WO2023071592A1

WO2023071592A1 - Network structure search method for ultra-large search space, system and medium

Info

Publication number: WO2023071592A1
Application number: PCT/CN2022/119120
Authority: WO
Inventors: 谭明奎; 国雍; 陈耀佛
Original assignee: 华南理工大学; 人工智能与数字经济广东省实验室(广州)
Priority date: 2021-10-27
Filing date: 2022-09-15
Publication date: 2023-05-04
Also published as: CN114065003A

Abstract

A network structure search method for an ultra-large search space, a system and a medium. The method comprises the following steps: constructing a target data set for training a neural network; determining a neural network search space for a target task; when gradually expanding the search space, searching for an efficient network structure by using a reinforcement learning algorithm; and according to the target data set, training the search network structure obtained by searching, so as to obtain a final search network structure. The means of gradually increasing search space can effectively reduce search difficulty, improve search efficiency and performance, and can be widely used in the field of artificial intelligence.

Description

Network structure search method, system and medium for super large search space

technical field

The invention relates to the field of artificial intelligence, in particular to a network structure search method, system and medium for a super large search space.

Background technique

In recent years, deep neural networks have been widely used in different tasks. With the increase in the number of parameters and the number of network layers and the effective utilization of the GPU, the neural network model has significantly improved in accuracy and efficiency. However, most of the current neural networks are manually designed, and the design process relies on rich experience in network structure design, which brings high design costs. This makes it difficult for deep neural network models to be applied to many real-world engineering tasks.

To address the above issues, neural network search attempts to automatically design efficient neural network architectures. In order to search for efficient network architectures, existing methods must explore the search space by sampling enough network structures. However, the search space is often very large (often containing billions of candidate network architectures), and under the condition of limited computing resources, only a small part of the network structure can be sampled from the search space, which greatly restricts the exploration of the search space. efficiency. In order to find a better network structure, it is necessary to explore the entire search space as much as possible. Given a limited number of network structure sampling times, this problem places very high requirements on the sampling accuracy during the search process.

Contents of the invention

In order to solve one of the technical problems in the prior art at least to a certain extent, the purpose of the present invention is to provide a network structure search method, system and medium for a super large search space.

The technical scheme adopted in the present invention is:

A network structure search method for super large search space, comprising the following steps:

Construct the target dataset for training the neural network;

Determine the neural network search space for the target task;

In the process of gradually expanding the search space, a reinforcement learning algorithm is used to search for an efficient network structure;

The search network structure obtained by searching is trained according to the target data set to obtain the final search network structure.

Further, the target data set for training the neural network is constructed, including:

Collect pictures from the target task scene, and classify the pictures to build the target data set;

Divide the labeled target data set into three parts: training set, verification set and test set.

Further, said determining the neural network search space for the target task includes:

Divide the computing units constituting the deep convolutional neural network model into standard computing units and downsampling computing units;

Set the search space of the computing unit.

Further, in the process of gradually expanding the search space, the reinforcement learning algorithm is used to search for an efficient network structure, including:

A1. Use a single-layer bidirectional long-short-term memory network to construct a meta-controller to generate a network architecture α~π(α,θ); where α is the generated network structure, π is the strategy learned by the meta-controller, and θ is Meta-controller network parameters;

A2. Construct a hypernetwork model oriented to image recognition tasks, the hypernetwork model is obtained by stacking several computing units; each computing unit contains multiple candidate operations between the input feature and the output feature;

A3. Construct the initial search space according to the candidate operations, gradually add candidate operations during the search process, and expand the new search space on the basis of the previous search space; the whole search process includes K search stages, each stage corresponds to a different size The search space Ω _i ;

A4. Use the meta-controller to generate a sub-network architecture α～π(α,θ) in the current search space, and activate the candidate operation weight w _α corresponding to the sub-network structure in the super-network model; and train on the target data set Hypernetwork model;

A5. Use the meta-controller to generate a sub-network architecture α～π(α,θ) in the current search space, and obtain the weight w _α of the sub-network model by inheriting the weight of the corresponding operation in the super-network model; in the divided verification data Set the performance index R(α,w _α ) obtained from the set test and use it as a reward value to update the weight θ of the meta-controller;

A6. Repeat step A3 to step A5 until all K candidate operations have been added to the search space.

Further, the search network structure obtained by searching is trained according to the target data set to obtain the final search network structure, including:

Obtain a high-performance network structure based on the trained meta-controller model;

According to the target data set, the obtained network structure model is trained to obtain the final search network structure model.

Further, the training of the obtained network structure includes:

The stochastic gradient descent algorithm is used to train the obtained network structure model until the network structure model converges.

Further, the step A3 specifically includes:

If the initial search space Ω ₀ has not yet been constructed, a candidate operation is randomly selected from all candidate operations to construct the initial search space;

If the current search space has been constructed, a candidate operation is randomly selected from the remaining candidate operations and added to the search space Ω _i-1 to construct a new search space Ω _i .

Another technical scheme adopted in the present invention is:

A network structure search system for a super large search space, including:

Dataset building blocks for constructing target datasets for training neural networks;

A search space determination module is used to determine the neural network search space for the target task;

The space search module is used to search for an efficient network structure using a reinforcement learning algorithm in the process of gradually expanding the search space;

The model training module is used to train the search network structure obtained by searching according to the target data set to obtain the final search network structure.

Another technical scheme adopted in the present invention is:

A network structure search system for a super large search space, including:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the above method.

Another technical scheme adopted in the present invention is:

A storage medium stores a processor-executable program therein, and the processor-executable program is used to execute the above method when executed by a processor.

The beneficial effects of the present invention are: the method of gradually increasing the search space adopted by the present invention can effectively reduce the search difficulty and improve the search efficiency and performance.

Description of drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following describes the accompanying drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the accompanying drawings in the following introduction are only In order to clearly describe some embodiments of the technical solutions of the present invention, those skilled in the art can also obtain other drawings based on these drawings without creative work.

Fig. 1 is a flow chart of the steps of a network structure search method for a super large search space in an embodiment of the present invention;

FIG. 2 is a schematic diagram of a gradually expanded network structure search space in an embodiment of the present invention;

Fig. 3 is a schematic diagram of the comparison of the search methods between the standard neural network search method and the automatic network structure search method for the super large search space proposed by the embodiment of the present invention.

Detailed ways

Embodiments of the present invention are described in detail below, examples of which are shown in the drawings, wherein the same or similar reference numerals designate the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by referring to the figures are exemplary only for explaining the present invention and should not be construed as limiting the present invention. For the step numbers in the following embodiments, it is only set for the convenience of illustration and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art sexual adjustment.

In the description of the present invention, it should be understood that the orientation descriptions, such as up, down, front, back, left, right, etc. indicated orientations or positional relationships are based on the orientations or positional relationships shown in the drawings, and are only In order to facilitate the description of the present invention and simplify the description, it does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and thus should not be construed as limiting the present invention.

In the description of the present invention, several means one or more, and multiple means two or more. Greater than, less than, exceeding, etc. are understood as not including the original number, and above, below, within, etc. are understood as including the original number. If the description of the first and second is only for the purpose of distinguishing the technical features, it cannot be understood as indicating or implying the relative importance or implicitly indicating the number of the indicated technical features or implicitly indicating the order of the indicated technical features relation.

In the description of the present invention, unless otherwise clearly defined, words such as setting, installation, and connection should be understood in a broad sense, and those skilled in the art can reasonably determine the specific meanings of the above words in the present invention in combination with the specific content of the technical solution.

The overall scheme of the present invention is shown in Fig. 1, the first step needs to construct the target data set and set the search space of the target task. Then construct an initial relatively small search space and search in it, gradually expand the search space by gradually increasing candidate calculation operations, and use the previously learned network structure to perform more accurate sampling in a larger search space , so that the search space can be more fully explored under a limited number of sampling times. This method can reduce the search difficulty in the search process to find a better network structure. Finally, the network structure obtained from the search is trained on the target data set until convergence, and the model can be deployed and applied to the target task. Each step in the figure will be described in detail below in conjunction with Figure 1:

S1: Collect and label the target dataset for training the neural network.

S1-1: Collect pictures from the target task scene and classify the pictures to build a data set;

S1-2: Divide the labeled data set into three parts: training set, verification set, and test set.

S2: Set the neural network search space for the target task.

S2-1: For computer vision tasks, the computing units of the deep convolutional neural network model can be divided into two categories: standard computing units and downsampling computing units. The search space of these two types of computing units is the same, but the standard computing unit maintains the spatial resolution of the input features, while the downsampling computing unit reduces the spatial resolution of the output features to half of the input. A complete convolutional neural network can be constructed by stacking computing units several times.

S2-2: Set the search space of the computing unit: each computing unit contains 7 nodes, including 2 input nodes, 4 intermediate nodes, and 1 output node. There can be K=8 different candidate operations between two nodes, including 3×3 depth separable convolution, 5×5 depth separable convolution, 3×3 maximum pooling, 3×3 average pooling, Operations such as 3×3 hole convolution, 5×5 hole convolution, jump operation and empty operation. After each convolution operation, a batch normalization budget is performed and a ReLU activation function is used.

S3: Search for an efficient network structure in a gradually expanding search space using a reinforcement learning algorithm.

S3-1: Construct a meta-controller using a single-layer bidirectional long-short-term memory network to generate a network architecture α∼π(α,θ), where α is the network structure and π is the strategy learned by the controller. Initialize the meta-controller network parameters θ;

S3-2: Construct a hypernetwork model for image recognition tasks, where the weight parameter is denoted as w. The entire hypernetwork model is obtained by stacking several computing units, including standard computing units and downsampling computing units; as shown in Figure 2, each computing unit (ie, a node in the graph) contains multiple Candidate operations (that is, connecting lines between nodes in the figure), only one candidate operation will be activated during the training and testing phases of the network model, and will eventually be used for the target task;

S3-3: The whole search process includes K search stages, and each stage corresponds to a search space Ω _i of a different size. If the initial search space Ω ₀ has not been constructed, a candidate operation is randomly selected from all candidate operations to construct the initial search space. Because it is impossible to train a hypernetwork model without parameters, it is stipulated that the first added operation must have parameters (such as convolution). If the current search space has been constructed, a candidate operation is randomly selected from the remaining candidate operations and added to the search space Ω _i-1 to construct a new search space Ω _i . Refer to Figure 2 for a schematic diagram of this process, where the connecting lines of different colors between nodes represent different candidate operations. In the initial stage, there is only one candidate operation in the entire search space, and the search space is relatively small, so it is easier to find a network structure with excellent performance in this space; as the search stage increases, a random operation is continuously selected from the remaining candidate operations. species are added to the search space to form a new search space;

Preheat the newly added candidate operation weights: In order to improve the fairness of competition between different computing operations, the super network is trained by randomly sampling the network structure, and each computing operation is trained with equal probability. In this way, candidate network structures that incorporate new computational operations can achieve comparable performance to other network structures. By operating the warm-up method, the search process becomes more stable and the search performance is significantly improved. Specifically, a sub-network architecture α∼p(α,Ω _i ) is randomly sampled from the search space, where p represents a uniform distribution and Ω _i represents the search space of the i-th stage. The parameter weights corresponding to the sub-network model and the super-network model are shared. After randomly sampling the sub-network structure, the generated sub-network model weight w _α is trained on the divided training set by stochastic gradient descent algorithm.

S3-4: Train the supernetwork model: use the meta-controller to generate a subnetwork architecture α∼π(α,θ) in the current search space, and activate its corresponding candidate operation weight w _α in the supernetwork model. Select a batch of sample data from the divided training data set, and train the generated sub-network model through the stochastic gradient descent algorithm.

S3-5: Train the meta-controller: use the meta-controller to generate a sub-network architecture α～π(α,θ) in the current search space, and obtain the weight w of the sub-network model by inheriting the weight of the corresponding operation in the super-network model _α ; select a batch of sample data from the divided verification data set, and test the performance index R(α,w _α ) of the generated sub-network model on the target task on the sample data; the performance index R( α, w _α ) as the reward value to update the weight θ of the meta-controller through the reinforcement learning policy gradient algorithm;

S3-6: Repeat steps S3-3 to S3-5 until all K candidate operations have been added to the search space.

S4: Infer the network structure and train the model on the target dataset.

S4-1: Based on a trained meta-controller model, a high-performance network structure can be inferred. Given K candidate operations, select the policy π(α,θ) learned in the final stage (i.e., the stage with the largest search space) as the final policy for sampling the network structure. Based on this strategy, 10 network structures are sampled, and then the network structure with the highest classification accuracy on the verification set is selected as the final search network structure.

S4-2: Use the stochastic gradient descent algorithm to train the searched network architecture model to convergence on the divided training data set.

In summary, compared with the prior art, the method of the embodiment of the present invention has the following beneficial effects:

Existing neural network search methods usually search directly in a fixed super-large search space. Due to the huge search difficulty brought about by the super-large search space, they can only find suboptimal network structures. Different from the existing methods, the method of gradually increasing the search space adopted by this scheme can effectively reduce the search difficulty and improve the search efficiency and performance. As shown in Figure 3, once this scheme finds some good network structures in a small search space, by gradually expanding the search space, it is more likely to find a network structure that is similar to the optimal network structure found in the previous subspace. Candidate subspaces with higher similarity (gray circles). Therefore, it is more likely to find a better network structure (at least with similar performance to the optimal network structure of the previous subspace) when sampling in the new subspace, and reduce sampling to a very poor network structure and waste sampling opportunities to Improve sampling accuracy. This scheme searches in the gradually expanding search space, which is essentially a multi-stage search process, and the candidate subspace will evolve adaptively with the search process. Since the candidate subspace is relatively small in each search stage, this method can perform more accurate sampling and achieve high-performance network structures in a larger search space.

The beneficial effects of the method of this embodiment are demonstrated below in combination with experimental data.

The automatic network structure search method for super-large search space proposed in this embodiment can reduce the difficulty of neural network search in the initial stage. As the search stage progresses, this solution will gradually increase the search space, and use the The structure of assisted the next stage of the search. Table 1 and Table 2 show the comparison results with the best existing methods on the CIFAR-10 dataset and the ImageNet dataset, respectively. After applying this scheme, the search cost can be reduced and the search performance can be improved on two commonly used image recognition datasets.

Table 1

Table 2

This embodiment also provides a network structure search system oriented to a super large search space, including:

A network structure search system oriented to a super-large search space in this embodiment can execute a network structure search method oriented to a super-large search space provided by the method embodiments of the present invention, and can perform any combination of implementation steps of the method embodiments. Corresponding functions and beneficial effects of the method.

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the method shown in FIG. 1 .

A network structure search system oriented to a very large search space in this embodiment can execute a network structure search method oriented to a very large search space provided by the method embodiments of the present invention, and can perform any combination of implementation steps of the method embodiments, with Corresponding functions and beneficial effects of the method.

The embodiment of the present application also discloses a computer program product or computer program, where the computer program product or computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the computer device can read the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device executes the method shown in FIG. 1 .

This embodiment also provides a storage medium, which stores an instruction or program capable of executing a super-large search space-oriented network structure search method provided by the method embodiment of the present invention. When the instruction or program is executed, the method can be executed. Any combination of implementation steps in the embodiments has the corresponding functions and beneficial effects of the method.

In some alternative implementations, the functions/operations noted in the block diagrams may occur out of the order noted in the operational diagrams. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks may sometimes be executed in the reverse order, depending upon the functionality/operations involved. Furthermore, the embodiments presented and described in the flowcharts of the present invention are provided by way of example in order to provide a more comprehensive understanding of the technology. The disclosed methods are not limited to the operations and logical flow presented herein. Alternative embodiments are contemplated in which the order of various operations is changed and in which sub-operations described as part of larger operations are performed independently.

Furthermore, although the invention has been described in the context of functional modules, it should be understood that one or more of the described functions and/or features may be integrated into a single physical device and/or unless stated to the contrary. or software modules, or one or more functions and/or features may be implemented in separate physical devices or software modules. It will also be appreciated that a detailed discussion of the actual implementation of each module is not necessary to understand the present invention. Rather, given the attributes, functions and internal relationships of the various functional blocks in the devices disclosed herein, the actual implementation of the blocks will be within the ordinary skill of the engineer. Accordingly, those skilled in the art can implement the present invention set forth in the claims without undue experimentation using ordinary techniques. It is also to be understood that the particular concepts disclosed are illustrative only and are not intended to limit the scope of the invention which is to be determined by the appended claims and their full scope of equivalents.

If the functions described above are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the essence of the technical solution of the present invention or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute all or part of the steps of the methods described in various embodiments of the present invention. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc., which can store program codes. .

The logic and/or steps represented in the flowcharts or otherwise described herein, for example, can be considered as a sequenced listing of executable instructions for implementing logical functions, can be embodied in any computer-readable medium, For use with instruction execution systems, devices, or devices (such as computer-based systems, systems including processors, or other systems that can fetch instructions from instruction execution systems, devices, or devices and execute instructions), or in conjunction with these instruction execution systems, devices or equipment used. For the purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate or transmit a program for use in or in conjunction with an instruction execution system, device or device.

More specific examples (non-exhaustive list) of computer-readable media include the following: electrical connection with one or more wires (electronic device), portable computer disk case (magnetic device), random access memory (RAM), Read Only Memory (ROM), Erasable and Editable Read Only Memory (EPROM or Flash Memory), Fiber Optic Devices, and Portable Compact Disc Read Only Memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program can be printed, since the program can be read, for example, by optically scanning the paper or other medium, followed by editing, interpretation or other suitable processing if necessary. The program is processed electronically and stored in computer memory.

It should be understood that various parts of the present invention can be realized by hardware, software, firmware or their combination. In the embodiments described above, various steps or methods may be implemented by software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented by any one or combination of the following techniques known in the art: Discrete logic circuits, ASICs with suitable combinational logic gates, programmable gate arrays (PGAs), field programmable gate arrays (FPGAs), etc.

In the above description of this specification, the description with reference to the terms "one embodiment/example", "another embodiment/example" or "certain embodiments/example" means that the description is described in conjunction with the embodiment or example. A particular feature, structure, material, or characteristic is included in at least one embodiment or example of the invention. In this specification, schematic representations of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the described specific features, structures, materials or characteristics may be combined in any suitable manner in any one or more embodiments or examples.

Although the embodiments of the present invention have been shown and described, those skilled in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principle and spirit of the present invention. The scope of the invention is defined by the claims and their equivalents.

The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the above-mentioned embodiments, and those skilled in the art can also make various equivalent deformations or replacements without violating the spirit of the present invention. Equivalent modifications or replacements are all within the scope defined by the claims of the present application.

Claims

A network structure search method oriented to an ultra-large search space, characterized in that it comprises the following steps:

Construct the target dataset for training the neural network;

Determine the neural network search space for the target task;

In the process of gradually expanding the search space, a reinforcement learning algorithm is used to search for an efficient network structure;

The search network structure obtained by searching is trained according to the target data set to obtain the final search network structure.
A kind of network structure search method facing super large search space according to claim 1, it is characterized in that, described construction is used for training the target data set of neural network, comprises:

Collect pictures from the target task scene, and classify the pictures to build the target data set;

Divide the labeled target data set into three parts: training set, verification set and test set.
A network structure search method oriented to a super large search space according to claim 1, wherein said determining the neural network search space used for the target task comprises:

Divide the computing units constituting the deep convolutional neural network model into standard computing units and downsampling computing units;

Set the search space of the computing unit.
A network structure search method for a super large search space according to claim 1, characterized in that, in the process of gradually expanding the search space, using a reinforcement learning algorithm to search for an efficient network structure, including:

A1. Use a single-layer bidirectional long-short-term memory network to construct a meta-controller to generate a network architecture α~π(α,θ); where α is the generated network structure, π is the strategy learned by the meta-controller, and θ is Parameters of the meta-controller network;

A2. Construct a hypernetwork model oriented to image recognition tasks, the hypernetwork model is obtained by stacking several computing units; each computing unit contains multiple candidate operations between the input feature and the output feature;

A3. Construct the initial search space according to the candidate operations, gradually add candidate operations during the search process, and expand the new search space on the basis of the previous search space; the whole search process includes K search stages, each stage corresponds to a different size The search space Ω i ;

A4. Use the meta-controller to generate a sub-network architecture α～π(α,θ) in the current search space, and activate the candidate operation weight w α corresponding to the sub-network structure in the super-network model; and train on the target data set Hypernetwork model;

A5. Use the meta-controller to generate a sub-network architecture α～π(α,θ) in the current search space, and obtain the weight w α of the sub-network model by inheriting the weight of the corresponding operation in the super-network model; in the divided verification data Set the performance index R(α,w α ) obtained from the set test and use it as a reward value to update the weight θ of the meta-controller;

A6. Repeat step A3 to step A5 until all K candidate operations have been added to the search space.
A kind of network structure search method facing super large search space according to claim 1, it is characterized in that, the described search network structure obtained by searching according to the target data set is trained to obtain the final search network structure, comprising:

Obtain a high-performance network structure based on the trained meta-controller model;

According to the target data set, the obtained network structure is trained to obtain the final search network structure model.
A network structure search method oriented to a super large search space according to claim 5, wherein said training the obtained network structure comprises:

The stochastic gradient descent algorithm is used to train the obtained network structure model until the network structure model converges.
A network structure search method oriented to an ultra-large search space according to claim 4, wherein said step A3 specifically includes:

If the initial search space Ω 0 has not yet been constructed, a candidate operation is randomly selected from all candidate operations to construct the initial search space;

If the current search space has been constructed, a candidate operation is randomly selected from the remaining candidate operations and added to the search space Ω i-1 to construct a new search space Ω i .
A network structure search system oriented to an ultra-large search space, characterized in that it includes:

Dataset building blocks for constructing target datasets for training neural networks;

A search space determination module is used to determine the neural network search space for the target task;

The space search module is used to search for an efficient network structure using a reinforcement learning algorithm in the process of gradually expanding the search space;

The model training module is used to train the search network structure model obtained by searching according to the target data set to obtain the final search network structure model.
A network structure search system oriented to an ultra-large search space, characterized in that it includes:

at least one processor;

at least one memory for storing at least one program;

When the at least one program is executed by the at least one processor, the at least one processor implements the method according to any one of claims 1-7.
A storage medium storing a processor-executable program therein, wherein the processor-executable program is used to execute the method according to any one of claims 1-7 when executed by a processor.