A method for constructing a neural network search architecture with a high-precision two-layer optimization method

Abstract

The present invention discloses a method for constructing a neural network search architecture of a high-precision double-layer optimization method. In the search space, the tendency of the operation receptive field is expanded, and only the receptive field of the tendency is selected; the ordinary search unit and the downsampling search unit are constructed and stacked to form a supernet; the architecture weight of each edge and each operation in the directed acyclic graph of the supernet is searched by the high-precision double-layer optimization method; the operation with the largest final weight in each edge is obtained by multiplying the architecture weight and the corresponding operation architecture weight as the final weight, and the corresponding final weight is used as the final weight of the edge, and the search unit is updated with the results of the edge and the operation to obtain the final model architecture. In the search space, the tendency of the operation receptive field is expanded, and only the receptive field of the tendency is selected, which reduces the exploration range and improves the search efficiency. The result of the high-precision double-layer optimization method is highly approximate to the optimal architecture weight, which avoids the enrichment of a large number of pooling operations in the optimization result.

Description

Technical Field

[0001] The present invention relates to the field of neural network architecture search, and in particular to a method for constructing a neural network search architecture using a high-precision double-layer optimization method. Background Art

[0002] Automated machine learning (AutoML) is a technology that aims to simplify machine learning tasks. Traditionally, the development of machine learning models requires expertise and experience, including processes such as data preprocessing, feature engineering, model selection, and tuning. AutoML aims to automate these processes, allowing more people to take advantage of machine learning technology without having to deeply understand its inner workings. The background of automated machine learning technology can be traced back to the initial attempts to automatically tune machine learning models, and modern AutoML goes a step further and attempts to automate the entire machine learning workflow. Traditional machine learning algorithms contain many hyperparameters that need to be manually adjusted to achieve optimal performance. Hyperparameter optimization algorithms can automatically search the parameter space and find the optimal parameter configuration. In deep learning, the design of the network structure is crucial to performance. Neural architecture search technology improves the performance of the model by automating the search for the structure and hyperparameters of the neural network. Feature engineering is a key step in machine learning, but it usually requires a lot of expertise and experience. Automated feature engineering technology attempts to automatically select, transform, and combine features to improve the performance of the model. AutoML can automatically select the model that is most suitable for a specific task, from simple linear models to complex deep learning models. In addition to the model development stage, AutoML can also automate the deployment and management of the model, making it easy to integrate into practical applications. The development of automated machine learning technology aims to lower the threshold for use, accelerate the model development process, and enable more people to benefit from the application of machine learning technology.

[0003] Neural Architecture Search (NAS) is a technology that uses machine learning methods to automatically design neural network structures. Traditionally, neural network structures are designed by human experts, usually based on experience and intuition, and require a lot of trials and iterations to achieve good performance. The goal of NAS is to automate this process through algorithms, making the design of neural networks more efficient and intelligent. With the growth of computing resources, especially the widespread use of dedicated hardware such as GPUs and TPUs, neural networks can be trained and evaluated more quickly, which provides more experimental conditions for NAS. Deep learning has achieved great success in fields such as computer vision and natural language processing, which has also promoted the demand for more complex and efficient neural network structures, and also spawned the demand for automatic design of neural networks. With the continuous evolution of neural network structures, the structural space of the network has become larger and more complex. The emergence of NAS technology can help find the optimal structure in this huge search space. The development of optimization algorithms such as reinforcement learning has given NAS technology more methods to search for the optimal network structure. As an emerging field, automated machine learning covers NAS, making the design and optimization process of machine learning models more intelligent and automated. The background of NAS technology is the result of the growth of computing resources, the development of deep learning, the expansion of search space, the progress of reinforcement learning, and the rise of automated machine learning. These factors have promoted the development and application of NAS technology, making the design of neural network structure more intelligent and efficient.

[0004] However, when researchers were conducting research on neural network architecture search, they found that the search efficiency of the constructed neural network search architecture was low, and the search results were not the optimal architecture or even extreme architecture. Summary of the invention

[0005] The present invention solves the technical problems in the prior art of low search efficiency and non-optimal search results by providing a neural network search architecture construction method with a high-precision double-layer optimization method.

[0006] The present invention provides a method for constructing a neural network search architecture of a high-precision double-layer optimization method, comprising:

[0007] Step 1: Define the operation space, expand the operation space by the tendency of the operation receptive field in the initial search space and select only the tendency receptive field;

[0008] Step 2: construct a normal search unit and a downsampling search unit in the operation space, stack the normal search unit and the downsampling search unit to form a supernet; the normal search unit and the downsampling search unit are both directed acyclic graphs of the same structure; the directed acyclic graph includes a plurality of nodes, each node represents a feature graph, the edges between the nodes are mixed operations composed of all operations in the operation space, the direction of the edge represents the direction of the information flow, and each edge and each operation in the edge has a corresponding architecture parameter;

[0009] Step 3: searching and obtaining the architecture weight of each edge and each operation in the directed acyclic graph of the supernet by a high-precision two-layer optimization method;

[0010] Step 4: The product of the architecture weight and the corresponding operation architecture weight is used as the final weight to obtain the operation with the largest final weight in each edge, and the corresponding final weight is used as the final weight of the edge. For units with input edges greater than two, only the first two edges with the largest final weight are retained. The results of the edge and operation are stacked and updated to update the ordinary search unit and the downsampling search unit until the required number of layers to obtain the final model architecture.

[0011] Specifically, constructing a normal search unit and a downsampling search unit in the operation space, stacking the normal search unit and the downsampling search unit to form a supernet, includes:

[0012] The common search unit and the down-sampling search unit are constructed in the operation space, and the operations on the edges and edges of the common search unit and the down-sampling search unit all use a partially decoupled operation weighting function as shown in the following formula:

[0013]

[0014] Among them, α is the architecture parameter, Sigmoid (α) is the architecture weight, the number of channels of the input edge of the search unit node is controllable, and the default is one-fourth of the input channels and the selected channels are randomly sampled, and then the channel sum after the edge output is spliced ​​with three-quarters of the input, and the ordinary search unit and the downsampling search unit are stacked to form the super network.

[0015] Specifically, the output feature map of the edge of the search unit is the result of weighted summation of the input feature map after sampling the input operation set and concatenating it with the unsampled channel, as shown in the following formula:

[0016]

[0017] where f(·) is the output of the partial channel connection of edge (i,j), and x i is the output of node i, S i,j is the channel sampling mask.

[0018] Specifically, the node feature graph is the weighted sum of the input edge results, as shown in the following formula:

[0019]

[0020] in, is a schema parameter.

[0021] Specifically, the high-precision two-layer optimization method is used to search and obtain the architecture weight of each edge and each operation in the directed acyclic graph of the supernet, including:

[0022] Initialize the model parameters to w * (γ0) and the architecture parameters are γ0, the training set images are sent to the supernet, the architecture parameters are fixed, the cross entropy loss is calculated according to the label after forward propagation, and the gradient is calculated by back propagation, and the model parameters are updated according to the gradient direction; the updated model parameters are fixed, and the cross entropy loss is calculated according to the label after forward propagation, and the gradient is calculated by back propagation, and the approximate gradients with respect to the architecture parameters are calculated according to the high-precision two-layer optimization method, and the architecture parameters of the supernet are updated until the architecture parameters converge, so as to obtain the final architecture weights of each edge and each operation in the directed acyclic graph of the supernet.

[0023] Specifically, sending the training set images into the supernet includes:

[0024] The training set images are randomly cropped, flipped, and normalized before being sent to the supernet.

[0025] Specifically, the gradient of the architecture parameters of the high-precision two-layer optimization method is an approximate numerical method, as shown in the following formula:

[0026]

[0027] in, is the symbol for finding the gradient, L val (w * ,γ) is the validation loss, L train (w,γ) is the training loss, γ is all the architecture parameters, w * is the optimal network weight, w is the model parameter, ξ, ε are the hyperparameters for adjusting the approximate formula,

[0028] One or more technical solutions provided in the present invention have at least the following technical effects or advantages:

[0029] 1. Define the operation space first, expand the operation space by the tendency of the operation receptive field in the initial search space and only select the receptive field of the tendency; then construct the ordinary search unit and the downsampling search unit in the operation space, stack the ordinary search unit and the downsampling search unit to form a supernet; then search and obtain the architecture weight of each edge and each operation in the directed acyclic graph of the supernet through the high-precision two-layer optimization method; finally, the operation with the largest final weight in each edge is obtained by multiplying the architecture weight and the corresponding operation architecture weight as the final weight, and the corresponding final weight is used as the final weight of the edge. The ordinary search unit and the downsampling search unit are stacked and updated with the results of the edge and the operation until the required number of layers to obtain the final model architecture. By expanding the operation space by the tendency of the operation receptive field in the initial search space and only selecting the receptive field of the tendency, the exploration scope of the search space is effectively reduced and the search efficiency is improved. It is reflected in that after expanding the three-dimensional inverted residual bottleneck operation, the search is completed in only six hours, and better results are achieved than the original search space under the premise of selecting the new operation. The results of the high-precision two-layer optimization method are highly approximate to the optimal architecture weight, avoiding the enrichment of a large number of pooling operations in the optimization results.

[0030] 2. In the present invention, the random sampling of some channels is sent to the edge, and the remaining channels do not participate in the calculation, so they do not occupy the video memory.

[0031] 3. The operation weighting function in the present invention uses sigmoid to ensure fair competition among operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A flowchart of a method for constructing a neural network search architecture for a high-precision dual-layer optimization method provided by an embodiment of the present invention;

[0033] Figure 2 A schematic diagram of a three-dimensional neural network search architecture for modeling a high-precision two-layer optimization method according to an embodiment of the present invention;

[0034] Figure 3 The search results of the present invention in the LUNA16 three-dimensional data set;

[0035] Figure 4 A heat map showing the effect of selecting operations for the high-precision dual-layer optimization method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The embodiment of the present invention solves the technical problems in the prior art of low search efficiency and non-optimal search results by providing a method for constructing a neural network search architecture with a high-precision double-layer optimization method.

[0037] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0038] See also Figure 1 The method for constructing a neural network search architecture of a high-precision two-layer optimization method provided by an embodiment of the present invention includes:

[0039] Step 1: Define the operation space. Expand the operation space according to the tendency of the operation receptive field in the initial search space and select only the receptive field with the tendency, thereby reducing the search range. The tendency of the receptive field is obtained by step 4. At the beginning, use the default operation space or a custom operation space; expand the search space according to the tendency of the receptive field selected in the training set in the original search space.

[0040] Step 2: Construct a normal search unit and a downsampling search unit in the operation space, stack the normal search unit and the downsampling search unit to form a supernet; the normal search unit and the downsampling search unit are both directed acyclic graphs with the same structure; the directed acyclic graph contains multiple nodes, each node represents a feature graph, and the edges between the nodes are mixed operations composed of all operations in the operation space. The direction of the edge represents the direction of the information flow, and each edge and each operation in the edge has corresponding architecture parameters;

[0041] This step is specifically described. A normal search unit and a downsampling search unit are constructed in the operation space, and the normal search unit and the downsampling search unit are stacked to form a supernet, including:

[0042] In the operation space, a normal search unit and a downsampling search unit are constructed. The operations on the edges and edges of the normal search unit and the downsampling search unit all use a partially decoupled operation weighting function as shown in the following formula:

[0043]

[0044] Among them, α is the architecture parameter, Sigmoid (α) is the architecture weight, the number of channels of the input edge of the search unit node is controllable, and the default is one-fourth of the input channels and the selected channels are randomly sampled. The sum of the channels after the edge output is concatenated to three-quarters of the input, and the ordinary search unit and the downsampling search unit are stacked to form a supernet.

[0045] Among them, the output feature map of the edge of the search unit is the result of the weighted sum of the input feature map sampling input operation set and the concatenation of the unsampled channels, as shown in the following formula:

[0046]

[0047] where f(·) is the output of the partial channel connection of edge (i,j), and x i is the output of node i, S i,j is the channel sampling mask.

[0048] That is to say, the feature map of the side output is the concatenation of the output of the mixing operation and the unsampled channels of the side input.

[0049] The node feature graph is the weighted sum of the input edge results, as shown in the following formula:

[0050]

[0051] in, is a schema parameter.

[0052] Step 3: Search and obtain the architectural weight of each edge and each operation in the directed acyclic graph of the supernet through a high-precision two-layer optimization method;

[0053] This step is specifically described. The architecture weights of each edge and each operation in the directed acyclic graph of the supernet are obtained by searching through a high-precision two-layer optimization method, including:

[0054] The training set images are sent to the supernet, the architecture parameters are fixed, the cross entropy loss is calculated according to the label after forward propagation, and the gradient is calculated by back propagation; the weight parameters of the supernet are optimized according to the gradient direction, and the weight parameters are repeated until convergence to obtain the final architecture weight of each edge and each operation in the directed acyclic graph of the supernet, that is, the architecture parameters and weight parameters are alternately optimized until convergence to obtain the final architecture weight of each edge and each operation in the directed acyclic graph of the supernet.

[0055] Specifically, the training set images are fed into the supernet, including:

[0056] The training set images are randomly cropped, flipped, and normalized before being sent to the supernet.

[0057] Step 4: The final weight is obtained by multiplying the architecture weight and the corresponding operation architecture weight as the final weight, and the operation with the largest final weight in each edge is obtained. The corresponding final weight is used as the final weight of the edge. For units with input edges greater than two, only the first two edges with the largest final weight are retained. The results of the edge and operation are stacked and updated to the required number of layers for the normal search unit and the downsampling search unit to obtain the final model architecture. It should be noted that if the operation space needs to be expanded, new search results can be obtained by starting from step 1 again according to the receptive field tendency of the update unit, and the updated normal search unit and downsampling search unit are stacked to the required number of layers to obtain the final model architecture.

[0058] Specifically, the gradient of the architecture parameters of the high-precision two-layer optimization method is an approximate numerical method, as shown in the following formula:

[0059]

[0060] in, is the symbol for finding the gradient, Lval (w * ,γ) is the validation loss, L train (w,γ) is the training loss, γ is all the architecture parameters, w * is the optimal network weight, w is the model parameter, ξ, ε are the hyperparameters for adjusting the approximate formula,

[0061] Using w * (γ0) initialization, Fit the height w * (γ), thereby increasing the accuracy of subsequent numerical calculations. train and L val It is the cross entropy loss function, but regularization is added and the regularization maintains the physical properties of the loss, as shown in the following formula:

[0062]

[0063] L total =L val (w * (γ),γ)+w 0-1 L 0-1

[0064] Among them, L 0-1 is the regularization function, γ is all the architectural parameters including edges and operations, and w 0-1 is a hyperparameter for adjusting regularization. total YesL val Add the regularized loss function and calculate the L total To replace the original L val .

[0065] like Figure 2 As shown in the figure, (a) is the search framework and the final model framework, (b) is a schematic diagram of the search unit construction method, and (c) is a schematic diagram of the output of the edge in the search unit. If the unit in (a) is a search unit, then (a) is the search framework, and if the unit in (a) is the final unit, then (a) is the final model framework. (c) embodies the overall process of channel sampling, splicing, operation and edge weighting.

[0066] The search results of the present invention in the LUNA16 three-dimensional data set are as follows: Figure 3 As shown, the upper picture is a normal unit searched out from the LUNA16 3D data set, and the lower picture is a downsampled unit.

[0067] like Figure 4As shown in Figure 1, no pooling enrichment phenomenon occurred when the model architecture converged. The final results and the comparison search results in the LUNA16 three-dimensional dataset are shown in Table 1. The best result and the second best result in each column are shown in bold and underline format respectively. “-” means not mentioned.

[0068] Table 1 Comparison of the search results of the model (NAS-HBO) searched by the present invention and several existing methods

[0069]

[0070]

[0071] As shown in Table 1, the present invention obviously achieves the highest classification accuracy with the least model parameters. The high accuracy and F1 score indicate that the NAS-HBO of the present invention achieves a good balance between sensitivity and specificity.

[0072] The process of constructing a benign and malignant lung nodule classification model and performing lung cancer diagnosis by the present invention is as follows:

[0073] Step 1: Extract lung nodules from lung window images;

[0074] This step is specifically described to extract lung nodules from the lung window image, including:

[0075] The center of the lung window image is set to -300, the width of the lung window image is adjusted to [-1200; 600], the contrast of the lung window image is linearly transformed to [0; 1], the lung image is obtained by segmentation annotation, and the lung nodules are extracted using the lung nodule annotation file, whose size is 36×36×36.

[0076] Step 2: Input the lung nodules into the built supernet model training model to obtain the model weights and architecture parameters;

[0077] This step is specifically described. The lung nodules are input into the built supernet model training model to obtain the model weights and architecture parameters, including:

[0078] The network search space, i.e., the edges in each unit in the supernet, is set. In this embodiment of the present invention, 3×3×3 and 5×5×5 separable convolutions and dilated separable convolutions are selected; 3×3×3 maximum pooling and average pooling; identity mapping; a three-dimensional inverted residual bottleneck block with a receptive field of 5×5×5 and a dimensionality increase factor of 3 or 6, and the receptive field is determined by the tendency of the receptive field in the search history of the data set.

[0079] Build the normal search unit and downsampling search unit in the supernet, the topology is as follows Figure 2 As shown, the embodiments of the present invention use sigmoid weighting in both edge and in-edge operations.

[0080] Eight units are stacked as a supernet, of which one-third and two-thirds are downsampling search units and the rest are normal search units.

[0081] Initialize the model parameters and architecture parameters, fix the architecture parameters, train the model parameters until convergence, and select the number of rounds with the highest accuracy in the validation set as a hyperparameter for the two-layer optimization.

[0082] Specifically, a high-precision two-layer optimization method is used to train the supernet, initialize the model parameters and architecture parameters, fix the architecture parameters, and initialize the model parameters of the two-layer optimization: the number of rounds of training model parameters is the number of rounds in which the highest accuracy first appears in the validation set determined in the previous step, and then the two-layer optimization is started to alternately train the model parameters and architecture parameters.

[0083] Step 3: Select the model architecture through the model weights and architecture parameters to obtain the final normal unit and downsampling unit;

[0084] This step is explained in detail. The model architecture is selected through the model weights and architecture parameters to obtain the final normal units and downsampling units, including:

[0085] The architectural weight of the edge is multiplied by the architectural weight of the operation to obtain a metric parameter. The operation with the largest metric parameter for each edge is selected as the representative of the edge. The weights of the representatives of the node input edges are then compared. For each node, only the input edges with the top two metric parameters are retained. Finally, the normal unit and downsampling unit are obtained after the search decision.

[0086] Step 4: Through the common units and downsampling units, stack them to the specified layer to input the lung nodules, train the model weights, and obtain the final lung nodule benign and malignant classification model.

[0087] This step is specifically described. Through common units and downsampling units, they are stacked to the specified layer to input lung nodules, train the model weights, and obtain the final lung nodule benign and malignant classification model, including:

[0088] The common units and down-sampling units obtained in the previous step are stacked into 20 units, of which one-third and two-thirds are down-sampling units, and the rest are common units. The model parameters are trained to obtain the final classification model of benign and malignant pulmonary nodules.

[0089] Step 5: Send the lung nodules obtained in the medical auxiliary diagnosis process to the benign and malignant lung nodule classification model obtained in step 4 to obtain the discrimination results and confidence levels as a criterion for assisting doctors in diagnosing lung cancer.

[0090] It should be noted here that other classification models can also be constructed through the present invention, and corresponding classification work can be performed. The specific process is similar and will not be described in detail here.

[0091] In summary, the embodiment of the present invention proposes to randomly sample some channels and send them into the edge. The remaining channels do not participate in the calculation, so they do not occupy the video memory, and the operation weighting function uses sigmoid to make the operation compete fairly. In addition, in the two-layer optimization, the model parameters are pre-trained to convergence, thereby improving the search accuracy of the two-layer optimization, which is manifested in completely eliminating the problem of enrichment pooling operations in the search and refreshing the historical record of model accuracy.

[0092] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0093] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0094] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0095] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0096] The parts not described in detail in the embodiments of the present invention are all known technologies to those skilled in the art. Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention is described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions of the present invention, which should be included in the scope of the claims of the present invention.

Claims

1. A method for constructing a neural network search architecture for a high-precision two-layer optimization method, characterized in that: include: Step 1: Define the operation space, expand the operation space by the tendency of the operation receptive field in the initial search space and select only the tendency receptive field; Step 2: construct a normal search unit and a downsampling search unit in the operation space, stack the normal search unit and the downsampling search unit to form a supernet; the normal search unit and the downsampling search unit are both directed acyclic graphs of the same structure; the directed acyclic graph includes a plurality of nodes, each node represents a feature graph, the edges between the nodes are mixed operations composed of all operations in the operation space, the direction of the edge represents the direction of the information flow, and each edge and each operation in the edge has a corresponding architecture parameter; Step 3: searching and obtaining the architecture weight of each edge and each operation in the directed acyclic graph of the supernet by a high-precision two-layer optimization method; The method of searching and obtaining the architecture weight of each edge and each operation in the directed acyclic graph of the supernet by a high-precision two-layer optimization method includes: Initialize the model parameters and architecture parameters, send the training set images into the supernet, fix the architecture parameters, calculate the cross entropy loss according to the label after forward propagation, perform back propagation to calculate the gradient, and update the model parameters according to the gradient direction; fix the updated model parameters, calculate the cross entropy loss according to the label after forward propagation, perform back propagation to calculate the gradient, calculate the approximate gradient of the architecture parameters according to the high-precision two-layer optimization method, and update the architecture parameters of the supernet until the architecture parameters converge, and obtain the final architecture weight of each edge and each operation in the directed acyclic graph of the supernet; The gradient of the architecture parameters of the high-precision two-layer optimization method is an approximate numerical method, as shown in the following formula: in, is the symbol for finding the gradient, is the validation loss, is the training loss, γ is all the architecture parameters, and w * is the optimal network weight, w is the model parameter, ξ, ε are the hyperparameters for adjusting the approximate formula, Step 4: The product of the architecture weight and the corresponding operation architecture weight is used as the final weight to obtain the operation with the largest final weight in each edge, and the corresponding final weight is used as the final weight of the edge. For units with input edges greater than two, only the first two edges with the largest final weight are retained. The results of the edge and operation are stacked and updated to update the ordinary search unit and the downsampling search unit until the required number of layers to obtain the final model architecture.

2. The method for constructing a neural network search architecture for a high-precision two-layer optimization method according to claim 1, characterized in that: The step of constructing a normal search unit and a down-sampling search unit in the operation space, and stacking the normal search unit and the down-sampling search unit to form a supernet includes: The common search unit and the down-sampling search unit are constructed in the operation space, and the operations on the edges and edges of the common search unit and the down-sampling search unit all use a partially decoupled operation weighting function as shown in the following formula: Among them, α is the architecture parameter, Sigmoid (α) is the architecture weight, the number of channels of the input edge of the search unit node is controllable, and the default is one-fourth of the input channels and the selected channels are randomly sampled, and then the channel sum after the edge output is spliced ​​with three-quarters of the input, and the ordinary search unit and the downsampling search unit are stacked to form the super network.

3. The method for constructing a neural network search architecture for a high-precision two-layer optimization method according to claim 2, characterized in that: The output feature map of the edge of the search unit is the result of weighted summation of the input feature map after sampling the input operation set and concatenating the unsampled channels, as shown in the following formula: where f(·) is the output of the partial channel connection of edge (i,j), and x i is the output of node i, S i,j is the channel sampling mask.

4. The method for constructing a neural network search architecture for a high-precision two-layer optimization method according to claim 3, characterized in that: The node feature graph is the weighted sum of the input edge results, as shown in the following formula: in, is a schema parameter.

5. The method for constructing a neural network search architecture for a high-precision two-layer optimization method according to claim 1, characterized in that: The sending the training set images into the supernet includes: The training set images are randomly cropped, flipped, and normalized before being sent to the supernet.

Images