Incremental training and optimizing system and method for large-scale machine learning model
Through the incremental training and optimization system of large-scale machine learning models, the model dependency graph and optimization prompt library are used to generate training optimization paths and adjustment parameters, which solves the problem of inefficient training of traditional machine learning models and realizes efficient and reliable model training and optimization.
Patent Information
- Application Number
- CN202510430709.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-08
- Publication Date
- 2025-07-25
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Traditional machine learning model training methods are inefficient, difficult to adapt to scenarios with dynamic data changes and high real-time requirements, fail to make full use of the correlation between features, and lack effective optimization tips and auxiliary tools, resulting in poor model training and optimization results.
It provides a large-scale machine learning model incremental training and optimization system, including training monitoring interface, data processing module and model optimization auxiliary module, analyzes the correlation between features through model dependency graphs, combines the optimization prompt library to generate training optimization paths and adjust parameters, and uses a distributed computing environment to accelerate the training process.
It improves model training efficiency and optimization effect, enhances the adaptability and reliability of the model, simplifies the operation process, and improves the generalization ability and overall training quality of the model.
Smart Images
Figure CN120373411A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of artificial intelligence and machine learning. More specifically, the present invention relates to a system and method for incremental training and optimization of large-scale machine learning models. Background Art
[0002] In today's digital age, large-scale machine learning models play a crucial role in many fields. For example, in image recognition, natural language processing, intelligent recommendation systems, etc., their powerful data processing capabilities and accurate prediction capabilities have brought great convenience to people's lives and work. However, with the continuous growth of data volume and the increasing complexity of application scenarios, the traditional machine learning model training methods face many challenges. Traditional machine learning model training is usually based on static datasets. Once the model is trained and put into use, if it needs to be updated or optimized, it often requires retraining from scratch, which not only consumes a large amount of time and computing resources, but also is difficult to meet the requirements of rapid iteration in application scenarios with high real-time requirements. In addition, the structure of large-scale machine learning models is complex, the correlation relationships between various features are intricate, and the calculation logic is also relatively complex. This makes it difficult to efficiently analyze and utilize the relationships between these features during model training and optimization, further increasing the difficulty of model training and optimization.
[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: First, the model update efficiency is low, and it is difficult to adapt to scenarios with dynamic data changes and high real-time requirements; second, the correlation relationships between features are not fully utilized, resulting in less than ideal model training and optimization effects; third, there is a lack of effective optimization prompts and auxiliary tools, making the model optimization process lack systematicness and pertinence, and it is difficult to quickly locate and solve problems in the optimization process. Summary of the Invention
[0004] The present invention provides a system and method for incremental training and optimization of large-scale machine learning models.
[0005] In the first aspect of the present invention, a system for incremental training and optimization of large-scale machine learning models is provided, including:
[0006] A training monitoring interface for allowing a user to input training instructions and display training results;
[0007] A data processing module, which includes a data cache, a model dependency graph, a data feature library, an incremental training engine, a large-scale machine learning model, an optimization hint library, and a distributed computing environment. The model dependency graph is constructed based on each feature in the data cache and the data feature library, and is used to analyze the association relationship between each feature in the data feature library. The data feature library stores feature data, and the feature data is a training path with practical significance sampled on the model dependency graph. The optimization hint library stores optimization hints written for the large-scale machine learning model based on the feature data.
[0008] A model optimization assistance module, which includes a feature analysis sub-module, a model adjustment sub-module, an incremental training sub-module, and a performance evaluation sub-module. The feature analysis sub-module is used to use the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate a training optimization path for the training instruction with reference to the feature data in the data feature library based on the training instruction input through the training monitoring interface. The model adjustment sub-module is used to guide the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path. The incremental training sub-module is used to use the distributed computing environment to execute this set of adjustment parameters to update the corresponding data in the data cache. The performance evaluation sub-module is used to use the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instruction according to the training instruction and the corresponding data updated by the incremental training sub-module from the data cache.
[0009] Among them, the model dependency graph refers to a model structure with intricate interrelationships and complex calculation logics.
[0010] Furthermore, the training monitoring interface includes: a training instruction input sub-module, which is used for a user to input a training instruction.
[0011] An optimization path display sub-module, which is used to display the training optimization path obtained by the feature analysis sub-module.
[0012] A parameter display sub-module, which is used to display a set of adjustment parameters obtained by the model adjustment sub-module.
[0013] A training result display sub-module, which is used to display the result of the incremental training sub-module executing this set of adjustment parameters.
[0014] A performance evaluation display sub-module, which is used to display a set of performance evaluation results obtained by the performance evaluation sub-module.
[0015] Furthermore, the optimization hint library includes: a feature optimization hint sub-library, which stores feature optimization hints written for the large-scale machine learning model based on the feature data.
[0016] A parameter adjustment hint sub-library, which stores parameter adjustment hints written for the large-scale machine learning model based on the feature data.
[0017] A performance evaluation hint sub-library, which stores performance evaluation hints written for the large-scale machine learning model based on the feature data.
[0018] Further, when the model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path, it first determines whether the training optimization path is in the data feature library. If it is in the data feature library, the parameter adjustment hints dedicated to the training optimization path in the parameter adjustment hint sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters; if it is not in the data feature library, the general parameter adjustment hints in the parameter adjustment hint sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters.
[0019] Further, each node in the model dependency graph represents an independent model module respectively, and the unidirectional edges between the nodes represent the dependency relationships of a model module, where the input of the tail node depends on the output of the head node.
[0020] Further, the distributed computing environment is a GPU distributed computing environment.
[0021] In the second aspect of the present invention, a method for incremental training and optimization of a large-scale machine learning model is provided, including:
[0022] The user inputs a training instruction through the training monitoring interface.
[0023] The feature analysis sub-module uses the optimization hints in the optimization hint library based on the training instruction to guide the large-scale machine learning model to generate a training optimization path for the training instruction with reference to the feature data in the data feature library.
[0024] The model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path.
[0025] The incremental training sub-module uses the distributed computing environment to execute this set of adjustment parameters to update the corresponding data in the data cache.
[0026] The performance evaluation sub-module uses the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instruction according to the training instruction and the corresponding data updated by the incremental training sub-module from the data cache.
[0027] The training results are presented through the training monitoring interface.
[0028] Furthermore, the training results presented through the training monitoring interface include the training optimization path, adjusted parameters, training execution results, and performance evaluation results.
[0029] In the third aspect of the present invention, there is provided a device for incremental training and optimization of a large-scale machine learning model, the device including: at least one memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the method for incremental training and optimization of a large-scale machine learning model described in the first aspect.
[0030] In the fourth aspect of the present invention, there is provided a computer-readable storage medium having stored thereon a computer program, which when executed by a processor implements the steps in the method for incremental training and optimization of a large-scale machine learning model described in the first aspect.
[0031] According to the above embodiments of the present invention, there are at least the following beneficial effects: The system can improve the training efficiency and optimization effect of a large-scale machine learning model. By parsing the complex associations between features through the model dependency graph and combining with the targeted guidance of the optimization hint library, the optimal training path and adjusted parameters can be automatically generated, reducing manual intervention and computational resource consumption. The use of a distributed computing environment can accelerate the incremental training process, and the cooperation between the data feature library and the incremental training engine can ensure data consistency and effectiveness when the model is updated, thereby enhancing the stability and reliability of model iteration.
[0032] The system can also enhance the transparency and controllability of the model optimization process. The training monitoring interface can display the optimization path, parameter adjustment, and performance evaluation results in real time, facilitating users to track the training progress. The linkage of the feature analysis, model adjustment, and performance evaluation modules can dynamically adapt to different training requirements and improve the generalization ability of the model. In addition, the classified management of the optimization hint library can provide precise guidance for different scenarios, making the model optimization more targeted, and ultimately improving the overall training quality and application effect. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, wherein:
[0034] Figure 1 FIG. is a schematic structural diagram of a system for incremental training and optimization of a large-scale machine learning model provided by an embodiment of the present invention. DETAILED DESCRIPTION
[0035] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and then implement the present invention, rather than limiting the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to be able to fully convey the scope of the present invention to those skilled in the art.
[0036] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, device, equipment, method, or computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: completely hardware, completely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.
[0037] It should be noted that any number of elements in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.
[0038] The following reference Figure 1 , Figure 1 is a schematic structural diagram of an incremental training and optimization system for a large-scale machine learning model provided in an embodiment of the present invention. As Figure 1 shown, a large-scale machine learning model incremental training and optimization system includes:
[0039] A training monitoring interface 101, which is used for a user to input training instructions and display training results;
[0040] A data processing module 102, which includes a data cache, a model dependency graph, a data feature library, an incremental training engine, a large-scale machine learning model, an optimization hint library, and a distributed computing environment. The model dependency graph is constructed according to each feature in the data cache and the data feature library, and is used to analyze the association relationship between each feature in the data feature library; the data feature library stores feature data, and the feature data is a training path with practical significance sampled on the model dependency graph; the optimization hint library stores optimization hints written for the large-scale machine learning model based on the feature data.
[0041] The model optimization assistance module 103 includes a feature analysis sub-module, a model adjustment sub-module, an incremental training sub-module, and a performance evaluation sub-module. The feature analysis sub-module is used to generate a training optimization path for the training instruction by guiding the large-scale machine learning model to refer to the feature data in the data feature library based on the training instruction input through the training monitoring interface by using the optimization tips in the optimization tip library; the model adjustment sub-module is used to guide the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path; the incremental training sub-module is used to execute this set of adjustment parameters by using the distributed computing environment to update the corresponding data in the data cache; the performance evaluation sub-module is used to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instruction according to the training instruction and the corresponding data updated by the incremental training sub-module from the data cache by using the optimization tips in the optimization tip library.
[0042] Among them, the model dependency graph refers to a model structure with intricate interrelationships and complex computational logics.
[0043] It should be noted that the present invention proposes a large-scale machine learning model incremental training and optimization system, the core of which is to achieve efficient incremental training and optimization of the large-scale machine learning model through the collaborative work of the training monitoring interface, the data processing module, and the model optimization assistance module. The training monitoring interface is the interface for users to interact with the system. Users can input training instructions through this interface, and the system displays training results through this interface. The data processing module is the core part of the system, including multiple components such as a data cache, a model dependency graph, a data feature library, an incremental training engine, a large-scale machine learning model, an optimization tip library, and a distributed computing environment. The model dependency graph is constructed based on the features in the data cache and the data feature library and is used to analyze the association relationships between the features in the data feature library. The feature data stored in the data feature library is a training path with practical significance sampled on the model dependency graph, while the optimization tip library stores optimization tips written for the large-scale machine learning model based on the feature data. The model optimization assistance module includes a feature analysis sub-module, a model adjustment sub-module, an incremental training sub-module, and a performance evaluation sub-module, and these sub-modules are respectively responsible for generating training optimization paths, adjusting model parameters, performing incremental training, and evaluating model performance.
[0044] Specifically, the training monitoring interface is the interface for users to interact with the system. Users can input training instructions through the training instruction input sub-module, and the system displays the training optimization path, adjusted parameters, training execution results, and performance evaluation results through the optimization path display sub-module, parameter display sub-module, training result display sub-module, and performance evaluation display sub-module. The data cache in the data processing module is used to store the temporary data generated during the training process. The model dependency graph is a directed graph, where each node in the graph represents an independent model module, and the unidirectional edges between the nodes represent the dependency relationships between the model modules. The input of the tail node depends on the output of the head node. The feature data stored in the data feature library is sampled from the model dependency graph, and these feature data have practical significance and can be used as training paths. The optimization hint library includes a feature optimization hint sub-library, a parameter adjustment hint sub-library, and a performance evaluation hint sub-library, which store feature optimization hints, parameter adjustment hints, and performance evaluation hints respectively. The feature analysis sub-module of the model optimization assistance module uses the optimization hints in the optimization hint library based on the training instructions to guide the large-scale machine learning model to generate a training optimization path with reference to the feature data in the data feature library. The model adjustment sub-module guides the large-scale machine learning model to generate adjusted parameters according to the training optimization path. The incremental training sub-module uses the distributed computing environment to execute the adjusted parameters to update the corresponding data in the data cache. The performance evaluation sub-module uses the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate performance evaluation results.
[0045] Preferably, the construction process of the large-scale machine learning model can be further refined. The construction of the model dependency graph is based on the feature data in the data cache and the data feature library. By analyzing the correlation relationships between the features, a directed graph is constructed, where the nodes in the graph represent independent model modules, and the edges represent the dependency relationships between the modules. When the feature analysis sub-module generates a training optimization path, it can generate the optimal training path through algorithm analysis by combining the training instructions, the hints in the optimization hint library, and the feature data in the data feature library. When the model adjustment sub-module generates adjusted parameters, it can generate a set of adjusted parameters suitable for the current training task according to the training optimization path and the hints in the parameter adjustment hint sub-library. When the incremental training sub-module performs incremental training, it can use the GPU resources in the distributed computing environment to accelerate the training process through parallel computing and improve the training efficiency. When the performance evaluation sub-module generates performance evaluation results, it can evaluate the model performance through a series of evaluation metrics according to the training instructions, the hints in the optimization hint library, and the data updated by the incremental training sub-module, and provide a detailed performance evaluation report for the user.
[0046] In some embodiments, the training monitoring interface includes: a training instruction input sub-module for the user to input training instructions;
[0047] An optimized path display sub-module, which is used to display the training optimization path obtained by the feature analysis sub-module;
[0048] A parameter display sub-module, which is used to display a set of adjustment parameters obtained by the model adjustment sub-module;
[0049] A training result display sub-module, which is used to display the result of the incremental training sub-module executing this set of adjustment parameters;
[0050] A performance evaluation display sub-module, which is used to display a set of performance evaluation results obtained by the performance evaluation sub-module.
[0051] It should be noted that the training monitoring interface is a key part of the interaction between the user and the system in the incremental training and optimization system of large-scale machine learning models. It not only allows users to input training instructions, but also can display information such as the optimization path, adjustment parameters, execution results, and performance evaluation results during the training process. The training instruction input sub-module is the starting point of the interaction between the user and the system, and users can input specific training task instructions through it. The optimized path display sub-module is used to present the training optimization path generated by the system according to the training instructions, helping users understand the optimization direction of the training process. The parameter display sub-module displays the adjustment parameters generated by the model adjustment sub-module, enabling users to intuitively see the changes in model parameters. The training result display sub-module is used to display the training results after the incremental training sub-module executes the adjustment parameters, while the performance evaluation display sub-module displays the performance evaluation results generated by the performance evaluation sub-module, providing feedback on the model performance to users.
[0052] Specifically, the training monitoring interface consists of multiple sub-modules, and each sub-module has a clear function. The training instruction input sub-module is a user interface component through which users can input specific instructions for training tasks, such as specifying parameters such as the training dataset, training objective, and optimization algorithm. The optimized path display sub-module is used to display the training optimization path generated by the feature analysis sub-module. This path is generated based on the prompts in the optimization prompt library and the feature data in the data feature library, indicating the directions that need to be focused on and optimized during the model training process. The adjustment parameters displayed by the parameter display sub-module are generated by the model adjustment sub-module according to the training optimization path. These parameters may include the learning rate, regularization coefficient, weights of the neural network, etc., and are used to guide the adjustment and optimization of the model. The training result display sub-module displays the results after the incremental training sub-module executes the adjustment parameters, and these results may include indicators such as the accuracy of the model and the value of the loss function. The performance evaluation display sub-module displays the performance evaluation results generated by the performance evaluation sub-module according to the prompts in the optimization prompt library and the updated data. These results usually include key performance indicators such as the generalization ability and convergence speed of the model.
[0053] Preferably, the implementation of the training monitoring interface can be refined in various ways. For example, the training instruction input sub-module can be designed as a graphical interface, where users can select training data sets, set training parameters, etc. by clicking and dragging the mouse. The optimization path display sub-module can display the training optimization path in the form of a chart, such as intuitively presenting the optimization steps and dependencies in the model training process through a flow chart or a tree diagram. The parameter display sub-module can display the adjusted parameters in the form of a table, facilitating users to compare and analyze the changes of different parameters. The training result display sub-module can update the metric change curve during the training process in real time, such as the change curve of accuracy with the number of training epochs, enabling users to intuitively see the effect of model training. The performance evaluation display sub-module can generate a detailed performance evaluation report, including the performance of the model on different test sets, comparative analysis with other models, etc., providing users with comprehensive performance feedback.
[0054] In some embodiments, the optimization hint library includes: a feature optimization hint sub-library that stores feature optimization hints written for the large-scale machine learning model based on the feature data;
[0055] a parameter adjustment hint sub-library that stores parameter adjustment hints written for the large-scale machine learning model based on the feature data;
[0056] a performance evaluation hint sub-library that stores performance evaluation hints written for the large-scale machine learning model based on the feature data.
[0057] It should be noted that the optimization hint library is an important part of the large-scale machine learning model incremental training and optimization system, which provides guidance and suggestions for the optimization process of the system. The optimization hint library includes a feature optimization hint sub-library, a parameter adjustment hint sub-library, and a performance evaluation hint sub-library. These sub-libraries provide specific hint information for the feature selection, parameter adjustment, and performance evaluation of the model respectively. These hint information are generated based on the feature data in the data feature library, which can help the system complete the model optimization task more efficiently. Through these optimization hints, the system can better understand how to adjust the model to adapt to new data and training requirements, thereby improving the performance and adaptability of the model.
[0058] Specifically, the three sub-libraries in the optimization hint library each undertake different functions. The feature optimization hint sub-library stores hints for feature selection and processing, such as which features have a greater impact on the model performance and which features can be removed to reduce the computational complexity. These hints may include feature importance scores, feature combination suggestions, etc. The parameter adjustment hint sub-library provides specific suggestions for model parameter adjustment, such as the adjustment strategy of the learning rate, the selection of regularization parameters, etc. These hints can help the system dynamically adjust parameters during the training process to achieve better convergence effects. The performance evaluation hint sub-library contains hints for evaluating the model performance, such as which performance metrics to select to measure the model effect and how to set the test set to verify the generalization ability of the model. These hints can help the system more comprehensively evaluate the model performance and further optimize the model according to the evaluation results.
[0059] Preferably, the hint information in the optimization hint library can be generated and refined in various ways. For example, the hints in the feature optimization hint sub-library can be generated by analyzing the feature data in the data feature library and combining machine learning algorithms (such as feature importance evaluation algorithms). Specifically, tree-based algorithms (such as random forests) can be used to evaluate the contribution of each feature to the model performance and generate feature optimization hints according to the contribution values. The hints in the parameter adjustment hint sub-library can be generated based on historical training data and model performance feedback. For example, according to the relationship between the learning rate and the convergence speed in the previous training process, hints for learning rate adjustment are generated. The hints in the performance evaluation hint sub-library can be generated according to different application scenarios and model types. For example, for classification models, evaluation hints for metrics such as accuracy, recall rate, and F1 score can be generated; for regression models, hints for metrics such as mean squared error and mean absolute error can be generated. In addition, these hint information can be updated regularly to adapt to the changes in data distribution and the adjustment of model structure.
[0060] In some embodiments, when the model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path, it first determines whether the training optimization path is in the data feature library. If it is in the data feature library, the parameter adjustment hints dedicated to the training optimization path in the parameter adjustment hint sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters; if it is not in the data feature library, the general parameter adjustment hints in the parameter adjustment hint sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters.
[0061] It should be noted that when generating adjustment parameters, the model adjustment sub-module will select different parameter adjustment prompts according to whether the training optimization path exists in the data feature library. If the training optimization path is recorded in the data feature library, then the parameter adjustment prompt dedicated to this path will be used to guide the model to generate adjustment parameters; if it does not exist, the general parameter adjustment prompt will be used. This mechanism can ensure that the model can utilize a more precise adjustment strategy when facing a known optimization path, and can also perform effective parameter adjustment through general prompts when facing a new optimization path, thereby improving the adaptability and optimization efficiency of the model.
[0062] Specifically, the operation logic of the model adjustment sub-module is based on the judgment of the training optimization path. The training optimization path refers to the direction and steps for guiding model training optimization generated by the feature analysis sub-module. The data feature library stores the feature data related to model training and its optimization path information. When the model adjustment sub-module receives the training optimization path, it will first check whether this path exists in the data feature library. If it exists, it means this is a known optimization path, and at this time, the parameter adjustment prompt dedicated to this path will be obtained from the parameter adjustment prompt sub-library. These dedicated prompts are usually customized according to the characteristics and requirements of specific optimization paths, and can more precisely guide the adjustment of model parameters. If the training optimization path is not in the data feature library, it means this is a new optimization path, and at this time, the general parameter adjustment prompt will be used. These general prompts are applicable to multiple optimization scenarios and can provide the model with basic adjustment directions and strategies.
[0063] Preferably, the model adjustment sub-module can further refine the operation steps when generating adjustment parameters. For example, when judging whether the training optimization path exists in the data feature library, it can quickly search through a hash table or index mechanism to improve the search efficiency. For the dedicated parameter adjustment prompt, its generation process can be based on historical training data and model performance feedback, and the optimal parameter adjustment strategy can be learned through machine learning algorithms (such as reinforcement learning). For the general parameter adjustment prompt, it can be formulated based on common machine learning theories and practical experiences. For example, according to the type of the model (such as neural network, decision tree, etc.) and the training objective (such as classification, regression, etc.), the appropriate parameter adjustment range and strategy can be selected. In addition, when generating adjustment parameters, the current state of the model (such as the current parameter value, training round, etc.) can also be combined for dynamic adjustment to ensure that the adjustment parameters can better adapt to the training needs of the model.
[0064] In some embodiments, each node in the model dependency graph represents an independent model module, and the unidirectional edge between nodes represents the dependency relationship of a model module, where the input of the tail node depends on the output of the head node.
[0065] It should be noted that the model dependency graph is an important part of the large-scale machine learning model incremental training and optimization system, which is used to represent the dependency relationships between various modules in the model. The model dependency graph is presented in the form of nodes and edges, where each node represents an independent model module, and the unidirectional edges between nodes represent the dependency relationships between modules. The input of the tail node depends on the output of the head node, and this structure can clearly show the computational logic and data flow direction of the model. By constructing the model dependency graph, the system can better analyze the correlation relationships between features, thereby optimizing the training process and improving the performance and efficiency of the model.
[0066] Specifically, the nodes and edges in the model dependency graph have clear definitions and functions. Nodes represent independent modules in the model, and these modules can be different layers of the model (such as hidden layers in a neural network), different functional components (such as feature extraction modules, classification modules, etc.), or different sub-models (such as in ensemble learning). Each node has its own input and output. The input comes from the output of other nodes or external data sources, and the output can be used as the input of other nodes. Edges represent the dependency relationships between nodes, and the direction of the unidirectional edge represents the direction of data flow. For example, in a deep neural network, the output of the previous layer is the input of the next layer, and this dependency relationship can be represented by the edges in the model dependency graph. The construction process of the model dependency graph is usually based on the structure definition of the model and the feature data in the feature database, and determines the connection method between nodes by analyzing the correlation relationships between features.
[0067] Preferably, the construction and use of the model dependency graph can be further refined. When constructing the model dependency graph, the overall structure of the model can be analyzed first to determine the functions and roles of each independent module. For example, in a complex deep learning model, the convolutional layer, pooling layer, fully connected layer, etc. can be regarded as independent nodes respectively. Then, according to the forward propagation and backward propagation processes of the model, determine the dependency relationships between nodes and represent them with unidirectional edges. When using the model dependency graph, the correlation relationships between features can be analyzed through graph traversal algorithms (such as depth-first search or breadth-first search), so as to provide suggestions for the optimization path of the feature analysis sub-module. In addition, the model dependency graph can be combined with the incremental training engine to dynamically adjust the training order and data update strategy according to the dependency relationships in the graph to improve the training efficiency. For example, during the incremental training process, if the input data of a certain node changes, the range of nodes that need to be recalculated can be quickly determined according to the model dependency graph, thereby reducing unnecessary calculations.
[0068] In some embodiments, the distributed computing environment is a GPU distributed computing environment.
[0069] It should be noted that the distributed computing environment is an important part of the large-scale machine learning model incremental training and optimization system for efficiently processing computing tasks. In this system, the distributed computing environment is specifically defined as a GPU distributed computing environment because the GPU (Graphics Processing Unit) has powerful parallel computing capabilities, which can significantly accelerate the training process of large-scale machine learning models, especially when dealing with complex model structures and massive data. By leveraging the GPU distributed computing environment, the system can distribute computing tasks to multiple GPUs for parallel execution, thereby improving training efficiency and reducing training time.
[0070] Specifically, the GPU in the distributed computing environment is a hardware device specifically designed for graphics and parallel computing. Compared with the traditional CPU (Central Processing Unit), the GPU has a large number of cores and can handle multiple computing tasks simultaneously, which makes it perform excellently in processing large-scale matrix operations and deep learning tasks. In this system, the GPU distributed computing environment usually consists of multiple GPUs, which can be multiple GPU cards on a single server or a GPU cluster distributed across multiple servers. By using the GPU distributed computing environment, the system can decompose the training tasks of large-scale machine learning models into multiple subtasks and distribute these subtasks to different GPUs for parallel execution. For example, in deep learning, the forward propagation and backward propagation calculations of the model can be assigned to different GPUs, and each GPU is responsible for processing a part of the data or model parameter calculations, thus significantly improving training efficiency.
[0071] Preferably, in order to better utilize the GPU distributed computing environment, the system can adopt various strategies to optimize the training process. For example, when building a large-scale machine learning model, different parts of the model (such as different layers or modules) can be assigned to different GPUs, and data and gradient information can be synchronized through the communication mechanism between GPUs (such as NVIDIA's NCCL library). In addition, the system can also dynamically adjust the GPU allocation strategy according to the model structure and data distribution to ensure the efficient use of computing resources. For example, for computationally intensive tasks, more GPU resources can be allocated; while for data-intensive tasks, the efficiency of data loading and transmission can be optimized. In practical applications, the system can dynamically adjust the number and allocation method of GPUs by monitoring the utilization rate of GPUs and the progress of training tasks, so as to achieve refined management of the training process.
[0072] The above-mentioned embodiments of the present invention have the following beneficial effects: This technical solution can improve the training efficiency and optimization accuracy of large-scale machine learning models. Through the training monitoring interface, instructions can be intuitively input and the training path, parameter adjustment, and performance evaluation results can be viewed in real time, simplifying the user operation process. The model dependency graph in the data processing module can analyze the complex associations between features, and combined with the data feature library and optimization hint library, it can automatically generate the optimal training strategy, reducing the cost of manual parameter tuning. The distributed computing environment can accelerate the incremental training process, and the incremental training engine can dynamically update the model parameters to ensure the efficiency and stability of the training process.
[0073] This technical solution can also enhance the intelligence and adaptability of model optimization. The feature analysis module can automatically generate a targeted training path based on the optimization hint library, and the model adjustment module can select dedicated or general parameter adjustment strategies according to the path matching situation, improving the optimization flexibility. The performance evaluation module can generate multi-dimensional evaluation results by combining the updated data to help users comprehensively understand the model performance. The modular design of the model dependency graph can clearly present the model structure dependency relationship, and the GPU distributed computing environment can further optimize the training speed, making the entire system more scalable and practical.
[0074] A method for incremental training and optimization of a large-scale machine learning model in some embodiments, the method includes:
[0075] S1 The user inputs a training instruction through the training monitoring interface;
[0076] The feature analysis sub-module uses the optimization hints in the optimization hint library based on the training instruction to guide the large-scale machine learning model to generate a training optimization path for the training instruction with reference to the feature data in the data feature library;
[0077] The model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path;
[0078] The incremental training sub-module uses the distributed computing environment to execute this set of adjustment parameters to update the corresponding data in the data cache;
[0079] The performance evaluation sub-module uses the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instruction according to the training instruction and the corresponding data updated by the incremental training sub-module from the data cache;
[0080] S2 Display the training results through the training monitoring interface.
[0081] It can be understood that the steps described in this method for incremental training and optimization of a large-scale machine learning model and referenceFigure 1 It corresponds to each module in the described large-scale machine learning model incremental training and optimization system. Therefore, the modules, features, and beneficial effects described above for the large-scale machine learning model incremental training and optimization system also apply to the large-scale machine learning model incremental training and optimization method and the operations included therein, which will not be elaborated here.
[0082] In some embodiments, the training results displayed through the training monitoring interface include the training optimization path, adjustment parameters, training execution results, and performance evaluation results.
[0083] It should be noted that the large-scale machine learning model incremental training and optimization method is an efficient training and optimization process based on a related system. This method starts the entire process by the user inputting training instructions through the training monitoring interface. Subsequently, the feature analysis sub-module uses the optimization tips in the optimization tip library to guide the large-scale machine learning model to refer to the feature data in the data feature library and generate a training optimization path for the training instructions. The model adjustment sub-module generates a set of adjustment parameters according to the training optimization path, and the incremental training sub-module uses the distributed computing environment to execute the adjustment parameters to update the corresponding data in the data cache. The performance evaluation sub-module generates performance evaluation results and displays the training results through the training monitoring interface. This process can effectively improve the training efficiency and performance of the large-scale machine learning model, and adapt to the dynamic data environment and real-time optimization requirements.
[0084] Specifically, several key terms and components involved in this method have clear definitions and functions. The training monitoring interface is the interface for user-system interaction. The user inputs training instructions through the training instruction input sub-module, and the system displays various information during the training process through the optimization path display sub-module, parameter display sub-module, training result display sub-module, and performance evaluation display sub-module. The feature analysis sub-module generates a training optimization path based on the training instructions and the tips in the optimization tip library, combined with the feature data in the data feature library. This path is the key guidance for optimizing model training. The model adjustment sub-module generates adjustment parameters according to the training optimization path, and these parameters are used to guide the adjustment and optimization of the model. The incremental training sub-module uses the distributed computing environment (such as the GPU distributed computing environment) to execute the adjustment parameters and update the data in the data cache, thereby realizing the incremental training of the model. The performance evaluation sub-module generates performance evaluation results according to the tips in the optimization tip library and the updated data, providing feedback on the model performance to the user.
[0085] Preferably, the implementation of the method can be further refined to improve efficiency and accuracy. For example, when the feature analysis sub-module generates a training optimization path, machine learning algorithms (such as decision trees, random forests, etc.) can be introduced to evaluate the importance of features and generate an optimization path based on the correlation between features. When the model adjustment sub-module generates adjustment parameters, it can dynamically adjust the parameter values in combination with the current state of the model (such as the current parameter values, the number of training rounds, etc.) to adapt to different training stages. When the incremental training sub-module performs incremental training, it can preferentially update the modules that have a greater impact on the model performance according to the dependency relationship between the nodes in the model dependency graph, thereby improving the training efficiency. When the performance evaluation sub-module generates performance evaluation results, it can adopt multiple performance metrics (such as accuracy, recall rate, F1 score, mean square error, etc.) to comprehensively evaluate the model performance and provide suggestions for further optimization based on the evaluation results. In addition, the system can also dynamically adjust the resource allocation in the distributed computing environment by monitoring the resource usage during the training process (such as GPU utilization, memory occupancy, etc.) to ensure the efficient progress of the training process.
[0086] The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), in-vehicle terminals (such as in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. The terminal devices shown in the present invention are only examples and should not impose any limitations on the functions and usage scopes of the embodiments of the present invention.
[0087] Furthermore, the storage medium of the embodiments of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions can be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. The aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROMs, Read-Only Memories), random access memories (RAMs, Random Access Memories), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.
[0088] The above description is only some preferred embodiments of the present invention and an explanation of the applied technical principles. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, but should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, the technical solutions formed by mutually replacing the above features with the technical features (but not limited to) having similar functions disclosed in the embodiments of the present invention.
Claims
1. An incremental training and optimization system for large-scale machine learning models, characterized in that, Including: A training monitoring interface for a user to input training instructions and display training results; A data processing module, which includes a data cache, a model dependency graph, a data feature library, an incremental training engine, a large-scale machine learning model, an optimization hint library, and a distributed computing environment. The model dependency graph is constructed based on each feature in the data cache and the data feature library to analyze the association relationships between the features in the data feature library. The data feature library stores feature data, which are training paths with practical significance sampled on the model dependency graph; The optimization hint library stores optimization hints written for the large-scale machine learning model based on the feature data; A model optimization assistance module, which includes a feature analysis sub-module, a model adjustment sub-module, an incremental training sub-module, and a performance evaluation sub-module. The feature analysis sub-module is used to use the optimization hints in the optimization hint library based on the training instructions input through the training monitoring interface to guide the large-scale machine learning model to generate a training optimization path for the training instructions with reference to the feature data in the data feature library; The model adjustment sub-module is used to guide the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path. The incremental training sub-module is used to use the distributed computing environment to execute the set of adjustment parameters to update the corresponding data in the data cache; The performance evaluation sub-module is used to use the optimization hints in the optimization hint library to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instructions according to the training instructions and the corresponding data updated by the incremental training sub-module from the data cache; Among them, the model dependency graph refers to a model structure with intricate interrelationships and complex calculation logics; 2. The large-scale machine learning model incremental training and optimization system according to claim 1, characterized in that, The training monitoring interface includes: a training instruction input sub-module for a user to input training instructions; An optimization path display sub-module for displaying the training optimization path obtained by the feature analysis sub-module; A parameter display sub-module for displaying a set of adjustment parameters obtained by the model adjustment sub-module; A training result display sub-module for displaying the result of the incremental training sub-module executing the set of adjustment parameters; A performance evaluation display sub-module for displaying a set of performance evaluation results obtained by the performance evaluation sub-module; 3. The large-scale machine learning model incremental training and optimization system according to claim 2, characterized in that, The optimization hint library includes: a feature optimization hint sub-library that stores feature optimization hints written for the large-scale machine learning model based on the feature data; A parameter adjustment hint sub-library that stores parameter adjustment hints written for the large-scale machine learning model based on the feature data; A performance evaluation hint sub-library that stores performance evaluation hints written for the large-scale machine learning model based on the feature data.
4. The large-scale machine learning model incremental training and optimization system according to claim 3, wherein When the model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path, it first determines whether the training optimization path is in the data feature library. If it is in the data feature library, the parameter adjustment prompts dedicated to the training optimization path in the parameter adjustment prompt sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters; If it is not in the data feature library, the general parameter adjustment prompts in the parameter adjustment prompt sub-library are used to guide the large-scale machine learning model to generate a set of adjustment parameters.
5. The large-scale machine learning model incremental training and optimization system according to claim 4, characterized in that Each node in the model dependency graph represents an independent model module respectively, and the unidirectional edge between nodes represents the dependency relationship of a model module, where the input of the tail node depends on the output of the head node.
6. The large-scale machine learning model incremental training and optimization system according to claim 5, characterized in that The distributed computing environment is a GPU distributed computing environment.
7. A method for incremental training and optimization of a large-scale machine learning model, characterized in that, It adopts the large-scale machine learning model incremental training and optimization system according to any one of claims 1-6, and includes the following steps: the user inputs a training instruction through the training monitoring interface; The feature analysis sub-module uses the optimization prompts in the optimization prompt library based on the training instruction to guide the large-scale machine learning model to generate a training optimization path for the training instruction with reference to the feature data in the data feature library; The model adjustment sub-module guides the large-scale machine learning model to generate a set of adjustment parameters according to the training optimization path; The incremental training sub-module uses the distributed computing environment to execute this set of adjustment parameters to update the corresponding data in the data cache; The performance evaluation sub-module uses the optimization prompts in the optimization prompt library to guide the large-scale machine learning model to generate a set of performance evaluation results for the training instruction according to the training instruction and the corresponding data updated by the incremental training sub-module from the data cache; The training result is displayed through the training monitoring interface.
8. The incremental training and optimization method for large-scale machine learning models according to claim 7, characterized in that The training result displayed through the training monitoring interface includes the training optimization path, adjustment parameters, training execution result and performance evaluation result.
9. An incremental training and optimization device for large-scale machine learning models, characterized in that, Including: One or more processors; A memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors are caused to implement the large-scale machine learning model incremental training and optimization method according to claim 7 or 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the large-scale machine learning model incremental training and optimization method according to claim 7 or 8.