Federal learning model training method

Abstract

The invention belongs to the technical field of distributed intelligent computing, and aims to provide a federal learning model training method. In the implementation process, a semi-asynchronous federated learning aggregation scheme with a fixed time period is constructed by setting a period aggregation duration and adopting a semi-asynchronous federated learning model training mode, so that the bottleneck node delay influence of a synchronous federated aggregation strategy in an equipment heterogeneous environment can be solved, and the reliability of the system is improved. Meanwhile, deflection of a global model due to the fact that part of equipment frequently uploads the local model in the complete asynchronous federation aggregation strategy is prevented; besides, under the environment that the computing power of the federated learning system is unbalanced, data in all local clients can participate in model training in a semi-asynchronous aggregation mode under the condition that global training time is fixed, so that the data utilization rate of the local clients participating in training in the federated learning system is effectively improved, and the training efficiency of the federated learning system is improved. And the convergence precision of model training is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of distributed intelligent computing, and specifically relates to a federated learning model training method. Background Art

[0002] As edge devices continue to improve their storage and computing capabilities, federated learning technology has been proposed to release computing power from traditional centralized cloud intelligence to end-side edge intelligence while protecting client data privacy. Specifically, federated learning technology can achieve collaborative model training between local devices and cloud server devices, ensuring that data does not leave the client, preventing the leakage of user privacy data.

[0003] However, in the process of using the prior art, the inventors found that the prior art has at least the following problems: In the actual application of existing federated learning technology, the federated learning system has various degrees of heterogeneity in the statistical distribution of data and the computing power configuration of the end-side devices. Among them, the heterogeneity of the types and quantities of data on the end-side devices will cause the model convergence accuracy to decrease, and the heterogeneity of the computing power configuration of the end-side devices will cause the computing time overhead of different end-side devices to differ. This heterogeneity will produce bottleneck nodes under the classic synchronous aggregation strategy, destroying the reliability of the training model convergence and prolonging the total time of the single-step global model aggregation update.

[0004] Therefore, how to alleviate the negative impact of client data and device heterogeneity on system training accuracy and speed during the training of the federated learning system, and achieve efficient, stable and effective model training and aggregation, is an urgent problem that needs to be solved in applying federated learning to practice. Summary of the invention

[0005] The present invention aims to solve the above technical problems at least to a certain extent, and provides a method for training a federated learning model.

[0006] In order to achieve the above object, the present invention adopts the following technical solutions: In a first aspect, the present invention provides a method for training a federated learning model, which is performed based on a federated learning system, wherein the federated learning system includes a central server and a plurality of local clients respectively connected to the central server in communication; the method includes: The central server receives the federated learning model training task and sets the initial global model according to the federated learning model training task , and then the initial global model The model to be trained is broadcasted to multiple local clients; the central server obtains the number of local clients K , initialize the delay status value of all local clients ( ) and training status values ( ), and set the periodic aggregation duration ; Any local client receives the central server at ( ) round of training, and according to its t Round training status value Perform local model training; as described in any local client t Round training status value , then any local client based on the local data set The model to be trained is t round of local model training; as described in the first round of local model training of any local client t Round training status value , then any local client continues the unfinished local model training; wherein, , T The central server broadcasts the total number of rounds of the model to be trained; In the t Before the end of the periodic aggregation duration of the round of training, any local client has completed the first t round of local model training, then its first t+ 1 round of training status value Set to 1 to make it t+ 1 round delay status value Set to 0 and complete the t The training model obtained after the round of local model training is uploaded to the central server; in the t After the periodic aggregation time of the round of training ends, if any local client fails to complete the t round of local model training, then its first t+ 1 round of training status value Set it to 0 and t+ 1 round delay status value Set to ; The central server receives the t All training models in round 1 training are divided into t The weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model ; The central server determines the updated global model Is it converged? If so, the updated global model as the trained global model, and end the federated learning model training task; if not, then update the updated global model The model is broadcasted to multiple local clients as a model to be trained.

[0007] In one possible design, any local client t Round training status value , any local client based on the local data set The model to be trained is t During the round of local model training, after completing the m The intermediate local model obtained after the local round of training is: ; In the formula, Indicates that the central server received by any local client is in ( ) The model to be trained is broadcasted during the training round; represents the learning rate when any local client performs local model training; F k () is the preset local loss function of federated learning; Represents the gradient calculation symbol; Indicates -1 The intermediate local model obtained after the local round of training; Indicates that any local client m Local data used by the local round.

[0008] In one possible design, any local client t Round training status value , when any local client continues to perform unfinished local model training, after completing the m The intermediate local model obtained after the local round of training is: ; In the formula, The initial global model represents the model being trained by any of the local clients.

[0009] In one possible design, any local client completes the t The training model obtained after a round of local model training is: ; In the formula, M Indicates the total number of local rounds of local model training performed by any local client; Indicates that any local client is based on a local data set The model to be trained is t During the round of local model training, after completing the M The intermediate local model obtained after the local round of training; Indicates that any local client continues to perform unfinished local model training. M The intermediate local model obtained after the local round of training.

[0010] In one possible design, the central server t The weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model ,include: The central server obtains the completion of the first t The local model update value after the round of local model training, and the central server t -1 round of model weighted aggregation processing of the global model update value, and calculate the cosine similarity factor between the local model update value and the global model update value; The central server is configured to execute the following operations according to the first t Round delay status value S k ( t ), calculate that any local client completes the t The delay discount factor of the training model obtained after a round of local model training; The central server calculates the weight aggregation parameter of any local client according to the cosine similarity factor and the delay discount factor; The central server performs a weight aggregation based on the weight aggregation parameter. t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model .

[0011] In one possible design, the cosine similarity factor between the local model update value and the global model update value is: ; In the formula, , complete the first t The local model update value after a round of local model training; , for the central server in t -Global model update value of 1 round of model weighted aggregation processing; Representation vector and vector The cosine of the angle between .

[0012] In one possible design, the delay discount factor is: ; In the formula, The preset expiration tolerance.

[0013] In a possible design, the weight aggregation parameter of any local client is: ; In the formula, is a cosine similarity factor between the local model update value and the global model update value; is the delay discount factor; and are the preset weight values ​​respectively.

[0014] In one possible design, the updated global model for: ; In the formula, , indicating that the central server is in t The sequence number set of local clients corresponding to all training models received in the round; p k is a weight aggregation parameter of any of the local clients.

[0015] In a second aspect, the present invention provides an electronic device, comprising: a memory for storing computer program instructions; and, A processor is used to execute the computer program instructions to complete the operation of the federated learning model training method as described in any one of the above.

[0016] In a third aspect, the present invention provides a computer program product, comprising a computer program or instructions, which, when executed by a computer, implements the federated learning model training method as described in any one of the above.

[0017] The beneficial effects of the present invention are: The present invention discloses a method for training a federated learning model, which can effectively deal with the negative impact of the heterogeneity of client data and computing power configuration commonly found in federated learning systems on the system training efficiency. It can fully utilize the characteristics of cloud-edge collaborative training in the distributed computing process based on the training model uploaded by the local client to achieve rapid convergence of the global model in the federated learning model training task. Specifically, during the implementation of the present invention, by setting the periodic aggregation time , and adopts a semi-asynchronous federated learning model training method to construct a semi-asynchronous federated learning aggregation solution with a fixed time period, which can solve the impact of bottleneck node delay in the synchronous federated aggregation strategy in the device heterogeneous environment, and also prevent the global model from being skewed due to frequent uploading of local models by some devices in the fully asynchronous federated aggregation strategy; in addition, in the environment of unbalanced computing power of the federated learning system, the present invention ensures that the data in all local clients can participate in the model training when the global training time is fixed through a semi-asynchronous aggregation method, thereby effectively improving the data utilization rate in the local clients participating in the training in the federated learning system and improving the convergence accuracy of the model training.

[0018] Other beneficial effects of the present invention will be further described in the specific embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of the federated learning model training method in Example 1; Figure 2 is a flow chart of the model weighted aggregation process in Example 1; Figure 3 is a schematic diagram of the training time period during the application of Example 1; Figure 4 It is a schematic diagram of the test results of the federated learning model training method in Example 1. DETAILED DESCRIPTION

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the present invention will be briefly introduced below in combination with the drawings and the description of the embodiments or the prior art. Obviously, the following description of the structure of the drawings is only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. It should be noted that the description of these embodiments is used to help understand the present invention, but does not constitute a limitation of the present invention.

[0021] Embodiment 1: This embodiment discloses a federated learning model training method, which is executed based on a federated learning system, wherein the federated learning system includes a central server and a plurality of local clients respectively connected to the central server in communication; Figure 1 As shown, the method may include but is not limited to the following steps: S1. The central server receives the federated learning model training task and sets the initial global model according to the federated learning model training task , and then the initial global model The model to be trained is broadcasted to multiple local clients; specifically, in this embodiment, the central server can set an initial global neural network model for a specific deep neural network model training task, which is not limited here. In this embodiment, after the local client receives the initial global model After that, the central server can send a confirmation message to the local client so that the local client can participate in the model training task. The central server obtains the number of local clients. K , that is, the number of local clients participating in the federated learning training, initialize the delay state value of all local clients ( ) and training status values ( ), and set the periodic aggregation duration Specifically, in this embodiment, the central server agrees on the periodic aggregation duration according to the overall situation of all local clients. .

[0022] S2. Any local client ( ) receives the central server at the ( ) round of training broadcasts the model to be trained, that is, receives the first t Wheel global model , and according to its t Round training status value Perform local model training; as described in any local client t Round training status value , then any local client based on the local data set The model to be trained is t round of local model training; as described in the first round of local model training of any local client t Round training status value , then any local client continues the unfinished local model training, that is, continues to complete the training in the first t The local model training tasks that have not been completed before the round starts; among them, , TThe total number of rounds of the model to be trained broadcasted by the central server, that is, the total number of training rounds when the federated learning model training task is completed.

[0023] Specifically, in step S2, any local client t Round training status value , any local client based on the local data set The model to be trained is t During the round of local model training, after completing the m The intermediate local model obtained after the local round of training is: ; In the formula, Indicates that the central server received by any local client is in ( ) The model to be trained is broadcasted during the training round; represents the learning rate when any local client performs local model training; F k () is the preset local loss function of federated learning; Represents the gradient calculation symbol; Indicates -1 The intermediate local model obtained after the local round of training; Indicates that any local client m Local data used by the local round.

[0024] In this embodiment, the total number of local rounds for any local client to perform local model training can be preset as M , that is, the number of rounds of local stochastic gradient descent training required for any local client to perform a model update process locally; where m=1,2,…, M .

[0025] In step S2, any local client t Round training status value , when any of the local clients continues to perform the unfinished local model training, assuming that any of the local clients performs the first m The local round of training is in t Within the periodic aggregation duration of the round of training, any local client completes the first m The intermediate local model obtained after the local round of training is: ; In the formula, The initial global model represents the model being trained by any of the local clients.

[0026] It should be noted that, at this time, it is assumed that any local client performs the m The local round of training is in t The aggregation duration of the training cycle.

[0027] S3. t Before the end of the periodic aggregation duration of the round of training, any local client has completed the first t round of local model training, then its first t+ 1 round of training status value Set to 1 to t+ 1 round delay status value Set to 0 and complete the t The training model obtained after the round of local model training is uploaded to the central server; in the t After the periodic aggregation time of the round of training ends, if any local client fails to complete the t round of local model training, then its first t+ 1 round of training status value Set it to 0 and t+ 1 round delay status value Set to .

[0028] Specifically, in step S3, any local client completes the t The training model obtained after a round of local model training is: ; In the formula, M Indicates the total number of local rounds of local model training performed by any local client; Indicates that any local client is based on a local data set The model to be trained is t During the round of local model training, after completing the M The intermediate local model obtained after the local round of training, that is, the local model to be uploaded after the local client that has not fallen behind has completed the local partial update; Indicates that any local client continues to perform unfinished local model training. M The intermediate local model obtained after the local round of training, that is, the one that falls behind in the global training The local client of the round completes the local local update and uploads the local model with outdated information.

[0029] S4. The central server receives t All training models of the round of training, that is, completing the first t The receiving task of the training model in round 1 is based on the participation in tThe weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model .

[0030] Specifically, in step S4, Figure 2 As shown, the central server participates in t The weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model ,include: S401. The central server obtains the completion of the first t The local model update value after the round of local model training, and the central server t -1 round of model weighted aggregation processing of the global model update value, and calculate the cosine similarity factor between the local model update value and the global model update value.

[0031] In step S401, the cosine similarity factor between the local model update value and the global model update value is: ; In the formula, , complete the first t The local model update value after a round of local model training; , for the central server in t -Global model update value of 1 round of model weighted aggregation processing; Representation vector and vector The cosine of the angle between .

[0032] S402. The central server receives the first t Round delay status value S k ( t ), calculate that any local client completes the t The training model obtained after a round of local model training The delay discount factor.

[0033] In step S402, the delay discount factor is: ; In the formula, It is a preset expiration tolerance, which is used to adjust the tolerance limit of the expiration degree of the training model uploaded by any local client during the model weighted aggregation processing.

[0034] S403. The central server calculates a weight aggregation parameter of any local client according to the cosine similarity factor and the delay discount factor.

[0035] In step S403, the weight aggregation parameter of any local client is: ; In the formula, is a cosine similarity factor between the local model update value and the global model update value; is the delay discount factor; and are preset weight values, which are used to adjust the influence of delay discount factor and model similarity on aggregation weight. For example, Set to 0.5, Set to 0.5.

[0036] S404. The central server performs the weight aggregation based on the weight aggregation parameters. t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model .

[0037] In step S404, the updated global model for: ; In the formula, , indicating that the central server is in t The sequence number set of local clients corresponding to all training models received in the round; p k is a weight aggregation parameter of any of the local clients.

[0038] Based on the above steps S401 to S404, this embodiment takes into account the delay information of the training model uploaded by the local client during the configuration of the weight aggregation parameters of the model weighted aggregation processing, and adjusts the aggregation weight of the model uploaded by the lagging local client by introducing the delay discount factor of the training model and the cosine similarity factor of the model, thereby alleviating the negative impact of the expired training model information on the convergence of the global model.

[0039] S5. The central server determines the updated global model Is it converged? If so, the updated global model as the trained global model, and end the federated learning model training task; if not, then update the updated global model The training model is broadcasted to multiple local clients. Specifically, in this embodiment, the central server obtains the updated global model. After that, the updated global model is tested using the preset test data set. Perform a performance convergence test to determine the updated global model Whether to converge.

[0040] This embodiment can effectively deal with the negative impact of the heterogeneity of client data and computing power configuration commonly found in federated learning systems on the system training efficiency. It can fully utilize the characteristics of cloud-edge collaborative training in the distributed computing process based on the training model uploaded by the local client to achieve rapid convergence of the global model in the federated learning model training task. Specifically, during the implementation of this embodiment, by setting the periodic aggregation time , and adopts a semi-asynchronous federated learning model training method to construct a semi-asynchronous federated learning aggregation solution with a fixed time period, which can solve the impact of bottleneck node delay in the synchronous federated aggregation strategy in the device heterogeneous environment, and also prevent the global model from being skewed due to frequent uploading of local models by some devices in the fully asynchronous federated aggregation strategy; in addition, in the environment of unbalanced computing power of the federated learning system, this embodiment uses a semi-asynchronous aggregation method to ensure that the data in all local clients can participate in the model training when the global training time is fixed, thereby effectively improving the data utilization rate in the local clients participating in the training in the federated learning system and improving the convergence accuracy of the model training.

[0041] As an example, Figure 3 A schematic diagram of the training time period of the method in the present embodiment during application is given. In addition, the method proposed in the present embodiment is tested in combination with specific data, and the process is as follows: Get the MNIST dataset (MNIST stands for Mixed National Institute of Standards and Technology database. The MNIST dataset is a binary image dataset used to train various image processing systems and is widely used in training and testing in machine learning). The MNIST dataset consists of 60,000 28*28 grayscale images. There are 10 categories in total, and each category consists of 6,000 handwritten digital images. The dataset is divided into 100 clients using Dirichlet distribution, so that the datasets between different local clients meet the non-independent and identically distributed assumption (Non-IID). In order to evaluate and verify the performance of the method proposed in this embodiment in terms of training accuracy and training speed, the currently popular classic federated learning aggregation algorithm FedAvg and the fully asynchronous federated learning algorithm FedAsync are used as benchmark test models, and the iteration of the machine learning loss function with the number of rounds is used as the evaluation indicator of the model. The test results are as follows: Figure 4 shown.

[0042] based on Figure 4 From the test results shown, it can be seen that this embodiment realizes the aggregation and update of the federated learning model by adopting a semi-asynchronous aggregation strategy starting from a fixed time period, prevents the bottleneck phenomenon of client devices by introducing an asynchronous mechanism, realizes the effectiveness of distributed intelligent computing by weighted aggregation based on the delay discount factor and the cosine similarity factor of the model, and ensures the rapid convergence of the model under data heterogeneity. This mechanism fully improves the participation of local clients in the training process and the data utilization rate in the federated learning system.

[0043] Embodiment 2: Based on Embodiment 1, this embodiment discloses an electronic device, which may be a smart phone, a tablet computer, a laptop computer, or a desktop computer, etc. The electronic device may be called a user terminal, a portable terminal, a desktop terminal, etc. The electronic device includes: a memory for storing computer program instructions; and, A processor is used to execute the computer program instructions to complete the operations of the federated learning model training method as described in any one of Example 1.

[0044] Embodiment 3: Based on any one of Examples 1 to 2, this embodiment discloses a computer program product, including a computer program or instructions, which, when executed by a computer, implements the federated learning model training method as described in any one of Example 1.

[0045] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0046] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the above embodiments, a person skilled in the art should understand that the technical solutions described in the above embodiments can still be modified, or some of the technical features can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for training a federated learning model, characterized in that: Based on a federated learning system, the federated learning system includes a central server and a plurality of local clients respectively connected to the central server in communication; the method includes: The central server receives the federated learning model training task and sets the initial global model according to the federated learning model training task , and then the initial global model The model to be trained is broadcasted to multiple local clients; the central server obtains the number of local clients K , initialize the delay status value of all local clients ( ) and training status values ( ), and set the periodic aggregation duration ; Any local client receives the central server at ( ) round of training, and according to its t Round training status value Perform local model training; as described in any local client t Round training status value , then any local client based on the local data set The model to be trained is t round of local model training; as described in the first round of local model training of any local client t Round training status value , then any local client continues the unfinished local model training; wherein, , T The central server broadcasts the total number of rounds of the model to be trained; In the t Before the end of the periodic aggregation duration of the round of training, any local client has completed the first t round of local model training, then its first t+ 1 round of training status value Set to 1 to make it t+ 1 round delay status value Set to 0 and complete the t The training model obtained after the round of local model training is uploaded to the central server; in the t After the periodic aggregation time of the round of training ends, if any local client fails to complete the t round of local model training, then its first t+ 1 round of training status value Set it to 0 and t+ 1 round delay status value Set to ; The central server receives the t All training models in round 1 training are divided into t The weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model ; The central server determines the updated global model Is it converged? If so, the updated global model as the trained global model, and end the federated learning model training task; if not, then update the updated global model The model is broadcasted to multiple local clients as a model to be trained.

2. A federated learning model training method according to claim 1, characterized in that: Any local client t Round training status value , any local client based on the local data set The model to be trained is t During the round of local model training, after completing the m The intermediate local model obtained after the local round of training is: ； In the formula, Indicates that the central server received by any local client is in ( ) The model to be trained is broadcasted during the training round; represents the learning rate when any local client performs local model training; F k () is the preset local loss function of federated learning; Represents the gradient calculation symbol; Indicates -1 The intermediate local model obtained after the local round of training; Indicates that any local client m Local data used by the local round.

3. A federated learning model training method according to claim 2, characterized in that: Any local client t Round training status value , when any local client continues to perform unfinished local model training, after completing the m The intermediate local model obtained after the local round of training is: ； In the formula, The initial global model represents the model being trained by any of the local clients.

4. A federated learning model training method according to claim 3, characterized in that: Any local client completes t The training model obtained after a round of local model training is: ； In the formula, M Indicates the total number of local rounds of local model training performed by any local client; Indicates that any local client is based on a local data set The model to be trained is t During the round of local model training, after completing the M The intermediate local model obtained after the local round of training; Indicates that any local client continues to perform unfinished local model training. M The intermediate local model obtained after the local round of training.

5. A federated learning model training method according to claim 1, characterized in that: The central server participates in t The weight aggregation parameters are set based on the specified information of the local client of the round of training, and the weight aggregation parameters are used to train the first t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model ,include: The central server obtains the completion of the first t The local model update value after the round of local model training, and the central server t -1 round of model weighted aggregation processing of the global model update value, and calculate the cosine similarity factor between the local model update value and the global model update value; The central server is configured to execute the following operations according to the first t Round delay status value S k ( t ), calculate that any local client completes the t The delay discount factor of the training model obtained after a round of local model training; The central server calculates the weight aggregation parameter of any local client according to the cosine similarity factor and the delay discount factor; The central server performs a weight aggregation based on the weight aggregation parameter. t All training models received in the round are subjected to model weighted aggregation processing to obtain the updated global model .

6. A federated learning model training method according to claim 5, characterized in that: The cosine similarity factor between the local model update value and the global model update value is: ； In the formula, , complete the first t The local model update value after a round of local model training; , for the central server in t -Global model update value of 1 round of model weighted aggregation processing; Representation vector and vector The cosine of the angle between .

7. A federated learning model training method according to claim 5, characterized in that: The delay discount factor is: ； In the formula, The preset expiration tolerance.

8. A federated learning model training method according to claim 5, characterized in that: The weight aggregation parameters of any local client are: ； In the formula, is a cosine similarity factor between the local model update value and the global model update value; is the delay discount factor; and are the preset weight values ​​respectively.

9. A federated learning model training method according to claim 5, characterized in that: The updated global model for: ； In the formula, , indicating that the central server is in t The sequence number set of local clients corresponding to all training models received in the round; p k is a weight aggregation parameter of any of the local clients.

Images