Self-adaptive collaborative optimization electromagnetic signal intelligent model distributed learning training method

Abstract

The invention discloses a self-adaptive collaborative optimization electromagnetic signal intelligent model distributed learning training method, belongs to the technical field of radio signal processing, and solves the problem of poor self-adaptive collaborative optimization capability of an existing intelligent model. The method comprises the following steps: a scheduling node allocates a training node to each training task, and allocates a training sample set and the same electromagnetic intelligent model to be trained to each training node executing the same training task; the training nodes use the training sample set to perform multi-round distributed training on the to-be-trained electromagnetic intelligent model; when each round of training is executed, the scheduling node generates a current round aggregation parameter according to the current round final training parameters of all the training nodes executing the same training task, and the current round aggregation parameter serves as an initial training parameter for the corresponding training node to execute the next round of training until a preset round of training is completed; and the scheduling node screens out an optimal aggregation parameter from each round of aggregation parameters of each training task to obtain a trained electromagnetic intelligent model of the corresponding training task.

Description

Technical Field

[0001] The present invention relates to the technical field of radio signal processing, and in particular, to a distributed learning and training method for an intelligent model of electromagnetic signals with adaptive collaborative optimization. Background Art

[0002] In the context of the big data era, the breakthrough of the extreme performance of intelligent models often relies on the infusion of a large amount of training data, which greatly increases the pressure of centralized training of traditional intelligent models, thus causing a series of problems such as low model training efficiency and inefficient utilization of computing resources. At present, distributed learning and training of intelligent models is an effective means to solve the problems brought by the explosive growth of training data. It distributes data and models to different training nodes, and the scheduling node coordinates the computing resources of the training nodes to achieve collaborative training among distributed multi-nodes. Existing distributed training methods for intelligent models usually require consistent data partitioning on each training node, with balanced distribution of quantity, category, etc.

[0003] However, in reality, it is very difficult to ensure that each training node obtains the same amount of data during the process of the scheduling node partitioning data to the training nodes. For classification tasks, this is prone to the "short board" effect, and the collaborative performance of the model is restricted. In addition, traditional methods cannot adaptively select optimization strategies according to different task types, nor can they adaptively allocate computing resources.

[0004] Therefore, how to research a distributed learning and training method for an intelligent model of electromagnetic signals that can achieve adaptive collaborative optimization is a technical problem that urgently needs to be solved at present. Summary of the Invention

[0005] In view of the above analysis, an embodiment of the present invention aims to provide a distributed learning and training method for an intelligent model of electromagnetic signals with adaptive collaborative optimization to solve the problem of poor adaptive collaborative optimization ability of existing intelligent models.

[0006] The present invention discloses a distributed learning and training method for an intelligent model of electromagnetic signals with adaptive collaborative optimization. The distributed learning and training method includes: The scheduling node assigns training nodes to each training task respectively, and distributes the training sample set and the same electromagnetic intelligent model to be trained to each training node executing the same training task; The training node performs multiple rounds of distributed training on the electromagnetic intelligent model to be trained using the training sample set; for each round of training executed, the scheduling node generates the current round of aggregated parameters based on the current round of final training parameters of all training nodes executing the same training task, and uses them as the initial training parameters for the corresponding training node to execute the next round of training until the training of the preset number of rounds is completed; The scheduling node respectively selects the optimal aggregation parameters from the aggregation parameters of each round of each training task, and respectively constructs an electromagnetic intelligent model with the optimal aggregation parameters of each training task as the model parameters, as the trained electromagnetic intelligent model of the corresponding training task.

[0007] On the basis of the above solution, the present invention has also made the following improvements: Further, every time a round of training is executed, the scheduling node also updates the training sample set of the training nodes by performing the following operations: The scheduling node updates the training sample sets of the training nodes corresponding to each training task according to the time when all the training nodes executing the same training task feedback the final training parameters of the current round, and distributes the aggregation parameters of the current round of each training task and the updated training sample sets to the corresponding training nodes respectively.

[0008] Further, the scheduling node updates the training sample set of the training nodes by performing the following operations: According to the time when all the training nodes executing the same training task feedback the final training parameters of the current round, calculate the average time when all the training nodes executing the same training task feedback the final training parameters of the current round, and the time difference between the earliest feedback and the latest feedback; Calculate the average training sample size that each training node executing the same training task should obtain under the average distribution method; According to the time when each training node feedbacks the final training parameters of the current round, the average time, the time difference and the average training sample size, recalculate the updated training sample size of each training node executing the same training task; According to the updated training sample size of each training node, re-partition the electromagnetic signal training set of the corresponding training task to obtain the updated training sample set of each training node.

[0009] Further, the updated training sample size of the training node k executing the same training task is ; where represents the average training sample size that each training node executing the same training task should obtain under the average distribution method, represents the average time when all the training nodes executing the same training task feedback the final training parameters of the current round, represents the time difference between the earliest feedback and the latest feedback, represents the training node the time when the final training parameters of the current round are feedback.

[0010] Further, the scheduling node respectively selects the optimal aggregation parameters from the aggregation parameters of each round of each training task, and executes: The scheduling node respectively obtains the prediction accuracy rates of the electromagnetic intelligent models with the aggregation parameters of each round as the model parameters on the electromagnetic signal verification set, and takes the aggregation parameter with the highest prediction accuracy rate as the optimal aggregation parameter.

[0011] Furthermore, the distributed learning and training method further includes: The scheduling node clears the training nodes corresponding to each training task and shuts down the training tasks.

[0012] Furthermore, for each round of training, the training node obtains the final training parameters of the current round in the following manner: The training node initializes the training parameters of the electromagnetic intelligent model with the initial training parameters of the current round, performs multiple trainings on the electromagnetic intelligent model based on the training sample set to obtain the final training parameters of the current round, and sends the final training parameters of the current round to the scheduling node.

[0013] Furthermore, the performing multiple trainings on the electromagnetic intelligent model based on the training sample set to obtain the final training parameters of the current round includes: The training node performs multiple trainings on the electromagnetic intelligent model based on the training sample set. For each training, the loss function of the corresponding training task is calculated, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation result of the loss function until the number of trainings in one round is reached, so as to obtain the final training parameters of the current round.

[0014] Furthermore, for each round of training, the scheduling node generates the aggregation parameters of the current round in the following manner: The scheduling node performs weighted average aggregation on the final training parameters of the current round of all training nodes performing the same training task to generate the aggregation parameters of the current round of the corresponding training task.

[0015] Furthermore, the aggregation parameters of the current round are generated in the following manner: For each training task, the scheduling node respectively verifies the prediction accuracy rates of the electromagnetic intelligent models with the final training parameters of the current round of each training node as the model parameters on the electromagnetic signal verification set, and determines the weighting ratios of the final training parameters of the current round of each training node based on the prediction accuracy rates; And based on the determined weighting ratios, performs weighted average aggregation on the final training parameters of the current round of all training nodes performing the same training task.

[0016] Compared with the prior art, the present invention can at least achieve one of the following beneficial effects: An adaptive collaborative optimization-based distributed learning and training method for an electromagnetic signal intelligent model according to an embodiment of the present invention designs an adaptive training resource allocation mechanism, and also designs an adaptive training optimization strategy and a model parameter aggregation strategy according to different types of training tasks, solving problems such as performance bottlenecks caused by inconsistent data allocation in traditional distributed training methods, inability to adaptively select optimization strategies according to task types, and inability to adaptively allocate training resources. It can autonomously optimize collaborative strategies, improve training efficiency, and break through the model performance bottleneck under distributed training.

[0017] Therefore, this method can address problems such as performance bottlenecks caused by inconsistent data allocation in existing methods, inability to adaptively select optimization strategies according to task types, and inability to adaptively allocate training resources, so as to achieve distributed adaptive collaborative optimization training of intelligent models.

[0018] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combined solutions. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The drawings are only for the purpose of showing specific embodiments and are not considered to be a limitation of the present invention. Throughout the drawings, the same reference signs denote the same components; Figure 1 It is a flowchart of the distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] The following will specifically describe the preferred embodiments of the present invention in conjunction with the drawings, where the drawings form a part of this application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0021] A specific embodiment of the present invention discloses a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization, and the flowchart is as Figure 1 shown.

[0022] Step S1: The scheduling node assigns training nodes to each training task respectively, and distributes the training sample set and the same electromagnetic intelligent model to be trained to each training node executing the same training task.

[0023] In this embodiment, the user can create several training tasks according to the training requirements of the electromagnetic intelligent model and import them into the scheduling node. Each training task includes a task type, an electromagnetic signal training set, an electromagnetic signal verification set, and an electromagnetic intelligent model to be trained. Currently, in the field of electromagnetic signal processing, according to different task objectives, the task types of training tasks can be divided into classification tasks, regression tasks, and reconstruction tasks. If there are other task types that can use the distributed training method in this embodiment in subsequent academic research, they should also be included in the protection scope of this application. Specifically, classification tasks in the field of electromagnetic signal processing can achieve signal source identification, modulation mode identification, signal frequency band classification, etc. Regression tasks in the field of electromagnetic signal processing can achieve signal intensity prediction, signal propagation delay estimation, signal frequency offset estimation, etc. Reconstruction tasks in the field of electromagnetic signal processing can achieve signal denoising, signal missing data recovery, signal compression and reconstruction, etc. It should be emphasized that the training tasks of the above different task types in the field of electromagnetic signal processing are common training tasks in this field. The technical solution provided in this embodiment aims to train the model parameters of the electromagnetic intelligent model to be trained for the known task type, electromagnetic signal training set, electromagnetic signal verification set, and model structure of the electromagnetic intelligent model to be trained. The data sets and model structures of each training task can be implemented in an existing manner, and this embodiment does not make any limitations in this regard. In addition, it should be noted that in training tasks of different task types, there are differences in the labels of the electromagnetic signal training set and the electromagnetic signal verification set. The neural network structures of the electromagnetic intelligent models to be trained selected for training tasks of different task types are also different. Specifically, the label of the training sample for a classification task is a classification category and must be an integer; the label of the training sample for a regression task is the next value or other values of the training sample and can be a decimal; the label of the training sample for a reconstruction task depends on the model optimization situation.

[0024] Preferably, in this embodiment, considering that the training urgency of training tasks with different priority levels is different, the training task can also include a priority level, which is pre-configured by the user to ensure that the scheduling node preferentially executes training tasks with a high priority level. The specific implementation process of step S1 is described as follows.

[0025] Step S11: The scheduling node receives the training tasks and determines whether the idle training nodes can execute all the training tasks. If so, corresponding training nodes are allocated to each training task respectively; otherwise, training nodes are preferentially allocated to the training tasks with a high priority level. At the same time, the training tasks with a low priority level enter the waiting queue until the number of idle training nodes can meet the requirements of the training tasks with a low priority level.

[0026] In the specific implementation process, after receiving the training tasks created by the user, the scheduling node respectively identifies the indexes of each training task. In the specific implementation process, the scheduling node can directly allocate training nodes according to the training tasks; in addition, when there are sufficient idle nodes, it can also consider feeding back the aggregated idle training nodes to the user, and the user allocates the training nodes and feeds back the allocation results of the training nodes to the scheduling node. Whether the training nodes are allocated by the scheduling node or the user, their allocation ideas are the same. The difference is that when there are sufficient idle training nodes, the user can independently select an appropriate number of training nodes according to the training requirements and hand them over to the scheduling node for allocation. Specifically, if the number of idle training nodes cannot meet the requirements of the user's training tasks, the scheduling node can preferentially allocate training nodes for high-priority training tasks. If the number of idle training nodes is still insufficient to meet the needs of high-priority training tasks, it can be considered to stop some training nodes that are executing low-priority training tasks and expand them into idle training nodes. In the specific implementation process, it can be judged whether the training node (computing device) is in an idle state by judging its available computing power resources.

[0027] Preferably, in this embodiment, the training nodes can also be allocated in the following manner. The scheduling node respectively calculates the proportion of the number of training samples in the electromagnetic signal training set of each training task to the total number of training samples in the electromagnetic signal training sets of all training tasks, and uses this as the allocation ratio of the training nodes for different training tasks to allocate training nodes for each training task and ensure that each training task is allocated at least 1 training node. Assume that the number of idle training nodes is K, i represents the index of the training task, and the number of training samples in the electromagnetic signal training set corresponding to the i-th training task is ; where represents the task type, , and represents the total number of task types. Exemplarily, in this embodiment, takes 3, "1" represents the classification task, "2" represents the regression task, and "3" represents the reconstruction task. Then, the initial allocation process of the training nodes can be expressed as: (1) where represents the number of training nodes allocated to the training task with the task type and the index i . , is the floor operation.

[0028] Step S12: Divide the electromagnetic signal training sets in each training task and distribute different training sample sets to each training node executing the same training task.

[0029] Preferably, in this embodiment, distributing different training sample sets to each training node performing the same training task can make the data in the training sample sets allocated to each training node have different distribution characteristics, thereby enriching data diversity and effectively enhancing the generalization ability of the electromagnetic model to be trained.

[0030] Exemplarily, for each training task, if the training task is assigned to a single training node, the electromagnetic signal training set in the training task can be used as the training sample set and distributed to the corresponding training node; if the training task is assigned to two or more training nodes, the electromagnetic signal training set in the training task can be evenly divided according to the number of training nodes to which the training task is assigned, and each of the divided training samples is distributed to the corresponding training node.

[0031] In addition, it is also necessary to distribute the same electromagnetic intelligent model to be trained to each training node performing the same training task to achieve distributed training. The specific implementation process of distributed training is described in the following.

[0032] Step S2: The training nodes use the training sample set to perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained; for each round of training executed, the scheduling node generates the aggregation parameters of the current round based on the final training parameters of all training nodes performing the same training task in the current round, and uses them as the initial training parameters for the corresponding training nodes to execute the next round of training until the training of the preset number of rounds is completed.

[0033] Step S21: For each round of training executed by the training nodes, the training parameters of the electromagnetic intelligent model are initialized using the initial training parameters of the current round, and the electromagnetic intelligent model is trained multiple times based on the training sample set to obtain the final training parameters of the current round, and the final training parameters of the current round are sent to the scheduling node.

[0034] Specifically, it should be noted that when each training node executes the first round of training, the initial training parameters of the first round can be determined in an existing manner, and the training parameters of the electromagnetic intelligent model are initialized using the initial training parameters of the first round; starting from the second round of training, the aggregation parameters of the previous round of the same training task are used as the initial training parameters of the current round, and the training parameters of the electromagnetic intelligent model are initialized using the initial training parameters of the current round, that is, starting from the second round, the training parameters of the electromagnetic intelligent model are initialized to the aggregation parameters of the previous round of the same training task. This means that the starting point of each round of training is the comprehensive result of the training results of all training nodes in the previous round, so that the training parameters of the electromagnetic intelligent model can be updated and optimized during each round of training.

[0035] Specifically, during each round of training, after initializing the training parameters of the electromagnetic intelligent model with the initial training parameters of the current round, the training node performs multiple trainings on the electromagnetic intelligent model based on the training sample set. For each training, the loss function of the corresponding training task is calculated, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation result of the loss function until the number of trainings in one round is reached.

[0036] Preferably, in this embodiment, the model training optimization method for classification tasks is improved, and the specific description is as follows. For classification tasks, the cross-entropy loss function is used to train and optimize the electromagnetic intelligent model in the first round of training; and starting from the second round of training, in each round of training, a loss function that combines cross-entropy loss and adaptive constraint loss is used to train and optimize the electromagnetic intelligent model.

[0037] The loss function that combines cross-entropy loss and adaptive constraint loss Is expressed as: (2) Where, Represents the loss weighting ratio, , As a hyperparameter, it can be preset manually in advance; Represents the cross-entropy loss function, Represents the adaptive constraint loss function.

[0038] For regression tasks, during each round of training, a conventional regression loss function is used to train and optimize the electromagnetic intelligent model. The regression loss function Is expressed as: (3) Where, Represents the parametric function of the electromagnetic intelligent model for regression tasks, Is the training sample The corresponding true label; Represents the training parameters of the electromagnetic intelligent model for regression tasks.

[0039] For reconstruction tasks, during each round of training, a conventional reconstruction loss function is used to train and optimize the electromagnetic intelligent model. The reconstruction loss function Is expressed as: (4) Where, Represents the training parameters of the electromagnetic intelligent model for reconstruction tasks.

[0040] As can be seen from the above, for regression and reconstruction tasks, during the first round of training, each training node can initialize the training parameters of the electromagnetic intelligent model in the existing manner, and then perform multiple rounds of training of the electromagnetic intelligent model in the first round. For each training, the corresponding loss function is calculated, and the training parameters of the model are optimized based on the calculation results of the loss function until the current round of training ends (reaching the number of training times in one round), and the final training parameters of the first round are obtained. Starting from the second round of training, the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the previous round fed back by the scheduling node (i.e., the aggregated parameters of the same training task in the previous round) for each round of training, and multiple rounds of training of the electromagnetic intelligent model are performed in the same manner to obtain the final training parameters of the current round.

[0041] For classification tasks, during the first round of training, each training node first initializes the training parameters of the electromagnetic intelligent model in the existing manner. Then, during multiple rounds of training of the electromagnetic intelligent model in the first round, the corresponding loss function (cross-entropy loss function) is calculated for each training, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation results of the loss function until the first round of training ends (reaching the number of training times in one round), and the final training parameters of the first round are obtained. Starting from the second round of training, the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the same training task in the previous round for each round of training. During multiple rounds of training performed in the current round, the loss function that combines the cross-entropy loss and the adaptive constraint loss is calculated for each training, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation results of the loss function that combines the cross-entropy loss and the adaptive constraint loss until the current round of training ends, and the final training parameters of the current round are obtained. That is, starting from the second round of training, in addition to calculating the cross-entropy loss function for each training, the adaptive constraint loss function also needs to be calculated according to the aggregated parameters and the training parameters of this training so as to calculate the loss function of this time based on the combined cross-entropy loss and adaptive constraint loss and update the training parameters of the electromagnetic intelligent model when performing the next training according to the loss function .

[0042] It should be noted that in view of the performance bottleneck problem caused by inconsistent data distribution during the training process of the electromagnetic intelligent model for classification tasks, on the basis of the existing cross-entropy loss function, the adaptive constraint loss function is creatively proposed to optimize the model performance. The adaptive constraint loss function is related to the aggregated parameters (distributed training model parameters) obtained by the training node in the previous round of training and the training parameters of each training performed by the training node during the distributed training process in the current round (the second round and subsequent rounds). It should be emphasized that the cross-entropy loss function Requires a label; while the adaptive constraint function Does not require a label and relies on the training output of the electromagnetic intelligent model for constraint, as specifically described below.

[0043] In this embodiment, the electromagnetic intelligent model for the classification task sequentially includes an input layer, a hidden layer, and a softmax layer. The input layer and the hidden layer are collectively referred to as the previous layer. Starting from the second round of training, the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the same training task in the previous round , at this time, the training node performs the first training of the current round and obtains the output of the electromagnetic intelligent model for the training sample after the previous layer processing , represents the training sample, represents the current aggregated parameters of the electromagnetic intelligent model, represents the parameterization function of the previous layer, and the dimension of the vector is equal to the dimension of the classification categories. At the same time, when the training node performs each training of the current round, it also obtains the output of the electromagnetic intelligent model for the training sample after the previous layer processing , represents the training parameters of the electromagnetic intelligent model for this training. It should be noted that starting from the second round of training, for the first training in each round of training, since the training parameters of the electromagnetic intelligent model are initialized to the aggregated parameters of the same training task in the previous round , therefore, the and obtained in the first training are the same value. At the same time, since the training parameters are optimized and updated in each training, therefore, in each subsequent training process of the training node, it is necessary to obtain the output of the electromagnetic intelligent model for the training sample after the previous layer processing , and based on , calculate the adaptive constraint loss function for this training to obtain a loss function that combines the cross-entropy loss and the adaptive constraint loss, and update the training parameters for the next training accordingly.

[0044] Next, perform selective softening on , respectively, and the softening formula is: (5) (6) Among them, represents the vector The element corresponding to the th classification category in represents the softening result of and represents the softening temperature. represents the set of classification categories with unbalanced sample sizes of classification categories in the training samples of the training nodes. For selective softening is also performed to obtain .

[0045] It should be noted that the conventional softmax processes all classification categories and does not have a softening temperature parameter. In this embodiment, for selective softening is performed. The essence of selective softening is also softmax processing, but a softening temperature hyperparameter is added. Therefore, softmax processing is only performed on the set of classification categories with unbalanced sample sizes of classification categories in the training samples; that is, the selectivity is reflected in the need to select the part with unbalanced sample sizes of the classification categories output by the electromagnetic intelligent model. That is, if there are a total of 10 classification categories and the labels of the classification categories are {0, 1, 2, …, 9}, if the sample numbers of classification categories 2 and 4 are small, that is, unbalanced, then ={2, 4}. Therefore, represents the set of classification categories with partially unbalanced sample sizes in the training samples assigned to the training node.

[0046] Specifically, in a classification task, during the process of the scheduling node distributing the electromagnetic signal training set to the training nodes, if the training node 1 has a relatively large number of training samples of category a and the training node 2 has a relatively small number of training samples of category a, then the electromagnetic intelligent models on the training node 1 and the training node 2 have different classification capabilities for category a during their respective training processes. Since the training node 1 processes more training samples of category a, it performs better in the classification task of category a. In contrast, due to the insufficient number of training samples of category a on the training node 2, its classification performance for category a is slightly inferior. Therefore, the training parameters obtained by training the training node 2 may not reach the optimal state. When performing weighted averaging of the parameters on the scheduling node, it will definitely be affected. Therefore, it is necessary to use the adaptive constraint loss function proposed in this embodiment to constrain such losses.

[0047] Therefore, if a training node processes a small number of training samples of a certain category, its classification performance for the corresponding category is inferior to that of other training nodes. This performance difference will be reflected in the vector output by the model, specifically manifested as the vector The values of the elements in the corresponding classification categories may not be as reliable as the corresponding structures of the remaining training nodes. If the number of training samples in different categories is balanced, the finally calculated adaptive constraint loss function will degenerate into the cross-entropy loss function because the cross-entropy loss function can effectively measure the difference between the probability distribution predicted by the model and the probability distribution of the true labels, thereby guiding the training and optimization of the model.

[0048] Finally, by and calculate the selective constraint loss : (7) Step S22: The scheduling node performs weighted average aggregation on the current round of final training parameters of all training nodes executing the same training task to generate the current round of aggregation parameters for the corresponding training task.

[0049] Preferably, for each training task, the scheduling node respectively verifies the prediction accuracy of the electromagnetic intelligent model with the current round of final training parameters of each training node as the model parameters on the electromagnetic signal verification set, determines the weighting ratio of the current round of final training parameters of each training node based on the prediction accuracy; and performs weighted average aggregation on the current round of final training parameters of all training nodes executing the same training task according to the determined weighting ratio. In the specific implementation process, the weighted average aggregation strategy for different training tasks can be independently selected according to the task type. Exemplarily, for classification tasks, the scheduling node can respectively verify the classification accuracy of the electromagnetic intelligent model with the current round of final training parameters of each training node as the model parameters on the electromagnetic signal verification set to determine the weighting ratio of the current round of final training parameters of each training node. For regression tasks or reconstruction tasks, the scheduling node can respectively verify the mean square error of the model parameters with the current round of final training parameters fed back by each training node on the electromagnetic signal verification set to determine the weighting ratio of the training parameters of each training node.

[0050] Specifically, both the training parameters and the aggregation parameters are the model parameter sets of the electromagnetic intelligent model (including weight parameters and bias parameters of each layer, etc.). Since the electromagnetic intelligent models assigned to multiple training nodes executing the same training task are the same, then the aggregation parameters obtained by weighted average of the model parameters of each layer respectively also include the model parameter set of the electromagnetic intelligent model.

[0051] After the scheduling node receives the current round of final training parameters fed back by the corresponding training nodes, it can perform weighted average aggregation according to the following formula to obtain the aggregation parameters : (8) Among them, represents the training node executing the training task and is the weighted ratio of the final training parameters of the current round of the training node is the final training parameter of the current round fed back by the training node k.

[0052] Based on the foregoing description, it can be seen that the weighted ratio is autonomously adjusted according to the task type . For classification tasks ( ), , represents the recognition accuracy of the electromagnetic intelligent model fed back by the training node k on the electromagnetic signal validation set.

[0053] For regression or reconstruction tasks ( ), , represents the training node and is the mean square error of the electromagnetic intelligent model fed back on the electromagnetic signal validation set.

[0054] Step S23: The scheduling node respectively determines whether each training task has reached the training of the preset number of rounds. If not, it distributes the current round aggregation parameters of each training task to the corresponding training nodes respectively and jumps to the next round of training; otherwise, the training ends.

[0055] Preferably, to optimize the training effect, while generating the aggregation parameters for each round of training, the scheduling node can also update the training sample set for the next round of training of each training task. The specific implementation process is described as follows.

[0056] Step S24: The scheduling node updates the training sample sets of the training nodes for the corresponding training tasks according to the time when all the training nodes executing the same training task feed back the final training parameters of the current round, and distributes the current round aggregation parameters and the updated training sample sets of each training task to the corresponding training nodes respectively.

[0057] Specifically, the scheduling node calculates the average time when all the training nodes executing the same training task feed back the final training parameters of the current round according to the time when all the training nodes executing the same training task feed back the final training parameters of the current round , and the time difference between the earliest feedback and the latest feedback .

[0058] Suppose the time when the training node feeds back the final training parameters of the current round. At this time, the average time when all the training nodes executing the same training task feed back the final training parameters of the current round is expressed as: (9) The scheduling node calculates the average number of training samples that each training node should obtain when executing the same training task under the average distribution method. : (10) The scheduling node, according to the time when each training node feeds back the final training parameters of the current round, , and , recalculates the updated training sample amounts of each training node executing the same training task. Specifically, the updated training sample amount of training node k is . Finally, based on the updated training sample amounts of each training node, the electromagnetic signal training set of the corresponding training task is repartitioned to obtain the updated training sample sets of each training node, and the updated training sample sets and aggregation parameters are distributed to the corresponding training nodes.

[0059] It should be noted that due to the differences in the computing power resources of training nodes, the training speeds of different training nodes are different. If a training node trains slowly in the previous round of training, that is, the time to feedback the model to the scheduling node is late, then in the new round of training, the scheduling node will allocate less data volume to it, so as to achieve an improvement in the overall training speed. Data is reallocated in each round to achieve dynamic update adjustment of samples, so as to enhance the adaptability of the model to different samples.

[0060] Step S3: The scheduling node respectively screens out the optimal aggregation parameters from the aggregation parameters of each round of each training task, and constructs an electromagnetic intelligent model with the optimal aggregation parameters of each training task as the model parameters, as the trained electromagnetic intelligent model of the corresponding training task.

[0061] Specifically, the scheduling node respectively obtains the prediction accuracy rates of the electromagnetic intelligent models with the aggregation parameters of each round as the model parameters on the electromagnetic signal verification set, and takes the aggregation parameter with the highest prediction accuracy rate as the optimal aggregation parameter.

[0062] In the specific implementation process, the prediction process of the electromagnetic intelligent model with the aggregation parameters of each round as the model parameters on the electromagnetic signal verification set can be executed after generating each round of aggregation parameters, or can be executed after obtaining the aggregation parameters of all preset rounds, and the specific description is as follows.

[0063] Exemplarily, for each round of aggregated parameters generated, the scheduling node can verify the training task by checking the prediction accuracy of the electromagnetic intelligent model with the current round of aggregated parameters as the model parameters on the electromagnetic signal validation set. If it is higher than the prediction accuracy of the optimal aggregated parameters saved in the historical rounds, the current round of aggregated parameters is updated to the optimal aggregated parameters; otherwise, no update is performed. Specifically, the aggregated parameters are tested on the electromagnetic signal validation set, and the performance of the optimal aggregated parameters saved in the historical rounds on the electromagnetic signal validation set is compared. When this is the case, the recognition accuracy is compared. If it is higher than that in the historical round, it is updated to and saved; otherwise, it is not saved. When this is the case, the mean square error is compared. If it is lower than that in the historical round, then it is updated to and saved; otherwise, it is not saved. Alternatively, after the scheduling node obtains the aggregated parameters for all preset rounds, it respectively obtains the prediction accuracy of the electromagnetic intelligent model with each round of aggregated parameters as the model parameters on the electromagnetic signal validation set, and takes the aggregated parameters with the highest prediction accuracy as the optimal aggregated parameters.

[0064] Preferably, the method may further include the following steps.

[0065] Step S4: The scheduling node clears the training nodes corresponding to each training task and shuts down the training tasks.

[0066] In summary, this embodiment provides a distributed learning and training method for an electromagnetic signal intelligent model with adaptive collaborative optimization. The method includes: The scheduling node in the distributed training system adaptively allocates training nodes for tasks according to the task priority and the number of idle training nodes, and then distributes the electromagnetic signal intelligent model and electromagnetic signal training samples to the training nodes; after each training node receives the model and training samples, it adaptively selects an optimization strategy according to the type of the task to which the node belongs; when the training on the node reaches the specified number of rounds, the model is fed back to the scheduling node; the scheduling node adaptively selects an aggregation strategy according to the type of the task to which the fed-back model belongs, adaptively reallocates the sample amount according to the time speed of the node-fed-back model to regulate the training speed, and then redistributes the aggregated parameters and the newly allocated samples to the training nodes, repeating this process until the task reaches the specified number of distributed training rounds; the task model saved by the scheduling node is the electromagnetic signal intelligent model obtained by training. The solution of the embodiment of the present invention can adaptively select optimization and aggregation strategies according to the training task type, and adaptively allocate training resources, providing a new technical approach for the distributed learning and training of the electromagnetic signal intelligent model.

[0067] Those skilled in the art can understand that all or part of the processes for implementing the methods of the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a disk, an optical disc, a read-only memory or a random access memory, etc.

[0068] As described above, only the preferred specific embodiments of the present invention are provided, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. An adaptive collaborative optimization-based distributed learning and training method for an intelligent model of electromagnetic signals, characterized in that, The described distributed learning and training method includes: The scheduling node assigns training nodes to each training task respectively, and distributes the training sample set and the same electromagnetic intelligent model to be trained to each training node executing the same training task. The training nodes perform multiple rounds of distributed training on the electromagnetic intelligent model to be trained using the training sample set. For each round of training executed, the scheduling node generates the aggregated parameter of the current round based on the final training parameters of the current round of all training nodes executing the same training task, and uses it as the initial training parameter for the corresponding training nodes to execute the next round of training until the training of the preset number of rounds is completed. The scheduling node respectively screens out the optimal aggregated parameters from the aggregated parameters of each round of each training task, and constructs an electromagnetic intelligent model with the optimal aggregated parameters of each training task as the model parameters, which is used as the trained electromagnetic intelligent model of the corresponding training task.

2. The distributed learning and training method of the electromagnetic signal intelligent model with adaptive collaborative optimization according to claim 1, characterized in that For each round of training executed, the scheduling node also updates the training sample set of the training nodes by performing the following operations: The scheduling node updates the training sample sets of the training nodes executing the corresponding training task according to the time when all training nodes executing the same training task feedback the final training parameters of the current round, and distributes the aggregated parameter of the current round and the updated training sample set of each training task to the corresponding training nodes respectively.

3. The distributed learning and training method of the electromagnetic signal intelligent model with adaptive collaborative optimization according to claim 2, wherein The scheduling node updates the training sample set of the training nodes by performing the following operations: According to the time when all training nodes executing the same training task feedback the final training parameters of the current round, calculate the average time when all training nodes executing the same training task feedback the final training parameters of the current round, and the time difference between the earliest feedback and the latest feedback. Calculate the average training sample quantity that each training node executing the same training task should obtain under the average distribution method. According to the time when each training node feedbacks the final training parameters of the current round, the average time, the time difference, and the average training sample quantity, recalculate the updated training sample quantity of each training node executing the same training task. Based on the updated training sample quantity of each training node, re-partition the electromagnetic signal training set of the corresponding training task to obtain the updated training sample set of each training node.

4. The distributed learning and training method of the electromagnetic signal intelligent model with adaptive collaborative optimization according to claim 3, characterized in that, The updated training sample size of the training node k that executes the same training task is ; where represents the average number of training samples that each training node executing the same training task should obtain under the average distribution method, represents the average time for all training nodes executing the same training task to feedback the final training parameters of the current round, represents the time difference between the earliest feedback and the latest feedback, represents the training node the time to feedback the final training parameters of the current round.

5. The distributed learning and training method for the intelligent model of electromagnetic signals with adaptive collaborative optimization according to any one of claims 1-4, characterized in that, The scheduling node respectively screens out the optimal aggregated parameters from the aggregated parameters of each round of each training task and executes: The scheduling node respectively obtains the prediction accuracy rate of the electromagnetic intelligent model with each round of aggregated parameters as the model parameters on the electromagnetic signal verification set, and takes the aggregated parameter with the highest prediction accuracy rate as the optimal aggregated parameter.

6. The distributed learning and training method of the electromagnetic signal intelligent model with adaptive collaborative optimization according to claim 5, characterized in that, The described distributed learning and training method further includes: The scheduling node clears the training nodes corresponding to each training task and closes the training task.

7. The distributed learning and training method for the intelligent model of electromagnetic signals with adaptive collaborative optimization according to claim 5, characterized in that, For each round of training executed, the training nodes obtain the final training parameters of the current round through the following method: The training nodes initialize the training parameters of the electromagnetic intelligent model with the initial training parameters of the current round, perform multiple trainings on the electromagnetic intelligent model based on the training sample set, obtain the final training parameters of the current round, and send the final training parameters of the current round to the scheduling node.

8. The distributed learning and training method of the electromagnetic signal intelligent model with adaptive collaborative optimization according to claim 7, characterized in that, The multiple trainings on the electromagnetic intelligent model based on the training sample set to obtain the final training parameters of the current round include: The training nodes perform multiple trainings on the electromagnetic intelligent model based on the training sample set. For each training, the loss function of the corresponding training task is calculated, and the training parameters of the electromagnetic intelligent model are optimized based on the calculation results of the loss function until the number of trainings in one round is reached, and the final training parameters of the current round are obtained.

9. The distributed learning and training method for the intelligent model of electromagnetic signals with adaptive collaborative optimization according to claim 8, wherein For each round of training executed, the scheduling node generates the aggregation parameters of the current round in the following manner: The scheduling node performs weighted average aggregation on the final training parameters of the current round of all training nodes that execute the same training task to generate the aggregation parameters of the current round of the corresponding training task.

10. The distributed learning and training method for the intelligent model of electromagnetic signals with adaptive collaborative optimization according to claim 9, characterized in that Generate the aggregation parameters of the current round in the following manner: For each training task, the scheduling node respectively verifies the prediction accuracy of the electromagnetic intelligent model with the final training parameters of the current round of each training node as the model parameters on the electromagnetic signal verification set, and determines the weighting ratio of the final training parameters of the current round of each training node based on the prediction accuracy; And based on the determined weighting ratio, perform weighted average aggregation on the final training parameters of the current round of all training nodes that execute the same training task.

Images