A proxy system in a federated learning platform

Abstract

The application discloses a kind of agent systems in joint learning platform, belong to computer distributed computing technology field, wherein local client and cloud are interactive, local client includes industry AI front-end background, industry AI training configuration module, local training agent and DT platform resource controller, cloud includes basic AI joint learning engine, basic AI ecological model library, industry AI database and DT computing decision module, DT computing decision module in cloud and local training agent in local client interact, local training agent and DT platform resource controller interact, local training agent also interacts with AI joint learning engine in cloud;Local model training module in local training agent runs first from DT platform resource controller to obtain monitoring information, then carries out this round iteration training, can very good to do local training and joint learning platform combination, improve the efficiency of local training.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer distributed computing, and particularly relates to a proxy system in a federated learning platform. BACKGROUND

[0002] Training data cannot be far from its source, that is, the server cannot collect data. The reasons may include privacy issues, such as unwillingness to upload photos; regulatory barriers (HIPAA, GDPR, etc.); and actual engineering limitations, such as expensive, slow or unreliable network connections, or too large data volume, etc. Federated learning can solve the above problems. Federated learning refers to a series of algorithms, although the server cannot collect data, but can collect the parameters of the model, the server coordinates the edge device to participate in training, each edge device has training data, each edge device trains a local model with its own data, and uploads the encrypted or unencrypted parameters to the server, and the server averages or weighted averages the collected parameters and broadcasts them to each edge device. When using the federated learning platform to perform local training proxy, the local training interval and the small amount of data are prone to occur, resulting in low overall local training efficiency. SUMMARY

[0003] The purpose of the present application is to provide a proxy system in a federated learning platform, which has the advantage of improving the efficiency of local training, to solve the problems raised in the background.

[0004] To achieve the above purpose, the present application provides the following technical scheme:

[0005] A proxy system in a federated learning platform, comprising a local client and a cloud, the local client comprising an industry AI front-end background, an industry AI training configuration module, a local training proxy and a DT platform resource controller, and the cloud comprising a basic AI federated learning engine, a basic AI ecological model library, an industry AI database and a DT computing decision module, the DT computing decision module being connected with the local training proxy, the local training proxy being interactive with the DT platform resource controller, and the local training proxy further being interactive with the AI federated learning engine.

[0006] The local training proxy comprises a proxy local deployment module, a training data acquisition module, a local model training module, an upload and download model, an abnormal processing strategy module, a monitoring information feedback module and a distributed computing module, the local model training module being interactive with the monitoring information feedback module.

[0007] Preferably, the local training proxy is deployed in the local client, and the local training proxy is a computing service. When a user uses data assets to participate in a specified federated learning task, a proxy instance will be executed for the local computing task.

[0008] Preferably, the local model training module in the local training agent interacts with the DT platform resource controller during the training process, adjusts the iteration strategy based on monitoring information, such as whether to perform the current iteration, and interacts with the AI ​​joint learning engine to upload and download data models.

[0009] Preferably, during the local training agent process, the agent is first started in the agent deployment module. After the local training agent is successfully deployed, the AI ​​joint learning engine starts the local training agent and then sends information such as the industry AI training configuration module and the data model and data volume in the industry AI database to the agent, that is, to the local training agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is user-configured, the user-configured algorithm is used.

[0010] Preferably, when the agent local deployment module deploys the local training agent, it interacts with the industry AI database and the industry AI training configuration module in advance with the input end of the AI ​​federated learning engine. The entry point of the industry AI database is not password-protected, and the AI ​​federated learning engine can use the algorithm configured by the user in the local training agent by default.

[0011] Preferably, the training data acquisition module obtains processed training data according to the data acquisition interface provided by the user. After the agent starts, it first calls the industry AI front-end and back-end API interface to obtain a certain training dataset based on parameters such as task name, algorithm name, and data type. At this time, the industry AI back-end will receive a training dataset request. After receiving the request, the industry AI back-end will query the corresponding industry AI database or local server. After finding the corresponding training data, the industry AI back-end will return the training data to the agent. During the process of the industry AI back-end finding the corresponding training data, the industry AI participant, i.e., the industry AI back-end, needs to perform a key-value mapping on the data. K is the interface input parameter used to obtain the corresponding V training data. The agent does not interfere with the data storage method, i.e., the data is converted into the algorithm requirements.

[0012] Preferably, the local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating the training. Obtaining monitoring information from the DT platform resource controller involves sending a request to the DT platform resource controller to obtain monitoring information such as CPU usage, memory usage, and disk space remaining. The agent uses this monitoring information to determine whether to continue the current joint iteration process. Then, the current round of iterative training is performed; that is, when the monitoring information meets the training criteria, the agent executes the current iteration of training, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. At this point, the AI ​​joint learning engine saves the aggregated intermediate model. Finally, training terminates; once the normal joint training ends and the termination conditions are met, the agent uploads the final model to the AI ​​joint learning engine, the AI ​​joint learning engine saves the model to the model library, and the agent calls the AI ​​joint learning engine's joint termination interface to end the training.

[0013] Preferably, in the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If it is a non-voluntary exit, such as a network error or node interruption, no check is needed, and training can continue directly after recovery. If it is a voluntary exit, the agent needs to be rejoined in the task process.

[0014] Preferably, during the operation of the monitoring information feedback module, the agent feeds back the monitoring information obtained from the platform resource controller before the current iteration, such as CPU usage, memory usage, disk space, etc., and the heartbeat information and model error obtained after the iteration, to the industry AI backend.

[0015] Preferably, when the distributed computing module runs, it first issues a request, that is, it issues a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. The responding targets form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations. Every once in a while, the jointly computed dataset is merged and uploaded, that is, the model is uploaded to the local model training module.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: The agent system in the federated learning platform proposed in this invention allows for interaction between a local client and the cloud. The local client includes an industry AI front-end / back-end, an industry AI training configuration module, a local training agent, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computational decision module. The DT computational decision module in the cloud interacts with the local training agent in the local client, the local training agent interacts with the DT platform resource controller, and the local training agent also interacts with the AI ​​federated learning engine in the cloud. When the local model training module in the local training agent runs, it first obtains monitoring information from the DT platform resource controller and then performs the current round of iterative training. This effectively combines local training with the federated learning platform, improving the efficiency of local training. Attached Figure Description

[0017] Figure 1 This is a block diagram showing the overall position of the present invention;

[0018] Figure 2 This is a block diagram of the local training agent of the present invention;

[0019] Figure 3 This is a block diagram illustrating the principle of the agent local deployment module of the present invention;

[0020] Figure 4 This is a block diagram illustrating the principle of the training data acquisition module of the present invention.

[0021] Figure 5 This is a block diagram illustrating the principle of the local model training module of the present invention.

[0022] Figure 6 This is a block diagram illustrating the principle of the exception handling strategy module of the present invention.

[0023] Figure 7 This is a block diagram illustrating the principle of the distributed computing module of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] Example 1

[0026] Please see the appendix Figure 1A proxy system in a federated learning platform includes a local client and a cloud. The local client includes an industry AI front-end and back-end, an industry AI training configuration module, a local training proxy, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computing and decision module. The DT computing and decision module is connected to the local training proxy, the local training proxy interacts with the DT platform resource controller, and the local training proxy also interacts with the AI ​​federated learning engine.

[0027] Please see the appendix Figure 2 The local training agent includes a local deployment module, a training data acquisition module, a local model training module, a model upload and download module, an exception handling strategy module, a monitoring information feedback module, and a distributed computing module. The local model training module and the monitoring information feedback module interact with the local model training module. The local training agent is deployed on the local client and is a computing service. When a user uses data assets to participate in a specified joint learning task, a proxy instance will execute the local computing task.

[0028] During the training process, the local model training module in the local training agent interacts with the DT platform resource controller and adjusts the iteration strategy based on monitoring information, such as whether to perform the current iteration. The local model training module also interacts with the AI ​​joint learning engine to upload and download data models.

[0029] Please see the appendix Figure 3 During the local training agent deployment process, the agent is started first. After the local training agent is successfully deployed, the AI ​​federated learning engine starts the local training agent and sends information such as the data model and data volume from the industry AI training configuration module and the industry AI database to the agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is a user-configured algorithm, the user-configured algorithm will be used. When deploying the local training agent, the agent's local deployment module interacts with the industry AI database and the industry AI training configuration module beforehand. The entry point of the industry AI database does not have a password, and the AI ​​federated learning engine can default to using the user-configured algorithm in the local training agent.

[0030] Please see the appendix Figure 4The training data acquisition module obtains processed training data based on the user-provided data acquisition interface. After the agent starts, it first calls the industry AI front-end and back-end API interfaces to obtain a certain training dataset based on parameters such as task name, algorithm name, and data type. At this time, the industry AI back-end will receive the training dataset request. After receiving the request, the industry AI back-end will query the corresponding industry AI database or local server. After finding the corresponding training data, the industry AI back-end will return the training data to the agent. During the process of the industry AI back-end finding the corresponding training data, the industry AI participants, i.e., the industry AI back-end, need to perform a key-value mapping on the data. K is the interface input parameter used to obtain the corresponding V training data. The agent does not interfere with the data storage method, i.e., the data is converted into the algorithm requirements.

[0031] Please see the appendix Figure 5 The local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating training. Obtaining monitoring information from the DT platform resource controller involves sending requests to the controller for monitoring data such as CPU usage, memory usage, and disk space. The agent uses this information to determine whether to continue the current joint iteration process. Then, the agent performs the current round of iterative training. When the monitored information meets the training criteria, the agent executes the current iteration, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. At this point, the AI ​​joint learning engine saves the aggregated intermediate model. Finally, training terminates. Once the joint training is complete and the termination conditions are met, the agent uploads the final model to the AI ​​joint learning engine, which saves the model to its model library. The agent then calls the AI ​​joint learning engine's joint termination interface to end the training.

[0032] Please see the appendix Figure 6 In the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If it is an involuntary exit, such as a network error or node interruption, no check is needed, and training can continue directly after recovery. If it is an involuntary exit, the agent needs to be rejoined to the task process.

[0033] During the operation of the monitoring information feedback module, the agent feeds back the monitoring information obtained from the platform resource controller before the current iteration, such as CPU usage, memory usage, disk space, etc., as well as the heartbeat information and model error obtained after the iteration, to the industry AI backend.

[0034] Please see the appendix Figure 7When the distributed computing module runs, it first sends a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. Targets that respond form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations. Every once in a while, the jointly computed dataset is merged and uploaded, that is, the model is uploaded to the local model training module.

[0035] Example 2

[0036] Please see the appendix Figure 1 A proxy system in a federated learning platform includes a local client and a cloud. The local client includes an industry AI front-end and back-end, an industry AI training configuration module, a local training proxy, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computing and decision module. The DT computing and decision module is connected to the local training proxy, the local training proxy interacts with the DT platform resource controller, and the local training proxy also interacts with the AI ​​federated learning engine.

[0037] Please see the appendix Figure 2 The local training agent includes a local deployment module, a training data acquisition module, a local model training module, a model upload and download module, an exception handling strategy module, a monitoring information feedback module, and a distributed computing module. The local model training module and the monitoring information feedback module interact with the local model training module. The local training agent is deployed on the local client and is a computing service. When a user uses data assets to participate in a specified joint learning task, a proxy instance will execute the local computing task.

[0038] During the training process, the local model training module in the local training agent interacts with the DT platform resource controller and adjusts the iteration strategy based on monitoring information, such as whether to perform the current iteration. The local model training module also interacts with the AI ​​joint learning engine to upload and download data models.

[0039] Please see the appendix Figure 3 During the local training process, the agent deployment module first starts the agent. After the local training agent is successfully deployed, the AI ​​federated learning engine starts the local training agent and then sends information such as the industry AI training configuration module and the data model and data volume in the industry AI database to the agent, i.e., to the local training agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is a user-configured algorithm, then the user-configured algorithm will be used.

[0040] Please see the appendix Figure 4The training data acquisition module obtains processed training data based on the user-provided data acquisition interface. After the agent starts, it first retrieves data directly from the industry AI database or local server based on metadata information such as server address, database name, and table name provided by the industry user, and integrates it into a training dataset. This process does not require the agent to obtain server address or database permissions.

[0041] Please see the appendix Figure 5 The local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating training. Obtaining monitoring information from the DT platform resource controller involves sending requests to the controller for monitoring data such as CPU usage, memory usage, and disk space. The agent uses this information to determine whether to continue the current joint iteration process. Then, the agent performs the current round of iterative training. When the monitored information meets the training criteria, the agent executes the current iteration, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. At this point, the AI ​​joint learning engine saves the aggregated intermediate model. Finally, training terminates. Once the joint training is complete and the termination conditions are met, the agent uploads the final model to the AI ​​joint learning engine, which saves the model to its model library. The agent then calls the AI ​​joint learning engine's joint termination interface to end the training.

[0042] Please see the appendix Figure 6 In the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If it is an involuntary exit, such as a network error or node interruption, no check is needed, and training can continue directly after recovery. If it is an involuntary exit, the agent needs to be rejoined to the task process.

[0043] During the operation of the monitoring information feedback module, the agent feeds back the monitoring information obtained from the platform resource controller before the current iteration, such as CPU usage, memory usage, disk space, etc., as well as the heartbeat information and model error obtained after the iteration, to the industry AI backend.

[0044] Please see the appendix Figure 7 When the distributed computing module runs, it first sends a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. Targets that respond form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations. Every once in a while, the jointly computed dataset is merged and uploaded, that is, the model is uploaded to the local model training module.

[0045] Example 3

[0046] Please see the appendix Figure 1 A proxy system in a federated learning platform includes a local client and a cloud. The local client includes an industry AI front-end and back-end, an industry AI training configuration module, a local training proxy, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computing and decision module. The DT computing and decision module is connected to the local training proxy, the local training proxy interacts with the DT platform resource controller, and the local training proxy also interacts with the AI ​​federated learning engine.

[0047] Please see the appendix Figure 2 The local training agent includes a local deployment module, a training data acquisition module, a local model training module, a model upload and download module, an exception handling strategy module, a monitoring information feedback module, and a distributed computing module. The local model training module and the monitoring information feedback module interact with the local model training module. The local training agent is deployed on the local client and is a computing service. When a user uses data assets to participate in a specified joint learning task, a proxy instance will execute the local computing task.

[0048] During the training process, the local model training module in the local training agent interacts with the DT platform resource controller and adjusts the iteration strategy based on monitoring information, such as whether to perform the current iteration. The local model training module also interacts with the AI ​​joint learning engine to upload and download data models.

[0049] Please see the appendix Figure 3 During the local training process, the agent deployment module first starts the agent. After the local training agent is successfully deployed, the AI ​​federated learning engine starts the local training agent and then sends information such as the industry AI training configuration module and the data model and data volume in the industry AI database to the agent, i.e., to the local training agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is a user-configured algorithm, then the user-configured algorithm will be used.

[0050] Please see the appendix Figure 4 The training data acquisition module obtains processed training data based on the data acquisition interface provided by the user. After the agent starts, it first obtains data directly from the industry AI database or local server based on the metadata information such as server address, database name, and table name provided by the industry user, and integrates it into a training dataset without the agent needing to obtain server address permissions or database permissions.

[0051] Please see the appendix Figure 5The local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating training. Obtaining monitoring information from the DT platform resource controller involves sending requests to the controller for monitoring data such as CPU usage, memory usage, and disk space. The agent uses this information to determine whether to continue the current joint iteration process. Then, the agent performs the current round of iterative training. When the monitored information meets the training criteria, the agent executes the current iteration, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. The AI ​​joint learning engine then saves the aggregated intermediate model. Finally, training terminates. Alternatively, the industry backend can actively end training. Based on the acquired monitoring information, such as reaching a certain accuracy, the user can tell the agent to actively exit the training information. In this case, the industry backend calls the agent's termination training interface, which in turn calls the AI ​​joint learning engine's exit joint training interface to remove the participant from the joint training list, thus ending the agent's training.

[0052] Please see the appendix Figure 6 In the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If it is an involuntary exit, such as a network error or node interruption, no check is needed, and training can continue directly after recovery. If it is an involuntary exit, the agent needs to be rejoined to the task process.

[0053] During the operation of the monitoring information feedback module, the agent feeds back the monitoring information obtained from the platform resource controller before the current iteration, such as CPU usage, memory usage, disk space, etc., as well as the heartbeat information and model error obtained after the iteration, to the industry AI backend.

[0054] Please see the appendix Figure 7 When the distributed computing module runs, it first sends a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. Targets that respond form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations. Every once in a while, the jointly computed dataset is merged and uploaded, that is, the model is uploaded to the local model training module.

[0055] Example 4

[0056] Please see the appendix Figure 1A proxy system in a federated learning platform includes a local client and a cloud. The local client includes an industry AI front-end and back-end, an industry AI training configuration module, a local training proxy, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computing and decision module. The DT computing and decision module is connected to the local training proxy, the local training proxy interacts with the DT platform resource controller, and the local training proxy also interacts with the AI ​​federated learning engine.

[0057] Please see the appendix Figure 2 The local training agent includes a local deployment module, a training data acquisition module, a local model training module, a model upload and download module, an exception handling strategy module, a monitoring information feedback module, and a distributed computing module. The local model training module and the monitoring information feedback module interact with the local model training module. The local training agent is deployed on the local client and is a computing service. When a user uses data assets to participate in a specified joint learning task, a proxy instance will execute the local computing task.

[0058] During the training process, the local model training module in the local training agent interacts with the DT platform resource controller and adjusts the iteration strategy based on monitoring information, such as whether to perform the current iteration. The local model training module also interacts with the AI ​​joint learning engine to upload and download data models.

[0059] Please see the appendix Figure 3 During the local training process, the agent deployment module first starts the agent. After the local training agent is successfully deployed, the AI ​​federated learning engine starts the local training agent and then sends information such as the industry AI training configuration module and the data model and data volume in the industry AI database to the agent, i.e., to the local training agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is a user-configured algorithm, then the user-configured algorithm will be used.

[0060] Please see the appendix Figure 4 The training data acquisition module obtains processed training data based on the data acquisition interface provided by the user. In the process of finding the corresponding training data in the industry AI backend, the industry AI participants, i.e., the industry AI backend, need to perform a K-V mapping on the data. K is the interface input parameter used to obtain the corresponding V training data. The agent does not interfere with the data storage method, i.e., the data is converted into the algorithm requirements.

[0061] Please see the appendix Figure 5The local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating training. Obtaining monitoring information from the DT platform resource controller involves sending requests to the controller for monitoring data such as CPU usage, memory usage, and disk space. The agent uses this information to determine whether to continue the current joint iteration process. Then, the agent performs the current round of iterative training. When the monitored information meets the training criteria, the agent executes the current iteration, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. The AI ​​joint learning engine then saves the aggregated intermediate model. Finally, training terminates. Alternatively, the industry backend can actively end training. Based on the acquired monitoring information, such as reaching a certain accuracy, the user can tell the agent to actively exit the training information. In this case, the industry backend calls the agent's termination training interface, which in turn calls the AI ​​joint learning engine's exit joint training interface to remove the participant from the joint training list, thus ending the agent's training.

[0062] Please see the appendix Figure 6 In the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If it is an involuntary exit, such as a network error or node interruption, no check is needed, and training can continue directly after recovery. If it is an involuntary exit, the agent needs to be rejoined to the task process.

[0063] During the operation of the monitoring information feedback module, the agent feeds back the monitoring information obtained from the platform resource controller before the current iteration, such as CPU usage, memory usage, disk space, etc., as well as the heartbeat information and model error obtained after the iteration, to the industry AI backend.

[0064] Please see the appendix Figure 7 When the distributed computing module runs, it first sends a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. Targets that respond form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations. Every once in a while, the jointly computed dataset is merged and uploaded, that is, the model is uploaded to the local model training module.

[0065] In summary, the agent system in the federated learning platform proposed in this invention involves interaction between a local client and the cloud. The local client includes an industry AI front-end / back-end, an industry AI training configuration module, a local training agent, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computational decision module. The DT computational decision module in the cloud interacts with the local training agent in the local client, the local training agent interacts with the DT platform resource controller, and the local training agent also interacts with the AI ​​federated learning engine in the cloud. When the local model training module in the local training agent runs, it first obtains monitoring information from the DT platform resource controller and then performs the current round of iterative training. This effectively combines local training with the federated learning platform, improving the efficiency of local training.

[0066] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0067] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An agent system in a federated learning platform, comprising a local client and a cloud, characterized in that, The local client includes an industry AI front-end and back-end, an industry AI training configuration module, a local training agent, and a DT platform resource controller. The cloud includes a basic AI federated learning engine, a basic AI ecosystem model library, an industry AI database, and a DT computing and decision-making module. The DT computing and decision-making module is connected to the local training agent, and the local training agent interacts with the DT platform resource controller. The local training agent also interacts with the AI ​​federated learning engine. The local training agent includes a local deployment module, a training data acquisition module, a local model training module, a model upload and download module, an exception handling strategy module, a monitoring information feedback module, and a distributed computing module. The local model training module and the monitoring information feedback module interact with the local model training module. The local model training module in the local training agent interacts with the DT platform resource controller during the training process, adjusts the iteration strategy based on monitoring information, and interacts with the AI ​​joint learning engine to upload and download data models. In the event of a node interruption, the exception handling strategy module determines whether the agent should rejoin the training task. If the interruption is due to non-voluntary exit, network abnormality, or node interruption, no further check is required, and training can continue directly after recovery. If you exit voluntarily, you need to rejoin the task process; During the local training agent process, the agent is first started. After the local training agent is successfully deployed, the AI ​​joint learning engine starts the local training agent and then sends information such as the data model and data volume of the industry AI training configuration module and the industry AI database to the agent, that is, to the local training agent. If the local training agent has an algorithm, it can be called directly. If the algorithm built into the local training agent is user-configured, the user-configured algorithm is used. When deploying the local training agent, the agent local deployment module interacts with the industry AI database and the industry AI training configuration module in advance with the input of the AI ​​federated learning engine. The industry AI database does not have a password, and the AI ​​federated learning engine defaults to using the algorithm configured by the user in the local training agent.

2. The agent system in a federated learning platform according to claim 1, characterized in that, The local training agent is deployed on the local client. The local training agent is a computing service. When a user uses data assets to participate in a specified joint learning task, a proxy instance will execute the local computing task.

3. The agent system in a federated learning platform according to claim 1, characterized in that, The training data acquisition module obtains processed training data based on the user-provided data acquisition interface. After the agent starts, it first calls the industry AI front-end and back-end API interface to obtain a certain training dataset based on the task name, algorithm name, and data type parameters. At this time, the industry AI back-end will receive the training dataset request. After receiving the request, the industry AI back-end will query the corresponding industry AI database or local server. After finding the corresponding training data, the industry AI back-end will return the training data to the agent. During the process of the industry AI back-end finding the corresponding training data, the industry AI participant, i.e., the industry AI back-end, needs to perform a key-value mapping on the data. K is the interface input parameter used to obtain the corresponding V training data. The agent does not interfere with the data storage method, i.e., the data is converted into the algorithm requirements.

4. The agent system in a federated learning platform according to claim 1, characterized in that, The local model training module operates through three steps: obtaining monitoring information from the DT platform resource controller, performing the current round of iterative training, and finally terminating the training. Obtaining monitoring information from the DT platform resource controller involves sending requests to the controller to obtain monitoring information such as CPU usage, memory usage, and disk space. The agent uses this information to determine whether to continue the current joint iteration process. Then, the current round of iterative training is performed. When the monitored information meets the training criteria, the agent executes the current iteration, uploads the trained model to the AI ​​joint learning engine, and downloads the aggregated model from the AI ​​joint learning engine. At this point, the AI ​​joint learning engine saves the aggregated intermediate model. Finally, training terminates. Once the joint training is successfully completed and the termination conditions are met, the agent uploads the final model to the AI ​​joint learning engine, which saves the model to its model library. The agent then calls the AI ​​joint learning engine's joint termination interface to end the training.

5. The agent system in a federated learning platform according to claim 1, characterized in that, During the operation of the monitoring information feedback module, the agent will receive the monitoring information obtained from the platform resource controller before this iteration.

6. The agent system in a federated learning platform according to claim 1, characterized in that, When the distributed computing module runs, it first issues a request to the AI ​​database module, the industry AI training configuration module, and the basic AI ecosystem model library, requesting them to perform joint computation. Targets that respond form a group with the distributed computing module and then perform joint computation. Different targets perform different data computations, and every once in a while, the jointly computed dataset is merged and uploaded, i.e., the model is uploaded to the local model training module.

Images