Multi-modal large model training strategy determination method and system

Abstract

The invention discloses a multi-modal large model training strategy determination method and system, and particularly relates to the technical field of multi-modal data analysis. Comprising a training task complexity module, a task matching model determination module, a model training multi-modal data acquisition module, a model training performance monitoring module, a model training strategy early warning processing module, a model training multi-modal fusion effect analysis module and a model training multi-modal fusion effect evaluation module. According to the method, the optimal execution large model of each training task is calculated, the multi-modal data is collected, the performance stability index is obtained through calculation, and the modal fusion effect evaluation coefficient is obtained through analysis, so that a quantitative index is provided for evaluating the training task of the large model, the overall training efficiency is improved, and the task execution risk early warning index is calculated; real-time task training risk early warning measures are provided, the rationality of evaluation of the execution condition of the training tasks is improved, and automatic early warning management of large-model training task abnormities is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of multimodal data analysis, and more specifically, to a method and system for determining a multimodal large model training strategy. Background Art

[0002] Multimodal data, that is, data that combines multiple types of data such as text, images, and sounds, can be effectively integrated by processing and integrating different types of data, thereby improving the overall performance of the model and providing richer and more comprehensive information for machine learning models.

[0003] The current training strategy ensures the quality and diversity of input data through cleaning, normalization and enhancement in the data preprocessing stage, thereby providing the model with richer information. According to the characteristics and requirements of the task, a suitable network structure is designed to ensure that the model can effectively capture the key features in the data. Through these comprehensive optimization measures, the performance and generalization ability of the model are effectively improved, enabling it to adapt and process more flexibly when facing different scenarios and tasks.

[0004] However, it still has some shortcomings in actual use. For example, the fusion method between different modal data is often relatively simple and lacks deep integration, which makes the model unable to fully utilize the complementarity between multimodal data.

[0005] The existing multimodal large model training data processing is complex. When the large model training process of the task has abnormalities such as training time and training speed, it may lead to instability and performance degradation in the model training process. The existing risk identification of large model task training cannot identify potential training problems by comparing historical training data with current training data. Summary of the invention

[0006] In order to overcome the above-mentioned defects of the prior art, an embodiment of the present invention provides a method and system for determining a multimodal large model training strategy, which are used to solve the problems raised in the above-mentioned background technology.

[0007] To achieve the above object, the present invention provides the following technical solution: a multimodal large model training strategy determination system, comprising:

[0008] Training task complexity module: used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory usage of the training task.

[0009] Task matching model determination module: used to receive the multimodal complexity index transmitted by the training task complexity module, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, use the large model to execute the training task, and number the large model training tasks.

[0010] Model training multimodal data acquisition module: used to collect multimodal data for large model training tasks, and the multimodal data includes training process quality indicators and training efficiency indicators.

[0011] Model training performance monitoring module: used to calculate the performance stability index of large model training tasks based on the training process quality indicators and training efficiency indicators.

[0012] Model training strategy warning processing module: used to obtain the quality indicators and training efficiency indicators of each iterative training process of the large model training task, calculate the task execution risk warning index of the large model training task, and screen out abnormal training strategies for warning.

[0013] Model training multimodal fusion effect analysis module: used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task based on the performance stability index of the large model training task.

[0014] Model training multimodal fusion effect evaluation module: used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

[0015] Preferably, the calculation formula of the multimodal complexity index is:

[0016]

[0017] in, It is expressed as the multimodal complexity index of the jth training task of the i-th large model, It is represented as the memory usage of the jth training task of the i-th large model. Expressed as the modality fusion efficiency of the jth training task of the i-th large model, fs max It represents the maximum amount of modal data that can be processed by the large model, and n represents the total number of large models.

[0018] Preferably, the task matching model determination module is specifically:

[0019] Step S31: Obtain the large model multimodal complexity index of each training task, sort the large model multimodal complexity index of each training task from large to small, select the large model corresponding to the maximum value of the multimodal complexity index, and mark it as the execution large model of each training task;

[0020] Step S32: According to the execution large model of each training task, the large model training tasks are numbered as 1, 2, ...g, ...m in sequence.

[0021] Preferably, the training process quality indicators include: training loss value, training accuracy, CPU utilization, and the training efficiency indicators include training iteration time and training processing data volume.

[0022] Preferably, the model training performance monitoring module is specifically:

[0023] Step S51: The calculation formula of the training process quality index is:

[0024]

[0025] Among them, XK g It is represented as the training process quality indicator of the g-th large model training task, bq g It is expressed as the training accuracy of the g-th large model training task, bc g It is expressed as the CPU utilization of the g-th large model training task, bs g It is represented as the training loss value of the g-th large model training task, BS 预 Represented as the preset training loss value, λ 1 , 2 , 3 They are respectively represented as the compensation factors of training accuracy, CPU utilization, and training loss value, and λ 1 +λ 2 +λ 3 =1;

[0026] Step S52: The calculation formula of the training efficiency index is:

[0027]

[0028] Among them, XZ g It is expressed as the training efficiency index of the g-th large model training task, wn g Represents the amount of training data processed by the g-th large model training task, wn max It is expressed as the maximum amount of training processing data, wn min It is expressed as the minimum amount of training data, wt g Represents the training iteration time of the g-th large model training task, wt max Expressed as the maximum training iteration time, wt min Represented as the minimum value of training iteration time;

[0029] Step S53: The calculation formula of the performance stability index is:

[0030]

[0031] Among them, β g It is expressed as the performance stability index of the g-th large model training task, XK g It is expressed as the training process quality indicator of the g-th large model training task, XZ g It is expressed as the training efficiency index of the g-th large model training task.

[0032] Preferably, the model training strategy warning processing module is specifically:

[0033] Step S61: Calculate the task execution risk warning index by using the training process quality index of the first iteration and the training process quality index of the s-th iteration of the large model training task, the training efficiency index of the first iteration and the training efficiency index of the s-th iteration:

[0034]

[0035] Among them, g It is represented as the task execution risk warning index of the g-th large model training task, XK g s It is expressed as the training process quality index of the sth iteration of the gth large model training task, XZ g s It is represented as the training efficiency index of the sth iteration of the gth large model training task, d is represented as the total number of iterations, σ g It is expressed as the standard deviation of the training process quality index of the g-th large model training task, ε g It is expressed as the standard deviation of the training efficiency index of the g-th large model training task, XK g 1 It is represented as the training process quality indicator of the first iteration of the g-th large model training task, XZ g 1 It is expressed as the training efficiency index of the first iteration of the g-th large model training task;

[0036] Step S62: Obtain the task execution risk warning index of the large model training task, and compare it with the preset task execution risk warning index. If the task execution risk warning index of a large model training task is less than the preset task execution risk warning index, it indicates that there is an abnormality in the training strategy of the large model, and the abnormal large model number should be sent to the management personnel immediately. Otherwise, it indicates that there is no abnormality in the training strategy of the large model.

[0037] Preferably, the calculation formula of the modal fusion effect evaluation coefficient is:

[0038]

[0039] Among them, θ g It is expressed as the modality fusion effect evaluation coefficient of the g-th large model training task, β g It is expressed as the performance stability index of the g-th large model training task, β 预 It is expressed as a preset performance stability index, and e is expressed as a natural constant.

[0040] Preferably, the model training multimodal fusion effect evaluation module is specifically:

[0041] Obtain the modal fusion effect evaluation coefficient of the large model training task, and compare it with the preset modal fusion effect evaluation coefficient. If the modal fusion effect evaluation coefficient of a large model training task is less than the preset modal fusion effect evaluation coefficient, it means that the large model fusion performance effect does not meet expectations, and relevant personnel should be notified to optimize the model performance. Otherwise, it means that the large model fusion performance effect meets expectations.

[0042] A method for determining a multimodal large model training strategy comprises the following steps:

[0043] Step S01: Training task complexity: used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory usage of the training task;

[0044] Step S02: Task matching model determination: used to receive the multimodal complexity index transmitted by the training task complexity step, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, execute the training task by the large model, and number the large model training tasks;

[0045] Step S03: Model training multimodal data acquisition: used to acquire multimodal data of large model training tasks, wherein the multimodal data includes training process quality indicators and training efficiency indicators;

[0046] Step S04: Model training performance monitoring: used to calculate the performance stability index of the large model training task based on the training process quality index and the training efficiency index;

[0047] Step S05: Model training strategy early warning processing: used to obtain the quality index and training efficiency index of each iterative training process of the large model training task, calculate the task execution risk early warning index of the large model training task, and screen out abnormal training strategies for early warning;

[0048] Step S06: Model training multimodal fusion effect analysis: used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task according to the performance stability index of the large model training task;

[0049] Step S07: Model training multimodal fusion effect evaluation: used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

[0050] Technical effects and advantages of the present invention:

[0051] 1. The present invention provides a method and system for determining a multimodal large model training strategy. The multimodal complexity index of the large model of each training task is calculated by the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory occupancy of the training task, so as to obtain the optimal execution large model of each training task. The multimodal data of the large model training task is collected, and the performance stability index of the large model training task is calculated according to the training process quality index and the training efficiency index. The modal fusion effect evaluation coefficient of the large model training task is further analyzed and compared with the preset modal fusion effect evaluation coefficient. If the modal fusion effect evaluation coefficient of a large model training task is less than the preset modal fusion effect evaluation coefficient, it indicates that the large model fusion performance effect does not meet expectations, and the relevant personnel should be notified to optimize the model performance. Otherwise, it indicates that the large model fusion performance effect meets expectations. The optimal execution of the large model training data provides quantitative indicators for evaluating the performance stability of the large model training task and the training data fusion effect, which is helpful to comprehensively evaluate the performance status of the large model training task, and help the relevant personnel understand the performance status of the large model training task, and make timely adjustments and optimizations, thereby improving the overall training efficiency.

[0052] 2. The present invention provides a method and system for determining a multimodal large model training strategy. The training process quality index is calculated by the training loss value, training accuracy, and CPU utilization rate. The training efficiency index is calculated by the training iteration time and the amount of training processed data. The task execution risk warning index of the large model training task is calculated according to the training process quality index and the training efficiency index of each iteration number, and compared with the preset task execution risk warning index. If the task execution risk warning index of a large model training task is less than the preset task execution risk warning index, it indicates that there is an abnormality in the training strategy of the large model, and the abnormal large model number should be immediately sent to the management personnel. Otherwise, it indicates that there is no abnormality in the training strategy of the large model. The key indicators in the large model training process can be monitored in real time, thereby providing real-time task training risk warning measures, improving the rationality of the evaluation of the execution of the training task, and realizing the automatic warning management of the abnormal large model training task. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1 It is a schematic diagram of the system module process connection of the present invention.

[0054] Figure 2It is a schematic diagram of the connection of the method steps of the present invention. DETAILED DESCRIPTION

[0055] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0056] See also Figure 1 As shown, the present invention provides a multimodal large model training strategy determination system, including a training task complexity module, a task matching model determination module, a model training multimodal data acquisition module, a model training performance monitoring module, a model training strategy early warning processing module, a model training multimodal fusion effect analysis module, and a model training multimodal fusion effect evaluation module.

[0057] The training task complexity module is connected to the task matching model determination module, the task matching model determination module is connected to the model training multimodal data acquisition module, the model training multimodal data acquisition module is connected to the model training performance monitoring module, the model training performance monitoring module is connected to the model training strategy early warning processing module and the model training multimodal fusion effect analysis module, and the model training multimodal fusion effect analysis module is connected to the model training multimodal fusion effect evaluation module.

[0058] The training task complexity module is used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory occupancy of the training task.

[0059] In a possible design, the calculation formula of the multimodal complexity index is:

[0060]

[0061] in, It is expressed as the multimodal complexity index of the jth training task of the i-th large model, It is represented as the memory usage of the jth training task of the i-th large model. Expressed as the modality fusion efficiency of the jth training task of the i-th large model, fs max It represents the maximum amount of modal data that can be processed by the large model, and n represents the total number of large models.

[0062] The task matching model determination module is used to receive the multimodal complexity index transmitted by the training task complexity module, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, use the large model to execute the training task, and number the large model training tasks.

[0063] In a possible design, the task matching model determination module is specifically:

[0064] Step S31: Obtain the large model multimodal complexity index of each training task, sort the large model multimodal complexity index of each training task from large to small, select the large model corresponding to the maximum value of the multimodal complexity index, and mark it as the execution large model of each training task;

[0065] Step S32: According to the execution large model of each training task, the large model training tasks are numbered as 1, 2, ...g, ...m in sequence.

[0066] The model training multimodal data acquisition module is used to collect multimodal data of large model training tasks, and the multimodal data includes training process quality indicators and training efficiency indicators.

[0067] In a possible design, the training process quality indicators include: training loss value, training accuracy, and CPU utilization; the training efficiency indicators include training iteration time and training processing data volume.

[0068] The model training performance monitoring module is used to receive the multimodal data transmitted by the model training multimodal data acquisition module, and calculate the performance stability index of the large model training task.

[0069] In a possible design, the model training performance monitoring module is specifically:

[0070] Step S51: The calculation formula of the training process quality index is:

[0071]

[0072] Among them, XK g It is represented as the training process quality indicator of the g-th large model training task, bq g It is expressed as the training accuracy of the g-th large model training task, bc g It is expressed as the CPU utilization of the g-th large model training task, bs g It is represented as the training loss value of the g-th large model training task, BS 预 Represented as the preset training loss value, λ 1 , 2 , 3They are respectively represented as the compensation factors of training accuracy, CPU utilization, and training loss value, and λ 1 +λ 2 +λ 3 =1;

[0073] Step S52: The calculation formula of the training efficiency index is:

[0074]

[0075] Among them, XZ g It is expressed as the training efficiency index of the g-th large model training task, wn g Represents the amount of training data processed by the g-th large model training task, wn max It is expressed as the maximum amount of training processing data, wn min It is expressed as the minimum amount of training data, wt g Represents the training iteration time of the g-th large model training task, wt max Expressed as the maximum training iteration time, wt min Represented as the minimum value of training iteration time;

[0076] Step S53: The calculation formula of the performance stability index is:

[0077]

[0078] Among them, β g It is expressed as the performance stability index of the g-th large model training task, XK g It is expressed as the training process quality indicator of the g-th large model training task, XZ g It is expressed as the training efficiency index of the g-th large model training task.

[0079] The model training strategy warning processing module is used to obtain the quality indicators and training efficiency indicators of each iterative training process of the large model training task, calculate the task execution risk warning index of the large model training task, and screen out abnormal training strategies for warning.

[0080] In a possible design, the model training strategy warning processing module is specifically:

[0081] Step S61: Calculate the task execution risk warning index by using the training process quality index of the first iteration and the training process quality index of the s-th iteration of the large model training task, the training efficiency index of the first iteration and the training efficiency index of the s-th iteration:

[0082]

[0083] Among them, gIt is represented as the task execution risk warning index of the g-th large model training task, XK g s It is expressed as the training process quality index of the sth iteration of the gth large model training task, XZ g s It is represented as the training efficiency index of the sth iteration of the gth large model training task, d is represented as the total number of iterations, and σ g It is expressed as the standard deviation of the training process quality index of the g-th large model training task, ε g It is expressed as the standard deviation of the training efficiency index of the g-th large model training task, XK g 1 It is represented as the training process quality indicator of the first iteration of the g-th large model training task, XZ g 1 It is expressed as the training efficiency index of the first iteration of the g-th large model training task;

[0084] Step S62: Obtain the task execution risk warning index of the large model training task, and compare it with the preset task execution risk warning index. If the task execution risk warning index of a large model training task is less than the preset task execution risk warning index, it indicates that there is an abnormality in the training strategy of the large model, and the abnormal large model number should be sent to the management personnel immediately. Otherwise, it indicates that there is no abnormality in the training strategy of the large model.

[0085] In this embodiment, it should be specifically stated that the calculation formula of the standard deviation of the training process quality indicator is:

[0086] Among them, ΔXK g It is represented as the mean of the training process quality indicators of the g-th large model training task,

[0087] The calculation formula of the standard deviation of the training efficiency index is:

[0088] Where ΔXZ g It is expressed as the mean of the training efficiency index of the g-th large model training task,

[0089] The model training multimodal fusion effect analysis module is used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task according to the performance stability index of the large model training task.

[0090] In a possible design, the calculation formula of the modal fusion effect evaluation coefficient is:

[0091]

[0092] Among them, θ g It is expressed as the modality fusion effect evaluation coefficient of the g-th large model training task, β g It is expressed as the performance stability index of the g-th large model training task, β 预 It is expressed as a preset performance stability index, and e is expressed as a natural constant.

[0093] The model training multimodal fusion effect evaluation module is used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

[0094] In a possible design, the model training multimodal fusion effect evaluation module is specifically:

[0095] Obtain the modal fusion effect evaluation coefficient of the large model training task, and compare it with the preset modal fusion effect evaluation coefficient. If the modal fusion effect evaluation coefficient of a large model training task is less than the preset modal fusion effect evaluation coefficient, it means that the large model fusion performance effect does not meet expectations, and relevant personnel should be notified to optimize the model performance. Otherwise, it means that the large model fusion performance effect meets expectations.

[0096] See also Figure 2 As shown, in this embodiment, it should be specifically explained that the present invention provides a method for determining a multimodal large model training strategy, comprising the following steps:

[0097] Step S01: Training task complexity: used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory usage of the training task;

[0098] Step S02: Task matching model determination: used to receive the multimodal complexity index transmitted by the training task complexity step, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, execute the training task by the large model, and number the large model training tasks;

[0099] Step S03: Model training multimodal data acquisition: used to acquire multimodal data of large model training tasks, wherein the multimodal data includes training process quality indicators and training efficiency indicators;

[0100] Step S04: Model training performance monitoring: used to calculate the performance stability index of the large model training task based on the training process quality index and the training efficiency index;

[0101] Step S05: Model training strategy early warning processing: used to obtain the quality index and training efficiency index of each iterative training process of the large model training task, calculate the task execution risk early warning index of the large model training task, and screen out abnormal training strategies for early warning;

[0102] Step S06: Model training multimodal fusion effect analysis: used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task according to the performance stability index of the large model training task;

[0103] Step S07: Model training multimodal fusion effect evaluation: used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

[0104] In this embodiment, it should be specifically explained that the present invention calculates the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of the modal data, and the memory occupancy of the training task, thereby obtaining the optimal execution large model of each training task, and by collecting the multimodal data of the large model training task, the performance stability index of the large model training task is calculated according to the training process quality index and the training efficiency index, and the modal fusion effect evaluation coefficient of the large model training task is further analyzed to obtain the modal fusion effect evaluation coefficient of the large model training task, which is compared with the preset modal fusion effect evaluation coefficient. If the modal fusion effect evaluation coefficient of a large model training task is less than the preset modal fusion effect evaluation coefficient, it indicates that the large model fusion performance effect does not meet expectations, and the relevant personnel should be notified to optimize the model performance. Otherwise, it indicates that the large model fusion performance effect meets expectations. The optimal training data for executing the large model is used to provide quantitative indicators for evaluating the performance stability of the large model training task and the training data fusion effect, which is helpful to comprehensively evaluate the performance status of the large model training task, and help the relevant personnel understand the performance status of the large model training task, and make timely adjustments and optimizations, thereby improving the overall training efficiency.

[0105] The present invention calculates the training process quality index through the training loss value, training accuracy and CPU utilization rate, calculates the training efficiency index through the training iteration time and the training processing data volume, and obtains the task execution risk warning index of the large model training task according to the training process quality index and the training efficiency index of each iteration number, and compares it with the preset task execution risk warning index. If the task execution risk warning index of a large model training task is less than the preset task execution risk warning index, it indicates that the training strategy of the large model is abnormal, and the abnormal large model number should be immediately sent to the management personnel. Otherwise, it indicates that there is no abnormality in the training strategy of the large model. The key indicators in the large model training process can be monitored in real time, thereby providing real-time task training risk warning measures, improving the rationality of the evaluation of the execution of the training task, and realizing the automatic warning management of the abnormality of the large model training task.

[0106] Finally: The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A multimodal large model training strategy determination system, characterized in that: include: Training task complexity module: used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory usage of the training task; Task matching model determination module: used to receive the multimodal complexity index transmitted by the training task complexity module, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, execute the training task by the large model, and number the large model training tasks; Model training multimodal data acquisition module: used to collect multimodal data of large model training tasks, the multimodal data includes training process quality indicators and training efficiency indicators; Model training performance monitoring module: used to calculate the performance stability index of large model training tasks based on the training process quality index and training efficiency index; Model training strategy warning processing module: used to obtain the quality indicators and training efficiency indicators of each iterative training process of the large model training task, calculate the task execution risk warning index of the large model training task, and screen out abnormal training strategies for warning; Model training multimodal fusion effect analysis module: used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task based on the performance stability index of the large model training task; Model training multimodal fusion effect evaluation module: used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

2. A multimodal large model training strategy determination system according to claim 1, characterized in that: The calculation formula of the multimodal complexity index is: in, It is expressed as the multimodal complexity index of the jth training task of the i-th large model, It is represented as the memory usage of the jth training task of the i-th large model. Expressed as the modality fusion efficiency of the jth training task of the i-th large model, fs max It represents the maximum amount of modal data that can be processed by the large model, and n represents the total number of large models.

3. A multimodal large model training strategy determination system according to claim 1, characterized in that: The task matching model determination module is specifically: Step S31: Obtain the large model multimodal complexity index of each training task, sort the large model multimodal complexity index of each training task from large to small, select the large model corresponding to the maximum value of the multimodal complexity index, and mark it as the execution large model of each training task; Step S32: According to the execution large model of each training task, the large model training tasks are numbered as 1, 2, ...g, ...m in sequence.

4. The multimodal large model training strategy determination system according to claim 1, characterized in that: The training process quality indicators include: training loss value, training accuracy, and CPU utilization; the training efficiency indicators include training iteration time and training processing data volume.

5. The multimodal large model training strategy determination system according to claim 1, characterized in that: The model training performance monitoring module is specifically: Step S51: The calculation formula of the training process quality index is: Among them, XK g It is represented as the training process quality indicator of the g-th large model training task, bq g It is expressed as the training accuracy of the g-th large model training task, bc g It is expressed as the CPU utilization of the g-th large model training task, bs g It is represented as the training loss value of the g-th large model training task, BS 预 represents the preset training loss value, λ1, λ2, and λ3 represent the compensation factors of training accuracy, CPU utilization, and training loss value, respectively, and λ1+λ2+λ3=1; Step S52: The calculation formula of the training efficiency index is: Among them, XZ g It is expressed as the training efficiency index of the g-th large model training task, wn g Represents the amount of training data processed by the g-th large model training task, wn max It is expressed as the maximum amount of training processing data, wn min It is expressed as the minimum amount of training data, wt g Represented as the training iteration time of the g-th large model training task, wt max Expressed as the maximum training iteration time, wt min Represented as the minimum value of training iteration time; Step S53: The calculation formula of the performance stability index is: Among them, β g It is expressed as the performance stability index of the g-th large model training task, XK g It is expressed as the training process quality indicator of the g-th large model training task, XZ g It is expressed as the training efficiency index of the g-th large model training task.

6. A multimodal large model training strategy determination system according to claim 1, characterized in that: The model training strategy early warning processing module is specifically: Step S61: Calculate the task execution risk warning index by using the training process quality index of the first iteration and the training process quality index of the s-th iteration of the large model training task, the training efficiency index of the first iteration and the training efficiency index of the s-th iteration: Among them, g It is represented as the task execution risk warning index of the g-th large model training task, XK g s It is expressed as the training process quality index of the sth iteration of the gth large model training task, XZ g s It is represented as the training efficiency index of the sth iteration of the gth large model training task, d is represented as the total number of iterations, σ g It is expressed as the standard deviation of the training process quality index of the g-th large model training task, ε g It is expressed as the standard deviation of the training efficiency index of the g-th large model training task, XK g 1 It is represented as the training process quality index of the first iteration of the g-th large model training task, XZ g 1 It is expressed as the training efficiency index of the first iteration of the g-th large model training task; Step S62: Obtain the task execution risk warning index of the large model training task, and compare it with the preset task execution risk warning index. If the task execution risk warning index of a large model training task is less than the preset task execution risk warning index, it indicates that there is an abnormality in the training strategy of the large model, and the abnormal large model number should be sent to the management personnel immediately. Otherwise, it indicates that there is no abnormality in the training strategy of the large model.

7. The multimodal large model training strategy determination system according to claim 1, characterized in that: The calculation formula of the modal fusion effect evaluation coefficient is: Among them, θ g It is expressed as the modality fusion effect evaluation coefficient of the g-th large model training task, β g It is expressed as the performance stability index of the g-th large model training task, β 预 It is expressed as a preset performance stability index, and e is expressed as a natural constant.

8. The multimodal large model training strategy determination system according to claim 1, characterized in that: The model training multimodal fusion effect evaluation module is specifically as follows: Obtain the modal fusion effect evaluation coefficient of the large model training task, and compare it with the preset modal fusion effect evaluation coefficient. If the modal fusion effect evaluation coefficient of a large model training task is less than the preset modal fusion effect evaluation coefficient, it means that the large model fusion performance effect does not meet expectations, and relevant personnel should be notified to optimize the model performance. Otherwise, it means that the large model fusion performance effect meets expectations.

9. A method for determining a multimodal large model training strategy, using a multimodal large model training strategy determination system as described in any one of claims 1 to 8, characterized in that: The following steps are involved: Step S01: Training task complexity: used to calculate the large model multimodal complexity index of each training task through the modal fusion efficiency of the large model, the maximum processing volume of modal data, and the memory usage of the training task; Step S02: Task matching model determination: used to receive the multimodal complexity index transmitted by the training task complexity step, screen out the large model corresponding to the maximum value of the multimodal complexity index of each task training, execute the training task by the large model, and number the large model training tasks; Step S03: Model training multimodal data acquisition: used to acquire multimodal data of large model training tasks, wherein the multimodal data includes training process quality indicators and training efficiency indicators; Step S04: Model training performance monitoring: used to calculate the performance stability index of the large model training task based on the training process quality index and the training efficiency index; Step S05: Model training strategy early warning processing: used to obtain the quality index and training efficiency index of each iterative training process of the large model training task, calculate the task execution risk early warning index of the large model training task, and screen out abnormal training strategies for early warning; Step S06: Model training multimodal fusion effect analysis: used to analyze and obtain the modal fusion effect evaluation coefficient of the large model training task according to the performance stability index of the large model training task; Step S07: Model training multimodal fusion effect evaluation: used to obtain the modal fusion effect evaluation coefficient of the large model training task, compare it with the preset modal fusion effect evaluation coefficient, and process it.

Images