Model training method and apparatus, and computer device and storage medium

Abstract

The present application belongs to the technical field of computers. Provided are a model training method and apparatus, and a computer device and a storage medium. The method comprises: acquiring a plurality of training samples and a processing model, wherein the processing model comprises a gating network and a plurality of expert networks (201); processing the plurality of training samples and tasks of the plurality of expert networks by means of the gating network in the processing model, and determining the sample number of training samples to be processed by each expert network in the processing model (202); on the basis of the sample numbers and reference numbers of the plurality of expert networks, determining a training loss (203); and on the basis of the training loss, training the processing model (204). By means of the technical solution, the utilization rate of expert networks can be improved, and it is also possible to prevent a reduction in the processing efficiency of a processing model caused by the excessive number of samples to be processed by some expert networks.

Description

Model training method, device, computer equipment and storage medium

[0001] This application claims priority to the Chinese patent application filed on February 29, 2024, with application number 202410232589.1, and invention name “Model training method, device, computer equipment and storage medium”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of computer technology, and in particular to a model training method, apparatus, computer equipment, and storage medium. Background Art

[0003] With the development of computer technology, the types of network models are increasing. A typical network model is the Mixture of Experts (MoE). In a MoE, each input is processed by the gating network in the MoE, which assigns it to the expert network that is most suitable for processing the input. Summary of the Invention

[0004] The present invention provides a model training method, apparatus, computer device, and storage medium. The technical solution is as follows:

[0005] In one aspect, a model training method is provided, comprising:

[0006] Acquire multiple training samples and a processing model, wherein the processing model includes a gating network and multiple expert networks;

[0007] Processing the plurality of training samples and the tasks of the plurality of expert networks through the gating network in the processing model to determine the number of training samples to be processed by each expert network in the processing model, wherein the task of each expert network represents the type of training samples that the expert network can process;

[0008] Determining a training loss based on the number of samples of the multiple expert networks and a reference number, where the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples of the multiple expert networks and the reference number;

[0009] The processing model is trained based on the training loss.

[0010] In another aspect, a model training device is provided, comprising:

[0011] An acquisition module, configured to acquire a plurality of training samples and a processing model, wherein the processing model includes a gated network and a plurality of expert networks;

[0012] a processing module, configured to process the plurality of training samples and the tasks of the plurality of expert networks through the gating network in the processing model, and determine the number of training samples to be processed by each expert network in the processing model, wherein the tasks of each expert network represent the types of training samples that the expert network can process;

[0013] a first determining module, configured to determine a training loss based on the number of samples of the multiple expert networks and a reference number, wherein the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples of the multiple expert networks and the reference number;

[0014] A training module is used to train the processing model based on the training loss.

[0015] In some embodiments, the processing module includes:

[0016] a first determining unit, configured to process, for any one of the plurality of training samples, the training sample and the tasks of the plurality of expert networks using the gating network in the processing model, and determine a matching result of the training sample, wherein the matching result includes a matching degree between the training sample and the plurality of expert networks;

[0017] The second determining unit is configured to determine the number of training samples to be processed by each expert network in the processing model based on the matching results of the multiple training samples.

[0018] In some embodiments, the first determination unit is used to perform feature extraction on any one of the multiple training samples through the gating network in the processing model to obtain sample features of the training sample, where the sample features represent the type of the training sample; for any one of the multiple expert networks, perform feature extraction on the task of the expert network through the gating network in the processing model to obtain task features of the expert network; and determine the similarity between the sample features and the task features of each expert network through the gating network in the processing model to obtain a matching result of the training sample.

[0019] In some embodiments, the processing module includes:

[0020] a processing unit configured to process, for any expert network in the processing model, the tasks of the expert network and the plurality of training samples through the gating network in the processing model, to obtain a matching result of the expert network, wherein the matching result includes a matching degree between the expert network and the plurality of training samples;

[0021] The third determining unit is configured to determine the number of training samples to be processed by the expert network based on the matching result of the expert network.

[0022] In some embodiments, the third determination unit is used to sort multiple matching degrees in the matching results of the expert network in descending order; select training samples whose matching degrees reach a reference degree as training texts to be processed by the expert network, and obtain the sample number of training samples to be processed by the expert network.

[0023] In some embodiments, the number of training samples to be processed by each expert network does not exceed the sample capacity of the expert network, and the sample capacity represents the maximum number of training samples that the expert network can process. The sample capacity is determined based on the performance of the expert network, and the performance of the expert network includes at least one of the efficiency or accuracy of the expert network in processing training samples.

[0024] In some embodiments, the apparatus further comprises:

[0025] a second determining module configured to determine, for any training sample from the plurality of training samples, a contribution of the training sample based on the training sample and the expert network that will process the training sample before assigning the training sample to any expert network for processing, wherein the contribution represents a degree of contribution of assigning the training sample to the expert network for processing to model training;

[0026] The training module is used to train the processing model based on the training loss and the contribution of the multiple training samples.

[0027] On the other hand, a computer device is provided, which includes a processor and a memory, wherein the memory is used to store at least one computer program, and the at least one computer program is loaded and executed by the processor to implement the model training method in the embodiment of the present application.

[0028] On the other hand, a computer-readable storage medium is provided, in which at least one computer program is stored. The at least one computer program is loaded and executed by a processor to implement the model training method in the embodiment of the present application.

[0029] On the other hand, a computer program product is provided, including a computer program, which is stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium, and the processor executes the computer program, so that the computer device executes the model training method provided in the above-mentioned various aspects or various optional implementations of various aspects.

[0030] An embodiment of the present application provides a model training method. For a processing network including multiple expert networks, during the training of this type of processing model, the number of training samples to be processed by each expert network can be determined based on multiple training samples and the tasks of each expert network. That is, the multiple training samples are allocated to the corresponding expert network according to the sample types that the expert network can process for training, so that each expert model can accurately learn the knowledge of each sample type, thereby more accurately processing samples of the corresponding type; and the training loss is determined by determining the number of samples and the reference number of the multiple expert networks, so that the training samples are distributed as evenly as possible to each expert network, that is, the number of training samples processed by each expert network is the same. This not only improves the utilization rate of the expert network, but also avoids the reduction of the processing efficiency of the processing model due to too many samples to be processed by some expert networks. It also enables each expert network to learn knowledge of multiple sample types, improves the generalization ability of the expert network, and thus more accurately processes samples based on the knowledge of multiple sample types, which is beneficial to improving the accuracy of the processing model. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0032] FIG1 is a schematic diagram of an implementation environment of a model training method provided according to an embodiment of the present application;

[0033] FIG2 is a flow chart of a model training method provided according to an embodiment of the present application;

[0034] FIG3 is a flow chart of another model training method provided according to an embodiment of the present application;

[0035] FIG4 is a block diagram of a model training device provided according to an embodiment of the present application;

[0036] FIG5 is a block diagram of another model training device provided according to an embodiment of the present application;

[0037] FIG6 is a structural block diagram of a terminal provided according to an embodiment of the present application;

[0038] FIG7 is a schematic diagram of the structure of a server provided according to an embodiment of the present application. DETAILED DESCRIPTION

[0039] In order to make the objectives, technical solutions and advantages of this application clearer, the implementation methods of this application will be further described in detail below with reference to the accompanying drawings.

[0040] In this application, the terms "first", "second", etc. are used to distinguish identical or similar items with substantially the same effects and functions. It should be understood that there is no logical or temporal dependency between "first", "second", and "nth", nor is there any limitation on the quantity and execution order.

[0041] In the present application, the term "at least one" means one or more, and the term "plurality" means two or more.

[0042] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application are all authorized by the user or fully authorized by all parties, and the collection, use, and processing of relevant data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the training samples involved in this application were obtained with full authorization.

[0043] In a hybrid expert model, each input is assigned by the gating network in the hybrid expert model to the expert network that is most suitable for processing it. However, this allocation method causes some expert networks to handle a large number of inputs, while other experts are idle. This underutilizes the hybrid expert network and reduces its processing efficiency.

[0044] The model training method provided in the embodiment of the present application can be executed by a computer device. In some embodiments, the computer device is a terminal or a server. First, taking the computer device as a server as an example, the implementation environment of the model training method provided in the embodiment of the present application is introduced. Figure 1 is a schematic diagram of the implementation environment of a model training method provided according to the embodiment of the present application. Referring to Figure 1, the implementation environment includes a terminal 101 and a server 102. The terminal 101 and the server 102 can be directly or indirectly connected via wired or wireless communication, and this application does not limit this.

[0045] In some embodiments, the terminal 101 is a smart phone, a tablet computer, a laptop computer, a desktop computer, a smart speaker, a smart watch, an intelligent voice interaction device, a smart home appliance, a car terminal, etc., but is not limited thereto. The terminal 101 runs an application that can obtain training samples. The training samples can be texts in multiple languages, images of multiple types, etc., and the embodiments of the present application are not limited to this. The application can be a communication application, a translation application, an academic application, or a medical application, etc., and the embodiments of the present application are not limited to this. The terminal 101 can send the text or image in the application to the server 102, so that the server 102 can use the received text or image as a training sample to train the processing model.

[0046] Those skilled in the art will appreciate that the number of the above-mentioned terminals may be more or less. For example, the above-mentioned terminal may be only one, or the above-mentioned terminals may be dozens or hundreds, or a larger number. The embodiments of the present application do not limit the number of terminals and device types.

[0047] In some embodiments, server 102 is an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN (Content Delivery Network), big data and artificial intelligence platforms. Server 102 is used to provide background services for applications that support the acquisition of training samples. In some embodiments, server 102 undertakes the primary computing work and terminal 101 undertakes the secondary computing work; alternatively, server 102 undertakes the secondary computing work and terminal 101 undertakes the primary computing work; alternatively, server 102 and terminal 101 use a distributed computing architecture for collaborative computing.

[0048] FIG2 is a flow chart of a model training method according to an embodiment of the present application. Referring to FIG2 , the present embodiment is described using the server as an example. The model training method includes the following steps:

[0049] 201. The server obtains multiple training samples and a processing model, where the processing model includes a gated network and multiple expert networks.

[0050] In the embodiments of the present application, the multiple training samples can be different types of text or different types of images. Different types of text can be text in different languages, or text in different fields or themes, which are not limited in the embodiments of the present application. Different types of images can be images for different diseases, images of different urgency, or images of different styles, etc., which are not limited in the embodiments of the present application. The urgency of an image refers to the urgency of processing the image or the importance of the image.

[0051] The processing model can be constructed based on the architecture of a hybrid expert network. The embodiments of the present application do not limit the specific architecture of the processing model. The processing model includes a gating network (routing) and multiple expert networks. The multiple expert networks are used to process different types of training samples respectively. Optionally, the types of samples processed by any two expert networks are not exactly the same. The gating network is used to assign multiple training samples to different expert networks for processing. The server obtains multiple training samples so that the processing model can be trained based on the multiple training samples.

[0052] 202. The server processes multiple training samples and tasks of multiple expert networks through the gating network in the processing model, and determines the number of training samples to be processed by each expert network in the processing model. The task of each expert network represents the type of training samples that the expert network can process.

[0053] In this embodiment of the present application, for any of the multiple training samples, the server processes the training sample and the tasks of the multiple expert networks using the gating network in the processing model, and assigns the multiple training samples to matching expert networks. In other words, the server assigns each training sample to an expert network capable of processing the type of training sample. The server then counts the number of training samples to be processed by each expert network.

[0054] 203. The server determines the training loss based on the number of samples and the reference number of the multiple expert networks, where the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples and the reference number of the multiple expert networks.

[0055] In this embodiment of the present application, the server determines a reference number based on the number of expert networks and the number of training samples in the processing model. Then, for any expert network, the server determines the expert network's numerical loss based on the number of samples in that expert network and the reference number. The server then aggregates the numerical losses of the multiple expert networks to obtain the training loss of the processing model.

[0056] 204. The server trains the processing model based on the training loss.

[0057] In an embodiment of the present application, the server trains the processing model with the goal of minimizing training loss. During the training of the processing model, the server can adjust the network parameters of the gating network in the processing model so that the gating network can evenly distribute multiple training samples to multiple expert networks; or it can also adjust the network parameters of each expert network in the processing model so that each expert network can accurately process the assigned training samples, which is not limited in this embodiment of the present application.

[0058] An embodiment of the present application provides a model training method. For a processing network including multiple expert networks, during the training of this type of processing model, the number of training samples to be processed by each expert network can be determined based on multiple training samples and the tasks of each expert network. That is, the multiple training samples are allocated to the corresponding expert network according to the sample types that the expert network can process for training, so that each expert model can accurately learn the knowledge of each sample type, thereby more accurately processing samples of the corresponding type; and the training loss is determined by determining the number of samples and the reference number of the multiple expert networks, so that the training samples are distributed to each expert network as evenly as possible, that is, the training samples processed by each expert network are as similar as possible. This not only improves the utilization rate of the expert network, but also avoids the reduction of the processing efficiency of the processing model due to too many samples to be processed by some expert networks. It also enables each expert network to learn knowledge of multiple sample types, improves the generalization ability of the expert network, and thus processes samples more accurately based on the knowledge of multiple sample types, which is conducive to improving the accuracy of the processing model.

[0059] FIG3 is a flow chart of another model training method provided according to an embodiment of the present application. Referring to FIG3 , the present embodiment of the present application is described using the server as an example. The model training method includes the following steps:

[0060] 301. The server obtains multiple training samples and a processing model, where the processing model includes a gated network and multiple expert networks.

[0061] In an embodiment of the present application, the multiple training samples can be different types of text or different types of images. The different types of text can be text in different languages, or text in different fields or themes, which is not limited in this embodiment of the present application. The different types of images can be images reflecting different symptoms, images of different urgency levels, or images of different styles, etc., which is not limited in this embodiment of the present application. The processing model includes a gated network and multiple expert networks. The gated network is used to assign training samples to appropriate expert networks. The expert network is used to process training samples of corresponding types.

[0062] For example, multiple training samples are texts in multiple languages, including English, Chinese, and French. The processing model is a large multilingual translation model. Expert Network A in the processing model processes English text; Expert Network B processes Chinese text; Expert Network C processes French text, and so on. Alternatively, multiple training samples are images reflecting different medical conditions, including images reflecting heart disease, images reflecting respiratory problems, and images reflecting liver disease. The processing model is a diagnostic model for triage. Expert Network A in the processing model processes images reflecting heart disease; Expert Network B processes images reflecting respiratory problems; and Expert Network C processes images reflecting liver disease.

[0063] 302. The server processes multiple training samples and tasks of multiple expert networks through the gating network in the processing model, and determines the number of training samples to be processed by each expert network in the processing model. The task of each expert network represents the type of training samples that the expert network can process.

[0064] In an embodiment of the present application, the server processes the tasks of the training samples and multiple expert networks by processing the gating network in the model to match the multiple training samples with the multiple expert networks, thereby assigning the multiple training samples to the matching expert networks for processing. The server then counts the number of training samples to be processed by each expert network. In the process of matching the training samples and the expert networks, the server can select a suitable expert network for each training sample from the perspective of the training samples. This method can be regarded as a forward matching method (forward selection method). Alternatively, the server can also select a suitable training sample for each expert network from the perspective of the expert network. This method can be regarded as a reverse matching method (reverse selection method). The embodiment of the present application does not limit the above-mentioned matching method.

[0065] In some embodiments, the server selects a suitable expert network to process each training sample from the perspective of the training sample. Accordingly, the process by which the server determines the sample quantity of the training samples to be processed by each expert network in the processing model includes: for any training sample among the multiple training samples, the server processes the training sample and the tasks of the multiple expert networks through the gating network in the processing model to determine the matching result of the training sample. The matching result includes the matching degree between the training sample and the multiple expert networks. Then, based on the matching results of the multiple training samples, the server determines the sample quantity of the training samples to be processed by each expert network in the processing model. The solution provided in the embodiment of the present application, from the perspective of the training sample, selects a matching expert network for each training sample, so that each training sample can be processed by an expert network suitable for processing the type to which the training sample belongs, thereby ensuring that the training sample can be accurately processed and improving the accuracy of the processing model.

[0066] In determining the matching result for a training sample, for each of the multiple training samples, the server extracts features from the training sample using the gating network in the processing model to obtain the sample features of the training sample. The sample features represent the type of the training sample. For each of the multiple expert networks, the server extracts features from the expert network's tasks using the gating network in the processing model to obtain the expert network's task features. The server then uses the gating network in the processing model to determine the similarity between the sample features and the task features of each expert network to obtain the matching result for the training sample.

[0067] In some embodiments, the server may calculate the matching result of the training sample using the following formula 1.

[0068] Formula 1:

[0069] g(x)=softmax(Wx+b)

[0070] Here, g(x) represents the matching result of the training sample. The matching result of the training sample is the probability distribution of the matching degree (similarity) between the training sample and multiple expert networks; softmax represents the matching function between the training sample and the expert network; x represents the training sample; W and b represent the network parameters of the gating network in the processing model. These are learnable network parameters that can be adjusted during the training process to adjust the probability distribution of the training sample assigned to different expert networks.

[0071] For example, multiple training samples are texts in multiple languages, including English and Chinese. The processing model is a large multilingual translation model. Expert network A in the processing model is used to process English text; expert network B is used to process Chinese text; expert network C is used to process French text, etc. The server uses the gating network in the processing model to determine the language present in each text, and then assigns the text to the expert network that can process the corresponding language for processing. For example, text containing English is assigned to expert network A for processing; text containing Chinese is assigned to expert network B for processing. For texts in multiple languages, the server can assign the text to multiple corresponding expert networks for processing, thereby aggregating the processing results of multiple expert networks to obtain the final processing result.

[0072] Alternatively, multiple training samples may be images representing different conditions, including those reflecting heart disease, respiratory problems, and liver disease. The processing model is a diagnostic model for triage. Expert Network A in the processing model processes images representing heart disease; Expert Network B processes images representing respiratory problems; and Expert Network C processes images representing liver disease. The server uses the gating network in the processing model to determine the type of condition the image represents and then assigns the image to an expert network capable of diagnosing that condition, resulting in a specific diagnosis. For example, the server assigns an image representing a heart disease to Expert Network A for processing.

[0073] In some embodiments, the server selects appropriate training samples for each expert network from the perspective of the expert network. Accordingly, the server determines the number of training samples to be processed by each expert network in the processing model. For each expert network in the processing model, the server processes the expert network's tasks and multiple training samples using the gating network in the processing model to obtain a matching result for the expert network. The matching result includes a degree of match between the expert network and the multiple training samples. Based on the matching result for the expert network, the server then determines the number of training samples to be processed by the expert network.

[0074] Among them, in the process of determining the sample number of training samples to be processed by the expert network based on the matching results of the expert network, the server sorts the multiple matching degrees in the matching results of the expert network in descending order. Then, the server selects the training samples whose matching degree reaches the reference degree as the training text to be processed by the expert network, and obtains the sample number of training samples to be processed by the expert network. The reference degree is the pre-set minimum matching degree, that is, the minimum standard. The solution provided in the embodiment of the present application, from the perspective of the expert network, selects multiple training samples with higher matching degrees for each expert network to process, so that each expert network can process training samples within its own capabilities, ensuring that the training samples can be accurately processed and improving the accuracy of the processing model.

[0075] In some embodiments, the server may select training samples for each expert network using the following formula 2.

[0076] Formula 2:

[0077] f(E,T)=top k (E,T)

[0078] Where E represents a set of multiple expert networks in the processing model; T represents a set of multiple training samples; and f(·) represents the matching function between the expert network and the training samples. This matching function is used to independently select the top k training samples with the highest matching degree for each expert network to process. k can be the sample capacity of the expert network. The sample capacity represents the maximum number of training samples that the expert network can process. The sample capacity depends not only on the processing power (performance) of the expert network, but also on the complexity and diversity of the training samples. In other words, the server can dynamically adjust the sample capacity of the expert network based on the processing power of the expert network and the complexity and diversity of the training samples. For example, the sample capacity of the expert network can be increased when processing more complex multilingual data.

[0079] The solution provided in the embodiment of the present application ensures that the load is evenly distributed among all expert networks by allowing the expert network to select training samples. This load balancing is crucial for improving the computational efficiency of the overall model, especially in large-scale data processing tasks. In addition, the reverse selection method also allows for more flexible allocation of resources such as computing training samples. For example, for some simpler training samples, only a few expert networks may be needed to effectively process them, while for complex training samples, more expert networks may be required to participate.

[0080] 303. The server determines the training loss based on the number of samples and the reference number of the multiple expert networks, where the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples and the reference number of the multiple expert networks.

[0081] In this embodiment of the present application, the server divides the number of training samples by the number of expert networks in the processing model to obtain a reference number. Then, for any expert network in the processing model, the server calculates the difference between the number of samples for that expert network and the reference number to obtain the training loss for that expert network. The server then sums the training losses for the multiple expert networks to obtain the training loss for the processing model.

[0082] In some embodiments, the server may calculate the training loss of the processing model using the following formula three.

[0083] Formula 3:

[0084] Where E represents the number (set) of multiple expert networks in the processing model; i represents the i-th expert network in the processing model; N i represents the number of samples assigned to the i-th expert network; T represents the total number of multiple training samples; L b represents the training loss, which balances the number of training samples processed by each expert network and is also called the balanced loss. This method considers the dynamic adjustment of the gating network's selection of expert networks for training samples. In the early stages of training, the processing model may tend to use a simpler selection strategy for rapid learning; as training progresses, more complex selection mechanisms can be gradually introduced to handle more complex data features. This balanced selection strategy allows for more efficient utilization of the individual expert networks in the processing model, improving the computational efficiency and processing power of the overall model. In particular, in large-scale multi-domain and multi-language processing tasks, the stable routing method can significantly improve performance through more balanced resource allocation and efficient model utilization.

[0085] In some embodiments, the processing capacity of each expert network is limited. The server can limit the number of training samples allocated to each expert network based on the sample capacity of the expert network. The sample capacity represents the maximum number of training samples that the expert network can process. The number of training samples to be processed by each expert network must not exceed the sample capacity of the expert network. The sample capacity can be determined based on the performance of the expert network. The performance of the expert network includes at least one of the efficiency or accuracy of the expert network in processing training samples. For any expert network, the server can determine the efficiency and accuracy of the expert network in processing training samples based on the expert network's historical processing of training samples. Accordingly, the server can also adjust the sample capacity of the expert network based on the expert network's historical processing of training samples. For example, if expert network B demonstrates higher accuracy and efficiency in processing a specific type of text (such as technical documents), the server may allocate more of this type of text to expert network B. This method can also be viewed as a method of reversely selecting training samples from the perspective of the expert network. The server can train the processing model using this reverse selection of training samples. By allowing the expert network to reversely select training samples, by controlling the sample capacity balance of each expert network, it is possible to ensure that the training samples are evenly distributed among all expert networks. That is, the training samples processed by each expert network are made as similar as possible. This not only improves the utilization rate of the expert network, but also avoids the situation where some expert networks reduce the processing efficiency of the processing model due to too many samples to be processed. It also enables each expert network to learn knowledge of multiple sample types, improves the generalization ability of the expert network, and thus processes samples more accurately based on the knowledge of multiple sample types, which is conducive to improving the accuracy of the processing model.

[0086] The server can dynamically adjust the sample size of each expert network through reinforcement learning. Alternatively, additional metrics can be introduced to evaluate the performance of the expert network, such as processing speed, accuracy, and resource consumption, so that these factors can be taken into account when selecting the expert model.

[0087] 304. For any training sample among the multiple training samples, before assigning the training sample to any expert network for processing, the server determines the contribution of the training sample based on the training sample and the expert network that will process the training sample. The contribution indicates the degree of contribution of assigning the training sample to the expert network for processing to the model training.

[0088] In an embodiment of the present application, the server can also combine the sample selection method of reinforcement learning and graph theory algorithm to significantly improve the efficiency and accuracy of the processing model. Among them, for any training sample among multiple training samples, the server determines the contribution of the training sample based on the training sample and the expert network that will process the training sample, so as to adjust the distribution according to the contribution during the training process. That is, the server can use the contribution of the expert network in processing the training sample as supervision information to perform self-supervised learning (training) on ​​the processing model. During the training process, the contribution of the training sample processed by the expert network is adjusted to the situation of the expert network distributing the training sample, so that the situation of the expert network processing the training sample after the adjustment distribution can bring a higher contribution to the model training, so that each training sample can be processed by the appropriate expert network, thereby helping to improve the accuracy of the processing model.

[0089] In some embodiments, the server may use the following formula 4 to calculate the contribution of assigning a training sample to a certain expert network to process the entire model training.

[0090] Formula 4:

[0091] Q(s,a)=R(s,a)+γ∑s′P(s′∣s,a)maxa′Q(s′,a′)

[0092] Here, s represents the network parameters and training sample of the processing model, that is, the state currently participating in training; a represents the assignment (selection) action of "assigning this training sample to a specific expert network"; R(s,a) represents the immediate reward for selecting action a in state s; Q(s,a) represents the decision selection function, that is, the expected return of selecting action a in state s; γ represents the weighting factor; ∑s′P(s′|s,a) represents the state transition probability; s′ represents the next state to transition to after selecting action a; and a′ represents the optimal action selected in state s′. This method not only considers the matching of individual training samples to expert networks but also applies graph theory analysis to account for the interactions and collaborative work between expert models, thereby optimizing the performance of the entire processing model.

[0093] 305. The server trains the processing model based on the training loss and the contribution of multiple training samples.

[0094] In an embodiment of the present application, the server trains the processing model with the goal of minimizing training loss and maximizing the contribution of multiple training samples. During the training of the processing model, the server can adjust the network parameters of the gating network in the processing model so that the gating network can evenly distribute multiple training samples to multiple expert networks; or it can also adjust the network parameters of each expert network in the processing model so that each expert network can accurately process the assigned training samples, which is not limited in this embodiment of the present application.

[0095] An embodiment of the present application provides a model training method. For a processing network including multiple expert networks, during the training of this type of processing model, the number of training samples to be processed by each expert network can be determined based on multiple training samples and the tasks of each expert network. That is, the multiple training samples are allocated to the corresponding expert network according to the sample types that the expert network can process for training, so that each expert model can accurately learn the knowledge of each sample type, thereby more accurately processing samples of the corresponding type; and the training loss is determined by determining the number of samples and the reference number of the multiple expert networks, so that the training samples are distributed to each expert network as evenly as possible, that is, the training samples processed by each expert network are as similar as possible. This not only improves the utilization rate of the expert network, but also avoids the reduction of the processing efficiency of the processing model due to too many samples to be processed by some expert networks. It also enables each expert network to learn knowledge of multiple sample types, improves the generalization ability of the expert network, and thus processes samples more accurately based on the knowledge of multiple sample types, which is conducive to improving the accuracy of the processing model.

[0096] FIG4 is a block diagram of a model training device according to an embodiment of the present application. The model training device is used to execute the steps of the above-mentioned model training method. Referring to FIG4 , the model training device includes: an acquisition module 401, a processing module 402, a first determination module 403, and a training module 404.

[0097] An acquisition module 401 is used to acquire multiple training samples and a processing model, where the processing model includes a gated network and multiple expert networks;

[0098] Processing module 402, configured to process the plurality of training samples and the tasks of the plurality of expert networks through the gating network in the processing model, and determine the number of training samples to be processed by each expert network in the processing model, where the tasks of each expert network represent the types of training samples that the expert network can process;

[0099] A first determining module 403 is configured to determine a training loss based on the number of samples of the multiple expert networks and a reference number, where the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the difference between the number of samples of the multiple expert networks and the reference number;

[0100] The training module 404 is used to train the processing model based on the training loss.

[0101] In some embodiments, FIG5 is a block diagram of another model training device provided according to an embodiment of the present application. Referring to FIG5 , the processing module 402 includes:

[0102] A first determining unit 4021 is configured to process, for any one of the plurality of training samples, the tasks of the training sample and the plurality of expert networks using a gating network in the processing model, and determine a matching result of the training sample, the matching result including a matching degree between the training sample and the plurality of expert networks;

[0103] The second determining unit 4022 is configured to determine the number of training samples to be processed by each expert network in the processing model based on the matching results of the multiple training samples.

[0104] In some embodiments, referring to FIG5 , the first determination unit 4021 is configured to extract features of any one of the multiple training samples by processing the gating network in the model to obtain sample features of the training sample, where the sample features represent the type of the training sample; extract features of the tasks of any one of the multiple expert networks by processing the gating network in the model to obtain task features of the expert network; and determine the similarity between the sample features and the task features of each expert network by processing the gating network in the model to obtain a matching result of the training sample.

[0105] In some embodiments, referring to FIG5 , the processing module 402 includes:

[0106] The processing unit 4023 is configured to process, for any expert network in the processing model, the tasks of the expert network and the multiple training samples through the gating network in the processing model to obtain a matching result of the expert network, where the matching result includes a matching degree between the expert network and the multiple training samples;

[0107] The third determining unit 4024 is configured to determine the number of training samples to be processed by the expert network based on the matching result of the expert network.

[0108] In some embodiments, referring to FIG5 , the third determining unit 4024 is configured to sort the multiple matching degrees in the matching results of the expert network in descending order; select training samples whose matching degrees reach the reference degree as the training texts to be processed by the expert network, and obtain the number of training samples to be processed by the expert network.

[0109] In some embodiments, the number of training samples to be processed by each expert network does not exceed the sample capacity of the expert network. The sample capacity represents the maximum number of training samples that the expert network can process. The sample capacity is determined based on the performance of the expert network. The performance of the expert network includes at least one of the efficiency or accuracy of the expert network in processing training samples.

[0110] In some embodiments, referring to FIG5 , the apparatus further includes:

[0111] A second determining module 405 is configured to determine, for any training sample from the plurality of training samples, a contribution of the training sample based on the training sample and the expert network that will process the training sample before assigning the training sample to any expert network for processing, where the contribution indicates the extent to which assigning the training sample to the expert network for processing contributes to model training;

[0112] The training module 404 is used to train the processing model based on the training loss and the contribution of multiple training samples.

[0113] An embodiment of the present application provides a model training device. For a processing network including multiple expert networks, during the training of this type of processing model, the number of training samples to be processed by each expert network can be determined based on multiple training samples and the tasks of each expert network. That is, the multiple training samples are allocated to the corresponding expert network according to the sample types that the expert network can process for training, so that each expert model can accurately learn the knowledge of each sample type, thereby more accurately processing samples of the corresponding type; and the training loss is determined by determining the number of samples and the reference number of the multiple expert networks, so that the training samples are distributed to each expert network as evenly as possible, that is, the number of training samples processed by each expert network is the same. This not only improves the utilization rate of the expert network, but also avoids the reduction of the processing efficiency of the processing model due to too many samples to be processed by some expert networks. It also enables each expert network to learn knowledge of multiple sample types, improves the generalization ability of the expert network, and thus more accurately processes samples based on the knowledge of multiple sample types, which is conducive to improving the accuracy of the processing model.

[0114] It should be noted that the model training device provided in the above embodiment only uses the division of the above functional modules as an example when running an application. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. In addition, the model training device provided in the above embodiment and the model training method embodiment are based on the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

[0115] In the embodiments of the present application, the computer device can be configured as a terminal or a server. When the computer device is configured as a terminal, the terminal can be used as the execution subject to implement the technical solution provided in the embodiments of the present application. When the computer device is configured as a server, the server can be used as the execution subject to implement the technical solution provided in the embodiments of the present application. The technical solution provided in the present application can also be implemented through interaction between the terminal and the server. The embodiments of the present application do not limit this.

[0116] Figure 6 is a block diagram of a terminal 600 according to an embodiment of the present application. Terminal 600 may be a portable mobile terminal, such as a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. Terminal 600 may also be referred to as user equipment, portable terminal, laptop terminal, desktop terminal, or other similar names.

[0117] Typically, the terminal 600 includes a processor 601 and a memory 602 .

[0118] The processor 601 may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor 601 may be implemented in at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array). The processor 601 may also include a main processor and a coprocessor. The main processor is a processor for processing data in the awake state, also known as a CPU (Central Processing Unit); the coprocessor is a low-power processor for processing data in the standby state. In some embodiments, the processor 601 may be integrated with a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the display screen. In some embodiments, the processor 601 may also include an AI (Artificial Intelligence) processor, which is used to process computing operations related to machine learning.

[0119] The memory 602 may include one or more computer-readable storage media, which may be non-transitory. The memory 602 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices and flash memory storage devices. In some embodiments, the non-transitory computer-readable storage medium in the memory 602 is used to store at least one computer program, which is used to be executed by the processor 601 to implement the model training method provided in the method embodiment of the present application.

[0120] In some embodiments, terminal 600 may optionally include a peripheral device interface 603 and at least one peripheral device. Processor 601, memory 602, and peripheral device interface 603 may be connected via a bus or signal lines. Each peripheral device may be connected to peripheral device interface 603 via a bus, signal lines, or circuit boards. Specifically, the peripheral device may include at least one of a radio frequency circuit 604, a display screen 605, a camera assembly 606, an audio circuit 607, and a power supply 608.

[0121] The peripheral device interface 603 can be used to connect at least one I / O (Input / Output)-related peripheral device to the processor 601 and the memory 602. In some embodiments, the processor 601, the memory 602, and the peripheral device interface 603 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 601, the memory 602, and the peripheral device interface 603 can be implemented on separate chips or circuit boards, which is not limited in this embodiment.

[0122] The radio frequency circuit 604 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency circuit 604 communicates with communication networks and other communication devices via electromagnetic signals. The radio frequency circuit 604 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. In some embodiments, the radio frequency circuit 604 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, etc. The radio frequency circuit 604 can communicate with other terminals via at least one wireless communication protocol. Such wireless communication protocols include, but are not limited to, the World Wide Web, metropolitan area networks, intranets, various generations of mobile communication networks (2G, 3G, 4G, and 5G), wireless local area networks, and / or WiFi (Wireless Fidelity) networks. In some embodiments, the radio frequency circuit 604 may also include circuits related to NFC (Near Field Communication), which is not limited in this application.

[0123] Display screen 605 is used to display a user interface (UI). This UI may include graphics, text, icons, videos, or any combination thereof. If display screen 605 is a touchscreen display, it is also capable of collecting touch signals on or above the surface of display screen 605. These touch signals can be input as control signals to processor 601 for processing. Display screen 605 can also be used to provide virtual buttons and / or a virtual keyboard, also known as soft buttons and / or a soft keyboard. In some embodiments, there can be a single display screen 605, located on the front panel of terminal 600. In other embodiments, there can be at least two display screens 605, located on different surfaces of terminal 600 or in a foldable design. In still other embodiments, display screen 605 can be a flexible display screen, located on a curved or foldable surface of terminal 600. Display screen 605 can also be configured as a non-rectangular, irregular shape, also known as a special-shaped screen. Display screen 605 can be made of materials such as LCD (Liquid Crystal Display) and OLED (Organic Light-Emitting Diode).

[0124] The camera assembly 606 is used to capture images or videos. In some embodiments, the camera assembly 606 includes a front camera and a rear camera. Typically, the front camera is arranged on the front panel of the terminal, and the rear camera is arranged on the back of the terminal. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize panoramic shooting and VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera assembly 606 may also include a flash. The flash can be a single-color temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.

[0125] The audio circuit 607 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals that are input into the processor 601 for processing, or input into the radio frequency circuit 604 to achieve voice communication. For the purpose of stereo sound collection or noise reduction, there may be multiple microphones, each located in different parts of the terminal 600. The microphone may also be an array microphone or an omnidirectional collection microphone. The speaker is used to convert electrical signals from the processor 601 or the radio frequency circuit 604 into sound waves. The speaker may be a traditional thin film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert electrical signals into sound waves audible to humans, but also convert electrical signals into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio circuit 607 may also include a headphone jack.

[0126] Power supply 608 is used to power various components in terminal 600. Power supply 608 can be AC ​​power, DC power, a disposable battery, or a rechargeable battery. When power supply 608 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is charged via a wired line, while a wireless rechargeable battery is charged via a wireless coil. The rechargeable battery can also support fast charging technology.

[0127] In some embodiments, the terminal 600 further includes one or more sensors 609 , including but not limited to: an acceleration sensor 610 , a gyroscope sensor 611 , a pressure sensor 612 , an optical sensor 613 , and a proximity sensor 614 .

[0128] The accelerometer 610 can detect the magnitude of acceleration along the three coordinate axes of the coordinate system established by the terminal 600. For example, the accelerometer 610 can be used to detect the components of gravity acceleration along the three coordinate axes. The processor 601 can control the display screen 605 to display the user interface in a landscape or portrait view based on the gravity acceleration signal collected by the accelerometer 610. The accelerometer 610 can also be used to collect game or user motion data.

[0129] The gyroscope sensor 611 can detect the orientation and rotation angle of the terminal 600. It can also work with the accelerometer 610 to collect the user's 3D movements of the terminal 600. Based on the data collected by the gyroscope sensor 611, the processor 601 can implement the following functions: motion sensing (for example, changing the UI based on the user's tilt operation), image stabilization during shooting, game control, and inertial navigation.

[0130] The pressure sensor 612 can be provided on the side frame of the terminal 600 and / or below the display screen 605. When the pressure sensor 612 is provided on the side frame of the terminal 600, it can detect the user's gripping signal of the terminal 600. The processor 601 performs left and right hand recognition or shortcut operations based on the gripping signal collected by the pressure sensor 612. When the pressure sensor 612 is provided below the display screen 605, the processor 601 controls the operable controls on the UI interface based on the user's pressure operation on the display screen 605. Operable controls include at least one of a button control, a scroll bar control, an icon control, and a menu control.

[0131] The optical sensor 613 is used to detect ambient light intensity. In one embodiment, the processor 601 can control the display brightness of the display screen 605 based on the ambient light intensity detected by the optical sensor 613. Specifically, when the ambient light intensity is high, the display brightness of the display screen 605 is increased; when the ambient light intensity is low, the display brightness of the display screen 605 is decreased. In another embodiment, the processor 601 can also dynamically adjust the shooting parameters of the camera assembly 606 based on the ambient light intensity detected by the optical sensor 613.

[0132] Proximity sensor 614, also known as a distance sensor, is typically located on the front panel of terminal 600. Proximity sensor 614 is used to detect the distance between the user and the front of terminal 600. In one embodiment, when proximity sensor 614 detects that the distance between the user and the front of terminal 600 is gradually decreasing, processor 601 controls display screen 605 to switch from the screen-on state to the screen-off state. When proximity sensor 614 detects that the distance between the user and the front of terminal 600 is gradually increasing, processor 601 controls display screen 605 to switch from the screen-off state to the screen-on state.

[0133] Those skilled in the art will appreciate that the structure shown in FIG6 does not limit the terminal 600 and may include more or fewer components than shown, or combine certain components, or adopt a different component arrangement.

[0134] FIG7 is a schematic diagram of the structure of a server provided according to an embodiment of the present application. The server 700 may have relatively large differences due to different configurations or performances, and may include one or more processors (Central Processing Units, CPU) 701 and one or more memories 702, wherein the memory 702 stores at least one computer program, and the at least one computer program is loaded and executed by the processor 701 to implement the model training method provided by the above-mentioned various method embodiments. Of course, the server may also have components such as a wired or wireless network interface, a keyboard, and an input and output interface for input and output. The server may also include other components for implementing device functions, which will not be described in detail here.

[0135] The embodiment of the present application further provides a computer-readable storage medium, wherein the computer-readable storage medium stores at least one computer program, and the at least one computer program is loaded and executed by a processor of a computer device to implement the operations performed by the computer device in the model training method of the above embodiment. For example, the computer-readable storage medium can be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), a magnetic tape, a floppy disk, an optical data storage device, etc.

[0136] The present application also provides a computer program product, including a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the model training methods provided in the various optional implementations described above.

[0137] Those skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware, or by a program to instruct the relevant hardware, and the program may be stored in a computer-readable storage medium, which may be a read-only memory, a disk, or an optical disk, etc.

[0138] The above description is merely an optional embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. A model training method, comprising: Acquire multiple training samples and a processing model, wherein the processing model includes a gating network and multiple expert networks; Processing the plurality of training samples and the tasks of the plurality of expert networks through the gating network in the processing model to determine the number of training samples to be processed by each expert network in the processing model, wherein the task of each expert network represents the type of training samples that the expert network can process; Determining a training loss based on the number of samples of the multiple expert networks and a reference number, where the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples of the multiple expert networks and the reference number; The processing model is trained based on the training loss.

2. The method according to claim 1, wherein The processing of the plurality of training samples and the tasks of the plurality of expert networks by the gating network in the processing model to determine the number of training samples to be processed by each expert network in the processing model includes: For any training sample among the multiple training samples, processing the training sample and the tasks of the multiple expert networks through the gating network in the processing model to determine a matching result of the training sample, wherein the matching result includes a matching degree between the training sample and the multiple expert networks; Based on the matching results of the multiple training samples, the number of training samples to be processed by each expert network in the processing model is determined.

3. The method according to claim 2, wherein: For any one of the multiple training samples, processing the training sample and the tasks of the multiple expert networks through the gating network in the processing model to determine a matching result of the training sample includes: For any training sample among the multiple training samples, extract features of the training sample through the gating network in the processing model to obtain a sample feature of the training sample, where the sample feature represents a type of the training sample; For any expert network among the multiple expert networks, extracting features of the task of the expert network through the gating network in the processing model to obtain task features of the expert network; The similarity between the sample features and the task features of each expert network is determined through the gating network in the processing model to obtain the matching result of the training sample.

4. The method according to claim 1, wherein The processing of the plurality of training samples and the tasks of the plurality of expert networks by the gating network in the processing model to determine the number of training samples to be processed by each expert network in the processing model includes: For any expert network in the processing model, processing the task of the expert network and the multiple training samples through the gating network in the processing model to obtain a matching result of the expert network, wherein the matching result includes a matching degree between the expert network and the multiple training samples; Based on the matching results of the expert network, the number of training samples to be processed by the expert network is determined.

5. The method according to claim 4, wherein The determining, based on the matching results of the expert network, the number of training samples to be processed by the expert network includes: sorting the multiple matching degrees in the matching results of the expert network in descending order; The training samples whose matching degree reaches the reference degree are selected as the training texts to be processed by the expert network, and the number of training samples to be processed by the expert network is obtained.

6. The method according to claim 1, wherein The number of training samples to be processed by each expert network does not exceed the sample capacity of the expert network. The sample capacity represents the maximum number of training samples that the expert network can process. The sample capacity is determined based on the performance of the expert network. The performance of the expert network includes at least one of the efficiency or accuracy of the expert network in processing training samples.

7. The method according to claim 1, wherein The method further comprises: For any training sample among the multiple training samples, before assigning the training sample to any expert network for processing, determining a contribution of the training sample based on the training sample and the expert network that will process the training sample, the contribution indicating the extent to which assigning the training sample to the expert network for processing contributes to model training; The step of training the processing model based on the training loss includes: The processing model is trained based on the training loss and the contributions of the multiple training samples.

8. A model training device, comprising: An acquisition module, configured to acquire a plurality of training samples and a processing model, wherein the processing model includes a gated network and a plurality of expert networks; a processing module, configured to process the plurality of training samples and the tasks of the plurality of expert networks through the gating network in the processing model, and determine the number of training samples to be processed by each expert network in the processing model, wherein the tasks of each expert network represent the types of training samples that the expert network can process; a first determining module, configured to determine a training loss based on the number of samples of the multiple expert networks and a reference number, wherein the reference number is the number of training samples allocated to each expert network when the multiple training samples are evenly distributed, and the training loss represents the gap between the number of samples of the multiple expert networks and the reference number; A training module is used to train the processing model based on the training loss.

9. A computer device comprising a processor and a memory, wherein the memory is used to store at least one computer program, and the at least one computer program is loaded by the processor and executes the model training method described in any one of claims 1 to 7.

10. A computer-readable storage medium, wherein the computer-readable storage medium is used to store at least one computer program, and the at least one computer program is used to execute the model training method described in any one of claims 1 to 7.

11. A computer program product, comprising a computer program, wherein when the computer program is executed by a processor, the model training method according to any one of claims 1 to 7 is implemented.

Images