A model training sample dynamic optimization method, device, equipment and storage medium

Abstract

The application discloses a model training sample dynamic optimization method and device, equipment and a storage medium. The method comprises the following steps: obtaining pre-training process data corresponding to a target training model obtained by pre-training; obtaining first sample training data and first sample prediction data corresponding to the first sample training data based on a preset training period; determining a quality evaluation score corresponding to the first sample training data based on a preset evaluation index, the pre-training process data, the first sample training data and the first sample prediction data; optimizing the first sample training data based on a quality evaluation threshold and the quality evaluation score to obtain second sample training data, so as to continue training the target training model based on the second sample training data in the next preset training period. The application can realize automatic and efficient optimization of model training samples, improve the adaptability to a complex model training process, and improve the training efficiency and quality of the model.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method, apparatus, device, and storage medium for dynamic optimization of model training samples. Background Technology

[0002] In the training process of artificial intelligence models, the quality of training data has a decisive impact on model performance. Effective optimization of training data has important theoretical research significance and wide-ranging engineering application value.

[0003] Currently, traditional model training sample optimization methods typically only perform static data cleaning (such as manual annotation and validation, rule filtering, etc.) before training to optimize training data. However, traditional model training sample optimization methods are poorly adaptable to the complex and dynamically evolving model training process and cannot cope with new sample quality issues exposed during training due to changes in model cognition, resulting in low training efficiency and quality. Summary of the Invention

[0004] This invention provides a method, apparatus, device, and storage medium for dynamic optimization of model training samples, enabling automatic and efficient optimization of model training samples. This improves the adaptability to complex and dynamically evolving model training processes and addresses new sample quality issues exposed during training due to changes in model cognition, thereby significantly improving the training efficiency and quality of the model.

[0005] According to one aspect of the present invention, a method for dynamic optimization of model training samples is provided, the method comprising: Obtain the pre-training process data corresponding to the target training model obtained from pre-training; Based on a preset training period, the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data are obtained. Based on the preset evaluation indicators, the pre-training process data, the first sample training data, and the first sample prediction data, the quality evaluation score corresponding to the first sample training data is determined. The first sample training data is optimized based on the quality assessment threshold and the quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

[0006] According to another aspect of the present invention, a model training sample dynamic optimization apparatus is provided, the apparatus comprising: The first data acquisition module is used to acquire the pre-training process data corresponding to the target training model obtained through pre-training. The second data acquisition module is used to acquire the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data based on a preset training period. The evaluation score determination module is used to determine the quality evaluation score corresponding to the first sample training data based on the preset evaluation index, the pre-training process data, the first sample training data, and the first sample prediction data. The training data optimization module is used to optimize the first sample training data based on the quality assessment threshold and the quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

[0007] According to another aspect of the present invention, an electronic device is provided, the electronic device comprising: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model training sample dynamic optimization method according to any embodiment of the present invention.

[0008] According to another aspect of the present invention, a computer-readable storage medium is provided, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the model training sample dynamic optimization method according to any embodiment of the present invention.

[0009] The technical solution of this invention provides a data foundation for subsequent steps by acquiring pre-training process data corresponding to the target training model obtained through pre-training. Based on a preset training period, it acquires first sample training data corresponding to the target training model and first sample prediction data corresponding to the first sample training data, providing the latest and targeted input for subsequent quality assessment, ensuring the timeliness and relevance of the assessment. Based on preset assessment indicators, the pre-training process data, the first sample training data, and the first sample prediction data, it determines the quality assessment score corresponding to the first sample training data, achieving a comprehensive quality assessment of the training samples and greatly improving the reliability of the assessment score. Based on the quality assessment threshold and the quality assessment score, it optimizes the first sample training data to obtain second sample training data, which is then used to continue training the target training model in the next preset training period, thereby improving the training quality of the model. This invention achieves automatic and efficient optimization of model training samples by evaluating and dynamically optimizing training samples during model training, and continuing model training based on the optimized training samples. This enables the model to continuously identify and suppress interference from low-quality data during training, thereby improving its adaptability to complex and dynamically evolving model training processes. It can also address new sample quality issues exposed during training due to changes in model cognition, greatly improving the training efficiency and quality of the model.

[0010] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of the present invention, nor is it intended to limit the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a flowchart of a dynamic optimization method for model training samples provided in Embodiment 1 of the present invention; Figure 2 This is a flowchart of a dynamic optimization method for model training samples provided in Embodiment 2 of the present invention; Figure 3 This is a schematic diagram of the structure of a model training sample dynamic optimization device provided in Embodiment 3 of the present invention; Figure 4 This is a schematic diagram of the structure of an electronic device that implements the model training sample dynamic optimization method of the present invention. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0014] It should be noted that the terms "first," "second," "target," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0015] Example 1 Figure 1 This is a flowchart illustrating a method for dynamically optimizing model training samples according to Embodiment 1 of the present invention. This embodiment is applicable to situations where training samples are optimized during model training. This method can be executed by a dynamic optimization device for model training samples, which can be implemented in hardware and / or software. This dynamic optimization device for model training samples can be configured in an electronic device. Figure 1 As shown, the method includes: S110. Obtain the pre-training process data corresponding to the target training model obtained from pre-training.

[0016] The target training model can refer to a model that has acquired preliminary discriminative ability after pre-training. The pre-training process data can refer to a multi-dimensional data set that is continuously recorded and stored during the pre-training phase of the target training model, reflecting the interaction between the model and the training samples.

[0017] Specifically, a target training model with preliminary discrimination ability can be obtained through pre-training, and the pre-training process data corresponding to the target training model can be automatically collected and saved after the target training model is generated, which can provide a data foundation for backtracking analysis in subsequent steps.

[0018] For example, S110 may include: performing a preset number of pre-training rounds on the original training model based on the initialized training dataset and initialized model parameters to generate a target training model; extracting data from the target training model to obtain the pre-training process data corresponding to the target training model, wherein the pre-training process data includes: sample identification data, model output data, model loss data, sample feature data, and model parameter data.

[0019] The initial training dataset can refer to the unfiltered set of raw training samples used for model training. Initial model parameters can refer to the initial values ​​set for the weights and biases of each layer (such as convolutional layers and fully connected layers) of the neural network before training begins. The raw training model can refer to a machine learning model that has not yet been trained on the sample dataset and does not possess effective discriminative ability. Sample label data can refer to the index information used to uniquely identify, locate, and track each training sample in the dataset and process data; it is the unique ID of the sample in the dataset. Model output data can refer to the raw data generated in the last layer after the model performs forward propagation on the input samples, used for task prediction. For example, for a classification task, the model output data can be a predicted probability distribution vector. Model loss data can refer to a quantized (scalar) value calculated using a loss function based on the model output data and the true labels of the samples, used to reflect the difference between the model's current prediction and the actual situation. Sample feature data can refer to the feature vectors extracted from the intermediate layers (non-output layers) of the model, representing a high-level abstract representation of the input samples. Model parameter data can refer to the values ​​of all parameters in each layer of the model, such as the weight matrix and bias vector.

[0020] Specifically, the sample weights in the initial training dataset can be set to 1 by default. The original training model is initialized based on the initial model parameters. Based on the initial training dataset, a weighted loss function (such as weighted cross-entropy) can be used to pre-train the initialized original training model for a preset number of rounds to generate a target training model with preliminary discrimination capabilities. Data extraction is performed on the training process of the target training model. Sample identification data, model output data, model loss data, sample feature data, and model parameter data are collected after each training cycle in the pre-training loop. This obtains the pre-training process data corresponding to the target training model, thereby avoiding the training instability caused by dynamic adjustment of samples from the original training model. This ensures that the entire optimization process starts from a relatively robust benchmark, providing a reliable data foundation for subsequent steps.

[0021] S120. Based on a preset training period, obtain the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data.

[0022] The preset training period can refer to a pre-set model training period used to automatically trigger the sample quality assessment and optimization process. The first sample training data can refer to the set of samples in the entire training set currently being trained when the preset training period arrives. The first sample prediction data can refer to the output data generated by the target training model at the current moment (i.e., the preset training period) after performing a forward propagation calculation on the first sample training data, which includes all signals valuable for assessing sample quality.

[0023] Specifically, in the main training phase after pre-training, a preset training period can be used as a quality assessment frequency parameter to periodically trigger the training sample assessment process. That is, after iteratively training the target training model for a preset training period, the sample set used by the target training model within the preset training period is used as the first sample training data, and the latest model output of the first sample training data within the preset training period is obtained, that is, the first sample prediction data corresponding to the first sample training data. This provides the latest and targeted input for subsequent quality assessment, ensuring the timeliness and relevance of the assessment.

[0024] For example, S120 may include: in response to the training period of the target training model reaching a preset training period, determining the first sample training data corresponding to the target training model in the current training period; and determining the first sample prediction data corresponding to the first sample training data based on the forward propagation result of the target training model in the current training period.

[0025] The forward propagation results can refer to data such as output values, loss function values, and intermediate layer activation values ​​calculated by the model through the forward propagation process.

[0026] Specifically, after determining the target training model, iterative training is performed on the target training model according to a preset training cycle. When the target training model completes the iterative training of the preset training cycle, the sample training dataset of the target training model within the preset training cycle is determined as the first sample training data. Based on the forward propagation result of the last training cycle of the target training model within the preset training cycle, that is, the current training cycle, the first sample prediction data corresponding to the first sample training data is obtained, thereby ensuring the timeliness and accuracy of the information on which the subsequent quality assessment is based.

[0027] S130. Based on the preset evaluation indicators, pre-training process data, first sample training data and first sample prediction data, determine the quality evaluation score corresponding to the first sample training data.

[0028] The preset evaluation metrics can refer to predefined, multi-dimensional evaluation metrics used to quantitatively measure the quality of a single training sample. The quality evaluation score can refer to the comprehensive score calculated for each sample in the first training data by integrating multiple dimensions of the preset evaluation metrics.

[0029] Specifically, based on the pre-training process data and the first sample prediction data, each sample in the first sample training data can be evaluated in multiple dimensions according to preset evaluation indicators, and a comprehensive score can be given based on the evaluation results to determine the quality evaluation score corresponding to each sample in the first sample training data, thereby achieving a comprehensive quantitative evaluation of sample quality.

[0030] For example, S130 may include: determining multiple evaluation scores for each sample in the first sample training data based on preset evaluation metrics, pre-training process data, first sample training data, and first sample prediction data, wherein the preset evaluation metrics include: prediction confidence, label consistency, gradient influence, historical performance stability, and neighborhood consistency; and determining a quality evaluation score for each sample in the first sample training data based on the multiple evaluation scores for each sample in the first sample training data.

[0031] Among these, the evaluation score can refer to a quantitative score reflecting the performance of a sample on a pre-defined evaluation metric. Prediction confidence can refer to a quantitative value representing the degree of certainty shown by the model in its prediction of a single sample. Label consistency can refer to a measure of the consistency between the model's predicted category and the sample's true label. Gradient influence can refer to a measure of the magnitude or sensitivity of the gradient of a single training sample to the model's loss function, reflecting the sample's potential influence on the current model parameter updates. Historical performance stability can refer to a measure of the frequency or stability with which a sample is correctly predicted by the model over multiple training epochs. Neighborhood consistency can refer to the proportion of a sample that shares the same label as its K nearest neighbors in the feature space extracted by the model.

[0032] Specifically, based on the pre-training process data and the first sample prediction data, the prediction confidence, label consistency, gradient influence, historical performance stability, and neighborhood consistency of each sample in the first sample training data can be determined. This allows for the acquisition of the indicator evaluation scores of each sample in the first sample training data under different preset evaluation indicators. Furthermore, based on the multiple indicator evaluation scores corresponding to each sample in the first sample training data, a comprehensive evaluation is performed on each sample in the first sample training data to determine the quality evaluation score corresponding to each sample. This avoids evaluation bias that may be caused by a single indicator, improves the reliability of the evaluation results, and provides reliable technical support for the optimization of subsequent training samples.

[0033] For example, determining the quality assessment score corresponding to each sample in the first sample training data based on the evaluation scores of multiple indicators corresponding to each sample in the first sample training data includes: for each sample in the first sample training data, normalizing the evaluation scores of multiple indicators corresponding to the sample, and taking a weighted average of the normalized evaluation scores to obtain the quality assessment score corresponding to the sample.

[0034] Specifically, the evaluation scores of multiple indicators for each sample in the first training data can be normalized to eliminate instability caused by different units of measurement and numerical ranges. The normalized evaluation scores are then weighted and averaged. It should be noted that the sum of the weight coefficients of each evaluation score is 1. The weights can be preset according to the model training requirements. Finally, a quality evaluation score is generated for each sample, thus ensuring the accuracy of the evaluation score.

[0035] S140. Optimize the first sample training data based on the quality assessment threshold and quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

[0036] The quality assessment threshold can be a pre-set critical value used to determine whether the quality assessment score meets the model training requirements. The second training sample data can be sample data that is optimized from the first training sample data to better suit the current model training environment.

[0037] Specifically, the quality assessment threshold can be compared with the quality assessment score corresponding to each sample in the first sample training data. Samples in the first sample training data with quality assessment scores lower than the quality assessment threshold are removed to obtain the second sample training data. The model training and sample evaluation optimization for the next preset training cycle are then performed based on the second sample training data. This enables the model to continuously identify and avoid interference from low-quality data during the training process, thereby improving the efficiency and quality of model training.

[0038] In this embodiment, by acquiring the pre-training process data corresponding to the target training model obtained through pre-training, a data foundation can be provided for subsequent steps. Based on a preset training cycle, the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data are acquired, providing the latest and targeted input for subsequent quality assessment, ensuring the timeliness and relevance of the assessment. Based on preset assessment indicators, pre-training process data, first sample training data, and first sample prediction data, the quality assessment score corresponding to the first sample training data is determined, realizing a comprehensive quality assessment of the training samples and greatly improving the reliability of the assessment score. Based on the quality assessment threshold and quality assessment score, the first sample training data is optimized to obtain second sample training data, so that the target training model can continue to be trained based on the second sample training data in the next preset training cycle, thereby improving the training quality of the model. This invention achieves automatic and efficient optimization of model training samples by evaluating and dynamically optimizing training samples during model training, and continuing model training based on the optimized training samples. This enables the model to continuously identify and suppress interference from low-quality data during training, thereby improving its adaptability to complex and dynamically evolving model training processes. It can also address new sample quality issues exposed during training due to changes in model cognition, greatly improving the training efficiency and quality of the model.

[0039] Example 2 Figure 2 This is a flowchart of a dynamic optimization method for model training samples provided in Embodiment 2 of the present invention. Based on the above embodiments, this embodiment optimizes the step of "optimizing the first sample training data based on the quality assessment threshold and quality assessment score to obtain the second sample training data". Explanations of terms that are the same as or corresponding to those in the above embodiments are not repeated here.

[0040] See Figure 2 The alternative method for dynamic optimization of model training samples provided in this embodiment specifically includes the following steps: S210. Obtain the pre-training process data corresponding to the target training model obtained from pre-training.

[0041] S220. Based on a preset training period, obtain the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data.

[0042] S230. Based on the preset evaluation indicators, pre-training process data, first sample training data and first sample prediction data, determine the quality evaluation score corresponding to the first sample training data.

[0043] S240. Based on the quality assessment threshold and quality assessment score, determine the target adjustment weight corresponding to each sample in the first sample training data.

[0044] The target adjustment weight can be a coefficient used to control the magnitude of the influence of training samples on model parameter updates in the next training cycle.

[0045] Specifically, the quality assessment threshold can be directly compared with the quality assessment score corresponding to each sample in the first sample training data, and the target adjustment weight corresponding to each sample in the first sample training data can be determined based on the comparison result. This avoids the information loss and potential bias that may be caused by directly deleting the sample, and preserves the possibility of re-evaluating and utilizing the sample in the future.

[0046] For example, S240 may include: for each sample in the first sample training data, in response to the quality assessment score corresponding to the sample being greater than or equal to the quality assessment threshold, determining the target adjustment weight corresponding to the sample as a first adjustment weight; in response to the quality assessment score corresponding to the sample being less than the quality assessment threshold, determining the target adjustment weight corresponding to the sample as a second adjustment weight, wherein the first adjustment weight is greater than the second adjustment weight.

[0047] The first adjustment weight can refer to the preset fixed weight when the sample quality is acceptable. The second adjustment weight can refer to the preset fixed weight when the sample quality is unacceptable.

[0048] Specifically, for each sample in the first training data, if the quality assessment score of the sample is greater than or equal to the quality assessment threshold, then the target adjustment weight corresponding to the sample is determined as the first adjustment weight; if the quality assessment score of the sample is less than the quality assessment threshold, then the target adjustment weight corresponding to the sample is determined as the second adjustment weight. The first adjustment weight is greater than the second adjustment weight, thereby greatly eliminating the impact of low-quality samples on model training, while avoiding information loss and potential bias that may result from directly deleting samples. It should be noted that the first adjustment weight can be 1, while the second adjustment weight is a value close to 0, such as 0.01.

[0049] S250. Adjust the weights of the first sample training data based on the target adjustment weights to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

[0050] Specifically, the default weight of the sample in the first training data is 1. The weight can be adjusted for each sample in the first training data according to the target to obtain the second training data. The model training and sample evaluation optimization are then carried out in the next preset training cycle based on the second training data, so that the model can continuously identify and avoid interference from low-quality data during the training process, thereby improving the efficiency and quality of model training.

[0051] The technical solution of this embodiment determines the target adjustment weight corresponding to each sample in the first sample training data based on the quality assessment threshold and quality assessment score; and adjusts the weights of the first sample training data based on the target adjustment weights to obtain the second sample training data, thereby improving the quality of the training data. This invention achieves soft removal of training data by adjusting the weights of the first sample training data to obtain the second sample training data, thus avoiding the permanent loss of information and model bias that may be caused by brute-force deletion of samples, and preserving the possibility of future re-evaluation and utilization of the sample.

[0052] Example 3 Figure 3 This is a schematic diagram of a model training sample dynamic optimization device provided in Embodiment 3 of the present invention. Figure 3 As shown, the device includes: a first data acquisition module 310, a second data acquisition module 320, an evaluation score determination module 330, and a training data optimization module 340; The first data acquisition module 310 is used to acquire the pre-training process data corresponding to the pre-trained target training model. The second data acquisition module 320 is used to acquire the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data based on a preset training period. The evaluation score determination module 330 is used to determine the quality evaluation score corresponding to the first sample training data based on the preset evaluation index, the pre-training process data, the first sample training data and the first sample prediction data. The training data optimization module 340 is used to optimize the first sample training data based on the quality assessment threshold and the quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

[0053] In this embodiment, by acquiring the pre-training process data corresponding to the target training model obtained through pre-training, a data foundation can be provided for subsequent steps. Based on a preset training period, the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data are acquired, providing the latest and targeted input for subsequent quality assessment, ensuring the timeliness and relevance of the assessment. Based on preset assessment indicators, the pre-training process data, the first sample training data, and the first sample prediction data, a quality assessment score corresponding to the first sample training data is determined, achieving a comprehensive quality assessment of the training samples and greatly improving the reliability of the assessment score. Based on the quality assessment threshold and the quality assessment score, the first sample training data is optimized to obtain second sample training data, so that the target training model can continue to be trained based on the second sample training data in the next preset training period, thereby improving the training quality of the model. This invention achieves automatic and efficient optimization of model training samples by evaluating and dynamically optimizing training samples during model training, and continuing model training based on the optimized training samples. This enables the model to continuously identify and suppress interference from low-quality data during training, thereby improving its adaptability to complex and dynamically evolving model training processes. It can also address new sample quality issues exposed during training due to changes in model cognition, greatly improving the training efficiency and quality of the model.

[0054] Optionally, the first data acquisition module 310 is specifically used for: performing a preset number of pre-training rounds on the original training model based on the initialized training dataset and initialized model parameters to generate a target training model; extracting data from the target training model to obtain the pre-training process data corresponding to the target training model, wherein the pre-training process data includes: sample identification data, model output data, model loss data, sample feature data, and model parameter data.

[0055] Optionally, the second data acquisition module 320 is specifically used for: in response to the training period of the target training model reaching a preset training period, determining the first sample training data corresponding to the target training model in the current training period; and determining the first sample prediction data corresponding to the first sample training data based on the forward propagation result of the target training model in the current training period.

[0056] Optionally, the evaluation score determination module 330 includes: The first score determination unit is used to determine the evaluation scores of multiple indicators for each sample in the first sample training data based on the preset evaluation indicators, the pre-training process data, the first sample training data, and the first sample prediction data. The preset evaluation indicators include: prediction confidence, label consistency, gradient influence, historical performance stability, and neighborhood consistency. The second score determination unit is used to determine the quality assessment score corresponding to each of the samples in the first sample training data based on the multiple indicator evaluation scores corresponding to each of the samples in the first sample training data.

[0057] Optionally, the second score determination unit is specifically used to: for each sample in the first sample training data, normalize the multiple indicator evaluation scores corresponding to the sample, and perform a weighted average of the normalized multiple indicator evaluation scores to obtain the quality evaluation score corresponding to the sample.

[0058] Optionally, the training data optimization module 340 includes: The weight determination unit is used to determine the target adjustment weight corresponding to each sample in the first sample training data based on the quality assessment threshold and the quality assessment score. The weight adjustment unit is used to adjust the weights of the first sample training data based on the target adjustment weights to obtain the second sample training data.

[0059] Optionally, the weight determination unit is specifically configured to: for each sample in the first sample training data, in response to the quality assessment score corresponding to the sample being greater than or equal to the quality assessment threshold, determine the target adjustment weight corresponding to the sample as a first adjustment weight; in response to the quality assessment score corresponding to the sample being less than the quality assessment threshold, determine the target adjustment weight corresponding to the sample as a second adjustment weight, wherein the first adjustment weight is greater than the second adjustment weight.

[0060] The above-described device can execute the model training sample dynamic optimization method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of executing the model training sample dynamic optimization method.

[0061] Example 4 Figure 4This is a schematic diagram of the structure of an electronic device implementing the model training sample dynamic optimization method of this invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (e.g., helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0062] like Figure 4 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded into the RAM 13 from storage unit 18. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0063] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0064] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as dynamic optimization methods for model training samples.

[0065] In some embodiments, the model training sample dynamic optimization method can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the model training sample dynamic optimization method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the model training sample dynamic optimization method by any other suitable means (e.g., by means of firmware).

[0066] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication unit 19, or installed from storage unit 18, or installed from ROM 12. When the computer program is executed by processor 11, it performs the functions defined in the methods of the embodiments of the present invention.

[0067] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0068] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0069] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0070] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0071] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0072] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0073] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0074] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for dynamic optimization of model training samples, characterized in that, include: Obtain the pre-training process data corresponding to the target training model obtained from pre-training; Based on a preset training period, the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data are obtained. Based on the preset evaluation indicators, the pre-training process data, the first sample training data, and the first sample prediction data, the quality evaluation score corresponding to the first sample training data is determined. The first sample training data is optimized based on the quality assessment threshold and the quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

2. The method according to claim 1, characterized in that, The step of obtaining the pre-training process data corresponding to the target training model obtained through pre-training includes: Based on the initial training dataset and initial model parameters, the original training model is pre-trained for a preset number of rounds to generate the target training model; Data extraction is performed on the target training model to obtain the pre-training process data corresponding to the target training model. The pre-training process data includes: sample identification data, model output data, model loss data, sample feature data, and model parameter data.

3. The method according to claim 1, characterized in that, The step of acquiring the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data based on a preset training period includes: In response to the training period of the target training model reaching a preset training period, the first sample training data of the target training model in the current training period is determined; Based on the forward propagation results of the target training model in the current training cycle, the first sample prediction data corresponding to the first sample training data is determined.

4. The method according to claim 1, characterized in that, The step of determining the quality assessment score corresponding to the first sample training data based on preset evaluation indicators, the pre-training process data, the first sample training data, and the first sample prediction data includes: Based on preset evaluation metrics, the pre-training process data, the first sample training data, and the first sample prediction data, multiple evaluation scores are determined for each sample in the first sample training data. The preset evaluation metrics include: prediction confidence, label consistency, gradient influence, historical performance stability, and neighborhood consistency. Based on the evaluation scores of multiple metrics corresponding to each sample in the first sample training data, a quality evaluation score is determined for each sample in the first sample training data.

5. The method according to claim 4, characterized in that, The step of determining the quality assessment score corresponding to each sample in the first sample training data based on the multiple indicator evaluation scores corresponding to each sample in the first sample training data includes: For each sample in the first sample training data, the evaluation scores of multiple indicators corresponding to the sample are normalized, and the normalized evaluation scores of multiple indicators are weighted and averaged to obtain the quality evaluation score corresponding to the sample.

6. The method according to claim 1, characterized in that, The step of optimizing the first sample training data based on the quality assessment threshold and the quality assessment score to obtain the second sample training data includes: Based on the quality assessment threshold and the quality assessment score, the target adjustment weight corresponding to each sample in the first sample training data is determined; The weights of the first sample training data are adjusted based on the target to obtain the second sample training data.

7. The method according to claim 6, characterized in that, The step of determining the target adjustment weight corresponding to each sample in the first sample training data based on the quality assessment threshold and the quality assessment score includes: For each sample in the first sample training data, in response to the quality assessment score corresponding to the sample being greater than or equal to the quality assessment threshold, the target adjustment weight corresponding to the sample is determined as the first adjustment weight; In response to the sample's quality assessment score being less than the quality assessment threshold, the target adjustment weight corresponding to the sample is determined as a second adjustment weight, wherein the first adjustment weight is greater than the second adjustment weight.

8. A device for dynamic optimization of model training samples, characterized in that, include: The first data acquisition module is used to acquire the pre-training process data corresponding to the target training model obtained through pre-training. The second data acquisition module is used to acquire the first sample training data corresponding to the target training model and the first sample prediction data corresponding to the first sample training data based on a preset training period. The evaluation score determination module is used to determine the quality evaluation score corresponding to the first sample training data based on the preset evaluation index, the pre-training process data, the first sample training data, and the first sample prediction data. The training data optimization module is used to optimize the first sample training data based on the quality assessment threshold and the quality assessment score to obtain the second sample training data, so as to continue training the target training model in the next preset training cycle based on the second sample training data.

9. An electronic device, characterized in that, The electronic device includes: At least one processor; and A memory communicatively connected to the at least one processor; wherein, The memory stores a computer program that can be executed by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the model training sample dynamic optimization method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause a processor to execute the method for dynamic optimization of model training samples as described in any one of claims 1-7.

Images