CN112106077A - Method, apparatus, storage medium, and computer program product for network structure search - Google Patents

Abstract

The present application provides a network structure search method, including: a general map training step: training a first general map according to a first network structure and training data to generate a second general map; network structure training step: according to the first general map The network structure determines several test sub-graphs from the second overall graph; tests the several test sub-graphs through test data to generate feedback results; determines jumper constraints according to the feedback results; according to the feedback results and The jumper constraints update the first network structure. In the training process of the neural network, different training stages have different requirements on the jumper density. The jumper constraints in the above method are determined based on the feedback results in the current training stage. Therefore, the jumper density in the above method is different from the current training. Stage correlation makes the current jumper density of the controller more suitable for the current training stage, so that the training efficiency can be improved while achieving good training results.

Description

Method, apparatus, storage medium and computer program product for network structure search

Copyright notice

The disclosure of this patent document contains material that is subject to copyright protection. This copyright belongs to the copyright owner. The copyright owner has no objection to the reproduction by anyone of the patent document or the patent disclosure as it exists in the official records and archives of the Patent and Trademark Office.

technical field

The present application relates to the field of artificial intelligence (Artificial Intelligence, AI), and more particularly, to a method, apparatus, storage medium and computer program product for searching a network structure.

Background technique

Neural network is the foundation of AI. As the performance of neural network continues to improve, its network structure becomes more and more complex. A neural network may need to be trained before it can be used normally. The training process of the neural network is mainly to adjust the operations in each layer of the neural network and the connection between the layers, so that the neural network can output correct results. . The above connection relationship may also be referred to as a jumper (skip) or a shortcut (shortcut).

One way to train a neural network is to use Efficient Neural Architecture Search (ENAS) to train the neural network. In the process of training the neural network with ENAS, the controller continuously samples the network structure of the neural network, tries the influence of different network structures on the output results, and uses the output results of the network structure obtained by the previous sampling to determine the next sampling the network structure until the neural network converges.

Compared with the method of manually debugging the neural network, using ENAS to train the neural network can improve the training efficiency of the neural network, but the training efficiency of using the ENAS to train the neural network still needs to be improved.

SUMMARY OF THE INVENTION

Embodiments of the present application provide a method and apparatus for searching a network structure, a computer storage medium, and a computer program product.

A first aspect provides a network structure search method, comprising: a general graph training step: training a first general graph according to a first network structure and training data to generate a second general graph; network structure training step: according to the The first network structure determines several test subgraphs from the second overall graph; tests the several test subgraphs through test data to generate feedback results; determines jumper constraints according to the feedback results; The result and the jumper constraints update the first network structure.

The above method can be applied to a chip, a mobile terminal or a server. In the training process of neural network, different training stages have different requirements for jumper density; for example, in the initial training stage of some neural networks, in order to explore the search space as much as possible to avoid the bias of the network structure, it is necessary to use A controller with a larger jumper density is used for training; in the later training stage of some neural networks, the randomness of the neural network has been greatly reduced, and the controller with a smaller jumper density can be used for training without exploring the entire search space. , in order to reduce the consumption of resources (including computing power resources and time resources). Since the jumper constraint in the above method is determined based on the feedback results in the current training stage, the jumper constraint in the above method is related to the current training stage, so that the current jumper density of the controller is more suitable for the current training stage, Therefore, it is possible to reduce resource consumption and improve training efficiency while obtaining good training results, which is especially suitable for mobile devices.

In a second aspect, an apparatus for searching a network structure is provided, the apparatus is configured to execute the method in the above-mentioned first aspect.

In a third aspect, a network structure search apparatus is provided, the apparatus includes a memory and a processor, the memory is used for storing instructions, the processor is used for executing the instructions stored in the memory, and the memory is stored in the memory. Execution of the instructions causes the processor to perform the method of the first aspect.

In a fourth aspect, a chip is provided, the chip includes a processing module and a communication interface, the processing module is configured to control the communication interface to communicate with the outside, and the processing module is further configured to implement the method of the first aspect.

In a fifth aspect, a computer-readable storage medium is provided on which a computer program is stored, and when the computer program is executed by a computer, the computer causes the computer to implement the method of the first aspect.

In a sixth aspect, there is provided a computer program product comprising instructions that, when executed by a computer, cause the computer to implement the method of the first aspect.

Description of drawings

1 is a schematic flowchart of a network structure search method provided by the present application;

Fig. 2 is the schematic diagram of a kind of general drawing and sub drawing provided by this application;

3 is a schematic flowchart of another network structure search method provided by the present application;

Fig. 4 is a schematic flowchart of yet another network structure search method provided by the present application;

FIG. 5 is a schematic diagram of a network structure search apparatus provided by the present application.

Detailed ways

The technical solutions in the embodiments of the present application will be described below with reference to the accompanying drawings.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the technical field to which this application belongs. The terms used herein in the specification of the application are for the purpose of describing specific embodiments only, and are not intended to limit the application.

In the description of this application, it should be understood that the terms "first" and "second" are only used to describe different objects, and should not be construed as having other limitations. For example, the "first general drawing" and the "second general drawing" represent two different general drawings, but there is no other definition. In addition, in the description of the present application, "plurality" means two or more, unless otherwise expressly and specifically defined.

In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the term "connection" should be understood in a broad sense, for example, it may be a fixed connection, a detachable connection, or an integral connection; It is a mechanical connection, it can also be an electrical connection or can communicate with each other; it can be a direct connection or an indirect connection through an intermediate medium, and it can be the internal communication between the two elements or the interaction relationship between the two elements. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific situations.

The following disclosure provides many different embodiments or examples for implementing the concepts of the present application. In order to simplify the disclosure of the present application, specific components or steps are described in the following examples. Of course, the following examples are not intended to limit the application. Furthermore, this application may repeat reference numerals and/or reference letters in various instances for the purpose of simplicity and clarity and does not in itself represent a relationship between the various embodiments and/or arrangements discussed.

The embodiments of the present application will be described in detail below. The embodiments described below are exemplary and are only used to explain the present application, but should not be construed as a limitation on the present application.

In recent years, machine learning algorithms, especially deep learning algorithms, have been rapidly developed and widely used. As the model performance continues to improve, the model structure becomes more and more complex. In non-automated machine learning algorithms, these structures need to be manually designed and debugged by machine learning experts, and the process is very cumbersome. Moreover, as application scenarios and model structures become more and more complex, it becomes increasingly difficult to obtain optimal models in application scenarios. In this case, Auto Machine Learning (AutoML) has received extensive attention from academia and industry, especially Neural Architecture Search (NAS).

Specifically, network structure search is a technique that uses algorithms to automate the design of neural network models. The network structure search is to search for the structure of the neural network model. In the embodiments of the present application, the neural network model to be searched for the network structure is a convolutional neural network (Convolutional Neural Networks, CNN).

The problem to be solved by network structure search is to determine the operations between nodes in the neural network model. Different combinations of operations between nodes correspond to different network structures. Further, the nodes in the neural network model can be understood as feature layers in the neural network model. An operation between two nodes refers to an operation required to transform feature data on one of the nodes into feature data on the other node. The operations mentioned in this application may be other neural network operations such as convolution operations, pooling operations, or fully connected operations. It can be considered that the operations between two nodes constitute the operation layer between the two nodes. Usually, there are multiple searchable operations on the operation layer between two nodes, that is, there are multiple candidate operations. The purpose of the network structure search is to identify an operation at each operation layer.

For example, define conv3*3, conv5*5, depthwise3*3, depthwise5*5, maxpool3*3, averagepool3*3, etc. as the search space. That is, each layer of the network structure operates by sampling among these six choices.

As shown in Figure 1, after NAS establishes a search space, it usually uses a controller (a kind of neural network) to sample a network structure A in the search space, and then trains a child network with architecture A (a child network with architecture A) To determine the amount of feedback such as the accuracy R, which may also be referred to as the predicted value; then, compute the gradient of p and scale it with R to update the controller, ie, update R as reward controller.

Then, a new network structure is obtained by sampling in the search space with the updated controller, and the above steps are performed cyclically until a convergent sub-network is obtained.

In the example of FIG. 1 , the controller may be a recurrent neural network (Recurrent Neural Network, RNN), or a CNN or a Long-Short Term Memory (Long-Short Term Memory, LSTM) neural network. This application does not limit the specific form of the controller.

However, training the sub-network structure to convergence is time-consuming. Therefore, various methods for improving the efficiency of NAS have emerged in the related art, such as efficient architecture search by network transformation based on network transformation, and ENAS via parameter sharing (ENAS via parameter sharing) based on weight sharing. Among them, ENAS based on weight sharing is widely used.

As shown in FIG. 2 , the overall graph is composed of operations represented by each node and jumpers between the operations, wherein the operations represented by each node may be all operations in the search space. In the process of using ENAS based on weight sharing, the controller can search the network structure from the search space, for example, determine the operation of each node and the connection relationship between nodes from the search space, Identify a subgraph in the overall graph. The operations connected by the bold arrows in FIG. 2 are an example of the final subgraph, wherein node 1 is the input node of the final subgraph, and node 3 and node 6 are the output nodes of the final subgraph.

In the process of using ENAS based on weight sharing, after sampling a network structure each time, after determining a subgraph from the overall graph based on the network structure, it is no longer directly trained to convergence, but a small It is trained on batches of data, for example, the parameters of the subgraph can be updated using the Back Propagation (BP) algorithm to complete one iteration. Since the subgraph belongs to the general graph, updating the parameters of the subgraph is equivalent to updating the parameters of the general graph. After many iterations, the total graph can finally converge. Note that the convergence of the total graph is not equivalent to the convergence of the subgraphs.

After training the general graph, the parameters of the general graph can be fixed, and then the controller can be trained. For example, a network structure can be searched in the search space by the controller, and a subgraph can be obtained from the overall graph based on the network structure, test data (valid) can be input into the subgraph to obtain a predicted value, and the controller can be updated with the predicted value. .

Since ENAS based on weight sharing shares the parameters that can be shared each time it searches the network structure, the efficiency of network structure search is improved. For example, in the example shown in Figure 2, when the network structure was searched last time, node 1, node 3 and node 6 were searched and the searched network structure was trained. This time, node 1, node 2, node 3 and Node 6, then the parameters related to node 1, node 3 and node 6 obtained last time can be applied to the training of the network structure searched this time. In this way, the efficiency can be improved through weight sharing.

ENAS can improve the efficiency of NAS by more than 1000 times. However, in the process of actual use, the following problems will occur: the predicted value of the sub-graph changes all the time. With the progress of the training process, the prediction of the sub-graph determined by the controller The result is more and more accurate, that is, the predicted value of the subgraph increases gradually; while the coefficient of the jumper constraint item of the parameter update formula of the controller is fixed, so the constraint force generated by the jumper constraint term increases with the subgraph. The training process of the controller decreases continuously; the jumper constraint item reflects the jumper density of the controller, and the gradual decrease of the constraint force generated by the jumper constraint means that the jumper density of the controller gradually increases, and too many jumpers will Increasing the number of floating-point operations per second (FLOPS) of the controller reduces the update efficiency of the controller, which in turn affects the efficiency of determining the final subgraph. In addition, if the initial value of the jumper constraint item is set to a larger value, since the predicted value of the subgraph is small in the initial stage of the overall graph training, the constraint force of the jumper constraint item will be too large, making the update control The controller cannot fully explore the search space, resulting in a large bias in the network structure of the controller.

Based on this, the present application provides a network structure search method, as shown in FIG. 3 , the method includes:

The whole graph (whole graph) training step S12: according to the first network structure and the training (train) data, the first overall graph is trained, and the second overall graph is generated;

Network structure training step S14: Determine several test sub-graphs from the second overall graph according to the first network structure; test the several test sub-graphs through test data to generate feedback results; determine according to the feedback results A jumper constraint item; the first network structure is updated according to the feedback result and the jumper constraint item.

The first network structure may be the network structure of the controller in any training stage. For example, the first network structure may be the network structure of the controller that has never been updated, or the first network structure may be the controller that has been updated several times. the network structure of the device.

In this application, "several" refers to at least one, eg, several times refers to at least one time, and several test submaps refer to at least one test submap.

The first general graph may be a neural network including a preset number of layers. For example, if the preset number of layers is 4, and each layer of neural network corresponds to a search space containing 6 operations, the first general graph may be composed of 24 operations and this The network structure formed by the connection relationship between 24 operations, each layer of network structure contains 6 operations.

The second general graph may not be a converged general graph, however, after training on the training data, the randomness of the second general graph is generally smaller than the randomness of the first general graph.

When training the first general graph, the first network structure can determine the first training subgraph in the first general graph; input a batch of data in the training data into the first training subgraph to generate the first training result; according to The first training result trains the first general graph to generate the second general graph. For example, the first overall graph may be trained using the first training result and a Back Propagation (BP) algorithm.

The first network structure can use the method shown in FIG. 2 to determine the first training subgraph from within the first overall graph, and update the first network structure using the method shown in FIG. 2 . After cyclically executing the overall graph training step and the network structure training step several times, a final overall graph and a final network structure (ie, the final controller) are generated; the final overall graph is determined by the final network structure in the final overall graph Subgraph, the final subgraph is the network structure that conforms to the preset scene.

Next, the process of updating the first network structure will be described in detail.

The first network structure may be an LSTM neural network, and the search space includes, for example, conv3*3, conv5*5, depthwise3*3, depthwise5*5, maxpool3*3 and averagepool3*3.

If the preset number of layers of the network structure to be searched is 20, a 20-layer first general graph needs to be constructed, and each layer of the first general graph includes all operations of the search space. Each layer of the network structure to be searched corresponds to a time step of the LSTM neural network. Without considering the jumper, the LSTM neural network needs to execute 20 time steps. The unit (cell) will output a hidden state (hidden state), which can be encoded (encoding) and mapped to a vector with a dimension of 6, which corresponds to the search space and 6 dimensions correspond to Six operations in the search space; then, the vector is processed by the softmax function and becomes a probability distribution, and the LSTM neural network performs sampling (sample) according to this probability distribution to obtain the operation of the current layer of the network structure to be searched. The above process is repeated to obtain a network structure (ie, subgraph).

FIG. 4 shows an example of determining a network structure.

The blank rectangle represents the cell of the LSTM neural network, the square containing "conv3*3" and other contents represents the operation of the layer in the network structure to be searched, and the circle represents the connection between layers.

The LSTM neural network can perform an encoding operation on the hidden state, and map it to a vector with a dimension of 6. The vector is transformed into a probability distribution through a normalized exponential function (softmax), and sampling is performed according to this probability distribution. Get the operation of the current layer.

For example, during the execution of the first time step, the input quantity (e.g., a random value) to the unit of the LSTM neural network is normalized into a vector by the softmax function, which is subsequently translated into an operation (conv3×3); conv3×3 is used as the input of the unit of the LSTM neural network when executing the second time step, and the hidden state generated by the first time step is also used as the input when executing the second time step. After processing, circle 1 is obtained, and circle 1 indicates that the output of the current operation layer (operation layer corresponding to node 2) and the output of the first operation layer (operation layer corresponding to node 1) are concatenated together.

Similarly, circle 1 is used as the input of the unit of the LSTM neural network when executing the third time step, and the hidden state generated by the second time step is also used as the input when executing the third time step. The amount was processed to obtain sep5×5. By analogy, a network structure is finally obtained.

Then, a subgraph is determined from the first overall graph based on the network structure, a batch of data in the training data is input into the above subgraph, and a training result is generated, so as to train the first overall graph based on the training result. Therefore, the subgraphs determined by the first network structure from the first overall graph may be referred to as training subgraphs.

For example, the controller updates the parameters of the training subgraph according to the training results and the BP algorithm, and completes one iteration. Since the training subgraph belongs to the first overall graph, updating the parameters of the training subgraph is equivalent to updating the parameters of the first overall graph; then , the controller can re-determine a training subgraph from the first overall graph after one iteration, input another batch of training data into the training subgraph, generate another training result, and use the training result and the BP algorithm to update the training subgraph again. Train the subgraph and complete another iteration. After all the training data have been used, a second overall graph is obtained.

After acquiring the second general graph, the parameters of the second general graph can be fixed, and then the controller can be trained. The first network structure can determine a subgraph from the second overall graph according to the method shown in FIG. 4, and this subgraph can be called a test subgraph; input a batch of data in the test data into the test subgraph to obtain a feedback result. (eg, predicted value). The first network structure can be updated directly with the feedback result, or the first network structure can be updated with the average value of multiple feedback results, wherein the multiple test results are obtained by inputting multiple batches of data in the test data into the test subgraph .

In the process of updating the first network structure, the jumper constraint item may be determined according to the feedback result, and then the first network structure is updated according to the jumper line constraint item and the feedback result.

Since the jumper constraints are related to the feedback results of the current training stage, the current jumper density determined based on the jumper constraints is more suitable for the current training stage, so that the network structure with higher reliability can be searched and improved at the same time. training efficiency.

Optionally, the size of the jumper constraint item is positively correlated with the size of the feedback result.

In this application, "positive correlation" means that the value of one parameter increases as the value of another parameter increases, or the value of one parameter decreases as the value of another parameter decreases.

In the initial training stage, the controller needs to fully explore the search space, so as to avoid large deviations that cause the total graph to fail to converge in the subsequent iteration process. Therefore, the jumper density of the controller should not be too small, that is, the value of the jumper constraint should not be too large. big. After several iterations, the randomness of the total graph decreases, i.e., the probability that some operations may be sampled decreases, in which case, continuing to use controller sampling with a higher density of jumpers will result in a decrease in training efficiency and A waste of computing power, therefore, requires a larger jumper constraint to update the controller.

Since the feedback result (prediction accuracy of the test subgraph) usually increases as the training stage progresses, the size of the jumper constraint is positively correlated with the size of the feedback result so that the value of the jumper constraint increases with the training stage. The performance of the network structure is continuously increased, so as to achieve a balance between the performance (ie, the performance of the subgraph) and the training efficiency and computing power in the search process of the network structure.

Optionally, the jumper constraint term includes cos(1-R _k ) ⁿ , where R _k is the feedback result, and n is a hyperparameter related to the application scenario. The value of n can be a real number greater than 0. For example, the value of n can be 10, 20, 30, 40, 50, 60, 70, or 100. Optionally, n can also take a value greater than or equal to 100. value.

其中，α＝cos(1-R_k)ⁿ，λ为超参数，θ_c为第一网络结构的参数，q为预设的期望跳线密度，p为当前跳线密度。当前跳线密度可以是基于测试子图的反馈结果得到的。所述反馈结果包括测试子图对测试数据的预测正确率。Optionally, the jumper constraint term includes the KL divergence (Kullback-Leibler divergence) between the current jumper density and the preset desired jumper density. For example, the jumper constraint is

Wherein, α=cos(1-R _k ) ⁿ , λ is a hyperparameter, θ _c is a parameter of the first network structure, q is a preset desired patch cord density, and p is the current patch cord density. The current patch density can be obtained based on the feedback results of the test subgraph. The feedback result includes the prediction accuracy rate of the test subgraph on the test data.

Optionally, the processor may update the first network structure according to formula (1).

Among them, at is the sampled operation (operation) in the _t -th time step, P(at |a ( _{t -1):1} ; θ _c ) is the probability of sampling the operation, m is the first The number of feedback results used when the network structure is updated once, T is the number of layers in the second overall graph, λ is a hyperparameter, which is generally set to 0.0001 in classification tasks, and different values can be set according to specific tasks.

The meaning of formula (1) is: while maximizing R _k , minimize the KL divergence; that is, maximize R _k while keeping the current patch density consistent with the expected patch density.

In the prior art, since R _k gradually increases with the convergence of the overall graph, and since the penalty force generated by λ remains unchanged, q can usually only be set to 0.4 when constraining the jumper in the actual iterative process. -0.6. Among them, q is calculated according to (the number of all current jumpers)/(the number of all connectable jumpers), and the value is between 0 and 1. In the initial state, the jumpers are randomly connected lines, and the jumper density is 0.5.

Compared with the prior art, formula (1) adds a coefficient α, and α is positively correlated with R _k . In the initial training stage, the predicted values generated by the controller from the test subgraphs determined in the overall graph are not accurate enough. Therefore, _Rk is small, α is also small, and the penalty of the jumper constraint is small, and the updated controller With a large jumper density, it can fully explore the search space and avoid large deviations in the initial training phase; as the training phase progresses, the accuracy of the prediction values generated by the controller from the test subgraphs determined in the overall graph increases , R _k gradually increases, α also increases gradually, the penalty of the jumper constraint term also increases gradually, the jumper density of the updated controller is small, and the search space can no longer be fully explored (due to the randomness of the overall graph In addition, since the updated controller no longer fully explores the search space, there is no need for too many FLOPS, thus reducing computing power consumption. It can be seen from the above that α is an adaptive coefficient, which can balance the performance of the network structure (that is, the performance of the subgraph) and the computing power consumption during the search process of the network structure, and is especially suitable for mobile devices with weak processing capabilities.

It should be noted that the jumper constraints including α in the above are only examples, and any jumper constraints that can be adaptively adjusted according to the training stage fall within the protection scope of the present application.

Examples of the network structure search method provided by the present application are described in detail above. It can be understood that, in order to realize the above-mentioned functions, the apparatus for searching the network structure includes corresponding hardware structures and/or software modules for executing each function. Those skilled in the art should easily realize that the present application can be implemented in hardware or a combination of hardware and computer software with the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein. Whether a function is performed by hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each particular application, but such implementations should not be considered beyond the scope of this application.

The present application can divide the functional units of the apparatus for network structure search according to the above method examples. For example, each function can be divided into each functional unit, or two or more functions can be integrated into one functional unit. The above-mentioned functional units may be implemented in the form of hardware, or may be implemented in the form of software. It should be noted that the division of units in this application is schematic, and is only a logical function division, and other division methods may be used in actual implementation.

FIG. 5 shows a schematic structural diagram of an apparatus for searching a network structure provided by the present application. The dotted line in Figure 5 indicates that this unit is an optional unit. The apparatus 500 can be used to implement the methods described in the above method embodiments. The apparatus 500 may be a software module, chip, terminal device or other electronic device.

The apparatus 500 includes one or more processing units 501, and the one or more processing units 501 can support the apparatus 500 to implement the method in the method embodiment corresponding to FIG. 3 . The processing unit 501 may be a software processing unit, a general purpose processor or a special purpose processor. The processing unit 501 may be used to control the apparatus 500, execute a software program (eg, a software program for implementing the method described in the first aspect), and process data (eg, a predicted value). The apparatus 500 may further include a communication unit 505 to enable input (reception) and output (transmission) of signals.

For example, the apparatus 500 may be a software module, and the communication unit 505 may be an interface function of the software module. The software module may run on a processor or control circuit.

For another example, the apparatus 500 may be a chip, and the communication unit 505 may be an input and/or output circuit of the chip, or the communication unit 505 may be a communication interface of the chip, and the chip may be used as a component of a terminal device or other electronic device .

In the apparatus 500, the processing unit 501 may execute:

The general map training step: according to the first network structure and training data, the first general map is trained to generate the second general map;

The network structure training step: determine several test sub-graphs from the second overall graph according to the first network structure; test the several test sub-graphs through test data to generate feedback results; Line constraint item; update the first network structure according to the feedback result and the jumper line constraint item.

Optionally, the size of the jumper constraint item is positively correlated with the size of the feedback result.

Optionally, the jumper constraint item includes cos(1-R _k ) ⁿ , where R _k is the feedback result, and n is a hyperparameter related to an application scenario.

Optionally, 0<n≤100.

Optionally, the patch cord constraint item includes a KL divergence between the current patch cord density and a preset desired patch cord density.

其中，α＝cos(1-R_k)ⁿ，λ为超参数，θ_c为所述第一网络结构的参数，q为预设的期望跳线密度，p为当前跳线密度。Optionally, the jumper constraint item is

Wherein, α=cos(1-R _k ) ⁿ , λ is a hyperparameter, θ _c is a parameter of the first network structure, q is a preset desired patch cord density, and p is a current patch cord density.

Optionally, the current patch cord density is obtained based on the several test submaps.

Optionally, the feedback result includes a prediction accuracy rate of the test data by the several test subgraphs.

Optionally, the processing unit 501 is specifically configured to: determine a first training subgraph in the first overall graph through the first network structure; input a batch of data in the training data into the first training a subgraph to generate a first training result; training the first overall graph according to the first training result to generate the second overall graph.

Optionally, the processing unit 501 is further configured to: generate a final overall graph and a final network structure after cyclically executing the overall graph training step and the network structure training step several times; A final subgraph is determined in the final overall graph, and the final subgraph is a network structure conforming to a preset scene.

Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process and the effect of the above devices and units may refer to the relevant descriptions in the method embodiments corresponding to FIG. 1 to FIG. 4 . For brevity, details are not repeated here.

As an optional implementation manner, each of the above steps may be completed by a logic circuit in the form of hardware or an instruction in the form of software. For example, the processing unit 501 may be a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), or a field programmable gate array (field programmable gate array). array, FPGA) or other programmable logic devices such as discrete gates, transistor logic devices or discrete hardware components.

The apparatus 500 may include one or more storage units 502 in which a program 504 (eg, a software program including the method described in the second aspect) is stored, and the program 504 may be executed by the processing unit 501 to generate instructions 503 to cause the processing unit 501 to The method described in the above method embodiment is performed according to the instruction 503 . Optionally, the storage unit 502 may also store data (eg, predicted values and patch density). Optionally, the processing unit 501 may also read the data stored in the storage unit 502 , the data may be stored at the same storage address as the program 504 , or the data may be stored at a different storage address from the program 504 .

The processing unit 501 and the storage unit 502 may be provided separately, or may be integrated together, for example, integrated on a single board or a system on chip (system on chip, SOC).

The present application also provides a computer program product, which implements the method described in any embodiment of the present application when the computer program product is executed by the processing unit 501 .

The computer program product can be stored in the storage unit 502 , such as a program 504 , and the program 504 is finally converted into an executable object file that can be executed by the processing unit 501 through processing processes such as preprocessing, compilation, assembly, and linking.

The computer program product can be transmitted from one computer-readable storage medium to another computer-readable storage medium, eg, from a website site, computer, server, or data center over a wireline (eg, coaxial cable, optical fiber, digital subscriber line ( digital subscriber line, DSL) or wireless (such as infrared, wireless, microwave, etc.) to another website site, computer, server or data center.

The present application also provides a computer-readable storage medium (eg, the storage unit 502 ), on which a computer program is stored, and when the computer program is executed by a computer, implements the method described in any of the embodiments of the present application. The computer program can be a high-level language program or an executable object program.

The computer-readable storage medium may be a magnetic medium (eg, a floppy disk, a hard disk, a magnetic tape), an optical medium (eg, a digital video disc (DVD)), or a semiconductor medium (eg, a solid state disk (SSD)) Wait. For example, the computer-readable storage medium may be volatile memory or non-volatile memory, or the computer-readable storage medium may include both volatile memory and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable memory Except programmable read-only memory (electrically EPROM, EEPROM) or flash memory. Volatile memory may be random access memory (RAM), which acts as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM) ), double data rate synchronous dynamic random access memory (double data rate SDRAM, DDR SDRAM), enhanced synchronous dynamic random access memory (enhanced SDRAM, ESDRAM), synchronous link dynamic random access memory (synchlink DRAM, SLDRAM) And direct memory bus random access memory (direct rambus RAM, DR RAM).

It should be understood that, in each embodiment of the present application, the size of the sequence number of each process does not mean the sequence of execution, and the execution sequence of each process should be determined by its function and internal logic, and should not Implementation constitutes any limitation.

The term "and/or" in this article is only an association relationship to describe the associated objects, indicating that there can be three kinds of relationships, for example, A and/or B, it can mean that A exists alone, A and B exist at the same time, independently There are three cases of B. In addition, the character "/" in this document generally indicates that the related objects are an "or" relationship.

The systems, devices, and methods disclosed in the embodiments provided in this application may be implemented in other ways. For example, some features of the method embodiments described above may be omitted, or not implemented. The apparatus embodiments described above are only illustrative, and the division of units is only a logical function division. In actual implementation, there may be other division methods, and multiple units or components may be combined or integrated into another system. In addition, the coupling between the various units or the coupling between the various components may be direct coupling or indirect coupling, and the above-mentioned coupling includes electrical, mechanical or other forms of connection.

In a word, the above descriptions are only some embodiments of the present application, and are not intended to limit the protection scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of this application shall be included within the protection scope of this application.

Claims (23)

1. A network structure search method, comprising:

and (3) general diagram training: training the first general diagram according to the first network structure and the training data to generate a second general diagram;

network structure training: determining a plurality of test subgraphs from the second general graph according to the first network structure; testing the plurality of test subgraphs through the test data to generate a feedback result; determining a jumper wire constraint item according to the feedback result; and updating the first network structure according to the feedback result and the jumper constraint item.

2. The method of claim 1, wherein the size of the jumper constraint term is positively correlated to the size of the feedback result.

3. The method of claim 2, wherein the jumper constraint term comprises cos (1-R)_k)ⁿ，R_kAnd n is a hyper-parameter related to the application scene for the feedback result.

4. The method of claim 3, wherein 0< n ≦ 100.

5. The method according to any one of claims 1 to 4, wherein the jumper constraint term includes a KL divergence between a current jumper density and a preset desired jumper density.

6. The method of claim 5, wherein the step of removing the metal oxide layer comprises removing the metal oxide layer from the metal oxide layerIn that, the jumper constraint term is

Wherein α ═ cos (1-R)_k)ⁿλ is a hyperparameter, θ_cQ is a preset expected jumper density and p is a current jumper density, wherein q is a parameter of the first network structure.

7. The method of claim 5 or 6, wherein the current patch cord density is derived based on the number of test subgraphs.

8. The method of any of claims 1 to 7, wherein the feedback results comprise predicted correctness of the test data by the number of test subgraphs.

9. The method of any one of claims 1 to 8, wherein training the first total graph according to the first network structure and the training data to generate a second total graph comprises:

determining a first training subgraph within the first population graph over the first network structure;

inputting a batch of data in the training data into the first training subgraph to generate a first training result;

and training the first general diagram according to the first training result to generate the second general diagram.

10. The method of any one of claims 1 to 9, further comprising:

after the overall graph training step and the network structure training step are executed for a plurality of times in a circulating mode, generating a final overall graph and a final network structure;

and determining a final sub-graph in the final general graph through the final network structure, wherein the final sub-graph is a network structure conforming to a preset scene.

11. A network structure search apparatus, comprising a processing unit configured to perform:

and (3) general diagram training: training the first general diagram according to the first network structure and the training data to generate a second general diagram;

network structure training: determining a plurality of test subgraphs from the second general graph according to the first network structure; testing the plurality of test subgraphs through the test data to generate a feedback result; determining a jumper wire constraint item according to the feedback result; and updating the first network structure according to the feedback result and the jumper constraint item.

12. The apparatus of claim 11, wherein the size of the jumper constraint term is positively correlated to the size of the feedback result.

13. The apparatus of claim 12, wherein the jumper constraint term comprises cos (1-R)_k)ⁿ，R_kAnd n is a hyper-parameter related to the application scene for the feedback result.

14. The apparatus of claim 13, wherein 0< n ≦ 100.

15. The apparatus of any of claims 11-14, wherein the jumper constraint term comprises a KL divergence between a current jumper density and a preset desired jumper density.

16. The apparatus of claim 15, wherein the jumper constraint term is

Wherein α ═ cos (1-R)_k)ⁿλ is a hyperparameter, θ_cQ is a preset expected jumper density and p is a current jumper density, wherein q is a parameter of the first network structure.

17. The apparatus of claim 15 or 16, wherein the current jumper density is derived based on the number of test subgraphs.

18. The apparatus of any of claims 11 to 17, wherein the feedback results comprise predicted correctness of the test data by the number of test subgraphs.

19. The apparatus according to any one of claims 11 to 18, wherein the processing unit is specifically configured to:

determining a first training subgraph within the first population graph over the first network structure;

inputting a batch of data in the training data into the first training subgraph to generate a first training result;

and training the first general diagram according to the first training result to generate the second general diagram.

20. The apparatus according to any one of claims 11 to 19, wherein the processing unit is further configured to:

after the overall graph training step and the network structure training step are executed for a plurality of times in a circulating mode, generating a final overall graph and a final network structure;

and determining a final sub-graph in the final general graph through the final network structure, wherein the final sub-graph is a network structure conforming to a preset scene.

21. A network structure search device characterized by comprising: a memory for storing instructions and a processor for executing the instructions stored by the memory, and execution of the instructions stored in the memory causes the processor to perform the method of any of claims 1 to 10.

22. A computer storage medium, having stored thereon a computer program which, when executed by a computer, causes the computer to perform the method of any one of claims 1 to 10.

23. A computer program product comprising instructions which, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 10.

Images