Dataset generation method and apparatus, electronic device, medium, and program product

Abstract

Embodiments of the present disclosure disclose a data set generation method and device, electronic equipment, medium and program product. A specific embodiment of the method comprises the following determination steps based on a sample data set and a corresponding initial scaling factor set: generating a sample scaling data set according to the sample data set and the initial scaling factor set; generating a neural tangent kernel loss value and a prediction output loss value for the sample scaling data set; generating a target loss value according to the neural tangent kernel loss value and the prediction output loss value; in response to determining that the target loss value meets a convergence condition, determining the initial scaling factor set as an updated scaling factor set; and generating a target sample data set according to the sample data set and the updated scaling factor set. The embodiment is related to artificial intelligence and can accurately determine the target sample data set from the sample data set.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computer technology, and more specifically to dataset generation methods, apparatus, electronic devices, media, and program products. Background Technology

[0002] Currently, with the development of deep learning, a large amount of training datasets is needed for model training. However, the cost of collecting, storing, and analyzing data is very high, so selecting a core dataset has become an important aspect of deep learning. The typical approach to generating a core sample dataset is to select the target dataset from the sample datasets based on various evaluation metrics discovered empirically by experts.

[0003] However, the inventors discovered that when using the above method to generate the core sample dataset, the following technical problems often arise:

[0004] The selected target dataset may only achieve the model's optimal results at a local time, ignoring the model's needs for the dataset at different training stages. This makes it difficult to achieve the optimal results throughout the entire model training process, resulting in lower model accuracy and increased system storage load.

[0005] The information disclosed in this background section is only intended to enhance the understanding of the background of the inventive concept, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0006] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0007] Some embodiments of this disclosure provide dataset generation methods, apparatuses, electronic devices, media, and program products to address the technical problems mentioned in the background section above.

[0008] In a first aspect, some embodiments of this disclosure provide a dataset generation method, comprising: based on a sample dataset and a corresponding initial scaling factor set, performing the following determination steps: generating a sample scaling dataset according to the sample dataset and the initial scaling factor set; generating a neural tangent kernel loss value and a predicted output loss value for the sample scaling dataset; generating a target loss value according to the neural tangent kernel loss value and the predicted output loss value; in response to determining that the target loss value satisfies the convergence condition, determining the initial scaling factor set as an updated scaling factor set; and generating a target sample dataset according to the sample dataset and the updated scaling factor set.

[0009] Optionally, before generating the target sample dataset based on the sample dataset and the updated scaling factor set, the method further includes: updating the initial scaling factor set in response to determining that the target loss value does not meet the convergence condition, obtaining the updated scaling factor set, and determining the updated scaling factor set as the initial scaling factor set to perform the determination step again.

[0010] Optionally, the above-mentioned generation of the neural tangent kernel loss value and prediction output loss value for the sample scaled dataset includes: generating a first neural tangent kernel loss value for the sample dataset; generating a second neural tangent kernel loss value for the sample scaled dataset; determining the neural tangent kernel difference between the second neural tangent kernel loss value and the first neural tangent kernel loss value; performing norm processing on an initial scaling factor set to obtain a scaling value, wherein the initial scaling factor set is a scaling factor set corresponding to the sample dataset; and generating the neural tangent kernel loss value based on the neural tangent kernel difference and the scaling value.

[0011] Optionally, the above-mentioned generation of neural tangent kernel loss value and prediction output loss value for the sample scaled dataset further includes: generating a first prediction output value for the sample dataset; generating a second prediction output value for the sample scaled dataset; determining the prediction difference between the second prediction output value and the first prediction output value; generating a target prediction output loss value based on the prediction difference; and performing an equivalent transformation on the target prediction output loss value to obtain the prediction output loss value.

[0012] Optionally, generating a target loss value based on the neural tangent nucleus loss value and the predicted output loss value includes: performing a weighted summation on the neural tangent nucleus loss value and the predicted output loss value to obtain a weighted sum value as the target loss value.

[0013] Optionally, generating the sample scaling dataset based on the sample dataset and the initial scaling factor set includes: vectorizing each sample data in the sample dataset to generate a sample vector corresponding to the sample data, thus obtaining a sample vector set; and generating the sample scaling dataset based on the initial scaling factor set and the sample vector set.

[0014] Optionally, the above method further includes: based on the target sample dataset, performing the following training steps: inputting the target sample dataset into the initial model to obtain a prediction result set; determining the error ratio according to the prediction result set and the target sample label set, wherein the target sample label set is the label set corresponding to the target sample dataset; in response to determining that the error ratio is less than a preset error threshold, determining the initial model as a trained model; in response to determining that the error ratio is greater than or equal to the preset error threshold, determining that the initial model has not reached the preset optimization objective, and adjusting the relevant parameters of the initial model, using the adjusted initial model as the initial model, re-acquiring the sample dataset as the target sample dataset, and performing the above training steps again.

[0015] Optionally, generating the target sample dataset based on the above sample dataset and the updated scaling factor set includes: selecting sample data from the above sample dataset that are greater than or equal to a preset threshold in the corresponding updated scaling factor set, and using them as target sample data to obtain the target sample dataset, wherein the preset threshold is determined based on the model training acceleration ratio.

[0016] Secondly, some embodiments of this disclosure provide a dataset generation apparatus, including: an execution unit configured to perform the following determination steps based on a sample dataset and a corresponding initial scaling factor set: generating a sample scaling dataset according to the sample dataset and the initial scaling factor set; generating a neural tangent kernel loss value and a prediction output loss value for the sample scaling dataset; generating a target loss value according to the neural tangent kernel loss value and the prediction output loss value; and determining the initial scaling factor set as an updated scaling factor set in response to determining that the target loss value satisfies a convergence condition; and a generation unit configured to generate a target sample dataset according to the sample dataset and the updated scaling factor set.

[0017] Optionally, the dataset generation apparatus may further include an update unit. The update unit may be configured to: update the initial scaling factor set in response to determining that the target loss value does not meet the convergence condition, to obtain an updated scaling factor set, and to determine the updated scaling factor set as the initial scaling factor set, so as to perform the determination step described above again.

[0018] Optionally, the execution unit can be configured to: generate a first tangent kernel loss value for the aforementioned sample dataset; generate a second tangent kernel loss value for the aforementioned scaled sample dataset; determine the tangent kernel difference between the second tangent kernel loss value and the first tangent kernel loss value; perform norm processing on an initial scaling factor set to obtain a scaling value, wherein the initial scaling factor set is a scaling factor set corresponding to the aforementioned sample dataset; and generate a tangent kernel loss value based on the tangent kernel difference and the scaling value.

[0019] Optionally, the execution unit can be configured to: generate a first predicted output value for the above-mentioned sample dataset; generate a second predicted output value for the above-mentioned sample scaled dataset; determine the prediction difference between the second predicted output value and the first predicted output value; generate a target predicted output loss value based on the prediction difference; and perform an equivalent transformation on the target predicted output loss value to obtain a predicted output loss value.

[0020] Optionally, the execution unit can be configured to perform a weighted summation on the above-mentioned neural tangent kernel loss value and the above-mentioned prediction output loss value to obtain a weighted sum value as the target loss value.

[0021] Optionally, the execution unit can be configured to: perform vectorization processing on each sample data in the above sample dataset to generate a sample vector corresponding to the above sample data, thereby obtaining a sample vector set; and generate a sample scaling dataset based on the above initial scaling factor set and the above sample vector set.

[0022] Optionally, the dataset generation device may further include a training unit and a determination unit. The training unit may be configured to perform the following training steps based on the target sample dataset: inputting the target sample dataset into an initial model to obtain a prediction result set; determining an error ratio based on the prediction result set and the target sample label set, wherein the target sample label set is the label set corresponding to the target sample dataset; and determining the initial model as a trained model in response to the determination error ratio being less than a preset error threshold. The determination unit may be configured to determine that the initial model has not reached a preset optimization objective in response to the determination error ratio being greater than or equal to a preset error threshold, and to adjust the relevant parameters of the initial model, using the adjusted initial model as the initial model, re-acquiring the sample dataset as the target sample dataset, and then re-executing the above training steps.

[0023] Optionally, the generation unit can be configured to: select sample data from the above sample dataset that are greater than or equal to a preset threshold in the corresponding updated scaling factor set, and use them as target sample data to obtain the target sample dataset, wherein the above preset threshold is determined based on the model training acceleration ratio.

[0024] Thirdly, some embodiments of this disclosure provide an electronic device, including: one or more processors; and a storage device having one or more programs stored thereon, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the method described in any of the implementations of the first aspect above.

[0025] Fourthly, some embodiments of this disclosure provide a computer-readable medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0026] Fifthly, some embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the method described in any of the implementations of the first aspect above.

[0027] The above embodiments of this disclosure have the following beneficial effects: the dataset generation method of some embodiments of this disclosure can accurately determine the target sample dataset from the sample dataset. Specifically, the reason for the low accuracy of the related models and the increased system storage load is that the selected target dataset may only achieve the optimal result of the model at a local time, ignoring the model's need for the dataset at different training stages, making it difficult to achieve the optimal result throughout the entire model training process, resulting in low model accuracy and increased system storage load. Based on this, the dataset generation method of some embodiments of this disclosure firstly performs the following determination steps based on the sample dataset and the corresponding initial scaling factor set: Step 1, generate a sample scaling dataset according to the above sample dataset and initial scaling factor set. Here, the obtained sample scaling dataset can characterize the importance of the sample scaling dataset for model training, facilitating subsequent updates to the initial scaling factor set. Step 2, generate the neural tangent kernel loss value and the prediction output loss value for the sample scaling dataset. Here, determining the neural tangent kernel loss value and predicted output loss value in the sample dataset and the sample scaled dataset allows us to determine the errors in the predicted output and neural tangent kernel between the sample scaled dataset and the sample dataset. This not only improves the speed at which the subsequent model learns the features of the target sample dataset but also improves the model's accuracy on the target sample dataset. The third step is to generate the target loss value based on the neural tangent kernel loss value and the predicted output loss value. Here, the target loss value is described from both aspects, combining the neural tangent kernel loss value and the predicted output loss value, which facilitates subsequent model training. The fourth step, in response to the determination that the target loss value satisfies the convergence condition, is to determine the initial scaling factor set as the updated scaling factor set. Here, when the target loss value satisfies the convergence condition, the determined updated scaling factor set can characterize the importance of the sample dataset for subsequent model training, so that the subsequently obtained target sample dataset is more accurate. Then, based on the above sample dataset and updated scaling factor set, the target sample dataset is generated. Here, the amount of data in the target dataset can be controlled by setting a preset threshold, which reduces the amount of training data needed for subsequent model retraining, thereby shortening the training time of subsequent models. Thus, by scaling the sample dataset corresponding to the initial scaling factor set, the error between the sample dataset and the sample dataset in the neural tangent kernel loss value and the predicted output loss value is determined, thereby identifying the target dataset. This improves the accuracy of determining the target dataset and increases the training speed of the model by training it with a smaller dataset. Attached Figure Description

[0028] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0029] Figure 1 This is a schematic diagram illustrating an application scenario of a dataset generation method according to some embodiments of the present disclosure;

[0030] Figure 2 This is a flowchart of some embodiments of the dataset generation method according to this disclosure;

[0031] Figure 3 These are flowcharts of other embodiments of the dataset generation method according to this disclosure;

[0032] Figure 4 This is a schematic diagram of an application scenario of some other embodiments of the dataset generation method according to this disclosure;

[0033] Figure 5 This is a schematic diagram of the structure of some embodiments of the dataset generation apparatus according to the present disclosure;

[0034] Figure 6 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure. Detailed Implementation

[0035] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0036] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0037] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0038] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0039] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0040] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] Figure 1 This is a schematic diagram illustrating an application scenario of a dataset generation method according to some embodiments of the present disclosure.

[0042] exist Figure 1 In the application scenario, electronic device 101 can first perform the following determination steps based on the sample dataset and the corresponding initial scaling factor set: First, generate a sample scaling dataset 104 based on the sample dataset 102 and the initial scaling factor set 103. Second, generate a neural tangent kernel loss value 105 and a prediction output loss value 106 for the sample scaling dataset 104. Third, generate a target loss value 107 based on the neural tangent kernel loss value 105 and the prediction output loss value 106. Fourth, in response to determining that the target loss value 107 satisfies the convergence condition, determine the initial scaling factor set 103 as the updated scaling factor set 108. Then, generate a target sample dataset 109 based on the sample dataset 102 and the updated scaling factor set 108.

[0043] It should be noted that the aforementioned electronic device 101 can be either hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is software, it can be installed in the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0044] It should be understood that Figure 1 The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0045] Continue to refer to Figure 2 The diagram illustrates a flow 200 of some embodiments of a dataset generation method according to the present disclosure. The dataset generation method includes the following steps:

[0046] Step 201: Based on the sample dataset and the corresponding initial scaling factor set, perform the following determination steps:

[0047] Step 2011: Generate a sample scaling dataset based on the sample dataset and the initial scaling factor set.

[0048] In some embodiments, the execution entity of the above-described dataset generation method (e.g.) Figure 1 The electronic device 101 shown can generate a sample-scaled dataset based on the aforementioned sample dataset and initial scaling factor set. The sample data in the aforementioned sample dataset can be data used to train the model. For example, the type of the aforementioned sample dataset can be at least one of the following: text, audio, video, or image. The initial scaling factor in the aforementioned initial scaling factor set can characterize the importance of the corresponding sample data to model training. The initial scaling factor in the aforementioned initial scaling factor set can be randomly generated values ​​within the range of 0 and 1. The closer the initial scaling factor is to 0, the smaller the impact of the corresponding sample data on model training. The closer the initial scaling factor is to 1, the greater the impact of the corresponding sample data on model training. The sample-scaled data in the aforementioned sample-scaled dataset can be data that assigns different weights to the corresponding sample data. The higher the weight, the greater the impact of the corresponding sample data on model training; the lower the weight, the smaller the impact of the corresponding sample data on model training.

[0049] As an example, the aforementioned execution entity can first extract features from each sample data in the aforementioned sample dataset to generate a feature tensor corresponding to the sample data, thus obtaining a feature tensor set. Secondly, it can vectorize each scaling factor in the aforementioned scaling factor set to generate a scaling factor vector, thus obtaining a scaling factor vector set. Finally, it can superimpose the aforementioned feature tensor set and the corresponding scaling factor vector set to obtain the sample scaled dataset.

[0050] In some optional implementations of certain embodiments, generating a sample scaling dataset based on the sample dataset and the initial scaling factor set includes:

[0051] The first step is to vectorize each sample data in the above sample dataset to generate a corresponding sample vector, thus obtaining a sample vector set. Here, the sample vector can be a vector representing the semantic information of the sample data.

[0052] As an example, the aforementioned execution entity can input each sample data in the sample dataset into an encoding network to generate a sample vector corresponding to the sample data, thus obtaining a sample vector set. The encoding network can be a CNN (Convolutional Neural Networks) model.

[0053] The second step is to generate a sample scaling dataset based on the initial scaling factor set and the sample vector set mentioned above.

[0054] As an example, the aforementioned execution entity can determine the sample scaling data by multiplying the initial scaling factor set and the corresponding sample vector.

[0055] Step 2012: Generate the neural tangent kernel loss value and the prediction output loss value for the sample scaled dataset.

[0056] In some embodiments, the execution entity can generate a neural tangent kernel loss value and a prediction output loss value for a sample scaled dataset. The neural tangent kernel loss value characterizes the inner product of the gradients of the function corresponding to the neural network model with respect to the network model parameters. The prediction output loss value characterizes the error between the model's predicted output value and the true value.

[0057] As an example, the aforementioned execution entity can first input the aforementioned scaled dataset into an initial model to obtain predicted output values ​​for the scaled dataset. Then, it determines the loss value between the predicted output values ​​and the corresponding true values ​​in the scaled dataset. The initial model can be an image model, a video model, or an audio model. For example, an image model could be an image classification model. A video model could be an object recognition model. An audio model could be a speech denoising model. Finally, using a neural tangent kernel function, the neural tangent kernel loss value for the aforementioned scaled dataset is determined.

[0058] In some optional implementations of certain embodiments, the generation of the neural tangent kernel loss value and the prediction output loss value for the sample scaled dataset may include the following steps:

[0059] The first step is to generate the first neural tangent kernel loss value for the aforementioned sample dataset. This first neural tangent kernel loss value characterizes the loss value of the inner product of the neural network model's function with respect to the model parameters for the sample dataset. The aforementioned neural network model can be any deep learning model. For example, it could be an image classification model or a video object recognition model.

[0060] As an example, the aforementioned execution entity can input the aforementioned sample dataset into the first neural tangent kernel loss function to obtain the first neural tangent kernel loss value. The aforementioned first neural tangent kernel loss function can be:

[0061]

[0062] Where Θ0(X, X) represents the first tangent kernel loss function with respect to the sample dataset at the initial time, where the initial time can be the time when the function corresponding to the neural network model is initialized. X represents the sample dataset. Let f0(X) represent the derivative of the function corresponding to the neural network model with respect to the model parameters θ at the initial time step. T This represents the transpose of the output of the neural network model with respect to the sample dataset at the initial time.

[0063] At the initial moment, the function corresponding to the neural network model can be:

[0064]

[0065] Where f0(x) represents the function corresponding to the neural network model at the initial time. x represents the current input data. g(X) represents the sample scaling data, which can be represented as g(X) = m⊙X. ⊙ represents the Hadamard product. m represents the initial scaling factor set. I represents the identity matrix. η represents the training learning rate. t represents the time parameter. Θ0 represents the tangent kernel matrix of the function corresponding to the neural network model at the initial time. Y represents the sample label corresponding to the sample data. Let Θ0(x, g(X)) represent the inverse function value of the neural network model at the initial time step. Let Θ0(x, g(X)) represent the tangent kernel value of the neural network model corresponding to the current data and the scaled sample data at the initial time step. Let f0(g(X)) represent the function value of the neural network model corresponding to the scaled sample data at the initial time step.

[0066] The function corresponding to the time parameter in the neural network model can be:

[0067] The second step is to generate a second neural tangent kernel loss value for the aforementioned scaled dataset. This second neural tangent kernel loss value characterizes the loss value of the function corresponding to the neural network model with respect to the inner product of the model parameters for the scaled dataset.

[0068] As an example, the aforementioned execution entity can input the aforementioned scaled sample dataset into the second neural tangent kernel loss function to obtain the second neural tangent kernel loss value. The aforementioned second neural tangent kernel function can be:

[0069]

[0070] Where Θ0(g(X), g(X)) represents the second tangent kernel loss function at the initial time with respect to the second tangent kernel loss value of the sample scaled dataset. f0(g(X)) represents the derivative of the function corresponding to the neural network model with respect to the scaled dataset at the initial time step, obtained by taking the derivative of the model parameters. T This represents the transpose of the function result of the neural network model with respect to the sample-scaled dataset at the initial time.

[0071] The third step is to determine the difference between the loss value of the second tangential nucleus and the loss value of the first tangential nucleus.

[0072] The fourth step is to perform norm processing on the initial scaling factor set to obtain the scaling values. Here, the initial scaling factor set is the scaling factor set corresponding to the aforementioned sample dataset. The norm processing can be first norm processing.

[0073] Fifth step: Based on the above difference in the tangential nucleus and the above scaling value, generate the tangential nucleus loss value.

[0074] As an example, the aforementioned execution entity can input the aforementioned neural tangent kernel difference and the aforementioned scaling value into the neural tangent kernel function to obtain the neural tangent kernel loss value. The neural tangent kernel loss function can be expressed as:

[0075]

[0076] in, Let represent the neural tangent kernel loss function. ||||2 represents the second norm value. ||m||1 represents the first norm value of the initial scaling factor set.

[0077] Optionally, the above-mentioned generation of neural tangent kernel loss values ​​and prediction output loss values ​​for a sample scaled dataset may further include the following steps:

[0078] The first step is to generate the first predicted output value for the above sample dataset. This first predicted output value represents the output value of the function corresponding to the neural network model for the sample dataset.

[0079] As an example, the aforementioned execution entity can input the aforementioned sample dataset into the first prediction output function to obtain the first prediction output value. The first prediction output function can be expressed as:

[0080]

[0081] Where f(x|X) represents the first prediction output function.

[0082] The second step is to generate a second predicted output value for the aforementioned scaled dataset. This second predicted output value can represent the output value of the function corresponding to the neural network model for the scaled dataset.

[0083] As an example, the aforementioned execution entity can input the aforementioned scaled sample dataset into the second prediction output function to obtain the second prediction output value. The aforementioned second prediction output function can be expressed as:

[0084]

[0085] Here, f(x|g(X)) represents the second prediction output function.

[0086] The third step is to determine the prediction difference between the second prediction output value and the first prediction output value.

[0087] The fourth step is to generate the target prediction output loss value based on the above prediction difference.

[0088] As an example, the aforementioned execution entity can input the prediction difference into the target prediction output loss function to obtain the target prediction output loss value. The target prediction output loss function can be:

[0089]

[0090] in, Let represent the target prediction output loss function. E represents the expectation of the first and second prediction output loss values. Let p(x) represent the square of the second norm. Let p(x) represent the data distribution of sample point x.

[0091] The fifth step involves performing an equivalent transformation on the target prediction output loss value to obtain the prediction output loss value. This equivalent transformation can be achieved by converting the target prediction output loss value into the form of a neural tangent kernel function. Since the prediction output function can be represented by a neural tangent kernel function, the target prediction output loss function is represented by the neural tangent kernel function to obtain the prediction output loss function. The prediction output function represented by the neural tangent kernel function can be expressed as follows:

[0092]

[0093] Here, f(x) represents the predicted output function.

[0094] As an example, the aforementioned execution entity can input the target prediction output loss value into the prediction output loss function to obtain the prediction output loss value. The prediction output loss function can be expressed as:

[0095]

[0096] Step 2013: Generate the target loss value based on the neural tangential nucleus loss value and the predicted output loss value.

[0097] In some embodiments, the execution entity can generate a target loss value based on the neural tangent kernel loss value and the predicted output loss value. This target loss value can characterize the error between the fused predicted output value and the true value.

[0098] As an example, the aforementioned execution entity may, in response to determining the optimization result of the subsequent model with respect to the model gradient, determine the neural tangent kernel loss value as the target loss value. In response to determining the predicted output result of the subsequent model, the predicted output loss value may be determined as the target loss value.

[0099] In some optional implementations of certain embodiments, generating the target loss value based on the neural tangent kernel loss value and the predicted output loss value includes:

[0100] The aforementioned neural tangent kernel loss value and the aforementioned predicted output loss value are weighted and summed to obtain a weighted sum value, which is used as the target loss value. The target loss function corresponding to the aforementioned target loss value can be expressed as:

[0101]

[0102] in, Let λ represent the target loss function, and λ represent the weight values ​​corresponding to the predicted output loss function. The weight values ​​can be determined based on the specific sample dataset.

[0103] Step 2014: In response to determining that the target loss value satisfies the convergence condition, the initial scaling factor set is determined as the updated scaling factor set.

[0104] In some embodiments, the execution entity may determine the initial scaling factor set as the updated scaling factor set in response to determining that the target loss value satisfies the convergence condition. The convergence condition may be that the target loss value obtained by inputting the scaled sample dataset into the target loss function tends to a constant loss value. The updated scaling factor set may be a scaling factor set obtained by updating the initial scaling factor set.

[0105] Step 202: Generate the target sample dataset based on the sample dataset and the updated scaling factor set.

[0106] In some embodiments, the execution entity may generate a target sample dataset based on the sample dataset and the updated scaling factor set. The target sample dataset may represent the key features that the model can learn, be consistent with the key features that the model can learn on the sample dataset, and have approximately the same training effect on the model.

[0107] As an example, the aforementioned execution entity can first filter out updated scaling factors from the set of updated scaling factors that are greater than or equal to a pre-set ratio threshold, thus obtaining a filtered set of scaling factors. The pre-set ratio threshold can be 0.7. Then, the sample dataset corresponding to the filtered set of scaling factors is determined as the target sample dataset.

[0108] In some optional implementations of certain embodiments, generating the target sample dataset based on the sample dataset and the updated scaling factor set may include the following steps:

[0109] From the aforementioned sample dataset, sample data with updated scaling factors greater than or equal to a preset threshold are selected as target sample data, thus obtaining the target sample dataset. The preset threshold is determined based on the model training acceleration ratio. This model training acceleration ratio can be a percentage of the target sample dataset relative to the total sample dataset. For example, the model training acceleration ratio could be 0.6.

[0110] As an example, the aforementioned execution entity can first determine the model training acceleration ratio. Then, using the aforementioned model training acceleration ratio, it can determine a subset of initial scaling factors that fall within the range of the aforementioned model training acceleration ratio. Finally, using the subset of initial scaling factors, it can determine a preset threshold.

[0111] Optionally, the aforementioned implementing entity may also perform the following steps:

[0112] The first step, based on the target sample dataset mentioned above, is to perform the following training steps:

[0113] The first sub-step involves inputting the target sample dataset into the initial model to obtain the prediction result set. The initial model can be, but is not limited to, one of the following: an image-based model, an audio-based model, a video-based model, or a text-based model. For example, the image-based model could be an image classification model. The audio-based model could be an audio denoising model. The video-based model could be a pedestrian detection model. The text-based model could be a sentiment analysis model. The prediction result set can be the output set of the initial model.

[0114] The second sub-step involves determining the error ratio based on the aforementioned prediction result set and target sample label set. Here, the target sample labels in the target sample label set can be the label set corresponding to the aforementioned target sample data. For example, if the target sample data in the target sample dataset could be an image whose main subject is a cat, then the label corresponding to the target sample data could be "cat."

[0115] As an example, the aforementioned execution entity can first determine the number of errors in the prediction result set where the prediction result is different from the corresponding target sample label. Then, it can determine the ratio of the number of errors to the prediction result set, which is used as the error ratio.

[0116] The third sub-step involves determining the initial model as a completed training model in response to the error ratio being less than a preset error threshold. The preset error threshold can be a pre-determined error value. For example, the prediction error threshold could be 0.75.

[0117] The second step involves determining that the error ratio is greater than or equal to a preset error threshold, identifying that the initial model has not reached the preset optimization objective, adjusting the relevant parameters of the initial model, using the adjusted initial model as the initial model, and re-acquiring the sample dataset as the target sample dataset to execute the above training steps again.

[0118] The above embodiments of this disclosure have the following beneficial effects: the dataset generation method of some embodiments of this disclosure can accurately determine the target sample dataset from the sample dataset. Specifically, the reason for the low accuracy of the related models and the increased system storage load is that the selected target dataset may only achieve the optimal result of the model at a local time, ignoring the model's need for the dataset at different training stages, making it difficult to achieve the optimal result throughout the entire model training process, resulting in low model accuracy and increased system storage load. Based on this, the dataset generation method of some embodiments of this disclosure firstly performs the following determination steps based on the sample dataset and the corresponding initial scaling factor set: Step 1, generate a sample scaling dataset according to the above sample dataset and initial scaling factor set. Here, the obtained sample scaling dataset can characterize the importance of the sample scaling dataset for model training, facilitating subsequent updates to the initial scaling factor set. Step 2, generate the neural tangent kernel loss value and the prediction output loss value for the sample scaling dataset. Here, determining the neural tangent kernel loss value and predicted output loss value in the sample dataset and the sample scaled dataset allows us to determine the errors in the predicted output and neural tangent kernel between the sample scaled dataset and the sample dataset. This not only improves the speed at which the subsequent model learns the features of the target sample dataset but also improves the model's accuracy on the target sample dataset. The third step is to generate the target loss value based on the neural tangent kernel loss value and the predicted output loss value. Here, the target loss value is described from both aspects, combining the neural tangent kernel loss value and the predicted output loss value, which facilitates subsequent model training. The fourth step, in response to the determination that the target loss value satisfies the convergence condition, is to determine the initial scaling factor set as the updated scaling factor set. Here, when the target loss value satisfies the convergence condition, the determined updated scaling factor set can characterize the importance of the sample dataset for subsequent model training, so that the subsequently obtained target sample dataset is more accurate. Then, based on the above sample dataset and updated scaling factor set, the target sample dataset is generated. Here, the amount of data in the target dataset can be controlled by setting a preset threshold, which reduces the amount of training data needed for subsequent model retraining, thereby shortening the training time of subsequent models. Thus, by scaling the sample dataset corresponding to the initial scaling factor set, the error between the sample dataset and the sample dataset in the neural tangent kernel loss value and the predicted output loss value is determined, thereby identifying the target dataset. This improves the accuracy of determining the target dataset and increases the training speed of the model by training it with a smaller dataset.

[0119] Further reference Figure 3 The diagram illustrates a flow 300 of another embodiment of the dataset generation method according to this disclosure. The dataset generation method includes the following steps:

[0120] Step 301: Based on the sample dataset and the corresponding initial scaling factor set, perform the following determination steps:

[0121] Step 3011: Generate a sample scaling dataset based on the sample dataset and the initial scaling factor set.

[0122] Step 3012: Generate the neural tangent kernel loss value and the prediction output loss value for the sample scaled dataset.

[0123] Step 3013: Generate the target loss value based on the neural tangent nucleus loss value and the predicted output loss value.

[0124] Step 3014: In response to determining that the target loss value satisfies the convergence condition, the initial scaling factor set is determined as the updated scaling factor set.

[0125] In some embodiments, the specific implementation of steps 3011-3014 and the resulting technical effects can be found in [reference needed]. Figure 2 Steps 2011-2014 in the corresponding embodiments will not be repeated here.

[0126] Step 302: In response to the determination that the target loss value does not meet the convergence condition, the initial scaling factor set is updated to obtain the updated scaling factor set, and the updated scaling factor set is determined as the initial scaling factor set to perform the determination step again.

[0127] In some embodiments, the aforementioned execution entity (e.g., Figure 4 The illustrated electronic device 401 can, in response to determining that the target loss value does not meet the convergence condition, update the initial scaling factor set to obtain an updated scaling factor set, and determine the updated scaling factor set as the initial scaling factor set to perform the determination step again. The updated scaling factor set can be the scaling factor set obtained by updating the initial scaling factor set. The update can be the target loss function corresponding to the target loss value. The initial scaling factor parameters are optimized by gradient update to obtain the updated scaling factor.

[0128] Step 303: Generate the target sample dataset based on the sample dataset and the updated scaling factor set.

[0129] In some embodiments, the specific implementation of step 303 and its resulting technical effects can be found in [reference needed]. Figure 2 Step 203 in the corresponding embodiment will not be repeated here.

[0130] from Figure 3 It can be seen from this that, with Figure 2 Compared to the description of some corresponding embodiments, Figure 3The flowchart 300 of the dataset generation method in some corresponding embodiments further emphasizes the specific steps where the target loss value does not meet the convergence condition. Therefore, the scheme described in these embodiments, by continuously updating the initial scaling factor that does not meet the convergence condition until it does, and then selecting the target sample dataset from the sample dataset corresponding to the updated scaling factor set, can make the determined target sample dataset more comprehensive and accurate.

[0131] Figure 4 This is a schematic diagram of an application scenario of another embodiment of the dataset generation method according to some embodiments of the present disclosure.

[0132] exist Figure 4 In the application scenario described, electronic device 401 can first perform the following determination steps based on the sample dataset and the corresponding initial scaling factor set: First, generate a sample scaling dataset 404 based on the sample dataset 402 and the initial scaling factor set 403. Second, generate a neural tangent kernel loss value 405 and a prediction output loss value 406 for the sample scaling dataset 404. Third, generate a target loss value 407 based on the neural tangent kernel loss value 405 and the prediction output loss value 406. Fourth, in response to determining that the target loss value 407 satisfies the convergence condition, determine the initial scaling factor set 403 as the updated scaling factor set 408. Then, in response to determining that the target loss value 407 does not satisfy the convergence condition, update the initial scaling factor set 403 to obtain the updated scaling factor set 408, and determine the updated scaling factor set 408 as the initial scaling factor set 403, to perform the above determination steps again. Finally, generate the target sample dataset based on the sample dataset 402 and the updated scaling factor set 408. In this application scenario, the target sample dataset mentioned above can be target sample dataset 409.

[0133] It should be noted that the aforementioned electronic device 401 can be either hardware or software. When the electronic device is hardware, it can be implemented as a distributed cluster consisting of multiple servers or terminal devices, or as a single server or a single terminal device. When the electronic device is software, it can be installed in the hardware devices listed above. It can be implemented as, for example, multiple software programs or software modules used to provide distributed services, or as a single software program or software module. No specific limitations are made here.

[0134] It should be understood that Figure 4 The number of electronic devices shown is merely illustrative. Any number of electronic devices can be used depending on the implementation requirements.

[0135] Further reference Figure 5As an implementation of the methods shown in the above figures, this disclosure provides some embodiments of a dataset generation apparatus, which are similar to... Figure 2 Corresponding to the method embodiments shown, the device can be specifically applied to various electronic devices.

[0136] like Figure 5 As shown, a dataset generation apparatus 500 includes an execution unit 501 and a generation unit 502. The execution unit 501 is configured to perform the following determination steps based on a sample dataset and a corresponding initial scaling factor set: generating a sample scaling dataset based on the sample dataset and the initial scaling factor set; generating a neural tangent kernel loss value and a prediction output loss value for the sample scaling dataset; generating a target loss value based on the neural tangent kernel loss value and the prediction output loss value; and determining the initial scaling factor set as an updated scaling factor set in response to determining that the target loss value satisfies the convergence condition. The generation unit 502 is configured to generate a target sample dataset based on the sample dataset and the updated scaling factor set.

[0137] In some optional implementations of certain embodiments, the dataset generation apparatus 500 further includes an update unit (not shown in the figure). The update unit may be further configured to: update the initial scaling factor set in response to determining that the target loss value does not meet the convergence condition, obtain an updated scaling factor set, and determine the updated scaling factor set as the initial scaling factor set to perform the determination step again.

[0138] In some optional implementations of certain embodiments, execution unit 501 may be further configured to: generate a first tangent kernel loss value for the aforementioned sample dataset; generate a second tangent kernel loss value for the aforementioned scaled sample dataset; determine the tangent kernel difference between the second tangent kernel loss value and the first tangent kernel loss value; perform norm processing on an initial scaling factor set to obtain a scaling value, wherein the initial scaling factor set is a scaling factor set corresponding to the aforementioned sample dataset; and generate a tangent kernel loss value based on the tangent kernel difference and the scaling value.

[0139] In some optional implementations of certain embodiments, the execution unit 501 may be further configured to: generate a first prediction output value for the sample dataset; generate a second prediction output value for the sample scaled dataset; determine the prediction difference between the second prediction output value and the first prediction output value; generate a target prediction output loss value based on the prediction difference; and perform an equivalent transformation on the target prediction output loss value to obtain the prediction output loss value.

[0140] In some optional implementations of some embodiments, the execution unit 501 may be further configured to: perform a weighted summation on the above-mentioned neural tangent kernel loss value and the above-mentioned prediction output loss value to obtain a weighted sum value as the target loss value.

[0141] In some optional implementations of some embodiments, the execution unit 501 may be further configured to: perform vectorization processing on each sample data in the above sample dataset to generate a sample vector corresponding to the above sample data, thereby obtaining a sample vector set; and generate a sample scaled dataset based on the above initial scaling factor set and the above sample vector set.

[0142] In some optional implementations of certain embodiments, the dataset generation apparatus 500 further includes a training unit and a determination unit (not shown in the figure). The training unit can be configured to perform the following training steps based on the target sample dataset: inputting the target sample dataset into an initial model to obtain a prediction result set; determining an error ratio based on the prediction result set and the target sample label set, wherein the target sample label set is the label set corresponding to the target sample dataset; and determining the initial model as a trained model in response to the determination that the error ratio is less than a preset error threshold. The determination unit can be configured to: in response to the determination that the error ratio is greater than or equal to a preset error threshold, determining that the initial model has not reached a preset optimization objective, using the adjusted initial model as the initial model, re-acquiring the sample dataset as the target sample dataset, and adjusting the relevant parameters of the initial model to perform the above training steps again.

[0143] In some optional implementations of some embodiments, the generation unit 502 may be further configured to: select sample data from the above sample dataset that are greater than or equal to a preset threshold in the corresponding updated scaling factor set, and use them as target sample data to obtain a target sample dataset, wherein the preset threshold is determined based on the model training acceleration ratio.

[0144] It is understandable that the units described in the device 500 are related to the reference. Figure 2 The steps in the described method correspond accordingly. Therefore, the operations, features, and beneficial effects described above for the method also apply to the device 500 and the units contained therein, and will not be repeated here.

[0145] The following is for reference. Figure 6 It illustrates electronic devices suitable for implementing some embodiments of this disclosure (e.g., Figure 1 A schematic diagram of the structure of electronic device 101)600 in the middle. Figure 6The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of this disclosure.

[0146] like Figure 6 As shown, electronic device 600 may include a processing device (e.g., a central processing unit, a graphics processor, etc.) 601, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 602 or a program loaded from storage device 608 into random access memory (RAM) 603. RAM 603 also stores various programs and data required for the operation of electronic device 600. Processing device 601, ROM 602, and RAM 603 are interconnected via bus 604. Input / output (I / O) interface 605 is also connected to bus 604.

[0147] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, magnetic tapes, hard disks, etc.; and communication devices 609. Communication device 609 allows electronic device 600 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 An electronic device 600 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 6 Each box shown can represent a device or multiple devices as needed.

[0148] In particular, according to some embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, some embodiments of this disclosure include a computer program product comprising a computer program carried on a 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 a communication device 609, or installed from a storage device 608, or installed from a ROM 602. When the computer program is executed by the processing device 601, it performs the functions defined above in the methods of some embodiments of this disclosure.

[0149] It should be noted that, in some embodiments of this disclosure, the computer-readable medium described above may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In some embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In some embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0150] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0151] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device. The aforementioned computer-readable medium carries one or more programs that, when executed by the electronic device, cause the electronic device to: based on the sample dataset and the corresponding initial scaling factor set, perform the following determining steps: generate a sample scaling dataset according to the aforementioned sample dataset and the initial scaling factor set; generate a neural tangent kernel loss value and a predicted output loss value for the sample scaling dataset; generate a target loss value based on the neural tangent kernel loss value and the predicted output loss value; in response to determining that the target loss value satisfies the convergence condition, determine the initial scaling factor set as the updated scaling factor set; and generate a target sample dataset according to the aforementioned sample dataset and the updated scaling factor set.

[0152] Computer program code for performing operations of some embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0153] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0154] The units described in some embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, a processor may be described as including an execution unit and a generation unit. The names of these units do not necessarily limit the unit itself. For example, an execution unit may be described as "a unit that, based on a sample dataset and a corresponding initial scaling factor set, performs the following determining steps: generating a sample scaling dataset according to the sample dataset and the initial scaling factor set; generating a neural tangent kernel loss value and a predicted output loss value for the sample scaling dataset; generating a target loss value according to the neural tangent kernel loss value and the predicted output loss value; and, in response to determining that the target loss value satisfies the convergence condition, determining the initial scaling factor set as the updated scaling factor set."

[0155] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0156] Some embodiments of this disclosure also provide a computer program product, including a computer program that, when executed by a processor, implements any of the dataset generation methods described above.

[0157] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for generating a dataset, comprising: Based on the sample dataset and the corresponding initial scaling factor set, the following determination steps are performed: Based on the sample dataset and the initial scaling factor set, a sample scaling dataset is generated, wherein the sample dataset is of at least one of the following types: text, audio, video, or image; Generating a neural tangent kernel loss value and a predicted output loss value for a sample-scaled dataset includes: inputting the sample-scaled dataset into an initial model to obtain a predicted output value for the sample-scaled dataset; determining the loss value between the predicted output value and the true value corresponding to the sample-scaled dataset; and using a neural tangent kernel function to determine the neural tangent kernel loss value for the sample-scaled dataset; wherein the neural tangent kernel loss value represents the inner product value between the gradients of the function corresponding to the neural network model with respect to the network model parameters. The target loss value is generated based on the neural tangent nucleus loss value and the predicted output loss value. In response to the determination that the target loss value satisfies the convergence condition, the initial scaling factor set is determined as the updated scaling factor set; The target sample dataset is generated based on the sample dataset and the updated scaling factor set.

2. The method according to claim 1, wherein, Before generating the target sample dataset based on the sample dataset and the updated scaling factor set, the method further includes: In response to the determination that the target loss value does not meet the convergence condition, the initial scaling factor set is updated to obtain the updated scaling factor set, and the updated scaling factor set is determined as the initial scaling factor set to perform the determination step again.

3. The method according to claim 1, wherein, The generation of neural tangent kernel loss values ​​and prediction output loss values ​​for the sample scaled dataset includes: Generate the first neural tangent kernel loss value for the sample dataset; Generate a second neural tangent kernel loss value for the scaled dataset of the samples; Determine the difference between the second tangential nucleus loss value and the first tangential nucleus loss value; The initial scaling factor set is subjected to norm processing to obtain scaling values, wherein the initial scaling factor set is the scaling factor set corresponding to the sample dataset; The neural tangential nucleus loss value is generated based on the difference between the neural tangential nuclei and the scaling value.

4. The method according to claim 3, wherein, The generation of the neural tangent kernel loss value and prediction output loss value for the sample scaled dataset also includes: Generate a first predicted output value for the sample dataset; Generate a second predicted output value for the scaled sample dataset; Determine the prediction difference between the second predicted output value and the first predicted output value; Based on the predicted difference, generate the target predicted output loss value; The target prediction output loss value is transformed into an equivalent value to obtain the prediction output loss value.

5. The method according to claim 1, wherein, The process of generating a target loss value based on the neural tangent nucleus loss value and the predicted output loss value includes: The loss value of the neural tangent nucleus and the loss value of the predicted output are weighted and summed to obtain a weighted sum value, which is used as the target loss value.

6. The method according to claim 1, wherein, The step of generating a sample scaling dataset based on the sample dataset and the initial scaling factor set includes: Each sample data in the sample dataset is vectorized to generate a sample vector corresponding to the sample data, thus obtaining a sample vector set; A sample scaling dataset is generated based on the initial scaling factor set and the sample vector set.

7. The method according to claim 1, wherein, The method further includes: Based on the target sample dataset, perform the following training steps: Input the target sample dataset into the initial model to obtain the prediction result set; The error ratio is determined based on the prediction result set and the target sample label set, wherein the target sample label set is the label set corresponding to the target sample dataset; In response to the determination that the error ratio is less than a preset error threshold, the initial model is determined as the trained model; In response to determining that the error ratio is greater than or equal to a preset error threshold, the initial model is determined to have failed to achieve the preset optimization objective. The relevant parameters of the initial model are adjusted, and the adjusted initial model is used as the initial model. A new sample dataset is then acquired as the target sample dataset to perform the training steps again.

8. The method according to claim 1, wherein, The step of generating the target sample dataset based on the sample dataset and the updated scaling factor set includes: From the sample dataset, sample data with an updated scaling factor greater than or equal to a preset threshold are selected as target sample data, and the target sample dataset is obtained. The preset threshold is determined based on the model training acceleration ratio.

9. A dataset generation apparatus, comprising: The execution unit is configured to perform the following determining steps based on the sample dataset and the corresponding initial scaling factor set: generating a sample scaling dataset according to the sample dataset and the initial scaling factor set, wherein the type of the sample dataset is at least one of the following: text, audio, video, or image; Generating a neural tangent kernel loss value and a predicted output loss value for a sample-scaled dataset includes: inputting the sample-scaled dataset into an initial model to obtain a predicted output value for the sample-scaled dataset; determining the loss value between the predicted output value and the true value corresponding to the sample-scaled dataset; determining the neural tangent kernel loss value for the sample-scaled dataset using a neural tangent kernel function; wherein the neural tangent kernel loss value represents the inner product value between the gradients of the function corresponding to the neural network model with respect to the network model parameters; generating a target loss value based on the neural tangent kernel loss value and the predicted output loss value; and, in response to determining that the target loss value satisfies the convergence condition, determining the initial scaling factor set as the updated scaling factor set. The generation unit is configured to generate a target sample dataset based on the sample dataset and the updated scaling factor set.

10. An electronic device, comprising: One or more processors; Storage device, on which one or more programs are stored, When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-8.

11. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-8.

12. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-8.

Images