A generative adversarial network architecture search method based on candidate sub-network pool guidance

Abstract

The application discloses a kind of generation of candidate subnetwork pool based on guide's generative adversarial network architecture search method, this method includes the steps: constructing weight sharing generator hypernetwork, and establishing candidate memory entry for candidate generator subnetwork;Through low-fidelity evaluation shrinkage generator search space;Periodically perform budget annealing interleaved evolutionary search in the process of shared weight training, parameter budget screening, performance evaluation, memory entry refresh and reliability score calculation are carried out to candidate subnetwork;According to smooth performance, stability deviation, evaluation information obsolescence and training feedback, candidate subnetwork pool is dynamically maintained;Then based on the search space after shrinkage, current high-quality candidate set and candidate subnetwork pool, annealing three-way sampling training is carried out, and the shared weight of weight sharing generator hypernetwork is updated.The application can improve the retention rate and reuse rate of historical reliable candidates, and alleviate the candidate ranking instability problem caused by the evaluation fluctuation of inherited weights.

Description

Technical Field

[0001] This invention relates to the field of network architecture search technology, and in particular to a generative adversarial network architecture search method guided by a candidate subnetwork pool. Background Technology

[0002] Generative Adversarial Networks (GANs) have significant applications in tasks such as image generation, data augmentation, and image-to-image translation. To reduce the cost of manually designing generator structures, existing technologies have developed neural network architecture search methods for GANs. These methods typically construct a supernetwork and sample, train, and evaluate candidate subnetworks under weight-sharing conditions, thereby automatically discovering superior generator structures within a large search space. Some methods also incorporate evolutionary search, multi-objective selection, or knowledge distillation into the architecture search process to improve search efficiency and candidate model performance.

[0003] However, existing technologies still have the following shortcomings in the search of generative adversarial network (GAN) architectures. First, the training process of GANs is inherently unstable. The performance statistics of candidate subnetworks are easily affected by factors such as the training phase, sampling randomness, and dynamic changes in the discriminator, resulting in large fluctuations in candidate rankings and unstable relationships between candidate architectures during the search process. Second, to improve the reliability of candidate ranking, existing methods typically require frequent re-evaluation of parent candidates or retained candidates. However, repeated evaluation significantly increases the time overhead of the search phase. Reducing the number of evaluations, on the other hand, can lead to insufficient timeliness of candidate performance information, causing inconsistencies between the ranking results in the search phase and the actual performance in subsequent independent retraining phases.

[0004] Furthermore, existing methods often focus more on a small number of candidate structures that perform well at the current stage, lacking mechanisms for long-term preservation, dynamic updating, and continuous utilization of historical high-quality candidate subnetworks. When some candidate structures perform well in the intermediate stage but are not retained in time, or are prematurely eliminated in later stages due to short-term fluctuations, high-quality candidate information is easily lost, reducing the search process's ability to utilize historical effective information. At the same time, existing methods usually lack a unified and coordinated mechanism for candidate maintenance, training sampling, and candidate refreshing, making it difficult to simultaneously consider search efficiency, candidate ranking stability, and subsequent independent retraining performance.

[0005] Furthermore, in the later stages of the search, as training sampling gradually shifts towards historical high-quality candidate sets, without a low-cost local exploration and diversity compensation mechanism, the supernetwork training is prone to over-concentration on a few well-performing candidate subnetworks, leading to an overly concentrated distribution of later sampling and a decrease in local structural diversity, thereby affecting the stability and generalization ability of the search results.

[0006] Therefore, there is an urgent need for a generative adversarial network architecture search method that can balance search efficiency, timeliness of candidate performance information, stability of candidate ranking, and consistency with subsequent independent retraining during the generative adversarial network architecture search process, and can continuously maintain and guide the utilization of historical high-quality candidate subnetworks. Summary of the Invention

[0007] Purpose of the Invention: The purpose of this invention is to provide a generative adversarial network architecture search method based on a candidate sub-network pool, in order to solve the technical problems in the existing generative adversarial network architecture search process, such as the difficulty in continuously accumulating and effectively utilizing candidate performance information, the difficulty in retaining and dynamically reusing historical high-quality candidate sub-networks, the large overhead of repeated evaluation of parent candidates and the tendency for insufficient timeliness of candidate performance information when reducing repeated evaluation, the tendency for excessive concentration of sampling distribution in later training, insufficient stability of candidate ranking, and poor consistency between search results and subsequent independent retraining results, so as to achieve continuous maintenance and guided utilization of historical high-quality candidate sub-networks.

[0008] Technical solution: A generative adversarial network architecture search method guided by a candidate subnetwork pool, comprising the following steps:

[0009] S1, construct a weight-shared generator supernetwork for generator architecture search, and initialize a set of candidate generator subnetworks based on a preset generator search space; at the same time, establish candidate memory entries for each candidate generator subnetwork.

[0010] S2, pre-train the weight-sharing generator supernetwork, and shrink the search space based on low-fidelity evaluation to obtain the shrunken search space. ;

[0011] S3 performs interleaved evolutionary search with budgeted annealing based on the shrunken generator search space and dynamically maintains a pool of candidate subnetworks.

[0012] S4, obtaining the current set of high-quality candidates in each round of interleaved evolutionary search. And complete the candidate subnetwork pool After dynamic maintenance, determine whether the current search process meets the termination condition;

[0013] If the termination condition is not met, then the search space is shrunk. Current high-quality candidate set and the candidate subnetwork pool after this round of maintenance An annealing three-way sampling distribution is constructed, and a training sub-network is selected according to the annealing three-way sampling distribution to update the shared weights of the weight-sharing generator supernetwork; the updated shared weights are then used for candidate evaluation, candidate memory entry update, candidate reliability score calculation, candidate sub-network pool maintenance, and interleaved evolutionary search.

[0014] If the termination condition is met, save the set of high-quality candidates and output the target generator architecture.

[0015] Furthermore, the candidate memory entry includes at least the candidate architecture encoding, inheritance weight evaluation result, parameter size, evaluation round, most recent training round, number of sampled training rounds, and number of historical evaluation rounds; wherein, the inheritance weight evaluation result includes the original FID value and IS value.

[0016] Furthermore, before the formal search phase begins, the weight-sharing generator supernetwork is pre-trained. After pre-training, the pre-trained shared weights are used to perform low-fidelity evaluation on the candidate generator subnetworks to obtain candidate performance estimation results. Based on the candidate performance estimation results, candidate selection constraints are constructed, and the generator search space is shrunk to obtain the shrunk search space. .

[0017] Furthermore, the pre-training phase searches the pre-defined generator space. A randomized candidate generator subnetwork is selected as the current activation path, and adversarial training is performed under the constraint of a fixed-structure discriminator to update the shared weights of the weight-sharing generator supernetwork. During adversarial training, the currently activated candidate generator subnetwork is fed with a random noise vector. As input, output the corresponding generated image sample;

[0018] Let the first The candidate architecture encoding obtained from sampling in each training round is: The shared weights of the weight-sharing generator supernetwork are: The random noise vector is The generated image sample is:

[0019] ,

[0020] in, Indicates that it is encoded by the architecture. Activate and inherit shared weights The generator subnetwork, Indicates the starting round of the formal search. This indicates the generation of image samples;

[0021] After pre-training, the updated shared weights are utilized. Perform low-fidelity evaluation on the candidate generator subnetwork to obtain approximate performance metrics of the candidate structure without requiring independent retraining;

[0022] Based on the approximate performance indicators, structural screening constraints are constructed, and the search space of the preset generator is applied. Structural feasibility is pruned, retaining only candidate structural regions that meet performance and parameter budget constraints, resulting in a shrunken search space. ;

[0023] The low-fidelity evaluation refers to the low-cost and fast estimation of the FID index, IS index and parameter size of the candidate generator subnetwork by limiting the number of generated samples under the condition of unchanged shared weights, in order to characterize the relative performance level of the candidate structure.

[0024] Furthermore, in step S3, the specific implementation process of dynamically maintaining the candidate sub-network pool is as follows:

[0025] After the training rounds reach the preset starting round for the formal search, the formal search phase begins in the shrunken search space. An interleaved evolutionary search is performed at a preset cycle to generate candidate generator subnetworks to be evaluated, and they are screened by annealing dynamic parameter budget constraints. The screened candidate generator subnetworks are then evaluated by inheritance weights to obtain the performance record for the current round. The performance record includes at least the FID index, IS index and parameter size.

[0026] The performance records are written into the corresponding candidate memory entries, and the smoothed FID value and stability deviation are updated based on the historical evaluation sequence, the most recent evaluation round, the most recent training round, and the number of sampled training sessions. The staleness of the evaluation information and the intensity of the training feedback are further determined to construct a candidate reliability score.

[0027] The current high-quality candidate set is obtained based on the FID index, IS index, parameter size and reliability score, and the candidate sub-network pool is dynamically maintained based on the current high-quality candidate set.

[0028] Furthermore, the specific steps for performing an interleaved evolutionary search according to a preset search cycle include:

[0029] S31, in the In the interleaved evolutionary search, the current search population... As the basic candidate set, and based on the structural perturbation mechanism, an extended candidate set is generated. Subsequently, the basic candidate set and the extended candidate set are merged to obtain the joint candidate set. :

[0030] ;

[0031] The structural disturbance mechanism includes one or more of the following: structural reorganization, structural intersection, structural variation, local path disturbance, or resampling.

[0032] S32, for candidate joint sets Apply annealing-type dynamic parameter budget constraints, the first Upper limit of parameter budget for round-robin evolutionary search for:

[0033] ,

[0034] in, The preset parameter size; These are non-incremental annealed dynamic coefficients used to gradually tighten parameter budget constraints as the search rounds progress;

[0035] Subsequently, the parameter budget constraint is used as a candidate selection constraint, retaining only candidate generator subnetworks whose parameter size does not exceed the upper limit of the parameter budget in the r-th round, thus obtaining the selected candidate set:

[0036] ,

[0037] in, This represents the architecture encoding of the i-th candidate generator subnetwork. Represents the candidate generator subnetwork The parameter scale, This represents the candidate set after filtering by parameter budget constraints;

[0038] S33, Candidate Set The candidate generator subnetwork inherits the shared weights of the current weight-sharing generator supernetwork and generates a preset number of image samples. Based on these image samples, the IS index is calculated, and based on the distribution difference between the image samples and real samples, the FID index is calculated. Simultaneously, the parameter scale of the candidate generator subnetwork is statistically analyzed to form a performance record for the current round. A phased differentiated refresh mechanism is used for weight inheritance performance evaluation: the new candidate set generated in this round... Directly perform inheritance weight performance evaluation; for the current search population If a candidate memory entry already exists in the historical candidate list, then it is determined whether the evaluation needs to be refreshed based on its most recent evaluation round, the number of historical evaluations, and the search phase.

[0039] S34, write the current round performance record into the corresponding candidate memory entry, and update the candidate state variables based on the candidate historical evaluation sequence. The candidate state variables include smoothed FID value, stability bias, evaluation information staleness and training feedback strength.

[0040] S35, calculate the candidate reliability score from the candidate state variables, and then apply the IS index, FID index, parameter size, and candidate reliability score to the filtered candidate set. By prioritizing and selecting candidates, the current set of high-quality candidates is obtained. .

[0041] Furthermore, the candidate reliability score is calculated based on the stability bias of the candidate generator sub-network, the staleness of the evaluation information, and the training feedback strength; wherein, the lower the staleness of the evaluation information, the newer the evaluation information; the staleness of the evaluation information is determined according to the evolutionary round corresponding to the most recent inherited weight performance evaluation of the candidate generator sub-network.

[0042] ,

[0043] in, Represents the candidate generator subnetwork In the The staleness of evaluation information in round evolutionary search

[0044] Represents the candidate generator subnetwork The evolutionary cycle corresponding to the most recent inheritance weight performance evaluation;

[0045] The training feedback strength is determined based on the number of times the candidate generator subnetwork has been sampled during recent shared weight training:

[0046] ,

[0047] in, Indicates the candidate generator subnetwork at the th The training feedback intensity corresponding to each training round Represents the candidate generator subnetwork The number of times the sample was trained during the recent training phase;

[0048] After normalizing the stability bias, evaluation information staleness, and training feedback strength, the candidate reliability score is obtained:

[0049] ,

[0050] in, Represents the candidate generator subnetwork In the Reliability score in round evolutionary search , and These represent the stability bias after normalization, the staleness of the evaluation information, and the intensity of the training feedback, respectively.

[0051] Furthermore, in step S3, the dynamic maintenance of the candidate sub-network pool includes candidate status updates, adding new candidates to the pool, candidate replacement, and retention of high-quality candidates.

[0052] For the candidate generator subnetwork Candidate Subnetwork Pool Historical state information is recorded using architecture encoding as an index, including at least the original FID value, IS value, parameter size, evaluation round, most recent training round, and number of evaluations. The number of sampled training times is obtained by associating with the corresponding candidate memory entries. The candidate sub-network pool is used to continuously model the state evolution of the candidate generator sub-network across search rounds.

[0053] Compared with the prior art, the significant advantages of this invention are as follows:

[0054] 1. This invention establishes candidate memory entries and a candidate sub-network pool, continuously recording the inherited weight evaluation results, parameter size, evaluation rounds, most recent training rounds, number of sampled training times, and number of historical evaluation times of the candidate generator sub-network. It also performs pooling, refreshing, replacement, and retention operations on historically reliable candidates, enabling the candidate generator sub-network to accumulate performance information and training feedback information across search rounds, thereby improving the retention rate and reuse rate of stable and high-quality candidates.

[0055] 2. This invention introduces a candidate reliability scoring mechanism, which models the reliability of the inherited weight evaluation results of the candidate generator sub-network based on smoothed FID value, stability bias, evaluation information staleness and training feedback strength. This reduces the impact of short-term performance fluctuations on candidate selection and candidate sub-network pool maintenance, thereby improving the stability of candidate ranking and candidate retention.

[0056] 3. This invention uses a budget annealing intermittent evolutionary search mechanism to periodically perform evolutionary search during the training of shared weights in the weight-sharing generator supernetwork, rather than searching independently after the supernetwork is trained; and through gradually tightening parameter budget constraints, it maintains structural diversity in the early stage of the search and enhances the pressure of selecting compact architectures in the later stage of the search, thereby balancing the exploratory nature of the search and the compactness of the model.

[0057] 4. This invention uses a phased differentiated refresh mechanism to perform real-time performance evaluation of newly generated candidates by inheriting weights and to selectively refresh historical candidates. This reduces the overhead of repeated evaluations while improving the timeliness of candidate performance information, thereby alleviating the consistency deviation between the inherited weight evaluation results in the search phase and the subsequent independent retraining performance.

[0058] 5. This invention incorporates the shrunken search space, the current high-quality candidate set, and the candidate subnetwork pool into the shared weight training process through the annealing three-way sampling training mechanism. As training progresses, the sampling ratio of historical reliable candidates in the candidate subnetwork pool is gradually increased, enabling reliable candidates to guide the shared weight optimization of the weight sharing generator supernetwork in reverse, forming a closed-loop optimization process in which interleaved evolutionary search and shared weight training mutually feedback each other. Attached Figure Description

[0059] Figure 1 This is an overall framework diagram of the present invention;

[0060] Figure 2 This is the search space graph in this invention;

[0061] Figure 3 This is a graph showing the trend of IS index changes during the architecture search process in this invention;

[0062] Figure 4 This is a graph showing the trend of FID index changes during the architecture search process in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0064] like Figure 1 The diagram illustrates the generative adversarial network architecture search method based on a candidate sub-network pool in this embodiment, which includes the following steps:

[0065] Step 1: Construct a weight-sharing generator supernetwork and initialize candidate memory entries;

[0066] like Figure 2 As shown, in this embodiment, under the constraints of a preset generator search space, a weight-shared generator supernetwork for generator architecture search is constructed, and a fixed-structure discriminator is used to perform adversarial training with the currently activated candidate generator subnetwork. The candidate generator structure is controlled by the candidate architecture encoding to activate its path in the weight-shared generator supernetwork, and the discriminator structure remains fixed and does not participate in the architecture search.

[0067] During the initialization phase, several candidate generator subnetworks that satisfy the initialization parameter size constraints are randomly sampled and generated, forming an initial candidate generator subnetwork set. At the same time For each candidate generator subnetwork, a candidate memory entry is established. This entry is indexed by the candidate generator subnetwork's architecture encoding and is used to uniformly store and update the structural information, performance evaluation results, and training feedback information of the candidate generator subnetwork across search rounds. The candidate memory entry continuously records the historical performance and training feedback information of the candidate generator subnetwork in subsequent search rounds, including at least the candidate architecture encoding, inherited weight evaluation results, parameter size, evaluation round, most recent training round, number of sampled training iterations, and number of historical evaluation iterations; the inherited weight evaluation results include the original FID value and IS value. Subsequent steps further update the smoothed FID value, stability bias, evaluation information staleness, and training feedback strength based on the candidate memory entries, and use these updates for reliability score calculation, candidate subnetwork pool maintenance, and annealing three-way sampling training.

[0068] Step 2: Pre-train the weight-sharing generator supernetwork and shrink the search space based on low-fidelity evaluation;

[0069] Before the formal search phase begins, the weight-sharing generator supernetwork is pre-trained to optimize the shared weights corresponding to the candidate generator subnetworks. In this embodiment, after completing the initial set of candidate generator subnetworks... After construction, the weight-sharing generator supernetwork is pre-trained. During the pre-training phase, candidate generator subnetworks are sampled from a predefined generator search space as the current activation path, and adversarial training is performed against a fixed-structure discriminator to optimize the shared weights of different candidate generator subnetworks. During adversarial training, the currently activated candidate generator subnetwork takes a random noise vector as input and outputs the corresponding generated image sample.

[0070] Let the first The candidate architecture encoding obtained from sampling in each training round is: The shared weights of the weight-sharing generator supernetwork are: The random noise vector is The generated image sample is:

[0071] ,

[0072] in, Indicates that it is encoded by the architecture. Activate and inherit shared weights The generator subnetwork, Indicates the starting round of the formal search. This indicates the generation of image samples.

[0073] After pre-training, a low-fidelity evaluation is performed on the candidate generator subnetworks. This low-fidelity evaluation refers to quickly estimating the FID, IS, and parameter sizes of the candidate generator subnetworks using the current shared weights and a small number of generated samples without performing complete independent retraining. The FID, IS, and parameter sizes are written into the corresponding candidate memory entries as the initial inherited weight evaluation record for that candidate generator subnetwork. Subsequently, based on the low-fidelity evaluation results and parameter budget constraints, the preset generator search space is filtered, retaining candidate structure regions that meet the conditions and reducing the subsequent sampling priority of low-confidence candidate structure regions, resulting in a shrunken generator search space. This shrinkage process preserves structural regions with the potential for high-quality generation and reduces the subsequent sampling priority of low-confidence candidate structural regions.

[0074] Step 3, based on the shrunken search space Periodically perform interleaved evolutionary search using budget annealing and dynamically maintain a pool of candidate subnetworks;

[0075] like Figure 1 As shown, when the pre-training rounds reach the preset starting round of the formal search... Then, the formal search phase begins. Compared to the approach of performing architecture search independently after the generator supernetwork has been trained, this embodiment performs interleaved evolutionary search during the training of the weight-sharing generator supernetwork, creating an alternating closed loop for candidate generation, inherited weight performance evaluation, candidate memory entry refresh, candidate subnetwork pool maintenance, and shared weight update. The specific steps are as follows:

[0076] Step 31, in the In round-robin evolutionary search, the current search population is denoted as... This serves as the basic candidate set. Based on the current search population... In the shrunken search space The internal mechanism generates a new candidate set for the current round based on structural perturbation. (i.e., expanding the candidate set). Among them, the structural perturbation mechanism includes one or more of the following: structural reorganization, crossover, mutation, local perturbation, or resampling.

[0077] Then, the current search population and the current round of new candidate set Merge to obtain the joint set of candidates for the current round. :

[0078] ;

[0079] in, This represents the current candidate population in the r-th round of search. Represents the new candidate set generated in the r-th round. This represents the joint set of candidates in the current round.

[0080] Step 32, for the candidate joint set Annealing-based dynamic parameter constraints are applied to limit the structural complexity of the candidate generator subnetwork. Then the... Upper limit of parameter budget for round-robin evolutionary search for:

[0081] ,

[0082] in, This is the preset parameter size. This is a non-incremental annealing dynamic coefficient, and the budget annealing coefficient varies with the search rounds, i.e., it gradually changes from the initial budget coefficient to the final budget coefficient as the interleaved evolutionary search rounds progress. The budget annealing coefficient can be determined using the following piecewise annealing function:

[0083] ,

[0084] in, , and These represent the initial, intermediate, and final values ​​of the budgeted annealing coefficient, respectively. and These represent the dividing rounds between the first annealing stage and the second annealing stage, respectively.

[0085] Subsequently, only candidate generator subnetworks that satisfy the parameter budget constraints are retained:

[0086] ,

[0087] in, This represents the architecture encoding of the i-th candidate generator subnetwork. Represents the candidate generator subnetwork The parameter scale, This represents the candidate set after being filtered by parameter budget constraints.

[0088] Step 33: For the candidate set after the current round has been filtered by parameter budget constraints The candidate generator subnetwork inherits the shared weights from the current weight-sharing generator supernetwork. The algorithm generates a preset number of image samples by inputting a random noise vector. Based on these image samples, it calculates the IS (Integer Identifier) ​​index and the FID (Fairness Identifier) ​​index based on the distribution difference between the image samples and the real samples. Simultaneously, it calculates the parameter size of the candidate generator subnetwork, thus forming a performance record for the current round's inherited weights. This performance record includes at least the IS index, FID index, and parameter size, calculated as follows:

[0089] ,

[0090] ,

[0091] ,

[0092] in, Represents the candidate generator subnetwork In the current shared weight Below, the original FID value is calculated based on the distribution difference between the generated image samples and the real samples. Represents the candidate generator subnetwork The IS value obtained under the current shared weights Represents the candidate generator subnetwork The parameter scale.

[0093] To reduce the overhead of repeated evaluation of historical candidate generator subnetworks and maintain the timeliness of candidate performance information, this embodiment adopts a phased differentiated refresh mechanism for inherited weight performance evaluation: the new candidate set generated in this round... Directly perform inheritance weight performance evaluation; for the current search population If a historical candidate generator subnetwork already contains candidate memory entries, then the decision on whether to perform an inheritance weight performance refresh evaluation is determined based on the most recent evaluation round, the number of historical evaluations, and the search phase of the historical candidate generator subnetwork.

[0094] The candidate set for which inheritance weight performance refresh evaluation needs to be performed is represented as follows:

[0095] ,

[0096] in, This represents the candidate set for which inheritance weight performance refresh evaluation needs to be performed. Represents the candidate generator subnetwork The evolutionary cycle corresponding to the most recent inheritance weight performance evaluation Indicates the re-evaluation interval for historical candidates. This indicates the initial evolutionary round for later high-frequency refresh evaluation.

[0097] For the candidate set The candidate generator subnetwork inherits the shared weights of the current weight-sharing generator supernetwork, generates a preset number of image samples, and calculates their IS and FID indices; for those not selected... Furthermore, for historical candidate generator subnetworks where candidate memory entries are still valid, their existing inheritance weight evaluation records are reused.

[0098] Step 34: After evaluation, write the original FID value, IS value, parameter size, evaluation epoch, and current training epoch of the candidate generator sub-network into the corresponding candidate memory entry. The original FID value refers to the FID value directly calculated by the candidate generator sub-network in a single inherited weight performance evaluation; the smoothed FID value refers to the performance state value obtained by exponentially moving average based on the historical original FID values ​​of the candidate generator sub-network. For the candidate generator sub-network being written into a candidate memory entry for the first time, initialize its smoothed FID value to the original FID value obtained in this inherited weight performance evaluation, and initialize the stability bias to 0.

[0099] ,

[0100] ,

[0101] For candidate generator subnetworks with existing historical records, update the smoothed FID values ​​and stability bias based on their historical original FID value sequences:

[0102] ,

[0103] ,

[0104] in, This is a smoothing coefficient, with a value range of 0 ≤ <1; This indicates the smoothed FID value before this update; Represents the candidate generator subnetwork The original FID value obtained in the most recent inheritance weight performance evaluation; Represents the candidate generator subnetwork Smoothed FID values; Represents the candidate generator subnetwork The stability bias is used to characterize the overall performance trend of the candidate across multiple evaluations.

[0105] Simultaneously, the staleness of the evaluation information is determined based on the most recent evaluation round, and the intensity of the training feedback is determined based on the training feedback information. The training feedback information includes the most recent training round and the number of sampled training iterations of the candidate generator sub-network. The intensity of the training feedback is used to quantify the training reach of the candidate generator sub-network during recent shared-weight training. Let the candidate generator sub-network... The obsolescence of the assessment information is as follows:

[0106] ,

[0107] Its training feedback intensity Represented as:

[0108] ,

[0109] in, Represents the candidate generator subnetwork The training feedback intensity corresponding to the t-th training round; This indicates the distance *r* from the candidate generator subnetwork in the current evolutionary round. The interval between the most recent evaluation rounds Represents the candidate generator subnetwork The evolutionary cycle corresponding to the most recent inheritance weight performance evaluation; Represents the candidate generator subnetwork The number of times the sample was trained during the recent training phase.

[0110] Step 35: Based on the stability bias, evaluation information staleness, and training feedback strength of the candidate generator sub-network, a candidate reliability score is calculated to characterize the long-term stability and shared weight training adaptability of the candidate generator sub-network. Specifically, after normalizing the stability bias, evaluation information staleness, and training feedback strength, the following is obtained: , and The candidate reliability score is defined as:

[0111] ,

[0112] in, This represents the reliability score of the candidate generator subnetwork in the r-th round of search; , , These represent the stability bias after normalization, the staleness of the evaluation information, and the strength of the training feedback, respectively. The lower the staleness of the evaluation information, the more recent the evaluation information of the candidate generator subnetwork.

[0113] Subsequently, based on the FID metric, IS metric, and reliability score of the candidate generator subnetwork, the candidate set after this round of screening is... The candidate subnetworks are non-dominated and sorted to obtain the current set of high-quality candidates. Among them, the FID and IS indices are used to characterize the generation quality of the candidate generator subnetwork, the parameter size serves as a parameter budget constraint to filter out infeasible candidates, and the reliability score is used to adjust the selection priority of the candidate generator subnetwork to reduce the impact of short-term inherited weight evaluation fluctuations on the candidate selection results.

[0114] In one optional implementation, for candidate generator subnetworks located in the same non-dominated layer, the crowding distance is first normalized to obtain... Then, the following selection and scoring methods are used to adjust the priority:

[0115] ,

[0116] in, Represents the candidate generator subnetwork Choose a rating, This represents the normalized crowding distance.

[0117] Therefore, the reliability score is not used as an independent Pareto optimization objective, but rather to enhance the candidate ranking and selection process and provide a reliability basis for the subsequent maintenance of candidate sub-network pools.

[0118] Step 36, based on the current high-quality candidate set Candidate memory entries and reliability scores are used to evaluate the candidate sub-network pool. Perform dynamic maintenance.

[0119] If the candidate generator subnetwork Already exists in the candidate subnetwork pool In the middle, for candidate sub-network pools The state records are incrementally updated, including synchronous updates of parameter size, evaluation rounds, most recent training rounds, and number of evaluations;

[0120] If the candidate generator subnetwork The candidate subnetwork pool does not exist. If the capacity limit is not reached, the candidate generator subnetwork will be moved to the next network. Added to candidate subnetwork pool ;

[0121] When candidate subnetwork pool When the capacity limit is reached, the simple sorting and replacement method based on a single performance indicator is no longer used. Instead, a candidate subnetwork pool maintenance score is constructed based on the original FID value, smoothed FID value, stability deviation, parameter size, and candidate reliability score of the candidate generator subnetwork. Based on the candidate subnetwork pool maintenance score, the retained and replaced objects in the candidate subnetwork pool are determined. This prioritizes the retention of candidate generator subnetworks with better overall performance, smaller performance fluctuations, and higher reliability scores, thereby realizing the dynamic updating of the candidate subnetwork pool.

[0122] The aforementioned candidate reliability score is used for candidate ranking and selection in the current candidate set; during the candidate sub-network pool maintenance phase, a candidate sub-network pool maintenance score is further constructed to determine the candidate retention objects and replacement objects when the candidate sub-network pool reaches its capacity limit.

[0123] For the candidate generator subnetwork The candidate sub-network pool maintenance score is defined as follows:

[0124] ,

[0125] in, Represents the candidate generator subnetwork The candidate sub-network pool maintenance score; , , , These represent the normalized smoothed FID value, stability bias, parameter size, and training feedback strength, respectively. The candidate subnetwork pool maintenance score is set between 0 and 1 to facilitate comparison of the overall maintenance priority of different candidate generator subnetworks. A higher candidate subnetwork pool maintenance score indicates higher reliability, better generation quality, smaller performance fluctuations, more compact parameter size, and more sufficient training feedback for the candidate generator subnetwork.

[0126] Write or refresh the candidate generator subnetworks that meet the performance and reliability requirements to the candidate subnetwork pool. When candidate subnetwork pool If the capacity limit is not reached, directly write to a candidate generator subnetwork that meets the conditions; when the candidate subnetwork pool... When the capacity limit is reached, the candidate sub-network pool is maintained based on the score. Replacement is performed by removing candidate generator subnetworks that have not been updated for a long time, have low reliability scores, or low overall maintenance scores, while retaining historically high-quality candidate generator subnetworks that have performed stably across search rounds.

[0127] Therefore, candidate subnetwork pool It not only preserves the high-quality candidates of the current round, but also continuously saves, refreshes and reuses the historical reliable candidate generator subnetwork across search rounds, providing candidate sources and sampling basis for subsequent annealing three-way sampling training.

[0128] Step 4: Perform annealing three-way sampling training guided by candidate sub-network pool, and output the target generator architecture when the termination condition is met.

[0129] The current set of high-quality candidates is obtained in each round of interleaved evolutionary search. And complete the candidate subnetwork pool After dynamic maintenance, it is determined whether the current search process meets the termination condition. Specifically, the termination condition is considered met when the number of evolutionary search rounds reaches the preset total number of evolutionary searches, or when the number of training rounds reaches the preset number of formal search rounds.

[0130] If the termination condition is met, the process of saving the high-quality candidate set and outputting the target generator architecture begins. During this process, only the formal candidate generator subnetworks that have undergone inheritance weight performance evaluation and have candidate memory entries are retained. Let the formal candidate set that has completed inheritance weight performance evaluation and has valid candidate memory entries up to the current search phase be denoted as . Formal candidate set From the current high-quality candidate set and candidate subnetwork pool The candidate generator subnetworks that have already undergone inheritance weight performance evaluation are merged, and duplicate candidates are removed; shadow candidates that have not undergone inheritance weight performance evaluation are not included in the formal candidate set. Subsequently, in the formal candidate set In this process, the optimal candidate for FID is determined based on the FID metric, and the optimal candidate for IS is determined based on the IS metric, as shown below:

[0131] ,

[0132] ,

[0133] in, This represents the candidate generator subnetwork with the best FID metric up to the current search phase. This represents the candidate generator subnetwork with the best IS metric up to the current search phase. A smaller FID metric indicates that the generated sample distribution is closer to the real sample distribution, while a larger IS metric indicates better quality and diversity of the generated samples.

[0134] After the search is completed, the candidate generator subnetwork with the best FID index is output first as the target generator architecture; alternatively, the candidate with the best IS index, the candidates in the final high-quality candidate set, and the global elite candidates across rounds can be used as alternative generator architectures for subsequent independent retraining or experimental verification.

[0135] If the current search process does not meet the termination condition, then the search space will be shrunk. Current high-quality candidate set and the candidate subnetwork pool after this round of maintenance Construct a three-way annealing sampling distribution and train the sub-network according to the distribution to update the shared weights. Let the encoding of the current training sub-network in the t-th training epoch be... The sampling process is as follows:

[0136] ,

[0137] Among them, the overall sampling distribution It consists of three sampling distributions:

[0138] ,

[0139] in, , , They represent the search space respectively. Current high-quality candidate set and candidate subnetwork pool The sampling distribution on; This represents the sampling probability in the search space, i.e., from the search space. The sampling probability is used to maintain the ability to explore structures; This represents the sampling probability of the candidate subnetwork pool, i.e., from the sampling branch outside the search space, it represents the probability of selecting a candidate subnetwork pool. The sampling probability is used to control the reuse intensity of historical high-quality structures.

[0140] To shift the training process from extensive exploration in the early stages to the reuse of high-quality candidates in the later stages, an annealing progress coefficient is defined. :

[0141] ,

[0142] Search space sampling probability and candidate subnetwork pool sampling probability They are changed in the following ways respectively:

[0143] ,

[0144] ,

[0145] in, This marks the start of the formal search round. The sampling ratio annealing length; , These represent the starting and ending values ​​of the sampling probability in the search space, respectively. , These represent the initial and final values ​​of the sampling probability for the candidate sub-network pool, respectively. As training progresses, Gradually reduce, This process is gradually improved, shifting the training process from structure exploration-driven to historically reliable candidate reuse-driven, balancing early-stage global exploration with later-stage reuse of high-quality historical candidates. Historically reliable candidates refer to candidate generator subnetworks that have demonstrated high stability after multiple rounds of evaluation.

[0146] When the sampling source is the current high-quality candidate set When sampling is performed, approximately uniform sampling is preferred to avoid over-concentration of training on a few high-quality candidates; when the sampling source is a candidate sub-network pool In this case, it is preferable to perform priority sampling with a uniform smoothing term based on the training sampling score, so that candidates with better performance, higher stability and newer information have a higher chance of training exposure, while retaining the necessary diversity.

[0147] The candidate subnetwork pool is trained using the above-mentioned three-way annealing sampling. The historically reliable candidate subnetworks gain higher training exposure opportunities in later stages, enabling the shared weights of the weight-sharing generator supernetwork to gradually adapt to reliable and high-quality candidate structures. The updated shared weights continue to be used in the next round of interleaved evolutionary search for candidate inherited weight performance evaluation, candidate memory entry refresh, reliability score calculation, and candidate subnetwork pool maintenance, and then return to step 3 to execute the next round of interleaved evolutionary search, thus forming a closed loop of interleaved evolutionary search, candidate memory entry refresh, candidate subnetwork pool maintenance, annealing three-way sampling training, shared weight update, and the next round of interleaved evolutionary search.

[0148] In an alternative implementation, a shadow candidate set exempt from real evaluation is also provided. Shadow candidate set From the current high-quality candidate set or candidate sub-network pool High-quality candidates are generated through crossover, mutation, or local perturbation, and only those are incorporated into the candidate sub-network pool. The corresponding sampling branches are used to enhance the local structural diversity of training sampling. Shadow candidate set It does not participate in real performance evaluation, formal candidate ranking, or the saving of the best candidate results.

[0149] In a specific embodiment, taking generator architecture search on the CIFAR-10 dataset as an example, the input image resolution is 32×32, the generator batch size is 256, the discriminator batch size is 256, the initial candidate generator subnetwork set size N is 32, and the current high-quality candidate set size K is 12. (Initialization phase parameter size upper limit) It can be set to 15.0MB; the weight-sharing generator supernetwork is pre-trained for about 40 training rounds, and enters the formal search phase from the 41st training round; in the formal search phase, an evolutionary update is performed every 10 training rounds, for a total of 17 rounds of interleaved evolutionary search.

[0150] In this specific embodiment, the candidate performance evaluation in the formal search phase can use 5000 generated images, with an evaluation batch size of 100; the dynamic parameter budget can use... =13.76MB =1.00、 =0.80、 =0.60、 =6、 =12 segmented annealing setting; early re-evaluation interval for parent candidates It can be set to 2, and the starting round will be refreshed later. It can be set to 9. In three-channel sampling, It can be annealed from 0.90 to 0.20. Annealing can be performed from 0.55 to 0.80, annealing length It can be set to 120 training rounds.

[0151] When the candidate generator architecture obtained by the above method is independently retrained, on the CIFAR-10 dataset, the generator parameter size is 8.785MB, and the FID=8.17 and IS=8.91±0.08 are achieved when evaluated with 50,000 generated images; on the CIFAR-100 dataset, the generator parameter size is 8.78MB, achieving FID=14.55 and IS=8.88±0.09; and on the STL-10 dataset, FID=18.67 and IS=10.27±0.12. These results demonstrate that the generator architecture obtained by this invention not only exhibits good independent retraining performance on the target dataset but also possesses a certain degree of adaptability under different data distributions and image resolution settings.

[0152] In summary, this invention, through candidate sub-network pool maintenance, annealed three-way sampling training, annealed dynamic parameter budget constraints, and a phased differentiated re-evaluation mechanism, balances candidate reuse, evaluation information staleness control, training diversity, and parameter scale control during the generative adversarial network architecture search process, and improves the usability of search results in subsequent independent retraining stages. Figure 3 and Figure 4 As shown, this embodiment verifies the changing trends of the IS and FID indices of the candidate generator subnetwork during the search process of the method of the present invention.

[0153] Specifically, during the formal search process, the changes in the IS index of the candidate generator subnetworks at each evolutionary stage are recorded. For example... Figure 3 As shown, the IS index of the candidate generator subnetwork generally shows an upward trend as the search process progresses, indicating that the present invention can gradually improve the generation quality and sample diversity of the candidate generator subnetwork during the search process.

[0154] Simultaneously, the changes in the FID index of the candidate generator subnetworks at each evolutionary stage are recorded. For example... Figure 4 As shown, the FID index of the candidate generator subnetwork generally shows a downward trend as the search process progresses, indicating that the present invention can gradually screen out candidate generator architectures whose generated sample distribution is closer to the real sample distribution.

[0155] comprehensive Figure 3 and Figure 4 It is understood that the present invention, through candidate sub-network pool maintenance, annealing three-way sampling training, annealing dynamic parameter budget constraint, and phased differentiated re-evaluation mechanism, can continuously discover candidate generator sub-networks with better performance during the search process and provide a target generator architecture for subsequent independent retraining.

Claims

1. A generative adversarial network architecture search method guided by a candidate subnetwork pool, characterized in that, Includes the following steps: S1, construct a weight-shared generator supernetwork for generator architecture search, and initialize a set of candidate generator subnetworks based on a preset generator search space; at the same time, establish candidate memory entries for each candidate generator subnetwork. S2, Before the formal search phase begins, the weight-sharing generator supernetwork is pre-trained; After pre-training, the shared weights from the pre-trained dataset are used to perform a low-fidelity evaluation of the candidate generator subnetwork to obtain candidate performance estimates. Based on these estimates, candidate selection constraints are constructed, and the generator search space is shrunk to obtain the shrunk search space. ; The pre-training phase searches the pre-defined generator space. A randomized candidate generator subnetwork is selected as the current activation path, and adversarial training is performed under the constraint of a fixed-structure discriminator to update the shared weights of the weight-sharing generator supernetwork. During adversarial training, the currently activated candidate generator subnetwork is fed with a random noise vector. As input, output the corresponding generated image sample; Let the first The candidate architecture encoding obtained from sampling in each training round is: The shared weights of the weight-sharing generator supernetwork are: The random noise vector is Then generate image samples. for: ， in, Indicates that it is encoded by the architecture. Activate and inherit shared weights The generator subnetwork, Indicates the starting round of the formal search; After pre-training, the updated shared weights are utilized. Perform low-fidelity evaluation on the candidate generator subnetwork to obtain approximate performance metrics of the candidate structure without requiring independent retraining; Based on the approximate performance indicators, structural screening constraints are constructed, and the search space of the preset generator is applied. Structural feasibility is pruned, retaining only candidate structural regions that meet performance and parameter budget constraints, resulting in a shrunken search space. ; The low-fidelity evaluation refers to: under the condition of unchanged shared weights, by limiting the number of generated samples, to estimate the FID index, IS index and parameter size of the candidate generator sub-network at low cost, so as to characterize the relative performance level of the candidate structure. S3 performs interleaved evolutionary search with budgeted annealing based on the shrunken generator search space and dynamically maintains a pool of candidate subnetworks. S4, obtaining the current set of high-quality candidates in each round of interleaved evolutionary search. And complete the candidate subnetwork pool After dynamic maintenance, determine whether the current search process meets the termination condition; If the termination condition is not met, then the search space is shrunk. Current high-quality candidate set and the candidate subnetwork pool after this round of maintenance An annealing three-way sampling distribution is constructed, and a training sub-network is selected according to the annealing three-way sampling distribution to update the shared weights of the weight-sharing generator supernetwork; the updated shared weights are then used for candidate evaluation, candidate memory entry update, candidate reliability score calculation, candidate sub-network pool maintenance, and interleaved evolutionary search. If the termination condition is met, save the set of high-quality candidates and output the target generator architecture.

2. The generative adversarial network architecture search method based on candidate subnetwork pooling as described in claim 1, characterized in that, The candidate memory entries include at least the candidate architecture encoding, inheritance weight evaluation results, parameter size, evaluation round, most recent training round, number of sampled training rounds, and number of historical evaluation rounds; wherein, the inheritance weight evaluation results include the original FID value and IS value.

3. The generative adversarial network architecture search method based on candidate sub-network pooling as described in claim 1, characterized in that, In step S3, the specific implementation process of dynamically maintaining the candidate sub-network pool is as follows: After the training rounds reach the preset starting round for the formal search, the formal search phase begins in the shrunken search space. An interleaved evolutionary search is performed at a preset cycle to generate candidate generator subnetworks to be evaluated, and they are screened by annealing dynamic parameter budget constraints. The screened candidate generator subnetworks are then evaluated by inheritance weights to obtain the performance record for the current round. The performance record includes at least the FID index, IS index and parameter size. The performance records are written into the corresponding candidate memory entries, and the smoothed FID value and stability deviation are updated based on the historical evaluation sequence, the most recent evaluation round, the most recent training round, and the number of sampled training sessions. The staleness of the evaluation information and the intensity of the training feedback are further determined to construct a candidate reliability score. The current high-quality candidate set is obtained based on the FID index, IS index, parameter size and reliability score, and the candidate sub-network pool is dynamically maintained based on the current high-quality candidate set.

4. The generative adversarial network architecture search method based on candidate sub-network pooling as described in claim 3, characterized in that, The specific steps for performing an interleaved evolutionary search according to a preset search cycle include: S31, in the In the interleaved evolutionary search, the current search population... As the basic candidate set, and based on the structural perturbation mechanism, an extended candidate set is generated. Subsequently, the basic candidate set and the extended candidate set are merged to obtain the joint candidate set. : ； The structural disturbance mechanism includes one or more of the following: structural reorganization, structural intersection, structural variation, local path disturbance, or resampling. S32, for candidate joint sets Apply annealing-type dynamic parameter budget constraints, the first Upper limit of parameter budget for round-robin evolutionary search for: ， in, The preset parameter size; These are non-incremental annealed dynamic coefficients used to gradually tighten parameter budget constraints as the search rounds progress; Subsequently, the parameter budget constraint is used as a candidate selection constraint, retaining only candidate generator subnetworks whose parameter size does not exceed the upper limit of the parameter budget in the r-th round, thus obtaining the selected candidate set: ， in, This represents the architecture encoding of the i-th candidate generator subnetwork. Represents the candidate generator subnetwork The parameter scale, This represents the candidate set after filtering by parameter budget constraints; S33, the filtered candidate set The candidate generator subnetwork inherits the shared weights of the current weight-sharing generator supernetwork and generates a preset number of image samples. Based on these image samples, the IS index is calculated, and based on the distribution difference between the image samples and real samples, the FID index is calculated. Simultaneously, the parameter scale of the candidate generator subnetwork is statistically analyzed to form a performance record for the current round. A phased differentiated refresh mechanism is used for weight inheritance performance evaluation: the new candidate set generated in this round... Directly perform inheritance weight performance evaluation; for the current search population If there is already a candidate generator subnetwork of candidate memory entries, then it is determined whether to perform an inheritance weight performance refresh evaluation based on the most recent evaluation round, the number of historical evaluations, and the search phase of the historical candidate generator subnetwork. S34, write the current round performance record into the corresponding candidate memory entry, and update the candidate state variables based on the candidate historical evaluation sequence. The candidate state variables include smoothed FID value, stability bias, evaluation information staleness and training feedback strength. S35, calculate the candidate reliability score from the candidate state variables, and then apply the IS index, FID index, parameter size, and candidate reliability score to the filtered candidate set. By prioritizing and selecting candidates, the current set of high-quality candidates is obtained. .

5. The generative adversarial network architecture search method based on candidate sub-network pooling as described in claim 1, characterized in that, The candidate reliability score is calculated based on the stability bias of the candidate generator sub-network, the staleness of the evaluation information, and the training feedback strength; where a lower staleness of the evaluation information indicates more recent evaluation information; the staleness of the evaluation information is determined according to the evolutionary round corresponding to the most recent inherited weight performance evaluation of the candidate generator sub-network. ， in, Represents the candidate generator subnetwork In the The staleness of evaluation information in round evolutionary search Represents the candidate generator subnetwork The evolutionary cycle corresponding to the most recent inheritance weight performance evaluation; The training feedback strength is determined based on the number of times the candidate generator subnetwork has been sampled during recent shared weight training: ， in, Indicates the candidate generator subnetwork at the th The training feedback intensity corresponding to each training round express Candidate Generator Subnetwork The number of times the sample was trained during the recent training phase; After normalizing the stability bias, evaluation information staleness, and training feedback strength, the candidate reliability score is obtained: ， in, Represents the candidate generator subnetwork In the Reliability score in round evolutionary search , and These represent the stability bias after normalization, the staleness of the evaluation information, and the intensity of the training feedback, respectively.

6. The generative adversarial network architecture search method based on candidate subnetwork pooling as described in claim 1, characterized in that, In step S3, the dynamic maintenance of the candidate sub-network pool includes candidate status updates, adding new candidates to the pool, candidate replacement, and retention of high-quality candidates. Candidate Subnetwork Pool Candidate generator subnetworks are recorded using architecture encoding as an index. The historical state information is obtained by associating it with the corresponding candidate memory entries to obtain the number of times the sampled training has been performed. The historical state information includes at least the original FID value, IS value, parameter size, evaluation round, most recent training round, and number of evaluations. The candidate sub-network pool is used to continuously model the state evolution of the candidate generator sub-network across search rounds. For the candidate generator subnetwork that writes candidate memory entries for the first time, its smoothed FID value and stability bias are initialized as follows: ， ； When the candidate generator subnetwork When the inheritance weight performance evaluation is performed again, the updated smoothed FID value and stability bias are as follows: ， ， in, This is the smoothing coefficient, and its value range is... ; This indicates the smoothed FID value before this update; Represents the candidate generator subnetwork The original FID value obtained in the most recent inheritance weight performance evaluation; Represents the candidate generator subnetwork Smoothed FID values; Represents the candidate generator subnetwork Stability deviation; If the candidate generator subnetwork Already exists in the candidate subnetwork pool In the middle, for candidate sub-network pools The state records are incrementally updated, including synchronous updates of parameter size, evaluation rounds, most recent training rounds, and number of evaluations; If the candidate generator subnetwork The candidate subnetwork pool does not exist. If the capacity limit is not reached, the candidate generator subnetwork will be moved to the next network. Added to candidate subnetwork pool ; When candidate subnetwork pool When the capacity limit is reached, a candidate sub-network pool maintenance score is constructed based on the original FID value, smoothed FID value, stability deviation, parameter size, and candidate reliability score of the candidate generator sub-network. The retained and replacement objects in the candidate sub-network pool are determined according to the candidate sub-network pool maintenance score, thereby realizing the dynamic update of the candidate sub-network pool.

7. The generative adversarial network architecture search method based on candidate subnetwork pool guidance according to claim 1, characterized in that, In step S4, the specific implementation process of selecting the training sub-network according to the three-way annealing sampling distribution is as follows: Let the encoding of the current training subnetwork in the t-th training round be... The sampling process is as follows: ， Among them, the overall sampling distribution It consists of three sampling distributions: ， in, , , These represent the shrunk search space, respectively. Current high-quality candidate set and candidate subnetwork pool The sampling distribution on; Indicates the sampling probability of the search space; This represents the sampling probability of the candidate subnetwork pool; Define annealing progress coefficient as follows: ， The sampling probability evolves as follows: ， ， in, This marks the start of the formal search round. The sampling ratio annealing length; , These represent the starting and ending values ​​of the sampling probability in the search space, respectively. , These represent the starting and ending values ​​of the sampling probability for the candidate sub-network pool, respectively. When the sampling source is the current high-quality candidate set When using this method, approximately uniform sampling is employed. When the sampling source is a candidate subnetwork pool At that time, priority sampling with a uniform smoothing term is performed based on the training sampling score.

Images