CN112784949A - Neural network architecture searching method and system based on evolutionary computation - Google Patents

Abstract

The invention provides a neural network architecture search method and system based on evolutionary computing, including: setting target requirements through an objective function according to target requirements and platform requirements; randomly generating N based on the set of sub-network modules according to the set size of the search space. Undirected cyclic graph, as a network search space for evolutionary optimization; under the guidance of heuristic information, combined with pheromone dynamic volatilization and probabilistic path selection mechanism, the ant colony algorithm is used to search randomly generated N undirected cyclic graphs N directed acyclic graph optimization paths form a candidate set; obtain the accuracy and inference delay of the N optimization paths in the candidate set through training and testing, and select the optimal result as the current optimal network structure; evaluate the current network architecture whether the target requirements are met. The invention has certain application flexibility and expansibility, and obtains a neural network model that achieves a good trade-off between accuracy and speed in the case of limited resources.

Description

A neural network architecture search method and system based on evolutionary computing

technical field

The invention relates to the technical field of architecture design and optimization of deep neural networks, in particular, to a method and system for searching neural network architectures based on evolutionary computing.

Background technique

Deep learning has a powerful automatic feature extraction function for unstructured data and a powerful automatic representation ability, so it has made major breakthroughs and progress in many fields such as machine translation, image recognition, speech recognition, target detection, and natural language processing. . Based on the importance of the design of neural network architectures to the feature representation of the data and the final performance, researchers have focused on designing various complex neural network architectures to obtain good data feature representations. However, the design of neural network architectures heavily relies on researchers' prior knowledge and experience, which requires a lot of time and effort. However, with the existing prior knowledge and fixed thinking paradigm of human beings, it is difficult to find a better network architecture to a certain extent, and it is difficult for beginners to make reasonable modifications to the network architecture according to their actual needs. Therefore, Neural Architecture Search (NAS) came into being. NAS aims to use algorithms to automatically design neural network architectures with optimal performance under limited computing resources, minimizing manual intervention. The research that the network architecture obtained by using the reinforcement learning method to achieve SOTA classification accuracy on the image classification task is considered to be the pioneering work of NAS, and it also shows that the idea of automatic network architecture design is feasible. Subsequently, the research work of large-scale evolutionary computation using evolutionary learning to obtain similar results has once again verified the feasibility of this idea. NAS has been rapidly applied in object detection, semantic segmentation, adversarial learning, building scale, and multi-objective optimization.

Since the NAS method requires strong computing power and consumes a huge amount of computing power, there have been attempts to reconstruct the search space to reduce the search range, reduce the search complexity, and accelerate the network architecture through strategies such as parameter sharing, model reuse, and gradient optimization. Search for research that reduces computational effort. Early NAS trains each candidate network architecture from scratch in the architecture search phase, resulting in a surge in computation; although a parameter sharing strategy is adopted to speed up the process of architecture search, it is likely to result in inaccurate ranking of candidate architectures, which will make NAS It is difficult to select the optimal network architecture from a large number of candidate architectures, further reducing the performance of the final searched network architecture. The one-shot-based differentiable neural network architecture search method relaxes the search space from discrete to continuous, so that gradient descent can be used to search the architecture and learn weights at the same time, which shortens the search time, but when the number of search rounds is too large, it will lead to The searched architecture contains many skip connections, making the network shallower. Shallow networks have fewer parameters to learn and have weaker expressive power, resulting in a sharp drop in network performance. Although the improved differentiable neural network architecture search method uses the early stop mechanism to directly control the number of skip connections, the timing of early stop is an important issue, and premature stop will lead to incomplete architecture search. Therefore, how to achieve a balance between performance and efficiency under resource-constrained conditions is an urgent problem to be solved.

Domestic patent CN111353313A discloses a sentiment analysis model construction method based on evolutionary neural network architecture search, including the following steps: group initialization; taking the embedding layer as the first layer, encapsulating several convolutional layer units, several pooling units and data. A fully connected unit ends with a fully connected unit, and M chromosomes are randomly generated; the fitness is evaluated by using the accuracy rate as the fitness function; the roulette method is used to select several chromosome individuals to form the first chromosome population; The long chromosome crossover method performs pairwise crossover of the chromosome individuals of the first chromosome population to obtain several chromosome individuals to form the second chromosome population; The connection units are added, replaced or deleted; the fitness of the chromosome individuals of the second chromosome population is calculated until the preset number of iterations is reached, and the chromosome individuals with the optimal neural network structure are selected by the fitness.

Domestic patent CN111144555A discloses a cyclic neural network architecture search method, system and medium based on an improved evolutionary algorithm. The method of the invention includes training a plurality of cyclic neural network sub-models to update shared weights; The history record table of network model performance; randomly sample from the population to generate samples, select the optimal model of the sample for mutation operation, remove the oldest or worst model in the population with a specified probability, and add the mutated child nodes to the population and history records Table; judge whether the preset end conditions are met, if not, continue to perform sample mutation, otherwise output the optimal model in the history table. The invention can speed up the process of searching for the cyclic neural network architecture, and simultaneously consider performance and search time when updating the population in each step, and can greatly improve the efficiency of the cyclic neural network architecture searching.

The domestic patent CN110728355A discloses a neural network architecture search method, device, computer equipment and storage medium, and relates to the technical field of deep learning, wherein the method may include: dividing the neural network architecture into M substructures, where M is a positive integer greater than 1; The topology in each substructure; the neural network architecture is obtained by connecting the topology in each substructure, which can improve the search speed.

The domestic patent CN110232434A discloses a neural network architecture evaluation method based on attribute graph optimization. The neural network architecture is modeled as an attribute graph, and a Bayesian graph neural network proxy model is constructed; a set of neural network architectures are randomly generated, trained and tested. , take this set of neural network architecture and the performance indicators corresponding to the test as the initial training set, and the training set is used to train the Bayesian graph neural network proxy model; according to the current training set, a new neural network candidate set is generated and trained by the evolutionary algorithm Bayesian graph neural network proxy model; and select a potential individual from the neural network candidate set by maximizing the acquisition function, then train and test the individual, and add it and the performance indicators corresponding to the test to the current training set; Under the constraint of fixed cost, the above steps are repeated until the best neural network architecture and the corresponding weights of the architecture are obtained in the current training set.

The foreign patent JP2020522035A discloses a method, system and device for determining the structure of a neural network. The method includes using a controller neural network with controller parameters to generate a batch of output sequences based on current values of the controller parameters. Output sequences in batches generate instances of a subconvolutional neural network (CNN), including multiple instances of the first convolutional unit with an architecture defined by the output sequence; train instances of the subCNN to perform image processing tasks, and evaluate The performance of the sub-CNN training instance, which is used to determine the performance metric of the sub-CNN training instance, and includes using the performance metric of the trained CNN training instance to adjust the current values of the controller parameters of the controller neural network.

Foreign patent WO2018081563A1 discloses a method, system and device for determining the architecture of a neural network. The method includes: using a controller neural network to generate a batch of output sequences, each output sequence in the batch is the architecture of a sub-neural network used to perform a specific neural network task; for each output sequence in the batch: through the respective The sub-neural network instance trains the architecture defined by the output sequence; evaluates the performance of the sub-neural network training instance on a specific neural network task to determine a measure of the performance of the sub-neural network training instance on the specific neural network task; and uses the sub-neural network A performance measure of the training instance to adjust the current values of the controller neural network controller parameters.

SUMMARY OF THE INVENTION

In view of the defects in the prior art, the purpose of the present invention is to provide a neural network architecture search method and system based on evolutionary computation.

According to a kind of neural network architecture search method based on evolutionary computation provided by the present invention, comprising:

Step S1: According to the target requirements and platform requirements, the target requirements are set through the target function, and the target requirements include: expected accuracy, inference delay, search space size and evolution times;

Step S2: According to the set size of the search space, randomly generate N undirected cyclic graphs based on the set of sub-network modules, as the network search space for evolutionary optimization;

Step S3: Under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the probability path selection mechanism, the ant colony algorithm is used to search for N directed acyclic graph optimization paths in the randomly generated N undirected acyclic graphs to form candidate set;

Step S4: Obtain the accuracy and inference delay of the N optimization paths in the candidate set through training and testing, and select the optimal result as the current optimal network structure;

Step S5: Evaluate whether the current network architecture meets the target requirements. When the current network architecture does not meet the preset target requirements, and when the speed and accuracy requirements are met, the optimization result is suspended, and the internal structure of all sub-network modules based on the current optimal network architecture Evolve mutation in real time, and continue to iterate until the preset target requirements are met; when the preset target requirements are met, the current optimal network result is output, otherwise, the search process is exited and the search exception is output.

Preferably, the step S1 includes:

The objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

Among them, the objective function is defined as a multi-objective search; Net represents the network obtained by the evolutionary algorithm; ACC(Net) represents the accuracy of the network; LAT(Net) represents the inference delay of the network; T represents the expected inference delay; A Indicates the expected accuracy rate; the expected accuracy rate is set according to the preset target requirements; the expected inference delay is set according to the mobile, embedded or general platform type.

Preferably, the step S2 includes:

The sub-network module is a node in an undirected cyclic graph; the sub-network module includes multiple types of sub-networks with M layers, and the sub-network type can expand the type of the sub-network module according to the target requirements and platform requirements and carry out choose;

The structure of the sub-network includes: multi-convolutional layers, ResNet blocks, depthwise separable convolution, inverted residual structure with linear bottleneck, and lightweight attention structure based on compression-excitation structure.

Preferably, the network search space includes: the network search space includes N search subspaces;

表示第i代子网络模块集合，

表示第i代中第j个子网络模块，

表示第i代搜索空间的边集合，子网络模块之间通过边连接；

表示第i代中第j个子网络模块和第k个子网络模块之间的边；

表示第i代第n个搜索空间；i表示迭代序号。in,

represents the set of i-th generation sub-network modules,

represents the jth subnetwork module in the ith generation,

Represents the edge set of the i-th generation search space, and the sub-network modules are connected by edges;

represents the edge between the jth subnetwork module and the kth subnetwork module in the ith generation;

Represents the n-th search space of the i-th generation; i represents the iteration number.

Preferably, the step S3 includes:

Step S3.1: select a point in each search subspace as the starting point, select the node farthest from the starting point as the ending point, and initialize the number of ants, the constant of pheromone intensity and the number of cycles;

Step S3.2: heuristic information calculation;

Among them, η _I,J (t) represents the heuristic information from node I to node J at time t; Dep _I,J , Wig _I,J , Con _I,J , Fil _I,J represent the depth of node J respectively , width, degree of connectivity and number of filters, ω represents the incentive factor, 0≤ω≤1; the reward mechanism is defined to motivate all nodes in the current optimal network architecture; the smaller the value of ω, the greater the heuristic information η, the initial Since there is no evolution in the network search space, the initial value of ω is set to 1;

Step S3.3: Probabilistic path selection;

表示第t时刻蚂蚁m由点I移动到点J的概率；allowed_m表示蚂蚁下一步可以选择的节点；α表示信息素启发因子，表示路径上残留的信息素在寻优过程中的作用，值越大则说明蚂蚁之间的协作能力越强，倾向于选择其它蚂蚁经过的路径；β表示期望值启发因子，表明准确性和推理时延在蚂蚁选择路径时的受重视程度，值越大，状态转移规律越接近贪心规则；τ_I，J(t)表示t时刻点I到点J路径上的信息素；τ_I,S(t)表示点I到allowed_m中任一点路径上的信息素；

表示I到allowed_m中任一点路径上的启发信息；in,

Represents the probability that the ant m moves from point I to point J at the t-th time; allowed _m represents the node that the ant can choose next; α represents the pheromone heuristic factor, which represents the role of the remaining pheromone on the path in the optimization process, value The larger the value, the stronger the cooperation ability between ants and the tendency to choose the path passed by other ants; β represents the expected value heuristic factor, which indicates the importance of accuracy and inference delay when the ants choose the path. The transition rule is closer to the greedy rule; τ _{I, J} (t) represents the pheromone on the path from point I to point J at time t; τ _{I, S} (t) represents the pheromone on the path from point I to any point in allowed _m ;

Indicates heuristic information on the path from I to allowed _m at any point;

Step S3.4: dynamic volatilization of pheromone;

η_i表示所有的启发信息，L表示当前网络下的结点总数；Among them, ρ _{I, J} (t) represents the volatility coefficient on the path from I to J at time t; η _{I, J} (t) represents the heuristic information on the path from I to J at time t;

η _i represents all the heuristic information, L represents the total number of nodes under the current network;

Step S3.5: pheromone increment calculation;

Among them, Q is the constant of pheromone intensity, which is the total amount of pheromone released by the ants on the path traversed in a cycle; η _m represents the total amount of enlightenment information received by the mth ant in this cycle;

Step S3.6: pheromone update;

τ _{I, J} (t+1)=(1-ρ)τ _{I, J} (t)+Δτ _{I, J} (t,t+1)

表示第m只蚂蚁在本次循环中在路径(I,J)上留下的信息素增量，Δτ_I,J(t,t+1)表示本次循环当中所有经过路径(I,J)的蚂蚁留下的信息素增量；K表示在本次循环中所有经过路径(I,J)的蚂蚁总数；where ρ is the dynamic volatility coefficient of pheromone;

Represents the pheromone increment left by the mth ant on the path (I, J) in this cycle, Δτ _{I, J} (t, t+1) represents all the paths (I, J) passed in this cycle The pheromone increment left by the ants; K represents the total number of ants passing through the path (I, J) in this cycle;

Step S3.7: Judgment of the end of optimization: when the optimization of all search subspaces reaches the maximum number of cycles, exit the cycle, and output the optimization results of all search subspaces as the contemporary candidate set; otherwise, repeat steps S3.2 to S3 .7 until the maximum number of cycles is reached.

Preferably, the evolution mutation in the step S5 includes: setting the excitation factor ω in all sub-network modules in the current optimal network architecture to be constant, and 0≤ω≤1; The internal structure of the sub-network module generates the next-generation sub-network module, and steps S2 to S5 are repeatedly executed until the preset target requirements are met.

A neural network architecture search system based on evolutionary computing provided by the present invention, comprising:

Module M1: According to the target requirements and platform requirements, the target requirements are set through the target function, and the target requirements include: expected accuracy, inference delay, search space size and evolution times;

Module M2: According to the set size of the search space, N undirected cyclic graphs are randomly generated based on the set of sub-network modules as the network search space for evolutionary optimization;

Module M3: Under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the probability path selection mechanism, the ant colony algorithm is used to search for N directed acyclic graphs in randomly generated N undirected acyclic graphs. candidate set;

Module M4: Obtain the accuracy and inference delay of N optimization paths in the candidate set through training and testing, and select the optimal result as the current optimal network structure;

Module M5: Evaluate whether the current network architecture meets the target requirements. When the current network architecture does not meet the preset target requirements, when the speed and accuracy requirements are met, the optimization result is suspended, and the internal structure of all sub-network modules based on the current optimal network architecture Evolve mutation in real time, and continue to iterate until the preset target requirements are met; when the preset target requirements are met, the current optimal network result is output, otherwise, the search process is exited and the search exception is output.

Preferably, the module M1 includes:

The objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

Among them, the objective function is defined as a multi-objective search; Net represents the network obtained by the evolutionary algorithm; ACC(Net) represents the accuracy of the network; LAT(Net) represents the inference delay of the network; T represents the expected inference delay; A Indicates the expected accuracy rate; the expected accuracy rate is set according to the preset target requirements; the expected inference delay is set according to the mobile, embedded or general platform type.

Preferably, the module M2 includes:

The sub-network module is a node in an undirected cyclic graph; the sub-network module includes multiple types of sub-networks with M layers, and the sub-network type can expand the type of the sub-network module according to the target requirements and platform requirements and carry out choose;

The structure of the sub-network includes: multiple convolutional layers, ResNet blocks, depthwise separable convolution, inverted residual structure with linear bottleneck, and lightweight attention structure based on compression-excitation structure;

The network search space includes: the network search space includes N search subspaces;

表示第i代子网络模块集合，

表示第i代中第j个子网络模块，

表示第i代搜索空间的边集合，子网络模块之间通过边连接；

表示第i代中第j个子网络模块和第k个子网络模块之间的边；

表示第i代第n个搜索空间；i表示迭代序号。in,

represents the set of i-th generation sub-network modules,

represents the jth subnetwork module in the ith generation,

Represents the edge set of the i-th generation search space, and the sub-network modules are connected by edges;

represents the edge between the jth subnetwork module and the kth subnetwork module in the ith generation;

Represents the n-th search space of the i-th generation; i represents the iteration number.

Preferably, the module M3 includes:

Module M3.1: Select a point in each search subspace as the starting point, select the node farthest from the starting point as the end point, and initialize the number of ants, the constant of pheromone intensity and the number of cycles;

Module M3.2: Heuristic information calculation;

Among them, η _I,J (t) represents the heuristic information from node I to node J at time t; Dep _I,J , Wig _I,J , Con _I,J , Fil _I,J represent the depth of node J respectively , width, degree of connectivity and number of filters, ω represents the incentive factor, 0≤ω≤1; the reward mechanism is defined to motivate all nodes in the current optimal network architecture; the smaller the value of ω, the greater the heuristic information η, the initial Since there is no evolution in the network search space, the initial value of ω is set to 1;

Module M3.3: Probabilistic Path Selection;

表示第t时刻蚂蚁m由点I移动到点J的概率；allowed_m表示蚂蚁下一步可以选择的节点；α表示信息素启发因子，表示路径上残留的信息素在寻优过程中的作用，值越大则说明蚂蚁之间的协作能力越强，倾向于选择其它蚂蚁经过的路径；β表示期望值启发因子，表明准确性和推理时延在蚂蚁选择路径时的受重视程度，值越大，状态转移规律越接近贪心规则；τ_I，J(t)表示t时刻点I到点J路径上的信息素；τ_I,S(t)表示点I到allowed_m中任一点路径上的信息素；

表示I到allowed_m中任一点路径上的启发信息；in,

Represents the probability that the ant m moves from point I to point J at the t-th time; allowed _m represents the node that the ant can choose next; α represents the pheromone heuristic factor, which represents the role of the remaining pheromone on the path in the optimization process, value The larger the value, the stronger the cooperation ability between ants and the tendency to choose the path passed by other ants; β represents the expected value heuristic factor, which indicates the importance of accuracy and inference delay when the ants choose the path. The transition rule is closer to the greedy rule; τ _{I, J} (t) represents the pheromone on the path from point I to point J at time t; τ _{I, S} (t) represents the pheromone on the path from point I to any point in allowed _m ;

Indicates heuristic information on the path from I to allowed _m at any point;

Module M3.4: dynamic volatilization of pheromone;

η_i表示所有的启发信息，L表示当前网络下的结点总数；Among them, ρ _{I, J} (t) represents the volatility coefficient on the path from I to J at time t; η _{I, J} (t) represents the heuristic information on the path from I to J at time t;

η _i represents all the heuristic information, L represents the total number of nodes under the current network;

Module M3.5: pheromone incremental calculation;

Among them, Q is the constant of pheromone intensity, which is the total amount of pheromone released by the ants on the path traversed in a cycle; η _m represents the total amount of enlightenment information received by the mth ant in this cycle;

Module M3.6: pheromone update;

τ _{I, J} (t+1)=(1-ρ)τ _{I, J} (t)+Δτ _{I, J} (t, t+1)

表示第m只蚂蚁在本次循环中在路径(I,J)上留下的信息素增量，Δτ_I，J(t，t+1)表示本次循环当中所有经过路径(I,J)的蚂蚁留下的信息素增量；K表示在本次循环中所有经过路径(I,J)的蚂蚁总数；where ρ is the dynamic volatility coefficient of pheromone;

Represents the pheromone increment left by the mth ant on the path (I, J) in this cycle, Δτ _{I, J} (t, t+1) represents all the paths (I, J) passed in this cycle The pheromone increment left by the ants; K represents the total number of ants passing through the path (I, J) in this cycle;

Module M3.7: Judgment of the end of optimization: when the optimization of all search subspaces reaches the maximum number of cycles, exit the loop, and output the optimization results of all search subspaces as the contemporary candidate set; otherwise, repeat triggering modules M3.2 to M3 .7 execute until the maximum number of cycles is reached;

The evolution mutation in the module M5 includes: setting the excitation factor ω in all sub-network modules in the current optimal network architecture to a constant, and 0≤ω≤1; at the same time randomly selecting mutation operations in the mutation set, and evolving the sub-network modules The internal structure generates the next-generation sub-network module, and repeatedly triggers the execution of modules M2 to M5 until the preset target requirements are met.

Compared with the prior art, the present invention has the following beneficial effects:

1. The present invention constructs a system-level search subspace with a sub-network module as the basic component through the construction idea of modular graph theory, which can effectively reduce the complexity of the search space, and speed up the architecture search process with a module-level search method, and improve the search performance. ;Using the heuristic optimization ability of the ant colony algorithm, it can prevent the architecture search from falling into a local optimum from a global perspective and lead to the degradation of network performance; Integrate the reward and mutation evolution mechanism, and evolve as comprehensively as possible through the structure-level method, which can avoid the architecture search. comprehensive.

2. The present invention can set the desired target through the objective function according to the actual needs of the application and the platform requirements, and is not limited by the application platform, and at the same time encourages the diversity of the structure of the modules of the entire network, and has certain application flexibility and scalability. properties, a neural network model that achieves a good trade-off between accuracy and speed can be obtained with limited resources.

Description of drawings

Other features, objects and advantages of the present invention will become more apparent by reading the detailed description of non-limiting embodiments with reference to the following drawings:

Figure 1 is a flowchart of a neural network architecture search method based on evolutionary computing.

Detailed ways

The present invention will be described in detail below with reference to specific embodiments. The following examples will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, for those skilled in the art, several changes and improvements can be made without departing from the inventive concept. These all belong to the protection scope of the present invention.

Example 1

According to a kind of neural network architecture search method based on evolutionary computation provided by the present invention, comprising:

Step S1: According to the target requirements and platform requirements, the target requirements are set through the target function, and the target requirements include: expected accuracy, inference delay, search space size and evolution times;

Step S2: According to the set size of the search space, randomly generate N undirected cyclic graphs based on the set of sub-network modules, as the network search space for evolutionary optimization;

Step S3: Under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the probability path selection mechanism, the ant colony algorithm is used to search for N directed acyclic graph optimization paths in the randomly generated N undirected acyclic graphs to form candidate set;

Step S4: Obtain the accuracy and inference delay of the N optimization paths in the candidate set through training and testing, and select the optimal result as the current optimal network structure;

Step S5: Evaluate whether the current network architecture meets the target requirements. When the current network architecture does not meet the preset target requirements, and when the speed and accuracy requirements are met, the optimization result is suspended, and the internal structure of all sub-network modules based on the current optimal network architecture Evolve mutation in real time, and continue to iterate until the preset target requirements are met; when the preset target requirements are met, the current optimal network result is output, otherwise, the search process is exited and the search exception is output.

Specifically, the step S1 includes:

The objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

Among them, the objective function is defined as a multi-objective search; Net represents the network obtained by the evolutionary algorithm; ACC(Net) represents the accuracy of the network; LAT(Net) represents the inference delay of the network; T represents the expected inference delay; A Indicates the expected accuracy rate; the expected accuracy rate is set according to the preset target requirements; the expected inference delay is set according to the mobile, embedded or general platform type.

Specifically, the step S2 includes:

The sub-network module is a node in an undirected cyclic graph; the sub-network module includes multiple types of sub-networks with M layers, and the sub-network type can expand the type of the sub-network module according to the target requirements and platform requirements and carry out choose;

The structure of the sub-network includes: multi-convolutional layers, ResNet blocks, depthwise separable convolution, inverted residual structure with linear bottleneck, and lightweight attention structure based on compression-excitation structure.

Specifically, the network search space includes: the network search space includes N search subspaces;

表示第i代子网络模块集合，

表示第i代中第j个子网络模块，

表示第i代搜索空间的边集合，子网络模块之间通过边连接；

表示第i代中第j个子网络模块和第k个子网络模块之间的边；

表示第i代第n个搜索空间；i表示迭代序号。in,

represents the set of i-th generation sub-network modules,

represents the jth subnetwork module in the ith generation,

Represents the edge set of the i-th generation search space, and the sub-network modules are connected by edges;

represents the edge between the jth subnetwork module and the kth subnetwork module in the ith generation;

Represents the n-th search space of the i-th generation; i represents the iteration number.

Specifically, the step S3 includes:

Step S3.1: select a point in each search subspace as the starting point, select the node farthest from the starting point as the ending point, and initialize the number of ants, the constant of pheromone intensity and the number of cycles;

Step S3.2: heuristic information calculation;

Among them, η _I,J (t) represents the heuristic information from node I to node J at time t; Dep _I,J , Wig _I,J , Con _I,J , Fil _I,J represent the depth of node J respectively , width, degree of connectivity and number of filters, ω represents the incentive factor, 0≤ω≤1; the reward mechanism is defined to motivate all nodes in the current optimal network architecture; the smaller the value of ω, the greater the heuristic information η, the initial Since there is no evolution in the network search space, the initial value of ω is set to 1;

Step S3.3: Probabilistic path selection;

表示第t时刻蚂蚁m由点I移动到点J的概率；allowed_m表示蚂蚁下一步可以选择的节点；α表示信息素启发因子，表示路径上残留的信息素在寻优过程中的作用，值越大则说明蚂蚁之间的协作能力越强，倾向于选择其它蚂蚁经过的路径；β表示期望值启发因子，表明准确性和推理时延在蚂蚁选择路径时的受重视程度，值越大，状态转移规律越接近贪心规则；τ_I，J(t)表示t时刻点I到点J路径上的信息素；τ_I,S(t)表示点I到allowed_m中任一点路径上的信息素；

表示I到allowed_m中任一点路径上的启发信息；in,

Represents the probability that the ant m moves from point I to point J at the t-th time; allowed _m represents the node that the ant can choose next; α represents the pheromone heuristic factor, which represents the role of the remaining pheromone on the path in the optimization process, value The larger the value, the stronger the cooperation ability between ants and the tendency to choose the path passed by other ants; β represents the expected value heuristic factor, which indicates the importance of accuracy and inference delay when the ants choose the path. The transition rule is closer to the greedy rule; τ _{I, J} (t) represents the pheromone on the path from point I to point J at time t; τ _{I, S} (t) represents the pheromone on the path from point I to any point in allowed _m ;

Indicates heuristic information on the path from I to allowed _m at any point;

Step S3.4: dynamic volatilization of pheromone;

η_i表示所有的启发信息，L表示当前网络下的结点总数；Among them, ρ _{I, J} (t) represents the volatility coefficient on the path from I to J at time t; η _{I, J} (t) represents the heuristic information on the path from I to J at time t;

η _i represents all the heuristic information, L represents the total number of nodes under the current network;

Step S3.5: pheromone increment calculation;

Among them, Q is the constant of pheromone intensity, which is the total amount of pheromone released by the ants on the path traversed in a cycle; η _m represents the total amount of enlightenment information received by the mth ant in this cycle;

Step S3.6: pheromone update;

τ _{I, J} (t+1)=(1-ρ)τ _{I, J} (t)+Δτ _{I, J} (t, t+1)

表示第m只蚂蚁在本次循环中在路径(I,J)上留下的信息素增量，Δτ_I，J(t，t+1)表示本次循环当中所有经过路径(I,J)的蚂蚁留下的信息素增量；K表示在本次循环中所有经过路径(I,J)的蚂蚁总数；where ρ is the dynamic volatility coefficient of pheromone;

Represents the pheromone increment left by the mth ant on the path (I, J) in this cycle, Δτ _{I, J} (t, t+1) represents all the paths (I, J) passed in this cycle The pheromone increment left by the ants; K represents the total number of ants passing through the path (I, J) in this cycle;

Step S3.7: Judgment of the end of optimization: when the optimization of all search subspaces reaches the maximum number of cycles, exit the cycle, and output the optimization results of all search subspaces as the contemporary candidate set; otherwise, repeat steps S3.2 to S3 .7 until the maximum number of cycles is reached.

Specifically, the evolution mutation in step S5 includes: setting the incentive factor ω in all sub-network modules in the current optimal network architecture to be constant, and 0≤ω≤1; The internal structure of the sub-network module generates the next-generation sub-network module, and steps S2 to S5 are repeatedly executed until the preset target requirements are met.

A neural network architecture search system based on evolutionary computing provided by the present invention, comprising:

Module M1: According to the target requirements and platform requirements, the target requirements are set through the target function, and the target requirements include: expected accuracy, inference delay, search space size and evolution times;

Module M2: According to the set size of the search space, N undirected cyclic graphs are randomly generated based on the set of sub-network modules as the network search space for evolutionary optimization;

Module M3: Under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the probability path selection mechanism, the ant colony algorithm is used to search for N directed acyclic graphs in randomly generated N undirected acyclic graphs. candidate set;

Module M4: Obtain the accuracy and inference delay of N optimization paths in the candidate set through training and testing, and select the optimal result as the current optimal network structure;

Module M5: Evaluate whether the current network architecture meets the target requirements. When the current network architecture does not meet the preset target requirements, when the speed and accuracy requirements are met, the optimization result is suspended, and the internal structure of all sub-network modules based on the current optimal network architecture Evolve mutation in real time, and continue to iterate until the preset target requirements are met; when the preset target requirements are met, the current optimal network result is output, otherwise, the search process is exited and the search exception is output.

Specifically, the module M1 includes:

The objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

Among them, the objective function is defined as a multi-objective search; Net represents the network obtained by the evolutionary algorithm; ACC(Net) represents the accuracy of the network; LAT(Net) represents the inference delay of the network; T represents the expected inference delay; A Indicates the expected accuracy rate; the expected accuracy rate is set according to the preset target requirements; the expected inference delay is set according to the mobile, embedded or general platform type.

Specifically, the module M2 includes:

The sub-network module is a node in an undirected cyclic graph; the sub-network module includes multiple types of sub-networks with M layers, and the sub-network type can expand the type of the sub-network module according to the target requirements and platform requirements and carry out choose;

The structure of the sub-network includes: multiple convolutional layers, ResNet blocks, depthwise separable convolution, inverted residual structure with linear bottleneck, and lightweight attention structure based on compression-excitation structure;

The network search space includes: the network search space includes N search subspaces;

表示第i代子网络模块集合，

表示第i代中第j个子网络模块，

表示第i代搜索空间的边集合，子网络模块之间通过边连接；

表示第i代中第j个子网络模块和第k个子网络模块之间的边；

表示第i代第n个搜索空间；i表示迭代序号。in,

represents the set of i-th generation sub-network modules,

represents the jth subnetwork module in the ith generation,

Represents the edge set of the i-th generation search space, and the sub-network modules are connected by edges;

represents the edge between the jth subnetwork module and the kth subnetwork module in the ith generation;

Represents the n-th search space of the i-th generation; i represents the iteration number.

Specifically, the module M3 includes:

Module M3.1: Select a point in each search subspace as the starting point, select the node farthest from the starting point as the end point, and initialize the number of ants, the constant of pheromone intensity and the number of cycles;

Module M3.2: Heuristic information calculation;

Among them, η _I,J (t) represents the heuristic information from node I to node J at time t; Dep _I,J , Wig _I,J , Con _I,J , Fil _I,J represent the depth of node J respectively , width, degree of connectivity and number of filters, ω represents the incentive factor, 0≤ω≤1; the reward mechanism is defined to motivate all nodes in the current optimal network architecture; the smaller the value of ω, the greater the heuristic information η, the initial Since there is no evolution in the network search space, the initial value of ω is set to 1;

Module M3.3: Probabilistic Path Selection;

表示第t时刻蚂蚁m由点I移动到点J的概率；allowed_m表示蚂蚁下一步可以选择的节点；α表示信息素启发因子，表示路径上残留的信息素在寻优过程中的作用，值越大则说明蚂蚁之间的协作能力越强，倾向于选择其它蚂蚁经过的路径；β表示期望值启发因子，表明准确性和推理时延在蚂蚁选择路径时的受重视程度，值越大，状态转移规律越接近贪心规则；τ_I，J(t)表示t时刻点I到点J路径上的信息素；τ_I,S(t)表示点I到allowed_m中任一点路径上的信息素；

表示I到allowed_m中任一点路径上的启发信息；in,

Represents the probability that the ant m moves from point I to point J at the t-th time; allowed _m represents the node that the ant can choose next; α represents the pheromone heuristic factor, which represents the role of the remaining pheromone on the path in the optimization process, value The larger the value, the stronger the cooperation ability between ants and the tendency to choose the path passed by other ants; β represents the expected value heuristic factor, which indicates the importance of accuracy and inference delay when the ants choose the path. The transition rule is closer to the greedy rule; τ _{I, J} (t) represents the pheromone on the path from point I to point J at time t; τ _{I, S} (t) represents the pheromone on the path from point I to any point in allowed _m ;

Indicates heuristic information on the path from I to allowed _m at any point;

Module M3.4: dynamic volatilization of pheromone;

η_i表示所有的启发信息，L表示当前网络下的结点总数；Among them, ρ _{I, J} (t) represents the volatility coefficient on the path from I to J at time t; η _{I, J} (t) represents the heuristic information on the path from I to J at time t;

η _i represents all the heuristic information, L represents the total number of nodes under the current network;

Module M3.5: pheromone incremental calculation;

Among them, Q is the constant of pheromone intensity, which is the total amount of pheromone released by the ants on the path traversed in a cycle; η _m represents the total amount of enlightenment information received by the mth ant in this cycle;

Module M3.6: pheromone update;

τ _{I, J} (t+1)=(1-ρ)τ _{I, J} (t)+Δτ _{I, J} (t,t+1)

表示第m只蚂蚁在本次循环中在路径(I,J)上留下的信息素增量，Δτ_I,J(t,t+1)表示本次循环当中所有经过路径(I,J)的蚂蚁留下的信息素增量；K表示在本次循环中所有经过路径(I,J)的蚂蚁总数；where ρ is the dynamic volatility coefficient of pheromone;

Represents the pheromone increment left by the mth ant on the path (I, J) in this cycle, Δτ _{I, J} (t, t+1) represents all the paths (I, J) passed in this cycle The pheromone increment left by the ants; K represents the total number of ants passing through the path (I, J) in this cycle;

Module M3.7: Judgment of the end of optimization: when the optimization of all search subspaces reaches the maximum number of cycles, exit the loop, and output the optimization results of all search subspaces as the contemporary candidate set; otherwise, repeat triggering modules M3.2 to M3 .7 execute until the maximum number of cycles is reached;

The evolution mutation in the module M5 includes: setting the excitation factor ω in all sub-network modules in the current optimal network architecture to a constant, and 0≤ω≤1; at the same time randomly selecting mutation operations in the mutation set, and evolving the sub-network modules The internal structure generates the next-generation sub-network module, and repeatedly triggers the execution of modules M2 to M5 until the preset target requirements are met.

Example 2

Example 2 is a modification of Example 1

The present invention proposes a neural network architecture search method and system based on evolutionary computation. The modular graph theory is used to construct the idea. On the basis of the neural network initialization module, the optimization ability and the reward evolution mechanism of the ant colony algorithm are used to search at the module level. Combining the structure-level evolution method to explore the optimal neural network architecture can solve the above technical problems by taking into account the speed and accuracy under the condition of limited resources.

Aiming at the problem that performance and efficiency are difficult to balance in the current neural network architecture search under the condition of limited resources, the purpose of the present invention is to provide a neural network architecture search method and system based on evolutionary computing. The invention takes the sub-network module as the basic component, randomly generates multiple network search subspaces (directed acyclic graph), and uses the depth, breadth, connection degree and number of filters of the sub-network module (node) as the sub-network module The complexity of the algorithm is combined with the incentive factor as the heuristic information. Through the dynamic volatilization of pheromone and the random probability path selection mechanism, the improved ant colony algorithm is used to search for optimization in the subspace, and at the same time, the reward and evolutionary mutation mechanism are continuously evolved to search for both performance and performance. Optimal neural network architecture for efficiency.

The evolutionary computation In the field of computer science, evolutionary computation (Evolutionary Computation) is a subfield of intelligent computation (Computational Intelligence) involving combinatorial optimization problems. Its evolutionary algorithm is influenced by the natural selection mechanism of "survival of the fittest" and the transmission law of genetic information in the process of biological evolution. This process is simulated iteratively through the program, and the problem to be solved is regarded as the environment. , the optimal solution is sought through natural evolution.

The evolutionary algorithm or "Evolutionary Algorithms" is an "algorithm cluster", although it has many variations, there are different ways of genetic gene expression, different crossover and mutation operators, references to special operators, and different methods of regeneration and selection, but they are all inspired by nature's biological evolution. Compared with the traditional calculus-based method and the exhaustive method and other optimization algorithms, the evolutionary algorithm is a mature global optimization method with high robustness and wide applicability, and has the characteristics of self-organization, self-adaptation and self-learning. , which is not limited by the nature of the problem, and can effectively deal with complex problems that are difficult to solve by traditional optimization algorithms.

The Neural Architecture Search (NAS) is one of the hotspots in deep learning research. NAS aims to design neural network architectures with optimal performance in an automated manner with minimal human intervention by using limited computing resources.

The above is one of the swarm intelligence algorithms in bionics. It was proposed in 1991 by Italian scholar M. Dorigo and others after being inspired by the foraging behavior of ants in the real world. During the ant foraging process, each ant releases a chemical substance called pheromone that can communicate with other ants on the path it walks. Over time, the pheromone will volatilize, but the shorter the path ants crawl, the slower the corresponding pheromone volatilization rate, and the concentration of the pheromone left on the path will be relatively higher. Ants can find this pheromone and sense its concentration, and can choose the path with the highest pheromone concentration with a higher probability. In this way, more ants will choose the path with high concentration of pheromone, and more and more ants will be attracted to this path, forming a positive feedback. Based on this principle, ants can quickly find the shortest path from the food source.

The invention relates to a neural network architecture search method and system based on evolutionary computation, including the steps of search target setting, search space initialization, ant colony optimization, target evaluation, evolution mutation and the like. First, according to the target requirements and platform requirements, set the search target parameters in terms of accuracy, inference delay, and evolution times; then initialize the search space, and randomly generate N search subspaces; Candidate set; then complete the target evaluation on the data set, and select the contemporary optimal network architecture from the candidate set; finally, implement evolution mutation for the internal structure of all sub-network modules of the contemporary optimal network architecture, and continue to iterate until the target requirements are met.

The search target setting, that is, according to the target requirements and platform requirements, the expected accuracy and inference delay are set through the defined target function, and parameters such as the size of the search space and the number of evolutions are set.

The search space initialization, that is, according to the set size of the search space, randomly generates N undirected cyclic graphs for the sub-network module set as the network search space for evolutionary optimization.

Specifically, the network search space is a hierarchical block undirected cyclic graph, each sub-network module represents a node in the undirected cyclic graph, allowing module-level search and module internal structure-level mutation and search .

Specifically, the sub-network module is a node in an undirected cyclic graph, which is defined as multiple types of sub-networks with M layers. The sub-network type can be expanded and selected according to target requirements and application platforms. Structures include but are not limited to multiple convolutional layers, ResNet blocks, depthwise separable convolutions, inverted residual structures with linear bottlenecks, lightweight attention structures based on compression-excitation structures, etc.

The ant colony optimization, that is, under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the probability path selection mechanism, through the ant colony algorithm to search N directed acyclic graphs in randomly generated N undirected acyclic graphs Find the optimal path to form the candidate set of this generation.

The target evaluation is to obtain the accuracy and inference delay of the N optimization paths in the candidate set through training and testing, select the optimal result as the optimal network architecture of this generation, and evaluate whether the target requirements are met.

The evolution mutation is to reward the internal structure of all sub-network modules in the optimal network architecture of this generation, and randomly select mutation operations in the mutation set to generate the next generation of sub-network modules, so as to simulate excellent individuals in nature. The process of leaving offspring is easier.

Specifically, the mutation set includes the operations of keeping the module structure unchanged, randomly selecting the convolution type, convolution kernel size, filter size, inserting a convolution layer, deleting a convolution layer, adding a connection, and deleting a connection.

A neural network architecture search method and system based on evolutionary computing, including the following steps: as shown in Figure 1,

Step 1, search target setting:

According to the target requirements and platform requirements, the expected accuracy and inference delay are set through the defined objective function, and parameters such as the size of the search space and the number of evolutions are also set. In order to seek a balance between accuracy and inference latency, the objective function is defined as a multi-objective search problem to find a neural network architecture with high accuracy and low thrust latency. The accuracy rate can be set according to the specific needs of the user, using TOP-1 or TOP-5; the inference delay can be set according to the type of mobile, embedded or general platform. The overall objective function is specifically formalized as follows:

s.t. LAT(Net)≤T&ACC(Net)≥A

Among them, Net represents the network obtained by the evolutionary algorithm, ACC(Net) represents the accuracy of the network, LAT(Net) represents the inference delay of the network, and T represents the desired target delay. A represents the expected accuracy. s.t.LAT(Net) representation

Step 2, search space initialization:

According to the set size of the search space, N undirected cyclic graphs are randomly generated from the set of sub-network modules as the network search space for evolutionary optimization.

形式化为:Each sub-network module represents a node in an undirected cyclic graph, and is defined as multiple types of sub-networks with M layers. The sub-network module type can be expanded and selected according to the target requirements and application platform requirements. Structures include but are not limited to multiple convolutional layers (1x1 conv + 2D conv filtered to 3*3 + 1x1 conv), ResNet blocks, depthwise separable convolutions, inverted residual structures with linear bottlenecks, compression-excitation structures based light Magnitude attention structure, etc. Edges represent connections between subnetwork modules. A search space consists of N search subspaces, namely

Formalized as:

表示第i代子网络模块集合，

表示第i代中第j个子网络模块，

表示第i代搜索空间的边集合，

表示第i代中第j个子网络模块和第k个子网络模块之间的边；

表示第i代第n个搜索空间；i表示迭代序号。in,

represents the set of i-th generation sub-network modules,

represents the jth subnetwork module in the ith generation,

represents the edge set of the ith generation search space,

represents the edge between the jth subnetwork module and the kth subnetwork module in the ith generation;

Represents the n-th search space of the i-th generation; i represents the iteration number.

Step 3, ant colony optimization:

Under the guidance of heuristic information, combined with the dynamic volatilization of pheromone and the random probability path selection mechanism, the ant colony algorithm is used to search for N directed and acyclic optimization paths in the randomly generated N undirected cyclic graphs, thus forming candidate set for this generation.

Step 3.1: Parameter initialization. Select a point in each search subspace as the starting point, automatically select the node farthest from the starting point as the end point, initialize the number of ants, pheromone intensity constant, cycle times, etc.

Step 3.2: Heuristic information calculation. The heuristic function has an important influence on the convergence and stability of the optimization process. In the ant colony algorithm, the heuristic information η is usually the inverse ratio of the distance between two points, but the standard of neural network architecture search balances accuracy and inference delay, and the optimization process should not only consider the internal sub-network module that affects the amount of calculation. The complexity of the structure, such as depth, width, connectivity, and number of filters, also needs to consider the incentives given by the reward mechanism, so the heuristic information is defined as

Among them, η _I,J (t) represents the heuristic information from node I to node J at time t; Dep _I,J , Wig _I,J , Con _I,J , Fil _I,J represent the depth of node J respectively , width, degree of connectivity and number of filters, ω represents the incentive factor, 0≤ω≤1, and the reward mechanism is defined to motivate all nodes in the optimal network architecture of this generation. The smaller the value of ω is, the greater the heuristic information η is. Since there is no evolution in the initial network search space, the initial value of ω is set to 1.

决定下一步往哪条路上移动。Step 3.3: Probabilistic path selection: In each step of the path selection in the subspace search process, affected by the pheromone heuristic factor and the expected value heuristic factor, the ant m follows the probability

Decide which way to move next.

表示第t时刻蚂蚁m由点I移动到点J的概率，allowed_m表示蚂蚁下一步可以选择的节点；信息素启发因子α，表示路径上残留的信息素在寻优过程中的作用，值越大则说明蚂蚁之间的协作能力越强，倾向于选择其它蚂蚁经过的路径；期望值启发因子β，表明准确性和推理时延在蚂蚁选择路径时的受重视程度，值越大，状态转移规律越接近贪心规则。τ_I，J(t)表示t时刻点I到点J路径上的信息素。τ_I,S(t)表示点I到allowed_m中任一点路径上的信息素。

Represents the probability that ant m moves from point I to point J at the t-th time, and allowed _m represents the node that the ant can choose next; The larger the value is, the stronger the cooperation ability between ants is, and they tend to choose the path passed by other ants; the expected value heuristic factor β indicates the importance of accuracy and inference delay when the ants choose the path. The larger the value is, the state transition law The closer to the greedy rule. τ _I,J (t) represents the pheromone on the path from point I to point J at time t. τ _I,S (t) represents the pheromone on the path from point I to any point in allowed _m .

Step 3.4: Dynamic volatilization of pheromone: The volatilization rate of pheromone in nature is a dynamic process, which will be changed over time by environmental factors such as temperature and humidity. In the optimization process of the neural network architecture, the more complex the structure of the node, the smaller the heuristic information; the node with a larger inference delay, the faster the evaporation speed, the less pheromone residue, so after all the ants reach the end, the information The prime dynamic volatilization coefficient ρ is calculated according to the following relationship,

η_i表示所有的启发信息，L表示该网络下的结点总数。ρ _{I, J} (t) represents the volatility coefficient on the path from I to J at time t, η _{I, J} (t) represents the heuristic information on the path from I to J at time t,

η _i represents all the heuristic information, L represents the total number of nodes under the network.

Step 3.5: Pheromone Incremental Calculation: The pheromone update model is an important part of the random search and fast convergence of the basic ant colony algorithm. According to the global optimal requirements, the ant-week model is modified and formalized as follows:

Q is the constant of pheromone intensity, which is the total amount of pheromone released by ants on the path passed in a cycle, which affects the convergence speed of the algorithm to a certain extent. η _m represents the total amount of heuristic information received by the mth ant in this cycle.

Step 3.6: Pheromone update: The amount of pheromone on each path is equal at the initial moment. After the ant completes a cycle, the pheromone will gradually volatilize over time, so the pheromone needs to be updated before the ant enters the next cycle. Concentration, formalized as follows:

τ _{I, J} (t+1)=(1-ρ)τ _{I, J} (t)+Δτ _{I, J} (t, t+1)

表示第m只蚂蚁在本次循环中在路径(I,J)上留下的信息素量，Δτ_I，J(t，t+1)表示本次循环当中所有经过路径(I,J)的蚂蚁留下的信息素增量。K表示在本次循环中所有经过路径(I,J)的蚂蚁总数。ρ is the pheromone dynamic volatility coefficient,

Represents the amount of pheromone left by the mth ant on the path (I, J) in this cycle, Δτ _{I, J} (t, t+1) represents all the pheromone passing through the path (I, J) in this cycle Pheromone increments left by ants. K represents the total number of ants passing through the path (I, J) in this cycle.

Step 3.7: Judgment of the end of optimization: when the optimization of all search subspaces reaches the maximum number of cycles, exit the cycle, output all search subspace optimization results as the contemporary candidate set, and enter step 4; otherwise, enter heuristic information calculation 3.2, continue cycle.

Step 4: Target evaluation. Obtain the accuracy and inference delay of N optimization paths in the candidate set through training and testing, and select the optimal result as the optimal network architecture of this generation. If required, then temporarily store the optimization results and go to step 5, otherwise discard the search subspace; if the maximum number of evolutions is reached and the target requirements are met, the results of the optimal network architecture of each generation in all evolutionary processes will be sorted and go to step 6 , otherwise exit the search process and output a search exception.

Step 5: Evolutionary mutation. Give rewards to all sub-network modules in the optimal network architecture of this generation, that is, set the incentive factor ω to a constant, 0≤ω≤1, the smaller the ω value, the greater the incentive effect. At the same time, the mutation operation is randomly selected in the mutation set, the internal structure of the sub-network module is evolved, the next-generation sub-network module is generated, and the process returns to step 2.

Step 6: Output the final result. Output the optimal neural network architecture.

Those skilled in the art know that, in addition to implementing the system, device and each module provided by the present invention in the form of pure computer readable program code, the system, device and each module provided by the present invention can be completely implemented by logically programming the method steps. The same program is implemented in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, and embedded microcontrollers, among others. Therefore, the system, device and each module provided by the present invention can be regarded as a kind of hardware component, and the modules included in it for realizing various programs can also be regarded as the structure in the hardware component; A module for realizing various functions can be regarded as either a software program for realizing a method or a structure within a hardware component.

Specific embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the above-mentioned specific embodiments, and those skilled in the art can make various changes or modifications within the scope of the claims, which do not affect the essential content of the present invention. The embodiments of the present application and features in the embodiments may be combined with each other arbitrarily, provided that there is no conflict.

Claims (10)

1. A neural network architecture searching method based on evolutionary computing is characterized by comprising the following steps:

step S1: setting target requirements through a target function according to the target requirements and the platform requirements, wherein the target requirements comprise: expected accuracy, reasoning time delay, search space size and evolution times;

step S2: according to the set size of the search space, randomly generating N directed cyclic graphs based on the sub-network module set to serve as a network search space for evolutionary optimization;

step S3: under the guidance of heuristic information, combining with a pheromone dynamic volatilization and probability path selection mechanism, searching N directed acyclic graphs for optimizing paths in N randomly generated acyclic graphs through an ant colony algorithm to form a candidate set;

step S4: acquiring the accuracy and reasoning time delay of N optimizing paths in the candidate set through training and testing, and selecting an optimal result as a current optimal network structure;

step S5: evaluating whether the current network architecture meets the target requirement, when the current network architecture does not meet the preset target requirement, and when the current network architecture meets the speed and precision requirements, pausing the optimization result, carrying out real-time evolution mutation on the internal structures of all the sub-network modules based on the current optimal network architecture, and continuing iteration until the preset target requirement is met; and when the preset target requirement is met, outputting the current optimal network result, otherwise, quitting the searching process and outputting the abnormal searching.

2. The evolutionary computing-based neural network architecture search method of claim 1, wherein the step S1 comprises:

the objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

wherein, the target function is defined as a multi-target search; net represents the network obtained by the evolutionary algorithm; acc (net) indicates the accuracy of the network; lat (net) represents the inference delay of the network; t represents the expected inference time delay; a represents the desired accuracy; the expected accuracy is set according to preset target requirements; the expected inference time delay is set according to mobile, embedded or general platform types.

3. The evolutionary computing-based neural network architecture search method of claim 1, wherein the step S2 comprises:

the sub-network modules are nodes in an undirected cyclic graph; the sub-network modules comprise a plurality of types of sub-networks with M layers, and the types of the sub-networks can be expanded and selected according to target requirements and platform requirements;

the structure of the sub-network comprises: multi-convolution layers, ResNet blocks, depth separable convolutions, reversed residual structure with linear bottlenecks, and lightweight attention structure based on compression-excitation structures.

4. The evolutionary computing-based neural network architecture search method of claim 3, wherein the network search space comprises: the network search space comprises N search subspaces;

wherein ,

representing the i-th generation of a collection of sub-network modules,

represents the jth sub-network module in the ith generation,

representing an edge set of the ith generation of search space, and connecting the sub-network modules through edges;

representing an edge between a jth sub-network module and a kth sub-network module in an ith generation;

representing the ith generation nth search space; i denotes an iteration number.

5. The evolutionary computing-based neural network architecture search method of claim 1, wherein the step S3 comprises:

step S3.1: selecting any point in each search subspace as a starting point, selecting a node farthest from the starting point as an end point, and initializing the ant number, the pheromone intensity constant and the cycle number;

step S3.2: calculating heuristic information;

wherein ,η_I，J(t) heuristic information from node I to node J at time t; dep_I，J、Wig_I，J、Con_I，J、Fil_I，JRespectively representing the depth, the width, the connectivity and the number of filters of the joint J, wherein omega represents an excitation factor, and omega is more than or equal to 0 and less than or equal to 1; prize-giving deviceThe excitation mechanism is defined to excite all nodes in the current optimal network architecture; the smaller the omega value is, the larger heuristic information eta is, and the initial value of omega is set to be 1 because evolution is not generated in the initial network search space;

step S3.3: selecting a probability path;

wherein ,

representing the probability that the ant m moves from the point I to the point J at the t-th moment; allowed_mRepresenting nodes which can be selected by ants in the next step; alpha represents pheromone elicitation factors, represents the effect of residual pheromones on the paths in the optimizing process, and the larger the value is, the stronger the cooperation capability among ants is, and the paths passed by other ants are prone to be selected; beta represents an expected value heuristic factor, which shows the accuracy and the degree of importance of reasoning time delay when ants select paths, and the larger the value is, the closer the state transition rule is to the greedy rule; tau is_I，J(t) pheromones on the path from point I to point J at time t; tau is_I，S(t) indicates points I to allowed_mPheromone on any point path;

representing I to allowed_mHeuristic information on the path of any point in the tree;

step S3.4: the pheromone is volatilized dynamically;

wherein ,ρ_I，J(t) represents the volatility coefficient on the path from I to J at the moment t; eta_I，J(t) heuristic information on the path from time I to J at t;

representing all initiation information, wherein L represents the total number of nodes in the current network;

step S3.5: performing pheromone increment calculation;

wherein Q is a pheromone strength constant, which is the total amount of pheromones released by ants on a path traveled in one cycle; eta_mRepresenting the total amount of heuristic information suffered by the mth ant in the cycle;

step S3.6: updating pheromone;

τ_I，J(t+1)＝(1-ρ)τ_I，J(t)+Δτ_I，J(t，t+1)

wherein rho is the pheromone dynamic volatility coefficient;

represents the pheromone increment left by the mth ant on the path (I, J) in the current cycle, delta tau_I，J(t, t +1) represents pheromone increment left by all ants passing through the path (I, J) in the current cycle; k represents the total number of ants passing through the paths (I, J) in the current cycle;

step S3.7: and (4) optimizing and judging: when the optimization of all the search subspaces reaches the maximum cycle times, the circulation is exited, and the optimization results of all the search subspaces are output as a current candidate set; otherwise, step S3.2 to step S3.7 are repeated until the maximum number of cycles is reached.

6. The evolutionary computing-based neural network architecture searching method of claim 1, wherein the evolving mutations in step S5 comprise: setting the excitation factors omega in all the sub-network modules in the current optimal network architecture as constants, wherein omega is more than or equal to 0 and less than or equal to 1; and simultaneously, randomly selecting mutation operation in the mutation set, promoting the internal structure of the sub-network module, generating a next generation sub-network module, and repeatedly executing the steps S2 to S5 until the preset target requirement is met.

7. An evolutionary computing-based neural network architecture search system, comprising:

module M1: setting target requirements through a target function according to the target requirements and the platform requirements, wherein the target requirements comprise: expected accuracy, reasoning time delay, search space size and evolution times;

module M2: according to the set size of the search space, randomly generating N directed cyclic graphs based on the sub-network module set to serve as a network search space for evolutionary optimization;

module M3: under the guidance of heuristic information, combining with a pheromone dynamic volatilization and probability path selection mechanism, searching N directed acyclic graphs for optimizing paths in N randomly generated acyclic graphs through an ant colony algorithm to form a candidate set;

module M4: acquiring the accuracy and reasoning time delay of N optimizing paths in the candidate set through training and testing, and selecting an optimal result as a current optimal network structure;

module M5: evaluating whether the current network architecture meets the target requirement, when the current network architecture does not meet the preset target requirement, and when the current network architecture meets the speed and precision requirements, pausing the optimization result, carrying out real-time evolution mutation on the internal structures of all the sub-network modules based on the current optimal network architecture, and continuing iteration until the preset target requirement is met; and when the preset target requirement is met, outputting the current optimal network result, otherwise, quitting the searching process and outputting the abnormal searching.

8. The evolutionary computing-based neural network architecture search system of claim 7, wherein the module M1 comprises:

the objective function formula is as follows:

s.t.LAT(Net)≤T&ACC(Net)≥A

wherein, the target function is defined as a multi-target search; net represents the network obtained by the evolutionary algorithm; acc (net) indicates the accuracy of the network; lat (net) represents the inference delay of the network; t represents the expected inference time delay; a represents the desired accuracy; the expected accuracy is set according to preset target requirements; the expected inference time delay is set according to mobile, embedded or general platform types.

9. The evolutionary computing-based neural network architecture search system of claim 7, wherein the module M2 comprises:

the sub-network modules are nodes in an undirected cyclic graph; the sub-network modules comprise a plurality of types of sub-networks with M layers, and the types of the sub-networks can be expanded and selected according to target requirements and platform requirements;

the structure of the sub-network comprises: the system comprises a multi-convolution layer, a ResNet block, a depth separable convolution, an inverted residual error structure with a linear bottleneck and a lightweight attention structure based on a compression-excitation structure;

the network search space includes: the network search space comprises N search subspaces;

wherein ,

representing an ith generation subnetworkA set of modules to be used in the method,

represents the jth sub-network module in the ith generation,

representing an edge set of the ith generation of search space, and connecting the sub-network modules through edges;

representing an edge between a jth sub-network module and a kth sub-network module in an ith generation;

representing the ith generation nth search space; i denotes an iteration number.

10. The evolutionary computing-based neural network architecture search system of claim 7, wherein the module M3 comprises:

module M3.1: selecting any point in each search subspace as a starting point, selecting a node farthest from the starting point as an end point, and initializing the ant number, the pheromone intensity constant and the cycle number;

module M3.2: calculating heuristic information;

wherein ,η_I，J(t) heuristic information from node I to node J at time t; dep_I，J、Wig_I，J、Con_I，J、Fil_I，JRespectively representing the depth, the width, the connectivity and the number of filters of the joint J, wherein omega represents an excitation factor, and omega is more than or equal to 0 and less than or equal to 1; the reward mechanism is defined as exciting all nodes in the current optimal network architecture; the smaller the omega value is, the larger heuristic information eta is, and the initial value of omega is set to be 1 because evolution is not generated in the initial network search space;

module M3.3: selecting a probability path;

wherein ,

representing the probability that the ant m moves from the point I to the point J at the t-th moment; allowed_mRepresenting nodes which can be selected by ants in the next step; alpha represents pheromone elicitation factors, represents the effect of residual pheromones on the paths in the optimizing process, and the larger the value is, the stronger the cooperation capability among ants is, and the paths passed by other ants are prone to be selected; beta represents an expected value heuristic factor, which shows the accuracy and the degree of importance of reasoning time delay when ants select paths, and the larger the value is, the closer the state transition rule is to the greedy rule; tau is_I，J(t) pheromones on the path from point I to point J at time t; tau is_I，S(t) indicates points I to allowed_mPheromone on any point path;

representing I to allowed_mHeuristic information on the path of any point in the tree;

module M3.4: the pheromone is volatilized dynamically;

wherein ,ρ_I，J(t) represents the volatility coefficient on the path from I to J at the moment t; eta_I，J(t) heuristic information on the path from time I to J at t;

representing all initiation information, wherein L represents the total number of nodes in the current network;

module M3.5: performing pheromone increment calculation;

wherein Q is a pheromone strength constant, which is the total amount of pheromones released by ants on a path traveled in one cycle; eta_mRepresenting the total amount of heuristic information suffered by the mth ant in the cycle;

module M3.6: updating pheromone;

τ_I，J(t+1)＝(1-ρ)τ_I，J(t)+Δτ_I，J(t，t+1)

wherein rho is the pheromone dynamic volatility coefficient;

represents the pheromone increment left by the mth ant on the path (I, J) in the current cycle, delta tau_I，J(t, t +1) represents pheromone increment left by all ants passing through the path (I, J) in the current cycle; k represents the total number of ants passing through the paths (I, J) in the current cycle;

module M3.7: and (4) optimizing and judging: when the optimization of all the search subspaces reaches the maximum cycle times, the circulation is exited, and the optimization results of all the search subspaces are output as a current candidate set; otherwise, repeatedly triggering the execution of the modules M3.2 to M3.7 until the maximum cycle number is reached;

the mutations evolved in the module M5 include: setting the excitation factors omega in all the sub-network modules in the current optimal network architecture as constants, wherein omega is more than or equal to 0 and less than or equal to 1; and meanwhile, randomly selecting mutation operation in the mutation set, promoting the internal structure of the sub-network module, generating a next generation sub-network module, and repeatedly triggering the execution of the modules M2 to M5 until the preset target requirement is met.

Images