WO2021184836A1

WO2021184836A1 - Method and apparatus for training recognition model, device, and readable storage medium

Info

Publication number: WO2021184836A1
Application number: PCT/CN2020/134029
Authority: WO
Inventors: 姜迪
Original assignee: 深圳前海微众银行股份有限公司
Priority date: 2020-03-20
Filing date: 2020-12-04
Publication date: 2021-09-23
Also published as: CN111428881A; CN111428881B

Abstract

A method and apparatus for training a recognition model, a device, and a readable storage medium, which relate to the field of financial technology. The method comprises the steps of: acquiring local training data, and training according to the training data to obtain a first recognition model (S10); sending model parameters of the first recognition model to a server for the server to perform federated learning according to model parameters sent by each client so as to obtain a federated learning result (S20); and receiving the federated learning result sent by the server, updating the first recognition model according to the federated learning result to obtain a corresponding second recognition model, and sending the second recognition model to the server, so that the server obtains a target recognition model according to the second recognition model sent by each client and a preset genetic algorithm (S30). In the method, migration learning, federated learning and a genetic algorithm are combined for training and obtaining a recognition model, which ensures that training data is not leaked during the training of the recognition model, and improves the recognition accuracy of the obtained recognition model.

Description

Training method, device, equipment and readable storage medium of recognition model

This application claims the priority of the Chinese patent application filed on March 20, 2020, with the application number 202010206241.7, titled "Recognition Model Training Method, Device, Equipment, and Readable Storage Medium", which is hereby incorporated in its entirety as refer to.

Technical field

This application relates to the field of data processing technology of financial technology (Fintech), and in particular to a training method, device, equipment and readable storage medium of a recognition model.

Background technique

With the development of computer technology, more and more technologies are applied in the financial field. The traditional financial industry is gradually changing to Fintech. Data processing technology is no exception. However, due to the security and real-time requirements of the financial industry, It also places higher requirements on data processing technology.

Machine learning technology is widely used in speech recognition. These technologies often rely heavily on massive speech data for training, which can easily leak the speech data, and thus cannot effectively protect the privacy of the users corresponding to the speech data. Moreover, the current speech recognition system can only provide very limited customization capabilities, which leads to low accuracy of speech recognition by the resulting speech recognition model. From this, it can be seen that the recognition accuracy of the currently trained recognition model is low, and the training data is likely to be leaked in the process of training the recognition model.

Technical solutions

The main purpose of this application is to provide a training method, device, equipment and readable storage medium for a recognition model, aiming to solve the technical problem of low recognition accuracy of the existing recognition model and easy leakage of training data during the training of the recognition model .

In order to achieve the above objective, the present application provides a method for training a recognition model, which is applied to a client, and the method for training a recognition model includes the steps:

Acquiring local training data, and training according to the training data to obtain a first recognition model;

Sending the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client to obtain a federated learning result, and return the federated learning result;

Receive the federated learning result sent by the server, update the first recognition model according to the federated learning result to obtain the corresponding second recognition model, and send the second recognition model to the server for The server obtains the target recognition model according to the second recognition model sent by each client and the preset genetic algorithm.

In addition, in order to achieve the above object, the present application also provides a training device for a recognition model. The training device for the recognition model is applied to a client, and the training device for the recognition model includes:

The acquisition module is used to acquire local training data;

The training module is used to train according to the training data to obtain the first recognition model;

The sending module is configured to send the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and returns the federated learning result ；

The receiving module is used to receive the federated learning result sent by the server;

An update module, configured to update the first recognition model according to the federated learning result to obtain a corresponding second recognition model;

The sending module is further configured to send the second recognition model to the server, so that the server obtains the target recognition model according to the second recognition model sent by each client and a preset genetic algorithm.

In addition, in order to achieve the above object, the present application also provides a recognition model training device, the recognition model training device is applied to the server, and the recognition model training device includes:

The receiving module is used to receive the model parameters corresponding to the first recognition model sent by each client;

The federated learning module is used to perform federated learning according to the model parameters to obtain federated learning results;

A sending module, configured to send the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model;

The receiving module is also used to receive the second recognition model sent by each client, and obtain the target recognition model according to each second recognition model and a preset genetic algorithm.

In addition, in order to achieve the above object, the present application also provides a training device for a recognition model. The training device for the recognition model includes a memory, a processor, and a recognition model stored in the memory and running on the processor. When the training program of the recognition model is executed by the processor, the steps of the training method of the recognition model corresponding to the federated learning server are implemented.

In addition, in order to achieve the above object, the present application also provides a computer-readable storage medium, the computer-readable storage medium stores a training program for a recognition model, and the training program for the recognition model is executed by a processor to achieve the above The steps of the training method of the recognition model described.

This application obtains the first recognition model based on the migration learning algorithm on the client side, trains according to the local training data, and sends the model parameters of the first recognition model to the server, receives the federated learning result sent by the server, and updates it according to the federated learning result The first recognition model obtains the corresponding second recognition model, and sends the second recognition model to the server, so that the server obtains the target recognition model according to the second recognition model sent by each client and the preset genetic algorithm . The recognition model is obtained through training of each client in the migration learning scenario, which improves the accuracy of the recognition model for recognizing the relevant information of each user, and this application supports the combination of multiple client recognition models in the federated learning scenario, which is effective In the case of protecting the privacy of the training data of each client corresponding to the user, the accuracy of the recognition data of the recognition model is further improved; this application supports the integration and optimization of multiple client recognition models in the evolutionary learning scenario through the genetic algorithm, and makes full use of each The client corresponds to the data value behind the recognition model, thereby further improving the recognition accuracy of the obtained recognition model.

Description of the drawings

FIG. 1 is a schematic flowchart of a first embodiment of a training method for a recognition model according to the present application;

2 is a schematic flowchart of a second embodiment of a training method for a recognition model according to the present application;

Fig. 3 is a functional schematic block diagram of the first embodiment of the training device for the recognition model of the present application;

Fig. 4 is a functional schematic block diagram of the second embodiment of the training device for the recognition model of the present application;

Fig. 5 is a schematic structural diagram of a hardware operating environment involved in a solution of an embodiment of the present application.

The realization, functional characteristics, and advantages of the purpose of this application will be further described in conjunction with the embodiments and with reference to the accompanying drawings.

Embodiments of the present invention

It should be understood that the specific embodiments described here are only used to explain the present application, and are not used to limit the present application.

The present application provides a method for training a recognition model. Refer to FIG. 1, which is a schematic flowchart of a first embodiment of a training method for a recognition model in this application.

The embodiment of the application provides an embodiment of the training method of the recognition model. It should be noted that although the logical sequence is shown in the flowchart, in some cases, the sequence shown here may be executed in a different order than here. Or the steps described.

In this embodiment, there are at least two clients, and the training method of the recognition model includes:

Step S10: Obtain local training data, and train according to the training data to obtain a first recognition model.

Each client obtains local training data, and inputs the training data into the basic model to obtain the first recognition model according to the local training data. In this embodiment, machine learning models such as decision trees, random forests, artificial neural networks, and Bayesian learning can be used as basic models. The local training data of the client is pre-stored. The type of local training data is determined by the recognition model that needs to be trained. For example, when the recognition model to be trained is a voice recognition model, the local training data is the voice data generated by the client corresponding to the user; when the recognition model to be trained is a face When recognizing the model, the local training data is the face image data generated by the client corresponding to the user. Specifically, the client can trigger the acquisition instruction through a pre-set timing task, and obtain local training data according to the acquisition instruction; the client can also obtain the local training data according to the acquisition instruction when it detects the acquisition instruction triggered by the client corresponding to the user. Training data.

Further, step S10 includes:

Step g: Obtain local training data, and train the first recognition model based on the training data based on the migration learning algorithm.

After the client obtains the local training data, the client obtains the first recognition model based on the migration learning algorithm and trains according to the obtained local training data. Transfer learning is a method of machine learning, which refers to a pre-trained model being reused in another task. Through migration learning, combined with the training data of other clients for joint modeling, under the premise of ensuring that the privacy of the training data of the two parties is not leaked, the trained recognition model can be applied to the party with missing features and annotations, thereby greatly improving It expands the application range of joint learning and effectively improves the predictive ability of the training recognition model. It should be noted that, in the process of obtaining the first recognition model through the migration learning algorithm, the training data obtained through the migration learning is input into the basic model to obtain the first recognition model. In this embodiment, the speech recognition model can be ASR (Speech Recognition Technology, speech recognition technology) corresponding model.

It should be noted that, during the migration learning process of the client, the client combines at least one other client's local training data. When there are multiple clients, the principles for obtaining the first recognition model for each client are the same, which will not be repeated here. There is a migration learning component in each client, and the migration learning algorithm is called through the migration learning component to obtain the first recognition model.

Step S20: Send the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and returns the federated learning result.

After each client obtains the first recognition model, the client obtains the model parameters of its own first recognition model, and sends the obtained model parameters to the server. It should be noted that the model parameters of the first recognition model of each client may be the same or different. When the server receives the model parameters sent by each client, the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and sends the federated learning result to each client. On the server side, there is a federated learning component, and the federated learning component can perform federated learning on the model parameters sent by each client to obtain the federated learning result.

For ease of understanding, this embodiment explains federated learning. Federated Learning (Federated Learning) is an emerging basic artificial intelligence technology. Its design goal is to ensure information security during the exchange of big data, protect terminal data and personal privacy data, and under the premise of legal compliance, in the presence of multiple parties or Carry out high-efficiency machine learning among multiple computing nodes. In the system architecture of federated learning, this embodiment uses a scenario containing two data owners (ie, enterprise A and enterprise B) as an example to introduce the system architecture of federated learning. The framework can be extended to scenarios that include multiple data owners. Suppose enterprise A and enterprise B want to jointly train a machine learning model, and their business systems each have relevant data of their respective users. In addition, company B also has label data that the model needs to predict. For data privacy protection and security considerations, enterprise A and enterprise B cannot directly exchange data, and can use a federated learning system to build models.

The use of federated learning to build a model includes two parts. The first part is: the alignment of encrypted samples. Since the user groups of the two companies are not completely overlapped, the system uses encryption-based user sample alignment technology to confirm the mutual users of both companies under the premise that company A and company B do not disclose their respective data, and does not expose users that do not overlap with each other, so that Combine the characteristics of these users for modeling. The second part is: encryption model training. After determining the common user group, the data can be used to train machine learning models. In order to ensure the confidentiality of the data during the training process, a third-party collaborator C needs to be used for encryption training. Taking the linear regression model as an example, the training process can be divided into the following 4 steps:

Step ①: Collaborator C distributes the public key to enterprise A and enterprise B to encrypt the data that needs to be exchanged during the training process.

Step ②: Encrypted interaction between enterprise A and enterprise B is used to calculate the intermediate result of the gradient.

Step ③: Enterprise A and Enterprise B respectively calculate based on the encrypted gradient value, and at the same time, Enterprise B calculates the loss based on its label data, and summarizes the results to collaborator C. Collaborator C calculates the total gradient value by summing up the results and decrypts it.

Step ④: Collaborator C sends the decrypted gradient value back to enterprise A and enterprise B respectively, and enterprise A and enterprise B update the parameters of their respective models according to the gradient value.

Iterate the above steps until the loss function corresponding to the linear regression model converges, and the loss function is a preset function, thus completing the entire training process. In the process of sample alignment and model training, the respective data of enterprise A and enterprise B are kept locally, and the data interaction during training will not cause data privacy leakage. Therefore, the two parties can realize the cooperative training model with the help of federated learning.

For different data sets, federated learning is divided into horizontal federated learning (horizontal federated learning), vertical federated learning and federated transfer learning (FmL).

Horizontal federated learning is to split the data set according to the horizontal (ie user dimension) when the user characteristics of the two data sets overlap more and the user overlaps less, and extract the user characteristics of the two parties that are the same but the users are not exactly the same. That part of the data for training. For example, there are two banks in different regions, and their user groups come from their respective regions, and their mutual intersections are small. However, their businesses are very similar, so the recorded user characteristics are the same. At this point, horizontal federated learning can be used to build a joint model.

Longitudinal federated learning is to split the data set according to the longitudinal direction (ie feature dimension) when the users of the two data sets overlap more and the user characteristics are less overlapped, and extract the ones where the two users are the same but the user characteristics are not exactly the same. Part of the data for training. For example, there are two different institutions, one is a bank in a certain place, and the other is an e-commerce company in the same place. Their user groups are likely to include most of the residents in that place, so the intersection of users is relatively large. However, because banks record the user's income and expenditure behavior and credit rating, while e-commerce stores the user's browsing and purchase history, their user characteristics have a small intersection. Vertical federated learning is to aggregate these different features in an encrypted state to enhance the model's capabilities. At present, many machine learning models such as logistic regression model, tree structure model and neural network model have gradually been confirmed to be able to build on this federal system.

Federated transfer learning is to use transfer learning to overcome the lack of data or labels when the user and user characteristics of the two data sets are less overlapped, without segmenting the data. For example, there are two different institutions, one is a bank in China, and the other is an e-commerce company in the United States. Due to geographical restrictions, the user groups of these two institutions have a very small intersection. At the same time, due to the different types of institutions, only a small part of the data characteristics of the two overlap. In this case, if you want to carry out effective federated learning, you must introduce transfer learning to solve the problem of small unilateral data and fewer label samples, so as to improve the effect of the model.

It should be noted that in the process of training the recognition model, training data of each client user is required, and these training data are stored locally in the client held by each user as the local training data of the corresponding user of the client. In the local training data, part of the training data belongs to the user's private data. Without the user's authorization, the user's private data cannot be obtained, that is, the user's private data cannot be used as training data to train the recognition model. At present, some applications in the client usually compulsorily insert clauses requiring the user to agree to the use of private data in advance in the user agreement, or obtain the user's private data in other unknowing ways, which significantly reduces the privacy of the user's private data. With federated learning, the federated learning result can be obtained without touching the user's original local training data, and each client uses the federated learning result to update the first recognition model.

Step S30: Receive the federated learning result sent by the server, update the first recognition model according to the federated learning result to obtain a corresponding second recognition model, and send the second recognition model to the server , So that the server can obtain the target recognition model according to the second recognition model sent by each client and the preset genetic algorithm.

Each client receives the federated learning result sent by the server, and updates the first recognition model according to the federated learning result to obtain the corresponding second recognition model, that is, the updated first recognition model is the second recognition model. After each client obtains the second recognition model, each client sends the second recognition model to the server. After the server receives the second recognition model sent by each client, the server obtains the target recognition model according to the second recognition model sent by each client and a preset genetic algorithm.

Genetic Algorithm (Genetic Algorithm) is a method of solving optimization problems through search. It first generates a certain amount of population randomly. Genetic Algorithm can include reproduce operators, crossover operators, mutation operators, etc. son. The crossover operator is the process of encoding the chromosomes of the members of the population and the chromosome codes of the two or two group members are crossed; the mutation operator is the process of mutating the chromosome codes after the crossover with a certain probability. In the server, there is a genetic algorithm component, through which the target recognition model can be obtained according to the second recognition model sent by each client.

Further, the step of receiving the federated learning result sent by the server, updating the first recognition model according to the federated learning result, and obtaining the corresponding second recognition model includes:

Step a: Receive a federated learning result sent by the server, and obtain a preset learning rate and a model parameter change amount in the federated learning result.

In step b, the updated model parameters are calculated according to the preset learning rate and the change amount of the model parameters, and the first recognition model is updated according to the updated model parameters to obtain the corresponding second recognition model.

Further, each client receives the federated learning result sent by the server, and each client obtains a preset learning rate and the amount of model parameter changes in the federated learning result. The preset learning rate can be set according to specific needs. The size of the preset learning rate is not specifically limited, and the preset learning rates corresponding to different clients may be the same or different. Specifically, each client obtains the model parameters of the first recognition model, calculates the product of the preset learning rate and the amount of model parameter change, and subtracts the calculated product from the model parameters of the first recognition model to obtain the updated Model parameters, and update the first recognition model according to the updated model parameters, that is, use the updated model parameters as the model parameters of the recognition model to obtain the second recognition model.

In this embodiment, the client is trained based on the migration learning algorithm to obtain the first recognition model according to the local training data, and the model parameters of the first recognition model are sent to the server, and the federated learning result sent by the server is received, and according to the federated learning result Update the first recognition model to obtain the corresponding second recognition model, and send the second recognition model to the server, so that the server can obtain target recognition according to the second recognition model sent by each client and the preset genetic algorithm Model. The recognition model obtained by training each client in the migration learning scenario improves the accuracy of the recognition model for recognizing the relevant information of each user, and this embodiment supports the combination of multiple client recognition models in the federated learning scenario. In the case of effectively protecting the privacy of the training data of each client corresponding to the user, the accuracy of the recognition data of the recognition model is further improved; this embodiment supports the integration and optimization of multiple client recognition models in the evolutionary learning scenario through the genetic algorithm, making full use of Each client corresponds to the data value behind the recognition model, thereby further improving the recognition accuracy of the obtained recognition model.

Further, a second embodiment of the training method of the recognition model of the present application is proposed. The difference between the second embodiment of the recognition model training method and the first embodiment of the recognition model training method is that, referring to FIG. 2, the recognition model training method further includes:

Step S40: Receive model parameters corresponding to the first recognition model sent by each client, and perform federated learning according to the model parameters to obtain a federated learning result.

The server receives the model parameters corresponding to the first recognition model sent by each client, and performs federated learning according to the model parameters to obtain the federated learning result. It should be noted that the process of federated learning has been described in detail in the foregoing embodiment, and will not be repeated here. In this embodiment, the federated learning result may be a model parameter change amount. Specifically, the model parameter change amount may be a gradient value calculated according to the loss function when the corresponding loss function converges during the federated learning process.

Step S50: Send the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model.

After the server obtains the federated learning result, the server sends the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model. Each client updates the first recognition model according to the federated learning result, and the process of obtaining the second recognition model has been described in detail in the foregoing embodiment, and will not be repeated here.

Step S60: Receive the second recognition model sent by each client, and obtain the target recognition model according to each second recognition model and a preset genetic algorithm.

The server receives the second recognition model sent by each client, and obtains the target recognition model according to each second recognition model and a preset genetic algorithm. It is understandable that the target recognition model is the optimal recognition model required. After the server obtains the target recognition model, the server can send the target recognition model to each client, so that each client can perform a corresponding recognition operation according to the target recognition model after receiving the recognition request. After each client receives the target recognition model, each client stores the received target recognition model.

Further, step S50 includes:

Step d, receiving the second recognition model sent by each client, selecting the parent recognition model from each of the received second recognition models, and selecting the target operator in the preset genetic algorithm.

Step e: Obtain a child recognition model corresponding to the parent recognition model through the parent recognition model and the target operator.

In step f, if it is detected that the offspring recognition model meets the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions is determined as the target recognition model.

Specifically, the server receives the second recognition model sent by each client, randomly selects the second recognition model from each of the received second recognition models to determine it as the parent recognition model, and randomly selects the calculation model in the preset genetic algorithm. The sub is determined as the target operator. When there are at least two parent recognition models, the target operator corresponding to each parent recognition model may be the same or different. After the server determines the parent recognition model and the target operator corresponding to the parent recognition model, the server obtains the child corresponding to the parent recognition model through the parent recognition model and the target operator corresponding to the parent recognition model Recognize the model, and check whether the resulting offspring recognition model meets the end condition. If the detected offspring recognition model satisfies the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions is obtained, and the offspring recognition model with the highest accuracy rate is determined as the target recognition model; if the obtained offspring recognition model is detected If the offspring recognition model does not meet the ending conditions, the server will use the obtained offspring recognition model as the parent recognition model for the next iteration, and determine the target operator corresponding to the next parent recognition model, and continue according to the parent recognition model and The target operator obtains the offspring recognition model corresponding to the parent recognition model, and when it is detected that the offspring recognition model satisfies the end condition, the offspring recognition model with the highest recognition accuracy rate among the satisfied conditions is determined as the target recognition model.

Further, step e includes:

Step e1: Obtain the to-be-processed model parameters corresponding to the parent recognition model, and determine the to-be-processed model parameters as parameter codes.

Step e2, processing the parameter encoding through the target operator to obtain a new parameter encoding.

Step e3: Correspondingly update the parent recognition model according to the new parameter code to obtain the offspring recognition model.

In this embodiment, the server obtains the model parameters in each second recognition model, that is, obtains the model parameters of the second recognition model as the parent recognition model, determines the obtained model parameters as the model parameters to be processed, and sets the parameters to be processed. Process model parameters to determine the parameter encoding of the genetic algorithm, and process the parameter encoding through the operator in the genetic algorithm, that is, process the encoding of the parameters to be processed through the target operator corresponding to the parent identification model, and obtain the new parameter encoding, and according to the new The parameter coding update corresponds to the parent recognition model, and the offspring recognition model is obtained. It is understandable that each client can also send the model parameters in the second recognition model to the server. It should be noted that in the process of processing the parameter encoding, the specific processing method used is determined by the corresponding target operator. For example, for a mutation operator, the offspring recognition model can only inherit a few specific model parameters of the parent recognition model. At this time, the genetic algorithm extracts the specific parameter code of the parent recognition model, and uses the extracted specific parameter code as the new parameter code . For the crossover operator, the parameter codes corresponding to the two parent recognition models can be integrated to obtain a new parameter code. The specific algorithm of integration is determined by the operator corresponding to the genetic algorithm. For example, the method of averaging the parameters can be used Coding for integration.

Further, when the server selects the parent recognition model in each second recognition model, the number of selected parent recognition models can be determined according to the current available resources in the server. When the available resources in the server are sufficient For a long time, all second recognition models can be determined as parent recognition models. It is understandable that as the available resources in the server increase, the number of selected parent identification models will increase accordingly.

Further, when detecting whether the progeny recognition model meets the end condition, it can be determined that the progeny recognition model meets the end condition when all the progeny recognition models meet the end condition, and if one of the progeny recognition models does not meet the end condition, It is determined that the offspring recognition model does not meet the end condition. It is also possible to determine that the child recognition model satisfies the end condition when the child recognition model with a preset ratio in the child recognition model is detected to meet the end condition; otherwise, it is determined that the child recognition model does not meet the end condition. At this time, the size of the preset ratio can be set according to specific needs, such as 60%, 75%, or 80%.

Further, step f includes:

Step f1: If it is detected that the number of iterations corresponding to the offspring identification model is greater than the preset number, it is determined that the offspring identification model meets the termination condition, and the pre-stored test data is used to obtain the identification of each offspring identification model that meets the termination condition Accuracy.

In step f2, the offspring recognition model with the highest recognition accuracy is selected and determined as the target recognition model.

Further, if the server detects that the number of iterations corresponding to the offspring identification model is greater than the preset number, it is determined that the offspring identification model meets the end condition. The preset number of times can be set according to specific needs, and this embodiment does not limit the size of the preset number of times. It is understandable that when the server obtains the offspring recognition model through the genetic algorithm for the first time, the corresponding iteration number is 1; when the server obtains the offspring identification model through the genetic algorithm for the second time, the corresponding iteration number is 2. , That is, the number of iterations is equal to the number of offspring identification models obtained through genetic algorithm. When it is determined that the offspring recognition model meets the end conditions, the server obtains the pre-stored test data, and inputs the test data into each offspring recognition model to determine the recognition accuracy of each offspring recognition model, and compares each child The recognition accuracy rate of the generation recognition model is selected, and the offspring recognition model with the highest recognition accuracy rate is selected as the target recognition model. Further, if it is detected that the number of iterations corresponding to the offspring identification model is less than or equal to the preset number, the server determines that the offspring identification model does not meet the end condition.

Further, step f also includes:

Step f3: If it is detected that the recognition accuracy rate corresponding to the offspring recognition model is greater than or equal to the preset accuracy rate, it is determined that the offspring recognition model meets the end condition.

Step f4, selecting the child recognition model with the highest recognition accuracy rate among the child recognition models and determining it as the target recognition model.

Further, every time the server obtains the offspring recognition model, the server obtains the pre-stored test data, inputs the test data into the offspring recognition model, obtains the recognition accuracy rate corresponding to the offspring recognition model, and detects each child Whether the recognition accuracy rate corresponding to the generation recognition model is greater than or equal to the preset accuracy rate. The size of the preset accuracy rate can be set according to specific needs, and this embodiment does not specifically limit the size of the preset accuracy rate. When the server detects that the corresponding recognition accuracy of each child recognition model is greater than or equal to the preset accuracy, the server determines that the child recognition model meets the end condition, and selects the child recognition with the highest recognition accuracy in the child recognition model The model is determined to be the target recognition model; when the server detects that the recognition accuracy rate corresponding to the offspring recognition model is less than the preset accuracy rate, the server determines that the offspring recognition model does not meet the end condition.

Further, the server can also determine that the child recognition model meets the end condition when the proportion of the child recognition model whose detection and recognition accuracy rate is greater than or equal to the preset accuracy rate to all the child recognition models in the current iteration process is greater than a specific ratio, Otherwise, it is determined that the offspring recognition model does not meet the end condition. Among them, this embodiment does not limit the size of the specific ratio.

Further, the server can also sort the child recognition models from high to low according to the recognition accuracy when determining that the child recognition model meets the end condition, and obtain the sorted child recognition model, and then the sorted child recognition model Obtain a preset number of offspring recognition models from front to back in the recognition model to determine the target offspring recognition model, obtain the model parameters of the target offspring recognition model, and linearly add the model parameters of each target offspring recognition model to obtain the target Model parameters, the recognition model corresponding to the target model parameters is determined as the target model. Among them, the size of the preset number can be set according to specific needs, for example, the preset number can be set to 2, 3, or 5. The specific process of linear addition can be to calculate the average value of the model parameters corresponding to each target offspring recognition model, and determine the average value as the target model parameter; the process of linear addition can also be to determine the corresponding value of each target offspring recognition model Weight, multiply the model parameters of each target offspring recognition model by the corresponding weight to get the product, and then add the products corresponding to the same model parameter in each target offspring recognition model to get the corresponding target model parameters, understandable Yes, the higher the recognition accuracy, the greater the weight of the model parameters corresponding to the target offspring recognition model.

In this embodiment, the server receives the model parameters corresponding to the first recognition model sent by each client, performs federated learning according to the model parameters, sends the result of federated learning to each client, and receives the second recognition sent by each client Model, obtain the target recognition model according to each second recognition model and the preset genetic algorithm. The recognition model obtained by training each client in the migration learning scenario improves the accuracy of the recognition model for recognizing the relevant information of each user, and this embodiment supports the combination of multiple client recognition models in the federated learning scenario. In the case of effectively protecting the privacy of the training data of each client corresponding to the user, the accuracy of the recognition data of the recognition model is further improved; this embodiment supports the integration and optimization of multiple client recognition models in the evolutionary learning scenario through the genetic algorithm, making full use of Each client corresponds to the data value behind the recognition model, thereby further improving the recognition accuracy of the obtained recognition model.

In addition, the present application also provides a training device for a recognition model. Referring to FIG. 3, the training device for the recognition model is applied to the client, and the training device for the recognition model includes:

The obtaining module 10 is used to obtain local training data;

The training module 20 is configured to train according to the training data to obtain a first recognition model;

The sending module 30 is configured to send the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and returns the federated learning result;

The receiving module 40 is configured to receive the federated learning result sent by the server;

The update module 50 is configured to update the first recognition model according to the federated learning result to obtain a corresponding second recognition model;

The sending module 30 is further configured to send the second recognition model to the server, so that the server obtains the target recognition model according to the second recognition model sent by each client and a preset genetic algorithm.

Further, the training module 20 is also used to obtain local training data, and based on a migration learning algorithm, train according to the training data to obtain a first recognition model.

Further, the update module 50 includes:

An acquiring unit for acquiring a preset learning rate and the model parameter change amount in the federated learning result;

A calculation unit, configured to calculate the updated model parameters according to the preset learning rate and the change amount of the model parameters;

The update unit updates the first recognition model according to the updated model parameters to obtain the corresponding second recognition model.

The specific implementation of the training device for the recognition model of this application is basically the same as the steps in the first embodiment of the training method for the recognition model described above, and will not be repeated here.

In addition, the present application also provides a training device for a recognition model. Referring to FIG. 4, the training device for the recognition model is applied to the server, and the training device for the recognition model includes:

The receiving module 60 is configured to receive model parameters corresponding to the first recognition model sent by each client;

The federated learning module 70 is configured to perform federated learning according to the model parameters to obtain federated learning results;

The sending module 80 is configured to send the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model;

The receiving module 60 is also configured to receive the second recognition model sent by each client, and obtain the target recognition model according to each second recognition model and a preset genetic algorithm.

Further, the receiving module 60 includes:

The receiving unit is configured to receive the second recognition model sent by each client;

The selection unit is used to select the parent recognition model among the received second recognition models, and select the target operator in the preset genetic algorithm;

The determining unit is configured to obtain the child recognition model corresponding to the parent recognition model through the parent recognition model and the target operator; if it is detected that the child recognition model satisfies the end condition, the end condition will be met The child recognition model with the highest recognition accuracy rate is determined as the target recognition model.

Further, the determining unit includes:

The first obtaining subunit is configured to obtain the to-be-processed model parameters corresponding to the parent recognition model;

The first determining subunit is used to determine the parameters of the model to be processed as parameter codes;

A processing subunit, configured to process the parameter encoding through the target operator to obtain a new parameter encoding;

The update subunit is used to update the parent recognition model corresponding to the new parameter encoding to obtain the child recognition model.

Further, the determining unit includes:

The second determining subunit is configured to determine that the child recognition model meets the ending condition if it is detected that the number of iterations corresponding to the offspring recognition model is greater than the preset number;

The second acquisition subunit is used to acquire the recognition accuracy rate of each offspring recognition model that meets the end condition by using the pre-stored test data;

The second determining subunit is also used to select the offspring recognition model with the highest recognition accuracy rate and determine it as the target recognition model.

Further, the determining unit is further configured to determine that the child recognition model satisfies the end condition if the recognition accuracy corresponding to the child recognition model is detected to be greater than or equal to a preset accuracy rate; In the model, the child recognition model with the highest recognition accuracy rate is selected as the target recognition model.

The specific implementation of the training device for the recognition model of the present application is basically the same as the steps in the second embodiment of the training method for the recognition model described above, and will not be repeated here.

In addition, this application also provides a training device for identifying models. As shown in FIG. 5, FIG. 5 is a schematic structural diagram of the hardware operating environment involved in the solution of the embodiment of the present application.

It should be noted that FIG. 5 can be a schematic structural diagram of the hardware operating environment of the training device for the recognition model. The training device for the recognition model may be a client or a server. The training device for the recognition model in the embodiment of the present application may be a terminal device such as a PC and a portable computer.

As shown in FIG. 5, the training device of the recognition model may include: a processor 1001, such as a CPU, a memory 1005, a user interface 1003, a network interface 1004, and a communication bus 1002. Among them, the communication bus 1002 is used to implement connection and communication between these components. The user interface 1003 may include a display screen (Display) and an input unit such as a keyboard (Keyboard), and the optional user interface 1003 may also include a standard wired interface and a wireless interface. The network interface 1004 may optionally include a standard wired interface and a wireless interface (such as a WI-FI interface). The memory 1005 may be a high-speed RAM memory, or a stable memory (non-volatile memory), such as a magnetic disk memory. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

Those skilled in the art can understand that the structure of the training device for the recognition model shown in FIG. 5 does not constitute a limitation on the training device for the recognition model, and may include more or less components than shown in the figure, or a combination of certain components, Or different component arrangements.

As shown in FIG. 5, the memory 1005 as a computer storage medium may include an operating system, a network communication module, a user interface module, and a training program for a recognition model. Among them, the operating system is a program that manages and controls the hardware and software resources of the training equipment of the recognition model, and supports the operation of the training program of the recognition model and other software or programs.

In the training device of the recognition model shown in FIG. 5, when the training device of the recognition model is the client, the user interface 1003 is mainly used to connect to the server and communicate with the server; when the training device of the recognition model is the server At the time, the user interface 1003 is mainly used to connect various clients and communicate with each client; the network interface 1004 is mainly used to communicate with the back-end server and perform data communication with the back-end server; the processor 1001 can be used to call the storage in the memory 1005 The training program of the recognition model, and the steps of the training method of the recognition model as described above are executed.

The specific implementation of the training device for the recognition model of the present application is basically the same as the foregoing embodiments of the training method for the recognition model, and will not be repeated here.

In addition, an embodiment of the present application also proposes a computer-readable storage medium, the computer-readable storage medium stores a training program for a recognition model, and the training program for the recognition model is executed by a processor to realize the recognition as described above. The steps of the model training method.

The specific implementation of the computer-readable storage medium of the present application is basically the same as the foregoing embodiments of the training method of the recognition model, and will not be repeated here.

It should be noted that in this article, the terms "include", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, It also includes other elements that are not explicitly listed, or elements inherent to the process, method, article, or device. If there are no more restrictions, the element defined by the sentence "including a..." does not exclude the existence of other identical elements in the process, method, article, or device that includes the element.

The serial numbers of the foregoing embodiments of the present application are for description only, and do not represent the superiority or inferiority of the embodiments.

Through the description of the above implementation manners, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。 Based on this understanding, the technical solution of this application essentially or the part that contributes to the existing technology can be embodied in the form of a software product, and the computer software product is stored in a storage medium (such as ROM/RAM, magnetic disk, The optical disc) includes several instructions to enable a terminal device (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) to execute the method described in each embodiment of the present application.

The above are only the preferred embodiments of the application, and do not limit the scope of the patent for this application. Any equivalent structure or equivalent process transformation made using the content of the description and drawings of the application, or directly or indirectly applied to other related technical fields , The same reason is included in the scope of patent protection of this application.

Claims

A training method for a recognition model is applied to a client, wherein the training method for the recognition model includes the following steps:

Acquiring local training data, and training according to the training data to obtain a first recognition model;

Sending the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and returns the federated learning result; and

Receive the federated learning result sent by the server, update the first recognition model according to the federated learning result to obtain the corresponding second recognition model, and send the second recognition model to the server for The server obtains the target recognition model according to the second recognition model sent by each client and the preset genetic algorithm.
The method for training a recognition model according to claim 1, wherein said obtaining local training data comprises:

Trigger the acquisition instruction through a pre-set timed task, and acquire local training data according to the acquisition instruction; or

When detecting that the client corresponds to the acquisition instruction triggered by the user, the local training data is acquired according to the acquisition instruction.
The method for training a recognition model according to claim 1, wherein the step of obtaining local training data and training according to the training data to obtain the first recognition model comprises:

Obtain local training data, and train the first recognition model based on the training data based on the migration learning algorithm.
The method for training a recognition model according to any one of claims 1 to 3, wherein the receiving the federated learning result sent by the server, and updating the first recognition model according to the federated learning result to obtain the corresponding The steps of the second recognition model include:

Receiving the federated learning result sent by the server, and obtaining the preset learning rate and the model parameter change amount in the federated learning result; and

The updated model parameters are calculated according to the preset learning rate and the change amount of the model parameters, and the first recognition model is updated according to the updated model parameters to obtain the corresponding second recognition model.
The method for training a recognition model according to claim 4, wherein the updated model parameters are calculated according to the preset learning rate and the amount of change in the model parameters, and the first model parameters are updated according to the updated model parameters. The recognition model to obtain the corresponding second recognition model includes:

Obtain the model parameters of the first recognition model, calculate the product of the preset learning rate and the amount of model parameter change, and subtract the calculated product from the model parameters of the first recognition model to obtain the updated Model parameters, and update the first recognition model according to the updated model parameters to obtain the second recognition model.
A training method for a recognition model, wherein the training method for the recognition model is applied to a server, and the training method for the recognition model includes the following steps:

Receiving model parameters corresponding to the first recognition model sent by each client, and performing federated learning according to the model parameters to obtain a federated learning result;

Sending the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model; and

The second recognition model sent by each client is received, and the target recognition model is obtained according to each second recognition model and a preset genetic algorithm.
The method for training a recognition model according to claim 6, wherein the federated learning result is a model parameter change amount, and the model parameter change amount is calculated according to the loss function when the corresponding loss function converges during the federated learning process The obtained gradient value.
8. The method for training a recognition model according to claim 6, wherein the target recognition model is a desired optimal recognition model.
The method for training a recognition model according to claim 6, wherein after the step of receiving the second recognition model sent by each client and obtaining the target recognition model according to each second recognition model and a preset genetic algorithm, the method further comprises :

The target recognition model is sent to each client, so that each client can perform a corresponding recognition operation according to the target recognition model after receiving the recognition request.
7. The method for training a recognition model according to claim 6, wherein the step of receiving the second recognition model sent by each client, and obtaining the target recognition model according to each second recognition model and a preset genetic algorithm comprises:

Receiving the second recognition model sent by each client, selecting the parent recognition model from each of the received second recognition models, and selecting the target operator in the preset genetic algorithm;

Obtain a child recognition model corresponding to the parent recognition model through the parent recognition model and the target operator; and

If it is detected that the offspring recognition model meets the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions is determined as the target recognition model.
10. The method for training a recognition model according to claim 10, wherein the step of obtaining a child recognition model corresponding to the parent recognition model through the parent recognition model and the target operator comprises:

Acquiring the to-be-processed model parameters corresponding to the parent recognition model, and determining the to-be-processed model parameters as parameter codes;

Processing the parameter encoding through the target operator to obtain a new parameter encoding; and

Correspondingly update the parent recognition model according to the new parameter encoding to obtain the offspring recognition model.
11. The method for training a recognition model according to claim 11, wherein the step of selecting a parent recognition model among the received second recognition models comprises:

The number of selected parent recognition models is determined according to the current available resources in the server.
10. The method for training a recognition model according to claim 10, wherein the step of detecting that the offspring recognition model satisfies an end condition comprises:

When all the offspring recognition models meet the end condition, it is determined that the offspring recognition model meets the end condition.
The method for training a recognition model according to any one of claims 10 to 13, wherein, if it is detected that the offspring recognition model satisfies the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions will be satisfied The steps to determine the target recognition model include:

If it is detected that the number of iterations corresponding to the offspring recognition model is greater than the preset number, it is determined that the offspring recognition model meets the end condition, and the pre-stored test data is used to obtain the recognition accuracy of each offspring recognition model that meets the end condition; as well as

The offspring recognition model with the highest recognition accuracy is selected as the target recognition model.
The method for training a recognition model according to any one of claims 10 to 13, wherein, if it is detected that the offspring recognition model satisfies the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions will be satisfied The steps to determine the target recognition model include:

If it is detected that the recognition accuracy rate corresponding to the offspring recognition model is greater than or equal to the preset accuracy rate, it is determined that the offspring recognition model meets the end condition; and

Among the offspring recognition models, the offspring recognition model with the highest recognition accuracy is selected and determined as the target recognition model.
The method for training a recognition model according to any one of claims 10 to 13, wherein, if it is detected that the offspring recognition model satisfies the end condition, the offspring recognition model with the highest recognition accuracy rate among the end conditions will be satisfied The steps to determine the target recognition model include:

If it is detected that the proportion of the child recognition model with a recognition accuracy greater than or equal to the preset accuracy to all child recognition models in the current iteration process is greater than a certain ratio, it is determined that the child recognition model meets the end condition; and

Among the offspring recognition models, the offspring recognition model with the highest recognition accuracy is selected and determined as the target recognition model.
A training device for a recognition model, wherein the training device for the recognition model is applied to a client, and the training device for the recognition model includes:

The acquisition module is used to acquire local training data;

The training module is used to train according to the training data to obtain the first recognition model;

The sending module is configured to send the model parameters of the first recognition model to the server, so that the server performs federated learning according to the model parameters sent by each client, obtains the federated learning result, and returns the federated learning result ；

The receiving module is used to receive the federated learning result sent by the server; and

An update module, configured to update the first recognition model according to the federated learning result to obtain a corresponding second recognition model;

The sending module is further configured to send the second recognition model to the server, so that the server obtains the target recognition model according to the second recognition model sent by each client and a preset genetic algorithm.
A training device for a recognition model, wherein the training device for the recognition model is applied to a server, and the training device for the recognition model includes:

The receiving module is used to receive the model parameters corresponding to the first recognition model sent by each client;

The federated learning module is used to perform federated learning according to the model parameters to obtain federated learning results; and

A sending module, configured to send the federated learning result to each client, so that each client can update the corresponding first recognition model according to the federated learning result to obtain the corresponding second recognition model;

The receiving module is also used to receive the second recognition model sent by each client, and obtain the target recognition model according to each second recognition model and a preset genetic algorithm.
A training device for a recognition model, wherein the training device for the recognition model includes a memory, a processor, and a training program for a recognition model that is stored on the memory and can run on the processor. When the training program is executed by the processor, the steps of the training method of the recognition model as described in any one of claims 1 to 16 are realized.
A computer-readable storage medium, wherein a training program for a recognition model is stored on the computer-readable storage medium, and when the training program for the recognition model is executed by a processor, the training program of any one of claims 1 to 16 is realized The steps of the training method of the recognition model.