WO2021036397A1 - Method and apparatus for generating target neural network model - Google Patents

Abstract

A method and apparatus for generating a target neural network model in the field of artificial intelligence. The method comprises: acquiring scoring weight values of a plurality of candidate operators and a plurality of neural network models; according to the scoring weight values of the plurality of candidate operators, the precision rate of each neural network model, and the delay of each neural network model, determining differentiable relationship information of a loss function with respect to scoring variables, the loss function being a joint function of the precision rates and the delays; updating the scoring weight values of the plurality of candidate operators according to the differentiable relationship information; generating a target neural network model according to the updated scoring weight values; and sending target model configuration information, the target model configuration information being used to configure the target neural network model. The described method may improve the efficiency of generating a target neural network model and reduce the consumption of computing resources and time.

Description

Method and device for generating target neural network model

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office, the application number is 201910814558.6, and the application name is "Method and Apparatus for Generating Target Neural Network Model" on August 30, 2019, the entire content of which is incorporated herein by reference. Applying.

Technical field

This application relates to the field of artificial intelligence, and in particular to a method and device for generating a target neural network model.

Background technique

Artificial Intelligence (AI) is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge, and use knowledge to obtain the best results. In other words, artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a similar way to human intelligence. Artificial intelligence is to study the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making. Research in the field of artificial intelligence includes robotics, natural language processing, computer vision, decision-making and reasoning, human-computer interaction, recommendation and search, and basic AI theories.

AutoMachine Learning (Automl) is a method of automatically designing convolutional neural networks. The target data set and evaluation criteria are input to the Automl system, and the Automl system will output the target neural network.

In the process of determining the target neural network, the Automl system needs to generate different neural network models, and separately train each neural network model to obtain the accuracy and delay of each neural network model, and use the accuracy and delay as The feedback information is used to update the parameters and generation rules of the neural network model according to the feedback information to determine the target neural network.

The above method of determining the target neural network requires sampling to obtain a large number of neural network models for training, which consumes more computing resources and time.

Summary of the invention

The present application provides a method and device for generating a target neural network model to improve the efficiency of generating a target neural network model and reduce the consumption of computing resources and time.

In the first aspect, an embodiment of the present application provides a method for generating a target neural network model. The method may include: obtaining score weights of a plurality of candidate operators and a plurality of neural network models; The scoring weight, the accuracy rate of each neural network model, and the delay of each neural network model determine the differentiable relationship information of the loss function with respect to the scoring variable, and the loss function is a joint function of the accuracy rate and the delay; Differentiable relationship information updates the score weights of the multiple candidate operators; generates a target neural network model according to the updated score weights; sends target model configuration information, the target model configuration information is used to configure the target neural network model.

In the solution of the present application, by determining the differentiable relationship information of the loss function on the scoring variable, and optimizing the scoring weight of the candidate operator based on the differentiable relationship information, the target neural network model can be generated based on the scoring weight of the candidate operator. Efficiency, reducing the consumption of computing resources and time.

In a possible design, each neural network model of the plurality of neural network models is determined by L candidate operators, and each candidate operator is one of the N candidate operators of the corresponding selection block, The differentiable relationship information of the loss function with respect to the scoring variable includes L*N differentiable relationship information, and each differentiable relationship information corresponds to a candidate operator of a selection block; the update of the differentiable relationship information is based on the differentiable relationship information. The scoring weights of multiple candidate operators include: updating the scoring weights of the candidate operators corresponding to each differentiable relationship information according to the L*N differentiable relationship information; where L and N are any positive Integer.

In the solution of this application, by determining the differentiable relationship information corresponding to each candidate operation of each selection block, and optimizing the scoring weight of each candidate operator according to the differentiable relationship information, the generation of the target neural network model can be improved. Efficiency, reducing the consumption of computing resources and time.

In a possible design, the method further includes: acquiring a target data set and a delay adjustment parameter; determining the loss function according to the delay adjustment parameter; updating the candidate operator's value according to the target data set Internal parameters; the multiple neural network models are determined according to the updated internal parameters and the score weights of the multiple candidate operators.

In the solution of this application, the target data set and time delay adjustment parameters are obtained, so that the generated target neural network model meets the corresponding functional requirements, the internal parameters of the candidate operators are updated according to the target data set, and the internal parameters of the candidate operators are updated according to the updated internal parameters. And the scoring weights of the candidate operations to determine multiple neural network models. According to the scoring weights of the candidate operators, the accuracy and delay of each neural network model, the differentiable relationship information of the loss function on the scoring variables is determined, and according to the differentiable Relational information optimizes the scoring weights of the candidate operators, thereby optimizing the internal parameters and scoring weights of the candidate operators, which can improve the efficiency of generating the target neural network model based on the scoring weights of the candidate operators, and reduce the consumption of computing resources and time .

In a possible design, the accuracy of each neural network model and the delay of each neural network model are used to determine the feasibility of the loss function with respect to the scoring variable according to the scoring weights of the multiple candidate operators. The micro-relation information includes: determining the m-th accuracy differentiable relation information of the l-th selection block according to the score weights of the N candidate operators of each selection block and the accuracy of each neural network model. The accuracy of each neural network model is determined based on the target data set; the sampling probability of each selection block of each neural network model is determined according to the score weights of the N candidate operators of each selection block and The sampling probability of each candidate operator of each selection block is related to the differentiable relationship information of the scoring variable; according to the sampling probability of the operator of the L-1 selection block of each neural network model, the m-th selection block of the l The differentiable relationship information of the sampling probabilities of the candidate operators with respect to the scoring variable and the delay of each neural network model determine the m-th delay differentiable relationship information of the l-th selection block, the L-1 Selection blocks include selection blocks other than the lth selection block among the L selection blocks; according to the mth accuracy rate of the lth selection block, the differentiable relationship information and the value of the lth selection block The m-th time delay differentiable relationship information determines the m-th differentiable relationship information of the l-th selection block. Among them, the value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N.

In the solution of the present application, the optimized scoring weight minimizes the accuracy rate error and the average network delay at the same time, thereby improving the accuracy of the target neural network model constructed based on the candidate model and reducing the network delay.

In a possible design, the updating the score weights of the candidate operators corresponding to each differentiable relationship information according to the L*N differentiable relationship information includes: separately updating the L*N differentiable relationship information Differentiable relationship information and the score weight of each candidate operator are input to the stochastic gradient descent algorithm, and the updated score weight of each candidate operator in each selection block is output.

In a possible design, the updating the internal parameters of the candidate operator according to the target data set includes: using the target data set to train a first neural network model, and the first neural network model is determined by L The number of candidate operators is determined, and each candidate operator is one of the N candidate operators of the corresponding selection block; the internal parameters of the L candidate operators are updated according to the training result; where L is any positive integer.

In a possible design, the method further includes: determining whether there is a first operator according to the updated scoring weight and a preset condition, and the first operator is a candidate operation for L selection blocks One or more of the operators; if there is a first operator, remove the first operator from the candidate operators of the selection block corresponding to the first operator.

In the solution of the present application, by excluding candidate operators with a smaller score weight from the candidate operators, the computing resources and time spent in obtaining the target neural model can be reduced.

In a possible design, after the update of the score weights of the multiple candidate operators according to the differentiable relationship information, the method further includes: judging whether a cut-off condition is satisfied, and if the cut-off condition is not satisfied, Then execute the step of training the first neural network model using the target data set; if the cut-off condition is met, execute the step of generating the target neural network model according to the updated scoring weight.

In a possible design, each candidate operator of each selection block has an equal probability of being selected as an operator of the first neural network model.

In the solution of this application, by setting the probability that each candidate operator of each selection block is selected as an operator of the first neural network model to be equal, the internal parameters of each candidate operator can be upgraded and updated The uniformity.

In a possible design, the method further includes: acquiring identification information of the speed measuring device; sending neural network model configuration information to the speed measuring device according to the identification information, and the neural network model configuration information is used to configure the speed measuring device. The multiple neural network models; receiving the time delay of each neural network model sent by the speed measuring device.

In the solution of this application, by acquiring the identification information of the speed measuring device, measuring the time delay of each neural network model on the corresponding speed measuring device, and optimizing the score weight of each candidate operator based on this, the target neural network can be generated The model meets the actual needs of use.

In a possible design, the neural network model configuration information includes multiple network structure coding information, and each network structure coding information is used to configure a neural network model of various operators and the connection relationship of the operators.

In a possible design, the generating the target neural network model according to the updated scoring weight includes: selecting, in each selection block, the candidate operator with the largest score in the selection block as the target neural network model. An operator of the network model, the internal parameter of the operator is the updated internal parameter.

In a possible design, the obtaining the target data set and the time delay adjustment parameter includes: receiving the target data set and the time delay adjustment parameter input by the user; or, receiving the data set selection information and the time delay adjustment parameter input by the user. The delay adjustment parameter determines the target data set according to the data set selection information; or, receives data set selection information and expected complexity information input by a user, and determines the target data set according to the data set selection information , Determining the delay adjustment parameter according to the expected complexity information.

In a second aspect, an embodiment of the present application provides a device for generating a target neural network model. The device for generating a target neural network model is used to execute the target neural network in the first aspect or any possible implementation of the first aspect. The generation method of the network model. Specifically, the device for generating the target neural network model may include a module for executing the method for generating the target neural network model in the first aspect or any possible implementation of the first aspect.

In a third aspect, an embodiment of the present application provides a terminal device. The terminal device includes a memory and a processor. The memory is used to store instructions. The execution of the instructions stored in the processor enables the processor to execute the method for generating the target neural network model in the first aspect or any possible implementation of the first aspect.

In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the first aspect or any one of the possible implementation manners of the first aspect is implemented. method.

In a fifth aspect, the present application provides a computer program product. The computer program product includes instructions that, when run on a computer, cause the computer to execute the method described in any one of the above-mentioned first aspects.

In a sixth aspect, the present application provides a chip including a processor and a memory, the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to execute the above-mentioned first aspect The method of any one of.

The method and device for generating a target neural network model of the present application obtain the scoring weights of multiple candidate operators and multiple neural network models, according to the scoring weights of the multiple candidate operators and the accuracy of each neural network model. The rate and the delay of each neural network model determine the differentiable relationship information of the loss function on the score variable. The loss function is a joint function of accuracy and delay. The scores of the multiple candidate operators are updated according to the differentiable relationship information The weight value is to generate the target neural network model according to the updated scoring weight value, and send the target model configuration information. The target model configuration information is used to configure the target neural network model. In this embodiment, the differentiable relationship information about the scoring variable is determined by the loss function Optimize the scoring weight of the operator according to the differentiable relationship information, which can improve the efficiency of generating the target neural network model and reduce the consumption of computing resources and time.

Description of the drawings

FIG. 1 is a schematic diagram of an artificial intelligence main body framework provided by an embodiment of this application;

2A is a schematic diagram of an application environment provided by an embodiment of the application;

2B is a schematic diagram of an application environment provided by an embodiment of this application;

FIG. 3 is a schematic diagram of a candidate model provided by an embodiment of the application;

4 is a schematic diagram of a convolutional neural network structure provided by an embodiment of the application;

FIG. 5 is a schematic diagram of a convolutional neural network structure provided by an embodiment of this application;

Fig. 6 is a method for generating a target neural network model according to an embodiment of the application;

FIG. 7 is another method for generating a target neural network model according to an embodiment of the application;

FIG. 8 is another method for generating a target neural network model according to an embodiment of the application;

FIG. 9 is a schematic diagram of a candidate model provided by an embodiment of the application;

FIG. 10 is a schematic diagram of a neural network model provided by an embodiment of this application;

FIG. 11A is a schematic diagram of a method for generating a target neural network model provided by an embodiment of this application;

FIG. 11B is a schematic diagram of updating the score weight of each candidate operator in the selection block 1 according to an embodiment of the application;

FIG. 12 is a flowchart of another method for generating a target neural network model according to an embodiment of the application;

FIG. 13 is a schematic structural diagram of a chip provided by an embodiment of the application;

FIG. 14 is a schematic diagram of a system architecture 300 provided by an embodiment of this application;

15 is a schematic block diagram of a device 1600 for generating a target neural network model provided by an embodiment of the application;

FIG. 16 is a schematic block diagram of an electronic device 1700 provided by this application.

detailed description

The following describes the technical solutions in the embodiments of the present application with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

Figure 1 shows a schematic diagram of an artificial intelligence main framework, which describes the overall workflow of the artificial intelligence system and is suitable for general artificial intelligence field requirements.

The following describes the above-mentioned artificial intelligence theme framework from the two dimensions of "intelligent information chain" (horizontal axis) and "IT value chain" (vertical axis).

"Intelligent Information Chain" reflects a series of processes from data acquisition to processing. For example, it can be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, the data has gone through the condensing process of "data-information-knowledge-wisdom".

The "IT value chain" is the industrial ecological process from the underlying infrastructure and information (providing and processing technology realization) of human intelligence to the system, reflecting the value that artificial intelligence brings to the information technology industry.

(1) Infrastructure:

The infrastructure provides computing power support for the artificial intelligence system, realizes communication with the outside world, and realizes support through the basic platform. Communicate with the outside through sensors; computing capabilities are provided by smart chips (hardware acceleration chips such as CPU, NPU, GPU, ASIC, FPGA); basic platforms include distributed computing frameworks and network related platform guarantees and support, which can include cloud storage and Computing, interconnection network, etc. For example, sensors communicate with the outside to obtain data, and these data are provided to the smart chip in the distributed computing system provided by the basic platform for calculation.

(2) Data

The data in the upper layer of the infrastructure is used to represent the data source in the field of artificial intelligence. The data involves graphics, images, voice, and text, as well as the Internet of Things data of traditional devices, including business data of existing systems and sensory data such as force, displacement, liquid level, temperature, and humidity.

(3) Data processing

Data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other methods.

Among them, machine learning and deep learning can symbolize and formalize data for intelligent information modeling, extraction, preprocessing, training, etc.

Reasoning refers to the process of simulating human intelligent reasoning in a computer or intelligent system, using formal information to conduct machine thinking and solving problems based on reasoning control strategies. The typical function is search and matching.

Decision-making refers to the process of making decisions based on intelligent information after reasoning, and usually provides functions such as classification, sorting, and prediction.

(4) General ability

After the above-mentioned data processing is performed on the data, some general capabilities can be formed based on the results of the data processing, such as an algorithm or a general system, for example, translation, text analysis, computer vision processing, speech recognition, image Recognition and so on.

(5) Smart products and industry applications

Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. It is an encapsulation of the overall solution of artificial intelligence, productizing intelligent information decision-making and realizing landing applications. Its application fields mainly include: intelligent manufacturing, intelligent transportation, Smart home, smart medical, smart security, autonomous driving, safe city, smart terminal, etc.

Referring to FIG. 2A, an embodiment of the present application provides a system architecture 200. The data collection device 260 is used to collect target data and store it in the database 230, and the training device 220 generates a target model/rule 201 based on the target data maintained in the database 230. The following will describe in detail how the training device 220 obtains the target model/rule 201 based on the target data. The target model/rule 201 can be applied to image classification, speech recognition, and the like.

A deep neural network can include multiple layers, and each layer can be set with the same or different operators. The operator can be a convolution operator, a pooling operator, a fully connected operator, etc. Exemplarily, the convolution operator may be a 3*3 convolution operator, a 5*5 convolution operator, or the like.

The operators of each layer in the deep neural network can be expressed in mathematical expressions

To describe: From the physical level, the operators of each layer in the deep neural network can be understood as five operations on the input space (the set of input vectors) to complete the transformation from the input space to the output space (that is, the row space of the matrix to Column space), these five operations include: 1. Dimension Up/Down; 2. Enlarge/Reduce; 3. Rotate; 4. Translation; 5. "Bend". The operations of 1, 2, and 3 are determined by

Complete, the operation of 4 is completed by +b, and the operation of 5 is realized by a(). The reason why the word "space" is used here is because the object to be classified is not a single thing, but a class of things. Space refers to the collection of all individuals of this kind of thing. Among them, W is a weight vector, and each value in the vector represents the weight value of a neuron in the layer of neural network. This vector W determines the spatial transformation from the input space to the output space described above, that is, the weight W of each layer controls how the space is transformed. The purpose of training a deep neural network is to finally obtain the weight matrix of all layers of the trained neural network (the weight matrix formed by the vector W of many layers). Therefore, the training process of the neural network is essentially the way of learning to control the spatial transformation, and more specifically the learning of the weight matrix.

Different combinations of the above five operations can implement pooling operations, convolution operations, and fully connected operations.

Because it is hoped that the output of the deep neural network is as close as possible to the value that you really want to predict, you can compare the current network's predicted value with the really desired target value, and then update each layer of neural network according to the difference between the two. The weight vector of the network (of course, there is usually an initialization process before the first update, which is to pre-configure parameters for each layer in the deep neural network). For example, if the predicted value of the network is high, adjust the weight vector to make it The prediction is lower and keep adjusting until the neural network can predict the target value you really want. Therefore, it is necessary to predefine "how to compare the difference between the predicted value and the target value". This is the loss function or objective function, which is used to measure the difference between the predicted value and the target value. Important equation. Among them, taking the loss function as an example, the higher the output value (loss) of the loss function, the greater the difference, then the training of the deep neural network becomes a process of reducing this loss as much as possible.

The training device 220 can determine the operators of each layer of the deep neural network and the internal parameters of the operators of each layer according to the target data set according to the method for generating the target neural network model of the embodiment of the present application, for example, the weights described above vector.

Take the required deep neural network including L layers as an example. As shown in Figure 3, the candidate model of the deep neural network includes L layers (also called selection blocks), and each layer includes N candidate operations. The scoring weight (α) between each candidate operator and the previous layer can be the same or different. For example, as shown in Figure 3, the input data is x ₀ , the output data of the first layer is x ₁ , the output data of the second layer is x ₂ ,..., the output data of the Lth layer is x _L , the first layer The scoring weight between the first candidate operator and the input data x _{0 is α 1,1} , and the scoring weight between the second candidate operator of the first layer and the input data x ₀ is α _1,2 ,..., the scoring weight between the Nth candidate operator of the first layer and the input data x _{0 is α 1,N} . Through the method for generating the target neural network model of the embodiment of the present application, one operator can be determined to form the target neural network model among the N candidate operators in each layer. The target neural network model can also be called target model/rule. Among them, L and N are arbitrary positive integers. Any candidate operator can be a convolution operator, a pooling operator, or a fully connected operation, etc.

It should be noted that in the candidate model shown in Fig. 3, L selection blocks are connected in series as a possible implementation method. L selection blocks can also be connected in parallel, or a combination of series and parallel. A directed acyclic graph, the directed acyclic graph can be used as a candidate model, and the embodiment of the present application uses the candidate model as shown in FIG. 3 as an example, which is not limited thereto.

The target model/rule obtained by the training device 220 can be applied to different systems or devices. In FIG. 2A, the execution device 210 is configured with an I/O interface 212 to perform data interaction with external devices, and the "user" can input data to the I/O interface 212 through the client device 240.

The execution device 210 can call data, codes, etc. in the data storage system 250, and can also store data, instructions, etc. in the data storage system 250.

The calculation module 211 uses the target model/rule 201 to process the input data, and returns the processing result to the client device 240 through the I/O interface 212, and provides it to the user.

At a deeper level, the training device 220 can generate corresponding target models/rules 201 based on different data for different targets, so as to provide users with better results.

In the case shown in FIG. 2A, the user can manually specify the input data in the execution device 210, for example, to operate in the interface provided by the I/O interface 212. In another case, the client device 240 can automatically input data to the I/O interface 212 and obtain the result. If the client device 240 automatically inputs data and needs the user's authorization, the user can set the corresponding authority in the client device 240. The user can view the result output by the execution device 210 on the client device 240, and the specific presentation form may be a specific manner such as display, sound, and action. The client device 240 may also serve as a data collection terminal to store the collected target data in the database 230.

It is worth noting that FIG. 2A is only a schematic diagram of a system architecture provided by an embodiment of the present application, and the positional relationship between the devices, devices, modules, etc. shown in the figure does not constitute any limitation. For example, in FIG. 2A, The data storage system 250 is an external memory relative to the execution device 210. In other cases, the data storage system 250 may also be placed in the execution device 210.

For another example, referring to FIG. 2B, an embodiment of the present application provides another system architecture 400. The system architecture 400 may include a client device 410 and a server 420. The client device 410 may establish a connection with the server 420, and the server 420 may use this The method for generating a target neural network model of the application embodiment generates a target model/rule, and provides the target model/rule to the client device 410. In some embodiments, the client device 410 may configure the target model/rule on the corresponding execution device, for example, an embedded neural network processor (Neural-network Processing Unit, NPU).

The method for generating the target neural network model of the embodiment of the present application may generate a convolutional neural network (CNN, Convolutional Neuron Network) as shown in FIG. 4 based on the candidate model shown in FIG. 3 (where L is 9). CNN is a deep neural network with a convolutional structure. It is a deep learning architecture. The deep learning architecture refers to the use of machine learning algorithms to perform multiple levels of learning at different abstract levels. As a deep learning architecture, CNN is a feed-forward artificial neural network. Each neuron in the feed-forward artificial neural network responds to overlapping regions in the input image.

As shown in FIG. 4, a convolutional neural network (CNN) may include an input layer, intermediate layers 121-127, and an output layer 130. The input layer 110 may be a candidate operator in the first selection block shown in FIG. 3, and the middle layer 121-127 may be a candidate in the second to eighth selection blocks shown in FIG. 3 Operator, the output layer 130 may be a candidate operator in the 9th selection block as shown in FIG. 3.

The middle layer 121-177 can be pooling, convolution or fully connected operation. The middle layer for pooling operation is the pooling layer, the middle layer for convolution operation is the convolutional layer, and the middle layer for full connection operation is Fully connected layer.

In one implementation, layer 121 is a convolutional layer, layer 122 is a pooling layer, layer 123 is a convolutional layer, layer 124 is a pooling layer, 125 is a convolutional layer, and 126 is a pooling layer; in another In the implementation manner, 121 and 122 are convolutional layers, 123 is a pooling layer, 124 and 125 are convolutional layers, and 126 is a pooling layer. That is, the output of the convolutional layer can be used as the input of the subsequent pooling layer, or as the input of another convolutional layer to continue the convolution operation.

It should be noted that the convolutional neural network shown in Figure 4 is only used as an example of a convolutional neural network. In specific applications, the convolutional neural network can also exist in the form of other network models, and this application can be used The method for generating the target neural network model of the embodiment (using 15 candidate models of selection blocks, the candidate models using a combination of series and parallel) generates, for example, a convolutional neural network as shown in FIG. 5, that is, multiple intermediate layers are parallel, The respectively extracted features are input to the middle layer 127 for processing.

In the present application, the target neural network model can be obtained by the method for generating the target neural network model as described below, and the target neural network model can be applied to face recognition, image classification, image super-resolution, and obstacle detection of the preceding vehicle. For example, when applied to face recognition, the target neural network model of the embodiment of the present application can process the input face image and output a vector for similarity matching. Applied to image classification, the target neural network model of the embodiment of the present application can process the input image to be classified and output a tensor (Tensor), which is used to determine the classification result. Applied to image super-resolution, the target neural network model of the embodiment of the present application can process the input sensor raw image and output a high-resolution image. For specific explanations of the method for generating the target neural network model in the embodiments of the present application, refer to the following embodiments.

FIG. 6 is a flowchart of a method for generating a target neural network model according to an embodiment of this application. As shown in FIG. 6, the method of this embodiment may be executed by the training device 220 or the processor of the training device 220 as shown in FIG. 2A. Executed, or may be executed by the server 420 or the processor of the server 420 as shown in FIG. 2B, and the method in this embodiment may include:

Step 101: Obtain score weights of multiple candidate operators and multiple neural network models.

Taking the candidate model shown in FIG. 3 as an example for illustration, step 101 can obtain the score weight of each candidate operator in the candidate model, for example, α _1,1 , α _1,2, etc. The scoring weights of the multiple candidate operators may be the scoring weights after the initialization operation, or the preset scoring weights, or they may be the last time the method for generating the target neural network model of the embodiment of this application is executed. The scoring weight in the candidate model. Based on each candidate operator in the candidate model, multiple neural network models can be determined. For example, a neural network model is the first candidate operator in the first selection block and the second candidate operation in the second selection block. The network model composed of the second candidate operator of the L-th selection block. In other words, the multiple neural network models may be neural network models obtained by sampling the candidate models as shown in FIG. 3.

Step 102: Determine the differentiable relationship information of the loss function with respect to the scoring variable according to the scoring weights of the multiple candidate operators, the accuracy rate of each neural network model, and the time delay of each neural network model, and the loss function is the accuracy rate And the joint function of the delay.

Exemplarily, the accuracy rate of each neural network model may be the accuracy rate of each neural network model in the above-mentioned multiple neural network models, and the accuracy rate of each neural network model may be the output and target of each neural network model. The Euclidean distance of the value. The delay of each neural network model can be the time required for each neural network model to propagate forward. The delay can be the delay of the neural network model measured on a real device. The real device can be as shown in Figure 2A. Execution equipment shown. _{According to the scoring weights of the multiple candidate operators and the accuracy of each neural network model, the differentiable relationship information of the accuracy function L acc with} respect to the scoring variable α can be determined. According to the scoring weights of the multiple candidate operators and the accuracy of each neural network model, The delay of a neural network model can determine the differentiable relationship information of the delay function T(α) with respect to the scoring variable α, and then the differentiable relationship information of the scoring variable α according to the accuracy function L _acc and the delay function T(α) The differentiable relationship information about the scoring variable α determines the differentiable relationship information of the loss function about the scoring variable α.

Step 103: Update the score weights of the multiple candidate operators according to the differentiable relationship information.

Through the above steps, the differentiable relationship information of the loss function on the score variable α can be obtained, and then the score weights of multiple candidate operators can be updated according to the differentiable relationship information to optimize the score weights of the candidate operators based on the loss function .

In some embodiments, each neural network model in the plurality of neural network models is determined by L candidate operators, and each candidate operator is one of the N candidate operators of the corresponding selection block, and the loss function The differentiable relationship information about the scoring variable includes L*N differentiable relationship information, and each differentiable relationship information corresponds to a candidate operator of a selection block. One possible implementation of the above step 103 is: according to the L*N differentiable relationship information, respectively update the score weight of the candidate operator corresponding to each differentiable relationship information; where L and N are any positive integers.

Step 104: Generate a target neural network model according to the updated scoring weight.

Take the candidate model shown in Figure 3 as an example for further illustration. Through the above steps, the score weights of the candidate operators in the candidate model can be updated, and then the operators can be selected in the updated candidate model to form the target. Neural network model.

Step 105: Send target model configuration information, where the target model configuration information is used to configure the target neural network model.

Taking the application scenario shown in FIG. 2A as an example, the training device 220 may send target model configuration information to the execution device 210, and the execution device 210 establishes a corresponding target neural network model according to the target model configuration information to use the target neural network model Realize corresponding data processing functions, such as face recognition.

Taking the application scenario shown in FIG. 2B as an example, the server 420 may send target model configuration information to the client device 410, and the client device 420 may establish a corresponding target neural network model according to the target model configuration information to use the target neural network model Realize corresponding data processing functions, such as face recognition. Alternatively, the client device 420 may configure the target neural network model on a corresponding execution device, such as an NPU.

The target model configuration information is used to configure each operator of the target neural network model and the connection relationship between the operators. You can also configure the internal parameters of each operator.

In this embodiment, by obtaining the scoring weights of multiple candidate operators and multiple neural network models, according to the scoring weights of the multiple candidate operators, the accuracy rate of each neural network model, and the time of each neural network model Determine the differentiable relationship information of the loss function on the scoring variable. The loss function is a joint function of accuracy and delay. The score weights of the multiple candidate operators are updated according to the differentiable relationship information, and the score weights are updated according to the updated scoring weights. The value generates the target neural network model, and sends the target model configuration information. The target model configuration information is used to configure the target neural network model. This embodiment determines the differentiable relationship information of the loss function with respect to the scoring variable, and optimizes the operation according to the differentiable relationship information The score weight of the symbol can improve the efficiency of generating the target neural network model and reduce the consumption of computing resources and time.

FIG. 7 is a flowchart of a method for generating a target neural network model according to an embodiment of this application. As shown in FIG. 7, the method of this embodiment can be executed by the training device 220 or the processor of the training device 220 as shown in FIG. 2A. Executed, or may be executed by the server 420 or the processor of the server 420 as shown in FIG. 2B, and the method in this embodiment may include:

Step 201: Obtain score weights, target data sets, and time delay adjustment parameters of multiple candidate operators.

For the explanation of obtaining the score weights of multiple candidate operators, refer to step 101 of the embodiment shown in FIG. 6, which will not be repeated here.

For example, the training device 220 may obtain the target data set and the delay adjustment parameters when receiving the model generation instruction. The model generation instruction may be sent by a terminal device, and the terminal device may be a client device as shown in FIG. 2A.

The target data set may include a plurality of training data, and the training data may include input data and corresponding target values. The target data set may be preset or input by the user. For example, the training device 220 receives the target data set sent by the terminal device. The delay adjustment parameter may be preset or input by the user. For example, the training device 220 receives the delay adjustment parameter sent by the terminal device.

For example, an achievable way is to receive the target data set and the delay adjustment parameter input by the user. For example, when there is no target data that meets the needs of the user in the database 230 as shown in FIG. 2A, the user may provide the target data set to the training device 220, so that the training device 220 obtains the target neural network model through the method of the embodiment of the present application. , For example, a neural network model for face recognition.

Another achievable way is to receive the data set selection information and time delay adjustment parameters input by the user, and determine the target data set according to the data set selection information. For example, there are multiple data sets in the database 230 as shown in FIG. 2A, and the training device 220 may use the data set corresponding to the data set selection information as the target data set according to the data set selection information input by the user.

Another achievable manner is to receive data set selection information and expected complexity information input by a user, determine the target data set according to the data set selection information, and determine the delay adjustment parameter according to the expected complexity information. For example, the expected complexity information can be in multiple levels such as high, medium, and low, and each level corresponds to a delay adjustment parameter. The training device 220 can determine the desired complexity information corresponding to the expected complexity information according to the expected complexity information input by the user. Delay adjustment parameters. For example, the higher the level, the smaller the value of the corresponding delay adjustment parameter.

Step 202: Determine a loss function according to the delay adjustment parameter, where the loss function is a joint function of accuracy and delay.

The joint function of the accuracy and the delay can be the sum of the accuracy function and the delay function, or the weighted sum of the accuracy function and the delay function. For example, the loss function is represented by L, and the accuracy function is represented by L _{Acc is} denoted, the time delay function is denoted by T, and the time delay adjustment parameter is denoted by β, and the loss function can be as the following formula (1).

L=L _acc +βT (1)

It should be noted that the loss function of the embodiment of the present application is differentiable with regard to the scoring weight variable. The specific form of the accuracy function L _acc can be flexibly set according to requirements, for example, it can be the Euclidean distance between the output of the neural network model and the target value.

Step 203: Update the internal parameters of the candidate operator according to the target data set.

For example, the training device 220 is provided with a candidate model as shown in FIG. 3. Before executing the method for generating the target neural network model in the embodiment of the present application, the internal parameters of each candidate operator in the candidate model can be initialized, and the training device 220 can select candidate operators in each layer to form a neural network model, and use the target data set to train the neural network model to update the internal parameters of each candidate operator constituting the neural network model .

Step 204: Determine multiple neural network models according to the updated internal parameters and the score weights of multiple candidate operators.

Taking the candidate model shown in Figure 3 as an example for further illustration, there are N candidate operators in each layer, and each candidate operator has a scoring weight. The training device 220 can be based on the candidate operators of each layer. For scoring weights, a plurality of candidate operators are selected from the candidate model to form a neural network model, and the internal parameters selected as the candidate operators of the neural network model may be the internal parameters updated through the above step 203. The training device 220 constructs multiple neural network models in the same manner.

Step 205: Determine the differentiable relationship information of the loss function with respect to the score variable according to the score weights of the multiple candidate operators, the accuracy rate of each neural network model, and the time delay of each neural network model.

The differentiable relationship information may be a derivative value.

The delay function of the embodiment of this application can be as the following formula (2)

Among them, θ _k represents a neural network model, and S is the traversal space formed by the multiple neural network models in step 104,

Represents the candidate operator selected by the lth layer of the neural network model θ _k , l takes 1 to L, and t(θ _k ) represents the time delay of the neural network model θ _k.

According to the score weights of multiple candidate operators and the delay of each neural network model, determine the differentiable relationship information of the delay function T(α) on the score variable α and the differentiability of the accuracy function L _acc on the score variable α Relationship information, and then according to formula (1), the differentiable relationship information of the loss function with respect to the score variable α can be determined.

An implementable way, the candidate model includes L selection blocks, each selection block includes N candidate operators, and the differentiable relationship information of the loss function on the score variable α includes L*N differentiable relationship information

The value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N), and each differentiable relation information corresponds to a selection block and a candidate operator. An implementation of the above step 205 is to determine the m-th accuracy differentiable relationship information of the l-th selection block according to the score weights of the N candidate operators of each selection block and the accuracy of each neural network model , The accuracy of each neural network model is determined according to the target data set. According to the scoring weights of the N candidate operators of each selection block, the sampling probability of the candidate operators of each selection block of each neural network model is determined (

The value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N) and the sampling probability of each candidate operator of each selection block is differentiable with respect to the score variable Relationship information

According to the sampling probability of the candidate operator of the L-1 selection block of each neural network model, the sampling probability of the m-th candidate operator of the l selection block, the differentiable relationship information about the scoring variable and each neural network model The delay determines the m-th delay differentiable relationship information of the l-th selection block, and the L-1 selection blocks include the selection blocks other than the l-th selection block among the L selection blocks. Determine the m-th differentiable relation information of the l-th selection block according to the m-th differentiable relation information of the accuracy rate of the l-th selection block and the m-th delay differentiable relation information of the l-th selection block. Among them, the value of l is a positive integer greater than or equal to 1 and less than or equal to L.

The above P(θ _k |α) can be expressed as the following formula (3).

According to formula (2) and the above formula (3) and formula (4), the _{derivative relationship of the delay function with respect to the scoring variable α l, m} is as follows: formula (5). It should be noted that the above formula (4) is _{an achievable way to determine p l, m} , and it may also be in other ways, and the embodiment of the present application is not limited thereto.

among them,

Candidate operators selected for the lth layer of the neural network model θ _k

The sampling probability.

The derivative relation of the accuracy rate function of the embodiment of the present application with respect to the scoring variable α _{l, m} can be as the following formula (6).

Among them, x _l represents the output of the selection block l (that is, the lth layer),

p _l,m represents the probability of selecting the m-th candidate operator in the l-th selection block.

Step 206: Update the score weights of multiple candidate operators according to the differentiable relationship information of the loss function on the score variable.

Through the above steps, the differentiable relationship information of the loss function on the score variable can be obtained, and then the score weights of multiple candidate operators can be updated according to the differentiable relationship information, so as to optimize the score weights of the candidate operators based on the loss function.

Step 207: Generate a target neural network model according to the updated scoring weight and the updated internal parameters.

Taking the candidate model shown in FIG. 3 as an example for further illustration, the internal parameters of the candidate operator in the candidate model and the score weight of the candidate operator can be updated through the above steps, and then the training device 220 can update In the latter candidate model, candidate operators are selected to form the target neural network model.

Step 208: Send target model configuration information, where the target model configuration information is used to configure the target neural network model.

For the explanation of step 208, refer to step 105 of the embodiment shown in FIG. 6, which will not be repeated here.

In this embodiment, by obtaining the scoring weights, target data sets, and delay adjustment parameters of multiple candidate operators, the loss function is determined according to the delay adjustment parameters. The loss function is a joint function of accuracy and delay. The data set updates the internal parameters of the candidate operators, determines multiple neural network models according to the updated internal parameters and the score weights of multiple candidate operators, and determines each neural network model according to the score weights of the multiple candidate operators The accuracy rate and the delay of each neural network model determine the differentiable relationship information of the loss function on the scoring variable, and update the score weights of multiple candidate operators according to the differentiable relationship information of the loss function on the scoring variable. The latter scoring weights and the updated internal parameters generate the target neural network model. In this embodiment, the differentiable relationship information of the loss function on the scoring variables is determined, and the scoring weights of the candidate operators are optimized according to the differentiable relationship information, which can improve The efficiency of generating the target neural network model reduces the consumption of computing resources and time.

FIG. 8 is a flowchart of another method for generating a target neural network model according to an embodiment of the application. As shown in FIG. 8, the method of this embodiment can be processed by the training device 220 or the training device 220 shown in FIG. 2A It may be executed by the server 420 or the processor of the server 420 as shown in FIG. 2B. The method in this embodiment may include:

Step 301: Obtain score weights, target data sets, and time delay adjustment parameters of multiple candidate operators.

Step 302: Determine a loss function according to the delay adjustment parameter, where the loss function is a joint function of accuracy and delay.

Wherein, the explanation of the above steps 301 to 302 can refer to step 201 and step 202 of the above embodiment, which will not be repeated here.

Step 303: Use the target data set to train the first neural network model, and update the internal parameters of the L candidate operators according to the training result.

The first neural network model is determined by L candidate operators, and each candidate operator is one of the N candidate operators of the corresponding selection block.

In some embodiments, each candidate operator in the candidate operator of each selection block has an equal probability of being selected as a candidate operator of the first neural network model.

Step 304: Determine multiple neural network models according to the updated internal parameters and the score weights of multiple candidate operators.

Step 305: Determine the differentiable relationship information of the loss function with respect to the scoring variable according to the scoring weight of the multiple candidate operators, the accuracy rate of each neural network model, and the time delay of each neural network model.

Step 306: Update the score weights of multiple candidate operators according to the differentiable relationship information.

Wherein, the explanations of the above steps 304 to 306 can refer to the steps 204 to 206 of the above embodiment, which will not be repeated here.

Step 307: Determine whether the first operator exists according to the updated scoring weight and the preset condition, if it exists, execute step 308, and if it does not exist, execute step 309.

The first operator is one or more of the candidate operators of the L selection blocks. The preset condition may be that the scoring weight is less than the preset threshold. When the updated scoring weight is less than the preset threshold, the candidate operator corresponding to the updated scoring weight is a first operator.

Step 308: Eliminate the first operator from the candidate operators of the selection block corresponding to the first operator.

The method of excluding the first operator may be to clear the internal parameters and the score weight of the first operator.

Step 309: It is judged whether the cut-off condition is met, if yes, step 310 is executed, and if not, step 303 is executed.

It should be noted that when the cut-off condition is not met and step 303 is executed, the L candidate operators constituting the first neural network model come from the remaining candidate operators after excluding the first operator.

Step 310: Generate a target neural network model according to the updated scoring weight and the updated internal parameters.

For example, in each selection block, the candidate operator with the largest scoring weight in the selection block is selected as a candidate operator of the target neural network model.

In this embodiment, by determining the differentiable relationship information of the loss function on the scoring variable, and optimizing the scoring weights of the candidate operators according to the differentiable relationship information, the efficiency of generating the target neural network model can be improved, and the consumption of computing resources and time can be reduced. In addition, by excluding candidate operators with smaller score weights from the candidate operators, the computational resources and time spent in obtaining the target neural model can be reduced.

Taking L=2 and N=4 as an example, the internal parameters of the update candidate operator and the scoring weight of the update candidate operator are exemplified.

The candidate model can be as shown in Fig. 9, input x _i , pass selection block 1, output x _j , pass selection block 2, output x _k . There are 4 candidate operators in selection block 1 and selection block 2, respectively, candidate operator 1, candidate operator 2, candidate operator 3, and candidate operator 4.

One optimization process for updating the internal parameters of candidate operators can be to randomly select a candidate operator from two selection blocks to form a neural network model, so as to select candidate operator 3 in selection block 1, and select in selection block 2. Candidate operator 2 is taken as an example to illustrate. The composed neural network model can be as shown in Figure 10, and the neural network model shown in Figure 10 can be trained using the target data set to update the candidate operator 3 of selection block 1. And select the internal parameters in the candidate operator 2 of block 2.

Taking the training device including an internal parameter optimization module, a differentiable relationship determination module, and a dynamic path decision module as an example, the scoring weights (α _1,1 , α _{1,2) of each candidate operator in the selection block 1 and the selection block 2} , Α _1,3 and α _1,4 , α _2,1 , α _2,2 , α _2,3 and α _2,4 ) are updated for illustration, as shown in Figure 11A and Figure 11B, the target neural network The model generation device can include an internal parameter optimization module, a differentiable relationship determination module, and a dynamic path decision module. The internal parameter optimization module can update the candidate operators of the selection block 1 and the candidate operators of the selection block 2 in the manner described above. The differentiable relationship determination module sends the neural network model configuration information to the speed measuring device, so that the speed measuring device restores the corresponding neural network model θ _k , and the speed measuring device measures the time delay t(θ _k ) of the _{neural network model θ k,} The time delay t(θ _k ) is fed back to the differentiable relationship determination module. The differentiable relationship determination module is based on the time delay t(θ _k ) of each neural network model, the accuracy of each neural network model, and each neural network model. The scoring weights of the candidate operators and the scoring weights of each candidate operator, use formula (3) to formula (5) to obtain the delay function about α _1,1 , α _1,2 , α _1,3 , α _{1, 4.} Derivative values of α _2,1 , α _2,2 , α _2,3 and α _2,4 , use formula (6) to obtain the accuracy function with respect to α _1,1 , α _1,2 , α _1,3 , Derivative values of α _1,4 , α _2,1 , α _2,2 , α _2,3 and α _2,4 , according to formula (1), the loss function is obtained according to α _1,1 , α _1,2 , α _{1 ,3} , α _1,4 , α _2,1 , α _2,2 , α _2,3 and α _2,4 are the derivative values, and the loss function is related to α _1,1 , α _1,2 , α _1,3 The derivative values of, α _1,4 , α _2,1 , α _2,2 , α _2,3 and α _2,4 are provided to the dynamic path decision module, and the dynamic path decision module converts α _1,1 , α _1,2 , α _1,3 , α _1,4 , α _2,1 , α _2,2 , α _2,3 and α _2,4 , and the loss function about α _1,1 , α _1,2 , α _1,3 , α _{The derivative values of 1,4} , α _2,1 , α _2,2 , α _2,3, and α _2,4 are input into the stochastic gradient descent algorithm to optimize the scoring weight of each candidate operator. For example, the optimized score weights are α′ _1,1 , α′ _1,2 , α′ _1,3 , α′ _1,4 , α′ _2,1 , α′ _2,2 , α′ _2,3 And α′ _2,4 . It should be noted that the device for generating the target neural network model may be the training device 220 or the processor of the training device 220 as shown in FIG. 2A, or the server 420 or the processor of the server 420 as shown in FIG. 2B.

In this embodiment, by determining the derivative value of the scoring variable of each candidate operator of the loss function, the scoring weight of each candidate operator in the selection block 1 and the selection block 2 is optimized, and the optimized scoring weight is minimized at the same time The accuracy error and the average network delay are reduced, thereby improving the accuracy of the target neural network model constructed based on the candidate model and reducing the network delay.

FIG. 12 is a flowchart of another method for generating a target neural network model according to an embodiment of the application. As shown in FIG. 12, the embodiment of the application relates to a device for generating a target neural network model and a speed measuring device. The device for generating the target neural network model may be the processor of the training device 220 or the training device 220 as shown in FIG. 2A, or the server 420 or the processor of the server 420 as shown in FIG. 2B. As shown in FIG. 11A, the speed measurement device may be a server or an internal chip of a server, or a terminal device or an internal chip of a terminal device. For example, the terminal device may be a wireless communication device or an Internet of Things (IoT) device. , Wearable devices or vehicle-mounted devices, mobile terminals, Customer Premise Equipment (CPE), etc. On the basis of any of the foregoing embodiments, the method for generating the target neural network model of the embodiment of the present application may further include:

Step 401: The generating device of the target neural network model obtains the identification information of the speed measuring device.

The identification information is used to identify the speed measurement device, and the identification information may be input by the user. For example, the user provides the identification information to the generating device of the target neural network model through the client device 240. The identification information may also be determined by the device for generating the target neural network model according to the target training set. For example, if the target training set is a training set for vehicle detection based on machine vision, the device for generating the target neural network model may be determined according to the target training set. In the application scenario, it is determined that the identification information of the speed measuring device is the identification information of the vehicle-mounted equipment.

The identification information of the speed measurement device can be the host name (Hostname), Media Access Control (MAC) address, IP address (Internet Protocol Address), user agent (user agent) information, domain name system (Domain Name System, DNS) ) Any one or a combination of information, International Mobile Subscriber Identity (IMSI).

Step 402: The device for generating the target neural network model sends the neural network model configuration information to the speed measuring device according to the identification information.

The neural network model configuration information is used to configure multiple neural network models. For example, the neural network model configuration information is used to configure multiple neural network models in step 101 of the embodiment shown in FIG. 6.

Exemplarily, the neural network model configuration information may include a plurality of network structure coding information, and each network structure coding information is used to configure a connection relationship between various operators and operators of a neural network model. For example, the network structure coding information of the neural network model shown in Fig. 10 can be 013 122, where 013 indicates that _{the operator between input x i} (node 0) to x _j (node 1) is candidate operator 3. 122 indicates that _{the operator between x j} (node 1) and output x _k (node 2) is candidate operator 2. For another example, the network structure coding information of the neural network model shown in Figure 10 can be 0010 0100, where 0010 corresponds to each candidate operator in selection block 1, 0 means not selected, 1 means selected, and 0100 corresponds to selection For each candidate operator in block 2, 0 means not selected, and 1 means selected.

It should be noted that the delay of the neural network model is usually related to the structure of the neural network model, that is, it is related to the connection relationship between the various operators of the neural network model and the operators. For the internal parameters of each operator, random initialization parameters can be used. Preset parameters can also be used, which can be flexibly set according to requirements.

Step 403: The speed measuring device restores multiple target neural network models according to the neural network model configuration information, and measures the time delay of each target neural network model.

For example, the speed measuring device can restore multiple neural network models in step 101 of the embodiment shown in FIG. 6 according to the neural network model configuration information. And measure the time delay of each target neural network model.

Step 404: The speed measuring device sends the time delay of each neural network model to the generating device of the target neural network model.

Correspondingly, the target neural network model generating device receives the time delay of each neural network model sent by the speed measuring device. The time delay of each neural network model may include model identification information and time delay t(θ _k ), and the model identification information is used to represent the neural network model θ _k .

The target neural network model generation device can determine the differentiable relationship information of the loss function with respect to the scoring variable through the above step 102, and then use the differentiable relationship information to optimize the score weight of each candidate operator to generate the target based on the optimized score weight. Neural network model.

Step 404', the speed measuring device determines the differentiable relationship information of the delay function on the score variable according to the score weights of the multiple candidate operators of each neural network model and the delay of each neural network model, and the speed measuring device sends the target neural network to the target neural network. The model generating device sends the differentiable relationship information of the delay function on the scoring variable.

Correspondingly, the target neural network model generating device receives the differentiable relationship information of the time delay function on the scoring variable sent by the speed measuring device, and the target neural network model generating device determines the loss function according to the differentiable relationship information of the delay function on the scoring variable The differentiable relationship information of the scoring variable is then used to optimize the scoring weight of each candidate operator to generate a target neural network model based on the optimized scoring weight.

Wherein, step 404 or step 404' can be executed after step 403.

In this embodiment, by measuring the time delay of each neural network model on the speed measuring device, the differentiable relationship information of the loss function with respect to the scoring variable is determined based on the measurement result of the speed measuring device, and then the differentiable relationship information is used to optimize each candidate operation To generate the target neural network model based on the optimized scoring weight based on the score weight of the symbol, the efficiency of generating the target neural network model can be improved, and the consumption of computing resources and time can be reduced.

FIG. 13 is a diagram of a chip hardware structure provided by an embodiment of the application. The algorithm based on the target neural network model of the embodiment of the present application can be implemented in the NPU chip shown in FIG. 13, and the target neural network model can be the target neural network model obtained by the above method embodiment.

Neural Network Processor NPU 50 NPU is mounted on the main CPU (Host CPU) as a coprocessor, and the Host CPU distributes tasks. The core part of the NPU is the arithmetic circuit 50. The arithmetic circuit 503 is controlled by the controller 504 to extract matrix data from the memory and perform multiplication operations.

In some implementations, the arithmetic circuit 503 includes multiple processing units (Process Engine, PE). In some implementations, the arithmetic circuit 503 is a two-dimensional systolic array. The arithmetic circuit 503 may also be a one-dimensional systolic array or other electronic circuit capable of performing mathematical operations such as multiplication and addition. In some implementations, the arithmetic circuit 503 is a general-purpose matrix processor.

For example, suppose there is an input matrix A, a weight matrix B, and an output matrix C. The arithmetic circuit fetches the corresponding data of matrix B from the weight memory 502 and caches it on each PE in the arithmetic circuit. The arithmetic circuit takes the matrix A data and matrix B from the input memory 501 to perform matrix operations, and the partial or final result of the obtained matrix is stored in the accumulator 508.

The unified memory 506 is used to store input data and output data. The weight data directly passes through the storage unit to access the controller 505 Direct Memory Access Controller, and the DMAC is transferred to the weight memory 502. The input data is also transferred to the unified memory 506 through the DMAC.

The BIU is the Bus Interface Unit, that is, the bus interface unit 510, which is used for the interaction between the AXI bus and the DMAC and the instruction fetch buffer 509.

The bus interface unit 510 (Bus Interface Unit, BIU for short) is used for the instruction fetch memory 509 to obtain instructions from an external memory, and is also used for the storage unit access controller 505 to obtain the original data of the input matrix A or the weight matrix B from the external memory.

The DMAC is mainly used to transfer the input data in the external memory DDR to the unified memory 506 or to transfer the weight data to the weight memory 502 or to transfer the input data to the input memory 501.

The vector calculation unit 507 has multiple arithmetic processing units, if necessary, further processing the output of the arithmetic circuit, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison and so on. Mainly used for non-convolution/FC layer network calculations in neural networks, such as Pooling, Batch Normalization, Local Response Normalization, etc.

In some implementations, the vector calculation unit 507 can store the processed output vector in the unified buffer 506. For example, the vector calculation unit 507 may apply a nonlinear function to the output of the arithmetic circuit 503, such as a vector of accumulated value, to generate an activation value. In some implementations, the vector calculation unit 507 generates a normalized value, a combined value, or both. In some implementations, the processed output vector can be used as an activation input to the arithmetic circuit 503, for example for use in a subsequent layer in a neural network.

The instruction fetch buffer 509 connected to the controller 504 is used to store instructions used by the controller 504;

The unified memory 506, the input memory 501, the weight memory 502, and the instruction fetch memory 509 are all On-Chip memories. The external memory is private to the NPU hardware architecture.

Among them, the calculation of each layer in the target neural network model can be executed by the vector calculation unit 507.

The target neural network model obtained in the foregoing embodiment of the present application can be applied to a server, a wearable device, a vehicle, or an automatic driving system, etc., to process input data and output a result based on the target neural network model.

Referring to FIG. 14, an embodiment of the present application provides a system architecture 300. The execution device 210 is implemented by one or more servers, and optionally, it cooperates with other computing devices, such as data storage, routers, load balancers and other devices; the execution device 210 can be arranged on one physical site or distributed in multiple On the physical site. The execution device 210 can use the data in the data storage system 250, or call the program code in the data storage system 250 to implement the processing method based on the above-mentioned target neural network model, and provide services for users.

The user can operate respective user devices (for example, the local device 301 and the local device 302) to interact with the execution device 210. Each local device can represent any computing device, such as personal computers, computer workstations, smart phones, tablets, smart cameras, smart cars or other types of cellular phones, media consumption devices, wearable devices, set-top boxes, game consoles, etc.

The local device of each user can interact with the execution device 210 through a communication network of any communication mechanism/communication standard. The communication network can be a wide area network, a local area network, a point-to-point connection, etc., or any combination thereof.

In another implementation, one or more aspects of the execution device 210 may be implemented by each local device. For example, the local device 301 may provide the execution device 210 with local data or feed back calculation results.

It should be noted that all functions of the execution device 210 can also be implemented by a local device. For example, the local device 301 implements the functions of the device 210 and provides services for its own users, or provides services for users of the local device 302.

Referring to FIG. 15, FIG. 15 is a schematic block diagram of a device 1600 for generating a target neural network model provided by this application. The generating device 1600 of the communication target neural network model includes an acquisition module 1601, a model generation module 1602, and a transceiver module 1602.

In an embodiment, the device 1600 for generating a target neural network model has the function of a training device or a server in the method embodiment. For example, the device 1600 for generating a target neural network model may execute the method of any one of the embodiments in FIGS. 6 to 8, or execute the method executed by the device for generating a target neural network model in the embodiment of FIG. 12. At this time, the units of the communication device 500 are respectively used to perform the following operations and/or processing.

The obtaining module 1601 is used to obtain the scoring weights of multiple candidate operators and multiple neural network models.

The model generation module 1602 is used to determine the differentiable relationship information of the loss function with respect to the scoring variables according to the scoring weights of the multiple candidate operators, the accuracy of each neural network model, and the time delay of each neural network model. The loss function is a joint function of accuracy and time delay.

The model generation module 1602 is further configured to update the score weights of the multiple candidate operators according to the differentiable relationship information.

The model generation module 1602 is also used to generate a target neural network model according to the updated scoring weight.

The transceiver module 1603 is configured to send target model configuration information, where the target model configuration information is used to configure the target neural network model

In some embodiments, each neural network model of the plurality of neural network models is determined by L candidate operators, and each candidate operator is one of the N candidate operators of the corresponding selection block, and The differentiable relationship information of the loss function on the scoring variable includes L*N differentiable relationship information, and each differentiable relationship information corresponds to a candidate operator of a selection block; the model generation module 1602 is used to: according to the L* N differentiable relation information, respectively update the score weight of the candidate operator corresponding to each differentiable relation information; where L and N are any positive integers.

In some embodiments, the acquisition module 1601 is further configured to: acquire a target data set and a delay adjustment parameter; determine the loss function according to the delay adjustment parameter; update the candidate operator according to the target data set The internal parameters of; the multiple neural network models are determined according to the updated internal parameters and the score weights of the multiple candidate operators.

In some embodiments, the model generation module 1602 is configured to determine the mth accuracy of the lth selection block according to the score weights of the N candidate operators of each selection block and the accuracy rate of each neural network model. Rate differentiable relationship information, the accuracy rate of each neural network model is determined according to the target data set; each neural network model is determined according to the score weight of the N candidate operators of each selection block The sampling probability of the operator of the selection block and the sampling probability of each candidate operator of each selection block. Differentiable relationship information about the scoring variable; according to the sampling probability of the L-1 selection block operators of each neural network model , The sampling probability of the m-th candidate operator of the l-th selection block and the differentiable relationship information of the scoring variable and the delay of each neural network model determine the m-th delay of the l-th selection block Micro-relationship information, the L-1 selection blocks include selection blocks other than the l-th selection block among the L selection blocks; and the m-th accuracy rate of the l-th selection block may be differentiable relation information And the m-th time delay differentiable relationship information of the l-th selection block to determine the m-th differentiable relationship information of the l-th selection block. Among them, the value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N.

In some embodiments, the model generation module 1602 is configured to: respectively input the L*N differentiable relationship information and the score weight of each candidate operator into the stochastic gradient descent algorithm, and output each updated The scoring weight of each candidate operator in the selection block.

In some embodiments, the acquisition module 1601 is configured to: use the target data set to train a first neural network model, the first neural network model is determined by L candidate operators, and each candidate operator is a corresponding Select one of the N candidate operators of the block; update the internal parameters of the L candidate operators according to the training result; where L is any positive integer.

In some embodiments, the model generation module 1602 is further configured to determine whether there is a first operator based on the updated scoring weight and a preset condition, and the first operator is the value of the L selection blocks. One or more of the candidate operators; if there is a first operator, remove the first operator from the candidate operators of the selection block corresponding to the first operator.

In some embodiments, the model generation module 1602 is further configured to determine whether a cut-off condition is satisfied after the score weights of the multiple candidate operators are updated according to the differentiable relationship information, if the cut-off condition is not satisfied , Then execute the step of training the first neural network model using the target data set; if the cut-off condition is met, execute the step of generating the target neural network model according to the updated scoring weight.

In some embodiments, each candidate operator of each selection block has an equal probability of being selected as an operator of the first neural network model.

In some embodiments, the acquisition module 1601 is further configured to: acquire identification information of the speed measurement device; according to the identification information, send neural network model configuration information to the speed measurement device through the transceiver module 1603, and the neural network The model configuration information is used to configure the multiple neural network models; the transceiver module 1603 is also used to receive the time delay of each neural network model sent by the speed measurement device.

In some embodiments, the neural network model configuration information includes a plurality of network structure coding information, and each network structure coding information is used to configure a connection relationship between various operators of a neural network model and the operators.

In some embodiments, the model generation module 1602 is configured to select, in each selection block, the candidate operator with the largest score in the selection block as an operator of the target neural network model, and the operation The internal parameters of the symbol are the updated internal parameters.

In some embodiments, the acquiring module 1601 is configured to: receive the target data set and the delay adjustment parameter input by the user through the transceiver module 1603; or, receive the data input by the user through the transceiver module 1603 The target data set is determined according to the data set selection information based on the set selection information and the delay adjustment parameters; or, the data set selection information and the expected complexity information input by the user are received through the transceiver module 1603, and the The data set selection information determines the target data set, and the delay adjustment parameter is determined according to the expected complexity information.

Optionally, the device 1600 for generating the target neural network model may also have other functions in the method embodiment at the same time. For similar description, reference may be made to the description of the foregoing method embodiment. To avoid repetition, I won’t repeat them here.

Optionally, the acquisition module 1601 and the model generation model 1602 may be a processor, and the transceiver module 1603 may be a transceiver. The transceiver includes a receiver and a transmitter, and has both sending and receiving functions.

Optionally, the acquisition module 1601 and the model generation model 1602 may be a processing device, and the functions of the processing device may be partially or fully implemented by software.

In a possible implementation manner, the functions of the processing device may be partially or fully implemented by software. At this time, the processing device may include a memory and a processor. The memory is used to store a computer program, and the processor reads and executes the computer program stored in the memory to execute the steps in each method embodiment.

Optionally, in a possible implementation manner, the processing device includes a processor. The memory for storing the computer program is located outside the processing device, and the processor is connected to the memory through a circuit/wire to read and execute the computer program stored in the memory.

In a possible implementation manner, the functions of the processing device may all be implemented by hardware. At this time, the processing device may include an input interface circuit, a logic circuit, and an output interface circuit. The input interface circuit is used to obtain the score weights of multiple candidate operators and multiple neural network models; the logic circuit is used to obtain the score weights of the multiple candidate operators and the accuracy of each neural network model. The rate and the delay of each neural network model determine the differentiable relationship information of the loss function with respect to the scoring variable, and the loss function is a joint function of the accuracy rate and the delay; the multiple candidate operations are updated according to the differentiable relationship information The target neural network model is generated according to the updated score weight; the output interface circuit is used to send target model configuration information, and the target model configuration information is used to configure the target neural network model.

In another embodiment, the device 1600 for generating the target neural network model may be a chip. At this time, the transceiver module 1601 may specifically be a communication interface or a transceiver circuit.

Refer to FIG. 16, which is a schematic structural diagram of an electronic device 1700 provided by this application. As shown in FIG. 16, the electronic device 1700 includes a processor 1701 and a transceiver 1702.

Optionally, the electronic device 1700 further includes a memory 1703. Among them, the processor 1701, the transceiver 1702, and the memory 1703 can communicate with each other through an internal connection path to transfer control signals and/or data signals.

Among them, the memory 1703 is used to store computer programs. The processor 1701 is configured to execute the computer program stored in the memory 1703, so as to realize the functions of the device 1600 for generating the target neural network model in the foregoing device embodiment.

Specifically, the processor 1701 may be used to perform the operations and/or processing performed by the acquisition module 1601 and the model generation model 1602 described in the apparatus embodiment (for example, FIG. 16), and the transceiver 1702 is used to perform the operations and/or processing performed by the transceiver module 1603. Perform operations and/or processing.

Optionally, the memory 1703 may also be integrated in the processor 1701 or independent of the processor 1701.

The electronic device of this embodiment can execute the method for generating the target neural network model of the foregoing method embodiment, and its technical principles and technical effects can be referred to the explanation of the foregoing embodiment, which will not be repeated here.

The present application also provides a computer-readable storage medium with a computer program stored on the computer-readable storage medium. When the computer program is executed by a computer, the computer executes the steps and/or processing in any of the above-mentioned method embodiments. .

This application also provides a computer program product. The computer program product includes computer program code. When the computer program code runs on a computer, the computer executes the steps and/or processing in any of the foregoing method embodiments.

The application also provides a chip including a processor. The memory for storing the computer program is provided independently of the chip, and the processor is used to execute the computer program stored in the memory to execute the steps and/or processing in any method embodiment.

Further, the chip may also include a memory and a communication interface. The communication interface may be an input/output interface, a pin, an input/output circuit, or the like.

The processor mentioned in the above embodiments may be an integrated circuit chip with signal processing capability. In the implementation process, the steps of the foregoing method embodiments can be completed by hardware integrated logic circuits in the processor or instructions in the form of software. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other Programming logic devices, discrete gates or transistor logic devices, discrete hardware components. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor or the like. The steps of the method disclosed in the embodiments of the present application may be directly embodied as being executed and completed by a hardware encoding processor, or executed and completed by a combination of hardware and software modules in the encoding processor. The software module can be located in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

The memory mentioned in the above embodiments may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Among them, the non-volatile memory can be read-only memory (ROM), programmable read-only memory (programmable ROM, PROM), erasable programmable read-only memory (erasable PROM, EPROM), and electrically available Erase programmable read-only memory (electrically EPROM, EEPROM) or flash memory. The volatile memory may be random access memory (RAM), which is used as an external cache. By way of exemplary but not restrictive description, many forms of RAM are available, such as static random access memory (static RAM, SRAM), dynamic random access memory (dynamic RAM, DRAM), and synchronous dynamic random access memory (synchronous DRAM, 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 connection dynamic random access memory (synchlink DRAM, SLDRAM) ) And direct memory bus random access memory (direct rambus RAM, DR RAM). It should be noted that the memories of the systems and methods described herein are intended to include, but are not limited to, these and any other suitable types of memories.

A person of ordinary skill in the art may realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed herein can be implemented by electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered beyond the scope of this application.

Those skilled in the art can clearly understand that, for the convenience and conciseness of description, the specific working process of the system, device and unit described above can refer to the corresponding process in the foregoing method embodiment, which will not be repeated here.

In the several embodiments provided in this application, it should be understood that the disclosed system, device, and method may be implemented in other ways. For example, the device embodiments described above are merely illustrative, for example, the division of the units is only a logical function division, and there may be other divisions in actual implementation, for example, multiple units or components may be combined or It can be integrated into another system, or some features can be ignored or not implemented. In addition, the displayed or discussed mutual coupling or direct coupling or communication connection may be indirect coupling or communication connection through some interfaces, devices or units, and may be in electrical, mechanical or other forms.

The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.

In addition, the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist alone physically, or two or more units may be integrated into one unit.

If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer readable storage medium. Based on this understanding, the technical solution of the present application essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including Several instructions are used to make a computer device (personal computer, server, or network device, etc.) execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage media include: U disk, mobile hard disk, read-only memory (read-only memory, ROM), random access memory (random access memory, RAM), magnetic disks or optical disks and other media that can store program codes. .

Claims (30)

1. A method for generating a target neural network model, which is characterized in that it includes:

   Obtain score weights of multiple candidate operators and multiple neural network models;

   According to the scoring weights of the multiple candidate operators, the accuracy of each neural network model, and the time delay of each neural network model, the differentiable relationship information of the loss function with respect to the scoring variable is determined, and the loss function is the accuracy and Joint function of time delay;

   Updating the score weights of the multiple candidate operators according to the differentiable relationship information;

   Generate the target neural network model according to the updated scoring weight;

   Send target model configuration information, where the target model configuration information is used to configure the target neural network model.
2. The method according to claim 1, wherein each neural network model in the plurality of neural network models is determined by L candidate operators, and each candidate operator is N candidate operations of the corresponding selection block One of the differentiable relationship information of the scoring variable of the loss function includes L*N differentiable relationship information, and each differentiable relationship information corresponds to a candidate operator of a selection block;

   The updating the score weights of the multiple candidate operators according to the differentiable relationship information includes:

   According to the L*N differentiable relationship information, respectively update the score weight of the candidate operator corresponding to each differentiable relationship information;

   Among them, L and N are arbitrary positive integers.
3. The method according to claim 2, wherein the method further comprises:

   Obtain target data set and time delay adjustment parameters;

   Determine the loss function according to the delay adjustment parameter;

   Update the internal parameters of the candidate operator according to the target data set;

   The multiple neural network models are determined according to the updated internal parameters and the score weights of the multiple candidate operators.
4. The method according to claim 3, wherein the loss function is determined according to the scoring weights of the multiple candidate operators, the accuracy rate of each neural network model, and the delay of each neural network model Information about the differentiable relationship of scoring variables, including:

   According to the scoring weights of the N candidate operators of each selection block and the accuracy of each neural network model, determine the m-th accuracy differentiable relationship information of the l selection block, and the accuracy of each neural network model The rate is determined according to the target data set;

   Determine the sampling probability of the operator of each selection block of each neural network model and the sampling probability of each candidate operator of each selection block with respect to the score variable according to the score weight of the N candidate operators of each selection block Differentiable relationship information;

   According to the sampling probability of the operator of the L-1 selection block of each neural network model, the sampling probability of the m-th candidate operator of the l selection block, the differentiable relationship information about the scoring variable and each neural network The delay of the model determines the m-th delay differentiable relation information of the l-th selection block, and the L-1 selection blocks include the selection blocks of the L selection blocks other than the l-th selection block ；

   According to the m-th accuracy differentiable relation information of the l-th selection block and the m-th delay differentiable relation information of the l-th selection block, determine the m-th possibility of the l-th selection block Micro-relationship information;

   Among them, the value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N.
5. The method according to claim 4, wherein the updating the score weight of the candidate operator corresponding to each differentiable relationship information according to the L*N differentiable relationship information includes:

   The L*N differentiable relationship information and the scoring weight of each candidate operator are respectively input to the stochastic gradient descent algorithm, and the updated scoring weight of each candidate operator in each selection block is output.
6. The method according to any one of claims 3 to 5, wherein the updating the internal parameters of the candidate operator according to the target data set comprises:

   Training a first neural network model using the target data set, the first neural network model being determined by L candidate operators, each candidate operator being one of the N candidate operators of the corresponding selection block;

   Update the internal parameters of the L candidate operators according to the training result;

   Among them, L is any positive integer.
7. The method according to claim 6, wherein the method further comprises:

   Determine whether there is a first operator according to the updated scoring weight and the preset condition, where the first operator is one or more of the candidate operators of the L selection blocks;

   If there is a first operator, remove the first operator from the candidate operators of the selection block corresponding to the first operator.
8. 8. The method according to claim 7, wherein after the updating the score weights of the multiple candidate operators according to the differentiable relationship information, the method further comprises:

   Determine whether the cut-off condition is met, and if the cut-off condition is not met, execute the step of using the target data set to train the first neural network model;

   If the cut-off condition is met, the step of generating the target neural network model according to the updated scoring weight is executed.
9. The method according to any one of claims 6 to 8, wherein each candidate operator of each selection block has an equal probability of being selected as an operator of the first neural network model.
10. The method according to any one of claims 1 to 9, wherein the method further comprises:

    Obtain identification information of the speed measuring device;

    Sending neural network model configuration information to the speed measurement device according to the identification information, where the neural network model configuration information is used to configure the multiple neural network models;

    The time delay of each neural network model sent by the speed measuring device is received.
11. The method according to claim 10, wherein the neural network model configuration information includes a plurality of network structure coding information, and each network structure coding information is used to configure each operator of a neural network model and the connection of the operators relationship.
12. The method according to any one of claims 3 to 11, wherein the generating a target neural network model according to the updated scoring weight value comprises:

    In each selection block, the candidate operator with the largest scoring weight in the selection block is selected as an operator of the target neural network model, and the internal parameter of the operator is the updated internal parameter.
13. The method according to any one of claims 3 to 12, wherein the obtaining the target data set and the delay adjustment parameter comprises:

    Receiving the target data set and the delay adjustment parameter input by the user; or,

    Receive the data set selection information and the time delay adjustment parameter input by the user, and determine the target data set according to the data set selection information; or,

    Receive data set selection information and expected complexity information input by a user, determine the target data set according to the data set selection information, and determine the delay adjustment parameter according to the expected complexity information.
14. A device for generating a target neural network model is characterized in that it comprises:

    The obtaining module is used to obtain the scoring weights of multiple candidate operators and multiple neural network models;

    The model generation module is used to determine the differentiable relationship information of the loss function with respect to the scoring variable according to the scoring weights of the multiple candidate operators, the accuracy of each neural network model, and the time delay of each neural network model. The loss function is a joint function of accuracy and delay;

    The model generation module is further configured to update the score weights of the multiple candidate operators according to the differentiable relationship information;

    The model generation module is also used to generate a target neural network model according to the updated scoring weight;

    The transceiver module is used to send target model configuration information, and the target model configuration information is used to configure the target neural network model.
15. The device according to claim 14, wherein each neural network model in the plurality of neural network models is determined by L candidate operators, and each candidate operator is N candidate operations of the corresponding selection block One of the differentiable relationship information of the scoring variable of the loss function includes L*N differentiable relationship information, and each differentiable relationship information corresponds to a candidate operator of a selection block;

    The model generation module is used to update the score weight of the candidate operator corresponding to each differentiable relationship information according to the L*N differentiable relationship information; where L and N are any positive integers.
16. The device according to claim 15, wherein the acquisition module is further configured to:

    Obtain target data set and time delay adjustment parameters;

    Determine the loss function according to the delay adjustment parameter;

    Update the internal parameters of the candidate operator according to the target data set;

    The multiple neural network models are determined according to the updated internal parameters and the score weights of the multiple candidate operators.
17. The device according to claim 16, wherein the model generation module is used for:

    According to the scoring weights of the N candidate operators of each selection block and the accuracy of each neural network model, determine the m-th accuracy differentiable relationship information of the l selection block, and the accuracy of each neural network model The rate is determined according to the target data set;

    Determine the sampling probability of the operator of each selection block of each neural network model and the sampling probability of each candidate operator of each selection block with respect to the score variable according to the score weight of the N candidate operators of each selection block Differentiable relationship information;

    According to the sampling probability of the operator of the L-1 selection block of each neural network model, the sampling probability of the m-th candidate operator of the l selection block, the differentiable relationship information about the scoring variable and each neural network The delay of the model determines the m-th delay differentiable relation information of the l-th selection block, and the L-1 selection blocks include the selection blocks of the L selection blocks other than the l-th selection block ；

    According to the m-th accuracy differentiable relation information of the l-th selection block and the m-th delay differentiable relation information of the l-th selection block, determine the m-th possibility of the l-th selection block Micro-relationship information;

    Among them, the value of l is a positive integer greater than or equal to 1 and less than or equal to L, and the value of m is a positive integer greater than or equal to 1 and less than or equal to N.
18. The device according to claim 17, wherein the model generation module is used for:

    The L*N differentiable relationship information and the scoring weight of each candidate operator are respectively input to the stochastic gradient descent algorithm, and the updated scoring weight of each candidate operator in each selection block is output.
19. The device according to any one of claims 16 to 18, wherein the acquisition module is configured to: use the target data set to train a first neural network model, and the first neural network model is operated by L candidates Each candidate operator is one of the N candidate operators of the corresponding selection block; according to the training result, the internal parameters of the L candidate operators are updated; where L is any positive integer.
20. The device according to claim 19, wherein the model generation module is further configured to:

    Determine whether there is a first operator according to the updated scoring weight and the preset condition, where the first operator is one or more of the candidate operators of the L selection blocks;

    If there is a first operator, remove the first operator from the candidate operators of the selection block corresponding to the first operator.
21. The device according to claim 20, wherein the model generation module is further configured to determine whether the cut-off condition is satisfied after the score weights of the multiple candidate operators are updated according to the differentiable relationship information , If the cut-off condition is not met, execute the step of using the target data set to train the first neural network model;

    If the cut-off condition is met, the step of generating the target neural network model according to the updated scoring weight is executed.
22. The device according to any one of claims 19 to 21, wherein each candidate operator of each selection block has an equal probability of being selected as an operator of the first neural network model.
23. The device according to any one of claims 14 to 21, wherein the acquisition module is further configured to: acquire identification information of the speed measurement device;

    Sending neural network model configuration information to the speed measuring device through the transceiver module according to the identification information, where the neural network model configuration information is used to configure the multiple neural network models;

    The transceiver module is also used to receive the time delay of each neural network model sent by the speed measurement device.
24. The device according to claim 23, wherein the neural network model configuration information includes a plurality of network structure coding information, and each network structure coding information is used to configure each operator of a neural network model and the connection of the operators relationship.
25. The device according to any one of claims 16 to 24, wherein the model generation module is configured to:

    In each selection block, the candidate operator with the largest scoring weight in the selection block is selected as an operator of the target neural network model, and the internal parameter of the operator is the updated internal parameter.
26. The apparatus according to any one of claims 16 to 25, wherein the acquisition module is configured to: receive the target data set and the delay adjustment parameter input by the user through the transceiver module; or,

    Receive the data set selection information and the time delay adjustment parameter input by the user through the transceiver module, and determine the target data set according to the data set selection information; or,

    The data set selection information and expected complexity information input by the user are received through the transceiver module, the target data set is determined according to the data set selection information, and the delay adjustment parameter is determined according to the expected complexity information.
27. An electronic device, characterized in that it comprises:

    One or more processors;

    Memory, used to store one or more programs;

    When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1-13.
28. A computer-readable storage medium, characterized by comprising a computer program, which when executed on a computer, causes the computer to execute the method according to any one of claims 1-13.
29. A computer program product, characterized in that, the computer program product includes instructions that, when running on a computer, cause the computer to execute the method according to any one of claims 1-13.
30. A chip, characterized by comprising a processor and a memory, the memory is used to store a computer program, and the processor is used to call and run the computer program stored in the memory to execute any one of claims 1-13. The method described in one item.

Images