-sharing strategy, the generator and discriminator inherit weights from the SuperNet-G and SuperNet-D networks, respectively, and the search process is divided into steps. A controller is proposed to control the search process of the generator and discriminator, both of which follow a multi-stage evolutionary process based on a two-factor cooperative mutation mechanism. Fast non-dominated sorting is used to obtain a set of Pareto front solutions under multiple set objectives. The first stage uses a fixed discriminator to search for a generator and adopts a many-to-one training strategy; the second stage uses the optimal generator found in the first stage to search for a discriminator and adopts a one-to-one training weight reset strategy to enhance the stability of the generative adversarial network training. The fixed discriminator, during the many-to-one training strategy, generates [a certain value] in each round. Several candidate generators are selected and their weights are shared using a weight-sharing strategy. The sampled generators are then subjected to adversarial training against an identical discriminator to obtain the optimal generator network. After obtaining the optimal generator, its weights are copied and labeled as... Then, it is trained one-to-one with different discriminators, and the weight reset strategy is started before the start of each round with the optimal generator network weight as the benchmark.

Images