An improved method and device for model training
By using matrix methods in model training to process data and generate model change reports, the training direction is determined, and the problem of inefficient model training is solved, and resource optimization and time saving is achieved.
Patent Information
- Application Number
- CN202411098639.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-12
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-12
AI Technical Summary
During the model training process, resources are wasted and inefficient due to frequent adjustments of data input and output. How to improve the efficiency of model training has become an urgent problem.
By filling the data to be trained in a matrix to the multi-dimensional data space, a model change report is generated, and the training direction is determined based on the report is determined to perform model training, reducing unnecessary data types and resource usage.
It reduces the time and resource consumption of model training, improves training efficiency, and avoids the tedious process of retraining after each parameter adjustment.
Smart Images

Figure CN118627573B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of model training technology, and in particular to an improved method and device for model training. Background Art
[0002] With the widespread application of large model technology in real society, the scale of training has become increasingly large, and the demand for computing resources has also increased dramatically, which not only leads to high material costs, but also makes the training process long and cumbersome.
[0003] Currently, in the process of model training and optimization, data is input into the model, output results are obtained, the output results are analyzed, analysis results are obtained, certain parameters are adjusted according to the analysis results, retraining is performed, and the performance of the model is evaluated again. However, the model needs to be retrained every time the parameters are adjusted, which consumes a lot of time and resources, resulting in low efficiency of model training.
[0004] Therefore, how to improve the efficiency of model training has become an urgent problem to be solved in this field. Summary of the invention
[0005] The present application provides an improved method and device for model training, aiming to improve the efficiency of model training.
[0006] In order to achieve the above objectives, this application provides the following technical solutions:
[0007] An improved method for model training, comprising:
[0008] Obtain multiple data to be trained;
[0009] For each of the to-be-trained data, the to-be-trained data is filled into the multi-dimensional data space of the matrix space in a matrix manner to obtain multiple information regions;
[0010] Inputting the plurality of information regions into the model respectively to obtain a training result for each of the information regions;
[0011] The training results of each of the information regions are stored in the target region of the matrix space in the matrix manner; the target region is another multi-dimensional data space;
[0012] Generate a model change report based on all the target areas and all the information areas;
[0013] Determine the relevant direction of the model training according to the model change report, and train the model according to the relevant direction of the model training.
[0014] Optionally, after obtaining the plurality of to-be-trained data, the method further includes:
[0015] Acquire the magnitude and precision requirement of the data to be trained; the magnitude of the data to be trained indicates the quantity of the data to be trained and the dimension of the data to be trained, and the precision requirement indicates the accuracy of the calculation result;
[0016] Constructing a data space according to the magnitude of the data to be trained and the accuracy requirement;
[0017] Transforming the data space to obtain an initial matrix space;
[0018] The storage space of the initial matrix space is reduced by using a sparse matrix method to obtain a matrix space; the matrix space includes a multidimensional data space.
[0019] Optionally, generating a model change report based on all the target areas and all the information areas includes:
[0020] For each of the information regions, comparing the information region with a target region corresponding to the information region to obtain a comparison result;
[0021] A model change report is generated based on all the comparison results.
[0022] Optionally, determining the relevant direction of the model training according to the model change report includes:
[0023] From the model change report, a region having the largest gap between the information region and the target region is selected and marked as a sensitive interaction region;
[0024] The sensitive interaction area is determined as a relevant direction for the model training.
[0025] Optionally, also include:
[0026] Filtering out, from the model change report, areas where there are differences between the information area and a target area corresponding to the information area, and marking them as activated information areas;
[0027] Generate a spatial logic chain reaction diagram according to the activation information area;
[0028] Screening out data with deviations from the spatial logic chain reaction diagram;
[0029] Analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of the model training;
[0030] The model is trained according to the data type of the data to be trained and the relevant direction of the model training.
[0031] An improved device for model training, comprising:
[0032] An acquisition unit, used for acquiring a plurality of to-be-trained data;
[0033] A filling unit is used for filling each of the to-be-trained data into the multi-dimensional data space of the matrix space in a matrix manner to obtain multiple information regions;
[0034] An input unit, used to input the plurality of information regions into the model respectively, to obtain a training result for each of the information regions;
[0035] A storage unit, used for storing the training results of each of the information regions in the matrix manner in sequence into a target region of the matrix space; the target region is another of the multidimensional data spaces;
[0036] A generating unit, configured to generate a model change report based on all the target areas and all the information areas;
[0037] A training unit is used to determine the relevant direction of the model training according to the model change report, and train the model according to the relevant direction of the model training.
[0038] Optionally, also include:
[0039] A data acquisition unit, used to acquire the magnitude and precision requirement of the data to be trained; the magnitude of the data to be trained indicates the quantity of the data to be trained and the dimension of the data to be trained, and the precision requirement indicates the accuracy of the calculation result;
[0040] A construction unit, used to construct a data space according to the magnitude of the data to be trained and the accuracy requirement;
[0041] A conversion unit, used for converting the data space to obtain an initial matrix space;
[0042] The reduction unit is used to reduce the storage space of the initial matrix space by using a sparse matrix method to obtain a matrix space; the matrix space includes a multidimensional data space.
[0043] Optionally, the generating unit is specifically used for:
[0044] For each of the information regions, comparing the information region with a target region corresponding to the information region to obtain a comparison result;
[0045] A model change report is generated based on all the comparison results.
[0046] Optionally, the training unit is specifically used for:
[0047] From the model change report, a region having the largest gap between the information region and the target region is selected and marked as a sensitive interaction region;
[0048] The sensitive interaction area is determined as a relevant direction for the model training.
[0049] Optionally, also include:
[0050] A screening unit, configured to screen out, from the model change report, an area where there is a difference between the information area and a target area corresponding to the information area, and mark the area as an activated information area;
[0051] An information generating unit, used for generating a spatial logic chain reaction diagram according to the activation information area;
[0052] A data screening unit, used for screening out data with deviations from the spatial logic chain reaction diagram;
[0053] An analysis unit, used to analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of the model training;
[0054] The model training unit is used to train the model according to the data type of the data to be trained and the relevant direction of the model training.
[0055] The technical solution provided by the present application obtains multiple data to be trained; for each data to be trained, fills the data to be trained into the multidimensional data space of the matrix space in a matrix manner to obtain multiple information areas; inputs the multiple information areas into the model respectively to obtain the training results of each information area; stores the training results of each information area in a matrix manner in the target area of the matrix space in sequence; generates a model change report based on all target areas and all information areas; determines the relevant directions of model training according to the model change report, and trains the model according to the relevant directions of model training. In the present application, the area where the model parameters change significantly is determined by the model change report, and the relevant directions of model training can be obtained by analyzing the area where the model parameters change, and the model can be trained by the data to be trained in the relevant directions of model training, and there is no need to train the model according to the data type of the data to be trained in each direction, which reduces the training time and resources of model training, thereby improving the training efficiency of model training. BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0057] Figure 1 A flowchart of an improved method for model training provided in an embodiment of the present application;
[0058] Figure 2 A flowchart of a method for constructing a data space provided in an embodiment of the present application;
[0059] Figure 3 A flowchart of another model training method provided in an embodiment of the present application;
[0060] Figure 4 A schematic diagram of the architecture of an improved device for model training provided in an embodiment of the present application. DETAILED DESCRIPTION
[0061] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.
[0062] In this application, the terms "comprises", "comprising" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, an element defined by the sentence "comprising a ..." does not exclude the presence of other identical elements in the process, method, article or device comprising the element.
[0063] like Figure 1 As shown, it is a flowchart of an improved method for model training provided in this application, comprising the following steps:
[0064] S101: Acquire a plurality of data to be trained.
[0065] The data to be trained refers to the data set used to train the machine learning model.
[0066] It should be noted that the data to be trained corresponds to the model to be trained. For example, if the model to be trained is an image recognition model, then the data to be trained is an image.
[0067] Optionally, in order to make data management and retrieval more convenient, and also to provide great convenience for subsequent processing and analysis, it is necessary to construct a data space in advance to store data, and then directly extract data from this data space. Therefore, in another embodiment of the present application, a method for constructing a data space is provided, such as Figure 2 As shown, the following steps are included:
[0068] S201: Obtain the magnitude and accuracy requirements of the data to be trained.
[0069] The magnitude of the data to be trained indicates the number of data to be trained and the dimension of the data to be trained, and the accuracy requirement indicates the accuracy of the calculation result.
[0070] Optionally, accuracy requirements include but are not limited to: centimeter level.
[0071] S202: Construct a data space according to the magnitude and accuracy requirements of the data to be trained.
[0072] The data space is constructed according to the magnitude and accuracy requirements of the data to be trained. Specifically, the data volume of the data space is obtained by first estimating the magnitude and accuracy requirements of the data to be trained; and the data space is constructed according to the data volume.
[0073] For example, a rectangular space area with a length of 15 meters, a width of 4 meters, and a height of 5 meters. The accuracy requirement is at the centimeter level. The data space is constructed according to the magnitude and accuracy requirements of the training data, and the resulting data space is: 1500*400*500.
[0074] For example, when the training data is an image, we can create an x-by-y two-dimensional data space, in which the x-axis represents the horizontal coordinate of the pixel in the image, increasing from left to right; the y-axis represents the vertical coordinate of the pixel, increasing from top to bottom. The value at each coordinate point represents the color value of the pixel at that location, which can be a grayscale value, RGB value, or a value in other color spaces.
[0075] It should be noted that, in the content of the data space, the coordinates of the data space can be used as identification bits, and the geometric features of the data can be marked through the identification bits.
[0076] S203: Convert the data space to obtain an initial matrix space.
[0077] Among them, the data space is converted. Specifically, the dimension of the matrix (determined by the number of rows and columns) and the element type (such as integer) are defined; a matrix space corresponding to the data space is constructed according to the dimension and element type of the matrix to obtain the initial matrix space.
[0078] It should be noted that the matrix space is an efficient data structure that can conveniently represent and operate multidimensional data. In order to facilitate data processing, the data space needs to be converted into the initial matrix space.
[0079] S204: Using a sparse matrix method to reduce the storage space of the initial matrix space to obtain a matrix space.
[0080] Among them, the matrix space includes multi-dimensional data space.
[0081] It should be noted that the sparse matrix method includes compressed storage method, linked storage method and index storage method. The compressed storage method stores non-zero elements; the linked storage method stores non-zero elements in the order of rows or columns, and records the position information of the row or column where the element is located; the index storage method stores the row index and column index of the non-zero element, as well as the element.
[0082] It can be seen that the initial matrix space is reduced by storing non-zero elements to reduce the use of storage space. Especially when the number of features is large but most of the eigenvalues are zero, the storage space of the initial matrix space is greatly reduced by using a sparse matrix.
[0083] In addition, the efficiency and speed of data processing can be improved through parallel computing and distributed storage technology. The above methods can provide an efficient and reliable data foundation for subsequent machine learning or data analysis tasks.
[0084] S102: For each to-be-trained data, fill the to-be-trained data into the multi-dimensional data space of the matrix space in a matrix manner to obtain multiple information regions.
[0085] The multidimensional data space includes multiple multidimensional coordinate systems, and each multidimensional coordinate system includes multiple dimensions.
[0086] It can be understood that the data to be trained is filled into the multidimensional data space of the matrix space in a matrix manner. Specifically, the data to be trained is filled into each dimension in the multidimensional coordinate system in a matrix manner to obtain an information region.
[0087] In addition, the dimension of the matrix space can be increased to distribute the data. By increasing the dimension, not only can the distribution and relationship of the data be displayed more intuitively, but also the retrieval efficiency and display effect of the data can be greatly improved.
[0088] For example, when the data to be trained is an image, the image is a two-dimensional matrix data in grayscale format, which is the horizontal coordinate of the pixel and the vertical coordinate of the pixel. The matrix space is a two-dimensional matrix space, and a third dimension can be introduced on the basis of the traditional two-dimensional matrix space. This newly added dimension can be called "image index" or "data layer", which is used to store and distinguish different images. Specifically, it can be set that when z=0, the first image is stored in each dimension of the multidimensional coordinate system; when z=1, the second image is stored in each dimension of the multidimensional coordinate system; and so on, until all the images to be trained are neatly arranged in this multidimensional data space.
[0089] It should be noted that increasing the dimension of the matrix space makes data management and retrieval more convenient, and also provides great convenience for subsequent image processing and analysis. For example, when training a model, data can be directly extracted from the multidimensional data space without additional data preprocessing or organization. In addition, this way of constructing a multidimensional data space is also scalable to a certain extent. More dimensions can be added as needed to store other types of data, such as labels, features, etc., thereby further enriching the expression of data.
[0090] S103: Inputting multiple information regions into the model respectively to obtain training results for each information region.
[0091] Optionally, multiple information regions are input into the model in sequence in row order or in column order to obtain a training result for each information region.
[0092] The training result refers to the specific changes of the model (such as the parameters of the model). For example, if the model is an image recognition model, the data to be trained is an image, and the image is input into the image recognition model to obtain the recognition result of the image and the adjusted model parameters.
[0093] It should be noted that the training results are still data in matrix form.
[0094] S104: The training results of each information region are stored in a matrix manner in sequence in the target region of the matrix space.
[0095] Among them, the target area is another multidimensional data space.
[0096] It can be understood that the training results of each information area are stored in a matrix manner in each dimension of a multidimensional coordinate system of another multidimensional data space.
[0097] In addition, the target area can be displayed to the user in a visual way, so that the user can more intuitively understand the training effect of the model, thereby more accurately evaluating the performance of the model and making corresponding improvements and optimizations as needed.
[0098] S105: Generate a model change report based on all target areas and all information areas.
[0099] Among them, the model change report indicates the change display diagram of the information area.
[0100] It can be understood that a model change report is generated through all target areas and all information areas, wherein the target area shows the information changes of each step of the training process (for example, the parameter changes of the model). The specific changes of the model can be clearly understood through the model change report.
[0101] Optionally, in another embodiment of the present application, the specific implementation of step S105 includes:
[0102] For each information region, the information region is compared with a target region corresponding to the information region to obtain a comparison result.
[0103] By comparing the information region with the target region corresponding to the information region, the change process of the corresponding information region, ie, the comparison result, can be learned.
[0104] Generate a model change report based on all alignment results.
[0105] It is understandable that the model change report can track the performance changes of the model at different stages, such as the performance improvement during the initial training stage and after tuning. This helps to determine whether the model is developing in the expected direction and whether there is potential for performance improvement.
[0106] S106: Determine relevant directions of model training according to the model change report, and train the model according to the relevant directions of model training.
[0107] Among them, the model training related direction refers to the reinforcement training direction that causes obvious changes in model parameters.
[0108] It can be understood that the model change report can be used to see the data type of the data to be trained that causes significant changes in model parameters, determine the relevant direction of model training based on the data type of the data to be trained, and perform intensive training on the model based on the relevant direction.
[0109] For example, if we want to identify cups, we need to constantly find different types of cups for training using traditional training methods. However, after extracting the common weight matrix of different types of cups based on the relevant direction of model training, we can quickly strengthen the model. This method is similar to the contour line in sketching techniques. We only need to simply outline the object (description matrix) to define the basic type of an object.
[0110] Optionally, in another embodiment of the present application, the specific implementation method of determining the relevant direction of model training according to the model change report in step S106 includes:
[0111] From the model change report, the areas with the largest gap between the information area and the target area are screened out and marked as sensitive interaction areas.
[0112] Among them, the sensitive interaction area is the area that causes obvious changes in model parameters.
[0113] It can be understood that from the model change report, the area with the largest gap between the information area and the target area is screened out, that is, the area that causes obvious changes in model parameters.
[0114] Sensitive interaction regions are identified as relevant directions for model training.
[0115] Among them, the sensitive interaction area is determined as the relevant direction of model training. Specifically, the data to be trained corresponding to the sensitive interaction area is obtained, the data type of the data to be trained is determined according to multiple data to be trained, and the data type of the data to be trained is determined as the relevant direction of model training.
[0116] Optionally, after step S106, in order to enhance the robustness of the model, the model may be further trained using other data types and related directions. Therefore, another model training method is provided in another embodiment of the present application, such as Figure 3 As shown, the following steps are included:
[0117] S301: Filter out areas where there are differences between an information area and a target area corresponding to the information area from the model change report, and mark them as activated information areas.
[0118] Among them, from the model change report, the area where there is a difference between the information area and the target area corresponding to the information area is screened out, that is, the area where the model parameter changes greatly is found, and the area where the model parameter changes greatly is determined as the activated information area.
[0119] For example, if the model is a curve model, the output y value after reading in x can be displayed in real time in the form of an image by the curve equation. For more complex ones, such as image recognition, the process based on model recognition, that is, the similarity scoring system at each step in the matching process can also be displayed in the form of an image (the actual calculation process is a matrix). When evaluating the model, it is very intuitive to find those areas with low matching degrees that affect the recognition results (that is, the activation information area).
[0120] S302: Generate a spatial logic chain reaction diagram according to the activation information area.
[0121] It is understandable that the Spatial Logic Chain Reaction Map is often used to visualize and analyze the relationships and influences between complex information regions. This kind of diagram can help understand the spatial layout of information and its associations, and is suitable for dealing with multiple activated information regions. Therefore, by generating a logic chain reaction map, we can subsequently find the data type and related directions of the data to be trained.
[0122] For example, when we use the model to determine whether an object is a cup, we need to make judgments in different areas of the image (usually cutting the image into small areas of x*x size for matching calculations). When the overall matching degree reaches the standard score, it is considered a successful recognition. Through the sensitive interaction area, we can simply outline the outline that needs to be matched in the image, thereby showing the judgment logic of the current model. That is, the spatial logic chain reaction diagram is a kind of graphical spatial weight matrix with a certain spatial judgment order.
[0123] S303: Filter out data with deviations from the spatial logic chain reaction diagram.
[0124] Among them, the deviated data are screened out from the spatial logic chain reaction diagram, that is, the data in which the training results in the target area deviate from the actual results.
[0125] S304: Analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of model training.
[0126] Among them, the data with deviations are analyzed, and the sources and causes of the deviation data are analyzed to obtain the data type and related direction of the data to be trained.
[0127] S305: Train the model according to the data type of the data to be trained and the relevant direction of the model training.
[0128] It should be noted that the data type of the data to be trained can be multiple types, and the model is trained in turn according to the data types and related directions of model training. Training with multiple data types can improve the generalization ability of the model, and training the model for specific data types and problem directions can effectively improve the efficiency of problem solving.
[0129] In summary, the model change report is used to determine the areas where the model parameters have changed significantly. By analyzing the areas where the model parameters have changed, the relevant directions of model training can be obtained. The model can be trained using the training data in the relevant directions of model training. There is no need to train the model according to the data type of the training data in each direction, which reduces the training time and resources of model training and improves the training efficiency of model training.
[0130] like Figure 4 As shown, it is a schematic diagram of the architecture of an improved device for model training provided in the present application, and the improved device includes: an acquisition unit 100, a filling unit 200, an input unit 300, a storage unit 400, a generation unit 500 and a training unit 600.
[0131] The acquisition unit 100 is used to acquire a plurality of to-be-trained data.
[0132] The filling unit 200 is used to fill each to-be-trained data into the multi-dimensional data space of the matrix space in a matrix manner to obtain multiple information regions.
[0133] The input unit 300 is used to input multiple information regions into the model respectively to obtain the training result of each information region.
[0134] The storage unit 400 is used to store the training results of each information area in a matrix manner in sequence into a target area of the matrix space; the target area is another multi-dimensional data space.
[0135] The generating unit 500 is used to generate a model change report based on all target areas and all information areas.
[0136] The generating unit 500 is specifically used for: for each information region, comparing the information region with the target region corresponding to the information region to obtain a comparison result; and generating a model change report based on all the comparison results.
[0137] The training unit 600 is used to determine the relevant direction of model training according to the model change report, and train the model according to the relevant direction of model training.
[0138] The training unit 600 is specifically used to: filter out the area with the largest gap between the information area and the target area from the model change report, and mark it as the sensitive interaction area; determine the sensitive interaction area as the relevant direction of model training.
[0139] In summary, the model change report is used to determine the areas where the model parameters have changed significantly. By analyzing the areas where the model parameters have changed, the relevant directions of model training can be obtained. The model can be trained using the training data in the relevant directions of model training. There is no need to train the model according to the data type of the training data in each direction, which reduces the training time and resources of model training and improves the training efficiency of model training.
[0140] Preferably, combined Figure 4 The improved device also includes: a data acquisition unit, a construction unit, a conversion unit and a reduction unit.
[0141] The data acquisition unit is used to acquire the magnitude and accuracy requirement of the data to be trained; the magnitude of the data to be trained indicates the number of the data to be trained and the dimension of the data to be trained, and the accuracy requirement indicates the accuracy of the calculation result.
[0142] The construction unit is used to construct a data space according to the magnitude and accuracy requirements of the data to be trained.
[0143] The conversion unit is used to convert the data space to obtain the initial matrix space.
[0144] The reduction unit is used to reduce the storage space of the initial matrix space by using a sparse matrix method to obtain a matrix space; the matrix space includes a multidimensional data space.
[0145] Preferably, combined Figure 4 The improved device also includes: a screening unit, an information generating unit, a data screening unit, an analyzing unit and a model training unit.
[0146] The screening unit is used to screen out the area where there is a difference between the information area and the target area corresponding to the information area from the model change report, and mark it as the activated information area.
[0147] The information generating unit is used to generate a spatial logic chain reaction diagram according to the activated information area.
[0148] The data screening unit is used to screen out the data with deviation from the spatial logic chain reaction diagram.
[0149] The analysis unit is used to analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of model training.
[0150] The model training unit is used to train the model according to the data type of the data to be trained and the relevant direction of the model training.
[0151] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can refer to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system or system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the scheme of this embodiment. Ordinary technicians in this field can understand and implement it without paying creative work.
[0152] Professionals may further appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in the above description according to function. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professionals and technicians may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.
[0153] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. An improved method for model training, characterized in that: include: Obtain multiple data to be trained; The data to be trained is an image; For each of the to-be-trained data, the to-be-trained data is filled into a multi-dimensional data space of a matrix space in a matrix manner to obtain a plurality of information regions; Inputting the plurality of information regions into the model respectively to obtain a training result for each of the information regions; The model is an image recognition model; the training result is the image recognition result and the adjusted model parameters; The training results of each of the information regions are stored in the target region of the matrix space in sequence in the matrix manner; The target area is another multi-dimensional data space; Generate a model change report based on all the target areas and all the information areas; Determining a relevant direction of the model training according to the model change report, and training the model according to the relevant direction of the model training; The step of training the model according to the relevant direction of the model training includes: From the model change report, a region having the largest gap between the information region and the target region is selected and marked as a sensitive interaction region; Determining the sensitive interaction area as a relevant direction for the model training; After determining the relevant direction of the model training according to the model change report and training the model according to the relevant direction of the model training, the method further includes: Filtering out, from the model change report, areas where there are differences between the information area and a target area corresponding to the information area, and marking them as activated information areas; Generate a spatial logic chain reaction diagram according to the activation information area; Screening out data with deviations from the spatial logic chain reaction diagram; Analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of the model training; The model is trained according to the data type of the data to be trained and the relevant direction of the model training.
2. The method according to claim 1, characterized in that After obtaining a plurality of training data, the method further includes: Acquire the magnitude and precision requirement of the data to be trained; the magnitude of the data to be trained indicates the quantity of the data to be trained and the dimension of the data to be trained, and the precision requirement indicates the accuracy of the calculation result; Constructing a data space according to the magnitude of the data to be trained and the accuracy requirement; Transforming the data space to obtain an initial matrix space; The storage space of the initial matrix space is reduced by using a sparse matrix method to obtain a matrix space; the matrix space includes a multidimensional data space.
3. The method according to claim 1, characterized in that: The generating of the model change report based on all the target areas and all the information areas includes: For each of the information regions, comparing the information region with a target region corresponding to the information region to obtain a comparison result; A model change report is generated based on all the comparison results.
4. An improved device for model training, characterized in that: include: An acquisition unit, used for acquiring a plurality of to-be-trained data; The data to be trained is an image; A filling unit is used for filling each of the to-be-trained data into the multi-dimensional data space of the matrix space in a matrix manner to obtain multiple information regions; An input unit, used to input the plurality of information regions into the model respectively, to obtain a training result for each of the information regions; The model is an image recognition model; the training result is the image recognition result and the adjusted model parameters; A storage unit, used for storing the training results of each of the information regions in the target region of the matrix space in sequence in the matrix manner; The target area is another multi-dimensional data space; A generating unit, configured to generate a model change report based on all the target areas and all the information areas; A training unit, used to determine a relevant direction of the model training according to the model change report, and train the model according to the relevant direction of the model training; Wherein, the training unit is specifically used for: From the model change report, a region having the largest gap between the information region and the target region is selected and marked as a sensitive interaction region; Determining the sensitive interaction area as a relevant direction for the model training; After the training unit, it also includes: A screening unit, configured to screen out, from the model change report, an area where there is a difference between the information area and a target area corresponding to the information area, and mark the area as an activated information area; An information generating unit, used for generating a spatial logic chain reaction diagram according to the activation information area; A data screening unit, used for screening out data with deviations from the spatial logic chain reaction diagram; An analysis unit, used to analyze the data with deviations to obtain the data type of the data to be trained and the relevant direction of the model training; The model training unit is used to train the model according to the data type of the data to be trained and the relevant direction of the model training.
5. The device according to claim 4, characterized in that Also includes: A data acquisition unit, used to acquire the magnitude and precision requirement of the data to be trained; the magnitude of the data to be trained indicates the quantity of the data to be trained and the dimension of the data to be trained, and the precision requirement indicates the accuracy of the calculation result; A construction unit, used to construct a data space according to the magnitude of the data to be trained and the accuracy requirement; A conversion unit, used for converting the data space to obtain an initial matrix space; A reduction unit, used for reducing the storage space of the initial matrix space by using a sparse matrix method to obtain a matrix space; The matrix space includes a multi-dimensional data space.
6. The device according to claim 4, characterized in that The generating unit is specifically used for: For each of the information regions, comparing the information region with a target region corresponding to the information region to obtain a comparison result; A model change report is generated based on all the comparison results.
