Convolutional Neural Network Architecture Search Method Based on Differentiable Sampler and Progressive Learning

Abstract

In view of the limitations of the prior art, the present invention proposes a convolutional neural network architecture search method based on a differentiable sampler and progressive learning. This method uses a differentiable sampler to directly sample and optimize the constructed super network, which can change the optimization goal of the architecture search from optimizing the super network to finding the optimal probability distribution function and minimizing the expectation of the loss function of the sub-network under the probability distribution; it can optimize the probability distribution function of the sub-network by evaluating the performance of the sub-network, thereby reducing the discretization error. At the same time, due to the adoption of a progressive learning strategy, a stable search can be performed in an exponentially increasing search space, thereby being able to obtain a more complex convolutional neural network architecture.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and deep learning technology, and specifically to a neural network architecture search technology, and more specifically to a convolutional neural network architecture search method based on a differentiable sampler and progressive learning. Background Art

[0002] In recent years, deep learning has developed rapidly. By utilizing the powerful feature representation capabilities of neural networks, deep learning has made breakthrough progress in tasks of various modalities such as image recognition. Among them, the architectural design of neural networks plays a vital role in their performance. However, with the continuous development of deep learning, the number of layers of neural networks continues to increase, and the neural operators become more and more complex, resulting in the structure of neural networks becoming more and more complex, and more and more hyperparameters need to be manually adjusted, which consumes a lot of manpower and material resources to design neural networks with excellent performance. Therefore, Neural Architecture Search (NAS), as a technology that uses algorithms to automatically search for neural network architectures to meet the needs of different tasks, has become a research hotspot in the field of automated machine learning.

[0003] Common NAS methods include reinforcement learning, evolutionary algorithms, and gradient optimization. NAS methods based on reinforcement learning and evolutionary algorithms are black-box optimization algorithms that can handle situations where the search space is discrete and not differentiable. However, the search time is long and the consumption of computing resources is large, so their efficiency is not as good as NAS methods based on gradient optimization. Among NAS methods based on gradient optimization, the mainstream is the Differentiable Architecture Search (DARTS) method. The DARTS method is based on the search space of repeated units, which reduces the search space of the entire network directly to the search space of two units, reducing the search space exponentially. The two units are normal units and reduction units. The normal unit can keep the number of channels and resolution of the feature map unchanged, while the reduction unit reduces the resolution of the feature map to half of the original and doubles the number of channels. The entire neural network is formed by stacking a certain number of normal units and reduction units. DARTS continuously stacks normal units and reduction units to build the entire network. All normal units share the same topology, and all reduction units also share the same topology, reducing the search space. However, the DARTS method has technical problems such as simple search space, large discretization error, large video memory consumption, and high video memory consumption.

[0004] The Chinese invention application with a publication date of October 13, 2020 and publication number CN111767983A: A discretized differentiable neural network search method based on an entropy loss function, which drives the sparsity and discretization of system elements (weights) by minimizing system entropy, and designs a new loss term based on the entropy function as a constrained loss term suitable for different target network structure configurations to reduce the discretization error. Although the scheme adds an entropy function, it still optimizes the loss of the hypernetwork on the validation set; and its architecture parameters represent the hard attention weights of the hypernetwork neural operators, and there is no sampling link in the optimization process; therefore, the effect of this scheme on reducing the discretization error of the hypernetwork is still relatively limited, and the existing technology still has certain limitations. Summary of the invention

[0005] In view of the limitations of the prior art, the present invention proposes a convolutional neural network architecture search method based on a differentiable sampler and progressive learning. The technical solution adopted by the present invention is:

[0006] A convolutional neural network architecture search method based on a differentiable sampler and progressive learning, comprising the following steps:

[0007] S1, constructing a hypernetwork, architectural parameters of the hypernetwork, and a differentiable sampler;

[0008] S2, in an incremental learning manner, using the differentiable sampler to perform sampling optimization on the super network according to the architecture parameters to obtain a desired network;

[0009] S3, retraining the expected network until the expected network converges.

[0010] Compared with the prior art, the present invention uses a differentiable sampler to directly sample and optimize the constructed super network, which can change the optimization goal of architecture search from optimizing the super network to finding the optimal probability distribution function and minimizing the expectation of the loss function of the sub-network under the probability distribution; it can evaluate the performance of the sub-network and optimize the probability distribution function of the sub-network, thereby reducing the discretization error. At the same time, due to the adoption of a progressive learning strategy, a stable search can be performed in an exponentially increasing search space, thereby being able to obtain a more complex convolutional neural network architecture.

[0011] As a preferred solution, the super network is divided into a first network module, a second network module and a third network module according to depth;

[0012] The output of the first network module is connected to the input of the second network module, and the output of the second network module is connected to the input of the third network module; the first network module and the second network module are respectively composed of two types of units, including a plurality of normal units and a reduction unit; the third network module is composed of a plurality of normal units;

[0013] The normal unit and the reduction unit each form a directed acyclic graph consisting of a plurality of nodes, wherein the nodes represent feature graphs, and the edges between the nodes represent neural operators that transform the feature graphs.

[0014] Furthermore, the architecture parameters characterize the probability distribution function of the subnetwork in the supernetwork, and are constructed in the following manner: the edges in the supernetwork are regarded as random variables, assuming that the edges in the supernetwork are independent of each other; a set of independent parameters are assigned to the edges in the supernetwork respectively and processed by the Softmax function to obtain the architecture parameters of the supernetwork.

[0015] Furthermore, the probability distribution function of the sub-network is expressed by the following formula:

[0016]

[0017] in,

[0018]

[0019] In the formula, A represents the subnetwork; p(A) represents the probability of sampling subnetwork A; k represents the unit number; i and j represent the numbers of the input node and output node of edge (i, j) respectively; E k,i,j represents the neural operator of edge (i, j) in the kth unit; α is the architecture parameter; O is the set of neural operators.

[0020] As a preferred solution, in the process of optimizing the super network using the differentiable sampler, step S2 includes a first stage, a second stage and a third stage of gradually increasing the sub-network search space;

[0021] In the first stage, all normal units and all reduction units in the super network share the same topological structure;

[0022] In the second stage, the hypernetwork and the architecture parameters of the hypernetwork need to be reconstructed according to the optimization results of the first stage before sampling optimization; in the hypernetwork reconstructed in the second stage, the normal units and the reduction units in the same module share the same topological structure, and different topological structures are allowed to exist between the normal units and the reduction units in different modules;

[0023] In the third stage, the super network and the architecture parameters of the super network need to be reconstructed according to the optimization results of the second stage and then sampled and optimized; in the super network reconstructed in the third stage, different topological structures are allowed to exist between all normal units and between all reduction units.

[0024] Furthermore, in step S2, during the forward propagation of the super network, only a number of neural operators with top sampling weights are calculated, and a Dropout operation is applied to the residual connection.

[0025] As a preferred solution, the differentiable sampler is implemented based on discrete distribution approximate sampling technology.

[0026] The present invention also provides the following contents:

[0027] A convolutional neural network architecture search system based on a differentiable sampler and progressive learning comprises a construction module, a sampling optimization module and a retraining module, wherein the sampling optimization module is connected to the construction module and the retraining module respectively, wherein:

[0028] The building module is used to build a hypernetwork, architecture parameters of the hypernetwork, and a differentiable sampler;

[0029] The sampling optimization module is used to perform sampling optimization on the super network using the differentiable sampler according to the architecture parameters in an incremental learning manner to obtain a desired network;

[0030] The retraining module is used to perform network retraining on the expected network until the expected network converges.

[0031] A medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the aforementioned convolutional neural network architecture search method based on a differentiable sampler and progressive learning are implemented.

[0032] A computer device comprises a medium, a processor and a computer program stored in the medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the aforementioned convolutional neural network architecture search method based on a differentiable sampler and progressive learning. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 A schematic diagram of the steps of a convolutional neural network architecture search method based on a differentiable sampler and progressive learning provided in Example 1 of the present invention;

[0034] Figure 2 is a general example of a hypernetwork;

[0035] Figure 3 A schematic diagram of the logical structure of a super network provided in Example 1 of the present invention;

[0036] Figure 4 A schematic diagram of the operation of the differentiable sampler provided in Example 1 of the present invention;

[0037] Figure 5 This is an example of the principle of the progressive learning strategy of Embodiment 1 of the present invention;

[0038] Figure 6 Schematic diagram of a convolutional neural network architecture search system based on differentiable samplers and progressive learning provided in Example 3 of the present invention. DETAILED DESCRIPTION

[0039] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0040] It should be clear that the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present application.

[0041] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application and the appended claims are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used herein refers to and includes any or all possible combinations of one or more associated listed items.

[0042] When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present application as detailed in the attached claims. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0043] In addition, in the description of this application, unless otherwise specified, "plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. The present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0044] In order to solve the limitations of the prior art, this embodiment provides a technical solution, and the technical solution of the present invention is further described below in conjunction with the drawings and embodiments.

[0045] Example 1

[0046] Please refer to Figure 1 , a convolutional neural network architecture search method based on differentiable sampler and progressive learning, comprising the following steps:

[0047] S1, constructing a hypernetwork, architectural parameters of the hypernetwork, and a differentiable sampler;

[0048] S2, in an incremental learning manner, using the differentiable sampler to perform sampling optimization on the super network according to the architecture parameters to obtain a desired network;

[0049] S3, retraining the expected network until the expected network converges.

[0050] Compared with the prior art, the present invention uses a differentiable sampler to directly sample and optimize the constructed super network, which can change the optimization goal of architecture search from optimizing the super network to finding the optimal probability distribution function, minimizing the expectation of the loss function of the sub-network under the probability distribution; by evaluating the performance of the sub-network, the probability distribution function of the sub-network is optimized, thereby reducing the discretization error. At the same time, due to the adoption of a progressive learning strategy, a stable search can be performed in an exponentially increasing search space, thereby being able to obtain a more complex convolutional neural network architecture.

[0051] Specifically, in the process of constructing the hypernetwork, for a given image classification task, we can first evaluate the size, resolution and other prior knowledge of the given data set and the difficulty of neural network training, and then construct a predefined hypernetwork containing all sub-networks based on the candidate neural operators in the search space.

[0052] In this embodiment, the hypernetwork is a convolutional neural network composed of a certain number of normal units and reduction units, in which the total number of units and the number of intermediate nodes in each unit are manually set and are unchangeable hyperparameters. The hypernetwork can be regarded as a directed acyclic graph. Among them, the nodes represent feature graphs, and the edges represent neural operators, which are variables to be optimized. The goal of this embodiment is to find suitable neural operators (such as deep separable convolution, residual connection pooling layer, etc.) for each edge of the hypernetwork, and achieve high-accuracy image classification tasks under the combination of these neural operators.

[0053] The difference between the normal unit and the reduction unit is that after the feature map passes through the normal unit, its width, height and number of channels remain unchanged, while after the feature map passes through the reduction unit, its width and height are reduced to 1 / 2 times of the original, and the number of channels is increased to 2 times of the original.

[0054] Among the nodes of the normal unit and the reduction unit, the first two nodes of each unit are input nodes, the last node is an output node, and the others are intermediate nodes.

[0055] See also Figure 2 , where circles represent nodes, which are feature graphs, including the middle nodes of normal units, the middle nodes of reduction units, and the output nodes of units. Edges represent neural operators, which can extract features from feature graphs connected to them. Edges of different shades (nodes and edges in the original graph are distinguished by different colors, but after adjustment, only the shades can be distinguished, which is only used as an example reference in this embodiment) represent different neural operators, which are variables to be optimized. Figure 2In the example, every 4 intermediate nodes and 1 output node form a unit.

[0056] The calculation formula of the intermediate node is as follows:

[0057]

[0058] Among them, x j is the output of the jth node; x i is the output of the i-th node; It is the calculation formula of edge (i, j), which represents a specific neural operator.

[0059] Each unit receives the output of the previous two units, and the output of the previous two units is represented as the two input nodes of the unit. The output node represents the output of the unit, which is also the input of the next unit. Therefore, the output of each unit is the concatenation of the feature maps of the intermediate nodes in the channel dimension. The output x of the normal unit and the reduction unit is out It can be expressed as follows:

[0060] x out =concat([x2,x3,...,x M ], dim = channels);

[0061] Among them, concat([x2,x3,...,x M ], dim=channels) represents the feature map x2,x3,...,x M The concatenation is performed in the channel dimension, and M represents the number of nodes.

[0062] As a preferred embodiment, please refer to Figure 3 , the super network is divided into a first network module, a second network module and a third network module according to depth;

[0063] The output of the first network module is connected to the input of the second network module, and the output of the second network module is connected to the input of the third network module; the first network module and the second network module are respectively composed of two types of units, including a plurality of normal units and a reduction unit; the third network module is composed of a plurality of normal units;

[0064] The normal unit and the reduction unit each form a directed acyclic graph consisting of a plurality of nodes, wherein the nodes represent feature graphs, and the edges between the nodes represent neural operators that transform the feature graphs.

[0065] Furthermore, the architecture parameters characterize the probability distribution function of the subnetwork in the supernetwork, and are constructed in the following manner: the edges in the supernetwork are regarded as random variables, assuming that the edges in the supernetwork are independent of each other; a set of independent parameters are assigned to the edges in the supernetwork respectively and processed by the Softmax function to obtain the architecture parameters of the supernetwork.

[0066] Furthermore, the probability distribution function of the sub-network is expressed by the following formula:

[0067]

[0068] in,

[0069]

[0070] In the formula, A represents the subnetwork; p(A) represents the probability of sampling subnetwork A; k represents the unit number; i and j represent the numbers of the input node and output node of edge (i, j) respectively; E k,i,j represents the neural operator of edge (i, j) in the kth unit; α is the architecture parameter; O is the set of neural operators.

[0071] For further information, see Figure 4 , the differentiable sampler is implemented based on discrete distribution approximate sampling technology, and can be constructed based on Gumbel-softmax technology. The process of sampling the sub-network according to the probability distribution function of the sub-network is continuous and differentiable, and can participate in the back propagation process of the loss function gradient, so the training efficiency is very high. Since this embodiment assumes that the edges in the super network are independent of each other when constructing the architecture parameters, the entire differentiable sampler is composed of independent samplers for each edge in the super network, and the differentiable samplers of each edge can be expressed by the following formula:

[0072]

[0073] Among them, S l is the lth element of the vector S, representing the component of the kth candidate neural operator after the deterministic action of the edge in the hypernetwork is continuously relaxed by the differentiable sampler; p l is the probability of selecting the kth candidate neural operator; g l is a random variable that follows the Gumbel distribution; τ is the temperature coefficient, which is used to control the degree of difference between the vector L and the One-Hot vector after continuous relaxation;

[0074] After obtaining the L vector of each edge, the forward calculation formula of each edge is:

[0075]

[0076] In the formula, edge (i,j) (x (i) ) represents the forward calculation formula of edge (i, j).

[0077] Furthermore, in the process of optimizing the super network using the differentiable sampler, step S2 includes a first stage, a second stage and a third stage of gradually increasing the sub-network search space;

[0078] In the first stage, all normal units and all reduction units in the super network share the same topological structure;

[0079] In the second stage, the hypernetwork and the architecture parameters of the hypernetwork need to be reconstructed according to the optimization results of the first stage before sampling optimization; in the hypernetwork reconstructed in the second stage, the normal units and the reduction units in the same module share the same topological structure, and different topological structures are allowed to exist between the normal units and the reduction units in different modules;

[0080] In the third stage, the super network and the architecture parameters of the super network need to be reconstructed according to the optimization results of the second stage and then sampled and optimized; in the super network reconstructed in the third stage, different topological structures are allowed to exist between all normal units and between all reduction units.

[0081] As a preferred embodiment, in order to save video memory consumption and speed up the search time in order to solve the problem of high video memory consumption in the prior art, in the process of forward propagation of the super network, step S2 only calculates several neural operators with high sampling weight rankings; and in order to suppress the dominance of residual connections in the early stage of weight sharing, a Dropout operation is also applied to the residual connections, which is dynamically adjusted with the search stage and the number of candidate neural operators.

[0082] Specifically, the core of the progressive learning strategy in this embodiment is to continuously reduce the complexity of the hypernetwork at the width level during the search process, while increasing the complexity of the hypernetwork at the depth level, so as to achieve a slow transition of the search space from simple to complex, and stabilize the architecture search process. The complexity of the hypernetwork at the width level refers to the number of candidate neural operators on each edge, and the complexity at the depth level refers to the degree of constraint on each unit structure. The importance of each neural operator can be measured by the probability of being sampled. A large sampling probability indicates that the neural operator is important. Correspondingly, if the sampling probability is small, it means that the neural operator is not that important. After each stage of optimization is completed, several neural operators with relatively small sampling probabilities on each edge will be deleted, which reduces the complexity of the hypernetwork at the width level.

[0083] Before starting the optimization in the first, second and third stages, the super network will be preheated: obtain training samples from the preset model parameter training set, sample sub-networks from the super network according to the probability distribution function of the sub-network represented by the architecture parameters, and calculate the loss function of the sub-network in the training samples, where the loss function uses the cross entropy function. Calculate the gradient of the loss function, and use the model parameter optimizer to optimize the parameters of the network model according to the gradient back propagation algorithm; in particular, this implementation will not optimize the architecture parameters of the super network during the warm-up phase.

[0084] After the warm-up round in each stage, the architecture parameters of the super network are trained: Similar to the warm-up phase, training samples are obtained from the architecture parameter training set, and sub-networks are sampled from the super network. The loss function of the sub-network in the training samples is calculated, and the architecture parameter optimizer is used to optimize the architecture parameters according to the gradient back propagation algorithm. Note that in the phase or round of optimizing the architecture parameters, the optimization of the model parameters and the optimization of the architecture parameters are performed iteratively.

[0085] See also Figure 5 , assuming that the initial number of candidate neural operators is 7, after the first stage, the number of neural operators on each edge of the unit is reduced from 8 to 5 according to the probability of being sampled, and the deleted neural operators are the ones with the smallest sampling probability; after the second stage, in order to further reduce the complexity of the hypernetwork at the width level, the number of candidate neural operators on each edge of the unit will be reduced from 5 to 3; and after the third stage, each edge of the unit retains the neural operator with the largest sampling probability.

[0086] At each stage of progressive learning, the model parameters and architecture parameters of the hypernetwork are reinitialized, and then the hypernetwork is retrained using the data set given by the task. Because the present embodiment uses a weight sharing strategy, when a subnetwork is sampled from a hypernetwork using a differentiable sampler, the model parameters of the subnetwork inherit the model parameters of the hypernetwork. Therefore, in the initial stage of training the network, since the existence of residual connections is conducive to quickly improving the performance of the network, the differentiable sampler will tend to choose residual connections, that is, in the probability distribution function, the probability of residual connections will become larger and larger. According to prior knowledge, although the existence of residual connections is conducive to the training of network models, it does not contribute to the final accuracy of the image classification task, so it is necessary to suppress residual connections. The present invention uses random discarding technology to constrain the strong phenomenon of residual connections. Specifically, when a residual connection is sampled by a differentiable sampler, it is not forward propagated with a certain probability, that is, the residual connection operation will be replaced by a zero operation. In one example, the probability of discarding a residual connection is 0.5.

[0087] At each stage of progressive learning, the differentiable sampler samples a subnetwork from the supernetwork, calculates the loss function of the subnetwork in a given task and calculates the gradient of the model parameters and the architecture parameters, and uses the back propagation algorithm to simultaneously optimize the model parameters and the architecture parameters until the number of optimization rounds in each stage reaches a preset number of rounds. In an optional embodiment, the number of rounds in each stage can be set to 25. After the last round of progressive learning, that is, after the end of the third stage, each edge of the unit will retain the neural operator with the highest sampling probability, thereby obtaining the desired network.

[0088] Furthermore, in each stage, the optimization objective of the differentiable sampler is expressed as:

[0089]

[0090]

[0091] in, is the probability distribution function of the sub-network, characterized by the architecture parameter α; L val and L train They are the validation set loss function and training set loss function of the sub-network during the optimization process; are the corresponding model parameters of sub-network A.

[0092] Example 2

[0093] This embodiment is a further description based on Embodiment 1 in combination with specific parameter settings and with a more specific example, wherein:

[0094] This embodiment performs architecture search on the CIFAR-10 dataset. Specifically, the CIFAR-10 dataset contains 60,000 images, of which 50,000 are training sets and 10,000 are validation sets, with an image resolution of 32 × 32 and 3 channels.

[0095] For the CIFAR-10 dataset: This embodiment predefines the number of channels of the three-stage hypernetwork to be [9, 13, 19], and in the progressive learning, the number of candidate neural operators retained in the three stages is [8, 5, 3], respectively.

[0096] The candidate neural operators include: 3×3 depthwise separable convolution, 5×5 depthwise separable convolution, 3×3 dilated convolution, 5×5 dilated convolution, 3×3 maximum pooling layer, 3×3 average pooling layer, residual connection, and zero operation.

[0097] The Dropout of the residual connections in the three stages are [0, 0.4, 0.7] respectively.

[0098] The forward calculation process of the hypernetwork only calculates the five candidate neural operators with the largest sampling weights.

[0099] The total number of the normal units and reduction units is 20.

[0100] The number of intermediate nodes in each unit is 4.

[0101] The learning rate of the hypernetwork model parameters is 0.25; the learning rate of the hypernetwork architecture parameters is 0.001; the optimizer of the hypernetwork model parameters is a stochastic gradient optimizer with a momentum of 0.9 and a cosine decay learning rate.

[0102] The optimizer for the hypernetwork architecture parameters is the ADMA optimizer with momentum (0.5-0.999) and no cosine decay learning rate.

[0103] The number of rounds in the three stages is 20 respectively, the number of warm-up rounds for the hypernetwork is 10, and the number of batches for the forward computation of the hypernetwork is 96.

[0104] The ratio of the training set to the validation set of the dataset is 0.5.

[0105] The expected network thus obtained in this embodiment can achieve an accuracy of 97.1% on the verification set after network retraining.

[0106] Example 3

[0107] A convolutional neural network architecture search system based on a differentiable sampler and progressive learning, see FIG6 , includes a construction module 1, a sampling optimization module 2 and a retraining module 3, wherein the sampling optimization module 2 is connected to the construction module 1 and the retraining module 3, respectively, wherein:

[0108] The construction module 1 is used to construct a hypernetwork, architecture parameters of the hypernetwork, and a differentiable sampler;

[0109] The sampling optimization module 2 is used to perform sampling optimization on the super network using the differentiable sampler according to the architecture parameters in an incremental learning manner to obtain a desired network;

[0110] The retraining module 3 is used to retrain the desired network until the desired network converges.

[0111] Example 4

[0112] A medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the steps of the convolutional neural network architecture search method based on a differentiable sampler and progressive learning in Example 1 are implemented.

[0113] Example 5

[0114] A computer device includes a medium, a processor, and a computer program stored in the medium and executable by the processor, wherein the computer program, when executed by the processor, implements the steps of the convolutional neural network architecture search method based on a differentiable sampler and progressive learning in Example 1.

[0115] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A convolutional neural network architecture search method based on differentiable samplers and progressive learning, which is used to automatically search for neural network architectures in image classification tasks, characterized in that: The following steps are involved: S1, constructing a hypernetwork, architectural parameters of the hypernetwork, and a differentiable sampler; S2, using the image data set, in a progressive learning manner, according to the architecture parameters, using the differentiable sampler to perform sampling optimization on the super network to obtain a desired network; The super network is divided into a first network module, a second network module and a third network module according to depth; The output of the first network module is connected to the input of the second network module, and the output of the second network module is connected to the input of the third network module; the first network module and the second network module are respectively composed of two types of units, including a plurality of normal units and a reduction unit; the third network module is composed of a plurality of normal units; The normal unit and the reduction unit each form a directed acyclic graph consisting of a plurality of nodes, wherein the nodes represent feature graphs, and the edges between the nodes represent neural operators that transform the feature graphs; In the process of optimizing the super network using the differentiable sampler, the first stage, the second stage and the third stage of gradually increasing the sub-network search space are included; In the first stage, all normal units and all reduction units in the super network share the same topological structure; In the second stage, the hypernetwork and the architecture parameters of the hypernetwork need to be reconstructed according to the optimization results of the first stage before sampling optimization; in the hypernetwork reconstructed in the second stage, the normal units and the reduction units in the same module share the same topological structure, and different topological structures are allowed to exist between the normal units and the reduction units in different modules; In the third stage, the super network and the architecture parameters of the super network need to be reconstructed according to the optimization results of the second stage before sampling optimization; in the super network reconstructed in the third stage, different topological structures are allowed to exist between all normal units and between all reduction units; S3, retraining the expected network until the expected network converges to obtain a final neural network architecture for the image classification task.

2. The convolutional neural network architecture search method based on differentiable sampler and progressive learning according to claim 1, characterized in that: The architecture parameters characterize the probability distribution function of the subnetwork in the supernetwork, and are constructed in the following manner: the edges in the supernetwork are regarded as random variables, and it is assumed that the edges in the supernetwork are independent of each other; a set of independent parameters are assigned to the edges in the supernetwork respectively and processed by the Softmax function to obtain the architecture parameters of the supernetwork.

3. The convolutional neural network architecture search method based on differentiable sampler and progressive learning according to claim 2, characterized in that: The probability distribution function of the subnetwork is expressed by the following formula: ； in, ； In the formula, A represents a subnetwork; p(A) Indicates sampling to subnetwork A probability; k Indicates the unit number; i and j Respectively represent the edge (i, j) The serial numbers of input nodes and output nodes; Indicates k Unit mid-edge (i, j) Neural operators; is the architecture parameter; O is a set of neural operators.

4. The convolutional neural network architecture search method based on differentiable sampler and progressive learning according to claim 3, characterized in that: In the process of forward propagation of the super network, step S2 only calculates a number of neural operators with top sampling weight rankings, and also applies a Dropout operation to the residual connection.

5. The convolutional neural network architecture search method based on differentiable sampler and progressive learning according to claim 1, characterized in that: The differentiable sampler is implemented based on discrete distribution approximate sampling technology.

6. A convolutional neural network architecture search system based on a differentiable sampler and progressive learning, applied to the convolutional neural network architecture search method based on a differentiable sampler and progressive learning as described in any one of claims 1 to 5, characterized in that: The system comprises a construction module (1), a sampling optimization module (2) and a retraining module (3), wherein the sampling optimization module (2) is connected to the construction module (1) and the retraining module (3), respectively, wherein: The building module (1) is used to build a hypernetwork, the architecture parameters of the hypernetwork and a differentiable sampler; The sampling optimization module (2) is used to perform sampling optimization on the super network using the differentiable sampler according to the architecture parameters in an incremental learning manner to obtain a desired network; The retraining module (3) is used to retrain the desired network until the desired network converges.

7. A medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the convolutional neural network architecture search method based on a differentiable sampler and progressive learning are implemented as described in any one of claims 1 to 5.

8. A computer device, characterized in that: It includes a medium, a processor, and a computer program stored in the medium and executable by the processor, and when the computer program is executed by the processor, the steps of the convolutional neural network architecture search method based on a differentiable sampler and progressive learning as described in any one of claims 1 to 5 are implemented.

Images