A joint training method and system based on real data and synthetic data

Abstract

The present application relates to the field of electric digital data processing, and particularly relates to a joint training method and system based on real data and synthetic data, the method comprising: obtaining a mixed sample set composed of real data and synthetic data; constructing a global correlation graph based on the mixed sample set, extracting global authenticity and global closeness, and adaptively determining an optimal number of neighbors for synthetic sample quality evaluation; for each synthetic sample, constructing a neighbor subgraph by K-neighbor method using the optimal number of neighbors, extracting subgraph authenticity and subgraph closeness, and obtaining a quality score by a fusion algorithm; giving a fixed weight to real samples and a corresponding quality score to synthetic samples as a weight, constructing a weighted loss function based on sample weights, and completing joint training of the model. The present application improves the model training effect through adaptive number of neighbors, multi-dimensional quality evaluation and weighted differentiated training.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing. In particular, it relates to a joint training method and system based on real and synthetic data. Background Technology

[0002] In the field of machine learning model training, the scarcity and imbalanced distribution of real-world labeled samples are common challenges, particularly in practical scenarios such as computer vision and industrial inspection. This directly leads to insufficient feature learning in the model, resulting in problems such as insufficient generalization ability and low training accuracy, severely restricting the effectiveness of model deployment. Synthetic data, due to its customizability, lack of privacy risks, and ability to simulate the characteristics of real data, has become an important means of supplementing real samples. Generative synthesis based on GANs and interpolation synthesis based on SMOTE are widely used. Therefore, joint training with synthetic and real data has become a mainstream path to improve model training performance.

[0003] Joint training with synthetic and real data can effectively fill gaps in the distribution of real samples, expand the training sample size, and allow the model to learn richer feature information. It is a key means to solve the problems of scarcity and uneven distribution of real samples, and its application value has been verified in various machine learning tasks. However, the quality of synthetic data varies greatly due to factors such as generation algorithms and parameter settings. If the effectiveness of synthetic data cannot be scientifically identified and used in a differentiated manner, the supplementary value of high-quality synthetic data will be difficult to realize, and poor-quality data will introduce noise that interferes with the model's learning of real sample features. Therefore, the effective evaluation and reasonable integration of synthetic data have become the core key to improving the joint training effect.

[0004] Current mainstream joint training schemes using synthetic and real data still have significant technical flaws, making it difficult to guarantee training effectiveness: Existing technologies mostly adopt a one-size-fits-all hybrid training mode without differentiating the quality of synthetic data. Poor-quality data can easily introduce noise and reduce model performance. At the same time, the core nearest neighbor parameter of sample association analysis uses fixed empirical values ​​and cannot be adaptively adjusted according to the actual distribution characteristics of the mixed data, resulting in insufficient accuracy of sample association mining. Furthermore, existing schemes lack consideration for the fusion of global and local features, and evaluate synthetic data from only a single dimension, making it difficult to fully reflect its fusion and adaptability with real data, and also unable to accurately judge its supplementary value to the distribution of real samples. Summary of the Invention

[0005] To address the problems of existing technologies that use a one-size-fits-all approach to training synthetic data, lack differentiated quality evaluation, fix the number of nearest neighbors, and lack the fusion of global and local features in single-dimensional evaluation, resulting in impaired model performance, inaccurate association mining, and inaccurate suitability assessment, this invention provides solutions in the following aspects.

[0006] In a first aspect, a joint training method based on real data and synthetic data includes: acquiring a mixed sample set of real data and synthetic data, wherein each sample in the mixed sample set consists of a data feature vector and a class label, with the class label of real samples marked as 1 and the class label of synthetic samples marked as 0; constructing a global association graph based on the mixed sample set, extracting the global realism and global density of the global association graph, and adaptively determining the optimal number of nearest neighbors required for quality evaluation of synthetic samples based on the global realism and global density; for each synthetic sample in the mixed sample set, constructing a nearest neighbor subgraph using the K-nearest neighbor method combined with the optimal number of nearest neighbors, extracting the subgraph realism and subgraph density of the nearest neighbor subgraph, and obtaining the quality score of each synthetic sample through a fusion algorithm; assigning weights to each sample in the mixed sample set, wherein the weight of real samples is 1 and the weight of synthetic samples is the corresponding quality score, setting a weighted loss function for joint training based on the sample weights, and using the weighted loss function to perform joint training of the model on the mixed sample set to complete model training.

[0007] Preferably, the steps for constructing a global association graph based on a mixed sample set include: Each sample in the mixed sample set is treated as a node in the global association graph, forming a node set, and at the same time, a subset of real nodes and a subset of synthetic nodes are obtained. Calculate the Euclidean distance between any two nodes, set a global edge threshold, and construct valid edges for node pairs whose Euclidean distance is less than or equal to the global edge threshold, forming a set of valid edges; The edge weight of each valid edge in the valid edge set is calculated by the reciprocal of the Euclidean distance between nodes, forming an edge weight set. The global relational graph is composed of the node set, the valid edge set, and the edge weight set.

[0008] Preferably, the global realism is the ratio of the sum of the core degrees of all nodes in the real node subset to the sum of the core degrees of all nodes in the node set. If the sum of the core degrees of all nodes in the node set is 0, then the global realism is 0. Here, the core degree of a node is the sum of the edge weights of all valid edges of the node. The global compactness is the mean of the improved local clustering coefficients of all nodes in the node set. The improved local clustering coefficients are calculated based on the basic clustering coefficients of the nodes and the importance coefficients of all triangles formed by the nodes. The basic clustering coefficients are the ratio of the total number of triangles actually formed by the nodes to the maximum number of triangles in the nearest neighbor subgraph. The importance coefficients of the triangles are the result of normalization based on the sum of the weights of the real edges in the triangle, where the weights of the real edges are the effective edges of the nodes corresponding to the real samples.

[0009] Preferably, the step of determining the optimal number of nearest neighbors required for synthetic sample quality assessment includes: The optimal nearest neighbor number is obtained by rounding down the ratio between the preset baseline positive integer and the product of global truth and global compactness.

[0010] Preferably, constructing a nearest neighbor subgraph includes: The current synthetic sample to be evaluated is taken as the central node, and the K nearest neighbor samples obtained by K-nearest neighbor method are taken as the nearest neighbor nodes, which together form the node set of the nearest neighbor subgraph; For the central node and its nearest neighbor nodes, a direct connection is used as an effective edge. For the nearest neighbor nodes, the effective edges in the global association graph are reused to form the effective edge set of the nearest neighbor subgraph. By adopting the edge weight calculation rules of the global association graph, the edge weight set of the nearest neighbor subgraph is obtained. The nearest neighbor subgraph is composed of the node set, the effective edge set, and the edge weight set.

[0011] Preferably, the subgraph realism is the ratio of the sum of the core degrees of real nodes in the nearest subgraph to the sum of the core degrees of all nodes in the nearest subgraph; The subgraph density is the improved local clustering coefficient of the central node corresponding to the synthetic sample, which is calculated by focusing only on the central node. The calculation method of the improved local clustering coefficient is the same as the calculation method of the improved local clustering coefficient of the node in the global association graph.

[0012] Preferably, the weighted loss function is calculated as follows: The sum of the products of the weights of each sample and the corresponding basic loss term is used as the total loss of the jointly trained model.

[0013] Secondly, a joint training system based on real data and synthetic data includes: a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned joint training method based on real data and synthetic data is implemented.

[0014] The present invention has the following effects: 1. This invention constructs a global association graph and extracts global truth and global density, and combines the two to adaptively determine the optimal number of nearest neighbors for synthetic sample quality assessment. This replaces the fixed empirical value setting method of the prior art, and makes the number of nearest neighbors adapt to the actual distribution characteristics of the mixed data, effectively improving the accuracy of sample association mining and laying a reliable parameter foundation for synthetic data quality assessment.

[0015] 2. This invention constructs nearest-neighbor subgraphs on the synthetic samples and extracts the subgraph realism and subgraph compactness, and then fuses them to obtain a synthetic data quality score. This achieves a multi-dimensional quality assessment of the fusion of global and local features of the synthetic data, breaking through the limitations of the single-dimensional assessment of existing technologies. It can comprehensively and accurately judge the fusion compatibility between synthetic data and real data, and effectively screen out high-quality synthetic data.

[0016] 3. This invention assigns corresponding training weights to synthetic data by assigning quality scores, and conducts joint training by setting a weighted loss function with fixed weights for real samples. This replaces the hybrid training mode of the existing technology, realizes differentiated training of synthetic data, fully leverages the supplementary value of high-quality synthetic data to real data, weakens the noise interference of low-quality synthetic data, effectively avoids damage to model performance, and improves the effect of joint training of the model. Attached Figure Description

[0017] Figure 1 This is a flowchart of steps S1-S4 in a joint training method based on real data and synthetic data according to an embodiment of the present invention.

[0018] Figure 2 This is a structural block diagram of a joint training system based on real data and synthetic data according to an embodiment of the present invention. Detailed Implementation

[0019] 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 some embodiments of the present invention, but not all embodiments.

[0020] Reference Figure 1 A joint training method based on real and synthetic data includes steps S1-S4, as follows: S1: Obtain a mixed sample set of real and synthetic data. Each sample in the mixed sample set consists of a data feature vector and a class label. The class label of the real sample is marked as 1, and the class label of the synthetic sample is marked as 0.

[0021] For the target application scenario, the raw real data directly obtained by physical sensors, historical records, etc., is first collected and preprocessed to adapt to the scenario to obtain real data that can be directly used for model training. At the same time, synthetic data is generated by artificial construction or algorithm simulation. The generated synthetic data maintains a high degree of similarity to the aforementioned real data in terms of statistical distribution, feature patterns, and semantic content, thereby ensuring that the synthetic data effectively supplements the real data.

[0022] The methods for generating synthetic data include, but are not limited to, the following three: First, using pre-trained generative models, such as generative adversarial networks and variational autoencoders, to generate samples that match the potential distribution of real data; second, generating data in a controllable simulation environment, such as a physical simulator, according to pre-set rules or parameterized models; and third, performing data augmentation operations such as noise addition and SMOTE (Synthetic Minority Oversampling Technique) interpolation on existing real data to generate new samples through reasonable transformation of the real data.

[0023] S2: Construct a global association graph based on the mixed sample set, extract the global truth and global compactness of the global association graph, and adaptively determine the optimal number of nearest neighbors required for the quality assessment of the synthetic sample based on the global truth and global compactness.

[0024] Constructing a global association graph involves mapping samples in the mixed sample set to graph nodes and the relationships between samples to valid edges and edge weights, thus transforming the mixed sample set from a numerical distribution to a topological structure. The global association features of all samples are then drawn using this method. The steps are as follows: Each sample in the mixed sample set is assigned a node in the global association graph, and all nodes together form the node set of the global association graph. At the same time, based on the category label of the sample, the nodes in the node set corresponding to the real samples are divided into a real node subset, and the nodes corresponding to the synthetic samples are divided into a synthetic node subset. Calculate the Euclidean distance between any two nodes in the node set. Pre-set a global edge threshold as the criterion for determining whether there is a valid association between nodes. Connect the node pairs whose Euclidean distance is less than or equal to the global edge threshold to form a valid edge set of the global association graph. Based on the principle that the smaller the distance between nodes, the higher the degree of association, the edge weight of each valid edge in the effective edge set is calculated according to the reciprocal of the Euclidean distance between nodes. The edge weights of all valid edges together form the edge weight set of the global association graph. The node set, effective edge set, and edge weight set constructed above together constitute a global association graph used to characterize the global association relationships of the mixed sample set.

[0025] Global truth score is used to quantify the global core dominance of real samples in the mixed sample set. Its value is the ratio of the sum of the core scores of all nodes in the real node subset to the sum of the core scores of all nodes in the global association graph node set. If the sum of the core scores of all nodes in the node set is 0, the global truth score is set to 0. The core score of any node is the sum of the edge weights of all valid edges of that node in the global association graph. The higher the core score, the stronger the association and the more core the topological position of that node in the global association graph.

[0026] Specifically, global realism satisfies the following relationship: ; In the formula, This represents the global truth of the global association graph, with a value range of [value range missing]. ; This represents the set of nodes in the global association graph. This represents a subset of the actual nodes in the global association graph; Represents the first in the subset of real nodes The core degree of each node corresponding to a real sample; Represents the first node in the node set. The coreness of each node.

[0027] Global density measures the overall topological stability of a mixed sample set. Its value is the mean of the improved local clustering coefficients of all nodes in the global association graph node set. The improved local clustering coefficient is an indicator that integrates the contribution of real samples to the node's neighborhood structure. It is calculated as follows: first, obtain the node's basic clustering coefficient; then, calculate and sum the importance coefficients of all triangles formed by that node; multiply the basic clustering coefficient by the sum of the triangle importance coefficients; and finally, normalize using a Sigmoid transformation formula to obtain the node's improved local clustering coefficient. This simultaneously reflects the structural density of the node's neighborhood and the support of real samples for the neighborhood structure. The basic clustering coefficient of a node is the ratio of the total number of triangles actually formed by that node to the maximum possible number of triangles that can be formed when all nodes in the nearest subgraph are connected pairwise. A larger basic clustering coefficient indicates a higher natural connectivity density in the node's local neighborhood.

[0028] Specifically, the improved local clustering coefficients satisfy the following relationship: ; In the formula, Indicates the first Improved local clustering coefficients for each node, with a range of values. ; Indicates the first The basic clustering coefficients of each node; Indicates the first The node participates in the formation of the first The importance coefficients of each triangle; Indicates the first The total number of nodes that participate in forming a triangle.

[0029] The importance coefficient of a triangle is an indicator that quantifies the contribution of a single triangle to the stability of the node neighborhood structure. It is calculated as follows: First, the real edges inside the triangle are defined as the effective edges of nodes containing real samples. The weights of all real edges inside the triangle are counted and summed. The sum of the real edge weights is then substituted into a modified Sigmoid formula for normalization to obtain the importance coefficient of the triangle. The larger the sum of the real edge weights, the higher the participation and support of real samples in the triangle, and the larger the corresponding importance coefficient of the triangle. This highlights the core contribution of triangles containing real samples to the stability of the global association graph topology.

[0030] Specifically, the importance coefficients of the triangle satisfy the following relationship: ; In the formula, Indicates the first The node participates in the formation of the first The importance coefficient of each triangle Indicates the first The node participates in the formation of the first The set of real sides of a triangle Indicates the first The edge weight of one edge in the set of real edges inside a triangle.

[0031] By combining the global distribution characteristics of the mixed sample set reflected by global realism and global density, the optimal number of nearest neighbors for the quality assessment of synthetic samples is adaptively determined through quantitative calculation. Specifically, a preset positive integer is used as the benchmark value for calculating the number of nearest neighbors. This benchmark value is then compared with the product of global realism and global density. The result is then rounded down, and the result is the optimal number of nearest neighbors. This calculation method adapts the optimal number of nearest neighbors to the actual data distribution of the mixed sample set. The higher the global realism and global density, the stronger the global dominance of the real samples and the more stable the topological structure of the mixed sample set. The smaller the number of nearest neighbors required, the more accurately the core nearest neighbor features of the synthetic samples are captured, avoiding the introduction of irrelevant samples that cause evaluation bias.

[0032] Specifically, the optimal nearest neighbor number satisfies the following relationship: ; in, Represents the number of best nearest neighbors; This represents the preset base positive integer, set to 10, which can be adjusted according to actual needs; This represents the global truth value of the global association graph. This indicates the global tightness of the global association graph.

[0033] S3: For each synthetic sample in the mixed sample set, the K-nearest neighbor method is used to construct a nearest neighbor subgraph by combining the optimal number of nearest neighbors. The subgraph truth and subgraph compactness of the nearest neighbor subgraph are extracted, and the quality score of each synthetic sample is obtained through the fusion algorithm.

[0034] The nearest neighbor subgraph is constructed by, specifically, for each synthetic sample to be evaluated, mining its local nearest neighbors based on the optimal number of nearest neighbors, reusing the effective edges and weight rules of the global association graph to construct the local topology, and then drawing the local association features of a single synthetic sample. The steps are as follows: The current synthetic sample to be evaluated is taken as the center node. Using Euclidean distance as the metric, the K nearest neighbor samples of the center node are selected from the mixed sample set by the K nearest neighbor method and taken as the nearest neighbor nodes. The center node and all the nearest neighbor nodes together constitute the node set of the nearest neighbor subgraph corresponding to the synthetic sample.

[0035] For all nodes in the node set of the nearest neighbor subgraph, two types of node pairs are formed: one type consists of center-nearest neighbor pairs formed by the center node and each of its nearest neighbor nodes, and the other type consists of nearest neighbor-nearest neighbor pairs formed by each pair of nearest neighbor nodes. A direct connection is established between center-nearest neighbor pairs, which serve as valid edges in the nearest neighbor subgraph. For nearest neighbor-nearest neighbor pairs, existing valid edges in the global association graph are reused as valid edges in the nearest neighbor subgraph. All valid edges together form the valid edge set of the nearest neighbor subgraph, thereby ensuring the consistency between the local topology and the global distribution characteristics.

[0036] Using the same edge weight calculation rules as the global association graph, the edge weight of each valid edge in the effective edge set of the nearest subgraph is calculated, and all edge weights together form the edge weight set of the nearest subgraph.

[0037] The node set, effective edge set, and edge weight set constructed above together constitute the nearest neighbor subgraph corresponding to the current synthetic sample to be evaluated.

[0038] The subgraph realism of the nearest neighbor subgraph is used to quantify the degree of fit between the synthetic sample and the real sample within its local nearest neighbor range. Its calculation method is consistent with the calculation logic of the global realism. It is taken as the ratio of the sum of the core degrees of the real sample nodes in the nearest neighbor subgraph to the sum of the core degrees of all nodes in the nearest neighbor subgraph. The higher the ratio, the stronger the proportion and dominance of the real sample in the local neighborhood of the synthetic sample, and the higher the local feature matching degree between the synthetic sample and the real sample. The calculation rule of the node core degree is the same as that of the node core degree in the global association graph. It is the sum of the edge weights of all valid edges of the node in the corresponding nearest neighbor subgraph.

[0039] The subgraph compactness of the nearest neighbor subgraph measures the structural stability and reliability of the synthetic sample itself in its local nearest neighbor topology. Its value is the improved local clustering coefficient of the central node corresponding to the synthetic sample. The calculation focuses only on the central node and does not involve the related calculations of other nodes in the nearest neighbor subgraph, so as to avoid the structural features of other nodes interfering with the judgment of the structural stability of the synthetic sample itself. Moreover, the specific calculation method of the improved local clustering coefficient of the central node is completely consistent with the calculation method of the improved local clustering coefficient of the node in the global association graph, ensuring the logical consistency between local feature calculation and global feature calculation.

[0040] S4: Assign weights to each sample in the mixed sample set, where the weight of the real sample is 1 and the weight of the synthetic sample is the corresponding quality score. Set the weighted loss function for joint training based on the sample weights, and use the weighted loss function to jointly train the model on the mixed sample set to complete the model training.

[0041] To achieve differentiated training between real samples and synthetic samples of varying quality, a weighted loss calculation method is used to determine the total loss of the joint training model. Specifically, the weight of each sample in the mixed sample set is multiplied by its corresponding basic loss term, and then the product results of all samples are summed. The summation result is the total loss of the joint training model. By matching the influence of sample loss on model training with the weight of the sample itself, real samples and high-quality synthetic samples, due to their higher weights, contribute more to the total loss through their loss terms. The model training will prioritize fitting the feature distribution of these types of samples, while the influence of low-quality synthetic samples, due to their lower weights, is weakened. This effectively avoids the interference of inferior synthetic data on model training.

[0042] Specifically, the weighted loss function satisfies the following relationship: ; in, This represents the total loss of the jointly trained model. Indicates the first The weights of each sample, Indicates the first Mean squared error of each sample This represents the total number of samples in the mixed sample set.

[0043] This invention also provides a joint training system based on real data and synthetic data. For example... Figure 2 As shown, the system includes a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a joint training method based on real and synthetic data according to the first aspect of the present invention. The system also includes other components well-known to those skilled in the art, such as a communication bus and a communication interface; their configuration and functions are known in the art and will not be described further here.

[0044] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A joint training method based on real data and synthetic data, characterized in that, include: Obtain a mixed sample set of real data and synthetic data. Each sample in the mixed sample set consists of a data feature vector and a class label. The class label of the real sample is marked as 1, and the class label of the synthetic sample is marked as 0. A global association graph is constructed based on a mixed sample set. The global truth and global compactness of the global association graph are extracted. The optimal number of nearest neighbors required for the quality evaluation of the synthetic sample is adaptively determined based on the global truth and global compactness. For each synthetic sample in the mixed sample set, the K-nearest neighbor method is used to construct a nearest neighbor subgraph by combining the optimal nearest neighbor number. The subgraph truth and subgraph compactness of the nearest neighbor subgraph are extracted, and the quality score of each synthetic sample is obtained through the fusion algorithm. Each sample in the mixed sample set is assigned a weight, where the weight of the real sample is 1 and the weight of the synthetic sample is the corresponding quality score. A weighted loss function for joint training is set based on the sample weights, and the model is jointly trained on the mixed sample set using the weighted loss function to complete the model training.

2. The joint training method based on real data and synthetic data according to claim 1, characterized in that, The steps for constructing a global association graph based on a mixed sample set include: Each sample in the mixed sample set is treated as a node in the global association graph, forming a node set, and at the same time, a subset of real nodes and a subset of synthetic nodes are obtained. Calculate the Euclidean distance between any two nodes, set a global edge threshold, and construct valid edges for node pairs whose Euclidean distance is less than or equal to the global edge threshold, forming a set of valid edges; The edge weight of each valid edge in the valid edge set is calculated by the reciprocal of the Euclidean distance between nodes, forming an edge weight set. The global relational graph is composed of the node set, the valid edge set, and the edge weight set.

3. The joint training method based on real data and synthetic data according to claim 2, characterized in that, The global realism is the ratio of the sum of the core degrees of all nodes in the subset of real nodes to the sum of the core degrees of all nodes in the set of nodes. If the sum of the core degrees of all nodes in the set of nodes is 0, then the global realism is 0. The core degree of a node is the sum of the edge weights of all valid edges of the node. The global compactness is the mean of the improved local clustering coefficients of all nodes in the node set. The improved local clustering coefficients are calculated based on the basic clustering coefficients of the nodes and the importance coefficients of all triangles formed by the nodes. The basic clustering coefficients are the ratio of the total number of triangles actually formed by the nodes to the maximum number of triangles in the nearest neighbor subgraph. The importance coefficients of the triangles are the result of normalization based on the sum of the weights of the real edges in the triangle, where the weights of the real edges are the effective edges of the nodes corresponding to the real samples.

4. The joint training method based on real data and synthetic data according to claim 1, characterized in that, The steps to determine the optimal number of nearest neighbors required for synthetic sample quality assessment include: The optimal nearest neighbor number is obtained by rounding down the ratio between the preset baseline positive integer and the product of global truth and global compactness.

5. The joint training method based on real data and synthetic data according to claim 1, characterized in that, Constructing a nearest neighbor subgraph includes: The current synthetic sample to be evaluated is taken as the central node, and the K nearest neighbor samples obtained by K-nearest neighbor method are taken as the nearest neighbor nodes, which together form the node set of the nearest neighbor subgraph; For the central node and its nearest neighbor nodes, a direct connection is used as an effective edge. For the nearest neighbor nodes, the effective edges in the global association graph are reused to form the effective edge set of the nearest neighbor subgraph. By adopting the edge weight calculation rules of the global association graph, the edge weight set of the nearest neighbor subgraph is obtained. The nearest neighbor subgraph is composed of the node set, the effective edge set, and the edge weight set.

6. The joint training method based on real data and synthetic data according to claim 1, characterized in that, The subgraph realism is the ratio of the sum of the coreities of real nodes in the nearest subgraph to the sum of the coreities of all nodes in the nearest subgraph. The subgraph density is the improved local clustering coefficient of the central node corresponding to the synthetic sample, which is calculated by focusing only on the central node. The calculation method of the improved local clustering coefficient is the same as the calculation method of the improved local clustering coefficient of the node in the global association graph.

7. The joint training method based on real data and synthetic data according to claim 1, characterized in that, The weighted loss function is calculated as follows: The sum of the products of the weights of each sample and the corresponding basic loss term is used as the total loss of the jointly trained model.

8. A joint training system based on real data and synthetic data, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement the joint training method based on real and synthetic data according to any one of claims 1-7.

Images