Model structure searching method and device, computer device and storage medium
By optimizing model parameters and policy selection probabilities using fractional gradient descent functions, the problem of slow search speed in traditional network structures is solved, and a fast and high-performance image processing model is constructed.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- ZHEJIANG DAHUA TECH CO LTD
- Filing Date
- 2026-03-05
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional network structure search methods require a lot of human experience and have a large search space, resulting in slow model building speed.
The model parameters and policy selection probability are optimized using a fractional gradient descent function. The model parameters and policy selection probability are adjusted using training and validation samples respectively until they stabilize, thus constructing an image processing model.
It realizes the potential of the model to quickly and accurately fit the current candidate structure, avoids local optima and oscillations, and builds an image processing model with superior performance and strong generalization ability.
Smart Images
Figure CN121787480B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of automated network structure search technology, and in particular to a model structure search method, apparatus, computer device, and storage medium. Background Technology
[0002] As artificial intelligence (AI) becomes increasingly integrated into daily life, more and more application scenarios can utilize AI models. However, general model structures are insufficient to address all application scenarios. Therefore, it is crucial to develop models tailored to different application scenarios, tasks, and datasets.
[0003] In traditional technologies, Neural Architecture Search (NAS) is typically used. Experienced AI algorithm engineers go through processes such as design, training, and testing to find the optimal neural network structure in a given search space under a specified performance evaluation strategy, thereby completing the model construction.
[0004] However, network structure search requires AI algorithm engineers to have extensive model building experience, and because the search space of network structure search contains a large amount of data, there is a problem of slow model building speed. Summary of the Invention
[0005] Therefore, it is necessary to provide a model structure search method, apparatus, computer equipment, and storage medium that can quickly construct models to address the aforementioned technical problems.
[0006] Firstly, this application provides a model structure search method, including:
[0007] Obtain training samples, validation samples, and the structure of the model to be searched;
[0008] For each training iteration, the model parameters in the model structure to be searched are adjusted based on the training samples until the model parameters in the model structure to be searched under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters.
[0009] Based on the validation samples, the policy selection probability of each feature processing strategy in the structure to be searched is adjusted until the policy selection probability in the structure to be searched in this training iteration tends to be stable; the policy selection probability is adjusted based on the fractional gradient descent function of the policy selection probability.
[0010] When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined based on the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0011] In one embodiment, the strategy selection probability of each feature processing strategy in the search model structure is adjusted based on the validation samples, including:
[0012] For each probability iteration, a fractional gradient descent function is selected based on the strategy. Based on the validation samples in this probability iteration, the probability gradient values corresponding to different feature processing strategies in the model structure to be searched are determined.
[0013] Based on the probability gradient value, adjust the strategy selection probability of the corresponding feature processing strategy.
[0014] In one embodiment, a fractional gradient descent function based on the strategy selection probability is used to determine the probability gradient values corresponding to different feature processing strategies in the model structure to be searched, based on the verification samples under the current probability iteration process.
[0015] For each feature processing strategy in the model structure to be searched, the probability gradient related parameters are determined based on the current strategy selection probability and the historical strategy selection probability of the corresponding feature processing strategy.
[0016] Based on the preset fractional order, probability gradient related parameters, and gamma function, the probability gradient values corresponding to different feature processing strategies in the model structure to be searched are determined.
[0017] In one embodiment, adjusting the strategy selection probability of the corresponding feature processing strategy based on the probability gradient value includes:
[0018] For each feature processing strategy in the model structure to be searched, the target learning rate of the corresponding feature processing strategy is determined based on the learning rate of the model structure to be searched and the strategy selection probability of the corresponding feature processing strategy.
[0019] The probability of selecting the corresponding feature processing strategy is adjusted based on the target learning rate and probability gradient value.
[0020] In one embodiment, the model structure to be searched includes at least one node; there is at least one node pointing relationship between different nodes; in the node pointing relationship from the first node to the second node, the first node corresponds to the first feature data, the node pointing relationship corresponds to the feature processing strategy, and the second node corresponds to the second feature data obtained after processing the first feature data using the feature processing strategy.
[0021] In one embodiment, the method further includes:
[0022] Acquire the image data to be processed;
[0023] Based on the image processing model, the image processing result of the image data is determined according to the image data.
[0024] Secondly, this application also provides a model structure search device, comprising:
[0025] The acquisition module is used to acquire training samples, validation samples, and the structure of the model to be searched.
[0026] The training module is used to adjust the model parameters in the search model structure based on the training samples for each training iteration until the model parameters in the search model structure under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters.
[0027] The validation module is used to adjust the strategy selection probability of each feature processing strategy in the search model structure based on the validation samples, until the strategy selection probability in the search model structure under the current training iteration tends to be stable; the strategy selection probability is adjusted based on the fractional gradient descent function of the strategy selection probability.
[0028] The module is used to determine the target feature processing strategy based on the strategy selection probability in the model structure to be searched, when the model parameters and strategy selection probabilities of the model structure to be searched tend to be stable under different iterative training processes, and to construct an image processing model based on the target feature processing strategy.
[0029] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:
[0030] Obtain training samples, validation samples, and the structure of the model to be searched;
[0031] For each training iteration, the model parameters in the model structure to be searched are adjusted based on the training samples until the model parameters in the model structure to be searched under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters.
[0032] Based on the validation samples, the policy selection probability of each feature processing strategy in the structure to be searched is adjusted until the policy selection probability in the structure to be searched in this training iteration tends to be stable; the policy selection probability is adjusted based on the fractional gradient descent function of the policy selection probability.
[0033] When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined based on the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0034] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:
[0035] Obtain training samples, validation samples, and the structure of the model to be searched;
[0036] For each training iteration, the model parameters in the model structure to be searched are adjusted based on the training samples until the model parameters in the model structure to be searched under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters.
[0037] Based on the validation samples, the policy selection probability of each feature processing strategy in the structure to be searched is adjusted until the policy selection probability in the structure to be searched in this training iteration tends to be stable; the policy selection probability is adjusted based on the fractional gradient descent function of the policy selection probability.
[0038] When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined based on the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0039] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:
[0040] Obtain training samples, validation samples, and the structure of the model to be searched;
[0041] For each training iteration, the model parameters in the model structure to be searched are adjusted based on the training samples until the model parameters in the model structure to be searched under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters.
[0042] Based on the validation samples, the policy selection probability of each feature processing strategy in the structure to be searched is adjusted until the policy selection probability in the structure to be searched in this training iteration tends to be stable; the policy selection probability is adjusted based on the fractional gradient descent function of the policy selection probability.
[0043] When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined based on the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0044] The aforementioned model structure search method, apparatus, computer equipment, and storage medium separate the optimization of model parameters from the optimization of structure policy probabilities in each iteration. Training samples and fractional gradient descent are used to optimize network parameters, ensuring the model can quickly and accurately fit the potential of current candidate structures. Its fractional-order characteristic allows for more flexible memory and long-range dependency capture, potentially avoiding local optima and accelerating internal convergence. Validation samples and fractional gradient descent are used to independently optimize policy selection probabilities. This allows the search process to be directly guided by validation performance, dynamically and smoothly evaluating and reinforcing excellent structural paths, further improving the stability of probability updates. This ensures that each iteration update is reliable and sufficient, avoiding oscillations or divergences caused by inappropriate alternation frequencies in traditional methods. When the model parameters and policy selection probabilities reach a stable state after multiple training iterations, the determined target feature processing strategy is selected with high confidence based on sufficient and stable evaluation. This ensures that the constructed image processing model not only has superior performance but also stronger generalization ability, and the overall search process is more efficient and reliable. Attached Figure Description
[0045] To more clearly illustrate the technical solutions in the embodiments or related technologies of this application, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0046] Figure 1 This is an application environment diagram of a model structure search method provided in this embodiment;
[0047] Figure 2A A flowchart illustrating the first model structure search method provided in this embodiment;
[0048] Figure 2B This is a schematic diagram of a node connection provided in this embodiment;
[0049] Figure 2C This is a schematic diagram of a basic module provided in this embodiment;
[0050] Figure 2D This is a schematic diagram illustrating the selection of a target feature processing strategy provided in this embodiment;
[0051] Figure 3This is a flowchart illustrating a model parameter adjustment step provided in this embodiment;
[0052] Figure 4 A flowchart illustrating the probability adjustment step for the first strategy selection provided in this embodiment;
[0053] Figure 5 This is a flowchart illustrating a probability gradient value determination step provided in this embodiment.
[0054] Figure 6 A flowchart illustrating the probability adjustment steps for the second strategy provided in this embodiment;
[0055] Figure 7 This is a flowchart illustrating an image processing step provided in this embodiment;
[0056] Figure 8 This is a flowchart illustrating the second model structure search method provided in this embodiment;
[0057] Figure 9 This is a structural block diagram of a model structure search device provided in this embodiment;
[0058] Figure 10 This is an internal structural diagram of a computer device provided in this embodiment. Detailed Implementation
[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0060] The model structure search method provided in this application embodiment can be applied to, for example... Figure 1 In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on a cloud or other network server. The computer device acquires training samples, validation samples, and the model structure to be searched; for each training iteration, based on the training samples, the model parameters in the model structure to be searched are adjusted until the model parameters in the model structure to be searched under this training iteration tend to stabilize; the model parameters are adjusted based on the fractional-order gradient descent function of the model parameters.
[0061] Based on the validation samples, the strategy selection probabilities of each feature processing strategy in the search model structure are adjusted until the strategy selection probabilities in the search model structure tend to stabilize during the current training iteration. The strategy selection probabilities are adjusted based on the fractional-order gradient descent function of the strategy selection probabilities. When the model parameters and strategy selection probabilities of the search model structure tend to stabilize under different training iterations, the target feature processing strategy is determined according to the strategy selection probabilities in the search model structure, and an image processing model is constructed based on the target feature processing strategy. The computer device can be either a terminal or a server. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can be smartwatches, smart bracelets, head-mounted devices, etc. Server 104 can be implemented using a standalone server or a server cluster composed of multiple servers.
[0062] In one exemplary embodiment, such as Figure 2A As shown, a model structure search method is provided, which can be applied to... Figure 1 Taking a computer device as an example, the explanation includes the following steps S201 to S204. Wherein:
[0063] S201 obtains training samples, validation samples, and the structure of the model to be searched.
[0064] In this context, training samples can be understood as a dataset used to adjust the model parameters of the model to be searched. Validation samples can be understood as a dataset used to evaluate model performance, guide structure search, and adjust the probability of strategy selection. It should be noted that in this embodiment, the ratio of training samples to validation samples can be 1:1. Optionally, both training samples and validation samples can be image data used as samples.
[0065] The model structure to be searched can be understood as a supernet composed of nodes and edges, representing the search space for all possible model structures. Each model structure includes at least one node; there is at least one node-to-node relationship between different nodes; in the node-to-node relationship where the first node points to the second node, the first node corresponds to the first feature data, the node-to-node relationship corresponds to the feature processing strategy, and the second node corresponds to the second feature data obtained after processing the first feature data using the feature processing strategy. The first node in the model structure can represent the input data, such as image data. The advantage of this setup is that the clear node-to-node relationships explicitly define the data flow propagation path, allowing each feature processing strategy to be viewed as a learnable discrete choice whose probability distribution can be optimized through gradient descent. This achieves end-to-end automated structure discovery, greatly enriching the model's performance capabilities and search space. Furthermore, its modular nature allows for efficient searching of optimal substructures for different computational budgets or performance requirements, ultimately significantly improving the efficiency and flexibility of structure search while ensuring powerful representation learning capabilities.
[0066] For example, such as Figure 2B The diagram illustrates node connections. Boxes represent nodes, dashed lines represent node-to-node relationships, and the letter 'A' on the dashed lines represents at least one selectable feature processing strategy. The feature processing strategies corresponding to different node-to-node relationships can include convolution operations, pooling operations, skip connections, mish activation functions, normalization operations, and none. Convolution operations can include 1×1 convolution (Conv), 3×3 convolution, and 5×5 convolution. Pooling operations can include max pooling and average pooling. Normalization operations can include batch normalization.
[0067] For example, in this embodiment, the second node can be characterized as: Among them, x (i) Representing the first node, x (j) Characterize the second node, The processing strategy for the k-th feature between the first node and the second node is represented.
[0068] In some embodiments, the model processing task corresponding to the model to be searched is determined; and the training samples, validation samples, and the structure of the model to be searched are determined based on the model processing task.
[0069] For example, when the model processing task is image processing, sample image data of a scene related to image processing is acquired; based on a preset partitioning ratio (e.g., 1:1), the sample image data is divided into training samples and validation samples; and a model structure to be searched that satisfies the model processing task is determined. The model processing task can specifically include at least one of image classification and image recognition tasks. Based on this, both the training samples and validation samples in this embodiment can be image data.
[0070] It should be noted that the method for determining the search model structure that satisfies the model processing task in this embodiment can be based on human experience or on a large number of experiments. The specific determination method is not limited in this embodiment.
[0071] It should be noted that in this embodiment, different nodes can also be combined to obtain basic modules, and then the corresponding search model structure can be constructed from the basic modules. For example... Figure 2C The diagram shown illustrates the basic module, which contains two input basic units (M). k-1 and M k-2 ), four nodes (N0, N1, N2, and N3) and one output basic unit (M) k1 The nodes and output units are connected based on the concatenation (Concat) function. The probability of strategy selection between different nodes can exist in the form of a probability distribution. In the figure, Op represents the operation index, that is, there are 9 feature processing strategies between different nodes.
[0072] For example, taking image recognition as an example, the process involves acquiring sample image data containing the target object; dividing the sample image data into training samples and validation samples based on a 1:1 partitioning ratio; and determining that the search model structure that satisfies the model processing task can include 24 interconnected basic modules, with the node pointing relationships between different nodes corresponding to 9 different feature processing strategies, and the strategy selection probability between node i and node j being used... This indicates that the nine feature processing strategies can be operated in a mixture according to the probability distribution. Used to represent.
[0073] For example, in this embodiment, the probability distribution of the probability of selecting a feature processing strategy among different nodes can be obtained by normalizing the probability distribution of the probability of selecting a feature processing strategy among nodes based on the normalization (Softmax) function shown in the following formula (1):
[0074] (1)
[0075] in, Characterizing the probability distribution, r i The probability distribution representing node i.
[0076] It should be noted that, in this embodiment, preprocessing can be performed on the corresponding image data before or after dividing the training and validation samples to ensure the processing efficiency of subsequent data processing. Preprocessing may include grayscale conversion to compress data volume, and Gaussian filtering to enhance target contours and improve algorithm performance, suppressing noise and weakening background information. Various preprocessing methods exist, and this embodiment does not limit the specific methods described herein.
[0077] For each training iteration, S202 adjusts the model parameters in the search model structure based on the training samples until the model parameters in the search model structure under this training iteration tend to be stable.
[0078] The training iteration process can be understood as a complete training process that adjusts the model parameters and policy selection probabilities. After completing one training iteration and moving to the next, the model parameters will change again in the next iteration due to the adjustment of the policy selection probabilities in the previous iteration. Therefore, the model parameters need to be readjusted in the next iteration until the model parameters and policy selection probabilities of the search model structure under different training iterations tend to stabilize. The training iteration process can include a model iteration process and a probability iteration process. The model iteration process is used to adjust the model parameters, and the probability iteration process is used to adjust the policy selection probabilities.
[0079] Here, model parameters can be understood as the trainable weights (such as the weights and biases of convolutional kernels) included in the feature processing strategy. The model parameters are adjusted based on the fractional gradient descent function of the model parameters; the fractional gradient descent function can be understood as an optimization algorithm based on fractional calculus, which considers not only the current gradient but also incorporates historical gradient information when updating variables, and has the property of being fully local.
[0080] In some embodiments, during each training iteration, for each model iteration, training samples are input into the model structure to be searched, and the model parameters of each feature processing strategy in the model structure to be searched are adjusted; based on the similarity between the model parameters in the current model iteration and the model parameters in the historical model iterations, it is determined whether the model parameters in the current model iteration tend to be stable; if the model parameters are unstable in the current model iteration, the steps of inputting training samples into the model structure to be searched and adjusting the model parameters of each feature processing strategy in the model structure to be searched are returned until the model parameters in the current model iteration tend to be stable, and it is determined that the model parameters in the model structure to be searched under the current training iteration tend to be stable, and the next step is executed.
[0081] For example, the method for determining whether the model parameters in the current model iteration are stable based on the similarity between the model parameters in the current iteration and those in historical model iterations is as follows: For each historical model iteration, based on the similarity between the model parameters in the current iteration and those in historical model iterations; if the similarity is less than a preset similarity threshold, the model parameters are determined to be unstable; otherwise, if the similarity is not less than the preset similarity threshold, the model parameters are determined to be stable; if the model parameters are stable in a consecutive preset number of historical model iterations, the model parameters in the searched model structure are determined to be stable. For example, if the model parameters are stable in the current iteration after three consecutive historical model iterations, it proves that the model parameters in the searched model structure are stable.
[0082] S203 adjusts the strategy selection probability of each feature processing strategy in the search model structure based on the validation samples until the strategy selection probability in the search model structure tends to stabilize during the current training iteration.
[0083] The policy selection probability can be understood as the probability of choosing the corresponding feature processing policy between two nodes. The policy selection probability is adjusted based on the fractional gradient descent function of the policy selection probability.
[0084] In some embodiments, for each probability iteration, validation samples are input into the model structure to be searched, and the strategy selection probabilities of each feature processing strategy in the model structure to be searched are adjusted. For the probability distributions corresponding to different feature processing strategies between two nodes, based on the similarity between the probability distributions between the two nodes in this probability iteration and the probability distributions between the two nodes in historical model iterations, it is determined whether the probability distribution in this probability iteration tends to be stable. If the probability distribution is unstable in this probability iteration, the step of inputting validation samples into the model structure to be searched and adjusting the strategy selection probabilities of each feature processing strategy in the model structure to be searched is returned until the probability distribution in the model structure to be searched in this training iteration tends to be stable, and the strategy selection probabilities in the model structure to be searched in this training iteration tend to be stable, then the next step is executed.
[0085] It should be noted that the implementation method of determining whether the probability distribution tends to be stable in the current probability iteration based on the similarity between the probability distribution between two nodes in the current probability iteration and the probability distribution between two nodes in the historical model iteration is the same as or similar to the implementation method in the above embodiment of determining whether the model parameters tend to be stable in the current model iteration based on the similarity between the model parameters in the current model iteration and the model parameters in the historical model iteration. This embodiment does not limit this.
[0086] S204: When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined according to the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0087] The target feature processing strategy can be understood as the feature processing strategy used to construct the image processing model.
[0088] In some embodiments, it is determined whether the model parameters and policy selection probabilities of the model structure to be searched tend to be stable under different iterative training processes; if any value of the model parameters and policy selection probabilities of the model structure to be searched is unstable under different iterative training processes, the process returns to step S202, which adjusts the model parameters of each feature processing strategy in the model structure to be searched based on the training samples until the model parameters in the model structure to be searched under the current training iteration tend to be stable; if the model parameters and policy selection probabilities of the model structure to be searched tend to be stable under different iterative training processes, the target feature processing strategy is determined based on the policy selection probabilities in the model structure to be searched, and an image processing model is constructed based on the target feature processing strategy.
[0089] For example, a way to determine whether the model parameters and policy selection probabilities of the model structure to be searched tend to be stable under different training iterations can be: determine whether there are consecutive training iterations in which the model parameters and policy selection probabilities of the model structure to be searched are similar; if so, it proves that the model parameters and policy selection probabilities of the model structure to be searched tend to be stable; if any value in the model parameters and policy selection probabilities of the model structure to be searched is not similar to the corresponding value in the previous training iteration in consecutive training iterations, it proves that the model parameters and policy selection probabilities of the model structure to be searched are unstable.
[0090] It should be noted that if the model parameters and policy selection probabilities of the model structure to be searched tend to stabilize under different training iterations, it proves that the model parameters will not change significantly due to the change in policy selection probabilities, and the policy selection probabilities will not change accordingly due to the change in model parameters. At this point, reliable model parameters have been found, and feature processing models can be accurately selected from different nodes.
[0091] For example, the method of determining the target feature processing strategy based on the selection probability of each strategy in the model structure to be searched, and constructing an image processing model based on the target feature processing strategy is as follows: for each feature processing strategy between different nodes, the target feature processing strategy is determined based on the selection probability of each strategy in the model structure to be searched; only the target feature processing strategy is retained between different nodes, and other feature processing strategies are deleted to obtain the image processing model.
[0092] Optionally, the method for determining the target feature processing strategy based on the selection probabilities of each strategy in the model structure to be searched can be as follows: First, select the feature processing strategy with the highest selection probability from the selection probabilities of each strategy in the model structure to be searched, and use this as the target feature processing strategy. Second, select a predetermined number of strategies with relatively high selection probabilities from the selection probabilities of each strategy in the model structure to be searched, and use these as the target selection probabilities. Then, fuse the feature processing strategies corresponding to each target selection probability to obtain the target feature processing strategy.
[0093] For example, such as Figure 2D The diagram shows the selection of target feature processing strategies. In the diagram, B represents several feature processing strategies that can exist between the two nodes. Taking node 0 and node 1 as an example, there are originally several feature processing strategies between node 0 and node 1. As the training iteration process proceeds, the probability of selecting a certain feature processing strategy increases. Finally, the feature processing strategy with the highest probability of selection is selected as the target feature processing strategy.
[0094] The aforementioned model structure search method separates the optimization of model parameters from the optimization of structure policy probabilities in each iteration. It uses training samples and fractional gradient descent to optimize network parameters, ensuring the model can quickly and accurately fit the potential of current candidate structures. Its fractional-order characteristic allows for more flexible memory and long-range dependency capture, potentially avoiding local optima and accelerating internal convergence. Validation samples and fractional gradient descent are used to independently optimize policy selection probabilities. This allows the search process to be directly guided by validation performance, dynamically and smoothly evaluating and reinforcing excellent structural paths, further improving the stability of probability updates. This ensures that each iteration update is reliable and sufficient, avoiding oscillations or divergences caused by inappropriate alternation frequencies in traditional methods. When the model parameters and policy selection probabilities reach a stable state after multiple training iterations, the determined target feature processing strategy is selected with high confidence based on sufficient and stable evaluation. This ensures that the constructed image processing model not only has superior performance but also stronger generalization ability, and the overall search process is more efficient and reliable.
[0095] Figure 3This is a flowchart illustrating the model parameter adjustment steps in one embodiment. This embodiment refines the step S202 of the above embodiment, which involves adjusting the model parameters of each feature processing strategy in the search model structure based on training samples. This embodiment provides an optional method for adjusting model parameters, including the following steps:
[0096] For each model iteration, S301 determines the model gradient value of the model structure to be searched based on the fractional gradient descent function of the model parameters and the training samples under this model iteration.
[0097] In one alternative embodiment, training samples are input into the model structure to be searched, and the model structure to be searched determines the model gradient value of the model structure to be searched during the current model iteration process based on the fractional gradient descent function of the model parameters.
[0098] In another alternative embodiment, the model gradient-related parameters are determined based on the current model parameters during the current model iteration and the historical model parameters during the historical model iteration; the model gradient value of the model structure to be searched is determined based on the preset fractional order, gradient model-related parameters, and gamma function.
[0099] For example, in this embodiment, the model gradient value of the model structure to be searched can be determined according to the following formula (2):
[0100] (2)
[0101] in, Characterizes the gradient value of the model. Characterizing the gamma function, The gradient-related parameters of the characterization model, w k W represents the current model parameters during this (k-th) model iteration. k-1 The historical model parameters are represented during the (k-1)th iteration. `n` represents the order (e.g., n=1), `M` represents the total number of iterations, `a` represents the lower bound of the integral (which can be a preset value used to influence the convergence result of the model structure being searched), and `ε` represents the preset compensation parameter. The compensation parameter `ε` can be a small non-negative number used to adjust the model structure within the range of `w`. k When =a, it avoids the situation where the structure of the model to be searched does not converge. Characterizing fractional order, for example .
[0102] S302 adjusts the model parameters of the model structure to be searched based on the model gradient value.
[0103] In one alternative embodiment, the difference between the model parameters and the model gradient values during the current model iteration is used as the model parameters of the adjusted model structure to be searched.
[0104] In another alternative embodiment, the model parameters for adjusting the model structure to be searched are determined based on the learning rate and model gradient values of the model structure to be searched.
[0105] For example, the product of the learning rate and the gradient value of the model structure to be searched is determined as the gradient value of the target model; the difference between the model parameters in the current model iteration process and the gradient value of the target model is used as the model parameters of the adjusted model structure to be searched.
[0106] For example, based on the model parameter update formula shown in formula (3), the model parameters of the model structure to be searched are adjusted according to the model gradient value:
[0107] (3)
[0108] in, The gradient value of the model, w k W represents the model parameters in this (k-th) model iteration process. k+1 Characterizes the model parameters in the next (k+1)th model iteration. It is the learning rate.
[0109] In this embodiment, a fractional gradient descent function is introduced as an optimizer to adjust model parameters. Compared to traditional integer gradient descent methods, it can capture the long-range dependence and historical trends of the gradient function more precisely and effectively. The inherent nonlocality of fractional calculus allows it to consider not only the current gradient but also gradient information from previous steps when updating parameters. This enables smoother navigation in complex non-convex gradient surfaces, effectively alleviating the problem of conventional methods easily getting trapped in local optima. This further accelerates the convergence of the training process and improves the final performance and stability of the discovered structure, making the entire search process more efficient and reliable.
[0110] Figure 4 This is a flowchart illustrating the strategy selection probability adjustment step in one embodiment. This embodiment refines the step S203 of the above embodiment, which adjusts the strategy selection probability of each feature processing strategy in the search model structure based on the validation samples. This embodiment provides an optional method for adjusting the strategy selection probability, including the following steps:
[0111] For each probability iteration, S401 selects a fractional gradient descent function based on the strategy and determines the probability gradient values corresponding to different feature processing strategies in the model structure to be searched based on the validation samples under this probability iteration.
[0112] In some embodiments, validation samples are input to the model structure to be searched, and the model structure to be searched determines the probability gradient value of the model structure to be searched in the current probability iteration process based on the fractional gradient descent function of the policy selection probability.
[0113] S402 adjusts the strategy selection probability of the corresponding feature processing strategy based on the probability gradient value.
[0114] In some embodiments, the difference between the strategy selection probability of the corresponding feature processing strategy and the probability gradient value during the current probability iteration is used as the strategy selection probability of the corresponding feature processing strategy in the adjusted search model structure.
[0115] In the above embodiments, fractional calculus is introduced. Based on the instantaneous gradient provided by the current validation gradient, historical update trajectories are comprehensively considered, thereby more comprehensively evaluating the long-term performance potential of each feature processing strategy. This effectively alleviates the drastic fluctuations in strategy probability caused by large gradient estimation variance in traditional methods, significantly improving the stability of the optimization process. Furthermore, it enhances the algorithm's ability to balance exploration and feature processing strategy usage, preventing the search from prematurely converging to suboptimal structural patterns. Ultimately, it can more reliably and efficiently guide the entire model structure towards better validation performance, discovering image processing models with stronger generalization capabilities.
[0116] Figure 5 This is a flowchart illustrating the probability gradient value determination step in one embodiment. This embodiment refines the step S401 in the previous embodiment, which selects a fractional gradient descent function based on the strategy and determines the probability gradient values corresponding to different feature processing strategies in the search model structure based on the validation samples from the current probability iteration process. This embodiment provides an optional method for determining the probability gradient value, including the following steps:
[0117] For each feature processing strategy in the model structure to be searched, S501 determines the probability gradient related parameters based on the current strategy selection probability and the historical strategy selection probability of the corresponding feature processing strategy.
[0118] Wherein, the current strategy selection probability is the strategy selection probability of the feature processing strategy in the current probability iteration process; the historical strategy selection probability is the strategy selection probability of the feature processing strategy in the historical probability iteration process.
[0119] In some embodiments, for each feature processing strategy in the model structure to be searched, the difference between the current strategy selection probability of the feature processing strategy and the historical strategy selection probability of the corresponding feature processing strategy is determined; the absolute value corresponding to the sum of the difference and the preset compensation parameter is determined; and the first derivative of the historical strategy selection probability is determined; and the probability gradient related parameters are determined based on the product between the absolute value and the first derivative.
[0120] S502 determines the probability gradient values corresponding to different feature processing strategies in the model structure to be searched based on the preset fractional order, the probability gradient related parameters, and the gamma function.
[0121] In some embodiments, based on a preset fractional order, a target gradient-related parameter is determined according to the probability-related parameter; and based on the target gradient-related parameter and the gamma function, the probability gradient values corresponding to different feature processing strategies in the model structure to be searched are determined.
[0122] For example, in this embodiment, the policy gradient value of the model structure to be searched can be represented by the following formula (4):
[0123] (4)
[0124] in, Characterizing the probability gradient value, Characterizing parameters related to probability gradients, Characterizing the gamma function, r k r represents the probability of selecting the current strategy during the current (k-th) probability iteration. k-1 The probability of selecting the historical strategy during the (k-1)th iteration of the model represents the historical process. n represents the order (e.g., n=1), M represents the total number of iterations, a represents the lower bound terminal of the integral (which can be a preset value used to influence the convergence result of the model structure being searched), and ε represents the preset compensation parameter. The compensation parameter ε can be a small non-negative number used to adjust the convergence of the model structure in the r-th iteration. k When =a, it avoids the situation where the structure of the model to be searched does not converge. Characterizing fractional order, for example .
[0125] In this embodiment, the probability gradient-related parameters are determined by combining the current strategy selection probability and the historical strategy selection probability of the feature processing strategy. Furthermore, the probability gradient values corresponding to different feature processing strategies are calculated using a preset fractional order, these parameters, and the gamma function, thereby more accurately reflecting the advantages and disadvantages and changing trends of each feature processing strategy and improving the overall performance and search efficiency of the model.
[0126] Figure 6This is a flowchart illustrating the strategy selection probability adjustment step in one embodiment. This embodiment refines the step S402 of the above embodiment, which adjusts the strategy selection probability of the corresponding feature processing strategy based on the probability gradient value. This embodiment provides an optional method for adjusting the strategy selection probability, including the following steps:
[0127] S601 determines the target learning rate of each feature processing strategy in the model structure to be searched, based on the learning rate of the model structure to be searched and the strategy selection probability of the corresponding feature processing strategy.
[0128] In some embodiments, for each feature processing strategy in the model structure to be searched, the strategy selection probability is used as the exponent of the learning rate; a target value is determined with the learning rate as the base and the strategy selection probability as the exponent; and this target value is used as the target learning rate of the corresponding feature processing strategy.
[0129] S602 adjusts the strategy selection probability of the corresponding feature processing strategy based on the target learning rate and probability gradient value.
[0130] In some embodiments, the product of the learning rate and the probability gradient value of the model structure to be searched is determined as the target probability gradient value; the difference between the strategy selection probability in the current probability iteration process and the target probability gradient value is used as the strategy selection probability of the adjusted corresponding feature processing strategy.
[0131] For example, based on the strategy selection probability update formula shown in the following formula (5), the strategy selection probability of the corresponding feature processing strategy is adjusted according to the target learning rate and the probability gradient value.
[0132] (5)
[0133] in, The gradient value of the representation policy, r k r represents the probability of strategy selection in this (k-th) probability iteration process. k+1 This represents the probability of policy selection in the next (k+1)th probability iteration. It is the learning rate.
[0134] In this embodiment, the global learning rate is scaled according to the policy probability, and its update step size is relatively conservative to consolidate and refine the current selection and prevent excessive oscillation. For low-probability policies that have not been fully explored, a relatively larger effective learning rate is assigned to encourage exploration and prevent the search process from getting trapped in local optima prematurely. This adaptive learning rate adjustment effectively balances the contradiction between exploration and policy, accelerates search convergence, improves the stability and efficiency of the search process, and ultimately guides the model to more reliably discover high-performance structures.
[0135] Figure 7 This is a flowchart illustrating the image processing steps in one embodiment. This embodiment refines the above embodiment and includes the following steps:
[0136] S701 acquires the image data to be processed.
[0137] In some embodiments, image data to be processed is acquired. It should be noted that, after acquiring the image data, this embodiment may further preprocess the image data to ensure accurate image processing operations. For example, image processing is described using image recognition as an example, where image data to be recognized is acquired.
[0138] S702 is based on an image processing model and determines the image processing result of the image data according to the image data.
[0139] In some embodiments, based on an image processing model, image processing operations matching the image processing model are performed on the image data to obtain the image processing result. For example, image processing is illustrated using image recognition as an example, where a target object is identified from the image data based on an image recognition model.
[0140] In this embodiment, the image processing model eliminates the tedious manual model design and tuning process of traditional methods. It directly utilizes the powerful feature representation and transformation capabilities of the optimal structure obtained through search to automatically adapt to various complex image processing needs, ensuring high accuracy of the processing results. At the same time, since the final model is a simplified and high-performance entity distilled from a large number of candidate strategies, its computational efficiency is usually higher, which is conducive to achieving low-latency, high-performance deployment and application in real-world scenarios.
[0141] In one embodiment, this embodiment provides an optional method for model training, using the application of this method to a server as an example for illustration. For example... Figure 8 As shown, the method includes the following steps:
[0142] S801 acquires training samples, validation samples, and the structure of the model to be searched;
[0143] In each training iteration, S802 determines the model gradient value of the model structure to be searched based on the fractional gradient descent function of the model parameters and the training samples under this model iteration.
[0144] The model structure to be searched includes at least one node; there is at least one node pointing relationship between different nodes; in the node pointing relationship from the first node to the second node, the first node corresponds to the first feature data, the node pointing relationship corresponds to the feature processing strategy, and the second node corresponds to the second feature data obtained after processing the first feature data using the feature processing strategy.
[0145] S803 adjusts the model parameters of the model structure to be searched based on the model gradient value until the model parameters in the model structure to be searched tend to stabilize during the current training iteration.
[0146] S804 determines the probability gradient related parameters for each feature processing strategy in the model structure to be searched, based on the current strategy selection probability and the historical strategy selection probability of the corresponding feature processing strategy.
[0147] Wherein, the current strategy selection probability is the strategy selection probability of the feature processing strategy in the current probability iteration process; the historical strategy selection probability is the strategy selection probability of the feature processing strategy in the historical probability iteration process.
[0148] S805 determines the probability gradient values corresponding to different feature processing strategies in the model structure to be searched based on preset fractional order, probability gradient related parameters, and gamma function.
[0149] S806 determines the target learning rate of the corresponding feature processing strategy based on the learning rate of the model structure to be searched and the strategy selection probability of the corresponding feature processing strategy.
[0150] S807 adjusts the strategy selection probability of the corresponding feature processing strategy according to the target learning rate and probability gradient value until the strategy selection probability in the search model structure under the current training iteration tends to stabilize.
[0151] When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, S808 determines the target feature processing strategy based on the strategy selection probabilities in the model structure to be searched, and constructs an image processing model based on the target feature processing strategy.
[0152] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0153] Based on the same inventive concept, this application also provides a model structure search apparatus for implementing the model structure search method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more embodiments of the model structure search apparatus provided below can be found in the limitations of the model structure search method described above, and will not be repeated here.
[0154] In one exemplary embodiment, such as Figure 9 As shown, a model structure search device is provided, comprising: an acquisition module 10, a training module 11, a verification module 12, and a construction module 13, wherein:
[0155] Module 10 is used to acquire training samples, validation samples, and the structure of the model to be searched.
[0156] Training module 11 is used to adjust the model parameters in the search model structure according to the training samples for each training iteration until the model parameters in the search model structure under the current training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters;
[0157] The verification module 12 is used to adjust the strategy selection probability of each feature processing strategy in the structure to be searched based on the verification samples, until the strategy selection probability in the structure to be searched in the current training iteration tends to be stable; the strategy selection probability is adjusted based on the fractional gradient descent function of the strategy selection probability.
[0158] Module 13 is used to determine the target feature processing strategy based on the strategy selection probability in the model structure to be searched, when the model parameters and strategy selection probabilities of the model structure to be searched tend to be stable under different iterative training processes, and to construct an image processing model based on the target feature processing strategy.
[0159] In some embodiments, the verification module is further configured to, for each probability iteration, determine the probability gradient value corresponding to different feature processing strategies in the model structure to be searched based on the fractional gradient descent function of the strategy selection probability and the verification samples under the current probability iteration, and adjust the strategy selection probability of the corresponding feature processing strategy according to the probability gradient value.
[0160] In some embodiments, the verification module is further configured to, for each feature processing strategy in the model structure to be searched, determine the probability gradient related parameters based on the current strategy selection probability of the feature processing strategy and the historical strategy selection probability of the corresponding feature processing strategy; and determine the probability gradient values corresponding to different feature processing strategies in the model structure to be searched based on the preset fractional order, probability gradient related parameters and gamma function.
[0161] In some embodiments, the verification module is further configured to, for each feature processing strategy in the model structure to be searched, determine the target learning rate of the corresponding feature processing strategy based on the learning rate of the model structure to be searched and the strategy selection probability of the corresponding feature processing strategy; and adjust the strategy selection probability of the corresponding feature processing strategy based on the target learning rate and the probability gradient value.
[0162] In some embodiments, the model structure to be searched in the model structure search device includes at least one node; there is at least one node pointing relationship between different nodes; in the node pointing relationship from the first node to the second node, the first node corresponds to the first feature data, the node pointing relationship corresponds to the feature processing strategy, and the second node corresponds to the second feature data obtained after processing the first feature data using the feature processing strategy.
[0163] In some embodiments, the model structure search apparatus further includes: a usage module for acquiring image data to be processed; and determining the image processing result of the image data based on the image processing model.
[0164] Each module in the aforementioned model structure search device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.
[0165] In one exemplary embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 10As shown, this computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a model structure search method.
[0166] Those skilled in the art will understand that Figure 10 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0167] In one embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.
[0168] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.
[0169] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0170] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.
[0171] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0172] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0173] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A model structure search method, characterized in that, The method includes: Obtain training samples, validation samples, and the structure of the model to be searched; For each training iteration, the model parameters in the model structure to be searched are adjusted based on the training samples until the model parameters in the model structure to be searched under this training iteration tend to be stable; the model parameters are adjusted based on the fractional gradient descent function of the model parameters. Based on the validation samples, the strategy selection probability of each feature processing strategy in the search model structure is adjusted until the strategy selection probability in the search model structure under this training iteration tends to be stable; the strategy selection probability is adjusted based on the fractional gradient descent function of the strategy selection probability. When the model parameters and strategy selection probabilities of the model structure to be searched tend to stabilize under different iterative training processes, the target feature processing strategy is determined according to the strategy selection probabilities in the model structure to be searched, and an image processing model is constructed according to the target feature processing strategy.
2. The method according to claim 1, characterized in that, The step of adjusting the strategy selection probability of each feature processing strategy in the search model structure based on the verification samples includes: For each probability iteration, a fractional gradient descent function is selected based on the strategy. Based on the validation samples in this probability iteration, the probability gradient values corresponding to different feature processing strategies in the model structure to be searched are determined. Based on the probability gradient value, the strategy selection probability of the corresponding feature processing strategy is adjusted.
3. The method according to claim 2, characterized in that, The fractional gradient descent function based on strategy selection probability determines the probability gradient values corresponding to different feature processing strategies in the model structure to be searched, based on the verification samples under this probability iteration process, including: For each feature processing strategy in the model structure to be searched, the probability gradient related parameters are determined based on the current strategy selection probability and the historical strategy selection probability of the corresponding feature processing strategy. Based on the preset fractional order, the probability gradient related parameters, and the gamma function, the probability gradient values corresponding to different feature processing strategies in the model structure to be searched are determined.
4. The method according to claim 2, characterized in that, The step of adjusting the strategy selection probability of the corresponding feature processing strategy based on the probability gradient value includes: For each feature processing strategy in the model structure to be searched, the target learning rate of the corresponding feature processing strategy is determined based on the learning rate of the model structure to be searched and the strategy selection probability of the corresponding feature processing strategy. The strategy selection probability of the corresponding feature processing strategy is adjusted based on the target learning rate and the probability gradient value.
5. The method according to any one of claims 1-4, characterized in that, The model structure to be searched includes at least one node; there is at least one node pointing relationship between different nodes; in the node pointing relationship from the first node to the second node, the first node corresponds to the first feature data, the node pointing relationship corresponds to the feature processing strategy, and the second node corresponds to the second feature data obtained after processing the first feature data using the feature processing strategy.
6. The method according to any one of claims 1-4, characterized in that, The method further includes: Acquire the image data to be processed; Based on the image processing model, the image processing result of the image data is determined according to the image data.
7. A model structure search device, characterized in that, The device includes: The acquisition module is used to acquire training samples, validation samples, and the structure of the model to be searched. The training module is used to adjust the model parameters in the search model structure according to the training samples for each training iteration until the model parameters in the search model structure tend to stabilize in the current training iteration; the model parameters are adjusted based on the fractional gradient descent function of the model parameters; The verification module is used to adjust the strategy selection probability of each feature processing strategy in the search model structure according to the verification samples, until the strategy selection probability in the search model structure tends to stabilize in the current training iteration; the strategy selection probability is adjusted based on the fractional gradient descent function of the strategy selection probability. The construction module is used to determine the target feature processing strategy based on the strategy selection probability in the search model structure when the model parameters and strategy selection probabilities of the search model structure tend to be stable under different iterative training processes, and to construct an image processing model based on the target feature processing strategy.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.