Task processing method and device, medium and product

Abstract

The invention provides a task processing method and device, a medium and a product, relates to the technical field of machine learning, and aims to solve the problem that a universal method has limitation in processing a cross-domain computing task. The task processing method comprises the following steps: dividing a cross-domain computing task into a plurality of sub-tasks according to a task area demand carried in the received cross-domain computing task; different subtasks are used for requesting different regional service nodes to perform task processing; sending the sub-task to a regional service node corresponding to each sub-task, so that the regional service node calls a plurality of computing nodes providing heterogeneous computing resources to process the sub-task; and receiving a processing sub-result of the sub-task sent by each regional service node, and determining a processing result of the cross-domain computing task according to the processing sub-result.

Description

Technical Field

[0001] The present application relates to the field of machine learning technology, and in particular to a task processing method, device, medium and product. Background Art

[0002] With the development of artificial intelligence technology represented by large artificial intelligence (AI) models, the amount of data and the demand for computing power are growing rapidly. Large-scale computing has gradually begun to adopt distributed computing. Building artificial intelligence development platforms in multiple regions can make full use of computing resources in different regions to achieve large-scale distributed computing. Due to the diversity of technology stacks such as computing frameworks of artificial intelligence development platforms, when training models based on artificial intelligence development platforms and / or reasoning computing tasks, the models between various artificial intelligence development platforms cannot be managed and shared with each other.

[0003] The general approach usually allows multiple AI development platforms to jointly train a global model without sharing local data, so that AI development platforms can collaborate in computing, train models, and process computing tasks in cross-domain scenarios.

[0004] However, the general method relies on homogeneous intelligent computing hardware resources to process computing tasks, while the intelligent computing hardware resources provided by the artificial intelligence development platform have different structures. When processing computing tasks, data between intelligent computing hardware resources cannot be directly shared and coordinated for scheduling. Therefore, when processing cross-domain tasks, the general method has different resource usage restrictions in different regions, which leads to limitations in the general method for processing cross-domain computing tasks. Summary of the invention

[0005] The embodiments of the present disclosure provide a task processing method, device, medium and product, aiming to solve the problem that general methods cannot schedule heterogeneous computing resources in cross-domain scenarios to realize the processing of cross-domain computing tasks.

[0006] In order to achieve the above purpose, this application adopts the following technical solutions:

[0007] In a first aspect, a task processing method is provided, including: dividing a cross-domain computing task into multiple subtasks according to the task area requirements carried in the received cross-domain computing task; different subtasks are used to request different regional service nodes to process the task; the subtask is sent to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtask; receiving the processing subresults of the subtasks sent by each regional service node, and determining the processing results of the cross-domain computing task according to the processing subresults.

[0008] Optionally, the cross-domain computing task is used to instruct training to obtain a cross-domain model suitable for different regions, and the processing result of the cross-domain computing task includes the trained cross-domain model.

[0009] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a first model parameter determined by each regional service node; the first model parameter is an average value of multiple second model parameters; different second model parameters are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing results of the cross-domain computing task based on the processing sub-results, including: determining the average value of the first model parameter determined by each regional service node as a third model parameter; and obtaining a cross-domain model through training based on the third model parameter.

[0010] Optionally, the processing sub-result of the subtask sent by each regional service node includes: a fourth model parameter determined by each regional service node; the fourth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the first gradient product; the first gradient product is the product of the learning rate and the first model gradient; the first model gradient is the sum of multiple second model gradients; different second model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing result of the cross-domain computing task based on the processing sub-result, including: determining the average value of the fourth model parameter determined by each regional service node as the fifth model parameter; and obtaining a cross-domain model through training based on the fifth model parameter.

[0011] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a third model gradient determined by each regional service node; the third model gradient is the sum of multiple fourth model gradients; different fourth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing results of the cross-domain computing task based on the processing sub-results, including: determining the sum of the third model gradients determined by each regional service node as the fifth model gradient; determining the difference between the model parameters obtained by the last training of the total service node and the second gradient product as the sixth model parameter; the second gradient product is the product of the learning rate and the fifth model gradient; and obtaining the cross-domain model based on the sixth model parameter training.

[0012] Optionally, the processing sub-results of the subtasks sent by each regional service node include: the sixth model gradient determined by the regional service node; the sixth model gradient is the sum of multiple seventh model gradients; different seventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing results of the cross-domain computing task based on the processing sub-results, including: determining the difference between the model parameters obtained by the last training of the computing node and the third gradient product as the seventh model parameter; the third gradient product is the product of the learning rate and the sixth model gradient; and obtaining the cross-domain model based on the seventh model parameter training.

[0013] Optionally, the processing sub-results of the subtask sent by each regional service node include: the eighth model parameter determined by the computing node; the eighth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fourth gradient product; the fourth gradient product is the product of the learning rate and the eighth model gradient; the eighth model gradient is the sum of the ninth model gradient; the ninth model gradient is the sum of multiple tenth model gradients; different tenth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing result of the cross-domain computing task based on the processing sub-results, including: obtaining a cross-domain model through training based on the eighth model parameters.

[0014] Optionally, the processing sub-results of the subtask sent by each regional service node include: the ninth model parameter determined by each regional service node; the ninth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fifth gradient product; the fifth gradient product is the product of the learning rate and the eleventh model gradient; different eleventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; determining the processing result of the cross-domain computing task based on the processing sub-results, including: determining the average value of the ninth model parameter determined by each regional service node as the tenth model parameter; and obtaining a cross-domain model based on the tenth model parameter training.

[0015] In a second aspect, a task processing device is provided, which includes: a communication unit and a processing unit; the processing unit is used to divide the cross-domain computing task into multiple sub-tasks according to the task area requirements carried in the received cross-domain computing task; different sub-tasks are used to request different regional service nodes to perform task processing; the communication unit is used to send the sub-task to the regional service node corresponding to each sub-task, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the sub-task; the processing unit is also used to receive the processing sub-results of the sub-tasks sent by each regional service node, and determine the processing results of the cross-domain computing task based on the processing sub-results.

[0016] Optionally, the cross-domain computing task is used to instruct training to obtain a cross-domain model suitable for different regions, and the processing result of the cross-domain computing task includes the trained cross-domain model.

[0017] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a first model parameter determined by each regional service node; the first model parameter is an average of multiple second model parameters; different second model parameters are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; a processing unit, specifically used to: determine the average of the first model parameters determined by each regional service node as a third model parameter; and obtain a cross-domain model based on training of the third model parameter.

[0018] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a fourth model parameter determined by each regional service node; the fourth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the first gradient product; the first gradient product is the product of the learning rate and the first model gradient; the first model gradient is the sum of multiple second model gradients; different second model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; a processing unit, specifically used to: determine the average value of the fourth model parameter determined by each regional service node as the fifth model parameter; and obtain a cross-domain model through training according to the fifth model parameter.

[0019] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a third model gradient determined by each regional service node; the third model gradient is the sum of multiple fourth model gradients; different fourth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; a processing unit, specifically used to: determine the sum of the third model gradients determined by each regional service node as the fifth model gradient; determine the difference between the model parameters obtained by the last training of the total service node and the second gradient product as the sixth model parameter; the second gradient product is the product of the learning rate and the fifth model gradient; and obtain a cross-domain model according to the sixth model parameter training.

[0020] Optionally, the processing sub-results of the subtasks sent by each regional service node include: the sixth model gradient determined by the regional service node; the sixth model gradient is the sum of multiple seventh model gradients; different seventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; a processing unit, specifically used to: determine the difference between the model parameters obtained by the last training of the computing node and the third gradient product as the seventh model parameter; the third gradient product is the product of the learning rate and the sixth model gradient; and obtain a cross-domain model according to the seventh model parameter training.

[0021] Optionally, the processing sub-results of the subtask sent by each regional service node include: the eighth model parameter determined by the computing node; the eighth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fourth gradient product; the fourth gradient product is the product of the learning rate and the eighth model gradient; the eighth model gradient is the sum of the ninth model gradient; the ninth model gradient is the sum of multiple tenth model gradients; different tenth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; the processing unit is specifically used to: obtain a cross-domain model according to the eighth model parameter training.

[0022] Optionally, the processing sub-results of the subtask sent by each regional service node include: the ninth model parameter determined by each regional service node; the ninth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fifth gradient product; the fifth gradient product is the product of the learning rate and the eleventh model gradient; different eleventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; a processing unit, specifically used to: determine the average value of the ninth model parameter determined by each regional service node as the tenth model parameter; and obtain a cross-domain model based on the tenth model parameter training.

[0023] In a third aspect, a task processing device is provided, comprising a memory and a processor; the memory is used to store computer execution instructions, and the processor and the memory are connected via a bus; when the task processing device is running, the processor executes the computer execution instructions stored in the memory, so that the task processing device performs the task processing method of the first aspect.

[0024] The task processing device may be an electronic device, or a part of an electronic device, such as a chip system in an electronic device. The chip system is used to support the electronic device to implement the functions involved in the first aspect and any possible implementation thereof, for example, to obtain and determine the data and / or information involved in the above-mentioned task processing method. The chip system includes a chip, and may also include other discrete devices or circuit structures.

[0025] In a fourth aspect, a computer-readable storage medium is provided, the computer-readable storage medium comprising computer execution instructions, and when the computer execution instructions are executed on a computer, the computer executes the task processing method described in the first aspect.

[0026] In a fifth aspect, a computer program product is also provided, which includes a computer program or instructions. When the computer instructions are executed on a task processing device, the task processing device executes the task processing method as described in the first aspect above.

[0027] It should be noted that the above-mentioned computer instructions may be stored in whole or in part on a computer-readable storage medium, wherein the computer-readable storage medium may be packaged together with the processor of the task processing device, or may be packaged separately from the processor of the task processing device, which is not limited in the embodiments of the present application.

[0028] The description of the second, third, fourth and fifth aspects of the present application can refer to the detailed description of the first aspect.

[0029] In the embodiments of the present application, the name of the task processing device does not limit the device or functional module itself. In actual implementation, these devices or functional modules may appear with other names. For example, the receiving unit may also be called a receiving module, a receiver, etc. As long as the functions of each device or functional module are similar to those of the present application, they fall within the scope of the claims of the present application and their equivalent technologies.

[0030] The technical solution provided by this application brings at least the following beneficial effects:

[0031] Based on any of the above aspects, an embodiment of the present application provides a task processing method. First, according to the task area requirements carried in the received cross-domain computing task, the cross-domain computing task is divided into multiple subtasks. Among them, different subtasks are used to request different regional service nodes to perform task processing. Then, the subtask is sent to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtask. Then, the processing sub-results of the subtasks sent by each regional service node are received, and the processing results of the cross-domain computing task are determined according to the processing sub-results.

[0032] As can be seen from the above, when processing tasks, the present application can divide the cross-domain computing task into multiple subtasks according to the needs of the task area carried in the cross-domain computing task. And send subtasks to the regional service node corresponding to each task, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process subtasks. In this way, when processing cross-domain computing tasks, the regional service node calls heterogeneous computing resources to process subtasks. The present application can also receive the processing sub-results of the subtasks sent by each regional service node, and determine the processing results of the cross-domain computing task based on the processing sub-results. In this way, the processing sub-results of each regional service node are brought together to realize the processing of cross-domain computing tasks in a cross-domain scenario.

[0033] The beneficial effects of the first, second, third, fourth and fifth aspects of the present application can all be referred to the analysis of the above beneficial effects, which will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1A schematic diagram of the structure of a task processing system provided in an embodiment of the present application;

[0035] Figure 2 A node relationship diagram provided for an embodiment of the present application;

[0036] Figure 3 A schematic diagram of a hardware structure of a communication device provided in an embodiment of the present application;

[0037] Figure 4 A flowchart of a task processing method provided in an embodiment of the present application;

[0038] Figure 5 A flowchart of another task processing method provided in an embodiment of the present application;

[0039] Figure 6 A flowchart of another task processing method provided in an embodiment of the present application;

[0040] Figure 7 A schematic diagram of a cross-domain model update provided in an embodiment of the present application;

[0041] Figure 8 A flowchart of another task processing method provided in an embodiment of the present application;

[0042] Fig. 9 A schematic diagram of another cross-domain model update provided in an embodiment of the present application;

[0043] Fig.10 A flowchart of another task processing method provided in an embodiment of the present application;

[0044] Fig.11 A schematic diagram of another cross-domain model update provided in an embodiment of the present application;

[0045] Fig.12 A flowchart of another task processing method provided in an embodiment of the present application;

[0046] Fig.13 A schematic diagram of another cross-domain model update provided in an embodiment of the present application;

[0047] Fig.14 A flowchart of another task processing method provided in an embodiment of the present application;

[0048] Fig.15 A schematic diagram of another cross-domain model update provided in an embodiment of the present application;

[0049] Fig.16 A flowchart of another task processing method provided in an embodiment of the present application;

[0050] Fig.17 A schematic diagram of another cross-domain model update provided in an embodiment of the present application;

[0051] Fig.18 A schematic diagram of the structure of a task processing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0053] It should be noted that, in the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0054] In order to facilitate a clear description of the technical solutions of the embodiments of the present application, in the embodiments of the present application, words such as "first" and "second" are used to distinguish between identical or similar items with substantially the same functions and effects. Those skilled in the art can understand that words such as "first" and "second" do not limit the quantity and order of execution.

[0055] Before introducing in detail the cross-domain model determination method provided by the present application, a brief introduction to the application scenarios and implementation environment involved in the present application is first given.

[0056] First, a brief introduction to the application scenarios involved in this application is given.

[0057] As described in the background technology, with the development of artificial intelligence technology represented by large artificial intelligence models, the amount of data and the demand for computing power are growing rapidly. Among them, the demand for computing power has shifted from general computing power to intelligent computing power with higher computing performance to accelerate model training and reasoning of computing tasks. However, the use of intelligent computing chips that provide a single intelligent computing power cannot meet large-scale computing. Therefore, large-scale computing has gradually begun to adopt distributed computing.

[0058] Building AI development platforms in multiple regions can make full use of computing resources in different regions to achieve large-scale distributed computing. However, the computing framework and other technology stacks of AI development platforms are diverse. Therefore, when training models based on AI development platforms and / or reasoning computing tasks, the models between various AI development platforms cannot be managed and shared in a coordinated manner.

[0059] The general approach usually allows multiple AI development platforms to jointly train a global model without sharing local data, so that AI development platforms can collaborate in computing, train models, and process computing tasks in cross-domain scenarios.

[0060] However, the general method relies on homogeneous intelligent computing hardware resources to process computing tasks. With the development of intelligent computing hardware resources (which can be intelligent computing chips), the intelligent computing hardware resources provided by artificial intelligence development platforms have different structures (also called heterogeneous intelligent computing hardware resources). Therefore, data between intelligent computing hardware resources cannot be directly shared, managed and scheduled in a coordinated manner.

[0061] In addition, the artificial intelligence development platform has the problem of uneven distribution of computing power hardware resources. Therefore, in large-scale calculations, there will be a mismatch between computing tasks and hardware computing power resources.

[0062] With the development of computing network technology, computing scheduling platforms based on cloud-network integration (including network-cloud scheduling mode and cloud-network scheduling mode) provide connection and computing services through the network, and can match computing hardware resources according to computing tasks to achieve unified management of heterogeneous computing hardware resources in the same artificial intelligence development platform. However, it cannot achieve unified scheduling of heterogeneous computing hardware resources in cross-domain scenarios.

[0063] As can be seen from the above, when the general method processes cross-domain tasks, the resource usage restrictions in different regions are different, which leads to certain limitations in the general method for processing cross-domain computing tasks. In response to the above problems, an embodiment of the present application provides a task processing method. First, the cross-domain computing task is divided into multiple subtasks according to the task area requirements carried in the received cross-domain computing task. Among them, different subtasks are used to request different regional service nodes to perform task processing. Then, the subtask is sent to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtask. Next, the processing sub-results of the subtasks sent by each regional service node are received, and the processing results of the cross-domain computing task are determined based on the processing sub-results.

[0064] As can be seen from the above, when processing tasks, the present application can divide the cross-domain computing task into multiple subtasks according to the needs of the task area carried in the cross-domain computing task. And send subtasks to the regional service node corresponding to each task, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process subtasks. In this way, when processing cross-domain computing tasks, the regional service node calls heterogeneous computing resources to process subtasks. The present application can also receive the processing sub-results of the subtasks sent by each regional service node, and determine the processing results of the cross-domain computing task based on the processing sub-results. In this way, the processing sub-results of each regional service node are brought together to realize the processing of cross-domain computing tasks in a cross-domain scenario.

[0065] The implementation environment of the above-mentioned cross-domain model determination method can be the cross-domain model determination system provided in the embodiment of the present application.

[0066] Figure 1 FIG. 1 shows a schematic diagram of a task processing system provided by an embodiment of the present application. Figure 1 As shown, the task processing system includes: a task processing device (also called a total service node, a global service node) 101, a subtask processing device 102 (also called a regional service node, a local service node), a regional computing device 103, and a communication device 104.

[0067] The task processing device 101 is communicatively connected with the subtask processing device 102 and the communication device 104 respectively.

[0068] The subtask processing device 102 is communicatively connected with the regional computing device 103 .

[0069] In practical applications, the task processing device 101 can be connected to any number of sub-task processing devices 102. For ease of understanding, Figure 1 An example in which one task processing device 101 is connected to two sub-task processing devices 102 is used for description.

[0070] In practical applications, the subtask processing device 102 can be connected to any number of regional computing devices 103. For ease of understanding, Figure 1 An example is given for describing the case where one subtask processing device 102 is connected to three regional computing devices 103 .

[0071] The task processing device 101 is used to update the cross-domain models applicable to different regions.

[0072] The subtask processing device 102 is used to update the regional model applicable to the current region.

[0073] The regional computing device 103 is used to execute the cross-domain computing task and determine the processing sub-result.

[0074] Optionally, the regional computing device 103 may be an intelligent computing power hardware resource computing node.

[0075] Intelligent computing hardware resources include: Graphics Processing Unit (GPU), Neural-Processing Unit (NPU), Tensor Processing Unit (TPU), Brain Processing Unit (BPU) and eXtended Processing Unit (xPU).

[0076] Optionally, the physical devices of the task processing device 101 and the subtask processing device 102 may be servers or other types of electronic devices, which is not limited in the embodiments of the present application.

[0077] Optionally, the above-mentioned server can be a server in a server cluster (consisting of multiple servers), or a chip in the server, or a system on a chip in the server, or can be implemented by a virtual machine (VM) deployed on a physical machine, which is not limited in the embodiments of the present application.

[0078] Optionally, the task processing system provided in the embodiment of the present application may also include an AI task management device (also referred to as an AI task management module), a computing power resource management device (also referred to as a computing power resource management module), a cross-domain computing control device (cross-domain computing control module), and a regional computing control device (local computing control module).

[0079] Among them, the AI ​​task management device is used to create cross-domain computing tasks, split cross-domain computing tasks, and monitor cross-domain computing tasks.

[0080] In one possible implementation, the AI ​​task management device can analyze the requirements of the cross-domain computing task and determine the relevant nodes in the cross-domain determination system that execute the cross-domain computing task. The cross-domain task related nodes include: the general management node (also called the global management node), the general service node, and the regional service node.

[0081] Computing resource management equipment is used to monitor and manage cross-domain computing clusters. Computing clusters include computing resources, storage resources, and network resources.

[0082] The regional computing control device is used to manage the regional computing cluster resources.

[0083] In one achievable manner, the regional computing control device and the computing power resource management device cooperate to complete the management of computing power resources. First, the regional computing control device manages the regional computing power cluster resources. Then, the regional computing control device sends the regional computing power cluster resources to the computing power resource management device. Then, the computing power resource management device aggregates the regional computing power cluster resources and manages them uniformly.

[0084] In one achievable manner, after the computing resource configuration required for the cross-domain computing task is determined, the computing resource management device can create the cross-domain computing resources, and the regional computing control device can create the regional computing resources.

[0085] The cross-domain computing control device can be used to implement the communication between the task processing device 101 and the subtask processing device 102. It includes: data, models and information of cross-domain computing subtasks.

[0086] The communication method between the task processing device 101 and the subtask processing device 102 is Google Remote Procedure Call (gRPC) communication.

[0087] In one achievable manner, the cross-domain computing control device may segment the relevant data according to the cross-domain computing subtasks.

[0088] Optionally, the cross-domain computing control device may divide the relevant data according to an equal distribution strategy.

[0089] Optionally, the cross-domain computing control device may perform weighting according to the total value of the computing power interacted by the subtask processing devices 102, and divide the relevant data according to the weighted value.

[0090] The implementation task processing device 101 may send the data and model required by the subtask processing device 102 to execute the cross-domain computing subtask.

[0091] Optionally, the cross-domain computing control device may support different cross-domain training strategies.

[0092] Cross-domain training strategies include: synchronous parameter update strategy, synchronous gradient update strategy, and asynchronous gradient update strategy.

[0093] Optionally, the cross-domain computing control device controls the manner in which the cross-domain model is updated.

[0094] Optionally, the regional computing control device is used to manage and schedule regional computing resources.

[0095] Optionally, the regional computing control device may determine a computing node that executes a cross-domain computing subtask.

[0096] Optionally, the regional computing control device may divide the cross-domain computing subtask into multiple regional computing subtasks.

[0097] Optionally, the regional computing control device may be used to implement communication between the subtask processing device 102 and the regional computing device 103 .

[0098] The communication mode between the subtask processing device 102 and the regional computing device 103 is gRPC communication.

[0099] In one practicable manner, the regional computing control device may segment the relevant data according to the regional computing subtasks.

[0100] Exemplarily, the regional computing control device may divide the relevant data according to an equal division strategy.

[0101] As another example, the cross-domain computing control device may perform weighting according to the computing power value of the regional computing device 103 and divide the related data according to the weighted value.

[0102] The subtask processing device 102 may send the data and model required by the regional computing device 103 to execute the regional computing subtask.

[0103] Optionally, the regional computing control device may support different regional training strategies (also referred to as local training strategies).

[0104] Regional training strategies include: synchronous parameter update strategy, synchronous gradient update strategy, and asynchronous gradient update strategy.

[0105] Optionally, the regional computing control device controls the manner in which the regional model is updated.

[0106] In some embodiments, Figure 1 Each device in can exist in the form of a node. Figure 2 Shows Figure 1 In one possible implementation, the task processing system includes a general management node, and the task processing device 101 exists in the form of a general service node, the subtask processing device 102 exists in the form of a regional service node, and the regional computing device 103 exists in the form of a regional computing node.

[0107] In one possible implementation, the task processing system includes a general management node, a general service node, and two regional service nodes. Each regional service node is connected to three regional computing nodes. Figure 2 shown.

[0108] There are a compute scheduler and four processors in the regional compute node.

[0109] The computing scheduler includes: GPU computing scheduling, NPU computing scheduling, and xPU computing scheduling.

[0110] Processors include: GPU, NPU, and xPU.

[0111] Regional computing nodes are grouped according to computing power type, and homogeneous intelligent computing power hardware resources are in the same computing node, that is, the processors in a regional computing node are the same.

[0112] In one achievable manner, the general service node, the regional service node, and the regional computing node interact with each other to jointly perform cross-domain computing tasks.

[0113] In one possible implementation, the general management node is used to perform heterogeneous resource management and computing task scheduling for cross-domain computing tasks, and control the operation flow of cross-domain computing tasks. The general management node is used to divide the cross-domain computing tasks into multiple cross-domain computing subtasks and assign them to regional service nodes.

[0114] In one practicable manner, the regional service node further divides the cross-domain computing subtask into regional computing subtasks and allocates the subtasks to the regional computing nodes.

[0115] In one implementation, within a regional computing node, a computing scheduler may allocate regional computing subtasks to processors.

[0116] As can be seen from the above, the task processing system can support various types of intelligent computing hardware resources, and through the aggregation and coordination of computing resources, it can support the simultaneous scheduling of cross-domain heterogeneous computing resources to perform cross-domain computing tasks, and can adapt to different underlying hardware and computing frameworks. It can effectively integrate resources and use them on demand to adapt to various computing demand scenarios, effectively exert the benefits of computing resources, and realize the unification of computing on multiple frameworks and multiple hardware, and improve the flexibility of scheduling computing resources.

[0117] The task processing equipment in the task processing system includes: Figure 3 The following are the components included. Figure 3 Taking the communication device shown in the figure as an example, the hardware structure of the task processing device is introduced.

[0118] Figure 3 , which is a schematic diagram of a hardware structure of a communication device provided in an embodiment of the present application. The communication device includes a processor 31, a memory 32, a communication interface 33, and a bus 34. The processor 31, the memory 32, and the communication interface 33 can be connected via a bus 34.

[0119] The processor 31 is the control center of the communication device, which can be a processor or a general term for multiple processing elements. For example, the processor 31 can be a general-purpose central processing unit (CPU) or other general-purpose processors. Among them, the general-purpose processor can be a microprocessor or any conventional processor.

[0120] As an embodiment, the processor 31 may include one or more CPUs, such as Figure 3 CPU0 and CPU1 are shown in the figure.

[0121] The memory 32 may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (RAM) or other type of dynamic storage device that can store information and instructions, an electrically erasable programmable read-only memory (EEPROM), a disk storage medium or other magnetic storage device, or any other medium that can be used to carry or store desired program code in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto.

[0122] In a possible implementation, the memory 32 may exist independently of the processor 31, and the memory 32 may be connected to the processor 31 via a bus 34 to store instructions or program codes. When the processor 31 calls and executes the instructions or program codes stored in the memory 32, the task processing method provided in the following embodiments of the present application can be implemented.

[0123] In the embodiment of the present application, for the communication device, the software programs stored in the memory 32 are different, so the functions implemented by the communication device are different. The functions performed by each device will be described in conjunction with the following flowchart.

[0124] In another possible implementation, the memory 32 may also be integrated with the processor 31 .

[0125] The communication interface 33 is used for the communication device to connect with other devices through a communication network, and the communication network may be Ethernet, wireless access network, wireless local area network (WLAN), etc. The communication interface 33 may include a receiving unit for receiving data and a sending unit for sending data.

[0126] The bus 34 may be an industry standard architecture (ISA) bus, a peripheral component interconnect (PCI) bus, or an extended industry standard architecture (EISA) bus. The bus may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 Only one thick line is used in the diagram, but this does not mean that there is only one bus or only one type of bus.

[0127] It should be pointed out that Figure 3 The structure shown in the figure does not constitute a limitation on the communication device, except Figure 3 In addition to the components shown, the communication device may include more or fewer components than shown, or combine certain components, or arrange the components differently.

[0128] The task processing method provided in the embodiment of the present application is described in detail below with reference to the accompanying drawings.

[0129] The task processing method provided in the embodiment of the present application is applied to Figure 1 The task processing device 101 in the task processing system shown in FIG. Figure 4 As shown, the task processing method provided in the embodiment of the present application includes:

[0130] S401. Divide the cross-domain computing task into multiple subtasks according to the task area requirements carried in the received cross-domain computing task.

[0131] Among them, different subtasks are used to request different regional service nodes to perform task processing.

[0132] In one possible implementation, when a cross-domain computing task involves multiple task areas, the task processing device can divide the cross-domain computing task into multiple subtasks according to the task area requirements carried in the received cross-domain computing task. The task can be handed over to the regional service node involved in the cross-domain computing task for processing, so as to realize the collaborative computing of multiple task areas. The task processing device can determine the processing result of the cross-domain computing task according to the processing sub-result of the regional service node.

[0133] In some embodiments, the cross-domain computing task involves multiple areas, and the task processing device can divide the cross-domain computing task into multiple subtasks according to the task area requirements carried in the cross-domain computing task.

[0134] Optionally, the task processing device can obtain cross-domain computing tasks from the AI ​​task management device.

[0135] Optionally, the AI ​​task management device is used to create cross-domain computing tasks.

[0136] Specifically, first, the AI ​​task management device can respond to the user's operation information. Then, the AI ​​task management device creates a cross-domain computing task and determines the task information of the cross-domain computing task. Then, the AI ​​task management device writes the task code based on the computing framework and determines the model, data set, training hyperparameters and training strategy required to run the cross-domain computing task. Then, the AI ​​task management device sends the cross-domain computing task to the task processing device, as well as the task information, model, data set, training hyperparameters and training strategy related to the cross-domain computing task.

[0137] The task information includes: the name of the cross-domain computing task, the application scenario of the cross-domain computing task, the creation time of the cross-domain computing task, and user information.

[0138] The computing frameworks include: TensorFlow, Python Torch (Pytorch) and Parallel Distributed Deep Learning (PaddlePaddle).

[0139] TensorFlow is an open source software library for machine learning and deep learning. It uses data flow graphs to represent computing tasks, in which tensors flow in the graph to transfer data and perform calculations.

[0140] Pytorch is a Python-based deep learning library that provides functions such as tensor calculations and dynamic neural networks.

[0141] PaddlePaddle is a deep learning platform for efficiently training and deploying deep learning models, supporting a variety of deep learning algorithms and application scenarios.

[0142] Exemplarily, the AI ​​task management device responds to user operation information. First, the AI ​​task management device creates a cross-domain computing task for full parameter fine-tuning of a Meta AI large language model - 7 billion parameters (Large Language Model Meta AI-7BillionParameters, LLaMA-7B), and determines the task information of the cross-domain computing task. Then, the AI ​​task management device can write task code based on the PyTorch framework, and determine the LLaMA-7B model, text data set, and Adaptive Moment Estimation with Weight Decay (Adaptive Moment Estimation with Weight Decay, AdamW) optimizer hyperparameters required to run the cross-domain computing task. Next, the AI ​​task management device sends the cross-domain computing task to the task processing device, as well as the task information, model, data set, training hyperparameters, training strategies, and computing resources related to the cross-domain computing task.

[0143] Among them, computing power resources include: 32 Type A smart computing cards (also called processors) and 32 Type B smart computing cards.

[0144] Exemplarily, when the task processing system includes 1 management node, 1 general service node, 2 regional service nodes, and 8 regional computing nodes, each regional computing node includes 8 Type A smart computing cards, 8 Type B smart computing cards, 8 Type B smart computing cards, 8 Type B smart computing cards, 8 Type A smart computing cards, 8 Type A smart computing cards, 8 Type A smart computing cards, and 8 Type B smart computing cards.

[0145] Optionally, computing power cards exchange information through high-speed communication between cards.

[0146] From the above, we can see that the task processing system adopts the distributed data parallel (DDP) mode to calculate cross-domain computing tasks.

[0147] In some embodiments, the cross-domain computing task is used to instruct training to obtain a cross-domain model applicable to different regions, and the processing result of the cross-domain computing task includes the trained cross-domain model.

[0148] S402: Send the subtask to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtask.

[0149] In one possible implementation, the task processing device may send subtasks to the regional service node corresponding to each task, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtasks. In this way, the task processing device may determine the processing result of the cross-domain computing task based on the processing subresults obtained by the regional service node processing the subtasks.

[0150] Combination Figure 1 The task processing device sends the subtask to the subtask processing device corresponding to each subtask, so that the subtask processing device calls multiple regional computing devices that provide heterogeneous computing resources to process the subtask.

[0151] Combination Figure 2 , the general service node sends the subtask to the regional service node corresponding to each subtask, so that the regional service node processes the subtask with multiple regional computing nodes that provide heterogeneous computing resources.

[0152] S403: Receive the processing sub-results of the sub-tasks sent by each regional service node, and determine the processing result of the cross-domain computing task according to the processing sub-results.

[0153] In one possible implementation, the task processing device may receive the processing sub-results of the sub-tasks sent by each regional service node, and determine the processing result of the cross-domain computing task according to the processing sub-results. In this way, the task processing device may determine the processing result of the cross-domain computing task according to the processing sub-results in a cross-domain scenario to determine the cross-domain model.

[0154] Combination Figure 1 The task processing device receives the subtask processing sub-result sent by each subtask processing device, and determines the processing result of the cross-domain computing task according to the processing sub-result.

[0155] Combination Figure 2 ,The general service node receives the subtask processing sub-results sent by each regional service node, and determines the processing result of the cross-domain computing task based on the processing sub-results.

[0156] In one achievable approach, combining Figure 2 , the general service node can send model information (also called global model information) to the regional service node. The regional service node can further send model information to the regional computing node. The regional computing node can execute the subtask and send the processing subresult to the regional service node (also called updating regional model information). The regional service node can aggregate multiple processing subresults and send the aggregated processing subresults to the general service node. The general service node can receive the processing subresults aggregated by each regional service node, and determine the processing result of the cross-domain computing task (also called the cross-domain model) according to the processing subresults.

[0157] Optionally, the general service node may continue to send the cross-domain model to the regional service node, and iteratively update the cross-domain model until the cross-domain model meets the model convergence condition.

[0158] Optionally, the total service node iteratively updates the cross-domain model until the number of model iterations is completed or the cross-domain computing task is completed.

[0159] For example, Figure 5 A flowchart of a task processing method is shown.

[0160] First, the user creates a cross-domain computing task on the AI ​​task management device. Then, the AI ​​task management device defines the cross-domain computing task based on the cross-domain computing task. Including: the name of the cross-domain computing task, the application scenario of the cross-domain computing task, the creation time of the cross-domain computing task, and user information. Then, the computing power resource management device can create cross-domain computing resources based on the task information of the cross-domain computing task. And send the cross-domain computing resources to the cross-domain computing control device. The computing resources include: cross-domain data and models. Then, the regional computing control device creates regional computing resources in conjunction with the computing power resource management device. And send the regional computing resources to the regional computing control device.

[0161] The cross-domain computing control device controls the execution of the cross-domain computing task, and the regional computing control device controls the execution of the cross-domain sub-computing task. Until the cross-domain computing task and the cross-domain sub-computing task are completed, the cross-domain computing resources and the regional computing resources are released.

[0162] AI task management devices can monitor cross-domain computing tasks, and users can view the status and results of cross-domain tasks.

[0163] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: a first model parameter determined by each regional service node, the first model parameter is an average value of multiple second model parameters, and different second model parameters are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks, combined with Figure 4 ,like Figure 6 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0164] S601: Determine an average value of the first model parameters determined by each regional service node as a third model parameter.

[0165] In one practicable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: the first model parameter determined by each regional service node, the task processing device may determine the average value of the first model parameter determined by each regional service node as the third model parameter, so as to determine the cross-domain model according to the third model parameter.

[0166] Specific, combined Figure 1 First, the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process the subtask and obtain the second model parameter. Then, the subtask processing device can determine the average value of multiple second model parameters as the first model parameter. Then, the subtask processing device sends the first model parameter to the task processing device.

[0167] Specific, combined Figure 1 , the task processing device determines an average value of multiple first model parameters as the third model parameter.

[0168] In one achievable approach, combining Figure 2 , the total service node S sends the model parameters to multiple regional service nodes L i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model parameters at the regional computation nodes The regional service node aggregates the model parameters of multiple regional computing nodes to obtain the regional model parameters M i The total service node aggregates the model parameters of multiple regional service computing nodes and determines the cross-domain model parameters.

[0169] Combination Figure 2 , the regional service node calls different computing nodes that provide heterogeneous computing resources to process subtasks and obtain different second model parameters.

[0170] The first model parameter is the average value of multiple second model parameters. The first model parameter satisfies the following formula:

[0171]

[0172] Among them, M i is used to represent the first model parameter, n is used to represent the number of regional computing nodes interacting with the regional service node, Used to represent the second model parameter.

[0173] The third model parameter is the average value of multiple first model parameters. The first model parameter satisfies the following formula:

[0174]

[0175] Wherein, M is used to represent the third model parameter, and m is used to represent the number of regional service nodes.

[0176] S602: Obtain a cross-domain model according to the third model parameter training.

[0177] In one achievable manner, after the task processing device determines the third model parameters, the task processing device may train the cross-domain model according to the third model parameters.

[0178] Optional, combined Figure 2 , the total service node can send the obtained cross-domain model to the regional service node and the regional computing node to continue updating the cross-domain model.

[0179] For example, Figure 7 A schematic diagram of a cross-domain model update is shown.

[0180] Among them, the cross-domain training strategy adopts a synchronous parameter update strategy, and the regional training strategy adopts a synchronous parameter update strategy.

[0181] The general service node sends the model parameters to the regional service nodes.

[0182] The regional service node sends the model parameters to the regional computing node.

[0183] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0184] The regional computing node determines the model parameters through forward propagation and back propagation, and uploads the model parameters to the regional service node.

[0185] The regional service node aggregates and updates the parameters uploaded by multiple regional computing nodes.

[0186] The regional service node uploads the updated parameters to the general service node.

[0187] The general service node aggregates and updates the parameters uploaded by multiple regional service nodes.

[0188] In the process of updating the model, forward propagation refers to the process of calculating and transmitting data layer by layer through each layer of the neural network from the input layer to finally obtain the output result of the model. In forward propagation, the input data can be converted into output results to achieve data classification and prediction.

[0189] In the process of updating the model, back propagation refers to the reverse calculation of the gradient through the chain rule based on the difference between the model's prediction results and the true label, and the gradient is propagated from the output layer back to each layer of the network to update the model parameters.

[0190] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: a fourth model parameter determined by each regional service node, the fourth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the first gradient product, the first gradient product is the product of the learning rate and the first model gradient, the first model gradient is the sum of multiple second model gradients, and different second model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks, combined Figure 4 ,like Figure 8 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0191] S801: Determine an average value of the fourth model parameter determined by each regional service node as a fifth model parameter.

[0192] In one achievable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: the fourth model parameter determined by each regional service node, the task processing device may determine the average value of the fourth model parameter determined by each regional service node as the fifth model parameter, so as to determine the cross-domain model according to the fifth model parameter.

[0193] Specific, combined Figure 1 First, the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process the subtask and obtain the second model gradient. Then, the subtask processing device can determine the sum of multiple second model gradients as the first model gradient. Next, the subtask processing device determines the product of multiple first model gradients and the learning rate as the first gradient product, and determines the difference between the model parameters obtained by the last training of the regional service node and the first gradient product as the fourth model parameter. Next, the subtask processing device sends the fourth model parameter to the task processing device.

[0194] Specific, combined Figure 1 , the task processing device determines the average value of the fourth model parameter determined by each regional service node as the fifth model parameter.

[0195] In one possible implementation, the general service node S sends the model parameters to multiple regional service nodes L. i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model gradient at the regional computation node The regional service node aggregates the model gradients of multiple regional computing nodes to obtain the regional model gradient G i The regional service node determines the regional model parameters according to the regional model gradient, and the general service node aggregates the model parameters of multiple regional service computing nodes to determine the cross-domain model parameters.

[0196] Combination Figure 2 ,The regional service node calls different computing nodes that provide heterogeneous computing resources to process subtasks and obtain different second model gradients.

[0197] The first model gradient is the sum of multiple second model gradients. The first model gradient satisfies the following formula:

[0198]

[0199] Among them, G i is used to represent the first model gradient, n is used to represent the number of regional computing nodes interacting with the regional service node, Used to represent the second model gradient.

[0200] The fourth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the first gradient product, and the first gradient product is the product of the learning rate and the first model gradient. The fourth model parameter satisfies the following formula:

[0201] M i =M i-1 -r×G i ;

[0202] Among them, M i-1 The model parameters obtained from the last training of the regional service node, r is used to represent the learning rate, G i Used to represent the first model gradient.

[0203] The fifth model parameter is an average value of a plurality of fourth model parameters.

[0204] S802: Obtain a cross-domain model according to the fifth model parameter training.

[0205] In one achievable manner, after the task processing device determines the fifth model parameter, the task processing device may train the cross-domain model according to the fifth model parameter.

[0206] Optional, combined Figure 2 , the total service node can send the obtained cross-domain model to the regional service node and the regional computing node to continue updating the cross-domain model.

[0207] For example, Fig. 9 A schematic diagram of a cross-domain model update is shown.

[0208] Among them, the cross-domain training strategy adopts a synchronous parameter update strategy, and the regional training strategy adopts a synchronous gradient update strategy.

[0209] The general service node sends the model parameters to the regional service nodes.

[0210] The regional service node sends the model parameters to the regional computing node.

[0211] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0212] The regional computing node determines the model gradient through forward propagation and uploads the model gradient to the regional service node.

[0213] The regional service node aggregates and updates the gradients uploaded by multiple regional computing nodes.

[0214] The regional service node updates the parameters according to the updated gradient.

[0215] The regional service node uploads the updated parameters to the general service node.

[0216] The general service node aggregates and updates the parameters uploaded by multiple regional service nodes.

[0217] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: a third model gradient determined by each regional service node, the third model gradient is the sum of multiple fourth model gradients, and different fourth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks, combined with Figure 4 ,like Fig.10 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0218] S1001. Determine the sum of the third model gradients determined by each regional service node as a fifth model gradient.

[0219] In one achievable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: the third model gradient determined by each regional service node, the task processing device may determine the sum of the third model gradients determined by each regional service node as the fifth model gradient, determine the sixth model parameter according to the fifth model gradient, and determine the cross-domain model according to the sixth model parameter.

[0220] Specific, combined Figure 1First, the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process the subtask and obtain the fourth model gradient. Then, the subtask processing device can determine the sum of multiple fourth model gradients as the third model gradient. Then, the subtask processing device sends the third model gradient to the task processing device.

[0221] Specific, combined Figure 1 , the task processing device determines the sum of the third model gradients determined by each regional service node as the fifth model gradient.

[0222] In one possible implementation, the general service node S sends the model parameters to multiple regional service nodes L. i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model gradient at the regional computation node The regional service node aggregates the model gradients of multiple regional computing nodes to obtain the regional model gradient G i The general service node aggregates the model gradients of multiple regional service nodes to obtain the cross-domain model gradient. The general service node determines the cross-domain model parameters based on the cross-domain model gradient.

[0223] Combination Figure 2 , the regional service node calls different computing nodes that provide heterogeneous computing resources to process subtasks and obtains different fourth model gradients.

[0224] The third model gradient is the sum of multiple fourth model gradients.

[0225] The fifth model gradient is the sum of multiple third model gradients. The fifth model gradient satisfies the following formula:

[0226]

[0227] Wherein, G is used to represent the fifth model gradient, and m is used to represent the number of regional service nodes.

[0228] S1002: Determine the difference between the model parameter obtained by the last training of the total service node and the second gradient product as the sixth model parameter.

[0229] The second gradient product is the product of the learning rate and the fifth model gradient.

[0230] In one achievable manner, the task processing device may determine the difference between the model parameter obtained in the last training and the second gradient product as the sixth model parameter, so as to train the cross-domain model according to the sixth model parameter.

[0231] Specific, combined Figure 1 First, the task processing device determines the product of the fifth model gradient and the learning rate as the second gradient product. Then, the task processing device determines the difference between the model parameter obtained by the last training of the total service node and the second gradient product as the sixth model parameter.

[0232] The sixth model parameter satisfies the following formula:

[0233] M=M ′ -r·G;

[0234] Among them, M ′ It is used to represent the model parameters obtained from the last training, and r is used to represent the learning rate.

[0235] S1003. Obtain a cross-domain model according to the sixth model parameter training.

[0236] In one achievable manner, after the task processing device determines the sixth model parameter, the task processing device may train the cross-domain model according to the sixth model parameter.

[0237] Optional, combined Figure 2 , the total service node can send the obtained cross-domain model to the regional service node and the regional computing node to continue updating the cross-domain model.

[0238] For example, Fig.11 A schematic diagram of a cross-domain model update is shown.

[0239] Among them, the cross-domain training strategy adopts a synchronous gradient update strategy, and the regional training strategy adopts a synchronous gradient update strategy.

[0240] The general service node sends the model parameters to the regional service nodes.

[0241] The regional service node sends the model parameters to the regional computing node.

[0242] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0243] The regional computing node determines the model gradient through forward propagation and uploads the model gradient to the regional service node.

[0244] The regional service node aggregates and updates the gradients uploaded by multiple regional computing nodes.

[0245] The regional service node uploads the updated gradient to the main service node.

[0246] The general service node aggregates and updates the gradients uploaded by multiple regional service nodes.

[0247] The general service node updates the parameters according to the gradient after the regional service nodes have updated them.

[0248] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: a sixth model gradient determined by the regional service node, the sixth model gradient is the sum of multiple seventh model gradients, and different seventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks, combined with Figure 4 ,like Fig.12 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0249] S1201. Determine the difference between the model parameter obtained by the last training of the computing node and the third gradient product as the seventh model parameter.

[0250] The third gradient product is the product of the learning rate and the sixth model gradient.

[0251] In one achievable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: in the case where the sixth model gradient determined by the regional service node, the task processing device can determine the third gradient product according to the sixth model gradient, and determine the difference between the model parameter obtained by the last training of the computing node and the third gradient product as the seventh model parameter, so as to determine the cross-domain model according to the seventh model parameter.

[0252] Specific, combined Figure 1 First, the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process the subtask and obtain the seventh model gradient. Then, the subtask processing device can determine the sum of multiple seventh model gradients as the sixth model parameter. Then, the subtask processing device sends the sixth model gradient to the task processing device.

[0253] Specific, combined Figure 1 The task processing device determines the product of the learning rate and the sixth model gradient as the third gradient product, and determines the difference between the model parameters obtained by the last training of the regional computing device and the third gradient product as the seventh model parameter.

[0254] In one possible implementation, the general service node S sends the model parameters to multiple regional service nodes L. i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model gradient at the regional computation node The general service node determines the model parameters according to the model gradients of the regional computing nodes.

[0255] Combination Figure 2 ,The regional service node calls different computing nodes that provide heterogeneous computing resources to process subtasks and obtains different seventh model gradients.

[0256] The sixth model gradient is the sum of multiple seventh model gradients.

[0257] The seventh model parameter satisfies the following formula:

[0258] M=M i ′ -r·G i ;

[0259] Among them, M i ′ It is used to represent the model parameters obtained from the last training of the regional computing node, and r is used to represent the learning rate.

[0260] S1202. Obtain a cross-domain model according to the seventh model parameter training.

[0261] In one achievable manner, after the task processing device determines the seventh model parameter, the task processing device may train the cross-domain model according to the seventh model parameter.

[0262] Optional, combined Figure 2 , the total service node can send the obtained cross-domain model to the regional service node and the regional computing node to continue updating the cross-domain model.

[0263] For example, Fig.13 A schematic diagram of a cross-domain model update is shown.

[0264] Among them, the cross-domain training strategy adopts an asynchronous gradient update strategy, and the regional training strategy adopts a synchronous gradient update strategy.

[0265] The general service node sends the model parameters to the regional service nodes.

[0266] The regional service node sends the model parameters to the regional computing node.

[0267] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0268] The regional computing node determines the model gradient through forward propagation and uploads the model gradient to the regional service node.

[0269] The regional service node aggregates and updates the gradients uploaded by multiple regional computing nodes.

[0270] The regional service node uploads the updated gradient to the main service node.

[0271] After the general service node obtains the gradient uploaded by the regional service node, it updates the parameters according to the gradient.

[0272] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: an eighth model parameter determined by the computing node, the eighth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the fourth gradient product, the fourth gradient product is the product of the learning rate and the eighth model gradient, the eighth model gradient is the sum of the ninth model gradient, the ninth model gradient is the sum of multiple tenth model gradients, and different tenth model gradients are obtained by the regional service node calling different computing nodes that provide heterogeneous computing resources to process subtasks, combined Figure 4 ,like Fig.14 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0273] S1401. Obtain a cross-domain model according to the eighth model parameter training.

[0274] In one practicable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: the eighth model parameter determined by the computing node, the task processing device can determine the cross-domain model according to the eighth model parameter.

[0275] Specific, combined Figure 1 , the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process subtasks and obtain the tenth model gradient. Then, the subtask processing device determines the sum of multiple tenth model gradients as the ninth model gradient. Then, the subtask processing device determines the sum of multiple ninth model gradients as the eighth model gradient. Then, the subtask processing device determines the product of the learning rate and the eighth model gradient as the fourth gradient product. Then, the subtask processing device determines the difference between the model parameters obtained by the last training of the regional service node and the fourth gradient product as the eighth model parameter. Then, the subtask processing device sends the eighth model parameter to the task processing device.

[0276] Specific, combined Figure 1 , the task processing device obtains a cross-domain model according to the eighth model parameter training.

[0277] In one possible implementation, the general service node S sends the model parameters to multiple regional service nodes L. i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model gradient at the regional computation node The regional service node aggregates the model gradients of multiple regional computing nodes to obtain the regional model gradient G i The general service node aggregates the model gradients of multiple regional service nodes to obtain the cross-domain model gradient. The general service node sends the cross-domain model gradient to the local computing node, and the local computing node determines the cross-domain model parameters based on the cross-domain model gradient.

[0278] Combination Figure 2 ,The regional service node calls different computing nodes that provide heterogeneous computing resources to process subtasks and obtain different tenth model gradients.

[0279] The ninth model gradient is the sum of multiple tenth model gradients.

[0280] The eighth model gradient is the sum of multiple ninth model gradients.

[0281] The eighth model parameter satisfies the following formula:

[0282]

[0283] Among them, M i-1 The model parameters obtained by the last training of the regional service node. r is used to represent the learning rate.

[0284] Optionally, the general service node may select a model from a regional computing node as the cross-domain model.

[0285] For example, Fig.15 A schematic diagram of a cross-domain model update is shown.

[0286] Among them, the cross-domain training strategy adopts a synchronous gradient update strategy, and the regional training strategy adopts a synchronous gradient update strategy.

[0287] The general service node sends the model parameters to the regional service nodes.

[0288] The regional service node sends the model parameters to the regional computing node.

[0289] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0290] The regional computing node determines the model gradient through forward propagation and uploads the model gradient to the regional service node.

[0291] The regional service node aggregates and updates the gradients uploaded by multiple regional computing nodes.

[0292] The regional service node can upload the updated gradient to the general service node.

[0293] After the general service node obtains the gradient uploaded by the regional service node, it sends the updated gradient to the regional service node.

[0294] The regional service node then sends the updated gradient to the regional computing node.

[0295] The regional computing node obtains the updated gradient from the regional service node, and after reverse transmission, reloads the model of the regional computing node.

[0296] In some embodiments, the processing sub-result of the subtask sent by each regional service node includes: a ninth model parameter determined by each regional service node, the ninth model parameter being the difference between the model parameter obtained by the last training of the regional service node and the fifth gradient product, the fifth gradient product being the product of the learning rate and the eleventh model gradient, and different eleventh model gradients being obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks, combined Figure 4 ,like Fig.16 As shown, in S403, the processing result of the cross-domain computing task is determined according to the processing sub-result, specifically including:

[0297] S1601: Determine an average value of the ninth model parameter determined by each regional service node as the tenth model parameter.

[0298] In one achievable manner, the task processing device determines the cross-domain model according to the subtask processing result sent by each regional service node. When the processing sub-result of the subtask sent by each regional service node includes: the ninth model parameter determined by each regional service node, the task processing device may determine the average value of the ninth model parameter determined by each regional service node as the tenth model parameter, so as to determine the cross-domain model according to the tenth model parameter.

[0299] Specific, combined Figure 1 First, the subtask processing device can call different regional computing devices that provide heterogeneous computing resources to process the subtask and obtain the eleventh model gradient. Then, the subtask processing device determines the product of the learning rate and the eleventh model gradient as the fifth gradient product. Then, the subtask processing device determines the difference between the model parameter obtained by the last training of the regional service node and the fifth gradient product as the ninth model parameter. Then, the subtask processing device sends the ninth model parameter to the task processing device.

[0300] Specific, combined Figure 1 , the task processing device determines an average value of multiple ninth model parameters as the tenth model parameter.

[0301] In one possible implementation, the general service node S sends the model parameters to multiple regional service nodes L. i The regional service node then sends the model parameters to the regional computing nodes that interact with it. Regional Compute Node According to the segmentation data Determine the model gradient at the regional computation node The regional service node determines the regional model parameters based on the model gradient of the regional computing node. The regional service node sends the updated regional model parameters to the regional computing node for further updating. After all regional computing nodes have completed the update, the regional service node sends the final model parameters to the general service node. The general service node aggregates the model parameters of multiple regional service nodes.

[0302] The ninth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the fifth gradient product, and the fifth gradient product is the product of the learning rate and the eleventh model gradient.

[0303] The task processing device determines an average value of the ninth model parameter determined by each regional service node as the tenth model parameter.

[0304] S1602. Obtain a cross-domain model according to the tenth model parameter training.

[0305] In one achievable manner, after the task processing device determines the tenth model parameter, the task processing device may train the cross-domain model according to the tenth model parameter.

[0306] Optionally, the general service node may send the obtained cross-domain model to the regional service node and the regional computing node to continue updating the cross-domain model.

[0307] For example, Fig.17 A schematic diagram of a cross-domain model update is shown.

[0308] Among them, the cross-domain training strategy adopts a synchronous parameter update strategy, and the regional training strategy adopts an asynchronous gradient update strategy.

[0309] After the regional computing node obtains the model, it loads the data and the model into the regional computing node.

[0310] The regional computing node determines the model gradient through forward propagation and uploads the model gradient to the regional service node.

[0311] The regional service node updates the parameters according to the gradients uploaded by the regional computing node and sends the updated parameters to the regional computing node.

[0312] The regional computing node can continue to update the model based on the parameters updated by the regional service node.

[0313] The regional service node can upload the parameters updated in the regional service node to the general service node.

[0314] The general service node aggregates and updates the parameters uploaded by multiple regional service nodes.

[0315] The general service node can send the parameters updated in the general service node to the regional service nodes, and further send them to the regional computing nodes.

[0316] The above mainly introduces the solution provided by the embodiment of the present application from the perspective of the method. In order to realize the above functions, it includes hardware structures and / or software modules corresponding to the execution of each function. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0317] The embodiment of the present application can divide the functional modules of the task processing device according to the above method example. For example, each functional module can be divided according to each function, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware or in the form of software functional modules. Optionally, the division of modules in the embodiment of the present application is schematic and is only a logical functional division. There may be other division methods in actual implementation.

[0318] Fig.18 FIG. 1 shows a schematic diagram of the structure of a task processing device provided in an embodiment of the present application. Fig.18 As shown, the task processing device includes: a communication unit 1801 and a processing unit 1802;

[0319] The processing unit 1802 is used to divide the cross-domain computing task into multiple subtasks according to the task area requirements carried in the received cross-domain computing task; different subtasks are used to request different regional service nodes to process the task; the communication unit is used to send the subtask to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes that provide heterogeneous computing resources to process the subtask; the processing unit is also used to receive the processing subresults of the subtasks sent by each regional service node, and determine the processing results of the cross-domain computing task according to the processing subresults.

[0320] Optionally, the cross-domain computing task is used to instruct training to obtain a cross-domain model suitable for different regions, and the processing result of the cross-domain computing task includes the trained cross-domain model.

[0321] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a first model parameter determined by each regional service node; the first model parameter is an average value of multiple second model parameters; different second model parameters are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: determine the average value of the first model parameter determined by each regional service node as a third model parameter; and obtain a cross-domain model based on training of the third model parameter.

[0322] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a fourth model parameter determined by each regional service node; the fourth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the first gradient product; the first gradient product is the product of the learning rate and the first model gradient; the first model gradient is the sum of multiple second model gradients; different second model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: determine the average value of the fourth model parameter determined by each regional service node as the fifth model parameter; and obtain a cross-domain model based on the fifth model parameter training.

[0323] Optionally, the processing sub-results of the subtasks sent by each regional service node include: a third model gradient determined by each regional service node; the third model gradient is the sum of multiple fourth model gradients; different fourth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: determine the sum of the third model gradients determined by each regional service node as the fifth model gradient; determine the difference between the model parameters obtained by the last training of the total service node and the second gradient product as the sixth model parameter; the second gradient product is the product of the learning rate and the fifth model gradient; and obtain a cross-domain model according to the sixth model parameter training.

[0324] Optionally, the processing sub-results of the subtask sent by each regional service node include: the sixth model gradient determined by the regional service node; the sixth model gradient is the sum of multiple seventh model gradients; different seventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: determine the difference between the model parameters obtained by the last training of the computing node and the third gradient product as the seventh model parameter; the third gradient product is the product of the learning rate and the sixth model gradient; and obtain a cross-domain model according to the seventh model parameter training.

[0325] Optionally, the processing sub-results of the subtask sent by each regional service node include: the eighth model parameter determined by the computing node; the eighth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fourth gradient product; the fourth gradient product is the product of the learning rate and the eighth model gradient; the eighth model gradient is the sum of the ninth model gradient; the ninth model gradient is the sum of multiple tenth model gradients; different tenth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: obtain a cross-domain model according to the eighth model parameter training.

[0326] Optionally, the processing sub-results of the subtask sent by each regional service node include: the ninth model parameter determined by each regional service node; the ninth model parameter is the difference between the model parameter obtained by the previous training of the regional service node and the fifth gradient product; the fifth gradient product is the product of the learning rate and the eleventh model gradient; different eleventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process subtasks; processing unit 1802 is specifically used to: determine the average value of the ninth model parameter determined by each regional service node as the tenth model parameter; and obtain a cross-domain model based on the tenth model parameter training.

[0327] An embodiment of the present application also provides a computer-readable storage medium, which includes computer execution instructions. When the computer execution instructions are executed on a computer, the computer executes the task processing method provided in the above embodiment.

[0328] The embodiment of the present application also provides a computer program, which can be directly loaded into a memory and contains software code. After being loaded and executed by a computer, the computer program can implement the task processing method provided in the above embodiment.

[0329] Those skilled in the art should be aware that in one or more of the above examples, the functions described in this application can be implemented using hardware, software, firmware, or any combination thereof. When implemented using software, these functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer-readable storage media and communication media, wherein communication media include any media that facilitates the transmission of computer programs from one place to another. Storage media can be any available media that can be accessed by general or special-purpose computers.

[0330] Through the description of the above implementation methods, technical personnel in the relevant field can clearly understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0331] In several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of the modules or units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the coupling or direct coupling or communication connection between each other shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms. The unit described as a separate component may or may not be physically separated, and the component displayed as a unit may be a physical unit or multiple physical units, that is, it may be located in one place, or it may be distributed in multiple different places. Some or all of the units can be selected according to actual needs to achieve the purpose of the scheme of this embodiment.

[0332] In addition, each functional unit in each embodiment of the present application can be integrated into a processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of a software functional unit. If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium. Based on such an understanding, the technical solution of the embodiment of the present application is essentially or the part that contributes to the general technology or all or part of the technical solution can be embodied in the form of a software product, which is stored in a storage medium, including several instructions to enable a device (which can be a single-chip microcomputer, chip, etc.) or a processor (processor) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: various media that can store program codes, such as a U disk, a mobile hard disk, a ROM, a RAM, a disk, or an optical disk.

[0333] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by a person skilled in the art within the technical scope disclosed in the present application should be included in the protection scope of the present application. Therefore, the protection scope of the present application shall be based on the protection scope of the claims.

Claims

1. A task processing method, characterized in that: The method comprises: According to the task area requirements carried in the received cross-domain computing task, the cross-domain computing task is divided into a plurality of subtasks; different subtasks are used to request different regional service nodes to perform task processing; Sending the subtask to a regional service node corresponding to each of the subtasks, so that the regional service node calls multiple computing nodes providing heterogeneous computing resources to process the subtask; The processing sub-result of the sub-task sent by each of the regional service nodes is received, and the processing result of the cross-domain computing task is determined according to the processing sub-result.

2. The method according to claim 1, characterized in that The cross-domain computing task is used to instruct training to obtain a cross-domain model suitable for different regions, and the processing result of the cross-domain computing task includes a trained cross-domain model.

3. The method according to claim 2, characterized in that The processing sub-result of the subtask sent by each of the regional service nodes includes: a first model parameter determined by each of the regional service nodes; the first model parameter is an average value of multiple second model parameters; different second model parameters are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; The determining the processing result of the cross-domain computing task according to the processing sub-result includes: determining an average value of the first model parameter determined by each of the regional service nodes as a third model parameter; The cross-domain model is obtained by training according to the third model parameters.

4. The method according to claim 2, characterized in that: The processing sub-result of the subtask sent by each of the regional service nodes includes: a fourth model parameter determined by each of the regional service nodes; the fourth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the first gradient product; the first gradient product is the product of the learning rate and the first model gradient; the first model gradient is the sum of multiple second model gradients; different second model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; the processing result of the cross-domain computing task is determined according to the processing sub-result, including: determining an average value of the fourth model parameter determined by each of the regional service nodes as a fifth model parameter; The cross-domain model is obtained by training according to the fifth model parameter.

5. The method according to claim 2, characterized in that: The processing sub-result of the subtask sent by each of the regional service nodes includes: a third model gradient determined by each of the regional service nodes; the third model gradient is the sum of multiple fourth model gradients; different fourth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; The determining the processing result of the cross-domain computing task according to the processing sub-result includes: Determining the sum of the third model gradients determined by each of the regional service nodes as a fifth model gradient; The difference between the model parameter obtained by the last training of the total service node and the second gradient product is determined as the sixth model parameter; the second gradient product is the product of the learning rate and the fifth model gradient; The cross-domain model is obtained by training according to the sixth model parameter.

6. The method according to claim 2, characterized in that The processing sub-result of the subtask sent by each of the regional service nodes includes: a sixth model gradient determined by the regional service node; the sixth model gradient is the sum of multiple seventh model gradients; different seventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; The determining the processing result of the cross-domain computing task according to the processing sub-result includes: Determine the difference between the model parameter obtained by the last training of the computing node and the third gradient product as the seventh model parameter; the third gradient product is the product of the learning rate and the sixth model gradient; The cross-domain model is obtained by training according to the seventh model parameter.

7. The method according to claim 2, characterized in that The processing sub-result of the subtask sent by each of the regional service nodes includes: the eighth model parameter determined by the computing node; the eighth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the fourth gradient product; the fourth gradient product is the product of the learning rate and the eighth model gradient; the eighth model gradient is the sum of the ninth model gradient; the ninth model gradient is the sum of the multiple tenth model gradients; different tenth model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; The determining the processing result of the cross-domain computing task according to the processing sub-result includes: The cross-domain model is obtained by training according to the eighth model parameter.

8. The method according to claim 2, characterized in that: The processing sub-result of the subtask sent by each of the regional service nodes includes: a ninth model parameter determined by each of the regional service nodes; the ninth model parameter is the difference between the model parameter obtained by the last training of the regional service node and the fifth gradient product; the fifth gradient product is the product of the learning rate and the eleventh model gradient; different eleventh model gradients are obtained by the regional service node calling different computing nodes providing heterogeneous computing resources to process the subtask; The determining the processing result of the cross-domain computing task according to the processing sub-result includes: determining an average value of the ninth model parameter determined by each of the regional service nodes as a tenth model parameter; The cross-domain model is obtained by training according to the tenth model parameter.

9. A task processing device, characterized in that: The device comprises: a communication unit and a processing unit; The processing unit is used to divide the cross-domain computing task into multiple subtasks according to the task area requirements carried in the received cross-domain computing task; different subtasks are used to request different regional service nodes to perform task processing; The communication unit is used to send the subtask to the regional service node corresponding to each subtask, so that the regional service node calls multiple computing nodes providing heterogeneous computing resources to process the subtask; The processing unit is further configured to receive a processing sub-result of the sub-task sent by each of the regional service nodes, and determine a processing result of the cross-domain computing task according to the processing sub-result.

10. A task processing device, characterized in that: include: A processor and a memory; wherein the memory is used to store one or more programs, and the one or more programs include computer-executable instructions. When the device is running, the processor executes the computer-executable instructions stored in the memory to enable the device to perform the method described in any one of claims 1 to 8.

11. A computer-readable storage medium, characterized in that: When the computer-executable instructions stored in the computer-readable storage medium are executed by a processor of a task processing device, the task processing device can perform the method according to any one of claims 1 to 8.

12. A computer program product, characterized in that The computer program product comprises: a computer program or instructions, and when the computer program or instructions are run on a computer, the computer is caused to perform the method according to any one of claims 1 to 8.

Images