CN114358302A

CN114358302A - Artificial intelligence AI training method, system and equipment

Info

Publication number: CN114358302A
Application number: CN202011626123.8A
Authority: CN
Inventors: 陈普
Original assignee: Huawei Cloud Computing Technologies Co Ltd
Current assignee: Huawei Cloud Computing Technologies Co Ltd
Priority date: 2020-10-14
Filing date: 2020-12-30
Publication date: 2022-04-15
Also published as: WO2022078280A1

Abstract

The application discloses an artificial intelligence AI training method which is applied to a guidance system. After a user of the cloud training system triggers the running of a training code on the local device, the method comprises the following steps: acquiring training task information according to the training code running on the local equipment, wherein the training code is used for training an AI model, and the AI model is developed and acquired by the user based on an AI framework installed on the local equipment; uploading the training task information to a cloud training system, and informing the cloud training system to execute a training task corresponding to the training task information. The method enables the developer to develop and train the AI model in the local device without being limited by the computing power of the local device.

Description

Artificial intelligence AI training method, system and equipment

Technical Field

The present application relates to the technical field of Artificial Intelligence (AI), and in particular, to a method, a system, and an apparatus for AI training.

Background

With the development of AI technology, AI is applied to various industries. Typical AI techniques include deep learning techniques, machine learning techniques, and the like, and the basic idea is to design an AI model and train the AI model based on training data, so that the trained AI model has certain functions, for example: object detection, object recognition, etc. The AI model is an algorithm implemented using AI techniques, such as: a deep learning model based on deep learning techniques.

In order to develop and train the AI model conveniently, various AI frameworks have appeared in the industry, and developers are generally used to install the AI framework on a local device, and develop and train the AI model based on the AI framework. Since training of the AI model requires computationally intensive support. The AI model developed by a developer at a local device based on the AI framework is typically trained locally, limited by the computing power of the local device.

Disclosure of Invention

The application provides an artificial intelligence AI training method, which uploads training task information to a cloud training system through the running of a training code in local equipment, and utilizes cloud resources to realize the training of an AI model, so that a developer can not be limited by the computing power of the local equipment when the developer executes the AI model on the local equipment.

In a first aspect, the present application provides an AI training method, which is applied to a guidance system, and when a user of a cloud training system triggers running of a training code on a local device, the guidance system executes the following steps: acquiring training task information according to a training code running on local equipment, wherein the training code is used for training an AI model, and the AI model is developed and obtained by a user based on an AI framework installed on the local equipment; and uploading the training task information to a cloud training system, and informing the cloud training system to execute a training task corresponding to the training task information.

By the method, when a user of the cloud training system needs to train the AI model built locally, the corresponding training code can be written by the editor on the local equipment, and after the user runs the training code locally, the guidance system in the application can acquire the training task information and upload the training task information to the cloud training system. The method can be used for training the AI model by using the resources on the cloud under the condition of keeping the habit of carrying out model development and model training locally by the user, thereby solving the problem of insufficient resources for training of local equipment and bringing convenience to the user.

In one possible implementation of the first aspect, the training task information includes information obtained from training code by an obtaining component in the AI framework, and information obtained from the local device according to the information in the training code. The acquisition component in the AI framework can acquire information in the modes of reading the training code, intercepting a calling API in the training code and the like in the running process of the training code, and the acquisition component acquires the information in the training code, so that the guidance system can acquire the training task information more quickly.

In one possible implementation of the first aspect, the aforementioned training task information comprises one or more of the following data: the training system comprises training parameters in the training codes, an AI model to be trained, training program logic used for training the AI model in the training codes, training environment information of local equipment and cloud training access information used for being connected with the cloud training system. The uploaded training task information enables the cloud training system to smoothly prepare a cloud training environment and execute a training task in the cloud environment.

In one possible implementation of the first aspect, the training environment information of the local device includes: version information of the AI framework, version information of a programming language of the training code.

Because the AI model and the training code are developed by the user based on the AI framework and the programming language installed in the local device, the version information of the AI framework in the local training environment information and the version information of the programming language of the training code are uploaded to the cloud training system, so that the cloud training system can prepare the version of the AI framework and the programming language matched with the AI model and the training code to be trained in advance when preparing the cloud environment.

In one possible implementation of the first aspect, the method performed by the boot system further comprises: receiving a training data acquisition request sent by a cloud training system in the process of executing the training task; and acquiring the training data and sending the training data to the cloud training system.

By the method, the guide system does not need to upload training data or upload all training data to the cloud training system before the cloud training system starts to execute the training task, so that too long waiting transmission time is avoided before the cloud training task starts, and the user experience is improved.

In one possible implementation of the first aspect, before notifying a cloud training system to execute a training task corresponding to the training task information, the method further includes: and receiving an environment preparation success response returned by the cloud training system.

In one possible implementation of the first aspect, the method further comprises: and receiving the AI model which is returned by the cloud training system and is finished in training. After receiving the trained AI model, the guidance system may store the trained AI model in the local device, and may also provide a prompt to the user, for example, prompt the user where the trained AI model is stored in the local device, so that the user may more conveniently obtain the trained AI model, and the user may conveniently apply the trained AI model subsequently.

In one possible implementation of the first aspect, the guidance system may be obtained from a cloud training system and installed in the local device. For example, an interface for downloading a guidance system may be provided in the cloud training system.

In a second aspect, the present application further provides an AI training method, which is applied to a cloud training system, and includes: acquiring training task information sent by a guidance system after a user triggers and runs a training code on local equipment, wherein the training task information comprises training environment information of the local equipment; performing preparation of a cloud training environment according to the training environment information; and executing the training task corresponding to the training task information based on the cloud training environment.

In one possible implementation of the second aspect, the training environment information of the local device includes: version information of an AI framework on which an AI model to be trained depends and version information of a programming language used by training code used to train the AI model.

In one possible implementation of the second aspect, the preparing of the cloud training environment according to the training environment information includes: and setting the AI framework and the programming language used for executing the training task in the cloud training environment according to the version information of the AI framework and the version information of the programming language of the training code.

In one possible implementation of the second aspect, the training task information further comprises: training parameters in the training codes, the AI model and training program logic used for training the AI model in the training codes; executing a training task corresponding to the training task information based on the cloud training environment, wherein the training task comprises: performing training of the AI model in the prepared cloud training environment according to the training parameters and the training program logic.

In one possible implementation of the second aspect, the training task information further comprises: cloud training access information, before performing preparation of a cloud training environment according to the training environment information, the method further comprising: and performing authentication and/or charging query on the training task corresponding to the training task information according to the cloud training access information.

The advantageous effects of the features in the second aspect and the possible implementation manners of the second aspect may refer to the advantageous effects of the corresponding features in the first aspect, and are not described herein again.

In a third aspect, the present application further provides a guidance system, including: the acquisition module is used for acquiring training task information according to a training code running on local equipment after a user of the cloud training system triggers the running of the training code on the local equipment, wherein the training code is used for training an AI model, and the AI model is developed and acquired by the user based on an AI framework installed on the local equipment; and the sending module is used for uploading the training task information to a cloud training system and informing the cloud training system to execute a training task corresponding to the training task information.

In one possible implementation of the third aspect, the training task information includes information obtained from the training code using an obtaining component in the AI framework, and information obtained from the local device according to information in the training code.

In one possible implementation of the third aspect, the training task information includes one or more of the following data: the training parameters in the training codes, the AI model, the training program logic in the training codes for training the AI model, the training environment information of the local device, and the cloud training access information for connecting with the cloud training system.

In one possible implementation of the third aspect, the training environment information of the local device includes: version information of the AI framework, version information of a programming language of the training code.

In a possible implementation of the third aspect, the system further includes a receiving unit, where the receiving unit is configured to receive a training data acquisition request sent by the cloud training system in a process of executing the training task; the acquisition unit is further used for acquiring the training data; the sending unit is further configured to send the training data to the cloud training system.

In a possible implementation of the third aspect, the system further includes a receiving unit, where the receiving unit is configured to receive an environment preparation success response returned by the cloud training system before the sending unit notifies the cloud training system to execute the training task corresponding to the training task information.

In a possible implementation of the third aspect, the system further includes a receiving unit, configured to receive the trained AI model returned by the cloud training system.

In one possible implementation of the third aspect, the guidance system is obtained from the cloud training system and installed in the local device.

In a fourth aspect, the present application further provides a cloud training system, including: the system comprises an environment preparation unit, a training unit and a training unit, wherein the environment preparation unit is used for acquiring training task information sent by a guide system after a user triggers and runs a training code on local equipment, and the training task information comprises training environment information of the local equipment; performing preparation of a cloud training environment according to the training environment information; and the training task execution unit is used for executing the training task corresponding to the training task information based on the cloud training environment.

In one possible implementation of the fourth aspect, the training environment information of the local device includes: and the version information of the AI framework depended by the AI model to be trained is used for training the version information of the programming language of the training code of the AI model.

In a possible implementation of the fourth aspect, the environment preparation unit is specifically configured to set, according to the version information of the AI framework and the version information of the programming language of the training code, the AI framework and the programming language used for executing the training task in the cloud training environment.

In one possible implementation of the fourth aspect, the training task information further includes: training parameters in the training codes, the AI model and training program logic used for training the AI model in the training codes; the training task execution unit is specifically configured to execute training of the AI model in the prepared cloud training environment according to the training parameters and the training program logic.

In one possible implementation of the fourth aspect, the training task information further includes: the cloud training access information, the environment preparation unit, further configured to: and performing authentication and/or charging query on the training task corresponding to the training task information according to the cloud training access information.

In a fifth aspect, the present application further provides a computing device comprising a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions to cause the computing device to perform the method described in the foregoing first aspect or possible implementation of the first aspect, or to perform the method described in the foregoing second aspect or possible implementation of the second aspect.

In a sixth aspect, the present application further provides a computer-readable storage medium storing computer program code, which, when executed by a computing device, performs the method of the foregoing first aspect or possible implementation of the first aspect, or performs the method of the foregoing second aspect or possible implementation of the second aspect. The computer readable storage medium includes, but is not limited to, volatile memory such as random access memory, non-volatile memory such as flash memory, hard disk (HDD), Solid State Disk (SSD).

In a seventh aspect, the present application further provides a computer program product comprising computer program code which, when executed by a computing device, performs the method provided in the foregoing first aspect or possible implementation of the first aspect, or performs the method provided in the foregoing second aspect or possible implementation of the second aspect. The computer program product may be a software installation package, which may be downloaded and executed on a computing device in case it is desired to use the method as provided in the aforementioned first aspect or possible implementation of the first aspect, or in case it is desired to use the method as provided in the aforementioned second aspect or possible implementation of the second aspect.

In an eighth aspect, the present application further provides an artificial intelligence AI system, including the guidance system described in the foregoing third aspect and possible implementations of the third aspect, and the cloud training system described in the foregoing fourth aspect and possible implementations of the fourth aspect.

Drawings

Fig. 1 is a schematic diagram of a system architecture according to an embodiment of the present application;

fig. 2 is a schematic structural diagram of a guidance system 106 and a cloud training system 120 according to an embodiment of the present disclosure;

fig. 3 is a schematic flowchart of an AI training method according to an embodiment of the present disclosure;

fig. 4 is a schematic flowchart of a process of performing authentication and charging query by a cloud training system according to training task information according to an embodiment of the present application;

fig. 5 is a schematic flowchart of a cloud training system executing a training task according to an embodiment of the present application;

fig. 6 is a schematic structural diagram of a computing device 300 according to an embodiment of the present application;

fig. 7 is a schematic structural diagram of a computing device 400 according to an embodiment of the present application.

Detailed Description

In order to more clearly describe the technical solutions of the embodiments of the present application, the following technical terms related to the present application are explained first:

AI training: the method is a process of updating parameters in the AI model by using training data so that the AI model learns the characteristics and rules in the training data and realizes specific application functions.

During the AI training process, a large amount of computing work needs to be performed, and a large amount of computing resources (i.e., computing power) and storage resources need to be consumed.

Training codes: the training code defines training parameters and training program logic in the AI training process, and may further include name or address information of the AI model to be trained, name or address information of a training data set, and the like. The training code also includes cloud training access information for connecting with a cloud training system at the cloud.

An AI framework: the method is a library (or tool) for developing, training and testing an AI model, wherein an AI framework comprises a plurality of packaged code components, each code component is a pre-constructed and optimized function set, and the code components in the AI framework are also called as function components in the application. Developers can develop and train AI models meeting application targets more quickly by using the AI framework without detailed knowledge of the underlying algorithm implementation when the AI models are constructed and trained. Currently, many AI frameworks are presented in the industry, and development and training of AI models using one or more AI frameworks brings great convenience to developers. Typical AI frameworks include deep learning frameworks, and various types of deep learning frameworks are in the industry, which provide differences in functionality, but all aim to provide developers with libraries (or tools) for the development, training, and testing of deep learning models.

AI techniques are rapidly evolving and are increasingly applied to more and more complex scenarios. At present, the structure of the AI model is more and more complex, and the performance requirements of many complex scenes on the AI model are also more and more high, so that at present, the AI model with a complex structure often needs to be trained by using a large amount of training data. However, the AI development patterns that developers are accustomed to are: the AI framework is installed on the local device and then the local editor is utilized, for example: and calling an AI framework to construct an AI model by using an Integrated Development Environment (IDE) of local application, and writing a training code. And then, local resources are utilized to execute training, and the trained AI model is obtained. Although some AI frameworks support performing distributed training on AI models at present, developers still often encounter the problem that locally developed AI models cannot have sufficient training resources locally.

It should be understood that a local or local device in this application refers to a device used by a developer to develop an AI model (e.g., a server, a virtual machine used by the developer), and/or other devices or clusters of devices belonging to the same owner as the device used by the developer to develop the AI model. Often, the local devices are in a relatively close physical environment, such as: a machine room. It should be noted that, in the case of AI development using a remote virtual machine (e.g., desktop cloud) by a client in a local terminal, the local terminal and the remote virtual machine may also be referred to as a local device.

In the cloud computing mode, a cloud service provider constructs a cloud environment by using a large amount of basic resources (such as computing resources, storage resources and communication resources), and the cloud service provider can provide various basic resources, platforms or application capabilities for cloud tenants, so that the cloud tenants can develop own services.

Based on the background, the application provides an AI training method, which can be used for training a locally constructed AI model for a developer by using basic resources on a cloud under the condition that the developer keeps the habit of local AI development, can solve the problem of insufficient local training resources, and can more conveniently and quickly train the AI model for the developer. It should be understood that the developers described herein may also be referred to as users, representing those who use the AI training methods of the present application. Since the user needs to register and purchase the cloud training service on the cloud platform before using the AI training method of the present application, the user (or called developer) in the present application is also a user of the cloud training system.

Before describing the specific embodiments of the present application, the system architecture applied in the present application is described. Fig. 1 is a schematic diagram of an exemplary system architecture provided in the present application, and as shown in fig. 1, the system architecture of the present application includes a local device 100 and a cloud training system 120 in a cloud environment. The local device 100 may be a server, a virtual machine, or a server cluster, a virtual machine cluster owned by a developer who performs AI development using the local device 100. The cloud training system 120 in the cloud environment may be a background system corresponding to a cloud training service provided by a cloud service provider, and the cloud training system 120 may perform cloud training on an AI model developed locally by a developer by using basic resources in the cloud environment.

The editor 102 and the AI framework 104 are installed in the local device 100, and the guidance system 106 is also installed in the local device 100. The AI framework 104 serves as a library used for AI development and is used for providing functional components for developers in the building and training processes of AI models. The editor 102 is used for a developer to perform editing of the training code, and the editor 102 may also be used for compiling the training code. The developer may invoke the components in the AI framework 104 through the editor 102 in the local device 100 for AI model construction and training. To train the constructed AI model, a developer typically develops training code in a local editor, writes training parameters required for training, and trains program logic to be used during training. Training program logic in the training code indicates how to train the AI model to be trained using the training data employed in the training and according to the set training parameters. Training of the AI model is started when the training code edited by the developer is run.

In order to make the AI model constructed by the developer be trained using the resources on the cloud, the local device is also installed with a guidance system 106. The guidance system 106 is configured to obtain training task information when the training code runs, and upload the training task information to the cloud training system 120 in the cloud environment, so that the cloud training system 120 executes a training task corresponding to the training task information. The training task information may include training parameters, training program logic, training environment information, cloud training access information, and the AI model to be trained, needed to perform training of the AI model. It should be understood that the guidance system 106 described above may be a separate software program installed in the local device 100 or a tool coupled with the AI framework 104. The boot system 106 is illustrated in FIG. 1 as a separate software program.

And the cloud training system 120 in the cloud environment is configured to receive the training task information sent by the guidance system 106, and prepare a cloud training environment by using the training task information, where the set cloud training environment is logically matched with the AI model and the training program. The cloud training system 120 is further configured to receive a training notification sent by the guidance system 106, and train the AI model of the developer in a set cloud training environment. After the training task is completed, the cloud training system may also return a response that the training was successful and/or return the trained AI model to the guidance system 106.

As can be seen from the above, in the present application, the guidance system 106 in the local device 100 is mainly used to interact with the cloud training system 120 in the cloud environment, so as to assign the training task corresponding to the training code developed in the local device 100 to the cloud training system 120 for execution.

The structural and functional partitioning of the guidance system 106 and the cloud training system 120 is described below in conjunction with fig. 2.

Since the structure of the AI model developed by the developer and the training code developed depend on the functional components in the AI framework, the AI framework 104 needs to be called when the training code runs. In an embodiment of the present application, an obtaining component 1042 is embedded in the AI framework 104 used for building and training the AI model, and the obtaining component 1042 can be a tool, a plug-in, or a patch in the AI framework 104.

The obtaining component 1042 is configured to communicate with the guidance system 106 and transmit the obtained initial information of the training task to the guidance system 106. In particular, the obtaining component 1042 is configured to obtain initial information of the training task from the training code when the training code is running. In this embodiment of the application, information acquired by the acquisition component 1042 in the AI framework is collectively referred to as initial information of a training task, and the initial information of the training task may include two types of information, one type of information is identification information or address information of a model or data that needs to be used when the training task is executed, for example: address information of an AI model to be trained, address information of training data, and the like; another is information that can be used directly when performing a training task, such as: training parameters, training program logic, etc. The obtaining component 1042 obtains initial information of the training task and sends the initial information to the guidance system 106.

The guidance system 106 may include a receiving unit 1061, an obtaining unit 1062, and a transmitting unit 1063. It should be understood that the above-described division of the units of the guidance system 106 is merely an example, and does not constitute a limitation of the guidance system 106 of the present application.

The receiving unit 1062 is configured to receive initial information of the training task sent by the obtaining component 1042. The receiving unit 1062 may also receive instructions or information sent by the cloud training system 120.

Since the initial information of the training task includes some identification information or address information of models or data to be used when the training task is performed, the obtaining unit 1064 in the guidance system 106 may be configured to obtain the training task information from the local device 100 according to the initial information of the training task received by the receiving unit 1062. For example: after the receiving unit 1062 receives the address information of the AI model acquired by the acquiring component 1042 of the AI frame, the acquiring unit 1064 acquires the AI model to be trained according to the address information of the AI model. It is to be appreciated that for initial information of training tasks of this type that can be directly used when performing the training tasks, the initial information of the training tasks obtained by the obtaining component 1042 is the same as the content of the corresponding training task information sent by the guidance system 106 to the cloud training system 120. In other words, after the receiving unit 1062 in the guidance system 106 receives the type of information from the obtaining component 1042, the obtaining unit 1064 does not need to further obtain the type of information, and the type of information can be used as the training task information that the guidance system is to send to the cloud training system 120. In the embodiment of the present application, data and information that are sent by the guidance system 106 to the cloud training system 120 for performing a training task are collectively referred to as training task information.

The sending unit 1066 is configured to send the training task information to the cloud training system 120. The sending unit 1066 is further configured to send a training notification and/or training data to the cloud training system 120.

As shown in fig. 2, the cloud training system 120 includes an environment preparation unit 122 and a training task execution unit 124.

The environment preparation unit 122 is configured to receive the training task information sent by the guidance system 1066, and prepare a cloud training environment according to the training task information, for example: and setting resources required to be used when the training task is executed on the cloud and a library dependent on the training. Optionally, the environment preparation unit 122 may also be configured to return a response message that the environment preparation is completed to the guidance system 106. Optionally, the environment preparation unit 122 may also be configured to send part of the training task information and/or training task execution instructions to the training task execution unit 124.

The training task execution unit 124 may execute the training task according to the training task execution instruction sent by the environment preparation unit 122 or the guidance system 106. The training task execution unit 124 mainly executes the training program logic in the training task information to train the AI model to be trained when executing the training task. The training task execution unit 124 is further configured to return a response that the training is successful and/or return a trained AI model to the lead unit.

Before a developer needs to use the cloud training system 120 described in this application to implement training of locally developed AI models by using resources on the cloud, the developer may purchase services for cloud training through a cloud platform of a cloud service provider, configure the total amount of resources available for the training of the cloud training system 120, set authentication key information, and the like. In some embodiments, the guidance system 106 and the acquisition component 1042 in the AI framework 104 can be software programs or tools developed by cloud service providers in kit for providing cloud services for cloud training. The developer, after purchasing and configuring the cloud trained services, may install the boot system 106 and the acquisition component 1042 described above on a local device. Thus, the developer may run the training code locally to initiate the training of the AI model by the cloud training system 120.

Fig. 2 described above is only one possible embodiment of the present application. In other embodiments, the functionality of the get component 1042 in the AI framework 104 can also be provided by the bootstrap system 106, i.e., the aforementioned action of the get component 1042 to get initial information of the training task according to the training code can be performed by the bootstrap system 106. In this case, the guidance system 106 may include four functional units, namely: an acquisition unit that executes the operation of the acquisition unit 1042, and the aforementioned reception unit 1062, acquisition unit 1064, and transmission unit 1066. The foregoing division of the functional units of the guidance system 106 is only an example, and different division manners may be provided, which are not described herein again.

The following embodiments of the present application describe the guidance system 106 without including the functionality of the acquisition component 1042 in the AI framework 104. It is to be appreciated that the contents of the embodiments described below can also be adaptively applied to scenarios in which the guidance system 106 includes the functionality of the acquisition component 1042 in the AI framework 104.

Fig. 3 is a schematic flowchart of an AI training method provided in an embodiment of the present application, and a specific implementation of the AI training method of the present application is specifically described below with reference to fig. 3. The AI training method can be performed in conjunction with the aforementioned guidance system 106, cloud training system 120, and acquisition component 1042 in AI framework 104.

S201: the developer develops and runs the training code.

Specifically, the developer may develop code for training the constructed AI model in an editor, which may be various IDEs in the industry, and the training code may include training parameters set by the developer, such as: learning rate, batch (batch) processing value, batch size, etc., and may also include the name of the training data set, the name of the AI model to be trained, and the training program logic. The training program logic can comprise training program logic written by the developer and can also comprise training program logic in an AI framework called by the developer. Algorithmic logic such as a loss function, optimizer, etc. may be included in the training program logic. It should be appreciated that the AI model built by the developer depends on the AI framework, as does the execution of the training program logic in the training code.

Because this application will be executed by the high in the clouds to the training task of AI model, when the developer was developing the training code, still need set up cloud training mode in the training code, for example: and representing the training mode as a cloud training mode by using codes. The developer may also write cloud training access information in training code, such as: cloud access address information, cloud authentication information, account information, and the like.

The training code developed by the developer can be started and run on the local device after being compiled by the compiler. After the training code is started to run, because the training code is provided with the cloud training mode, in some embodiments, the cloud training mode in the training code may trigger the acquisition component in the AI framework to perform an acquisition operation on the initial information of the training task.

S202: and the acquisition component in the AI framework acquires the initial information of the training task corresponding to the training code according to the training code and sends the initial information of the training task to the guide system.

The initial information of the training task that the acquisition component can acquire from the training code includes: address information of the AI model to be trained. When the obtaining component obtains the address information of the AI model to be trained, an Application Program Interface (API) when the model is loaded in the training code may be intercepted, so as to obtain the path information of the AI model to be trained in the local device.

The obtaining of initial information of the training task that the component can obtain from the training code further comprises: address information of the training data. When the acquisition component acquires the address information of the training data, the acquisition component can also acquire the path information of the training data in the local device by intercepting the API when the training data is loaded in the training code.

The obtaining of initial information of the training task that the component can obtain from the training code further comprises: some training program logic, cloud training access information, training parameters, training environment information, etc. The cloud training access information may include: cloud access address information, cloud authentication information and account information. The training parameters may include: learning rate, batch value, batch size, etc. employed during training. The training environment information may include: version information of the AI framework, programming language version information of the training code, some plug-in or library information of the programming language version or the AI framework version, the specification and amount of resources used to perform the training, etc. The above training environment information can be classified into two types. A kind of training environment information representing local equipment, which represents the environment information when the local equipment constructs AI model and develops training code, includes: version information of the AI framework, programming language version information of the training code, some plug-in or library information of the programming language version or the AI framework version, etc. Another type represents training environment information set in the local device, represents environment information when AI model training is performed, and includes: the resource specifications, amounts, etc. used to perform the training, such information is typically set by the user in the training code.

S203: and the guide system uploads the training task information to the cloud training system.

After the guidance system receives the initial information of the training task sent by the acquisition component in the AI framework, the guidance system can obtain the training task information to be sent according to the initial information of the training task and upload the training task information to the cloud training system. For example: and reading the AI model to be trained from the local equipment according to the address information of the AI model to be trained, and uploading the AI model to the cloud training system. Optionally, the guidance system may also actively detect and acquire some training task information, and this action may be performed by an acquisition unit in the guidance system. For example: when the acquisition component in the AI framework cannot acquire some training environment information (e.g., programming language version of the training code), the guidance system may acquire the training environment information by actively detecting the training environment in the local device. Thus, the training task information includes information obtained from the training code using an obtaining component in the AI framework, and information obtained from the local device based on the information in the training code.

Before the guidance system uploads the acquired training task information to the cloud training system, connection can be established with the cloud training system according to cloud training access information in the training task information, and the connection establishment can include operations of authentication, charging inquiry and the like.

Specifically, the flow diagram of establishing the connection between the guidance system and the cloud training system may specifically include, as shown in fig. 4, the following steps S2031 to S2036:

s2031: the guidance system sends an upload request to the cloud training system.

The cloud training access information included in the upload request includes cloud access information, account information of the local device, and authentication information. The cloud access information may be address information of the cloud training system, and the upload request may be sent to the cloud training system according to the address information of the cloud training system. The account information and the authentication information may be information registered and acquired by a developer when purchasing cloud training services in a cloud platform before using the scheme of the present application. For example: the account information may be a user name of the developer on the cloud platform, and the authentication information may be a key corresponding to the cloud training service acquired from the cloud platform.

It should be noted that in some cases, the upload request of the guidance system may include only the cloud access information, the account information of the local device, and the authentication information for access, authentication, and fee inquiry, and in other cases, the upload request may further include part or all of the aforementioned training task information. If the upload request only includes cloud access information, account information and authentication information, the guidance system may upload other training task information to the cloud training system after receiving a prompt that the authentication and charging query passes.

S2032: and the cloud training system receives the uploading request and sends the account information and the authentication information to the cloud authentication center.

S2033: and the cloud authentication center authenticates the training task requested by the uploading request according to the acquired account information and authentication information, and returns an authentication result.

The specific authentication method may be any available authentication method in the industry, and the present application does not limit this.

And after the cloud authentication center completes authentication, returning an authentication result to the cloud training system.

S2034: and the cloud training system sends the account information to a cloud charging center.

S2035: and the cloud charging center confirms the charge information of the account corresponding to the account information according to the account information and returns the charge information to the cloud training system.

The present application does not limit the execution sequence of steps S2034 to S2035 and steps S2032 to S2033. The execution of steps S2034 to S2035 may also be optional.

S2036: and the cloud training system returns a response that the authentication and charging inquiry is passed to the guide system and receives the training task information.

Under the condition that the authentication is passed and the account cost corresponding to the account information is greater than or equal to the preset threshold value, the cloud training system returns a response that the authentication and charging inquiry is passed to the guide system, and receives other training task information uploaded by the guide system, wherein the other training task information comprises the above-described training task information except the cloud training access information for cloud access, authentication and cost inquiry, such as: training parameters, training program logic, AI models to be trained, training environment information, and the like.

It should be understood that the above steps are optional, and in the case that the upload request includes other training task information, the cloud training system may directly receive the uploaded training task information without returning a response.

Under the condition that the authentication fails or the pre-stored cost of the account corresponding to the account information is less than the preset threshold, the cloud training system may return an upload request failure response to the guidance system, and may also return a request failure reason, for example: authentication fails and/or the pre-stored cost is insufficient.

After the steps S2031 to S2036 are performed, the guidance system may successfully send the training task information to the cloud training system.

It should be understood that the above steps S202-S203 are described based on one of the aforementioned embodiments of the present application (i.e., embodiments in which the guidance system performs the acquisition of the training task information in cooperation with the acquisition component in the AI framework). In another embodiment, the guidance system may include the function of the acquisition component in the AI framework, and both the steps S202 and S203 are performed by the guidance system.

After the cloud training system receives the training task information, the cloud training environment preparation and the training task amount execution work on the cloud can be carried out, which is specifically described in step S204 below:

s204: and the cloud training system executes the training task corresponding to the training task information according to the received training task information.

As shown in fig. 5, specifically, step S204 can be divided into the following steps:

s2041: the cloud training system prepares a cloud training environment according to the received training task information.

Before executing a training task, the cloud training system needs to prepare a cloud training environment, so that the cloud training environment is matched with a local training code and a local AI model to be trained. Specifically, the cloud training system needs to prepare an environment for cloud training according to training environment information in the training task information, where the training environment information may include: version information of the AI framework, version information of the programming language of the training code, some plug-in or library information of the programming language version or the AI framework version, resource specifications for performing the training.

The cloud training system needs to ensure that the AI framework and the programming language version that the cloud training needs to depend on are ready to be executed in the cloud environment according to the versions of the AI framework and the programming language of the training code. Various mainstream AI framework and programming language versions are typically included in the cloud environment, and thus, typically, the cloud training system only needs to detect and validate when preparing the cloud training environment, and does not need to temporarily install these versions.

The cloud training system also needs to ensure that the plug-ins or libraries are installed in the cloud environment according to information of some plug-ins or libraries of the programming language version or the AI framework version, generally, the plug-ins and the libraries needed by the mainstream AI framework and programming software are updated and downloaded in time in the cloud environment, and if the plug-ins and libraries needed for executing the training task are not installed in the cloud environment preparation stage, the plug-ins and libraries can be downloaded and installed in time.

The cloud training system also needs to prepare corresponding training resources on the cloud according to the information of the resource specification for executing training included in the training task information. For example: according to the information of the required resource specification, the relevant virtual machine and container are started at the cloud, and corresponding hardware resources, such as a Graphic Processing Unit (GPU) or an AI training chip, are mounted.

After the cloud training environment preparation is complete, the cloud training system may perform the following steps:

s2042: the cloud training system returns an environment preparation success response to the bootstrap system.

S2043: the guidance system sends a training notification to the cloud training system.

It is noted that, in other embodiments, steps S2042 and S2043 may not be executed. For example: the guidance system may notify the cloud training system to execute the training task when uploading the training task information, and the cloud training system may start to execute the training task after the cloud training system has executed the step S2041, so that the steps S2042 and S2043 are omitted.

S2044: the cloud training system executes a training task corresponding to the training task information in a cloud training environment.

The cloud training system can start a training container prepared with relevant resources to execute a training task, and specifically, when the training task is executed, functional components in a corresponding AI framework on the cloud are called according to a training program logic. Inputting training data into an AI model to be trained, calculating the training data by using each component in the model based on training resources, updating the values of parameters in the model according to some training parameters and training program logic, iterating in this way, stopping training the model until the training of the AI model reaches a training stopping condition, and obtaining the trained AI model, wherein the training stopping condition is as follows: the loss function converges to less than a preset threshold, or alternatively, the number of rounds of training reaches a preset value.

Because training data (for example, tens of thousands of pictures or tens of thousands of videos) used for training the AI model are more, when training task information is uploaded, a guide system reads a training data set in local equipment and uploads the training data set to a cloud training system at one time, which may result in higher transmission delay and longer environment preparation time of the cloud training system, and affect user experience. In some embodiments, the cloud training system may send a training data acquisition request to the bootstrap device at least once in the course of performing the training.

That is, optionally, the following steps may be further performed in the process of executing the training task:

s2045: the cloud training system sends a training data acquisition request to the guide system;

s2046: the guiding system reads the training data from the local equipment and sends the training data to the cloud training system.

In other embodiments, the training data used to train the AI model may also be pre-saved by the user where the cloud training system can read, such as: and (4) cloud storage.

It should be noted that, in the process of executing the training task, the cloud charging center may also continuously charge according to the duration of the resources used in the training, the resource specification, the number of the resources, and the like.

Through the step S204, the cloud training system can successfully train the AI model to obtain the trained AI model.

S205: and the cloud training system returns the trained AI model to the guide system.

It is to be noted that the step S205 is only a step performed in one case, and in other cases, the cloud training system may not return the trained AI model to the guidance system after the training task is performed. For example: the cloud training system may return a response that the training is successful to the bootstrap system, or return address information that the trained AI model is stored in the cloud environment to the bootstrap system. What the cloud training system returns to the bootstrap system after training is completed can be determined by a developer through preset. Optionally, the cloud training system may also return a charging ticket to the guidance system.

Through the steps S201 to S205, the developer writes and runs the training code locally, so that the AI model can be trained by using the resources of the cloud environment. The problem that the training of the AI model cannot be supported due to insufficient resources required by local training is solved. The method greatly facilitates developers, and the developers do not need to change the habits of locally constructing the AI model and developing the training codes under the condition of facing insufficient local resources. According to the scheme, a developer does not need to perform complex configuration and adaptation, and cloud training is quickly realized through cooperation of the guide system and the cloud training system.

The embodiment of the present application further provides the guidance system 106 shown in fig. 2, and in some embodiments, the guidance system 106 is specifically configured to perform the steps performed by the guidance systems shown in fig. 3 to fig. 5, and the functions of the functional units of the guidance system 106 are as described in the foregoing description of fig. 2, and are not described again here. In other embodiments, the guidance system 106 may also be specifically adapted to perform the functionality of the acquisition components in the guidance system and AI framework described above and illustrated in FIGS. 3-5.

The embodiment of the present application further provides the cloud training system 120 shown in fig. 2, where the cloud training system 120 may be specifically configured to perform the steps performed by the cloud training systems shown in fig. 3 to fig. 5, and the functions of each functional unit of the cloud training system 120 are as described in the foregoing description of fig. 2, and are not described again here.

The embodiment of the present application further provides a computing device 300 as shown in fig. 6, where the computing device 300 may be the aforementioned local device. Computing device 300 includes memory 301, processor 302, communication interface 303, and bus 304. The memory 301, the processor 302 and the communication interface 303 are connected to each other by a bus 304. It should be understood that the present application is not limited to the number of processors, memories in the computing device 300. Computing device 300 may also represent a cluster of devices made up of multiple servers or virtual machines.

The Memory 301 may be a Read Only Memory (ROM), a static Memory device, a dynamic Memory device, or a Random Access Memory (RAM). The memory 301 may store computer instructions that, when executed by the processor 302, stored in the memory 301, the processor 302 and the communication interface 303 perform some or all of the AI training methods described in fig. 3-5 and performed by the guidance system described previously. I.e., the computer instructions of the aforementioned guidance system 106 may be stored in the memory 301. The memory 301 may also store AI modules to be trained and training data.

The processor 302 may be a general-purpose Central Processing Unit (CPU), an Application Specific Integrated Circuit (ASIC), a Graphics Processing Unit (GPU), or any combination thereof. The processor 302 may include one or more chips, and the processor 302 may include an AI accelerator, such as: a neural Network Processor (NPU).

Communication interface 303 enables communication between computing device 300 and other devices or communication networks using transceiver modules, such as, but not limited to, transceivers. For example, a response that the training was successful or a trained AI model may be obtained through the communication interface 303.

Bus 304 may include a path that transfers information between components of computing device 300 (e.g., memory 301, processor 302, communication interface 303).

An embodiment of the present application further provides a computing device 400 as shown in fig. 7, where the computing device 400 may be a cloud server or a cloud server cluster provided by a cloud service provider, and may also be a virtual machine or a virtual machine cluster. Computing device 400 includes memory 401, processor 402, communication interface 403, and bus 404. The possible hardware structures of the memory 401, the processor 402, the communication interface 403 and the bus 404 and the relationship between the parts may be the same as or similar to the corresponding parts in the computing device 300, and are not described herein again. The memory 401 in the computing device 400 may store the aforementioned computer instructions included in the environment preparation unit 122 and the training task execution unit 124 in the cloud training system 120, and when the computer instructions stored in the memory 401 are executed by the processor 402, the processor 402 and the communication interface 403 perform part or all of the aforementioned AI training method performed by the cloud training system described in fig. 3-5.

The descriptions of the flows corresponding to the above-mentioned figures have respective emphasis, and for parts not described in detail in a certain flow, reference may be made to the related descriptions of other flows.

In the above embodiments, the implementation may be wholly or partially realized by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product implementing the AI training method described above includes one or more computer program instructions that, when loaded and executed on a computer, fully or partially perform the method flow of AI training according to fig. 3-5 described previously herein.

The computer may be a general purpose computer, a special purpose computer, a network of computers, or other programmable device. The computer instructions may be stored on a computer readable storage medium or transmitted from one computer readable storage medium to another, for example, from one website, computer, server, or data center to another website, computer, server, or data center via wire (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device, such as a server, a data center, etc., that incorporates one or more of the available media. The usable medium may be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., Digital Versatile Disk (DVD)), or a semiconductor medium (e.g., Solid State Disk (SSD)), among others.

Claims

1. An Artificial Intelligence (AI) training method is applied to a guidance system, and after a user of the cloud training system triggers the running of a training code on a local device, the method comprises the following steps:

acquiring training task information according to the training code running on the local equipment, wherein the training code is used for training an AI model, and the AI model is developed and acquired by the user based on an AI framework installed on the local equipment;

uploading the training task information to the cloud training system, and informing the cloud training system to execute a training task corresponding to the training task information.

2. The method of claim 1, wherein the training task information comprises information obtained from the training code using an obtaining component in the AI framework and information obtained from the local device based on information in the training code.

3. The method of claim 1 or 2, wherein the training task information comprises one or more of the following data: the training parameters in the training codes, the AI model, the training program logic in the training codes for training the AI model, the training environment information of the local device, and the cloud training access information for connecting with the cloud training system.

4. The method of claim 3, wherein the training environment information of the local device comprises: version information of the AI framework, and/or version information of a programming language of the training code.

5. The method according to any one of claims 1-4, further comprising:

receiving a training data acquisition request sent by the cloud training system in the process of executing the training task;

and acquiring the training data according to the training data acquisition request, and sending the training data to the cloud training system.

6. The method according to any one of claims 1-5, wherein before notifying the cloud training system to execute a training task corresponding to the training task information, the method further comprises:

and receiving an environment preparation success response returned by the cloud training system.

7. The method according to any one of claims 1-6, further comprising: and receiving the AI model which is returned by the cloud training system and is finished in training.

8. The method of any of claims 1-7, wherein the guidance system is obtained from the cloud training system and installed in the local device.

9. An AI training method is applied to a cloud training system and comprises the following steps:

acquiring training task information sent by a guidance system after a user triggers a running training code at local equipment, wherein the training task information comprises training environment information of the local equipment;

performing preparation of a cloud training environment according to the training environment information;

and executing the training task corresponding to the training task information based on the cloud training environment.

10. The method of claim 9, wherein the training environment information of the local device comprises: version information of an AI framework on which an AI model to be trained depends, and/or version information of a programming language used by the training code for training the AI model.

11. The method of claim 10, wherein the performing preparation of a cloud training environment from the training environment information comprises:

and setting the AI frame and the programming language used for executing the training task in the cloud training environment according to the version information of the AI frame and the version information of the programming language of the training code.

12. The method of any of claims 9-11, wherein the training task information further comprises: training parameters in the training codes, the AI model and training program logic used for training the AI model in the training codes;

the executing of the training task corresponding to the training task information based on the cloud training environment includes: performing training of the AI model in the prepared cloud training environment according to the training parameters and the training program logic.

13. The method of any of claims 9-12, wherein the training task information further comprises: cloud training access information, before performing preparation of a cloud training environment according to the training environment information, the method further comprising:

and performing authentication and/or charging query on the training task corresponding to the training task information according to the cloud training access information.

14. A guidance system, comprising:

the acquisition module is used for acquiring training task information according to a training code running on local equipment after a user of the cloud training system triggers the running of the training code on the local equipment, wherein the training code is used for training an AI model, and the AI model is developed and acquired by the user based on an AI framework installed on the local equipment;

and the sending module is used for uploading the training task information to the cloud training system and informing the cloud training system to execute the training task corresponding to the training task information.

15. The system of claim 14, wherein the training task information comprises information obtained from the training code using an obtaining component in the AI framework and information obtained from the local device based on information in the training code.

16. The system of claim 14 or 15, wherein the training task information comprises one or more of the following data: the training parameters in the training codes, the AI model, the training program logic in the training codes for training the AI model, the training environment information of the local device, and the cloud training access information for connecting with the cloud training system.

17. The system of claim 16, wherein the training environment information of the local device comprises: version information of the AI framework, and/or version information of a programming language of the training code.

18. The system according to any of claims 14-17, wherein the system further comprises a receiving unit,

the receiving unit is used for receiving a training data acquisition request sent by the cloud training system in the process of executing the training task;

the acquiring unit is further configured to acquire the training data according to the training data acquisition request;

the sending unit is further configured to send the training data to the cloud training system.

19. The system according to any one of claims 14-18, wherein the system further comprises a receiving unit,

the receiving unit is configured to receive an environment preparation success response returned by the cloud training system before the sending unit notifies the cloud training system to execute the training task corresponding to the training task information.

20. The system according to any of claims 14-19, wherein the system further comprises a receiving unit,

the receiving unit is used for receiving the AI model which is returned by the cloud training system and is finished in training.

21. The system of any one of claims 14-20, wherein the guidance system is obtained from the cloud training system and installed in the local device.

22. A cloud training system, comprising:

the system comprises an environment preparation unit, a training unit and a training unit, wherein the environment preparation unit is used for acquiring training task information sent by a guide system after a user triggers and runs a training code on local equipment, and the training task information comprises training environment information of the local equipment; performing preparation of a cloud training environment according to the training environment information;

and the training task execution unit is used for executing the training task corresponding to the training task information based on the cloud training environment.

23. The system of claim 22, wherein the training environment information of the local device comprises: version information of an AI framework on which the AI model to be trained depends, and/or version information of a programming language of the training code used to train the AI model.

24. The system according to claim 23, wherein the environment preparation unit is specifically configured to set an AI framework and a programming language used for executing a training task in the cloud training environment according to version information of the AI framework and version information of a programming language of the training code.

25. The system of any of claims 22-24, wherein the training task information further comprises: training parameters in the training codes, the AI model and training program logic used for training the AI model in the training codes;

the training task execution unit is specifically configured to execute training of the AI model in the prepared cloud training environment according to the training parameters and the training program logic.

26. The system of any of claims 22-25, wherein the training task information further comprises: the cloud training access information, the environment preparation unit, further configured to: and performing authentication and/or charging query on the training task corresponding to the training task information according to the cloud training access information.

27. A computing device comprising a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions to cause the computing device to perform the method of any preceding claim 1 to 8 or to perform the method of any preceding claim 9 to 13.

28. A computer-readable storage medium, characterized in that the computer-readable storage medium stores computer program code which, when executed by a computing device, performs the method of any of the preceding claims 1-8 or performs the method of any of the preceding claims 9-13.