Cluster-based federated learning reservation platform, reservation system and reservation method

Abstract

To provide a cluster-based associative learning reservation platform, a reservation system and a reservation method.SOLUTION: A reservation platform P, which is used by a first user end to input reservation information and perform training, includes a second user end having a data set, a main server and a slave server. The slave server is instructed to operate by the main server, which includes a service server. The slave server is composed of a backup host, and the service server receives the reservation information and then communicates with the backup host and the first user end, and the backup host includes a training end where the second user end is used to store a data set. After the first user end inputs the reservation information, the first user end and the backup host train synchronously, which greatly improves a success rate of the training.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to Federated Learning, and more particularly to a cluster-based federated learning booking platform, a booking system and a method thereof. [Background technology]

[0002] Federated learning is a machine learning technology. In traditional centralized machine learning, all relevant data is centralized on the same computing device during training to generate a model. In contrast, federated learning allows multiple devices to train a machine learning model without transmitting training data. Federated learning combines multi-node operations in a distributed system to distribute the work of training a model to multiple nodes. Summary of the Invention [Problem to be solved by the invention]

[0003] Distributed systems mainly distribute the computational load to relevant units when processing large amounts of data. However, in federated learning, in order to strengthen data privacy protection, training data is not concentrated on a single server but is instead distributed for computation, and each device trains the model directly.

[0004] Federated learning has a high advantage in terms of privacy, but it depends on the computing power of each device. Therefore, when training is performed simultaneously on multiple devices, the device with higher computing power must wait for the training of the device with lower computing power to finish before it can perform the next training, due to the difference in computing power between the devices. Furthermore, if the training of any device fails, all devices cannot complete the training. Therefore, the failure rate of training increases obviously as the number of devices increases, and other users are less motivated to participate in the training.

[0005] In view of the above circumstances, unlike the conventional federated learning platform, which uses devices with different locations and capabilities to perform the training, the present invention provides a federated learning reservation platform that provides services by combining the methods of using devices with different locations and capabilities and devices with the same capabilities but in different locations, and by adopting clusters, it realizes the feature of using devices with different capabilities in the same location, which greatly reduces the failure rate of training and allows participating users to reserve training anytime, anywhere.

[0006] The present invention is made by the inventors in consideration of the above problems and their main purpose is to provide a cluster-based federated learning reservation platform, that is, a platform for user ends to make reservations, at least one user end stores a data set in a backup host of the training end, when the training end receives reservation information from other user ends, the training end communicates with the backup host and the hosts of other user ends, so that scheduling can be made more flexible.

[0007] Another object of the present invention is to provide a cluster-based federated learning reservation system, which includes a first user end, a second user end, and a training end. After the training end obtains the data set of the second user end, when the first user end submits reservation information, the training end controls the host of the first user end and the data set backed up at the training end to perform training synchronously, so as to greatly improve the training efficiency.

[0008] Another object of the present invention is to provide a cluster-based federated learning reservation method, in which after the training end obtains the reservation information, the training end transmits the control information and the training model to the corresponding user end host and the backup host of the training end, which respectively carry out training, and the training end collects the training models respectively, and generates and returns a final training model, thereby improving the success rate of training.

[0009] In order to achieve the above-mentioned objective, according to one embodiment of the present invention, a cluster-based federated learning reservation platform is provided, which includes: a first user end inputs reservation information to perform training; a second user end having a data set; and a training end including a main server and a slave server, the slave server being instructed to operate by the main server, the main server including a service server, the slave server being a backup host, the service server receiving the reservation information and communicating with the backup host and the first user end based thereon, and the backup host being used by the second user end to store the data set.

[0010] In a preferred embodiment of the present invention, the training end includes a reservation interface through which the first user end transmits the reservation information, and the service server receives the reservation information and communicates with the backup host and the first user end based on the reservation information.

[0011] In a preferred embodiment of the present invention, the service server communicates and encrypts with the first user end through a first transmission protocol, and the first transmission protocol is HTTPS (HyperText Transfer Protocol Secure).

[0012] In a preferred embodiment of the present invention, the reservation interface communicates and encrypts with the service server through a second transmission protocol, and the second transmission protocol is SSH (Secure Shell Protocol).

[0013] In order to achieve the above-mentioned other object, according to one embodiment of the present invention, a cluster-based federated learning reservation system is provided, which includes a training end including a first user end for inputting reservation information, a second user end having a data set, a main server, a slave server, and a reservation interface, the slave server is instructed to operate by the main server, the main server includes a service server, the slave server is a backup host, the service server receives the reservation information through the reservation interface and communicates with the backup host and the first user end based on the reservation information, and the backup host is used by the second user end to store the data set.

[0014] In a preferred embodiment of the present invention, the service server transmits corresponding control information and a first initial training model to the backup host and the first user end based on the reservation information.

[0015] In a preferred embodiment of the present invention, the backup host performs training using the first initial training model and the data set based on the control information, generates a first training model, and returns the training model to the service server. The first user end performs training using the first initial training model based on the control information, generates a second training model, and returns the training model to the service server. The service server receives the first training model, calculates the second training model based on the first training model, generates a third training model, and returns the training model to the backup host and the first user end.

[0016] In a preferred embodiment of the present invention, when the first user end re-enters the reservation information into the reservation interface, the service server performs a test using the first training model, the second training model, and the third training model to generate a test result, and the test result is a second initial training model. The service server transmits control information and the second initial training model to the backup host and the first user end based on the reservation information, and the backup host performs training using the second initial training model and the data set based on the control information, generates a fourth training model, and returns it to the service server. Also, the first user end performs training using the second initial training model based on the control information, generates a fifth training model, and returns it to the service server, and the service server performs calculations based on the fourth training model and the fifth training model, generates a sixth training model, and returns it to the backup host and the first user end.

[0017] In a preferred embodiment of the invention, the quantity of said second user ends is equal to or greater than the quantity of said first user ends.

[0018] In order to achieve the above-mentioned further object, according to one embodiment of the present invention, a cluster-based federated learning reservation method is provided, which includes the steps of: when a service server at a training end receives reservation information from a first user end, the service server communicates with a backup host at the training end and the first user end based on the reservation information, the backup host is instructed to operate by the service server, and the backup host obtains a data set from a second user end; transmitting control information and a first initial training model from the service server to the backup host and the first user end; based on the control information, the backup host trains using the first initial training model and the data set, generates a first training model and returns it to the service server; based on the control information, the first user end trains using the first initial training model, generates a second training model and returns it to the service server; and calculating the first training model and the second training model from the service server, generates a third training model and returns it to the backup host and the first user end, and the second user end obtains the third training model through the backup host.

[0019] The service server performs testing using the first training model, the second training model, and the third training model to generate a test result, and sets the test result as a second initial training model. When the service server receives the reservation information again from the first user end, the service server transmits the control information and the second initial training model to the backup host and the first user end. Effect of the Invention

[0020] As described above, the present invention has the following advantages. The present invention provides services in a manner in which they are executed simultaneously by devices with different capabilities at different locations and by devices with different capabilities at the same location, and by employing clusters to realize the manner in which they are executed by devices with different capabilities at the same location, thereby greatly improving the success rate of training and enabling more flexible and efficient scheduling.

[0021] Other objects, configurations and effects of the present invention will become apparent from the following detailed description of the preferred embodiments of the present invention. [Brief description of the drawings]

[0022] [Figure 1] 1 illustrates an example of a configuration according to an embodiment of the present invention. [Diagram 2] 1 illustrates a schematic diagram of an example platform for an embodiment of the present invention. [Diagram 3] 1 illustrates a schematic diagram of an example of a system according to an embodiment of the present invention. [Figure 4] 2 shows a flow chart of a method according to an embodiment of the present invention. [Diagram 5] FIG. 2 is a schematic diagram showing steps of one embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0023] The present invention will be described below through embodiments of the invention, but the following embodiments do not limit the scope of the invention according to the claims. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention.

[0024] FIG. 1 is a schematic diagram of an example of the configuration of an embodiment of the present invention. FIG. 2 is a schematic diagram of an example of a platform of an embodiment of the present invention. As shown in FIG. 1, an embodiment of the present invention mainly realizes a service by using clusters to execute services on devices with different performance at the same location. The second user end U2 stores its data set U21 in a backup host T21. In addition, the reservation platform is used by the user end to reserve training, and its operation method will be described in detail. The reservation platform according to an embodiment of the present invention is used by the first user end U1 to input reservation information and make a reservation, and the reservation platform P includes the second user end U2 and the training end T.

[0025] In one embodiment, the first user end U1 can enter reservation information to reserve training via the host, server, or any electronic device capable of performing training at the first user end U1. The number of first user ends U1 is not limited and may be one or more. The reservation information includes the reservation date and time, but the present invention is not limited thereto.

[0026] In one embodiment, the second user end U2 has a data set U21. Preferably, the data set U21 can be stored in any host, server, or cloud end storage location of the second user end U2. The number of the second user ends U2 is not limited and may be one or more. Preferably, the number of the second user ends U2 is equal to or greater than the number of the first user ends U1, but the present invention is not limited thereto.

[0027] In one embodiment, the training end T comprises a main server T1 and a slave server T2, the main server includes a service server T11, and the slave server T2 includes a backup host T21. After the service server T11 receives the reservation information, it communicates with the backup host T21 and the first user end U1 according to the reservation information, and the backup host T21 is used by the second user end U2 to store its data set U21. The service server T11 communicates with and encrypts the first user end U1 according to a first transmission protocol, and the first transmission protocol may be HTTPS (HyperText Transfer Protocol Secure).

[0028] Preferably, the main server T1 and the slave server T2 are a cluster, so that the slave server T2 is instructed to operate by the main server T1. The number of servers is not limited.

[0029] Preferably, the training end T further includes a reservation interface T3, the first user end U1 inputs reservation information through the reservation interface T3, and the service server T11 can transmit corresponding control information and the first initial training model to the backup host T21 and the first user end U1 according to the reservation information. In one embodiment, the reservation interface T3 may be a reservation web page provided by another web server, or may be a reservation web page provided by the service server T11.

[0030] Preferably, the reservation interface T3 communicates and encrypts with the service server T11 through a second transmission protocol, which may be SSH (Secure Shell Protocol), but the present invention is not limited thereto.

[0031] In one embodiment, the reservation platform is applicable to, but not limited to, the medical industry, financial industry, or other private data that needs to be trained.

[0032] In one embodiment, Fig. 3 is a schematic diagram of an example of a system according to one embodiment of the present invention. As shown in Fig. 3, a reservation system S according to one embodiment of the present invention includes a first user end U1, a second user end U2, and a training end T. The first user end U1 and the second user end U2 are the same as those in the above-mentioned embodiment, so the description thereof will not be repeated. The operation of the system will be described below. Both the first user end U1 and the second user end U2 store the training data in a designated storage location at any host or cloud end. In one embodiment, the first user end U1 does not store the training data at the training end T, and the data set U21 of the second user end U2 is stored at the training end T, or the backup host T21 of the training end T is used as a space for storing the data set U21.

[0033] In another embodiment, the first user end U1 may store the data to be trained at the training end T, or may use the backup host T21 of the training end T as a space for storing the data. That is, the data of the first user end U1 and the second user end U2 may be trained in a cluster of the training end T, but the present invention is not limited thereto.

[0034] The training end T includes a main server T1, a slave server T2, and a reservation interface T3. In one embodiment, the main server T1 includes a service server T11, and the slave server T2 includes a backup host T21. The main server T1 and the slave server T2 are a cluster, that is, the slave server T2 is instructed to operate by the main server T1, but the present invention is not limited thereto.

[0035] Preferably, the first user end U1 inputs reservation information through the reservation interface T3, and after the service server T11 receives the reservation information through the reservation interface T3, the service server T11 transmits corresponding control information and a first initial training model to the backup host T21 and the first user end U1 based on the reservation information.

[0036] In one embodiment, the second user end U2 inputs reservation information through a reservation interface T3, but the present invention is not limited thereto.

[0037] Preferably, the control information includes an application program, and the backup host T21 and the first user end U1 are connected to the service server T11 through a security mechanism. In one embodiment, when the service server T11 is connected to the backup host T21, the service server T11 needs to have a host account of the backup host T21, and the host account is used as the authentication mechanism, but the present invention is not limited thereto.

[0038] In one embodiment, after the backup host T21 and the first user end U1 are installed according to the control information, the backup host T21 performs training according to the first initial training model and the data set U21, generates a first training model and sends it back to the service server T11; the first user end U1 also performs training according to the first initial training model, generates a second training model and sends it back to the service server T11; the service server T11 receives the first training model and the second training model, performs calculations according to them, generates a third training model and sends it back to the backup host T21 and the first user end U1.

[0039] In one embodiment, after the backup host T21 and the first user end U1 receive the third training model, they respectively train the third training model, generate a fourth training model and a fifth training model and send them back to the service server T11, the service server T11 receives the fourth training model and the fifth training model, performs calculations based on them, generates a sixth training model and sends them back to the backup host T21 and the first user end U1, and continues the above-mentioned flowchart until the training is completed.

[0040] In one embodiment, after the training is completed, the first user end U1 and the second user end U2 are notified by email, and the first user end U1 and the second user end U2 can respectively retrieve the final training model from the host, server, electronic device used for training, or a file folder at the storage location on the cloud end.

[0041] The second user end U2 uses the backup host T21 of the training end T as a personal rental space and shares data and uses this space to learn. In other words, the second user end U2 stores data in the backup host T21, or directly uses the backup host T21 as a space for storing and using primary data, which not only eliminates privacy issues but also is not limited by the computing power of the host, server, or electronic device used for training at the second user end U2, and further enhances the flexibility and efficiency of federated learning.

[0042] In one embodiment, after the training is completed, the service server T11 tests all the existing training models and takes the generated test result as the next initial model. For example, if all the existing training models are the first and second training models respectively returned by the backup host T21 and the first user end U1, and the third training model generated by the service server T11 by performing calculations based on the first and second training models, the service server T11 tests the first, second, and third training models and takes the test result as the second initial training model. Similarly, if the backup host T21 and the first user end U1 further return a fourth and fifth training model, and the service server T11 performs calculations based on the fourth and fifth training models to generate a sixth training model, the service server T11 tests the first, second, third, fourth, fifth, and sixth training models and takes the test result as the second initial training model. In other words, by continuing training using a model that is suitable for the user end as the initial model for the next training, it is possible to prevent the need to repeat training and improve training efficiency.

[0043] In one embodiment, when the first user end U1 re-enters reservation information into the reservation interface T3, the service server T11 transmits control information and a second initial training model generated after the previous training to the backup host T21 and the first user end U1 based on the reservation information. At this time, the backup host T21 trains using the second initial training model and the data set, and the first user end U1 trains using the second initial training model. The training method is the same as described above, so the description will not be repeated here.

[0044] In one embodiment, when the second user end U2 saves the dataset U21 in the backup host T21 at a scheduled or non-scheduled time, it can transmit the path data of the dataset and a backup command to the service server T11, and the service server T11 controls the backup host T21 instructed to acquire the dataset U21 from the second user end U2 according to the backup command, in one embodiment, the dataset U21 acquired by the backup host T21 can be a copy of the dataset U21 or a data shortcut of the dataset U21, that is, the dataset U21 acquired by the backup host T21 can be actual data or only one path, and the data format to be saved is not limited thereto.

[0045] In other circumstances, the second user end U2 authorizes the service server T11, and the designated backup host T21 retrieves the data set U21 from the second user end U2 at regular or irregular times, and retrieves the updated data set U21 from the second user end U2 immediately, but the present invention is not limited thereto.

[0046] In other circumstances, the service server T11 can inquire the second user end U2 at a scheduled or non-scheduled time whether it agrees to obtain the authority, and after receiving the backup command, controls the backup host T21 designated to obtain the data set U21 from the second user end U2 based on the backup command, but the present invention is not limited thereto.

[0047] In still another situation, the second user end U2 uses the backup host T21 as a storage space and directly stores the data set U21 in the backup host T21, but the present invention is not limited thereto.

[0048] 4 shows a flowchart of a method according to an embodiment of the present invention. As shown in FIG. 4, the reservation method according to an embodiment of the present invention is implemented in the above-mentioned reservation platform or reservation system, and the method includes the following steps: <Step S1>: When the service server at the training end receives reservation information from the first user end, the service server communicates with the backup host at the training end and the first user end based on the reservation information, the backup host is instructed to operate by the service server, and the backup host obtains a data set from the second user end. <Step S2>: The service server transmits control information and a first initial training model to the backup host and the first user end. <Step S3>: Based on the control information, the backup host performs training using the first initial training model and the data set, generates a first training model, and returns the first training model to the service server. <Step S4>: The first user end performs training using the first initial training model based on the control information, generates a second training model, and returns the second training model to the service server. <Step S5>: The first training model and the second training model are calculated from the service server, a third training model is generated and returned to the backup host and the first user end, and the second user end obtains the third training model through the backup host.

[0049] In step S1, when the service server T11 of the training end T receives reservation information from the first user end U1, the service server T11 communicates with the backup host T21 and the first user end U1 based on the reservation information. The backup host T21 is instructed to operate by the service server T11. In addition, the backup host T21 obtains a data set U21 from the second user end U2.

[0050] In step S2, the service server T11 transmits control information and a first initial training model to the backup host T21 and the first user end U1, and the backup host T21 and the first user end U1 execute the installation control information.

[0051] In step S3, the backup host T21 then performs training according to the control information using the first initial training model and the data set U21, generates a first training model, and returns it to the service server T11.

[0052] In step S4, similarly, the first user end U1 performs training with the first initial training model based on the control information, generates a second training model, and returns it to the service server T11. In one embodiment, steps S3 and S4 are respectively performed at the same time, but the present invention is not limited thereto.

[0053] In step S5, the service server T11 performs calculations using the first and second training models, generates a third training model, and returns it to the backup host T21 and the first user end U1. The second user end U2 obtains the third training model from the backup host T21, and at the same time, the service server T11 performs testing using the first, second, and third training models, and sets the test result as a second initial training model. That is, by setting the second initial training model as the initial model for the next training, when the service server T11 receives reservation information again from the first user end U1, it transmits control information and the second initial training model from the service server T11 to the backup host T21 and the first user end U1, and then executes steps S3 to S5.

[0054] To further explain an embodiment of the reservation method according to the present invention, reference should be made to accompanying Figure 5. Figure 5 is a schematic diagram showing the steps of an embodiment of the present invention. The first user ends U1 are medical institutions A and B, respectively. The second user ends U2 are medical institutions C and D, respectively. The backup host T21 at the training end T includes a backup host for a medical institution C and a backup host for a medical institution D.

[0055] The first user end U1 logs into the reservation interface T3 of the training end T through the host and inputs reservation information. For example, it is 17:00 on July 20th. At this time, the reservation interface T3 transmits the reservation information to the service server T11, and the service server T11 communicates with the host of the first user end U1 and the backup host T21 of the training end T based on the reservation information at 17:00 on July 20th, and transmits the control information and the first initial training model to the backup host T21 and the host of the first user end U1. After that, the backup host of the medical institution C and the backup host of the medical institution D of the training end T respectively train with the first initial training model and the data set U21 acquired by them, obtain the first training model and return it to the service server T11. The hosts of the medical institution A and the medical institution B respectively train with the first initial training model, obtain the second training model and return it to the service server T11.

[0056] Finally, the service server T11 calculates the first training model and the second training model, and returns the generated third training model to the host of medical institution A, the host of medical institution B, the backup host of medical institution C, and the backup host of medical institution D. The host of medical institution C can obtain the third training model via the backup host of medical institution C, and the host of medical institution D can obtain the third training model via the backup host of medical institution D.

[0057] In summary, the cluster-based federated learning reservation platform, reservation system and reservation method according to one embodiment of the present invention realizes a backup host at the training end to use clusters to run on devices with different performance at the same location, and the backup host obtains data sets from the second user end. After the first user end inputs the reservation information, the first user end and the backup host perform training synchronously, which greatly improves the success rate of training, makes scheduling more flexible and efficient, and certainly achieves the objectives of the present invention.

[0058] Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It is clear to those skilled in the art that various modifications and improvements can be made to the above embodiments. It is clear from the description of the claims that such modifications and improvements can also be included in the technical scope of the present invention. [Explanation of symbols]

[0059] S Reservation System S1 Step S2 Step S3 Step S4 Step S5 Step T Training End T1 Main Server T11 Service Server T2 Slave Server T21 Backup Host T3 Reservation Interface P Reservation Platform U1 First user end U2 2nd User End U21 dataset

Claims

1. A cluster-based federated learning reservation platform for a first user end to input reservation information and train, a second user end having a data set; a training end including a main server and a slave server, the slave server being instructed to operate by the main server, the main server including a service server, the slave server being a backup host, the service server receiving the reservation information and communicating with the backup host and the first user end based thereon, the backup host being a training end used by the second user end to store the data set; The service server transmits corresponding control information and a first initial training model to the backup host and the first user end based on the reservation information; the backup host performs training using the first initial training model and the data set based on the control information, generates a first training model, and sends the training back to the service server; the first user end performs training using the first initial training model based on the control information, generates a second training model, and sends the training back to the service server; the service server receives the first training model, calculates the second training model based on the first training model, generates a third training model, and sends the training back to the backup host and the first user end; A cluster-based federated learning reservation platform characterized by:

2. The cluster-based federated learning reservation platform of claim 1, characterized in that the training end has a reservation interface through which the first user end transmits the reservation information, and the service server receives the reservation information and communicates with the backup host and the first user end based on the reservation information.

3. The cluster-based federated learning reservation platform according to claim 1, characterized in that the service server communicates and encrypts with the first user end through a first transmission protocol, and the first transmission protocol is HTTPS (HyperText Transfer Protocol Secure).

4. The cluster-based federated learning reservation platform according to claim 2, characterized in that the reservation interface communicates and encrypts with the service server through a second transmission protocol, and the second transmission protocol is SSH (Secure Shell Protocol).

5. A first user end for inputting reservation information; a second user end having a data set; A training end including a main server, a slave server, and a reservation interface, the slave server being instructed to operate by the main server, the main server including a service server, the slave server being a backup host, the service server receiving the reservation information through the reservation interface and communicating with the backup host and the first user end based thereon, the backup host being a training end used by the second user end to store the data set; The service server transmits corresponding control information and a first initial training model to the backup host and the first user end based on the reservation information; the backup host performs training using the first initial training model and the data set based on the control information, generates a first training model, and sends the training back to the service server; the first user end performs training using the first initial training model based on the control information, generates a second training model, and sends the training back to the service server; the service server receives the first training model, calculates the second training model based on the first training model, generates a third training model, and sends the training back to the backup host and the first user end; A cluster-based federated learning reservation system.

6. When the first user end re-enters the reservation information into the reservation interface, the service server performs a test using the first training model, the second training model, and the third training model to generate a test result; and the test result is a second initial training model; The service server transmits control information and the second initial training model to the backup host and the first user end based on the reservation information; the backup host performs training using the second initial training model and the data set based on the control information; Generate a fourth training model and return it to the service server, and the first user end trains with the second initial training model according to the control information; Generate a fifth training model and return it to the service server; The service server performs a calculation based on the fourth training model and the fifth training model; generating a sixth training model and sending it back to the backup host and the first user end; The cluster-based federated learning reservation system according to claim 5.

7. The cluster-based federated learning reservation system according to claim 5, wherein the quantity of the second user ends is equal to or greater than the quantity of the first user ends.

8. When a service server at a training end receives reservation information from a first user end, the service server communicates with a backup host at the training end and the first user end according to the reservation information, and the backup host is instructed to operate by the service server, and the backup host obtains a data set from a second user end; transmitting control information and a first initial training model from the service server to the backup host and the first user end; the backup host performs training using the first initial training model and the data set based on the control information, generates a first training model, and returns the first training model to the service server; the first user end performs training using the first initial training model according to the control information, generates a second training model, and returns the second training model to the service server; Calculating the first training model and the second training model from the service server, generating a third training model and returning it to the backup host and the first user end, and the second user end obtaining the third training model through the backup host; the service server performs testing using the first training model, the second training model, and the third training model to generate a test result, and sets the test result as a second initial training model; and when the service server receives the reservation information again from the first user end, the service server transmits the control information and the second initial training model to the backup host and the first user end.

Images