Proxy model training method and device, server and storage medium
By acquiring multi-class modal data to generate a feature dictionary, determining the feature contribution, and constructing a composite feature dictionary, the surrogate model is trained, solving the problem of poor training effect of single modal data and achieving efficient training and accurate output in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNITED NETWORK COMM GRP CO LTD
- Filing Date
- 2025-12-30
- Publication Date
- 2026-05-08
AI Technical Summary
Existing technologies train surrogate models using modal data of a single category, resulting in poor training performance and making them unsuitable for complex application environments.
Multiple modal data of different categories are obtained from the modal data information dictionary, a feature dictionary of the corresponding modality is generated, the usage frequency and contribution of feature atoms are determined, target feature atoms are screened, a composite feature dictionary is constructed, and the surrogate model is trained based on the composite feature dictionary.
This improved the training performance of the surrogate model, making it suitable for complex application environments and enhancing the feedback accuracy and output accuracy of the surrogate model.
Smart Images
Figure CN121997044A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and in particular to a method, apparatus, server and storage medium for training a proxy model. Background Technology
[0002] Agent training, or AI Agent training, is the process by which a server cultivates an autonomous, goal-oriented intelligent agent. The core of an AI Agent is that it is endowed with the ability to proactively complete tasks, autonomously perceive the environment, make decisions, and execute those decisions to achieve its goals.
[0003] In the existing technology, traditional surrogate model training methods train surrogate models by using modal data of a single class and combining hierarchical feature fusion and dynamic attention mechanisms.
[0004] However, in existing technologies, the training method of training surrogate models using modal data of a single category has a single data type, resulting in poor training performance of surrogate models and making it unsuitable for complex application environments. Summary of the Invention
[0005] This application provides a surrogate model training method, apparatus, server, and storage medium to solve the problem that training a surrogate model using modal data of a single category has a single data type, resulting in poor training performance and unsuitability for complex application environments.
[0006] Firstly, this application provides a proxy model training method applied to a server, comprising:
[0007] When it is necessary to train a preset proxy model, modal data corresponding to multiple different categories of modalities are obtained from the modal data information dictionary;
[0008] Each modal data is encoded to generate a feature dictionary for the corresponding modality, where each feature dictionary is used to characterize the feature representation of the corresponding modality;
[0009] Determine the usage frequency of each feature atom in each feature dictionary;
[0010] The feature contribution of the corresponding mode is quantified based on the usage frequency of each feature atom.
[0011] Multiple target feature atoms are selected from the feature contribution of each mode;
[0012] Construct a composite feature dictionary based on each target feature atom;
[0013] The proxy model is trained based on the composite feature dictionary to complete the training of the proxy model.
[0014] In one possible design, determining the usage frequency of each feature atom in each feature dictionary includes: determining the multiple feature atoms contained in each feature dictionary; determining the number of times each feature atom is used in the corresponding feature dictionary; determining the total number of times all feature atoms are used in the corresponding feature dictionary based on the number of times each feature atom is used; and determining the usage frequency of each feature atom based on the number of times each feature atom is used and the total number of times it is used.
[0015] In one possible design, quantifying the feature contribution of the corresponding mode based on the usage frequency of each feature atom includes: setting a time decay factor and determining the usage time of each feature atom; setting a weight corresponding to the usage frequency of each feature atom based on the usage time and the time decay factor; and quantifying the feature contribution of the corresponding mode based on the weight and the usage frequency of each feature atom.
[0016] In one possible design, the step of selecting multiple target feature atoms from the feature contribution of each modality includes: setting a selection rule for the target feature atoms; adjusting the selection rule to complete the adjustment of the selection rule; and after the selection rule is adjusted, selecting multiple target feature atoms from the feature contribution of each modality.
[0017] In one possible design, after training the proxy model according to the composite feature dictionary to complete the training of the proxy model, the method further includes: after the proxy model is trained, testing the proxy model and performing the following steps: acquiring multiple test modality data, wherein each test modality data is configured with a standard question-answering file; performing question-answering tests on the proxy model according to each test modality data and the standard question-answering file to obtain the number of correct question-answering results; determining the feature contribution degree corresponding to each test modality data; determining the question-answering accuracy of the proxy model according to the number of correct question-answering results and the feature contribution degree corresponding to each test modality data; and updating the proxy model according to the question-answering accuracy to complete the update of the proxy model.
[0018] In one possible design, testing the proxy model based on each test modality data and the standard question-and-answer file to obtain the correct number of test results includes: inputting each test modality data into the proxy model for question-and-answer processing to obtain multiple question-and-answer results; comparing each question-and-answer result with the standard question-and-answer file to obtain comparison results; and determining the correct number corresponding to the multiple question-and-answer results based on the comparison results.
[0019] In one possible design, the method further includes: constructing local feature dictionaries and global feature dictionaries for each modality based on a pre-defined framework distributed algorithm and data from each modality; integrating the local feature dictionaries and global feature dictionaries into a feature dictionary corresponding to each modality; and aggregating the feature dictionaries into a composite feature dictionary using a pre-defined aggregation algorithm.
[0020] Secondly, this application provides a proxy model training device applied to a server, comprising:
[0021] The first acquisition module is used to acquire modal data corresponding to multiple different categories of modalities from the modal data information dictionary when it is necessary to train a preset agent model;
[0022] The encoding module is used to encode the data of each modality to generate the feature dictionary of the corresponding modality, where each feature dictionary is used to characterize the feature representation of the corresponding modality;
[0023] The first determining module is used to determine the usage frequency of each feature atom in each feature dictionary;
[0024] The quantization module is used to quantify the feature contribution of the corresponding mode based on the usage frequency of each feature atom.
[0025] The filtering module is used to filter multiple target feature atoms from the feature contribution of each mode;
[0026] The first building module is used to construct a composite feature dictionary based on each target feature atom;
[0027] The training module is used to train the proxy model based on the composite feature dictionary to complete the training of the proxy model.
[0028] Thirdly, this application provides a server, including: at least one processor and a memory;
[0029] The memory stores computer-executed instructions;
[0030] The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the surrogate model training method as described in the first aspect and various possible designs of the first aspect.
[0031] Fourthly, this application provides a computer storage medium storing computer execution instructions, which, when executed by a processor, implement the proxy model training method described in the first aspect and various possible designs of the first aspect.
[0032] The surrogate model training method, apparatus, server, and storage medium provided in this application, when training a preset surrogate model, obtain modal data corresponding to multiple different categories of modalities from a modal data information dictionary; encode each modal data to generate a feature dictionary for the corresponding modality; determine the usage frequency of each feature atom in each feature dictionary; quantify the feature contribution of the corresponding modality based on the usage frequency of each feature atom; select multiple target feature atoms from the feature contribution of each modality; construct a composite feature dictionary based on each target feature atom; and train the surrogate model based on the composite feature dictionary to complete the training of the surrogate model. By processing the modal data of various modalities to determine the composite feature dictionary constructed by each target feature atom, and training the surrogate model based on the composite feature dictionary, the training effect of the surrogate model is improved, and it is suitable for complex application environments. Attached Figure Description
[0033] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0034] Figure 1 This is a schematic diagram illustrating an application scenario of the proxy model training method provided in the embodiments of this application;
[0035] Figure 2 Schematic diagram of the proxy model training method provided in the embodiments of this application Figure 1 ;
[0036] Figure 3 Schematic diagram of the proxy model training method provided in the embodiments of this application Figure 2 ;
[0037] Figure 4 This is a schematic diagram of the structure of the proxy model training device provided in the embodiments of this application;
[0038] Figure 5 This is a schematic diagram of the hardware structure of the server provided in an embodiment of this application. Detailed Implementation
[0039] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0040] Agent model training, or AI agent training, is the process by which a server cultivates an autonomous, goal-oriented intelligent agent. The core of an AI agent is its ability to proactively complete tasks, autonomously perceiving its environment, making decisions, and executing those decisions to achieve its goals. Current technologies, particularly traditional agent model training methods, employ single-class modal data and techniques such as hierarchical feature fusion and dynamic attention mechanisms to train the agent model. However, this single-class modal data approach results in limited data variety, leading to poor training performance and unsuitability for complex application environments.
[0041] To address the aforementioned technical problems, this application proposes the following technical concept: Considering the modal data corresponding to each category of modality, the inventors generate a feature dictionary for each modality based on the modal data, construct a composite feature dictionary using each feature dictionary, and train the proxy model based on the composite feature dictionary to complete the training of the proxy model, thereby improving the training effect of the proxy model and making it suitable for complex application environments.
[0042] Figure 1 This is a schematic diagram illustrating an application scenario of the proxy model training method provided in this application embodiment.
[0043] like Figure 1 As shown, the scenario includes a display terminal 101 and a server 102.
[0044] The display terminal 101 can be a display screen or a personal computer or other terminal.
[0045] Server 102 can be a standalone server or a cluster of multiple servers.
[0046] Server 102, when maintenance personnel need to train the proxy model, obtains modal data corresponding to multiple different categories of modalities from the modal data information dictionary; encodes each modal data to generate a feature dictionary for the corresponding modality; determines the usage frequency of each feature atom in each feature dictionary; quantifies the feature contribution of the corresponding modality based on the usage frequency of each feature atom; selects multiple target feature atoms from the feature contribution of each modality; constructs a composite feature dictionary based on each target feature atom; trains the proxy model based on the composite feature dictionary to complete the training of the proxy model, and outputs a prompt indicating that the proxy model training is complete to the display terminal 101. The following detailed implementation example illustrates this process.
[0047] Figure 2 Schematic diagram of the proxy model training method provided in the embodiments of this application Figure 1 The execution entity in this embodiment can be Figure 1 The server in the illustrated embodiment is not specifically limited in this embodiment. Figure 2 As shown, the method includes:
[0048] S201: When it is necessary to train a preset proxy model, obtain modal data corresponding to multiple different categories of modalities from the modal data information dictionary.
[0049] For example, the modalities include text modal, audio modal, image modal, and gesture modal.
[0050] For example, the modal data corresponding to each mode is: The dictionary matrix corresponding to the modal data of the i-th modality is: , This refers to the exponentiation operation of a matrix G with s rows and m columns.
[0051] S202: Encode each modal data to generate a feature dictionary for the corresponding modality, where each feature dictionary is used to characterize the feature representation of the corresponding modality.
[0052] In this embodiment, the formula for encoding each modal data to generate the feature dictionary of the corresponding modality is as follows:
[0053]
[0054] In the formula, Let W be the feature dictionary for the i-th modality; W is the weight parameter matrix. The set of values for Z and W to obtain the minimum value; This is the dictionary matrix corresponding to the i-th modal data; Measured The degree of difference; These are regularization coefficients used to restrict the matrix. The degree of sparsity; Let be the L1 norm of W.
[0055] S203: Determine the usage frequency of each feature atom in each feature dictionary.
[0056] Specifically, step S203 includes steps a~d:
[0057] Step a: Determine the multiple feature atoms contained in each feature dictionary.
[0058] Step b: Determine the number of times each feature atom is used in the corresponding feature dictionary.
[0059] For example, the number of times each feature atom is used in the corresponding feature dictionary is set to n.
[0060] Step c: Determine the total number of times each feature atom is used in the corresponding feature dictionary based on the number of times each feature atom is used.
[0061] For example, the total number of times all feature atoms are used in the corresponding feature dictionary is set to N.
[0062] Step d: Determine the usage frequency of each feature atom based on the number of uses and the total number of uses.
[0063] In this embodiment, the formula for calculating the usage frequency of each feature atom based on its individual usage counts and total usage counts is as follows:
[0064]
[0065] In the formula, Let i be the frequency of use of the i-th feature atom. ; is the number of times the i-th feature atom is used; N is the total number of times all feature atoms are used.
[0066] also, It can also be defined as the contribution of the i-th feature atom.
[0067] S204: Quantify the feature contribution of the corresponding mode based on the usage frequency of each feature atom.
[0068] In this embodiment, the formula for calculating the feature contribution of the corresponding mode based on the usage frequency of each feature atom is as follows:
[0069]
[0070] In the formula, For the feature contribution of any mode, 0 ≤ ≤m; The frequency of use of the i-th feature atom in the feature dictionary corresponding to any modality data, 1≤ ≤m.
[0071] Specifically, step S204 includes steps a~c:
[0072] Step a: Set the time decay factor and determine the usage time of each feature atom.
[0073] For example, the time decay factor is set to γ. The usage time of each characteristic atom is determined to be t.
[0074] Step b: Set the weight corresponding to the usage frequency of each feature atom according to the usage time and time decay factor.
[0075] Specifically, each usage time is divided into near-term and far-term times, and the weights corresponding to the usage frequency of each feature atom are set according to the near-term and far-term times corresponding to each usage time, as well as the time decay factor.
[0076] Step c: Quantify the feature contribution of the corresponding mode based on the usage frequency of each weight and feature atom.
[0077] For example, if a time decay factor is introduced into the feature contribution calculation and more recent times are given higher weight, then the formula for calculating the feature contribution of the corresponding mode based on each weight and the usage frequency of each feature atom is as follows:
[0078]
[0079] In the formula, C m,i The feature contribution of feature atom i in mode m; The weight corresponding to the usage frequency of the i-th feature atom, set based on the time decay factor; It is the number of times the i-th feature atom in mode m is used at time t; It is a three-dimensional data index of mode-atom-time, used to record the number of times feature atom j is used in mode m at time t.
[0080] S205: Select multiple target feature atoms from the feature contribution of each mode.
[0081] Specifically, step S205 includes steps a~c:
[0082] Step a: Set the filtering rules for target feature atoms.
[0083] For example, the selection rule for target feature atoms is: Select And the feature contribution under different modes Characteristic atoms >0.75.
[0084] Step b: Adjust the filtering rules to complete the filtering rule adjustment.
[0085] Specifically, the feature contribution in the screening rules is adjusted according to actual needs to complete the adjustment of the screening rules.
[0086] Step c: After the screening rules are adjusted, select multiple target feature atoms from the feature contribution of each mode.
[0087] S206: Construct a composite feature dictionary based on the feature atoms of each target.
[0088] For example, the composite feature dictionary is as follows:
[0089]
[0090] In addition, the construction of the composite feature dictionary also includes steps a~c:
[0091] Step a: Based on the preset framework distributed algorithm and the data of each modality, construct the local feature dictionary and global feature dictionary for each modality.
[0092] In this embodiment, the preset framework distributed algorithm is a type of algorithm that runs on multiple interconnected distributed computer nodes without a central node, enabling multiple servers to work collaboratively to solve problems. Examples include consensus algorithms and distributed hash algorithms.
[0093] Step b: Integrate the local feature dictionaries and the global feature dictionaries into a feature dictionary corresponding to each modality.
[0094] Step c: Using a preset aggregation algorithm, aggregate the feature dictionaries into a composite feature dictionary.
[0095] In this embodiment, the preset aggregation algorithm can be either the median aggregation method or the arithmetic mean aggregation method.
[0096] S207: Train the surrogate model based on the composite feature dictionary to complete the training of the surrogate model.
[0097] Specifically, the surrogate model is trained based on the composite feature dictionary to complete the training of the surrogate model, which is then used for knowledge updates and learning optimization of the surrogate model.
[0098] In summary, the proxy model training method provided in this embodiment, when training a preset proxy model, obtains modal data corresponding to multiple different categories of modalities from the modal data information dictionary; encodes each modal data to generate a feature dictionary for the corresponding modality; determines the usage frequency of each feature atom in each feature dictionary; quantifies the feature contribution of the corresponding modality based on the usage frequency of each feature atom; selects multiple target feature atoms from the feature contribution of each modality; constructs a composite feature dictionary based on each target feature atom; and trains the proxy model based on the composite feature dictionary to complete the training of the proxy model. By processing the modal data of various modalities to determine the composite feature dictionary constructed by each target feature atom, and training the proxy model based on the composite feature dictionary, the training effect of the proxy model is improved, and it is suitable for complex application environments.
[0099] In addition, the surrogate model training method provided in this embodiment sets the selection rules for target feature atoms; adjusts the selection rules to complete the adjustment of the selection rules; after the selection rules are adjusted, selects multiple target feature atoms from the feature contribution of each modality, and uses the selected multiple target feature atoms to train the surrogate model, thereby improving the training effect of the surrogate model and thus improving the feedback accuracy of the surrogate model.
[0100] Furthermore, the surrogate model training method provided in this embodiment sets a time decay factor and determines the usage time of each feature atom; sets the weight corresponding to the usage frequency of each feature atom according to the usage time and the time decay factor; and quantifies the feature contribution of the corresponding mode according to the weight and the usage frequency of each feature atom. By quantifying the usage frequency of each feature atom and the feature contribution of the corresponding mode, quantitative data is provided for the subsequent training of the surrogate model.
[0101] Figure 3 Schematic diagram of the proxy model training method provided in the embodiments of this application Figure 2 In the embodiments of this application, in Figure 2 Based on the provided embodiments, a detailed explanation is given of the specific implementation method for updating the proxy model after S207. For example... Figure 3 As shown, the method includes:
[0102] S301: After the surrogate model is trained, test the surrogate model and perform the following steps:
[0103] S302: Acquire multiple test modal data, each of which is configured with a standard question and answer file.
[0104] S303: Based on the test modality data and standard question-and-answer file, perform question-and-answer tests on the proxy model to obtain the correct number of question-and-answer results.
[0105] Specifically, step S303 includes steps a~c:
[0106] Step a: Input the data of each test modality into the proxy model for question-and-answer processing to obtain multiple question-and-answer results;
[0107] Step b: Compare each question and answer result with the standard question and answer document to obtain the comparison results;
[0108] In this embodiment, each comparison result is either a correct comparison or an incorrect comparison.
[0109] Step c: Determine the correct number of multiple question-and-answer results based on the comparison results.
[0110] S304: Determine the feature contribution of each test modality data.
[0111] In this embodiment, the calculation of feature contribution has been explained in detail in step S204, and will not be repeated here.
[0112] S305: Determine the question-answering accuracy of the surrogate model based on the number of correct answers and the feature contribution of each test modality data.
[0113] In this embodiment, based on the number of correct answers and feature contribution of each test modality data, the formula for calculating the question-answering accuracy of the surrogate model is as follows:
[0114]
[0115] In the formula, For the question-answering accuracy of the proxy model; For the first The number of correct question-and-answer results under each test modality; Indicates will Mapped to the range of 0 to 1, it is essentially a coefficient; The feature contribution of the test mode.
[0116] in, This can be specifically interpreted as the first The number of correct question-and-answer results under each test modality; For the first Other quantities related to the number of correct question-and-answer results under each test modality.
[0117] S306: Update the proxy model based on the question-answering accuracy to complete the proxy model update.
[0118] In summary, the proxy model training method provided in this embodiment, after the proxy model is trained, tests the proxy model and performs the following steps: acquiring multiple test modal data, each of which is configured with a standard question-and-answer file; performing question-and-answer tests on the proxy model based on each test modal data and the standard question-and-answer file to obtain the number of correct question-and-answer results; determining the feature contribution degree corresponding to each test modal data; determining the question-and-answer accuracy of the proxy model based on the number of correct question-and-answer results and the feature contribution degree corresponding to each test modal data; and updating the proxy model based on the question-and-answer accuracy to complete the proxy model update. By testing the proxy model after training to obtain the question-and-answer accuracy and feeding the question-and-answer accuracy back to the proxy model for optimization and updating, the output accuracy of the subsequent proxy model is improved.
[0119] Figure 4 This is a schematic diagram of the surrogate model training device provided in an embodiment of this application. Figure 4 As shown, the surrogate model training device includes: a first acquisition module 401, an encoding module 402, a first determination module 403, a quantization module 404, a screening module 405, a first construction module 406, and a training module 407.
[0120] The first acquisition module 401 is used to acquire modal data corresponding to multiple different categories of modalities from the modal data information dictionary when it is necessary to train a preset proxy model;
[0121] The encoding module 402 is used to encode the modal data to generate a feature dictionary for the corresponding modality, wherein each feature dictionary is used to characterize the feature representation of the corresponding modality;
[0122] The first determining module 403 is used to determine the usage frequency of each feature atom in each feature dictionary;
[0123] Quantization module 404 is used to quantize the feature contribution of the corresponding mode based on the usage frequency of each feature atom.
[0124] The filtering module 405 is used to filter multiple target feature atoms from the feature contribution of each mode;
[0125] The first construction module 406 is used to construct a composite feature dictionary based on each target feature atom;
[0126] Training module 407 is used to train the surrogate model based on the composite feature dictionary to complete the training of the surrogate model.
[0127] In one possible implementation, the first determining module 403 specifically includes:
[0128] The first determining unit is used to determine the multiple feature atoms contained in each feature dictionary;
[0129] The second determining unit is used to determine the number of times each feature atom is used in the corresponding feature dictionary;
[0130] The third determining unit is used to determine the total number of times all feature atoms are used in the corresponding feature dictionary based on the number of times each feature atom is used.
[0131] The fourth determining unit is used to determine the usage frequency of each characteristic atom based on the number of uses and the total number of uses.
[0132] In one possible implementation, the quantization module 404 specifically includes:
[0133] The first setting unit is used to set the time decay factor and determine the usage time of each feature atom;
[0134] The second setting unit is used to set the weight corresponding to the usage frequency of each feature atom according to each usage time and time decay factor.
[0135] The quantization unit is used to quantize the feature contribution of the corresponding mode based on the usage frequency of each weight and feature atom.
[0136] In one possible implementation, the filtering module 405 specifically includes:
[0137] The setting unit is used to set the filtering rules for target feature atoms;
[0138] The adjustment unit is used to adjust the filtering rules in order to complete the adjustment of the filtering rules;
[0139] The filtering unit is used to filter multiple target feature atoms from the feature contribution of each mode after the filtering rules have been adjusted.
[0140] In one possible implementation, the device further includes:
[0141] The first testing module is used to test the proxy model after training, and performs the following steps:
[0142] The second acquisition module is used to acquire multiple test modality data, each of which is configured with a standard question and answer file;
[0143] The second testing module is used to perform question-and-answer tests on the proxy model based on the data of each test modality and the standard question-and-answer file, so as to obtain the correct number of question-and-answer results;
[0144] The second determination module is used to determine the feature contribution degree corresponding to each test modality data;
[0145] The third determination module is used to determine the question-answering accuracy of the proxy model based on the number of correct questions and answers and the feature contribution of each test modality data.
[0146] The update module is used to update the proxy model based on the question-answering accuracy to complete the proxy model update.
[0147] In one possible implementation, the second test module specifically includes:
[0148] The processing unit is used to input the data of each test modality into the proxy model for question-and-answer processing to obtain multiple question-and-answer results;
[0149] The comparison unit is used to compare each question and answer result with the standard question and answer document to obtain the comparison results;
[0150] The determination unit is used to determine the correct number of multiple question-and-answer results based on the comparison results.
[0151] In one possible implementation, the device further includes:
[0152] The second construction module is used to construct local feature dictionaries and global feature dictionaries for each modality based on the preset framework distributed algorithm and modal data.
[0153] The integration module is used to integrate the local feature dictionaries and the global feature dictionaries into a feature dictionary corresponding to each modality.
[0154] The aggregation module is used to aggregate various feature dictionaries into a composite feature dictionary using a preset aggregation algorithm.
[0155] The apparatus provided in this embodiment can be used to execute the technical solutions of the above method embodiments. Its implementation principle and technical effects are similar, and will not be described again here.
[0156] Figure 5 This is a schematic diagram of the hardware structure of the server provided in an embodiment of this application. Figure 5 As shown, the server in this embodiment includes a processor 501 and a memory 502; the memory stores computer execution instructions; at least one processor executes the computer execution instructions stored in the memory, causing at least one processor to execute the above-described proxy model training method.
[0157] Alternatively, the memory 502 can be either standalone or integrated with the processor 501.
[0158] When the memory 502 is configured independently, the server also includes a bus 503 for connecting the memory 502 and the processor 501.
[0159] This application also provides a computer storage medium storing computer execution instructions. When the processor executes the computer execution instructions, the above-mentioned proxy model training method is implemented.
[0160] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described agent model training method.
[0161] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or modules, and may be electrical, mechanical, or other forms.
[0162] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to implement the solution of this embodiment according to actual needs.
[0163] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing unit, or each module can exist physically separately, or two or more modules can be integrated into one unit. The unit composed of the above modules can be implemented in hardware or in the form of hardware plus software functional units.
[0164] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.
[0165] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.
[0166] The memory may include high-speed RAM, and may also include non-volatile storage (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk or optical disc, etc.
[0167] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.
[0168] The aforementioned storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage media can be any available medium accessible to general-purpose or special-purpose computers.
[0169] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. Both the processor and the storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components in an electronic device or host device.
[0170] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.
[0171] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.
Claims
1. A method for training a surrogate model, characterized in that, Applied to servers, including: When it is necessary to train a preset proxy model, modal data corresponding to multiple different categories of modalities are obtained from the modal data information dictionary; Each modal data is encoded to generate a feature dictionary for the corresponding modality, where each feature dictionary is used to characterize the feature representation of the corresponding modality; Determine the usage frequency of each feature atom in each feature dictionary; The feature contribution of the corresponding mode is quantified based on the usage frequency of each feature atom. Multiple target feature atoms are selected from the feature contribution of each mode; Construct a composite feature dictionary based on each target feature atom; The proxy model is trained based on the composite feature dictionary to complete the training of the proxy model.
2. The method according to claim 1, characterized in that, Determining the usage frequency of each feature atom in each feature dictionary includes: Identify the multiple feature atoms contained in each feature dictionary; Determine the number of times each feature atom is used in its corresponding feature dictionary; Based on the number of times each feature atom is used, determine the total number of times all feature atoms are used in the corresponding feature dictionary; The usage frequency of each feature atom is determined based on the number of times each atom is used and the total number of times it is used.
3. The method according to claim 1, characterized in that, The quantification of the feature contribution of the corresponding mode based on the usage frequency of each feature atom includes: Set the time decay factor and determine the usage time of each feature atom; The weights corresponding to the usage frequency of each feature atom are set according to the usage time and the time decay factor. The feature contribution of the corresponding mode is quantified based on the usage frequency of each weight and feature atom.
4. The method according to claim 1, characterized in that, The process of selecting multiple target feature atoms from the feature contribution values of each mode includes: Set the filtering rules for target feature atoms; The filtering rules are adjusted to complete the adjustment of the filtering rules; After the filtering rules are adjusted, multiple target feature atoms are selected from the feature contribution of each mode.
5. The method according to claim 1, characterized in that, After training the proxy model based on the composite feature dictionary to complete the training of the proxy model, the method further includes: After the proxy model is trained, it is tested, and the following steps are performed: Acquire multiple test modality data, each of which is configured with a standard question-and-answer file; Based on the test modality data and the standard question-and-answer file, the proxy model is subjected to question-and-answer testing to obtain the correct number of question-and-answer results; Determine the feature contribution of each test modality data; The question-answering accuracy of the proxy model is determined based on the number of correct answers and the feature contribution of each test modality data. The proxy model is updated based on the question-answering accuracy to complete the proxy model update.
6. The method according to claim 5, characterized in that, The step of testing the proxy model based on each test modality data and the standard question-and-answer file to obtain the correct number of test results includes: Each test modality data is input into the proxy model for question-and-answer processing to obtain multiple question-and-answer results; Each question and answer result is compared with the standard question and answer document to obtain the comparison results; The correct number of the multiple question-and-answer results is determined based on the comparison results.
7. The method according to any one of claims 1 to 6, characterized in that, Also includes: Based on the pre-defined framework distributed algorithm and the data of each modality, construct the local feature dictionary and global feature dictionary for each modality; The local feature dictionaries and the global feature dictionaries are integrated into a feature dictionary corresponding to each modality; By using a preset aggregation algorithm, the feature dictionaries are aggregated into a composite feature dictionary.
8. A proxy model training device, characterized in that, Applied to servers, including: The first acquisition module is used to acquire modal data corresponding to multiple different categories of modalities from the modal data information dictionary when it is necessary to train a preset agent model; The encoding module is used to encode the data of each modality to generate the feature dictionary of the corresponding modality, where each feature dictionary is used to characterize the feature representation of the corresponding modality; The first determining module is used to determine the usage frequency of each feature atom in each feature dictionary; The quantization module is used to quantify the feature contribution of the corresponding mode based on the usage frequency of each feature atom. The filtering module is used to filter multiple target feature atoms from the feature contribution of each mode; The first building module is used to construct a composite feature dictionary based on each target feature atom; The training module is used to train the proxy model based on the composite feature dictionary to complete the training of the proxy model.
9. A server, characterized in that, include: At least one processor and memory; The memory stores computer-executed instructions; The at least one processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the surrogate model training method as described in any one of claims 1 to 7.
10. A computer storage medium, characterized in that, The computer storage medium stores computer execution instructions, and when the processor executes the computer execution instructions, it implements the surrogate model training method as described in any one of claims 1 to 7.