WO2023125594A1 - Ai模型传输方法、装置、设备及存储介质 - Google Patents

Ai模型传输方法、装置、设备及存储介质 Download PDF

Info

Publication number
WO2023125594A1
WO2023125594A1 PCT/CN2022/142586 CN2022142586W WO2023125594A1 WO 2023125594 A1 WO2023125594 A1 WO 2023125594A1 CN 2022142586 W CN2022142586 W CN 2022142586W WO 2023125594 A1 WO2023125594 A1 WO 2023125594A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
information
conversion
platform
sending
Prior art date
Application number
PCT/CN2022/142586
Other languages
English (en)
French (fr)
Inventor
崇卫微
程思涵
Original Assignee
维沃移动通信有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 维沃移动通信有限公司 filed Critical 维沃移动通信有限公司
Publication of WO2023125594A1 publication Critical patent/WO2023125594A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04WWIRELESS COMMUNICATION NETWORKS
    • H04W24/00Supervisory, monitoring or testing arrangements
    • H04W24/02Arrangements for optimising operational condition
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F8/00Arrangements for software engineering
    • G06F8/40Transformation of program code
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04WWIRELESS COMMUNICATION NETWORKS
    • H04W28/00Network traffic management; Network resource management
    • H04W28/02Traffic management, e.g. flow control or congestion control
    • H04W28/06Optimizing the usage of the radio link, e.g. header compression, information sizing, discarding information

Definitions

  • the present application belongs to the technical field of communication, and in particular relates to an AI model transmission method, device, equipment and storage medium.
  • AI models are implemented in different ways (different frameworks are used), resulting in incompatibility of generated files.
  • Pytorch and TensorFlow are two framework platforms used to build AI network models.
  • the generated model is saved in the form of ".pth” and ".meta/.index", and these two files can only be read, called and other operations by their own framework, which makes the AI model not directly cross-framework platform or Use across users.
  • Embodiments of the present application provide an AI model transmission method, device, device, and storage medium, which can solve the problem of incomparable intelligent characteristics caused by the inability of different devices to communicate with each other when AI functions are introduced into the network in a wireless communication network scenario.
  • a well-implemented problem A well-implemented problem.
  • an AI model transmission method which is applied to a model receiving end device, and the method includes:
  • the model receiver device sends a first request message, where the first request message is used to request a first AI model, and the first request message includes AI model description information;
  • the model receiver device receives a first response message
  • the first response message includes one of the following:
  • the first AI model is a model obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and the first AI model can be used by the model receiving end device;
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission method which is applied to a model sending end device, and the method includes:
  • the model sending end device receives an AI model request message, the AI model request message is used to request a first AI model, and the AI model request message includes AI model description information;
  • the model sending end device performs a first operation according to the AI model request message
  • said performing the first operation includes at least one of the following:
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot converting the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission method which is applied to the model conversion platform, and the method includes:
  • the model conversion platform receives the second AI model sent by the model sending device
  • the model conversion platform converts the second AI model into a first AI model, and sends the first AI model to a model receiving device; or,
  • the model conversion platform sends first indication information or second indication information to the model receiving end device, the first indication information is used to indicate that the model sending end device cannot provide the second AI model, and the second indication The information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model;
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission method which is applied to a model sending device, and the method includes:
  • the model sending end device sends a third request message to the model conversion platform, where the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the second AI model corresponds to the second model representation method
  • the first AI model corresponds to the first model representation method
  • an AI model transmission method which is applied to a model conversion platform, and the method includes:
  • the model conversion platform receives a third request message sent by the model sending device, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the model conversion platform performs a second operation according to the third request message
  • said performing the second operation includes at least one of the following:
  • an AI model transmission device including:
  • a first sending unit configured to send a first request message, where the first request message is used to request a first AI model, where the first request message includes AI model description information;
  • a first receiving unit configured to receive a first response message
  • the first response message includes one of the following:
  • the first AI model is a model obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and the first AI model can be used by the model receiving end device;
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission device including:
  • the second receiving unit is configured to receive an AI model request message, the AI model request message is used to request the first AI model, and the AI model request message includes AI model description information;
  • a first executing unit configured to execute a first operation according to the AI model request message
  • said performing the first operation includes at least one of the following:
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot converting the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission device including:
  • the fourth receiving unit is configured to receive the second AI model sent by the model sending device
  • a first processing unit configured to convert the second AI model into a first AI model, and send the first AI model to a model receiving device;
  • a fifth sending unit configured to send first indication information or second indication information to the model receiver device, where the first indication information is used to indicate that the model sender device cannot provide the second AI model, and the first The second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model;
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • an AI model transmission device including:
  • a tenth sending unit configured to send a third request message to the model conversion platform, where the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the second AI model corresponds to the second model representation method
  • the first AI model corresponds to the first model representation method
  • an AI model transmission device including:
  • a tenth receiving unit configured to receive a third request message sent by the model sending device, where the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • a second executing unit configured to execute a second operation according to the third request message
  • said performing the second operation includes at least one of the following:
  • a model receiving end device includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and the programs or instructions are executed by the When executed by the processor, the steps of the method described in the first aspect are realized.
  • a model receiver device including a processor and a communication interface, wherein the communication interface is used to send a first request message, and the first request message is used to request a first AI model, so The first request message includes AI model description information; the communication interface is also used to receive a first response message; wherein, the first response message includes one of the following: the first AI model, the first AI
  • the model is a model obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and the first AI model can be used by the model receiving end device; the first indication information is used to indicate the model sending end device The second AI model cannot be provided; the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model; wherein, the first AI model corresponds to the first model representation method, the second AI model corresponds to the second model representation method.
  • a model sender device in a thirteenth aspect, includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and the program or instructions are executed by the The processor implements the steps of the method described in the second aspect or the fourth aspect when executed.
  • a model sending end device including a processor and a communication interface, wherein the communication interface is used to receive an AI model request message, and the AI model request message is used to request a first AI model, so
  • the AI model request message includes AI model description information;
  • the processor is configured to perform a first operation according to the AI model request message; wherein, the execution of the first operation includes at least one of the following: selecting or training to generate a second AI model, and send the second AI model to the model conversion platform, wherein the second AI model is converted by the model conversion platform to obtain the first AI model, and the first AI model can be used by the model receiving end device ; sending first indication information or second indication information, the first indication information is used to indicate that the model sending end device cannot provide the second AI model, and the second indication information is used to indicate the model conversion platform
  • the second AI model cannot be converted into the first AI model; wherein, the first AI model corresponds to the first model representation method, and the second AI model corresponds to the second model representation method.
  • the communication interface is used to send a third request message to the model conversion platform, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model; wherein, the second The AI model corresponds to the second model representation method, and the first AI model corresponds to the first model representation method.
  • a fifteenth aspect provides a model conversion platform, the model conversion platform includes a processor and a memory, the memory stores programs or instructions that can run on the processor, and the program or instructions are executed by the processor When executed, the steps of the method described in the third aspect or the fifth aspect are realized.
  • a model conversion platform including a processor and a communication interface, wherein the communication interface is used to receive the second AI model sent by the model sending end device; the processor is used to convert the second AI model The AI model is converted into a first AI model, and the first AI model is sent to the model receiving end device; the communication interface is also used to: send the first instruction information or the second instruction information to the model receiving end device, The first indication information is used to indicate that the model sending end device cannot provide the second AI model, and the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model.
  • AI model wherein, the first AI model corresponds to the first model representation method, and the second AI model corresponds to the second model representation method.
  • the communication interface is used to receive a third request message sent by the model sending end device, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the processor uses Performing the second operation according to the third request message; wherein the performing the second operation includes at least one of the following: determining whether the conversion operation of the second AI model can be completed; sending the model to the model sending terminal device Sending a second model conversion response message, where the second model conversion response message is used to indicate whether the model conversion platform can complete the conversion operation on the second AI model; convert the second AI model into the first AI model; Send the first AI model to the model receiving end device.
  • an AI model transmission system including: a model receiving end device, a model sending end device, and a model conversion platform, and the model receiving end device can be used to execute the AI model transmission method as described in the first aspect
  • the step of the model sending end device may be used to execute the steps of the AI model transmission method described in the second aspect or the fourth aspect.
  • a readable storage medium on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the method as described in the first aspect are implemented, or The steps of the method described in the second aspect, or the implementation of the steps of the method described in the third aspect, or the implementation of the steps of the method described in the fourth aspect, or the implementation of the steps of the method described in the fifth aspect.
  • a chip in a nineteenth aspect, there is provided a chip, the chip includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run a program or an instruction to implement the method described in the first aspect The method, or implement the method as described in the second aspect, or implement the method as described in the third aspect, or implement the method as described in the fourth aspect, or implement the method as described in the fifth aspect.
  • a computer program/program product is provided, the computer program/program product is stored in a storage medium, and the computer program/program product is executed by at least one processor to implement the The steps of the method, or realize the steps of the method as described in the second aspect, or realize the steps of the method as described in the third aspect, or realize the steps of the method as described in the fourth aspect, or realize the steps of the method as described in the fifth aspect The steps of the method.
  • the model receiver device sends a first request message, the first request message is used to request the first AI model, the first request message includes AI model description information, receives the first response message, and the second A response message includes the first AI model, the first indication information or the second indication information, the first AI model is obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and can be used by the model receiving end device
  • the model makes it possible for the AI model transfer and intercommunication of different platform frameworks, and provides a solution for the AI model transfer of devices from different manufacturers.
  • FIG. 1 is a block diagram of a wireless communication system to which an embodiment of the present application is applicable;
  • Fig. 2 is the schematic diagram of neuron
  • FIG. 3 is one of the schematic flow charts of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 4 is the second schematic flow diagram of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 5 is the third schematic flow diagram of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 6 is one of the interactive flow diagrams of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 7 is the second schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 8 is the third schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 9 is the fourth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • Fig. 10 is a schematic diagram of reporting capability information to the public transformation platform provided by the embodiment of the present application.
  • Fig. 11 is the fifth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 12 is the sixth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 13 is the fourth schematic flow diagram of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 14 is the fifth schematic flow diagram of the AI model transmission method provided by the embodiment of the present application.
  • Fig. 15 is the seventh schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 16 is the eighth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • FIG. 17 is one of the structural schematic diagrams of the AI model transmission device provided by the embodiment of the present application.
  • Fig. 18 is the second structural schematic diagram of the AI model transmission device provided by the embodiment of the present application.
  • Fig. 19 is the third structural schematic diagram of the AI model transmission device provided by the embodiment of the present application.
  • FIG. 20 is the fourth structural schematic diagram of the AI model transmission device provided by the embodiment of the present application.
  • Fig. 21 is the fifth structural schematic diagram of the AI model transmission device provided by the embodiment of the present application.
  • FIG. 22 is a schematic structural diagram of a communication device provided by an embodiment of the present application.
  • FIG. 23 is a schematic diagram of the hardware structure of a model receiver device implementing an embodiment of the present application.
  • FIG. 24 is a schematic diagram of a hardware structure of a model sending end device implementing an embodiment of the present application.
  • first, second and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific sequence or sequence. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application are capable of operation in sequences other than those illustrated or described herein and that "first" and “second” distinguish objects. It is usually one category, and the number of objects is not limited. For example, there may be one or more first objects.
  • “and/or” in the description and claims means at least one of the connected objects, and the character “/” generally means that the related objects are an "or” relationship.
  • LTE Long Term Evolution
  • LTE-Advanced LTE-Advanced
  • LTE-A Long Term Evolution-Advanced
  • CDMA Code Division Multiple Access
  • TDMA Time Division Multiple Access
  • FDMA Frequency Division Multiple Access
  • OFDMA Orthogonal Frequency Division Multiple Access
  • SC-FDMA Single-carrier Frequency Division Multiple Access
  • system and “network” in the embodiments of the present application are often used interchangeably, and the described technology can be used for the above-mentioned system and radio technology, and can also be used for other systems and radio technologies.
  • NR New Radio
  • the following description describes the New Radio (NR) system for illustrative purposes, and uses NR terminology in most of the following descriptions, but these techniques can also be applied to applications other than NR system applications, such as the 6th generation (6 th Generation, 6G) communication system.
  • 6G 6th Generation
  • Fig. 1 shows a block diagram of a wireless communication system to which the embodiment of the present application is applicable.
  • the wireless communication system includes a terminal 11 and a network side device 12 .
  • the terminal 11 can be a mobile phone, a tablet computer (Tablet Personal Computer), a laptop computer (Laptop Computer) or a notebook computer, a personal digital assistant (Personal Digital Assistant, PDA), a palmtop computer, a netbook, a super mobile personal computer (ultra-mobile personal computer, UMPC), mobile Internet device (Mobile Internet Device, MID), augmented reality (augmented reality, AR) / virtual reality (virtual reality, VR) equipment, robot, wearable device (Wearable Device) , vehicle equipment (VUE), pedestrian terminal (PUE), smart home (home equipment with wireless communication functions, such as refrigerators, TVs, washing machines or furniture, etc.), game consoles, personal computers (personal computers, PCs), teller machines or self-service Wearable devices include: smart watches, smart bracelets, smart headphones, smart glasses, smart jewelry (
  • the network side device 12 may include an access network device or a core network device, where the access network device 12 may also be called a radio access network device, a radio access network (Radio Access Network, RAN), a radio access network function, or Wireless access network unit.
  • RAN Radio Access Network
  • RAN Radio Access Network
  • Wireless access network unit Wireless access network unit
  • the access network device 12 may include a base station, a WLAN access point, or a WiFi node, etc., and the base station may be called a Node B, an evolved Node B (eNB), an access point, a Base Transceiver Station (Base Transceiver Station, BTS), a radio Base station, radio transceiver, Basic Service Set (BSS), Extended Service Set (ESS), Home Node B, Home Evolved Node B, Transmitting Receiving Point (TRP) or all As long as the same technical effect is achieved, the base station is not limited to a specific technical vocabulary. It should be noted that in this embodiment of the application, only the base station in the NR system is used as an example for introduction, and The specific type of the base station is not limited.
  • Core network equipment may include but not limited to at least one of the following: core network nodes, core network functions, mobility management entities (Mobility Management Entity, MME), access mobility management functions (Access and Mobility Management Function, AMF), session management functions (Session Management Function, SMF), User Plane Function (UPF), Policy Control Function (Policy Control Function, PCF), Policy and Charging Rules Function (PCRF), edge application service Discovery function (Edge Application Server Discovery Function, EASDF), unified data management (Unified Data Management, UDM), unified data storage (Unified Data Repository, UDR), home subscriber server (Home Subscriber Server, HSS), centralized network configuration ( Centralized network configuration, CNC), network storage function (Network Repository Function, NRF), network exposure function (Network Exposure Function, NEF), local NEF (Local NEF, or L-NEF), binding support function (Binding Support Function, BSF), application function (Application Function, AF), etc. It should be noted that, in the embodiment of the present application, only the core
  • AI models can be implemented in a variety of algorithms, such as neural networks, decision trees, support vector machines, and Bayesian classifiers. This application uses a neural network as an example for illustration, but does not limit the specific type of AI module.
  • a neural network consists of many neurons. X1, X2...Xn, etc. are the input values, and Y is the output result. Each neuron is also the place for calculations, and the results will continue to be passed to the next layer.
  • An input layer, a hidden layer, and an output layer composed of these many neurons are a neural network. The number of hidden layers, the number of neurons in each layer is the "network structure" of the neural network.
  • the neural network is composed of neurons, and the schematic diagram of the neurons is shown in FIG. 2 .
  • a1, a2,...aK that is, X1...Xn above
  • w is the weight (multiplicative coefficient)
  • b is the bias (additive coefficient)
  • ⁇ (.) is the activation function
  • z is the output value.
  • Common activation functions include Sigmoid, tanh, ReLU (Rectified Linear Unit, linear rectification function, corrected linear unit) and so on.
  • the combination of the parameter information of each neuron and the algorithm used is the "parameter information" of the entire network, which is also an important part of the AI model file.
  • an AI model refers to a file containing elements such as network structure and parameter information.
  • the trained AI model can be directly used again by its framework platform without repeated construction or learning, and can be directly judged, identified, etc. Intelligent function.
  • TensorFlow There are many implementation frameworks for neural networks, including TensorFlow, PyTorch, Keras, MXNet, Caffe2, etc.
  • Each framework has a different focus.
  • Caffe2 and Keras are high-level deep learning frameworks that can quickly verify models.
  • TensorFlow and PyTorch It is the underlying deep learning framework that can modify the underlying details of the neural network.
  • PyTorch focuses on supporting dynamic graph models
  • TensorFlow focuses on supporting multiple hardware and running fast
  • Caffe2 focuses on lightweight.
  • Each implementation framework will use its own method to describe the neural network and complete operations such as network construction, training, and inference.
  • the description methods of models under different implementation frameworks cannot be understood by other frameworks, resulting in the interoperability of the models between them.
  • ONNX is a relatively general AI model description language or model expression method.
  • ONNX itself is just a data structure, excluding implementation solutions, used to describe an AI network.
  • ONNX defines a set of standard formats that are independent of the environment and platform, providing a basis for the interoperability of AI models, so that AI models can be used interactively in different frameworks and environments.
  • ONNX describes each operator of the network as a node.
  • the input and output names of each node are globally unique, and the structure of the entire network is described by the matching relationship between the input and output names.
  • All weight parameters are regarded as input or output, and are also retrieved by name.
  • the specific weight value is stored in a separate location, and the corresponding parameter is obtained from the storage location according to the name of the input and output of each node.
  • the frameworks of each AI platform are independent of each other, and only a few of the generated models can be transformed into each other. As a result, their models cannot communicate with each other, and there is no way to use them on other frameworks, which is a big obstacle in actual use. Because the focus of each framework and even the supported development language are different, two nodes using different frameworks cannot transmit the information of the AI model, because the two nodes using different frameworks have different description methods for the network, and the data compression method Different, the file saving format is also different, and the network trained by other frameworks cannot be parsed. Similarly, different developers define and describe frameworks in different ways, and there are high barriers to each other, making it difficult to convert each other. Even in the same framework, different versions and other information will lead to different file saving formats. When two nodes with different framework versions transmit the AI model, the result of the analysis will be biased or cannot be parsed.
  • ONNX is intended to be developed as a general solution for describing AI models, it has been supported by some AI platforms, but not fully supported by all platforms. If you want to develop it into a fully universal standardized solution, you need all platforms (all terminals, network elements, and application servers) to be equipped with ONNX, which may cause unnecessary increases in hardware and software costs.
  • the model receiver device may be one or more terminals, which need to request a corresponding AI model from the model sender device due to AI service requirements.
  • the model sending end device can be a network element or a third-party service server in one or more communication networks, which has sufficient computing power to perform model training for AI services and provide the model to model demanders.
  • model receiving device Due to hardware and software conditions or cost constraints, the model receiving device only has the ability to understand a few AI model frameworks, and cannot understand AI models of various frameworks from various model sending devices, so it cannot be used directly.
  • a model conversion platform may be deployed in a unified manner (for example, terminal device manufacturers uniformly deploy a cloud platform), which converts the AI models from the model sender devices into the model types supported by them and sends them to Each model receiver device.
  • model conversion platform can also be deployed behind the model sender device, for example, the model sender device is uniformly deployed on the cloud platform.
  • the model transformation platform is a public transformation platform, and both the model sending end device and the model receiving end device can interact with the model transformation platform, for example, register and report their own attribute information to the model transformation platform, such as identification address, address information and /or supported AI framework information.
  • model receiving end device in the embodiment of the present application is not limited to a terminal, and the model receiving end device may also be other devices that request AI models.
  • the model sending end device is not limited to a network element or a third-party service server in the communication network, and may also be other devices that can provide AI models.
  • Figure 3 is one of the schematic flow diagrams of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 3, the method is applied to the model receiving end device, and the method includes:
  • Step 300 the model receiver device sends a first request message, where the first request message is used to request a first AI model, and the first request message includes AI model description information.
  • model receiving end device sends the first request message to request the first AI model.
  • the first request message includes AI model description information, where the AI model description information is used to indicate information of the requested AI model, that is, the AI model description information is used to describe the first AI model.
  • the first AI model corresponds to the first model representation method.
  • the AI model description information includes at least one of the following:
  • AI model type information such as analytic ID
  • AI model identification information such as model ID
  • AI model algorithm information that is, the algorithm used by the model, such as neural network, random forest, etc.
  • the training object information corresponding to the AI model such as a certain user, a certain area AOI;
  • the time information corresponding to the AI model means that the model is requested for a specific time period and time point;
  • AI model training accuracy requires information, such as model accuracy of 90%.
  • the first request message must carry AI model description information.
  • the first request message also includes at least one of the following:
  • Attribute information of the model receiver device including the identification information, address information and/or supported or selected AI framework information of the model receiver device;
  • the attribute information of the model sender device including the identification information, address information and/or supported or selected AI framework information of the model sender device;
  • the attribute information of the model conversion platform including the identification information, address information and/or conversion capability information of the model conversion platform, where the conversion capability information is used to indicate the type of AI framework supported, and the ability to convert these AI framework types into model receivers
  • the capability of the type of AI framework supported by the device should be able to support all mainstream frameworks;
  • the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model, such as delay requirement (transformation is completed and sent within 5s), accuracy requirement (transformation accuracy rate of more than 99%), etc.
  • the AI framework information is used to describe the implementation framework of the AI model.
  • the AI framework information includes the type information of the AI framework, network topology information, data compression method information, design motivation information, data format information, and language information used. at least one type of information.
  • AI frameworks such as TensorFlow, PyTorch, Keras, MXNet, Caffe2 and other frameworks.
  • the model receiver device sends a first request message, including:
  • the model receiver device sends the first request message to the model sender device; or,
  • the model receiver device sends the first request message to the model conversion platform.
  • the model receiving end device may send the first request message to the model sending end device or the model conversion platform;
  • the model receiving device sends the first request message to the model sending device.
  • Step 301 the model receiver device receives a first response message
  • the first response message includes one of the following:
  • the first AI model is a model obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and the first AI model can be used by the model receiving end device;
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the model receiver device After sending the first request message, the model receiver device will receive the first response message.
  • the content of the first response message may be that the request is accepted, that is, the model receiving end device can obtain the first AI model.
  • the content of the first response message may also be that the request is rejected, that is, the model receiving device cannot obtain the first AI model, and the model receiving device will receive the first indication information or the second indication information. Through the first indication information or the second indication information, the model receiver device can learn that the request is rejected.
  • the first AI model includes a file of complete network structure and parameter information of the AI model.
  • an AI model generated using TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • the first indication information is used to indicate that the model sender device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model from the model sender device into the requested AI model.
  • the first AI model is used to indicate that the model sender device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model from the model sender device into the requested AI model.
  • the model receiver device receiving the first response message includes:
  • the model receiver device acquires the first response message from the model conversion platform, where the first response message includes the first AI model or the first indication information or the second indication information.
  • the model receiving end device can obtain the first AI model or the first instruction information from the model conversion platform, Or the second instruction information.
  • the model receiver device receiving the first response message includes:
  • the model receiver device acquires the first response message from the model sender device, where the first response message includes the first indication information or the second indication information.
  • the model receiver device may obtain the first indication information or the second indication information from the model sender device.
  • the model receiver device sends a first request message, the first request message is used to request the first AI model, the first request message includes AI model description information, receives the first response message, and the second A response message includes the first AI model, the first instruction information or the second instruction information, the first AI model is obtained after the second AI model from the model sending device is transformed by the model conversion platform, and can be used by the model receiving device
  • the model makes it possible for the AI model transfer and intercommunication of different platform frameworks, and provides a solution for the AI model transfer of devices from different manufacturers.
  • the first request message sent by the model receiver device to the model sender device further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the information carried in the first request message sent by the model receiver device to the model sender device is different, Including the following situations:
  • model conversion platform behind the model receiver device means that the model conversion platform has the attribute information of the model receiver device, including the address information of the model receiver device, identification information, and/or, supported or selected AI Framework information; and, the model receiving end device has attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is behind the model receiving end device, which can save hard disk space for the model receiving end device.
  • the first request message sent by the model receiving device to the model sending device further includes at least one of the following:
  • the attribute information of the model conversion platform includes address information and identification information of the model conversion platform, which are used to inform the model sending end device where to send the AI model.
  • the address information or identification information of the model transformation platform must be carried.
  • the attribute information of the model conversion platform may also include conversion capability information, which is used to inform the model sending device of the AI framework types supported by the model conversion platform, and the ability to convert these AI framework types into the AI framework types supported by the model receiving device. ability.
  • model conversion platform should be able to support all mainstream frameworks, so the conversion capability information is optional.
  • the attribute information of the model receiver device includes address information, identification information, and/or supported or selected AI framework information of the model receiver device.
  • the address information or identification information of the model receiver device is used to inform the AI model to be fed back to the model receiver device, which can be sent implicitly and is optional.
  • the AI framework information supported or selected by the model receiver device is used to inform the model sender device, and then inform the model conversion platform of the framework information of the AI model required by the model receiver device through the model sender device.
  • the AI framework information supported or selected by the model receiver device may be known to the model conversion platform, that is, the AI framework information supported or selected by the model receiver device is Optional to carry.
  • Model conversion requirement information where the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model.
  • the model conversion requirement information is used to instruct the model conversion platform to perform model conversion requirement information, including at least one of the following: model conversion delay requirement information; model conversion accuracy requirement information.
  • the model conversion platform is behind the model sending end device or the model conversion platform is a public conversion platform
  • model conversion platform behind the model sending device means that the model conversion platform has the attribute information of the model sending device, including the address information of the model sending device, identification information, and/or supported or selected AI framework information; and, the model sending end device has attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is behind the model sending device, which can save hard disk space for the model sending device.
  • the model transformation platform is a public transformation platform, which means that the model transformation platform can obtain the attribute information of the model receiving end device and/or the attribute information of the model sending end device.
  • the model receiver device and the model sender device can also obtain attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is a public conversion platform, which can save hard disk space for the model sending end device and the model receiving end device.
  • the first request message sent by the model receiver device to the model sender device may not carry attribute information of the model transformation platform.
  • the first model sent by the model receiving device to the model sending device The request message also includes at least one of the following:
  • the attribute information of the model receiver device includes address information, identification information, and/or supported or selected AI framework information of the model receiver device.
  • the address information or identification information of the model receiver device is used to inform the AI model to be fed back to the model receiver device, which can be sent implicitly and is optional.
  • the AI framework information supported or selected by the model receiver device is used to inform the model sender device, and then inform the model conversion platform of the framework information of the AI model required by the model receiver device through the model sender device.
  • the AI framework information supported or selected by the model receiving end device is unknown to the model transformation platform, then the AI framework information supported or selected by the model receiving end device is necessary carried.
  • the model receiver device can carry the AI framework information it supports or selects in the first request message, inform the model sender device, and then notify the model via the model sender device Conversion platform, to indicate the framework of the AI model required by the model conversion platform model receiver device.
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the model conversion requirement information is used to instruct the model conversion platform to perform model conversion requirement information, including at least one of the following: model conversion delay requirement information; model conversion accuracy requirement information.
  • the first request message sent by the model receiving device to the model sending device further includes:
  • the address and/or identification information of the model receiver device is the address and/or identification information of the model receiver device.
  • the first request message sent by the model receiving device to the model sending device carries the AI model description information and the address and/or Identification information.
  • the AI framework information supported or selected by the model receiver device can be notified to the model conversion platform through the following two implementation methods to indicate the AI model required by the model receiver device on the model conversion platform frame:
  • the method further includes:
  • the model receiver device registers and reports attribute information of the model receiver device to the model conversion platform, and the attribute information includes AI framework information supported by the model receiver device.
  • model receiving end device sends the AI framework information supported or selected by the model receiving end device to the model conversion platform through registration and reporting.
  • the method also includes:
  • the model receiver device In the case of receiving the first AI framework request message sent by the model conversion platform, the model receiver device sends the AI framework information supported by the model receiver device to the model conversion platform, wherein the first An AI framework request message is used to request the AI framework information supported by the model receiver device.
  • the model conversion platform may send a first AI framework request message to the model receiving device to request information about the AI framework supported or selected by the model receiving device.
  • the model receiver device may directly send a first request message to the model conversion platform, and through the model conversion platform according to the first A request message, sending a second request message to the model sending device to request the model sending device to provide the AI model requested by the model receiving device.
  • the first request message further includes at least one of the following:
  • the attribute information of the model receiver device includes address information, identification information, and/or supported or selected AI framework information of the model receiver device.
  • the address information or identification information of the model receiver device is used to inform the model conversion platform to feed back the AI model to the model receiver device, which can be sent implicitly and is optional.
  • the AI framework information supported or selected by the model receiver device is used to indicate to the model conversion platform model the framework information of the AI model required by the receiver device.
  • the AI framework information supported or selected by the model receiver device may be known to the model conversion platform, that is, the AI framework information supported or selected by the model receiver device is Optional to carry.
  • the attribute information of the model sender device includes address information, identification information, and/or supported or selected AI framework information of the model sender device.
  • the address information or identification information of the model sender device is used to instruct the model conversion platform to send a second request message to the specified model sender device to request the model sender device to provide the AI model requested by the model receiver device.
  • the address information or identification information of the model sender device is optional.
  • the model transformation platform will select a model sender device that can provide a model conforming to the AI model description information, and send a second request message to the selected model sender device, With the selected request model, the sender device provides a model conforming to the description information of the AI model.
  • the AI framework information supported or selected by the model sending device is used to indicate to the model conversion platform the AI framework information of the model sent by the model sending device, so that the model conversion platform can judge whether the model conversion can be completed based on this information.
  • Model conversion requirement information where the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model.
  • delay requirements transformation is completed and sent within 5s
  • accuracy requirements conversion accuracy rate is over 99%
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes at least one of the following:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the content of the first response message may be that the request is accepted, that is, the model receiving end device receives the first AI model
  • the content of the first response message may also be that the request is rejected, that is, the model receiving end device receives the first AI model.
  • the device cannot obtain the first AI model, and the model receiving end device will receive the first indication information or the second indication information.
  • the specific content of the first indication information or the second indication information received by the model receiver device will be introduced below.
  • the model receiving end device obtains the second indication information from the model transformation platform.
  • the model conversion platform cannot complete the conversion operation of the second AI model, including:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model conversion delay requirement requested by the model receiving end device; or,
  • the conversion accuracy rate of the second AI model by the model conversion platform cannot meet the model conversion accuracy requirement requested by the model receiving end device.
  • the model receiver device acquires the second indication information from the model sender device;
  • the first condition includes at least one of the following:
  • the model sending end device determines that the model conversion platform cannot complete the conversion operation on the second AI model according to the conversion capability information of the model conversion platform;
  • the model sending end device receives the first model conversion response message sent by the model conversion platform, and the first model conversion response message indicates that the model conversion platform cannot complete the conversion operation on the second AI model.
  • the model receiving device acquires the first indication information from the model sending device or a model transformation platform.
  • the model receiver device sends a first request message and receives a first response message
  • the first response message includes the first AI model, first indication information or second indication information
  • the first AI model is from
  • the second AI model of the model sending device is converted by the model conversion platform and can be used by the model receiving device.
  • the model conversion platform can be deployed behind the model receiving device or at the model sending end
  • the back of the device, or the model conversion platform is a public conversion platform, which reduces the subsequent update cost, increases the reuse rate of the conversion platform, and increases the types of frameworks that support conversion, making AI model transfer and interoperability between different platform frameworks a It is possible, and it also provides a solution for AI model transfer of devices from different manufacturers.
  • FIG. 4 is the second schematic flow diagram of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 4, the AI model transmission method is applied to the model sending end device, and the method includes the following steps:
  • Step 400 the model sending end device receives an AI model request message, and the AI model request message is used to request the first AI model;
  • AI model request message can come from the model receiving end device or from the model conversion platform.
  • the AI model request message includes AI model description information, where the AI model description information is used to indicate the information of the requested AI model, that is, the AI model description information is used to describe the first AI model.
  • the first AI model corresponds to the first model representation method.
  • the AI model description information includes at least one of the following:
  • AI model training accuracy requires information.
  • this embodiment is applied to the device side of the model sending end, and describes the actions of the opposite end in the method embodiment of the device side of the receiving end of the aforementioned model.
  • Step 401 the model sending end device performs the first operation according to the AI model request message
  • the model sending end device After receiving the AI model request message, the model sending end device performs the first operation according to the AI model request message
  • said performing the first operation includes at least one of the following:
  • the model sending end device selects or trains to generate an AI model that matches the information.
  • the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • a training object such as UE
  • the analytics ID selecting the input data type according to the analytics ID
  • the model sending end device can select or train to generate the second AI model
  • the model sending end device will send the second AI model to the model conversion platform, wherein the second AI model is converted by the model conversion platform to obtain the first AI model model, the first AI model can be used by the model receiver device.
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate the model conversion the platform is unable to convert the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the model sending end device receives the AI model request message, performs the first operation according to the AI model request message, including selecting or training to generate the second AI model, and sends the second AI model to the model conversion platform,
  • sending the first instruction information or the second instruction information makes it possible to transfer and intercommunicate AI models of different platform frameworks, and provides a solution for transferring AI models of devices of different manufacturers.
  • the AI model request message may be a second request message or a first request message.
  • the model sending end device receives the AI model request message, including:
  • the model sending end device receives the second request message sent by the model transformation platform, wherein the second request information is the first request message sent by the model transformation platform to the model transformation platform according to the model receiving end device. request message confirmation;
  • the model sending end device receives the first request message sent by the model receiving end device.
  • the sending the second AI model to the model conversion platform includes:
  • the model sending end device sends a second response message to the model conversion platform, where the second response message includes the second AI model.
  • the model sending device when the model sending device receives the second request message sent by the model transformation platform, if the model sending device can select or train to generate the second AI model, the model sending device will send The model transformation platform sends a second response message, the second response message carries the second AI model, where the second AI model includes the complete network structure and parameter information of the AI model selected or generated by the model sending device.
  • the second response message further includes at least one of the following:
  • the attribute information of the model receiver device is the attribute information of the model receiver device.
  • the framework information of the second AI model is used to inform the model conversion platform of the framework used by the second AI model, so as to facilitate the conversion by the model conversion platform.
  • the AI model file is a file generated by TesnorFLow
  • the information can be "TensorFLow” + "2.5.3” (that is, the framework platform + version number).
  • the framework information of the second AI model does not need to be carried. That is, the framework information of the second AI model is optional.
  • the attribute information of the model receiver device includes: identification information, address information and/or supported or selected AI framework information of the model receiver device.
  • the identification information or address information of the model receiving end device is used to inform the model conversion platform where to send the converted model, such as the IP address or FQDN of the model receiving end device and other identification information.
  • the AI framework information supported or selected by the model receiver device is used to indicate which AI framework model the model conversion platform should convert the second AI model sent by the model sender device into.
  • the sending the second AI model to the model conversion platform includes:
  • the model sending end device sends a third request message to the model conversion platform, the third request message is used to request the model conversion platform to convert the second AI model into the first AI model; the third request message The second AI model is included.
  • the model sending device selects or trains to generate a second AI model, and sends a third request message to the model conversion platform to request the model conversion platform to convert the second AI model to the model conversion platform.
  • the AI model is transformed into a first AI model.
  • the second request message further includes at least one of the following:
  • the attribute information of the model transformation platform is the attribute information of the model transformation platform.
  • the model receiving end device sends the first request message to the model conversion platform
  • the first request message includes: AI model description information, attribute information of the model receiving end device, the model sending at least one of the attribute information of the terminal device and the model transformation requirement information.
  • the model conversion platform generates a second request message according to the first request message, and sends the second request message to the model sending device.
  • the second request message includes AI model description information and attribute information of the model conversion platform. This step is applicable to The model conversion platform is behind the model receiver device.
  • the attribute information of the model conversion platform includes address information and identification information of the model conversion platform, which are used to inform the model sending device where to send the AI model, b) such as the IP address and FQDN of the conversion platform. Since the transformation platform sends information to the sender first, the address information or identification information of the model transformation platform is optional.
  • the attribute information of the model conversion platform may also include conversion capability information, which is used to inform the model sending device of the AI framework types supported by the model conversion platform, and the ability to convert these AI framework types into the AI framework types supported by the model receiving device. ability.
  • conversion capability information is used to inform the model sending device of the AI framework types supported by the model conversion platform, and the ability to convert these AI framework types into the AI framework types supported by the model receiving device. ability.
  • the conversion platform should be able to support all mainstream frameworks, so the conversion capability information is optional.
  • the information carried by the first request message received by the model receiver device from the model receiver device is different , including the following situations:
  • model conversion platform behind the model receiver device means that the model conversion platform has the attribute information of the model receiver device, including the address information of the model receiver device, identification information, and/or, supported or selected AI Framework information; and, the model receiving end device has attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is behind the model receiving end device, which can save hard disk space for the model receiving end device.
  • the first request message also include at least one of the following:
  • the attribute information of the model conversion platform includes address information and identification information of the model conversion platform, which are used to inform the model sending end device where to send the AI model.
  • the address information or identification information of the model transformation platform must be carried.
  • the attribute information of the model conversion platform may also include conversion capability information, which is used to inform the model sending device of the AI framework types supported by the model conversion platform, and the ability to convert these AI framework types into the AI framework types supported by the model receiving device. ability.
  • model conversion platform should be able to support all mainstream frameworks, so the conversion capability information is optional.
  • the attribute information of the model receiver device includes address information, identification information, and/or supported or selected AI framework information of the model receiver device.
  • the address information or identification information of the model receiver device is used to inform the AI model to be fed back to the model receiver device, which can be sent implicitly and is optional.
  • the AI framework information supported or selected by the model receiver device is used to inform the model sender device, and then inform the model conversion platform of the framework information of the AI model required by the model receiver device through the model sender device.
  • the AI framework information supported or selected by the model receiver device may be known to the model conversion platform, that is, the AI framework information supported or selected by the model receiver device is Optional to carry.
  • Model conversion requirement information where the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model.
  • the model conversion requirement information is used to instruct the model conversion platform to perform model conversion requirement information, including at least one of the following: model conversion delay requirement information; model conversion accuracy requirement information.
  • the model conversion platform is behind the model sending device or the model conversion platform is a public conversion platform
  • model conversion platform behind the model sending device means that the model conversion platform has the attribute information of the model sending device, including the address information of the model sending device, identification information, and/or supported or selected AI framework information; and, the model sending end device has attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is behind the model sending device, which can save hard disk space for the model sending device.
  • the model transformation platform is a public transformation platform, which means that the model transformation platform can obtain the attribute information of the model receiving end device and/or the attribute information of the model sending end device.
  • the model receiver device and the model sender device can also obtain attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • the model conversion platform is a public conversion platform, which can save hard disk space for the model sending end device and the model receiving end device.
  • the first request message received by the model sending device and sent by the model receiving device may not carry attribute information of the model transformation platform.
  • the model sending end device receives the first request message sent by the model receiving end device, and the model transformation platform is behind the model sending end device or the model transformation platform is a public transformation platform
  • the first request message further includes at least one of the following:
  • the attribute information of the model receiver device includes address information, identification information, and/or supported or selected AI framework information of the model receiver device.
  • the address information or identification information of the model receiver device is used to inform the AI model to be fed back to the model receiver device, which can be sent implicitly and is optional.
  • the AI framework information supported or selected by the model receiver device is used to inform the model sender device, and then inform the model conversion platform of the framework information of the AI model required by the model receiver device through the model sender device.
  • the AI framework information supported or selected by the model receiving end device is unknown to the model transformation platform, then the AI framework information supported or selected by the model receiving end device is necessary carried.
  • the model receiver device can carry the AI framework information it supports or selects in the first request message, inform the model sender device, and then notify the model via the model sender device Conversion platform, to indicate the framework of the AI model required by the model conversion platform model receiver device.
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the model conversion requirement information is used to instruct the model conversion platform to perform model conversion requirement information, including at least one of the following: model conversion delay requirement information; model conversion accuracy requirement information.
  • the first request message further includes:
  • the address and/or identification information of the model receiver device is the address and/or identification information of the model receiver device.
  • the model sending device receives the first request message sent by the model receiving device, which carries the AI model description information and the address and address of the model receiving device. /or identifying information.
  • the AI framework information supported or selected by the model sending device can be notified to the model conversion platform through the following two implementation methods to indicate the AI model provided by the model sending device on the model conversion platform frame:
  • the method further includes:
  • the model sender device registers with the model conversion platform and reports attribute information of the model sender device, where the attribute information includes AI framework information supported by the model sender device.
  • model sending end device sends the AI framework information supported or selected by the model sending end device to the model transformation platform through registration and reporting.
  • the method also includes:
  • the model sending end device sends the model transformation platform the AI framework information supported by the model sending end device, wherein the second The AI framework request message is used to request the AI framework information supported by the model sending end device.
  • the model conversion platform may send a second AI framework request message to the model sending device to request the model sending device to feed back the supported or selected AI framework information.
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the model sending end device may send the first indication information or the second indication information to the model receiving end device.
  • the following describes the specific content of sending the first indication information or the second indication information by the model sending end device.
  • the sending the first indication information includes:
  • model sending end device does not have or cannot generate the second AI model, send the first indication information to the model receiving end device.
  • the model sending end device sends the first indication information to the model receiving end device.
  • the method also includes:
  • the model sending end device does not have or cannot generate the second AI model, it sends a fourth response message to the model conversion platform, and the fourth response message is used to indicate that the model sending end device cannot provide The second AI model.
  • model sending device receives the second request message sent by the model conversion platform, if the sending device does not have or cannot generate the AI model specified or supported by the model conversion platform, then the model conversion platform cannot perform conversion, then the model sending end device will respond to the conversion platform with "request not accepted", that is, the model sending end device sends a fourth response message to the model conversion platform, and the fourth response message is used to indicate that the model The sending end device cannot provide the second AI model.
  • the method also includes:
  • the model sending end device can also determine whether the model conversion platform can complete the conversion operation of the second AI model according to the conversion capability information of the model conversion platform, and if it is determined that the model conversion platform cannot complete the conversion operation of the second AI model The conversion operation, then send the second indication information to the model receiver device.
  • the method also includes:
  • the model sending device selects or trains to generate the second AI model
  • the second AI model is sent to the model conversion platform for conversion
  • the model conversion platform can send the model to the model sending device after the model conversion is successful.
  • a first model conversion response message is sent to indicate whether the conversion operation on the second AI model is completed.
  • the model conversion platform can also determine whether the conversion operation of the second AI model can be completed according to the conversion capability information before the model conversion, and send the first model conversion response message to the model sending device, indicating whether the conversion of the second AI model can be completed.
  • the sending the second indication information includes:
  • the first condition includes at least one of the following:
  • the model conversion platform According to the conversion capability information of the model conversion platform, it is determined that the model conversion platform cannot complete the conversion operation on the second AI model;
  • a first model conversion response message sent by the model conversion platform is received, and the first model conversion response message indicates that the model conversion platform cannot complete the conversion operation on the second AI model.
  • model conversion platform cannot complete the conversion operation of the second AI model, including:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model conversion delay requirement requested by the model receiving end device; or,
  • the conversion accuracy rate of the second AI model by the model conversion platform cannot meet the model conversion accuracy requirement requested by the model receiving end device.
  • the model sending end device receives the AI model request message, performs the first operation according to the AI model request message, including selecting or training to generate the second AI model, and sends the second AI model to the model conversion platform, Alternatively, the first instruction information or the second instruction information is sent, and the second AI model is converted by the model conversion platform to obtain a model that can be used by the model receiving end device.
  • the model conversion platform can be deployed behind the model receiving end device, or It can be deployed behind the model sending end device, or the model conversion platform is a public conversion platform, which reduces the subsequent update cost, increases the reuse rate of the conversion platform, and increases the types of frameworks that support conversion, making different platform frameworks It is possible to transfer and interoperate with other AI models, and it also provides a solution for the transfer of AI models of devices from different manufacturers.
  • FIG. 5 is the third schematic flowchart of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 5, the method is applied to the model transformation platform, and the method includes the following steps:
  • Step 500 the model conversion platform receives the second AI model sent by the model sending device
  • Step 501 the model conversion platform converts the second AI model into a first AI model, and sends the first AI model to a model receiving device;
  • the model conversion platform sends first indication information or second indication information to the model receiving end device, the first indication information is used to indicate that the model sending end device cannot provide the second AI model, and the second indication The information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model;
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • this embodiment is applied to the model transformation platform side, and describes the peer-to-peer actions of the aforementioned method embodiments on the model receiver device side and the model sender device side.
  • the method embodiment of the device side of the model receiving end and the method embodiment of the device side of the model sending end refer to the method embodiment of the device side of the model receiving end and the method embodiment of the device side of the model sending end, and will not be repeated here.
  • the model conversion platform will convert the second AI model into the first AI model that can be used by the model receiving end device, and convert the first AI model to the model conversion platform.
  • the AI model is sent to the model receiving device.
  • the model conversion platform If the model conversion platform cannot convert the second AI model into the first AI model, the model conversion platform sends the second instruction information to the model receiving end device.
  • the model conversion platform If the model conversion platform knows that the model sending device cannot provide the second AI model, the model conversion platform sends the first indication information to the model receiving device.
  • the model conversion platform receives the second AI model sent by the model sending end device, including:
  • the model conversion platform receives a second response message sent by the model sending device, where the second response message includes the second AI model.
  • the second response message further includes at least one of the following:
  • the attribute information of the model receiver device is the attribute information of the model receiver device.
  • the AI framework information supported by the model sending device is also the framework information of the second AI model, which is used to inform the model conversion platform of the framework used by the second AI model, so as to facilitate the model conversion platform to perform conversion.
  • the AI model file is a file generated by TesnorFLow
  • the information can be "TensorFLow” + "2.5.3” (that is, the framework platform + version number).
  • the framework information of the second AI model does not need to be carried. That is, the framework information of the second AI model is optional.
  • the attribute information of the model receiver device includes: identification information, address information and/or supported or selected AI framework information of the model receiver device.
  • the identification information or address information of the model receiving end device is used to inform the model conversion platform where to send the converted model, such as the IP address or FQDN of the model receiving end device and other identification information.
  • the AI framework information supported or selected by the model receiver device is used to indicate which AI framework model the model conversion platform should convert the second AI model sent by the model sender device into.
  • the model conversion platform receives the second AI model sent by the model sending end device, including:
  • the model conversion platform receives a third request message sent by the model sending device, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model.
  • the model conversion platform receives the second AI model sent by the model sending end device; converts the second AI model into the first AI model, and sends the first AI model to the model receiving end device , or send the first instruction information or the second instruction information to the model receiver device, which makes it possible to transfer and interoperate AI models of different platform frameworks, and also provides a solution for the transfer of AI models of devices from different manufacturers.
  • the method also includes at least one of the following:
  • the model conversion platform receives the attribute information of the model sender device reported by the model sender device, and the attribute information of the model sender device includes AI framework information supported by the model sender device;
  • the model transformation platform receives the attribute information of the model receiving end device reported by the model receiving end device, and the attribute information of the model receiving end device includes AI framework information supported by the model receiving end device.
  • the model transformation platform is a public transformation platform.
  • the model transformation platform is a public transformation platform, which means that the model transformation platform can obtain the attribute information of the model receiving end device and/or the attribute information of the model sending end device.
  • the model receiver device and the model sender device can also obtain attribute information of the model conversion platform, including address information, identification information, and/or conversion capability information of the model conversion platform.
  • model receiving end device sends the AI framework information supported or selected by the model receiving end device to the model conversion platform through registration and reporting.
  • the model sending end device sends the AI framework information supported or selected by the model receiving end device to the model conversion platform by registering and reporting.
  • the method also includes at least one of the following:
  • the model conversion platform may send a first AI framework request message to the model receiving device to request information about the AI framework supported or selected by the model receiving device.
  • the model transformation platform may send a second AI framework request message to the model sending device to request information about the AI framework supported or selected by the model sending device.
  • the method also includes:
  • the model conversion platform receives the first request message sent by the model receiver device
  • the model conversion platform sends a second request message to the model sending device
  • the second request information is determined according to the first request message.
  • the model receiver device may directly send a first request message to the model conversion platform, and the model conversion platform receives the model receiver After sending the first request message, the device sends a second request message to the model sending device to request the model sending device to provide the AI model requested by the model receiving device.
  • the first request message or the second request message includes: AI model description information, where the model description information includes at least one of the following:
  • AI model training accuracy requires information.
  • the first request message further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model
  • the second request message also includes at least one of the following:
  • the attribute information of the model conversion platform is the attribute information of the model conversion platform.
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes at least one of the following:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the method before the model conversion platform sends the second request message to the model sending device, the method further includes:
  • the model transformation platform determines that the model transformation platform has the ability to transform the AI framework supported by the model sending end device into the AI framework supported by the model receiving end device.
  • model conversion platform receives the first request message sent by the model receiver device, according to its own conversion capability information, the AI framework information supported by the model sender device, and the AI framework information supported by the model receiver device, Determine whether it has the ability to transform from the AI framework supported by the model sending device to the AI framework supported by the model receiving device.
  • the model conversion platform determines that it has the ability to convert from the AI framework supported by the model sending device to the AI framework supported by the model receiving device, the model conversion platform sends a second request message to the model sending device.
  • the method further includes:
  • the model conversion platform determines that the model conversion platform does not have the ability to convert the AI framework supported by the model sending end device into the AI framework supported by the model receiving end device, and the model conversion platform sends a third A response message, the third response message is used to indicate that the first request message is rejected.
  • the method further includes:
  • the model conversion platform determines the model sending end device.
  • the model conversion platform After the model conversion platform receives the first request message sent by the model receiver device, if the first request message does not carry the attribute information of the model sender device, the model conversion platform will, according to the transmission speed, The server location, line occupancy rate and other factors automatically select the model sending end device that can provide the AI model conforming to the AI model description information in the first request message. Alternatively, the model transformation platform may automatically select a model sending device that is more in line with the expectations of the model receiving device based on previous interaction transfer records.
  • the method further includes:
  • the model conversion platform sends a first model conversion response message to the model sending device, and the first model conversion response message is used to indicate whether the model conversion platform has completed the conversion operation on the second AI model.
  • the model conversion platform sends second indication information to the model receiver device, including;
  • the model conversion platform In a case where the model conversion platform cannot complete the conversion operation on the second AI model, the model conversion platform sends the second instruction information to the model receiving end device.
  • the model conversion platform cannot complete the conversion operation of the second AI model, including at least one of the following:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model provision delay requirement requested by the model receiving end device; or,
  • the conversion accuracy of the second AI model by the model conversion platform cannot meet the model accuracy requirement requested by the model receiving end device.
  • the method before the model conversion platform sends the first indication information to the model receiver device, the method further includes:
  • the model sending end device After receiving the fourth response message sent by the model sending device, the model transformation platform sends the first indication information to the model receiving device.
  • model transformation platform may be deployed behind the model receiving end device, and the model transformation platform may also be deployed behind the model sending end device, or the model transformation platform may be a public transformation platform.
  • the hardware cost of the model receiving end device can be saved. If the model conversion platform is deployed behind the model sending device, the hardware cost of the model sending device can be saved. If the model conversion platform is a public conversion platform, it can save hard disk space for the model sending device and the model receiving device.
  • the model conversion platform receives the second AI model sent by the model sending end device; converts the second AI model into the first AI model, and sends the first AI model to the model receiving end device , or send the first indication information or the second indication information to the model receiving end device, wherein the model conversion platform can be deployed behind the model receiving end device or behind the model sending end device, or the model conversion platform can be public
  • the conversion platform reduces the subsequent update cost, increases the reuse rate of the conversion platform, and increases the types of frameworks that support conversion, making it possible to transfer and interoperate AI models of different platform frameworks, and also provides support for equipment from different manufacturers.
  • AI model delivery provides a solution.
  • the conversion platform in the following embodiments shown in Figures 6 to 12 is the model conversion platform
  • the sending end is the model sending end device
  • the receiving end is the model receiving end device.
  • FIG. 6 is one of the interactive flow diagrams of the AI model transmission method provided by the embodiment of the present application.
  • step 1
  • the model receiver device sends an artificial intelligence model (AI model) request message to the model sender device, including at least one of the following information:
  • AI model artificial intelligence model
  • Model description information information used to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., the algorithm used by the model (such as neural network, random forest, etc.),
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • the time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as 90% model accuracy, etc.
  • Model receiving end device information including address information or identification information of the model receiving end device, which can be sent implicitly to feed back the model to the model receiving end device. (optional)
  • Transformation platform address or identification information Inform the model sending device where to send the model, such as the IP address and FQDN of the transformation platform. (required)
  • Conversion platform capability information inform the model sending device that the conversion platform supports AI framework types, and has the ability to convert these AI framework types into the AI framework types supported by the model receiving device. (The conversion platform should be able to support all mainstream frameworks, so it is "optional")
  • AI framework information of the model receiver device inform the model sender device, and then tell the conversion platform via the model sender device, what kind of framework the model receiver device itself needs. (Because the conversion platform belongs to the model receiver device, the information conversion platform should have it, so it is "optional")
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the model sending device sends the requested model information to the transformation platform.
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the model receiving device. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file which contains the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • AI model framework information inform the conversion platform of the framework used by the AI model file, so as to facilitate the conversion platform. (Can be “optional” if the conversion platform can judge for itself). If the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, the framework platform + version number).
  • Model receiver device information tell the conversion platform where to send the converted model, such as model receiver device IP address or FQDN and other identification information (optional);
  • Framework information of the device at the receiving end of the model the model that informs the conversion platform of which framework the model should be converted into. (optional).
  • the model file generated by TensorFLow needs to be converted into a model file recognizable by PyTorch, it can be in the form of "PyTorch” + "1.10.0" (framework platform + version number).
  • the conversion platform sends a response message to the model sending device:
  • the conversion platform can convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end device that the model conversion is accepted or successful;
  • the conversion platform cannot convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end that the device model conversion fails. It may further include a failure reason value, for example, because the AI model cannot be converted from the framework A model to the framework B model.
  • the model sender device sends request response information to the model receiver device (optional):
  • step 3.a Request acceptance; if "transformation is accepted or successful" is received (step 3.a), inform the receiving end that its request is reasonable and feasible and has been accepted.
  • step 3.b Request Denial. For example, when receiving "conversion failure" (step 3.b), inform the model sending end device that its request is rejected. The reason for the failure can also be further informed based on the feedback in "step 3.b", for example, the transformation platform does not support the transformation from AI framework A to AI framework B at the moment. Or, the request of the model receiving device is rejected due to the reason of the model sending device, for example, the model sending device cannot generate the required AI model.
  • step 4 and step 5 below are not limited. (It can be sent twice, there may be abnormal conditions, and the forwarding is unsuccessful)
  • the conversion platform converts the AI model into a model that can be used by the AI framework of the model receiver device.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • the conversion platform (“tf2onnx.conver") can convert TensorFlow model files into ONNX files with the suffix ".onnx".
  • the calculation graph Graph contains information such as network structure and parameters, and consists of some basic information such as names and four sets of lists.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TetensorProto type
  • node stores all the computing nodes in the model
  • input stores the name of the model input node.
  • output stores the model output node name
  • initializer stores the specific values of all network parameters of the model, including hyperparameters and input values.
  • each Node contains information such as its operation type and specified inputs and outputs names.
  • the node information of a Node can be the calculation process of a whole layer, or the calculation of a node. All nodes are connected together to form a graph, and the two arrays of inputs and outputs in the Node calculation node point to the relationship between the input and output nodes to construct the topology of the entire network, that is, the network structure.
  • the conversion platform sends the converted AI model to the model receiving device.
  • the conversion platform sends the converted AI model file to the model receiving device, and the AI model file is supported by the model receiving device.
  • the model receiving device can understand and use the AI model file .
  • AI framework information of the model sender device may also be included (for example, the model sender device supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • FIG. 7 is the second schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • the AI model transmission method includes the following steps:
  • step 1
  • the model receiver device sends an AI model request message to the transformation platform, including at least one of the following information:
  • Model description information It can indicate the information specified to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., and the algorithm used by the model (such as neural network, random forest etc.), the training object information corresponding to the model (such as a certain user, a certain area AOI), the time information corresponding to the model (model for a specific time period and time point), the accuracy requirements of model training (such as the model accuracy of 90% etc.) wait for at least one item. (required)
  • model type or ID information such as analytic ID, model ID), model name, etc.
  • the algorithm used by the model such as neural network, random forest etc.
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as the model accuracy of 90% etc.
  • Model receiving end device information It is convenient for the transformation platform to send results to include address information or identification information of the model receiving end device, which can be sent implicitly to feed the model back to the model receiving end device. (optional)
  • AI framework information of the model receiver device tell the conversion platform which framework model to convert the model into. If you need to convert the model file generated by TensorFLow into a model file recognizable by PyTorch, it can be in the form of "PyTorch” + "1.10.0" (framework platform + version number). (Optional, because the conversion platform belongs to the model receiver device, it should know the framework information of the model receiver device.)
  • Model sender device information Let the conversion platform send a model request to the specified model sender device, including the address information or identification information of the model sender device; send information" model sender device and request the model. (optional)
  • AI framework information of the model sender device inform the conversion platform of the framework platform of the model sent by the model sender device, and let the conversion platform judge whether the conversion can be completed. If the conversion cannot be completed, step 1a will be triggered. (optional)
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • Step 1a (choose one from Step 5):
  • the conversion platform judges whether the conversion can be performed and responds.
  • the conversion platform cannot convert from the "model sending device AI framework" specified by the model receiving device to the model receiving device AI platform framework. If the model conversion request information in 1e is requested, "request denied" will be returned.
  • Step 1b (optional):
  • the transformation platform judges whether to execute this step, and automatically selects the model sending end device
  • step 1 If the "model sender device information" has been specified in the request message in step 1, this step will not be performed, and the conversion platform will directly request the model from the specified model sender device.
  • the conversion platform will automatically select the "model description information" that meets the model receiver device in step 1 according to factors such as transmission speed, server location, and line occupancy The model sender device of the AI model. (Or, based on previous interaction transfer records, the model sender device that is more in line with the expectations of the model receiver device can be automatically selected)
  • the transformation platform sends a model request to the model sending device (automatically selected in 1b or specified in 1), including at least one of the following information:
  • Model description information It can indicate the information specified to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., and the algorithm used by the model (such as neural network, random forest etc.), the training object information corresponding to the model (such as a certain user, a certain area AOI), the time information corresponding to the model (model for a specific time period and time point), the accuracy requirements of model training (such as the model accuracy of 90% etc.) wait for at least one item. (required)
  • model type or ID information such as analytic ID, model ID), model name, etc.
  • the algorithm used by the model such as neural network, random forest etc.
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as the model accuracy of 90% etc.
  • Conversion platform information including the conversion platform address or identification information, telling the model sending device where to send the model, such as the conversion platform's IP address, FQDN, etc. (Since the conversion platform sends information to the model sending device first, it is optional)
  • Conversion platform capability information Inform the model sending device of the AI framework types supported by the conversion platform, and have the ability to convert these AI framework types into the AI framework types supported by the model receiving device. (The conversion platform should be able to support all mainstream frameworks, so it is "optional")
  • Step 3 (choose one from Step 3a):
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the conversion platform. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file a file containing elements such as the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt” suffix, which contains parameter information. (required)
  • the framework used by the model sending device inform the conversion platform of the framework used by the AI model file, so that the conversion platform can facilitate the conversion.
  • the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, in the form of framework platform + version number). (If the conversion platform can judge by itself, it can be "optional")
  • Step 3a (choose one from Step 3):
  • the model sender device responds to the transformation platform
  • model sending device does not have or cannot generate the AI model specified or supported by the conversion platform, then the conversion platform cannot perform the conversion, and the model sending device will respond to the conversion platform with "request not accepted".
  • step 3 that is, an AI model is successfully sent from the model sending device to the conversion platform, then the conversion platform converts the AI model into a model available to the receiving AI framework.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • model A of the TensorFlow framework to model B of the ONNX framework.
  • Python script provided by ONNX, "tf2onnx.conver” can convert TensorFlow model files into ONNX files with the suffix ".onnx”.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TinsorProto type
  • input nodes in the model are stored in input
  • output nodes in the model are stored in output
  • all weight parameters in the model are stored in initializer.
  • Step 5 (choose one from Step 1a):
  • the conversion platform sends the converted AI model to the model receiver device. Specifically, the conversion platform sends the converted AI model file to the model receiver device.
  • the AI model file is supported by the model receiver device. In other words, the model receiver device can understand and use the AI model file.
  • it may also include AI framework information of the model sender device (such as the IP address of the model sender device, which supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • AI framework information of the model sender device such as the IP address of the model sender device, which supports the TensorFlow framework.
  • FIG. 8 is the third schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 8, the AI model transmission method includes the following steps:
  • step 1
  • the model receiver device sends an artificial intelligence model (AI model) request message to the model sender device, including at least one of the following information:
  • AI model artificial intelligence model
  • Model description information information used to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., the algorithm used by the model (such as neural network, random forest, etc.),
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • the time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as 90% model accuracy, etc.
  • Model receiving end device information including address information or identification information of the model receiving end device, which can be sent implicitly to feed back the model to the model receiving end device. (optional)
  • AI framework information of the model receiver device inform the model sender device, and then tell the transformation platform via the model sender device, what kind of framework the model receiver device itself needs. (Because the conversion platform belongs to the model sender device, there is no such information conversion platform, so it must be selected)
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the model sending device sends the requested model information to the conversion platform.
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the model receiving device. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file which contains the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • AI model framework information inform the conversion platform of the framework used by the AI model file, so as to facilitate the conversion platform. (Can be “optional” if the conversion platform can judge for itself). If the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, in the form of framework platform + version number).
  • Model receiver device information tell the conversion platform where to send the converted model, such as model receiver device IP address or FQDN and other identification information (optional);
  • Framework information of the device at the receiving end of the model the model that informs the conversion platform of which framework the model should be converted into. (optional).
  • the model file generated by TensorFLow needs to be converted into a model file recognizable by PyTorch, it can be in the form of "PyTorch” + "1.10.0" (framework platform + version number).
  • the conversion platform sends a response message to the model sending device:
  • the conversion platform can convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end device that the model conversion is accepted or successful;
  • the conversion platform cannot convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end that the device model conversion fails. It can further include a value for the reason of failure, for example, because the AI model is not supported to be converted from the framework A model to the framework B model.
  • the model sender device sends request response information to the model receiver device (optional):
  • step 3.a Request acceptance; if "transformation is accepted or successful" is received (step 3.a), inform the receiving end that its request is reasonable and feasible and has been accepted.
  • step 3.b Request Denial. For example, when receiving "conversion failure" (step 3.b), inform the model sending end device that its request is rejected. The reason for the failure can also be further informed based on the feedback in "step 3.b", for example, the transformation platform does not support the transformation from AI framework A to AI framework B at the moment. Or, the request of the model receiving device is rejected due to the reason of the model sending device, for example, the model sending device cannot generate the required AI model.
  • step 4 and step 5 below is not limited. (It can be sent twice, there may be abnormal conditions, and the forwarding is unsuccessful).
  • the conversion platform converts the AI model into a model that can be used by the AI framework of the model receiver device.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • the conversion platform (“tf2onnx.conver") can convert TensorFlow model files into ONNX files with the suffix ".onnx".
  • the calculation graph Graph contains information such as network structure and parameters, and consists of some basic information such as names and four sets of lists.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TetensorProto type
  • node stores all the computing nodes in the model
  • input stores the name of the model input node.
  • output stores the model output node name
  • initializer stores the specific values of all network parameters of the model, including hyperparameters and input values.
  • each Node contains information such as its operation type and specified inputs and outputs names.
  • the node information of a Node can be the calculation process of a whole layer, or the calculation of a node. All nodes are connected together to form a graph, and the two arrays of inputs and outputs in the Node calculation node point to the relationship between the input and output nodes to construct the topology of the entire network, that is, the network structure.
  • the conversion platform sends the converted AI model to the model receiving device.
  • the conversion platform sends the converted AI model file to the model receiving device, and the AI model file is supported by the model receiving device.
  • the model receiving device can understand and use the AI model file .
  • AI framework information of the model sender device may also be included (for example, the model sender device supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • FIG. 9 is the fourth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • the AI model transmission method includes the following steps:
  • step 1
  • the model receiver device sends an artificial intelligence model (AI model) request message to the model sender device, including at least one of the following information:
  • AI model artificial intelligence model
  • Model description information information used to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., the algorithm used by the model (such as neural network, random forest, etc.),
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • the time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as 90% model accuracy, etc.
  • Model receiving end device information including address information or identification information of the model receiving end device, which can be sent implicitly to feed back the model to the model receiving end device. (optional)
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the model sending device sends the requested model information to the conversion platform.
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the model receiving device. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file which contains the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • AI model framework information inform the conversion platform of the framework used by the AI model file, so as to facilitate the conversion platform. (Can be “optional” if the conversion platform can judge for itself). If the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, in the form of framework platform + version number).
  • Model receiver device information tell the conversion platform where to send the converted model, such as the IP address or FQDN of the model receiver device;
  • Model sender device information including the address information or identification information of the model sender device, which can be sent implicitly to find the corresponding framework.
  • the conversion platform obtains the supportable framework information of both parties.
  • the conversion platform converts the AI model into a model that can be used by the AI framework of the model receiver device.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • the conversion platform (“tf2onnx.conver") can convert TensorFlow model files into ONNX files with the suffix ".onnx".
  • the calculation graph Graph contains information such as network structure and parameters, and consists of some basic information such as names and four sets of lists.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TetensorProto type
  • node stores all the computing nodes in the model
  • input stores the name of the model input node.
  • output stores the model output node name
  • initializer stores the specific values of all network parameters of the model, including hyperparameters and input values.
  • each Node contains information such as its operation type and specified inputs and outputs names.
  • the node information of a Node can be the calculation process of an entire layer, or the calculation of a node. All nodes are connected together to form a graph, and the two arrays of inputs and outputs in the Node calculation node point to the relationship between the input and output nodes to construct the topology of the entire network, that is, the network structure.
  • the conversion platform sends a response message to the model sending device:
  • the conversion platform can convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end device that the model conversion is accepted or successful;
  • the conversion platform cannot convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end that the device model conversion fails. It may further include a failure reason value, for example, because the AI model cannot be converted from the framework A model to the framework B model.
  • the model sender device sends request response information to the model receiver device (optional):
  • Request acceptance for example, when receiving "transformation accepted or successful", inform the receiving end that its request is reasonable and feasible and has been accepted.
  • b) Request Denial For example, when receiving "conversion failure", inform the model sending device that its request is rejected. The reason for the failure can also be further informed based on the feedback in "step 3.b", for example, the transformation platform does not support the transformation from AI framework A to AI framework B at the moment. Or, the request of the model receiving device is rejected due to the reason of the model sending device, for example, the model sending device cannot generate the required AI model.
  • step 4 and step 5 below are not limited. (It can be sent twice, there may be abnormal conditions, and the forwarding is unsuccessful)
  • the conversion platform sends the converted AI model to the model receiving device.
  • the conversion platform sends the converted AI model file to the model receiving device, and the AI model file is supported by the model receiving device.
  • the model receiving device can understand and use the AI model file .
  • AI framework information of the model sender device may also be included (for example, the model sender device supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • Fig. 10 is a schematic diagram of registering and reporting capability information to the public transformation platform provided by the embodiment of the present application.
  • step 1 may also include:
  • Step 01 the model receiver device registers and reports capability information to the public conversion platform, including the following information:
  • Model receiving end device information including the address information or identification information of the model receiving end device, which can be sent implicitly and used to feed back the model to the model receiving end device. (optional)
  • AI framework information of the model receiver device Tell the conversion platform what kind of framework the model receiver device itself needs.
  • the model sender device registers and reports capability information to the public conversion platform, including the following information:
  • Model sender device information including address information or identification information of the model sender device, which can be sent implicitly to feed back the model to the model receiver device.
  • AI framework information of the model sender device tell the conversion platform what kind of framework the model sender device needs.
  • the public transformation platform saves the capability information of the model receiver device and the model sender device.
  • FIG. 11 is the fifth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 11, the AI model transmission method includes the following steps:
  • step 1
  • the model receiver device sends an artificial intelligence model (AI model) request message to the model sender device, including at least one of the following information:
  • AI model artificial intelligence model
  • Model description information information used to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., the algorithm used by the model (such as neural network, random forest, etc.),
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • the time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as 90% model accuracy, etc.
  • Model receiving end device information including address information or identification information of the model receiving end device, which can be sent implicitly to feed back the model to the model receiving end device. (optional)
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the model sending device sends the requested model information to the transformation platform.
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the model receiving device. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file which contains the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • Model receiver device information tell the conversion platform where to send the converted model, such as the IP address or FQDN of the model receiver device;
  • the conversion platform sends response information to the model receiver device and the model sender device to obtain the model framework information:
  • the conversion platform converts the AI model into a model that can be used by the AI framework of the model receiver device.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • the conversion platform (“tf2onnx.conver") can convert TensorFlow model files into ONNX files with the suffix ".onnx".
  • the calculation graph Graph contains information such as network structure and parameters, and consists of some basic information such as names and four sets of lists.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TetensorProto type
  • node stores all the computing nodes in the model
  • input stores the name of the model input node.
  • output stores the model output node name
  • initializer stores the specific values of all network parameters of the model, including hyperparameters and input values.
  • each Node contains information such as its operation type and specified inputs and outputs names.
  • the node information of a Node can be the calculation process of a whole layer, or the calculation of a node. All nodes are connected together to form a graph, and the two arrays of inputs and outputs in the Node calculation node point to the relationship between the input and output nodes to construct the topology of the entire network, that is, the network structure.
  • the conversion platform sends the converted AI model to the model receiving device.
  • the conversion platform sends the converted AI model file to the model receiving device, and the AI model file is supported by the model receiving device.
  • the model receiving device can understand and use the AI model file .
  • AI framework information of the model sender device may also be included (for example, the model sender device supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • FIG. 12 is the sixth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 12, the AI model transmission method includes the following steps:
  • step 1
  • the model receiver device sends an artificial intelligence model (AI model) request message to the model sender device, including at least one of the following information:
  • AI model artificial intelligence model
  • Model description information information used to indicate the requested model, such as model type or ID information (such as analytic ID, model ID), model name, etc., the algorithm used by the model (such as neural network, random forest, etc.), The training object information corresponding to the model (such as a certain user, a certain area AOI), the time information corresponding to the model (model for a specific time period and time point), the accuracy requirements of model training (such as 90% model accuracy, etc.), etc. at least one. (required)
  • model type or ID information such as analytic ID, model ID), model name, etc.
  • the algorithm used by the model such as neural network, random forest, etc.
  • the training object information corresponding to the model such as a certain user, a certain area AOI
  • the time information corresponding to the model model for a specific time period and time point
  • the accuracy requirements of model training such as 90% model accuracy, etc.
  • Model receiving end device information including address information or identification information of the model receiving end device, which can be sent implicitly to feed back the model to the model receiving end device. (optional)
  • AI framework information of the model receiver device inform the model sender device, and then tell the conversion platform via the model sender device, what kind of framework the model receiver device itself needs.
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the model sending device sends the requested model information to the transformation platform.
  • the model sending device selects or trains the AI model requested by the model receiving device according to the request message from the model receiving device. Specifically, the model sending end device selects or trains according to the AI model description information in the request message to generate an AI model that matches the information. For example, the model sending end device executes the machine learning model training process on the relevant data of a training object (such as UE) in a specific area or time period (selecting the input data type according to the analytics ID) according to the neural network algorithm, and the accuracy of the generated model meets expectations Value AI model.
  • the method of model training is not limited in the present invention.
  • the model sending end device sends the requested model information to the conversion platform, including at least one of the following:
  • AI model file which contains the complete network structure and parameter information of the generated AI model. (required).
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt" suffix, which contains parameter information.
  • AI model framework information inform the conversion platform of the framework used by the AI model file, so as to facilitate the conversion platform. (Can be “optional” if the conversion platform can judge for itself). If the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, in the form of framework platform + version number).
  • Model receiver device information tell the conversion platform where to send the converted model, such as model receiver device IP address or FQDN and other identification information (optional);
  • Framework information of the device at the receiving end of the model tell the conversion platform which framework the model should be converted into. (optional). For example, if the model file generated by TensorFLow needs to be converted into a model file recognizable by PyTorch, it can be in the form of "PyTorch” + "1.10.0" (framework platform + version number).
  • the conversion platform sends a response message to the model sending device:
  • the conversion platform can convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end device that the model conversion is accepted or successful;
  • the conversion platform cannot convert the received model information into a model corresponding to the device frame of the model receiving end, the response message is used to inform the model sending end that the device model conversion fails. It may further include a failure reason value, for example, because the AI model cannot be converted from the framework A model to the framework B model.
  • the model sender device sends request response information to the model receiver device (optional):
  • step 3.a Request acceptance; if "transformation is accepted or successful" is received (step 3.a), inform the receiving end that its request is reasonable and feasible and has been accepted.
  • step 3.b Request Denial. For example, when receiving "conversion failure" (step 3.b), inform the model sending end device that its request is rejected. The reason for the failure can also be further informed based on the feedback in "step 3.b", for example, the transformation platform does not support the transformation from AI framework A to AI framework B at the moment. Or, the request of the model receiving device is rejected due to the reason of the model sending device, for example, the model sending device cannot generate the required AI model.
  • step 4 and step 5 below is not limited.
  • the conversion platform converts the AI model into a model that can be used by the AI framework of the model receiver device.
  • the conversion platform converts the received AI model file from the AI framework model of the device at the model sending end to a model available to the AI framework of the device at the model receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • the conversion platform (“tf2onnx.conver") can convert TensorFlow model files into ONNX files with the suffix ".onnx".
  • the calculation graph Graph contains information such as network structure and parameters, and consists of some basic information such as names and four sets of lists.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TetensorProto type
  • node stores all the computing nodes in the model
  • input stores the name of the model input node.
  • output stores the model output node name
  • initializer stores the specific values of all network parameters of the model, including hyperparameters and input values.
  • each Node contains information such as its operation type and specified inputs and outputs names.
  • the node information of a Node can be the calculation process of a whole layer, or the calculation of a node. All nodes are connected together to form a graph, and the two arrays of inputs and outputs in the Node calculation node point to the relationship between the input and output nodes to construct the topology of the entire network, that is, the network structure.
  • the conversion platform sends the converted AI model to the model receiving device.
  • the conversion platform sends the converted AI model file to the model receiving device, and the AI model file is supported by the model receiving device.
  • the model receiving device can understand and use the AI model file .
  • AI framework information of the model sender device may also be included (for example, the model sender device supports the TensorFlow framework). This information is used to indicate to the model receiver device the AI framework specifically used by the model sender device.
  • FIG. 13 is the fourth schematic flowchart of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 13, the AI model transmission method is applied to the model sending end device, and the AI model transmission method includes the following steps:
  • Step 1300 the model sending end device sends a third request message to the model conversion platform, the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the second AI model corresponds to the second model representation method
  • the first AI model corresponds to the first model representation method
  • model sending end device may directly send a third request message to the model conversion platform to request the model conversion platform to perform model conversion.
  • the model sending end device sends a third request message to the model conversion platform in certain scenarios, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model.
  • the AI model makes it possible to transfer and interoperate AI models of different platform frameworks, and also provides a solution for the transfer of AI models of devices from different manufacturers.
  • the third request message includes at least one of the following:
  • Attribute information of the device at the receiving end of the model
  • the second AI model includes files such as the complete network structure and parameter information of the AI model.
  • the attribute information of the model sending end device includes:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • the attribute information of the model receiver device includes:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • model transformation requirement information includes:
  • the method also includes:
  • the model sending end device receives the second model conversion response message sent by the model conversion platform, where the second model conversion response message is used to indicate whether the model conversion platform can complete the conversion operation on the second AI model.
  • the model sending end device After the model sending end device sends the third request message to the model conversion platform, it may receive a response message sent by the model conversion platform, that is, a second model conversion response message, which is used to indicate whether the model conversion platform can complete the conversion of the second model conversion platform. Transformation operation of AI model.
  • the method before the model sending device sends the third request message to the model conversion platform, the method further includes:
  • the model sending device receives a fourth request message sent by a third-party device, where the fourth request message is used to request the model sending device to send the second AI model to the model transformation platform.
  • the third-party device refers to a device that has no AI model and has reached an agreement with the model receiving device, and the third-party device triggers model conversion by sending a fourth request message to the model sending device.
  • the model sender device receives a fourth request message sent by a third-party device, the fourth request message is used to request the model sender device to send a second AI model to the model conversion platform to instruct the model conversion platform to perform The model is converted, and the model obtained after the model conversion is sent to the model receiving end device.
  • the embodiment of the present application provides a cross-platform framework, a cross-user/network element AI model transfer method, which is a method of setting up a conversion platform based on the model sending end.
  • This method saves hard disk space for the model sending end; reduces the subsequent update cost, increases the reuse rate of the conversion platform; increases the types of frameworks that support conversion; makes the AI model transfer and interoperability of different platform frameworks become a Possibly, it also provides a method for the AI model transfer of devices from different manufacturers; it can also make the model group sending, instead of being limited to one-to-one sending.
  • FIG 14 is the fifth schematic flow diagram of the AI model transmission method provided by the embodiment of the present application. As shown in Figure 14, the AI model transmission method is applied to the model conversion platform, and the AI model transmission method includes the following steps:
  • Step 1400 the model conversion platform receives a third request message sent by the model sending device, and the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • Step 1401 the model transformation platform executes a second operation according to the third request message
  • said performing the second operation includes at least one of the following:
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the model sending end device may directly send a third request message to the model conversion platform to request the model conversion platform to perform model conversion.
  • the model conversion platform After the model conversion platform receives the third request message sent by the model sending device, it determines whether it can complete the conversion of the second AI model according to its own conversion capability information, the framework information of the first AI model, and the framework information of the second AI model.
  • a model conversion operation sending a second model conversion response message to the model sending device, where the second model conversion response message is used to indicate whether the model conversion platform can complete the conversion operation on the second AI model.
  • the model conversion platform can complete the conversion operation of the second AI model, it converts the second AI model into the first AI model, and sends the first AI model to the model receiving device.
  • the model receiving end device can be multiple devices, so that the group sending of AI models can be realized, so that it is no longer limited to one-to-one sending.
  • model conversion platform cannot complete the conversion operation of the second AI model, then send a second model conversion response indicating that the model conversion platform cannot complete the conversion operation of the second AI model to the model sending device information.
  • the present application does not limit the sequence of sending the second model conversion response message to the model sending device and converting the second AI model into the first AI model.
  • the model conversion platform may encounter unpredictable failures, so that the conversion operation of the second AI model cannot be completed, and the model sending terminal The device sends a second model conversion response message for indicating that the model conversion platform cannot complete the conversion operation on the second AI model.
  • the third request message includes at least one of the following:
  • Attribute information of the device at the receiving end of the model
  • the attribute information of the model sending end device includes:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • the attribute information of the model receiver device includes:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • model transformation requirement information includes:
  • the embodiment of the present application provides a cross-platform framework, a cross-user/network element AI model transfer method, which is a method of setting up a conversion platform based on the model sending end.
  • This method saves hard disk space for the model sending end; reduces the subsequent update cost, increases the reuse rate of the conversion platform; increases the types of frameworks that support conversion; makes the AI model transfer and interoperability of different platform frameworks become a Possibly, it also provides a method for the AI model transfer of devices from different manufacturers; it can also make the model group sending, instead of being limited to one-to-one sending.
  • FIG. 15 is the seventh schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • the AI model transmission method includes the following steps:
  • step 1
  • the model sender sends a conversion request to the conversion platform, including at least one of the following information:
  • AI model file a file containing elements such as the complete network structure and parameter information of the generated AI model.
  • the AI model produced by TensorFlow will be saved as a file with a ".meta” suffix, which includes the network structure of the AI model; and a file with a ".ckpt” suffix, which contains parameter information. (required)
  • the framework used by the model sender inform the transformation platform of the framework used by the AI model file, so that the transformation platform can facilitate transformation. If the AI model file is a file generated by TesnorFLow, the information can be "TensorFLow” + "2.5.3” (that is, in the form of framework platform + version number). (If the conversion platform can judge by itself, it can be "optional")
  • Model receiving end information it is convenient for the conversion platform to send results to include address information or identification information of the receiving end, which can be sent implicitly to feed the model back to the receiving end. (optional)
  • AI framework information at the receiving end tell the conversion platform which framework model to convert the model into. If you need to convert the model file generated by TensorFLow into a model file recognizable by PyTorch, it can be in the form of "PyTorch” + "1.10.0" (framework platform + version number).
  • Requirement information for model conversion information used to guide the conversion platform for conversion, such as delay requirements (transformation is completed and sent within 5s), accuracy requirements (transformation accuracy rate of more than 99%), etc.
  • the conversion platform sends a response message to the sender:
  • the conversion platform can convert the received model information into a model corresponding to the receiving end framework, the response message is used to inform the sending end that the model conversion is accepted or successful;
  • the conversion platform cannot convert the received model information into a model corresponding to the framework of the receiving end, the response message is used to inform the sending end that the model conversion fails. It may further include a failure reason value, for example, because the conversion of the AI model from the framework A model to the framework B model is not supported.
  • step 1 that is, an AI model is successfully sent from the sender to the conversion platform
  • the conversion platform will convert the AI model into a model available to the AI framework of the receiver.
  • the conversion platform converts the received AI model file from the AI framework model at the sending end to a model available to the AI framework at the receiving end. For example, from model A of the TensorFlow framework to model B of the ONNX framework.
  • model A of the TensorFlow framework to model B of the ONNX framework.
  • Python script provided by ONNX, "tf2onnx.conver” can convert TensorFlow model files into ONNX files with the suffix ".onnx”.
  • node node
  • input ValueInfoProto type
  • output ValueInfoProto type
  • initializer TinsorProto type
  • input nodes in the model are stored in input
  • output nodes in the model are stored in output
  • all weight parameters in the model are stored in initializer.
  • the conversion platform sends the converted AI model to the receiving end. Specifically, the conversion platform sends the converted AI model file to the receiving end.
  • the AI model file is supported by the receiving end. In other words, the receiving end This AI model file can be understood and used.
  • the AI framework information of the sender may also be included (such as the IP address of the sender, which supports the TensorFlow framework). This information is used to indicate to the receiving end the specific AI framework used by the sending end.
  • FIG. 16 is an eighth schematic diagram of the interaction process of the AI model transmission method provided by the embodiment of the present application.
  • the method also includes:
  • the model sender device receives a fourth request message sent by a third-party device, the fourth request message is used to request the model sender device to send the second AI model to the model conversion platform, so that the model conversion platform will
  • the second AI model is converted into the first AI model and the first AI model is delivered to the model receiving end device.
  • the AI model transmission method provided in the embodiment of the present application may be executed by an AI model transmission device.
  • the AI model transmission device provided in the embodiment of the present application is described by taking the AI model transmission device executing the AI model transmission method as an example.
  • FIG. 17 is one of the structural schematic diagrams of the AI model transmission device provided by the embodiment of the present application. As shown in Figure 17, the AI model transmission device 1700 includes:
  • the first sending unit 1710 is configured to send a first request message, the first request message is used to request a first AI model, and the first request message includes AI model description information;
  • a first receiving unit 1720 configured to receive a first response message
  • the first response message includes one of the following:
  • the first AI model is a model obtained after the second AI model from the model sending end device is transformed by the model conversion platform, and the first AI model can be used by the model receiving end device;
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the first request message is used to request the first AI model, the first request message includes AI model description information, and the first response message is received, the first response message Including the first AI model, the first instruction information or the second instruction information, the first AI model is a model that can be used by the model receiving end device obtained after the second AI model from the model sending end device is transformed by the model transformation platform, This makes it possible to transfer and interoperate with AI models of different platform frameworks, and provides a solution for the transfer of AI models of devices from different manufacturers.
  • the first sending unit is configured to:
  • the first receiving unit is configured to:
  • the first response message from the model conversion platform, where the first response message includes the first AI model or the first indication information or the second indication information.
  • the first receiving unit is configured to:
  • the AI model description information includes at least one of the following:
  • AI model training accuracy requires information.
  • the first request message sent by the model receiver device to the model sender device further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the device also includes:
  • the first reporting unit is configured to register and report the attribute information of the model receiving end device to the model conversion platform, where the attribute information includes the AI framework information supported by the model receiving end device.
  • the device also includes:
  • the second sending unit is configured to send the AI framework information supported by the model receiving end device to the model conversion platform when receiving the first AI framework request message sent by the model conversion platform, wherein the The first AI framework request message is used to request the AI framework information supported by the model receiver device.
  • the first request message when the first request message is sent to the model transformation platform, the first request message further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes at least one of the following:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the second instruction information is obtained from the model conversion platform.
  • the model receiver device acquires the second indication information from the model sender device;
  • the first condition includes at least one of the following:
  • the model sending end device determines that the model conversion platform cannot complete the conversion operation on the second AI model according to the conversion capability information of the model conversion platform;
  • the model sending end device receives the first model conversion response message sent by the model conversion platform, and the first model conversion response message indicates that the model conversion platform cannot complete the conversion operation on the second AI model.
  • the model conversion platform cannot complete the conversion operation of the second AI model, including:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model conversion delay requirement requested by the model receiving end device; or,
  • the conversion accuracy rate of the second AI model by the model conversion platform cannot meet the model conversion accuracy requirement requested by the model receiving end device.
  • the first indication information is obtained from the model sending device or a model transformation platform.
  • the first response message includes the first AI model, the first indication information or the second indication information
  • the first AI model is from the model sending end
  • the second AI model of the device is transformed by the model conversion platform and can be used by the model receiving end device.
  • the model conversion platform can be deployed behind the model receiving end device or behind the model sending end device , or the model conversion platform is a public conversion platform, which reduces the subsequent update cost, increases the reuse rate of the conversion platform, and increases the types of frameworks that support conversion, making it possible for AI models of different platform frameworks to communicate with each other. It also provides a solution for AI model transfer of devices from different manufacturers.
  • the AI model transmission device provided in the embodiment of the present application can realize the various processes realized by the method embodiments in Fig. 2, Fig. 6-Fig. 12, and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • FIG. 18 is the second structural schematic diagram of the AI model transmission device provided by the embodiment of the present application. As shown in Figure 18, the AI model transmission device 1800 includes:
  • the second receiving unit 1810 is configured to receive an AI model request message, the AI model request message is used to request the first AI model, and the AI model request message includes AI model description information;
  • the first executing unit 1820 is configured to execute a first operation according to the AI model request message
  • said performing the first operation includes at least one of the following:
  • the first indication information is used to indicate that the model sending end device cannot provide the second AI model
  • the second indication information is used to indicate that the model conversion platform cannot converting the second AI model into the first AI model
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the first operation is performed according to the AI model request message, including selecting or training to generate a second AI model, sending the second AI model to the model conversion platform, or sending
  • the first instruction information or the second instruction information makes it possible to transfer and intercommunicate AI models of different platform frameworks, and provides a solution for the transfer of AI models of devices from different manufacturers.
  • the second receiving unit is configured to:
  • the sending the second AI model to the model conversion platform includes:
  • the sending the second AI model to the model conversion platform includes:
  • the third request message is used to request the model conversion platform to convert the second AI model into the first AI model; the third request message includes the second AI model Model.
  • the AI model description information includes at least one of the following:
  • AI model training accuracy requires information.
  • the second response message further includes at least one of the following:
  • the attribute information of the model receiver device is the attribute information of the model receiver device.
  • the first request message further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the device also includes:
  • the second reporting unit is configured to register and report the attribute information of the model sender device to the model conversion platform, where the attribute information includes AI framework information supported by the model sender device.
  • the device also includes:
  • the third sending unit is configured to send the AI framework information supported by the model sending end device to the model conversion platform when receiving the second AI framework request message sent by the model conversion platform, wherein the first The second AI framework request message is used to request the AI framework information supported by the model sending end device.
  • the second request message further includes at least one of the following:
  • the first request message sent to the model conversion platform further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate requirement information for conversion to obtain the first AI model.
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the sending the first indication information includes:
  • model sending end device does not have or cannot generate the second AI model, send the first indication information to the model receiving end device.
  • the device also includes:
  • a fourth sending unit configured to send a fourth response message to the model conversion platform when the model sending device does not or cannot generate the second AI model, and the fourth response message is used to indicate the The model sending end device cannot provide the second AI model.
  • the device also includes:
  • the first determining unit is configured to determine whether the model conversion platform can complete the conversion operation of the second AI model according to the conversion capability information of the model conversion platform.
  • the device also includes:
  • a third receiving unit configured to receive a first model conversion response message sent by the model conversion platform, where the first model conversion response message is used to indicate whether the model conversion platform has completed the conversion operation on the second AI model .
  • the sending the second indication information includes:
  • the first condition includes at least one of the following:
  • the model conversion platform According to the conversion capability information of the model conversion platform, it is determined that the model conversion platform cannot complete the conversion operation on the second AI model;
  • a first model conversion response message sent by the model conversion platform is received, and the first model conversion response message indicates that the model conversion platform cannot complete the conversion operation on the second AI model.
  • the model conversion platform cannot complete the conversion operation of the second AI model, including:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model conversion delay requirement requested by the model receiving end device; or,
  • the conversion accuracy rate of the second AI model by the model conversion platform cannot meet the model conversion accuracy requirement requested by the model receiving end device.
  • the first operation is performed according to the AI model request message, including selecting or training to generate a second AI model, sending the second AI model to the model conversion platform, or sending The first instruction information or the second instruction information
  • the second AI model is converted by the model conversion platform and can be used by the model receiving end device.
  • the model conversion platform can be deployed behind the model receiving end device or at the The rear of the model sending end device, or the model conversion platform is a public conversion platform, which reduces the subsequent update cost, increases the reuse rate of the conversion platform, and increases the types of frameworks that support conversion, making AI models of different platform frameworks Transfer and intercommunication has become possible, and it also provides a solution for the transfer of AI models of devices from different manufacturers.
  • the AI model transmission device provided by the embodiment of the present application can realize the various processes realized by the method embodiments in Fig. 3, Fig. 6-Fig. 12, and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • FIG. 19 is the third structural schematic diagram of the AI model transmission device provided by the embodiment of the present application. As shown in Figure 19, the AI model transmission device 1900 includes:
  • the fourth receiving unit 1910 is configured to receive the second AI model sent by the model sending device
  • the first processing unit 1920 is configured to convert the second AI model into the first AI model, and send the first AI model to the model receiving end device; or,
  • model conversion platform includes:
  • a fifth sending unit configured to send first indication information or second indication information to the model receiver device, where the first indication information is used to indicate that the model sender device cannot provide the second AI model, and the first The second indication information is used to indicate that the model conversion platform cannot convert the second AI model into the first AI model;
  • the first AI model corresponds to a first model representation method
  • the second AI model corresponds to a second model representation method
  • the second AI model sent by the model sending end device is received; the second AI model is converted into the first AI model, and the first AI model is sent to the model receiving end device, or Sending the first instruction information or the second instruction information to the model receiving end device makes it possible to transfer and communicate AI models of different platform frameworks, and also provides a solution for the transfer of AI models of devices from different manufacturers.
  • the fourth receiving unit is configured to:
  • the fourth receiving unit is configured to:
  • the device further includes a fifth receiving unit, configured to:
  • the attribute information of the model receiving end device reported by the model receiving end device is received, where the attribute information of the model receiving end device includes AI framework information supported by the model receiving end device.
  • the device further includes a sixth sending unit, configured to:
  • the second response message further includes at least one of the following:
  • the attribute information of the model receiver device is the attribute information of the model receiver device.
  • the device also includes:
  • a sixth receiving unit configured to receive the first request message sent by the model receiving end device
  • a seventh sending unit configured for the model conversion platform to send a second request message to the model sending device
  • the second request information is determined according to the first request message.
  • the first request message or the second request message includes: AI model description information, where the model description information includes at least one of the following:
  • AI model training accuracy requires information.
  • the first request message further includes at least one of the following:
  • Model conversion requirement information where the model conversion requirement information is used to indicate the requirement information for conversion to obtain the first AI model
  • the second request message also includes at least one of the following:
  • the attribute information of the model conversion platform is the attribute information of the model conversion platform.
  • the attribute information of the model receiver device includes at least one of the following:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • the attribute information of the model sending end device includes at least one of the following:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • model transformation requirement information includes at least one of the following:
  • the attribute information of the model conversion platform includes at least one of the following:
  • Transformation capability information of the model transformation platform is
  • the device also includes:
  • the second determination unit is configured to determine that the model conversion platform has the capability of converting the AI framework supported by the model sending device to the AI framework supported by the model receiving device.
  • the device also includes:
  • the third determination unit is used to determine that the model conversion platform does not have the ability to convert the AI framework supported by the model sending end device into the AI framework supported by the model receiving end device;
  • An eighth sending unit configured to send a third response message to the model receiver device, where the third response message is used to indicate that the first request message is rejected.
  • the apparatus further includes:
  • a fourth determining unit configured to determine the model sending end device.
  • the device also includes:
  • a ninth sending unit configured to send a first model conversion response message to the model sending device, where the first model conversion response message is used to indicate whether the model conversion platform has completed the conversion operation on the second AI model.
  • the fifth sending unit is configured to;
  • the model conversion platform cannot complete the conversion operation of the second AI model, including at least one of the following:
  • the model conversion platform does not support converting the second AI model from the AI framework of the model sending device to the AI framework of the model receiving device; or,
  • the conversion delay of the second AI model by the model conversion platform cannot meet the model provision delay requirement requested by the model receiving end device; or,
  • the conversion accuracy of the second AI model by the model conversion platform cannot meet the model accuracy requirement requested by the model receiving end device.
  • the device also includes:
  • a seventh receiving unit configured to receive a fourth response message sent by the model sending device, where the fourth response message is used to indicate that the model sending device cannot provide the second AI model.
  • the second AI model sent by the model sending end device is received; the second AI model is converted into the first AI model, and the first AI model is sent to the model receiving end device, or Send the first indication information or the second indication information to the model receiving end device, wherein the model conversion platform can be deployed behind the model receiving end device or behind the model sending end device, or the model conversion platform is a public conversion platform , while reducing subsequent update costs, increased the reuse rate of the conversion platform, increased the types of frameworks that support conversion, made it possible to transfer and interoperate AI models of different platform frameworks, and also provided AI models for devices from different manufacturers Pass provides a solution.
  • the AI model transmission device provided by the embodiment of the present application can implement the various processes realized by the method embodiments in Fig. 4-Fig. 12 and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • FIG. 20 is a fourth structural schematic diagram of the AI model transmission device provided by the embodiment of the present application. As shown in Figure 20, the AI model transmission device 2000 includes:
  • a tenth sending unit 2010, configured to send a third request message to the model conversion platform, where the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the second AI model corresponds to the second model representation method
  • the first AI model corresponds to the first model representation method
  • the third request message is used to request the model conversion platform to convert the second AI model into the first AI model , which makes it possible to transfer and interoperate AI models of different platform frameworks, and also provides a solution for the transfer of AI models of devices from different manufacturers.
  • the third request message includes at least one of the following:
  • Attribute information of the device at the receiving end of the model
  • the attribute information of the model sending end device includes:
  • the identification information of the model sending device is the identification information of the model sending device
  • AI framework information supported by the model sender device where the AI framework information supported by the model sender device corresponds to the second model representation method.
  • the attribute information of the model receiver device includes:
  • the AI framework information supported by the model receiver device, the AI framework information supported by the model receiver device corresponds to the first model representation method.
  • model transformation requirement information includes:
  • the device also includes:
  • An eighth receiving unit configured to receive a second model conversion response message sent by the model conversion platform, where the second model conversion response message is used to indicate whether the model conversion platform can complete the conversion operation on the second AI model.
  • the device also includes:
  • a ninth receiving unit configured for the model sending device to receive a fourth request message sent by a third-party device, where the fourth request message is used to request the model sending device to send a second AI model to the model conversion platform .
  • the embodiment of this application provides a cross-platform framework, a cross-user/network element AI model transfer scheme, which can save hard disk space for the model sender, reduce subsequent update costs, and increase the reuse of the conversion platform rate; increased the types of frameworks that support conversion; made it possible for AI model transfer and intercommunication of different platform frameworks, and also provided a method for AI model transfer of devices from different manufacturers; it also allows models to be distributed in groups, not limited to a pair A send.
  • the AI model transmission device provided by the embodiment of the present application can implement the various processes realized by the method embodiments in Fig. 13, Fig. 15-Fig. 16, and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • FIG. 21 is the fifth structural schematic diagram of the AI model transmission device provided by the embodiment of the present application. As shown in Figure 21, the AI model transmission device 2100 includes:
  • a tenth receiving unit 2110 configured to receive a third request message sent by the model sending device, where the third request message is used to request the model conversion platform to convert the second AI model into the first AI model;
  • the second execution unit 2120 is configured to execute a second operation according to the third request message
  • said performing the second operation includes at least one of the following:

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • General Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Software Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Communication Control (AREA)
  • Telephonic Communication Services (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

本申请公开了一种AI模型传输方法、装置、设备及存储介质,属于通信技术领域,本申请实施例的AI模型传输方法包括:模型接收端设备发送第一请求消息,第一请求消息用于请求第一AI模型,第一请求消息包括AI模型描述信息;模型接收端设备接收第一响应消息;其中,第一响应消息包括以下其中之一:第一AI模型,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型;第一指示信息,用于指示模型发送端设备无法提供第二AI模型;第二指示信息,用于指示模型转化平台无法将第二AI模型转换成第一AI模型;其中,第一AI模型对应第一模型表示方法,第二AI模型对应第二模型表示方法。

Description

AI模型传输方法、装置、设备及存储介质
相关申请的交叉引用
本申请要求于2021年12月28日提交的申请号为202111627453.3,发明名称为“AI模型传输方法、装置、设备及存储介质”的中国专利申请的优先权,其通过引用方式全部并入本申请。
技术领域
本申请属于通信技术领域,具体涉及一种AI模型传输方法、装置、设备及存储介质。
背景技术
人工智能(Artificial Intelligence,AI)模型根据实现方法不同(所选用的框架不同),导致所生成的文件并不互通理解,如Pytorch和TensorFlow是两种用来构建AI网络模型的框架平台,他们所生成的模型以“.pth”和“.meta/.index”的形式保存,而这两种文件只能被他们自己的框架所读取、调用等操作,这使得AI模型不能直接跨框架平台或者跨用户使用。
具体到无线通信网络场景中,如果网络引入AI功能,不同设备(网络设备、终端设备、应用设备等)来自不同的厂家或设备商,它们很可能使用不同的AI模型框架,则导致不同设备之间不能互通AI模型,从而导致智能化特性无法很好的实现。
发明内容
本申请实施例提供一种AI模型传输方法、装置、设备及存储介质,能够解决无线通信网络场景中网络引入AI功能的情况下,不同设备之间不能互通AI模型所导致的智能化特性无法较好地实现的问题。
第一方面,提供了一种AI模型传输方法,应用于模型接收端设备,该方法包括:
模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;
所述模型接收端设备接收第一响应消息;
其中,所述第一响应消息包括以下其中之一:
所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第二方面,提供了一种AI模型传输方法,应用于模型发送端设备,该方法包括:
模型发送端设备接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;
模型发送端设备根据所述AI模型请求消息,执行第一操作;
其中,所述执行第一操作包括以下至少一项:
选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第三方面,提供了一种AI模型传输方法,应用于模型转化平台,该方法包括:
模型转化平台接收模型发送端设备发送的第二AI模型;
所述模型转化平台将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;或者,
所述模型转化平台向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第四方面,提供了一种AI模型传输方法,应用于模型发送端设备,该方法包括:
模型发送端设备向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
第五方面,提供了一种AI模型传输方法,应用于模型转化平台,该方法包括:
模型转化平台接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
所述模型转化平台根据所述第三请求消息,执行第二操作;
其中,所述执行第二操作包括以下至少一项:
确定是否能够完成对所述第二AI模型的转化操作;
向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
将第二AI模型转化为第一AI模型;
向模型接收端设备发送所述第一AI模型。
第六方面,提供一种AI模型传输装置,包括:
第一发送单元,用于发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;
第一接收单元,用于接收第一响应消息;
其中,所述第一响应消息包括以下其中之一:
所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第七方面,提供一种AI模型传输装置,包括:
第二接收单元,用于接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;
第一执行单元,用于根据所述AI模型请求消息,执行第一操作;
其中,所述执行第一操作包括以下至少一项:
选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第八方面,提供一种AI模型传输装置,包括:
第四接收单元,用于接收模型发送端设备发送的第二AI模型;
第一处理单元,用于将所述第二AI模型转化为第一AI模型,并将所述第一AI 模型发送至模型接收端设备;或者,
第五发送单元,用于向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第九方面,提供一种AI模型传输装置,包括:
第十发送单元,用于向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
第十方面,提供一种AI模型传输装置,包括:
第十接收单元,用于接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
第二执行单元,用于根据所述第三请求消息,执行第二操作;
其中,所述执行第二操作包括以下至少一项:
确定是否能够完成对所述第二AI模型的转化操作;
向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
将第二AI模型转化为第一AI模型;
向模型接收端设备发送所述第一AI模型。
第十一方面,提供了一种模型接收端设备,该模型接收端设备包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如第一方面所述的方法的步骤。
第十二方面,提供了一种模型接收端设备,包括处理器及通信接口,其中,所述通信接口用于发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;所述通信接口还用于接收第一响应消息;其中,所述第一响应消息包括以下其中之一:所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
第十三方面,提供了一种模型发送端设备,该模型发送端设备包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处 理器执行时实现如第二方面或第四方面所述的方法的步骤。
第十四方面,提供了一种模型发送端设备,包括处理器及通信接口,其中,所述通信接口用于接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;所述处理器用于根据所述AI模型请求消息,执行第一操作;.其中,所述执行第一操作包括以下至少一项:选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
或者,所述通信接口用于向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
第十五方面,提供了一种模型转化平台,该模型转化平台包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如第三方面或第五方面所述的方法的步骤。
第十六方面,提供了一种模型转化平台,包括处理器及通信接口,其中,所述通信接口用于接收模型发送端设备发送的第二AI模型;所述处理器用于将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;所述通信接口还用于:向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
或者,所述通信接口用于接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;所述处理器用于根据所述第三请求消息,执行第二操作;其中,所述执行第二操作包括以下至少一项:确定是否能够完成对所述第二AI模型的转化操作;向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;将第二AI模型转化为第一AI模型;向模型接收端设备发送所述第一AI模型。
第十七方面,提供了一种AI模型传输系统,包括:模型接收端设备、模型发送端设备及模型转化平台,所述模型接收端设备可用于执行如第一方面所述的AI模型传输方法的步骤,所述模型发送端设备可用于执行如第二方面或第四方面所述的AI模型传输方法的步骤。
第十八方面,提供了一种可读存储介质,所述可读存储介质上存储程序或指令,所述程序或指令被处理器执行时实现如第一方面所述的方法的步骤,或者实现如第二方面所述的方法的步骤,或者实现如第三方面所述的方法的步骤,或者实现如第四方面所述的方法的步骤,或者实现如第五方面所述的方法的步骤。
第十九方面,提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和所述处理器耦合,所述处理器用于运行程序或指令,实现如第一方面所述的方法,或者实现如第二方面所述的方法,或者实现如第三方面所述的方法,或者实现如第四方面所述的方法,或者实现如第五方面所述的方法。
第二十方面,提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现如第一方面所述的方法的步骤,或者实现如第二方面所述的方法的步骤,或者实现如第三方面所述的方法的步骤,或者实现如第四方面所述的方法的步骤,或者实现如第五方面所述的方法的步骤。
在本申请实施例中,模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息,接收第一响应消息,第一响应消息包括第一AI模型、第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,从而使得不同平台框架的AI模型传递互通成为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
附图说明
图1是本申请实施例可应用的一种无线通信系统的框图;
图2为神经元的示意图;
图3为本申请实施例提供的AI模型传输方法的流程示意图之一;
图4为本申请实施例提供的AI模型传输方法的流程示意图之二;
图5为本申请实施例提供的AI模型传输方法的流程示意图之三;
图6为本申请实施例提供的AI模型传输方法的交互流程示意图之一;
图7为本申请实施例提供的AI模型传输方法的交互流程示意图之二;
图8为本申请实施例提供的AI模型传输方法的交互流程示意图之三;
图9为本申请实施例提供的AI模型传输方法的交互流程示意图之四;
图10为本申请实施例提供的向公共转化平台注册上报能力信息的示意图;
图11为本申请实施例提供的AI模型传输方法的交互流程示意图之五;
图12为本申请实施例提供的AI模型传输方法的交互流程示意图之六;
图13为本申请实施例提供的AI模型传输方法的流程示意图之四;
图14为本申请实施例提供的AI模型传输方法的流程示意图之五;
图15为本申请实施例提供的AI模型传输方法的交互流程示意图之七;
图16为本申请实施例提供的AI模型传输方法的交互流程示意图之八;
图17为本申请实施例提供的AI模型传输装置的结构示意图之一;
图18为本申请实施例提供的AI模型传输装置的结构示意图之二;
图19为本申请实施例提供的AI模型传输装置的结构示意图之三;
图20为本申请实施例提供的AI模型传输装置的结构示意图之四;
图21为本申请实施例提供的AI模型传输装置的结构示意图之五;
图22为本申请实施例提供的通信设备的结构示意图;
图23为实现本申请实施例的一种模型接收端设备的硬件结构示意图;
图24为实现本申请实施例的一种模型发送端设备的硬件结构示意图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员所获得的所有其他实施例,都属于本申请保护的范围。
本申请的说明书和权利要求书中的术语“第一”、“第二”等是用于区别类似的对象,而不用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换,以便本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施,且“第一”、“第二”所区别的对象通常为一类,并不限定对象的个数,例如第一对象可以是一个,也可以是多个。此外,说明书以及权利要求中“和/或”表示所连接对象的至少其中之一,字符“/”一般表示前后关联对象是一种“或”的关系。
值得指出的是,本申请实施例所描述的技术不限于长期演进型(Long Term Evolution,LTE)/LTE的演进(LTE-Advanced,LTE-A)系统,还可用于其他无线通信系统,诸如码分多址(Code Division Multiple Access,CDMA)、时分多址(Time Division Multiple Access,TDMA)、频分多址(Frequency Division Multiple Access,FDMA)、正交频分多址(Orthogonal Frequency Division Multiple Access,OFDMA)、单载波频分多址(Single-carrier Frequency Division Multiple Access,SC-FDMA)和其他系统。本申请实施例中的术语“系统”和“网络”常被可互换地使用,所描述的技术既可用于以上提及的系统和无线电技术,也可用于其他系统和无线电技术。以下描述出于示例目的描述了新空口(New Radio,NR)系统,并且在以下大部分描述中使用NR术语,但是这些技术也可应用于NR系统应用以外的应用,如第6代(6 th Generation,6G)通信系统。
图1示出本申请实施例可应用的一种无线通信系统的框图。无线通信系统包括终端11和网络侧设备12。其中,终端11可以是手机、平板电脑(Tablet Personal Computer)、膝上型电脑(Laptop Computer)或称为笔记本电脑、个人数字助理(Personal Digital Assistant,PDA)、掌上电脑、上网本、超级移动个人计算机(ultra-mobile personal computer,UMPC)、 移动上网装置(Mobile Internet Device,MID)、增强现实(augmented reality,AR)/虚拟现实(virtual reality,VR)设备、机器人、可穿戴式设备(Wearable Device)、车载设备(VUE)、行人终端(PUE)、智能家居(具有无线通信功能的家居设备,如冰箱、电视、洗衣机或者家具等)、游戏机、个人计算机(personal computer,PC)、柜员机或者自助机等终端侧设备,可穿戴式设备包括:智能手表、智能手环、智能耳机、智能眼镜、智能首饰(智能手镯、智能手链、智能戒指、智能项链、智能脚镯、智能脚链等)、智能腕带、智能服装等。需要说明的是,在本申请实施例并不限定终端11的具体类型。网络侧设备12可以包括接入网设备或核心网设备,其中,接入网设备12也可以称为无线接入网设备、无线接入网(Radio Access Network,RAN)、无线接入网功能或无线接入网单元。接入网设备12可以包括基站、WLAN接入点或WiFi节点等,基站可被称为节点B、演进节点B(eNB)、接入点、基收发机站(Base Transceiver Station,BTS)、无线电基站、无线电收发机、基本服务集(Basic Service Set,BSS)、扩展服务集(Extended Service Set,ESS)、家用B节点、家用演进型B节点、发送接收点(Transmitting Receiving Point,TRP)或所述领域中其他某个合适的术语,只要达到相同的技术效果,所述基站不限于特定技术词汇,需要说明的是,在本申请实施例中仅以NR系统中的基站为例进行介绍,并不限定基站的具体类型。核心网设备可以包含但不限于如下至少一项:核心网节点、核心网功能、移动管理实体(Mobility Management Entity,MME)、接入移动管理功能(Access and Mobility Management Function,AMF)、会话管理功能(Session Management Function,SMF)、用户平面功能(User Plane Function,UPF)、策略控制功能(Policy Control Function,PCF)、策略与计费规则功能单元(Policy and Charging Rules Function,PCRF)、边缘应用服务发现功能(Edge Application Server Discovery Function,EASDF)、统一数据管理(Unified Data Management,UDM),统一数据仓储(Unified Data Repository,UDR)、归属用户服务器(Home Subscriber Server,HSS)、集中式网络配置(Centralized network configuration,CNC)、网络存储功能(Network Repository Function,NRF),网络开放功能(Network Exposure Function,NEF)、本地NEF(Local NEF,或L-NEF)、绑定支持功能(Binding Support Function,BSF)、应用功能(Application Function,AF)等。需要说明的是,在本申请实施例中仅以NR系统中的核心网设备为例进行介绍,并不限定核心网设备的具体类型。
首先对本申请涉及的相关内容进行介绍。
一、人工智能(AI)及AI模型
人工智能目前在各个领域获得了广泛的应用。AI模型有多种算法实现方式,例如神经网络、决策树、支持向量机、贝叶斯分类器等。本申请以神经网络为例进行说明,但是并不限定AI模块的具体类型。
一个神经网络包括多个神经元。X1,X2…Xn等为输入值,Y为输出结果,一个个神经元也是进行运算的地方,结果会继续传入到下一层。这些众多神经元组成的一输入层、隐藏层、输出层就是一个神经网络。隐藏层的数量,每一层神经元的数量就是神经网络的“网 络结构”。
其中,神经网络由神经元组成,神经元的示意图如图2所示。其中a1,a2,…aK(即上文的X1…Xn)为输入,w为权值(乘性系数),b为偏置(加性系数),σ(.)为激活函数,z为输出值。常见的激活函数包括Sigmoid、tanh、ReLU(Rectified Linear Unit,线性整流函数,修正线性单元)等等。每一个神经元的参数信息和所用算法组合在一起就是整个网络的“参数信息”,也是AI模型文件中很重要的一部分。
在实际使用过程中,一个AI模型指的是一个包含网络结构和参数信息等元素的文件,经过训练的AI模型可被其框架平台直接再次使用,无需重复构建或者学习,直接进行判断,识别等智能化功能。
二、网络框架
神经网络有很多种实现框架,包括TensorFlow,PyTorch,Keras,MXNet,Caffe2等,每种框架的侧重点各不相同,例如,Caffe2与Keras是高层的深度学习框架可以快速地验证模型,TensorFlow与PyTorch是底层的深度学习框架可以实现对神经网络底层细节的修改。
又例如,PyTorch的重点在于支持动态图模型,TensorFlow重点在于支持多种硬件,运行速度快,Caffe2在于轻量级等。每一个实现框架都会使用自己的方法对神经网络进行描述,完成网络的搭建、训练、推断等操作。一般不同实现框架下模型的描述方法并不能被其他框架所理解,导致它们之间的模型不能互通使用。
三、开放神经网络交换(Open Neural Network Exchange,ONNX)
ONNX是一种比较通用的AI模型的描述语言或称之为模型的表达方法,ONNX本身只是一种数据结构,不包括实现方案,用于描述一个AI网络。ONNX定义了一组与环境和平台无关的标准格式,为AI模型的互操作性提供了基础,使AI模型可以在不同框架和环境下交互使用。
ONNX将网络的每个算子都描述成一个节点,每个节点的输入和输出的名称都是全局唯一的,通过输入和输出名称的匹配关系描述整个网络的结构。所有的权重参数都被视为输入或输出,也是通过名称检索,具体的权重数值存储在单独的位置,根据每个节点输入输出的名称去存储位置获取对应的参数。
目前各AI平台的框架相互独立,所产生的模型只有极少的几种可以相互转化,导致他们的模型不能互通,没有办法在其他框架上用,在实际使用中有很大阻碍。因为每个框架的侧重点甚至支持的开发语言不尽相同,导致使用不同框架的两个节点无法传递AI模型的信息,因为使用不同框架的两个节点对网络的描述方法不同,数据的压缩方法不同,文件保存格式也不同,无法解析其他框架训练好的网络。同样,不同的开发者对框架的定义、描述方式不同,相互之间壁垒较高,难以相互转换。即使在相同的框架内,不同的版本等信息也会导致文件保存格式不同,两个框架版本不同的节点传递AI模型的时候,会导致解析的结果偏差或者无法解析
另外,虽然ONNX有意被发展作为一种描述AI模型的通用方案,已经被一些AI平台所支持,但并未完全被所有平台所支持。若欲将其发展完全通用的标准化方案,则需要所有平台(所有终端、网元、应用服务器)都搭载ONNX,可能会造成不必要的软硬件成本的增加。
下面结合附图,通过一些实施例及其应用场景对本申请实施例提供的AI模型传输方法、装置、设备及存储介质进行详细地说明。
在本申请实施例中,模型接收端设备可以是一个或多个终端,其因AI业务需求需要向模型发送端设备请求相应的AI模型。
模型发送端设备可以是一个或多个通信网络中的网元或者第三方业务服务器,其具备足够的计算能力,可以进行AI业务的模型训练,并将模型提供给模型需求者使用。
模型接收端设备因软硬件条件或成本的限制,仅具有理解少数AI模型框架的能力,无法理解来自各类模型发送端设备的各类框架的AI模型,无法直接使用。但该一个或多个模型接收端设备后方可能会统一部署模型转化平台(如终端设备厂商统一部署云化平台),将来自模型发送端设备的AI模型转换成各自支持的模型类型并下发至每个模型接收端设备。
需要说明的是,模型转化平台也可以部署在模型发送端设备的后方,如模型发送端设备统一部署云化平台。
或者,模型转化平台为一个公共的转化平台,模型发送端设备和模型接收端设备均可与模型转化平台进行交互,例如向该模型转化平台注册上报自身的属性信息,如标识地址,地址信息和/或所支持的AI框架信息。
需要说明的是,本申请实施例中的模型接收端设备并不限于终端,模型接收端设备也可以是其他请求AI模型的设备。
模型发送端设备也不限于通信网络中的网元或者第三方业务服务器,也可以是其他可提供AI模型的设备。
图3为本申请实施例提供的AI模型传输方法的流程示意图之一,如图3所示,该方法应用于模型接收端设备,该方法包括:
步骤300、模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息。
可以理解的是,模型接收端设备发送第一请求消息,以请求第一AI模型。
所述第一请求消息包括AI模型描述信息,该AI模型描述信息用于指示所请求的AI模型的信息,即该AI模型描述信息用于描述第一AI模型。其中,第一AI模型对应第一模型表示方法。
可选地,所述AI模型描述信息包括以下至少一项:
AI模型类型信息,如analytic ID;
AI模型标识信息,如model ID;
AI模型名称信息;
AI模型算法信息,即模型所使用的算法,如神经网络,随机森林等;
AI模型对应的训练对象信息,如某个用户,某个区域AOI;
AI模型对应的时间信息,即表示请求针对特定时间段、时间点的模型;
AI模型训练精度要求信息,如模型准确度90%等。
需要说明的是,第一请求消息必须携带AI模型描述信息。
可选地,第一请求消息还包括以下至少一项:
模型接收端设备的属性信息,包括模型接收端设备的标识信息、地址信息和/或所支持或选用的AI框架信息;
模型发送端设备的属性信息,包括模型发送端设备的标识信息、地址信息和/或所支持或选用的AI框架信息;
模型转化平台的属性信息,包括模型转化平台的标识信息、地址信息和/或转化能力信息,其中,转化能力信息用于指示所支持AI框架类型,以及具备将这些AI框架类型转化为模型接收端设备所支持的AI框架类型的能力,可选地,转化平台应该可以支持所有主流框架;
模型转化要求信息,用于指示对转化得到所述第一AI模型的要求信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
在本申请实施例中,AI框架信息用于描述AI模型的实现框架,AI框架信息包括AI框架的类型信息、网络拓扑信息、数据压缩方法信息、设计动机信息、数据格式信息、所用语言信息中的至少一种信息。当前已知的AI框架有多种,例如TensorFlow,PyTorch,Keras,MXNet,Caffe2等框架。
可选地,所述模型接收端设备发送第一请求消息,包括:
所述模型接收端设备向模型发送端设备发送所述第一请求消息;或者,
所述模型接收端设备向所述模型转化平台发送所述第一请求消息。
可以理解的是,在模型转化平台部署在模型接收端设备的后方的情况下,模型接收端设备可以向模型发送端设备或模型转化平台发送所述第一请求消息;
在模型转化平台部署在模型发送端设备的后方或为公共的转化平台的情况下,所述模型接收端设备向模型发送端设备发送所述第一请求消息。
步骤301、所述模型接收端设备接收第一响应消息;
其中,所述第一响应消息包括以下其中之一:
所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
可以理解的是,模型接收端设备发送第一请求消息后,会接收到第一响应消息。
第一响应消息的内容可以是请求被接受,即模型接收端设备可以得到第一AI模型。
第一响应消息的内容也可以是请求被拒绝,即模型接收端设备无法得到第一AI模型,模型接收端设备会接收到第一指示信息或第二指示信息。通过第一指示信息或第二指示信息,模型接收端设备可以获知请求被拒绝。
可选地,第一AI模型包括AI模型的完整的网络结构和参数信息的文件。例如使用TensorFlow生成的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
其中,第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,第二指示信息用于指示是模型转化平台无法将来自模型发送端设备的第二AI模型转换成所请求的第一AI模型。
可选地,所述模型接收端设备接收第一响应消息,包括:
所述模型接收端设备从所述模型转化平台获取所述第一响应消息,所述第一响应消息包括所述第一AI模型或所述第一指示信息或所述第二指示信息。
由于来自模型发送端设备的第二AI模型需要经过模型转化平台转化后,才能生成第一AI模型,因此,模型接收端设备可以从模型转化平台得到第一AI模型,或者是第一指示信息,或者是第二指示信息。
可选地,所述模型接收端设备接收第一响应消息,包括:
所述模型接收端设备从所述模型发送端设备获取所述第一响应消息,所述第一响应消息包括所述第一指示信息或所述第二指示信息。
模型接收端设备可以从模型发送端设备得到第一指示信息,或者是第二指示信息。
在本申请实施例中,模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息,接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,从而使得不同平台框架的AI模型传递互通成为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
可选地,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
所述模型转化平台的属性信息;
所述模型接收端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
在本申请实施例中,根据模型转化平台与模型接收端设备以及模型发送端设备之间的关系,模型接收端设备向模型发送端设备发送的第一请求消息所携带的信息,有所区别,包括以下几种情形:
情形一,模型转化平台在所述模型接收端设备的后方
需要说明的是,模型转化平台在模型接收端设备的后方意味着模型转化平台具有模型接收端设备的属性信息,包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息;以及,模型接收端设备具有模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台在所述模型接收端设备的后方,可为模型接收端设备节省硬盘空间。
可选地,在所述模型转化平台在所述模型接收端设备的后方的情况下,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
1)所述模型转化平台的属性信息;
其中,模型转化平台的属性信息包括模型转化平台的地址信息,标识信息,用于告知模型发送端设备应该将AI模型发送到哪里。模型转化平台的地址信息或标识信息是必须携带的。
模型转化平台的属性信息还可以包括转化能力信息,用于告知模型发送端设备模型转化平台所支持的AI框架类型,以及具备将这些AI框架类型转化为模型接收端设备所支持的AI框架类型的能力。
需要说明的是,模型转化平台应该可以支持所有主流框架,故转化能力信息是可选携带的。
2)所述模型接收端设备的属性信息;
模型接收端设备的属性信息包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型接收端设备的地址信息或标识信息用于告知将AI模型反馈给模型接收端设备,可以隐式发送,是可选携带的。
模型接收端设备支持或选用的AI框架信息,用于告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台模型接收端设备需要的AI模型的框架信息。
由于模型转化平台在所述模型接收端设备的后方,因此,模型接收端设备支持或选用的AI框架信息对于模型转化平台可以是已知的,即模型接收端设备支持或选用的AI框架信息是可选携带的。
3)模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
模型转化要求信息用于指示模型转化平台进行模型转化的要求信息,包括以下至少一项:模型转化时延要求信息;模型转化准确率要求信息。
情形二,模型转化平台在所述模型发送端设备的后方或者所述模型转化平台为公共转 化平台
需要说明的是,模型转化平台在模型发送端设备的后方意味着模型转化平台具有模型发送端设备的属性信息,包括模型发送端设备的地址信息,标识信息,和/或,支持或所选用的AI框架信息;以及,模型发送端设备具有模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台在所述模型发送端设备的后方,可为模型发送端设备节省硬盘空间。
模型转化平台为公共转化平台,意味着模型转化平台可以获得模型接收端设备的属性信息和/或模型发送端设备的属性信息。模型接收端设备和模型发送端设备也可以获得模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台为公共转化平台,可为模型发送端设备和模型接收端设备节省硬盘空间。
因此,在此情形下,模型接收端设备向模型发送端设备发送的所述第一请求消息可以不携带模型转化平台的属性信息。
可选地,在所述模型转化平台在所述模型发送端设备的后方或者所述模型转化平台为公共转化平台的情况下,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
1)所述模型接收端设备的属性信息;
模型接收端设备的属性信息包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型接收端设备的地址信息或标识信息用于告知将AI模型反馈给模型接收端设备,可以隐式发送,是可选携带的。
模型接收端设备支持或选用的AI框架信息,用于告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台模型接收端设备需要的AI模型的框架信息。
在模型转化平台在所述模型发送端设备的后方的情况下,模型接收端设备支持或选用的AI框架信息对于模型转化平台是未知的,则模型接收端设备支持或选用的AI框架信息是必须携带的。
在模型转化平台为公共转化平台的情况下,模型接收端设备可以将其支持或选用的AI框架信息携带在第一请求消息中,告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台,以指示模型转化平台模型接收端设备需要的AI模型的框架。
2)模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
模型转化要求信息用于指示模型转化平台进行模型转化的要求信息,包括以下至少一项:模型转化时延要求信息;模型转化准确率要求信息。
情形三,模型转化平台为公共转化平台
可选地,在所述模型转化平台为公共转化平台的情况下,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括:
所述模型接收端设备的地址和/或标识信息。
需要说明的是,在模型转化平台为公共转化平台的情况下,模型接收端设备向模型发送端设备发送的第一请求消息中携带AI模型描述信息和所述模型接收端设备的地址和/或标识信息。
在模型转化平台为公共转化平台的情况下,模型接收端设备支持或选用的AI框架信息可以通过如下两种实施方式告知给模型转化平台,以指示模型转化平台模型接收端设备需要的AI模型的框架:
一种实施方式中,所述方法还包括:
所述模型接收端设备向所述模型转化平台注册上报所述模型接收端设备的属性信息,所述属性信息包括所述模型接收端设备支持的AI框架信息。
即模型接收端设备通过注册上报的方式向所述模型转化平台发送所述模型接收端设备支持或选用的AI框架信息。
另一种实施方式中,所述方法还包括:
在接收到所述模型转化平台发送的第一AI框架请求消息的情况下,所述模型接收端设备向所述模型转化平台发送所述模型接收端设备支持的AI框架信息,其中,所述第一AI框架请求消息用于请求所述模型接收端设备支持的AI框架信息。
即模型转化平台可以向模型接收端设备发送第一AI框架请求消息,用于请求模型接收端设备支持或选用的AI框架信息。
在一些可选的实施例中,在模型转化平台在模型接收端设备的后方的情况下,模型接收端设备可以直接向模型转化平台发送第一请求消息,并经由模型转化平台根据所述第一请求消息,向模型发送端设备发送第二请求消息,以请求模型发送端设备提供模型接收端设备请求的AI模型。
可选地,在所述模型接收端设备向所述模型转化平台发送所述第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
1)所述模型接收端设备的属性信息;
模型接收端设备的属性信息包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型接收端设备的地址信息或标识信息用于告知模型转化平台将AI模型反馈给模型接收端设备,可以隐式发送,是可选携带的。
模型接收端设备支持或选用的AI框架信息,用于向模型转化平台模型指示接收端设备需要的AI模型的框架信息。
由于模型转化平台在所述模型接收端设备的后方,因此,模型接收端设备支持或选用的AI框架信息对于模型转化平台可以是已知的,即模型接收端设备支持或选用的AI框架信息是可选携带的。
2)所述模型发送端设备的属性信息;
模型发送端设备的属性信息包括模型发送端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型发送端设备的地址信息或标识信息,用于指示模型转化平台向指定的模型发送端设备发送第二请求消息,以请求模型发送端设备提供模型接收端设备请求的AI模型。
模型发送端设备的地址信息或标识信息是可选携带的。
若第一请求消息未包括模型发送端设备的属性信息,则模型转化平台会选择能够提供符合AI模型描述信息的模型的模型发送端设备,并向选择的模型发送端设备发送第二请求消息,以选择的请求模型发送端设备提供符合AI模型描述信息的模型。
模型发送端设备支持或选用的AI框架信息,用于向模型转化平台指示模型发送端设备所发送的模型的AI框架信息,从而模型转化平台可以根据该信息判断是否可以完成模型转化。
3)模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
可选地,所述模型转化平台的属性信息包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息包括以下至少一项:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
在前述实施例中,已经提及第一响应消息的内容可以是请求被接受,即模型接收端设备接收到第一AI模型,第一响应消息的内容也可以是请求被拒绝,即模型接收端设备无法得到第一AI模型,模型接收端设备会接收到第一指示信息或第二指示信息。
下面介绍模型接收端设备会接收到第一指示信息或第二指示信息的具体内容。
可选地,在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,所述模型接收端设备从所述模型转化平台获取所述第二指示信息。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
可选地,在满足第一条件的情况下,所述模型接收端设备从所述模型发送端设备获取所述第二指示信息;
其中,所述第一条件包括以下至少一项:
所述模型发送端设备根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
所述模型发送端设备接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
可选地,在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,所述模型接收端设备从所述模型发送端设备或模型转化平台获取所述第一指示信息。
在本申请实施例中,模型接收端设备发送第一请求消息,并接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,其中,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
图4为本申请实施例提供的AI模型传输方法的流程示意图之二。如图4所示,该AI模型传输方法应用于模型发送端设备,该方法包括以下步骤:
步骤400、模型发送端设备接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型;
需要说明的是,AI模型请求消息可以来自模型接收端设备,也可以来自模型转化平台。
所述AI模型请求消息包括AI模型描述信息,该AI模型描述信息用于指示所请求的AI模型的信息,即该AI模型描述信息用于描述第一AI模型。其中,第一AI模型对应第一模型表示方法。
可选地,所述AI模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
需要说明的是,本实施例应用于模型发送端设备侧,描述的是前述模型接收端设备侧方法实施例的对端动作,为了简化描述,关于本实施例中与前述模型接收端设备侧方法实施例中的相同内容可以参考前述模型接收端设备侧方法实施例,在此不再赘述。
步骤401、模型发送端设备根据所述AI模型请求消息,执行第一操作;
模型发送端设备接收到AI模型请求消息后,根据所述AI模型请求消息,执行第一操作
其中,所述执行第一操作包括以下至少一项:
1)选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中,所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
可以理解的是,模型发送端设备根据AI模型请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。
例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。
值得说明的是,模型训练的方法在本申请中不做限定。
若模型发送端设备能够选择或训练生成第二AI模型,则模型发送端设备会向模型转化平台发送第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用。
2)发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
在本申请实施例中,模型发送端设备接收AI模型请求消息,根据所述AI模型请求消息,执行第一操作,包括选择或训练生成第二AI模型,向模型转化平台发送第二AI模型,或者,发送第一指示信息或第二指示信息,从而使得不同平台框架的AI模型传递互通成 为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
所述AI模型请求消息可以是第二请求消息,或者第一请求消息。
可选地,所述模型发送端设备接收AI模型请求消息,包括:
所述模型发送端设备接收所述模型转化平台发送的第二请求消息,其中,所述第二请求信息是所述模型转化平台根据所述模型接收端设备发送给所述模型转化平台的第一请求消息确定;
或者,
所述模型发送端设备接收所述模型接收端设备发送的第一请求消息。
可选地,所述在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
所述模型发送端设备向所述模型转化平台发送第二响应消息,所述第二响应消息包括所述第二AI模型。
可以理解的是,在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,若模型发送端设备能够选择或训练生成第二AI模型,则模型发送端设备会向模型转化平台发送第二响应消息,第二响应消息携带所述第二AI模型,此处,第二AI模型包括模型发送端设备所选择或生成的AI模型的完整的网络结构和参数信息。
可选地,所述第二响应消息还包括以下至少一项:
所述第二AI模型的框架信息;
所述模型接收端设备的属性信息。
其中,第二AI模型的框架信息,用于告知模型转化平台该第二AI模型所用框架,方便模型转化平台进行转化。
例如,AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。
如果模型转化平台可以自己判断,则所述第二AI模型的框架信息可以不需要携带。即所述第二AI模型的框架信息是可选的。
所述模型接收端设备的属性信息包括:模型接收端设备的标识信息、地址信息和/或,所支持或选用的AI框架信息。
其中,模型接收端设备的标识信息或地址信息用于告知模型转化平台应该将转化后的模型发送到哪里,如模型接收端设备的IP地址或FQDN等标识信息。
模型接收端设备所支持或选用的AI框架信息,用于指示模型转化平台应该将模型发送端设备发送的第二AI模型转化为何种AI框架的模型。
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
所述模型发送端设备向所述模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;所述第三请求消息包括所述第 二AI模型。
可以理解,模型发送端设备接收所述模型接收端设备发送的第一请求消息之后,选择或训练生成第二AI模型,并向模型转化平台发送第三请求消息,以请求模型转化平台将第二AI模型转化为第一AI模型。
可选地,在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,所述第二请求消息还包括以下至少一项:
所述模型转化平台的属性信息。
可以理解的是,在模型接收端设备向模型转化平台发送第一请求消息的情况下,该第一请求消息包括:AI模型描述信息,以及所述模型接收端设备的属性信息、所述模型发送端设备的属性信息和模型转化要求信息中的至少一项。
模型转化平台根据该第一请求消息,生成第二请求消息,并向模型发送端设备发送第二请求消息,该第二请求消息包括AI模型描述信息以及模型转化平台的属性信息,此步骤适用于模型转化平台在模型接收端设备的后方。
其中,模型转化平台的属性信息包括模型转化平台的地址信息,标识信息,用于告知模型发送端设备应该将AI模型发送到哪里,b)如转化平台的IP地址、FQDN等。由于是转化平台先向发送端发送信息,故模型转化平台的地址信息或标识信息是可选携带的。
模型转化平台的属性信息还可以包括转化能力信息,用于告知模型发送端设备模型转化平台所支持的AI框架类型,以及具备将这些AI框架类型转化为模型接收端设备所支持的AI框架类型的能力。转化平台应该可以支持所有主流框架,故转化能力信息是可选携带的。
在本申请实施例中,根据模型转化平台与模型接收端设备以及模型发送端设备之间的关系,模型发送端设备接收的来自模型接收端设备的第一请求消息所携带的信息,有所区别,包括以下几种情形:
情形一:模型转化平台在模型接收端设备的后方
需要说明的是,模型转化平台在模型接收端设备的后方意味着模型转化平台具有模型接收端设备的属性信息,包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息;以及,模型接收端设备具有模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台在所述模型接收端设备的后方,可为模型接收端设备节省硬盘空间。
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息,且所述模型转化平台在所述模型接收端设备的后方的情况下,所述第一请求消息还包括以下至少一项:
1)所述模型转化平台的属性信息;
其中,模型转化平台的属性信息包括模型转化平台的地址信息,标识信息,用于告知模型发送端设备应该将AI模型发送到哪里。模型转化平台的地址信息或标识信息是必须 携带的。
模型转化平台的属性信息还可以包括转化能力信息,用于告知模型发送端设备模型转化平台所支持的AI框架类型,以及具备将这些AI框架类型转化为模型接收端设备所支持的AI框架类型的能力。
需要说明的是,模型转化平台应该可以支持所有主流框架,故转化能力信息是可选携带的。
2)所述模型接收端设备的属性信息;
模型接收端设备的属性信息包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型接收端设备的地址信息或标识信息用于告知将AI模型反馈给模型接收端设备,可以隐式发送,是可选携带的。
模型接收端设备支持或选用的AI框架信息,用于告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台模型接收端设备需要的AI模型的框架信息。
由于模型转化平台在所述模型接收端设备的后方,因此,模型接收端设备支持或选用的AI框架信息对于模型转化平台可以是已知的,即模型接收端设备支持或选用的AI框架信息是可选携带的。
3)模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
模型转化要求信息用于指示模型转化平台进行模型转化的要求信息,包括以下至少一项:模型转化时延要求信息;模型转化准确率要求信息。
情形二:模型转化平台在模型发送端设备的后方或者模型转化平台为公共转化平台
需要说明的是,模型转化平台在模型发送端设备的后方意味着模型转化平台具有模型发送端设备的属性信息,包括模型发送端设备的地址信息,标识信息,和/或,支持或所选用的AI框架信息;以及,模型发送端设备具有模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台在所述模型发送端设备的后方,可为模型发送端设备节省硬盘空间。
模型转化平台为公共转化平台,意味着模型转化平台可以获得模型接收端设备的属性信息和/或模型发送端设备的属性信息。模型接收端设备和模型发送端设备也可以获得模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
模型转化平台为公共转化平台,可为模型发送端设备和模型接收端设备节省硬盘空间。
因此,在此情形下,模型发送端设备接收的由模型接收端设备发送的所述第一请求消息可以不携带模型转化平台的属性信息。
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息,且所述模型转化平台在所述模型发送端设备的后方或者所述模型转化平台为公共转化平台的情况下,所述第一请求消息还包括以下至少一项:
1)所述模型接收端设备的属性信息;
模型接收端设备的属性信息包括模型接收端设备的地址信息,标识信息,和/或,支持或选用的AI框架信息。
其中,模型接收端设备的地址信息或标识信息用于告知将AI模型反馈给模型接收端设备,可以隐式发送,是可选携带的。
模型接收端设备支持或选用的AI框架信息,用于告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台模型接收端设备需要的AI模型的框架信息。
在模型转化平台在所述模型发送端设备的后方的情况下,模型接收端设备支持或选用的AI框架信息对于模型转化平台是未知的,则模型接收端设备支持或选用的AI框架信息是必须携带的。
在模型转化平台为公共转化平台的情况下,模型接收端设备可以将其支持或选用的AI框架信息携带在第一请求消息中,告知给模型发送端设备,再经由模型发送端设备告知给模型转化平台,以指示模型转化平台模型接收端设备需要的AI模型的框架。
2)模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
模型转化要求信息用于指示模型转化平台进行模型转化的要求信息,包括以下至少一项:模型转化时延要求信息;模型转化准确率要求信息。
情形三:模型转化平台为公共转化平台
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息,且所述模型转化平台为公共转化平台的情况下,所述第一请求消息还包括:
所述模型接收端设备的地址和/或标识信息。
需要说明的是,在模型转化平台为公共转化平台的情况下,模型发送端设备接收所述模型接收端设备发送的第一请求消息中携带AI模型描述信息和所述模型接收端设备的地址和/或标识信息。
在模型转化平台为公共转化平台的情况下,模型发送端设备支持或选用的AI框架信息可以通过如下两种实施方式告知给模型转化平台,以指示模型转化平台模型发送端设备提供的AI模型的框架:
一种实施方式中,所述方法还包括:
所述模型发送端设备向所述模型转化平台注册上报所述模型发送端设备的属性信息,所述属性信息包括所述模型发送端设备支持的AI框架信息。
即模型发送端设备通过注册上报的方式向所述模型转化平台发送所述模型发送端设备支持或选用的AI框架信息。
另一种实施方式中,所述方法还包括:
在接收所述模型转化平台发送的第二AI框架请求消息的情况下,所述模型发送端设备向所述模型转化平台发送所述模型发送端设备支持的AI框架信息,其中,所述第二AI 框架请求消息用于请求所述模型发送端设备支持的AI框架信息。
即模型转化平台可以向模型发送端设备发送第二AI框架请求消息,用于请求模型发送端设备反馈支持或选用的AI框架信息。
可选地,所述模型转化平台的属性信息包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
在前述实施例中,已经提及模型发送端设备可以向模型接收端设备发送第一指示信息或第二指示信息。
下面介绍模型发送端设备发送第一指示信息或第二指示信息的具体内容。
可选地,所述发送第一指示信息,包括:
在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型接收端设备发送所述第一指示信息。
可以理解的是,在模型发送端设备接收到模型接收端设备发送的第一请求消息后,若模型发送端设备本地没有符合AI模型描述信息的AI模型,并且无法生成符合AI模型描述信息的AI模型,则模型发送端设备向模型接收端设备发送第一指示信息。
可选地,所述方法还包括:
在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型转化平台发送第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
可选理解的是,在模型发送端设备接收到模型转化平台发送的第二请求消息后,如果发送端设备没有或无法生成模型转化平台所指定或者支持的AI模型,那么模型转化平台 就无法进行转化,那么模型发送端设备会向转化平台做出“请求不接受”的响应,即模型发送端设备向所述模型转化平台发送第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
可选地,所述方法还包括:
根据所述模型转化平台的转化能力信息确定所述模型转化平台是否可以完成对所述第二AI模型的转化操作。
可选地,模型发送端设备还可以根据模型转化平台的转化能力信息确定模型转化平台是否可以完成对所述第二AI模型的转化操作,若确定模型转化平台无法完成对所述第二AI模型的转化操作,则向模型接收端设备发送第二指示信息。
可选地,所述方法还包括:
接收所述模型转化平台发送的第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
可以理解的是,在模型发送端设备选择或训练生成第二AI模型后,将该第二AI模型发送给模型转化平台进行转化,模型转化平台可以在模型转化成功之后后,向模型发送端设备发送第一模型转化响应消息,指示是否完成了对第二AI模型的转化操作。
或者,模型转化平台也可以在模型转化之前,根据转化能力信息,确定是否可以完成对第二AI模型的转化操作,并向模型发送端设备发送第一模型转化响应消息,指示是否能够完成对第二AI模型的转化操作。
可选地,所述发送第二指示信息,包括:
在满足第一条件的情况下,向所述模型接收端设备发送第二指示信息;
其中,所述第一条件包括以下至少一项:
根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
其中,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
在本申请实施例中,模型发送端设备接收AI模型请求消息,根据所述AI模型请求消息,执行第一操作,包括选择或训练生成第二AI模型,向模型转化平台发送第二AI模型,或者,发送第一指示信息或第二指示信息,第二AI模型经过模型转化平台转化后得到的、 能够被模型接收端设备使用的模型,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
图5为本申请实施例提供的AI模型传输方法的流程示意图之三。如图5所示,该方法应用于模型转化平台,该方法包括以下步骤:
步骤500、模型转化平台接收模型发送端设备发送的第二AI模型;
步骤501、所述模型转化平台将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;
或者,模型转化平台向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
需要说明的是,本实施例应用于模型转化平台侧,描述的是前述模型接收端设备侧和模型发送端设备侧的方法实施例的对端动作,为了简化描述,对于本实施例中与前述模型接收端设备侧方法实施例和模型发送端设备侧方法实施例中的相同内容可以参考前述模型接收端设备侧方法实施例和模型发送端设备侧方法实施例,在此不再赘述。
可以理解的是,若模型发送端设备提供了第二AI模型给模型转化平台,模型转化平台则将该第二AI模型转化为能够被模型接收端设备使用的第一AI模型,并将第一AI模型发送至模型接收端设备。
若模型转化平台无法将第二AI模型转换成第一AI模型,则模型转化平台向模型接收端设备发送第二指示信息。
若模型转化平台获知模型发送端设备无法提供所述第二AI模型,则模型转化平台向模型接收端设备发送第一指示信息。
可选地,所述模型转化平台接收模型发送端设备发送的第二AI模型,包括:
所述模型转化平台接收模型发送端设备发送的第二响应消息,所述第二响应消息包括所述第二AI模型。
可选地,所述第二响应消息还包括以下至少一项:
所述模型发送端设备支持的AI框架信息;
所述模型接收端设备的属性信息。
其中,模型发送端设备支持的AI框架信息也是第二AI模型的框架信息,用于告知模型转化平台该第二AI模型所用框架,方便模型转化平台进行转化。
例如,AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3” (即框架平台+版本号的形式)。
如果模型转化平台可以自己判断,则所述第二AI模型的框架信息可以不需要携带。即所述第二AI模型的框架信息是可选的。
所述模型接收端设备的属性信息包括:模型接收端设备的标识信息、地址信息和/或,所支持或选用的AI框架信息。
其中,模型接收端设备的标识信息或地址信息用于告知模型转化平台应该将转化后的模型发送到哪里,如模型接收端设备的IP地址或FQDN等标识信息。
模型接收端设备所支持或选用的AI框架信息,用于指示模型转化平台应该将模型发送端设备发送的第二AI模型转化为何种AI框架的模型。
可选地,所述模型转化平台接收模型发送端设备发送的第二AI模型,包括:
所述模型转化平台接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型。
在本申请实施例中,模型转化平台接收模型发送端设备发送的第二AI模型;将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备,或者向模型接收端设备发送第一指示信息或第二指示信息,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
可选地,所述方法还包括以下至少一项:
所述模型转化平台接收所述模型发送端设备上报的所述模型发送端设备的属性信息,所述模型发送端设备的属性信息包括所述模型发送端设备支持的AI框架信息;
所述模型转化平台接收所述模型接收端设备上报的所述模型接收端设备的属性信息,所述模型接收端设备的属性信息包括所述模型接收端设备支持的AI框架信息。
在本实施例中,所述模型转化平台为公共转化平台。
模型转化平台为公共转化平台,意味着模型转化平台可以获得模型接收端设备的属性信息和/或模型发送端设备的属性信息。模型接收端设备和模型发送端设备也可以获得模型转化平台的属性信息,包括模型转化平台的地址信息,标识信息,和/或转化能力信息。
即模型接收端设备通过注册上报的方式向所述模型转化平台发送所述模型接收端设备支持或选用的AI框架信息。模型发送端设备通过注册上报的方式向所述模型转化平台发送所述模型接收端设备支持或选用的AI框架信息。
可选地,所述方法还包括以下至少一项:
向模型接收端设备发送第一AI框架请求消息,接收所述模型接收端设备支持的AI框架信息;
向模型发送端设备发送第二AI框架请求消息,接收所述模型发送端设备支持的AI框架信息。
即模型转化平台可以向模型接收端设备发送第一AI框架请求消息,用于请求模型接收端设备支持或选用的AI框架信息。
模型转化平台可以向模型发送端设备发送第二AI框架请求消息,用于请求模型发送端设备支持或选用的AI框架信息。
可选地,所述方法还包括:
所述模型转化平台接收所述模型接收端设备发送的第一请求消息;
所述模型转化平台向模型发送端设备发送第二请求消息;
其中,所述第二请求信息根据所述第一请求消息确定。
在一些可选的实施例中,在模型转化平台在模型接收端设备的后方的情况下,模型接收端设备可以直接向模型转化平台发送第一请求消息,并模型转化平台接收所述模型接收端设备发送的第一请求消息之后,向模型发送端设备发送第二请求消息,以请求模型发送端设备提供模型接收端设备请求的AI模型。
可选地,所述第一请求消息或第二请求消息包括:AI模型描述信息,所述模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
可选地,所述第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
所述模型发送端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息;
所述第二请求消息还包括以下至少一项:
所述模型转化平台的属性信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息包括以下至少一项:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述 第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,所述模型转化平台的属性信息,包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型转化平台向模型发送端设备发送第二请求消息之前,所述方法还包括:
所述模型转化平台确定所述模型转化平台具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力。
可以理解的是,模型转化平台接收到模型接收端设备发送的第一请求消息之后,根据自身的转化能力信息、模型发送端设备支持的AI框架信息,以及模型接收端设备支持的AI框架信息,确定是否具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力。
若模型转化平台确定自身具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力,则模型转化平台向模型发送端设备发送第二请求消息。
可选地,所述模型转化平台接收所述模型接收端设备发送的第一请求消息之后,所述方法还包括:
所述模型转化平台确定所述模型转化平台不具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力,所述模型转化平台向所述模型接收端设备发送第三响应消息,所述第三响应消息用于指示所述第一请求消息被拒绝。
可选地,在所述第一请求消息中未携带所述模型发送端设备的属性信息的情况下,所述模型转化平台向模型发送端设备发送第二请求消息之前,所述方法还包括:
所述模型转化平台确定所述模型发送端设备。
需要说明的是,模型转化平台接收到模型接收端设备发送的第一请求消息之后,若述第一请求消息中未携带所述模型发送端设备的属性信息,则模型转化平台会根据传输速度,服务器位置,线路占有率等因素自动选择能够提供符合第一请求消息中的AI模型描述信息的AI模型的模型发送端设备。或者,模型转化平台可以根据之前的交互传递记录,自动选择更符合模型接收端设备期望的模型发送端设备。
可选地,所述模型转化平台接收模型发送端设备发送的第二AI模型之后,所述方法还包括:
所述模型转化平台向模型发送端设备发送第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
可选地,所述模型转化平台向所述模型接收端设备发送第二指示信息,包括;
在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,所述模型转化平台向所述模型接收端设备发送所述第二指示信息。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括以下至少一项:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型提供时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确度不能满足所述模型接收端设备请求的模型准确度要求。
可选地,所述模型转化平台向所述模型接收端设备发送第一指示信息之前,所述方法还包括:
接收模型发送端设备发送的第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
可选理解的是,在模型发送端设备接收到模型转化平台发送的第二请求消息后,如果发送端设备没有或无法生成模型转化平台所指定或者支持的AI模型,那么模型转化平台就无法进行转化,那么模型发送端设备会向模型转化平台做出“请求不接受”的响应,即模型发送端设备向所述模型转化平台发送第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。模型转化平台接收模型发送端设备发送的第四响应消息之后,向所述模型接收端设备发送第一指示信息。
需要说明的是,模型转化平台可以部署在所述模型接收端设备的后方,模型转化平台也可部署在所述模型发送端设备的后方,或者所述模型转化平台为公共转化平台。
若模型转化平台部署在所述模型接收端设备的后方,可以节省模型接收端设备的硬件成本。若模型转化平台部署在所述模型发送端设备的后方,可以节省模型发送端设备的硬件成本。若模型转化平台为公共转化平台,可为模型发送端设备和模型接收端设备节省硬盘空间。
在本申请实施例中,模型转化平台接收模型发送端设备发送的第二AI模型;将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备,或者向模型接收端设备发送第一指示信息或第二指示信息,其中,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
需要说明的是,下述如图6~图12所示实施例中的转化平台即模型转化平台,发送端 即模型发送端设备,接收端即模型接收端设备。
图6为本申请实施例提供的AI模型传输方法的交互流程示意图之一。
步骤1:
模型接收端设备向模型发送端设备发送人工智能模型(AI model)请求消息,包含信息如下至少一项:
a)模型的描述信息:用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
c)转化平台地址或标识信息:告知模型发送端设备该把模型发送到哪里,如转化平台的IP地址、FQDN等。(必选)
d)转化平台能力信息:告知模型发送端设备转化平台支持AI框架类型,以及具备将这些AI框架类型转化模型接收端设备所支持的AI框架类型的能力。(转化平台应该可以支持所有主流框架,故“可选”)
e)模型接收端设备AI框架信息:告知模型发送端设备,再经由模型发送端设备告诉转化平台,模型接收端设备自己需要什么样的框架。(因为转化平台是属于模型接收端设备的,所以该信息转化平台应该有,故“可选”)
f)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1的AI模型请求消息,模型发送端设备向转化平台发送所请求的模型信息。
首先,模型发送端设备根据模型接收端设备的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
a)AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。(必选)。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
b)AI模型框架信息,告知转化平台该AI模型文件所用框架,方便转化平台进行转化。(如果转化平台可以自己判断,可为“可选”)。如AI模型文件为TesnorFLow所生成 的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。
c)模型接收端设备信息:告知转化平台该把转化后模型发送到哪里,如模型接收端设备IP地址或FQDN等标识信息(可选);
d)模型接收端设备框架信息:告知转化平台该把模型转化成哪种框架的模型。(可选)。例如需将TensorFLow生成的模型文件转化成PyTorch可识别的模型文件,则这里可以是“PyTorch”+“1.10.0”(框架平台+版本号)的形式。
步骤3:
根据步骤2接收到的模型信息,转化平台向模型发送端设备发送响应信息:
a)若转化平台能够将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化被接受或成功;
b)若转化平台无法将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化失败。其中还可以进一步包括失败的原因值,例如因为不支持AI模型从框架A模型转换成框架B模型。
步骤4:
根据步骤3接收到的响应信息,模型发送端设备给模型接收端设备发送请求响应信息(可选):
a)请求接受;如当接受到“转化被接受或成功”时(步骤3.a),告知接受端其请求合理可行并已被接受。
b)请求拒绝。如当接收到“转化失败”时(步骤3.b),告知模型发送端设备其请求被拒绝。其中也可以根据“步骤3.b”的反馈,进一步告知其失败的原因,例如转化平台暂不支持从AI框架A转化为AI框架B。或者,因为模型发送端设备的原因,拒绝模型接收端设备的请求,如该模型发送端设备无法生成所需AI模型。
需要说明的是,该步骤4与下述步骤5不限定时间上的先后顺序。(可以发两次,可能有异常状况,转发不成功)
步骤5:
转化平台将AI模型转化成模型接收端设备AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。例如,通过ONNX提供的Python脚本,转化平台(“tf2onnx.conver”)可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息,生产者信息外,还有计算图信息等。计算图Graph(GraphProto类型)包含了网络结构和参数等信息,由一些名称等基本信息和四组列表组成。这四组数组为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型输入节点的名称,类型,形状等除数据以外的信息,output存放了模型输出节点名称,类型,形状等除数据以外的 信息,initializer存放了模型的所有网络参数的具体数值,包括超参数和输入值。具体地,每个Node中包含它的运算操作类型和指定的inputs和outputs名字等信息。一个Node的节点信息,可以是一整层的计算过程,也可以是一个节点的计算。所有node连接在一起组成graph,通过Node计算节点中的inputs和outputs两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤6:
转化平台将转化后的AI模型发送给模型接收端设备。
具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,还可包含模型发送端设备AI框架信息(如模型发送端设备支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
图7为本申请实施例提供的AI模型传输方法的交互流程示意图之二。如图7所示,该AI模型传输方法包括以下步骤:
步骤1:
模型接收端设备向转化平台发送AI模型的请求消息,包含信息如下至少一项:
a)模型的描述信息:可以表明指定用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)模型接收端设备信息:方便转化平台发送结果到包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
c)模型接收端设备AI框架信息:告知转化平台该把模型转化成哪种什么框架的模型。如需将TensorFLow生成的模型文件转化成PyTorch可识别的模型文件,则这里可以是“PyTorch”+“1.10.0”(框架平台+版本号)的形式。(可选,因为该转化平台属于模型接收端设备,所以应该知晓模型接收端设备的框架信息。)
d)模型发送端设备信息:让转化平台向指定模型发送端设备发送模型请求,包含模型发送端设备的地址信息或者标识信息;如果不发送该信息,则转化平台会自动选择符合“a.模型发送信息”的模型发送端设备并请求模型。(可选)
e)模型发送端设备AI框架信息:告知转化平台模型发送端设备所发送的模型的框架平台,让转化平台判断是否可以完成转化,如不能转化会触发步骤1a。(可选)
f)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤1a(与步骤5二选一):
根据步骤1的请求消息,转化平台判断能否进行转化,并做出响应。
例如转化平台无法从模型接收端设备指定的“模型发送端设备AI框架”转化为模型接收端设备AI平台框架,如转化平台暂不兼容该框架,转化平台任务过多超负荷,或不满足步骤1e中的模型转化要求信息,则会返回“请求拒绝”。
步骤1b(可选):
根据步骤1的请求消息,转化平台判断是否要执行此步骤,自动选择模型发送端设备
如果步骤1的请求消息中已经指定“模型发送端设备信息”,则不执行这一步,转化平台会直接向指定模型发送端设备请求模型。
如果步骤1的请求消息中未指定“模型发送端设备信息”,则转化平台会根据传输速度,服务器位置,线路占有率等因素自动选择有符合步骤1中模型接收端设备的“模型描述信息”AI模型的模型发送端设备。(或者,可以根据之前的交互传递记录,自动选择更符合模型接收端设备期望的模型发送端设备)
步骤2:
根据步骤1的请求消息,转化平台向模型发送端设备(1b自动选择的或者1中指定的)发送模型请求,包含信息如下至少一项:
a)模型的描述信息:可以表明指定用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)转化平台信息:包括转化平台地址或标识信息,告知模型发送端设备该把模型发送到哪里,如转化平台的IP地址、FQDN等。(由于是转化平台先向模型发送端设备发送信息,故可选)
c)转化平台能力信息:告知模型发送端设备转化平台支持的AI框架类型,以及具备将这些AI框架类型转化模型接收端设备所支持的AI框架类型的能力。(转化平台应该可以支持所有主流框架,故“可选”)
步骤3(与步骤3a二选一):
首先,模型发送端设备根据转化平台的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
a)所生成的模型:AI模型文件,包含所生成AI模型的完整的网络结构和参数信息 等元素的文件。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。(必选)
b)模型发送端设备所使用的框架:告知转化平台该AI模型文件所用框架,方便转化平台进行转化。如AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。(如果转化平台可以自己判断,可为“可选”)
步骤3a(与步骤3二选一):
根据步骤2的请求信息,模型发送端设备向转化平台做出响应的一种;
如果模型发送端设备没有或无法生成转化平台所指定或者支持的AI模型,那么转化平台就无法进行转化,则模型发送端设备会向转化平台做出“请求不接受”的响应。
步骤4:
如果步骤3执行,即有AI模型成功从模型发送端设备发送到转化平台,则转化平台将AI模型转化成接受端AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。通过ONNX提供的Python脚本,“tf2onnx.conver”可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息、生产者信息外,还有四组数组,它们为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型的输入节点,output存放了模型中所有的输出节点,initializer存放了模型的所有权重参数。通过Node计算节点中的input和output两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤5(与步骤1a二选一):
转化平台将转化后的AI模型下发给模型接收端设备,具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,并还可包含模型发送端设备AI框架信息(如模型发送端设备IP地址等支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
图8为本申请实施例提供的AI模型传输方法的交互流程示意图之三。如图8所示,该AI模型传输方法包括以下步骤:
步骤1:
模型接收端设备向模型发送端设备发送人工智能模型(AI model)请求消息,包含信息如下至少一项:
a)模型的描述信息:用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应 的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
c)模型接收端设备AI框架信息:告知模型发送端设备,再经由模型发送端设备告诉转化平台,模型接收端设备自己需要什么样的框架。(因为转化平台是属于模型发送端设备的,所以该信息转化平台没有,故必选)
d)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1的AI模型请求消息,模型发送端设备向转化平台发送所请求的模型信息。
首先,模型发送端设备根据模型接收端设备的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
a)AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。(必选)。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
b)AI模型框架信息,告知转化平台该AI模型文件所用框架,方便转化平台进行转化。(如果转化平台可以自己判断,可为“可选”)。如AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。
c)模型接收端设备信息:告知转化平台该把转化后模型发送到哪里,如模型接收端设备IP地址或FQDN等标识信息(可选);
d)模型接收端设备框架信息:告知转化平台该把模型转化成哪种框架的模型。(可选)。例如需将TensorFLow生成的模型文件转化成PyTorch可识别的模型文件,则这里可以是“PyTorch”+“1.10.0”(框架平台+版本号)的形式。
步骤3:
根据步骤2接收到的模型信息,转化平台向模型发送端设备发送响应信息:
a)若转化平台能够将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化被接受或成功;
b)若转化平台无法将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化失败。其中还可以进一步包括失败的原因 值,例如因为不支持AI模型从框架A模型转换成框架B模型。
步骤4:
根据步骤3接收到的响应信息,模型发送端设备给模型接收端设备发送请求响应信息(可选):
a)请求接受;如当接受到“转化被接受或成功”时(步骤3.a),告知接受端其请求合理可行并已被接受。
b)请求拒绝。如当接收到“转化失败”时(步骤3.b),告知模型发送端设备其请求被拒绝。其中也可以根据“步骤3.b”的反馈,进一步告知其失败的原因,例如转化平台暂不支持从AI框架A转化为AI框架B。或者,因为模型发送端设备的原因,拒绝模型接收端设备的请求,如该模型发送端设备无法生成所需AI模型。
需要说明的是,该步骤4与下述步骤5不限定时间上的先后顺序。(可以发两次,可能有异常状况,转发不成功)。
步骤5:
转化平台将AI模型转化成模型接收端设备AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。例如,通过ONNX提供的Python脚本,转化平台(“tf2onnx.conver”)可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息,生产者信息外,还有计算图信息等。计算图Graph(GraphProto类型)包含了网络结构和参数等信息,由一些名称等基本信息和四组列表组成。这四组数组为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型输入节点的名称,类型,形状等除数据以外的信息,output存放了模型输出节点名称,类型,形状等除数据以外的信息,initializer存放了模型的所有网络参数的具体数值,包括超参数和输入值。具体地,每个Node中包含它的运算操作类型和指定的inputs和outputs名字等信息。一个Node的节点信息,可以是一整层的计算过程,也可以是一个节点的计算。所有node连接在一起组成graph,通过Node计算节点中的inputs和outputs两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤6:
转化平台将转化后的AI模型发送给模型接收端设备。
具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,还可包含模型发送端设备AI框架信息(如模型发送端设备支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
图9为本申请实施例提供的AI模型传输方法的交互流程示意图之四。如图9所示,该AI模型传输方法包括以下步骤:
步骤1:
模型接收端设备向模型发送端设备发送人工智能模型(AI model)请求消息,包含信息如下至少一项:
a)模型的描述信息:用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
c)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1的AI模型请求消息,模型发送端设备向转化平台发送所请求的模型信息。
首先,模型发送端设备根据模型接收端设备的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
a)AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。(必选)。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
b)AI模型框架信息,告知转化平台该AI模型文件所用框架,方便转化平台进行转化。(如果转化平台可以自己判断,可为“可选”)。如AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。
c)模型接收端设备信息:告知转化平台该把转化后模型发送到哪里,如模型接收端设备IP地址或FQDN等标识信息;
d)模型发送端设备信息:包括模型发送端设备的地址信息或标识信息,可以隐式发送,用于寻找对应框架。
步骤3a:
根据步骤1、步骤2接收到的模型接收端设备、模型发送端设备信息,转化平台获取 双方的可支持框架信息。
步骤3b:
转化平台将AI模型转化成模型接收端设备AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。例如,通过ONNX提供的Python脚本,转化平台(“tf2onnx.conver”)可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息,生产者信息外,还有计算图信息等。计算图Graph(GraphProto类型)包含了网络结构和参数等信息,由一些名称等基本信息和四组列表组成。这四组数组为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型输入节点的名称,类型,形状等除数据以外的信息,output存放了模型输出节点名称,类型,形状等除数据以外的信息,initializer存放了模型的所有网络参数的具体数值,包括超参数和输入值。具体地,每个Node中包含它的运算操作类型和指定的inputs和outputs名字等信息。一个Node的节点信息,可以是一整层的计算过程,也可以是一个节点的计算。所有node连接在一起组成graph,通过Node计算节点中的inputs和outputs两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤4a:
根据步骤2接收到的模型信息,转化平台向模型发送端设备发送响应信息:
a)若转化平台能够将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化被接受或成功;
b)若转化平台无法将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化失败。其中还可以进一步包括失败的原因值,例如因为不支持AI模型从框架A模型转换成框架B模型。
步骤4b:
根据步骤3接收到的响应信息,模型发送端设备给模型接收端设备发送请求响应信息(可选):
a)请求接受;如当接受到“转化被接受或成功”时,告知接受端其请求合理可行并已被接受。
b)请求拒绝。如当接收到“转化失败”时,告知模型发送端设备其请求被拒绝。其中也可以根据“步骤3.b”的反馈,进一步告知其失败的原因,例如转化平台暂不支持从AI框架A转化为AI框架B。或者,因为模型发送端设备的原因,拒绝模型接收端设备的请求,如该模型发送端设备无法生成所需AI模型。
需要说明的是,该步骤4与下述步骤5不限定时间上的先后顺序。(可以发两次,可能有异常状况,转发不成功)
步骤5:
转化平台将转化后的AI模型发送给模型接收端设备。
具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,还可包含模型发送端设备AI框架信息(如模型发送端设备支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
图10为本申请实施例提供的向公共转化平台注册上报能力信息的示意图。
如图10所示,在步骤1之前,还可以包括:
步骤01,模型接收端设备向公共转化平台注册上报能力信息,包含信息如下:
a)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
b)模型接收端设备AI框架信息:告诉转化平台,模型接收端设备自己需要什么样的框架。
步骤02,
模型发送端设备向公共转化平台注册上报能力信息,包含信息如下:
a)模型发送端设备信息:包括模型发送端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
b)模型发送端设备AI框架信息:告诉转化平台,模型发送端设备自己需要什么样的框架。
步骤03:
公共转化平台保存模型接收端设备和模型发送端设备的能力信息。
图11为本申请实施例提供的AI模型传输方法的交互流程示意图之五。如图11所示,该AI模型传输方法包括以下步骤:
步骤1:
模型接收端设备向模型发送端设备发送人工智能模型(AI model)请求消息,包含信息如下至少一项:
a)模型的描述信息:用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
b)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
c)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1的AI模型请求消息,模型发送端设备向转化平台发送所请求的模型信息。
首先,模型发送端设备根据模型接收端设备的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
a)AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。(必选)。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
b)模型接收端设备信息:告知转化平台该把转化后模型发送到哪里,如模型接收端设备IP地址或FQDN等标识信息;
步骤3:
根据步骤2接收到的模型信息,转化平台向模型接收端设备和模型发送端设备发送响应信息,获取模型框架信息:
3a)向模型接收端设备请求获取模型接收端设备所支持的框架信息,并接受模型接收端设备框架信息。
3b)向模型发送端设备请求获取模型接收端设备所支持的框架信息,并接受模型发送端设备框架信息。
步骤4:
转化平台将AI模型转化成模型接收端设备AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。例如,通过ONNX提供的Python脚本,转化平台(“tf2onnx.conver”)可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息,生产者信息外,还有计算图信息等。计算图Graph(GraphProto类型)包含了网络结构和参数等信息,由一些名称等基本信息和四组列表组成。这四组数组为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型输入节点的名称,类型,形状等除数据以外的信息,output存放了模型输出节点名称,类型,形状等除数据以外的信息,initializer存放了模型的所有网络参数的具体数值,包括超参数和输入值。具体地,每个Node中包含它的运算操作类型和指定的inputs和outputs名字等信息。一个Node的节点信息,可以是一整层的计算过程,也可以是一个节点的计算。所有node连接在一起组成graph,通过Node计算节点中的inputs和outputs两个数组指向输入和输出的节点关 系,构建出整个网络的拓扑结构,也就是网络结构。
步骤5:
转化平台将转化后的AI模型发送给模型接收端设备。
具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,还可包含模型发送端设备AI框架信息(如模型发送端设备支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
图12为本申请实施例提供的AI模型传输方法的交互流程示意图之六。如图12所示,该AI模型传输方法包括以下步骤:
步骤1:
模型接收端设备向模型发送端设备发送人工智能模型(AI model)请求消息,包含信息如下至少一项:
c)模型的描述信息:用于指示所请求模型的信息,如模型类型或ID信息(如analytic ID,model ID),模型名字等,模型所使用的算法(如神经网络,随机森林等),模型对应的训练对象信息(如某个用户,某个区域AOI),模型对应的时间信息(针对特定时间段、时间点的模型),模型训练的精度要求(如模型准确度90%等)等至少一项。(必选)
d)模型接收端设备信息:包括模型接收端设备的地址信息或标识信息,可以隐式发送,用于将模型反馈给模型接收端设备。(可选)
e)模型接收端设备AI框架信息:告知模型发送端设备,再经由模型发送端设备告诉转化平台,模型接收端设备自己需要什么样的框架。(“必选“)
f)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1的AI模型请求消息,模型发送端设备向转化平台发送所请求的模型信息。
首先,模型发送端设备根据模型接收端设备的请求消息,选择或训练生成模型接收端设备所请求的AI模型。具体地,模型发送端设备根据请求消息中的AI模型描述信息选择或训练生成与该信息匹配的AI模型。例如,模型发送端设备按照神经网络算法,对某训练对象(如UE)在具体区域或时间段中的相关数据(根据analytics ID选择输入数据类型)执行机器学习模型训练过程,生成模型精度符合预期值的AI模型。模型训练的方法在本发明中不做限定。
其中,模型发送端设备向转化平台发送所请求的模型信息,包括以下至少一项:
g)AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。(必选)。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。
h)AI模型框架信息,告知转化平台该AI模型文件所用框架,方便转化平台进行转化。(如果转化平台可以自己判断,可为“可选”)。如AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。
i)模型接收端设备信息:告知转化平台该把转化后模型发送到哪里,如模型接收端设备IP地址或FQDN等标识信息(可选);
j)模型接收端设备框架信息:告知转化平台该把模型转化成哪种框架的模型。(可选)。例如需将TensorFLow生成的模型文件转化成PyTorch可识别的模型文件,则这里可以是“PyTorch”+“1.10.0”(框架平台+版本号)的形式。
步骤3:
根据步骤2接收到的模型信息,转化平台向模型发送端设备发送响应信息:
c)若转化平台能够将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化被接受或成功;
d)若转化平台无法将接收到的模型信息转化为模型接收端设备框架对应的模型,则所述响应消息用于告知模型发送端设备模型转化失败。其中还可以进一步包括失败的原因值,例如因为不支持AI模型从框架A模型转换成框架B模型。
步骤4:
根据步骤3接收到的响应信息,模型发送端设备给模型接收端设备发送请求响应信息(可选):
a)请求接受;如当接受到“转化被接受或成功”时(步骤3.a),告知接受端其请求合理可行并已被接受。
b)请求拒绝。如当接收到“转化失败”时(步骤3.b),告知模型发送端设备其请求被拒绝。其中也可以根据“步骤3.b”的反馈,进一步告知其失败的原因,例如转化平台暂不支持从AI框架A转化为AI框架B。或者,因为模型发送端设备的原因,拒绝模型接收端设备的请求,如该模型发送端设备无法生成所需AI模型。
需要说明的是,该步骤4与下述步骤5不限定时间上的先后顺序。
步骤5:
转化平台将AI模型转化成模型接收端设备AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从模型发送端设备AI框架模型转化成模型接收端设备AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。例如,通过ONNX提供的Python脚本,转化平台(“tf2onnx.conver”)可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息,生产者信息外,还有计算图信息等。计算图Graph(GraphProto类型)包含了网络结构和参数等信息,由一些名称等基本信息和四组列表组成。这四组数组为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型输入节点的名称,类型, 形状等除数据以外的信息,output存放了模型输出节点名称,类型,形状等除数据以外的信息,initializer存放了模型的所有网络参数的具体数值,包括超参数和输入值。具体地,每个Node中包含它的运算操作类型和指定的inputs和outputs名字等信息。一个Node的节点信息,可以是一整层的计算过程,也可以是一个节点的计算。所有node连接在一起组成graph,通过Node计算节点中的inputs和outputs两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤6:
转化平台将转化后的AI模型发送给模型接收端设备。
具体地,转化平台将转化后的AI模型文件发送给模型接收端设备,该AI模型文件是所述模型接收端设备所支持的,换句话说,模型接收端设备可以理解并使用该AI模型文件。
该步骤中,还可包含模型发送端设备AI框架信息(如模型发送端设备支持TensorFlow框架)。该信息用于向模型接收端设备指明模型发送端设备具体使用的AI框架。
在模型转化平台在模型发送端设备后方的场景下,本申请还提供如下的实施方式。
图13为本申请实施例提供的AI模型传输方法的流程示意图之四。如图13所示,该AI模型传输方法应用于模型发送端设备,该AI模型传输方法包括以下步骤:
步骤1300、模型发送端设备向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
可以理解的是,模型发送端设备可以直接向模型转化平台发送第三请求消息,以请求模型转化平台进行模型转换。
在本申请实施例中,模型发送端设备在某些场景下,向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型,,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
可选地,所述第三请求消息包括以下至少一项:
所述第二AI模型;
模型发送端设备的属性信息;
模型接收端设备的属性信息;
模型转化要求信息。
此处,第二AI模型包括AI模型的完整的网络结构和参数信息等文件。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与第二模型表示方法对应。
可选地,所述模型接收端设备的属性信息,包括:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型转化要求信息包括:
模型转化时延要求信息;
模型转化准确率要求信息。
对于上述各信息的理解,可以参考前述实施例中的描述,在此不再赘述。
可选地,所述方法还包括:
模型发送端设备接收模型转化平台发送的第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作。
模型发送端设备向模型转化平台发送第三请求消息之后,可以接收到模型转化平台发送的响应消息,即第二模型转化响应消息,用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作。
可选地,所述模型发送端设备向模型转化平台发送第三请求消息之前,所述方法还包括:
所述模型发送端设备接收第三方设备发送的第四请求消息,所述第四请求消息用于请求所述模型发送端设备向所述模型转化平台发送第二AI模型。
可选地,第三方设备是指没有AI模型,与模型接收端设备达成了协议的设备,第三方设备通过向模型发送端设备发送第四请求消息,触发模型转化。
所述模型发送端设备接收第三方设备发送的第四请求消息,所述第四请求消息用于请求所述模型发送端设备向所述模型转化平台发送第二AI模型,以指示模型转化平台进行模型转化,并将模型转化后得到的模型下发给模型接收端设备。
本申请实施例提供了一种跨平台框架,跨用户/网元的AI模型传递方法,是一种基于模型发送端设立转化平台的方法。此方法在为模型发送端节省了硬盘空间;减少了后续的更新成本的情况下,增加了转化平台的重复利用率;增加了支持转换的框架种类;使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了方法;还可以使模型群发,而不再限制于一对一的发送。
图14为本申请实施例提供的的AI模型传输方法的流程示意图之五,如图14所示,该AI模型传输方法应用于模型转化平台,该AI模型传输方法包括以下步骤:
步骤1400、模型转化平台接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
步骤1401、所述模型转化平台根据所述第三请求消息,执行第二操作;
其中,所述执行第二操作包括以下至少一项:
确定是否能够完成对所述第二AI模型的转化操作;
向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
将第二AI模型转化为第一AI模型;
向模型接收端设备发送所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
模型发送端设备可以直接向模型转化平台发送第三请求消息,以请求模型转化平台进行模型转换。
模型转化平台接收到模型发送端设备发送的第三请求消息后,根据自身的转化能力信息,第一AI模型的框架信息,第二AI模型的框架信息,确定是否能够完成对所述第二AI模型的转化操作,向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作。
若模型转化平台能够完成对所述第二AI模型的转化操作,则将第二AI模型转化为第一AI模型,并将第一AI模型下发给模型接收端设备。其中,模型接收端设备可以是多个设备,从而可以实现AI模型的群发,从而不再限制于一对一的发送。
若模型转化平台无法完成对所述第二AI模型的转化操作,则向模型发送端设备发送用于指示所述模型转化平台无法完成对所述第二AI模型的转化操作的第二模型转化响应消息。
需要说明的是,本申请不限制向所述模型发送端设备发送第二模型转化响应消息,以及将第二AI模型转化为第一AI模型的先后顺序。
例如,在将第二AI模型转化为第一AI模型的过程中,模型转化平台可能会遇到无法预知的故障,从而无法完成对所述第二AI模型的转化操作,也会向模型发送端设备发送用于指示所述模型转化平台无法完成对所述第二AI模型的转化操作的第二模型转化响应消息。
可选地,所述第三请求消息包括以下至少一项:
所述第二AI模型;
模型发送端设备的属性信息;
模型接收端设备的属性信息;
模型转化要求信息。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与第二模型表示方法对应。
可选地,所述模型接收端设备的属性信息,包括:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型转化要求信息包括:
模型转化时延要求信息;
模型转化准确率要求信息。
需要说明的是,对于上述各信息的理解可以参考前述实施例中的描述,在此不再赘述。
本申请实施例提供了一种跨平台框架,跨用户/网元的AI模型传递方法,是一种基于模型发送端设立转化平台的方法。此方法在为模型发送端节省了硬盘空间;减少了后续的更新成本的情况下,增加了转化平台的重复利用率;增加了支持转换的框架种类;使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了方法;还可以使模型群发,而不再限制于一对一的发送。
图15为本申请实施例提供的AI模型传输方法的交互流程示意图之七。如图15所示,该AI模型传输方法包括以下步骤:
步骤1:
模型发送端向转化平台发送转化请求,包含信息如下至少一项:
a)所生成的模型:AI模型文件,包含所生成AI模型的完整的网络结构和参数信息等元素的文件。例如使用TensorFlow生产的AI模型会保存为“.meta”后缀的文件,里面包括AI模型的网络结构;和“.ckpt”后缀的文件,包含了参数信息。(必选)
b)模型发送端所使用的框架:告知转化平台该AI模型文件所用框架,方便转化平台进行转化。如AI模型文件为TesnorFLow所生成的文件,则该信息可为“TensorFLow”+“2.5.3”(即框架平台+版本号的形式)。(如果转化平台可以自己判断,可为“可选”)
c)模型接收端信息:方便转化平台发送结果到包括接收端的地址信息或标识信息,可以隐式发送,用于将模型反馈给接收端。(可选)
d)接收端AI框架信息:告知转化平台该把模型转化成哪种什么框架的模型。如需将TensorFLow生成的模型文件转化成PyTorch可识别的模型文件,则这里可以是“PyTorch”+“1.10.0”(框架平台+版本号)的形式。
e)模型转化的要求信息:用于指导转化平台进行转化的信息,如时延要求(5s内转化完成并发送),准确性要求(转化准确率达99%以上)等。
步骤2:
根据步骤1接收到的模型信息,转化平台向发送端发送响应信息:
a)若转化平台能够将接收到的模型信息转化为接收端框架对应的模型,则所述响应消息用于告知发送端模型转化被接受或成功;
b)若转化平台无法将接收到的模型信息转化为接收端框架对应的模型,则所述响应消息用于告知发送端模型转化失败。其中还可以进一步包括失败的原因值,例如因为不支持AI模型从框架A模型转换成框架B模型。
步骤3:
如果步骤1执行,即有AI模型成功从发送端发送到转化平台,则转化平台将AI模型转化成接受端AI框架可用的模型。
具体地,转化平台将所接收的AI模型文件从发送端AI框架模型转化成接收端AI框架可用的模型。例如,从TensorFlow框架的模型A装换成ONNX框架的模型B。通过ONNX提供的Python脚本,“tf2onnx.conver”可以将TensorFlow的模型文件转化成“.onnx”后缀的ONNX文件。在这个文件中,除了版本信息、生产者信息外,还有四组数组,它们为node(NodeProto类型),input(ValueInfoProto类型),output(ValueInfoProto类型)和initializer(TensorProto类型),其中node中存放了模型中所有的计算节点,input存放了模型的输入节点,output存放了模型中所有的输出节点,initializer存放了模型的所有权重参数。通过Node计算节点中的input和output两个数组指向输入和输出的节点关系,构建出整个网络的拓扑结构,也就是网络结构。
步骤4:
转化平台将转化后的AI模型下发给接收端,具体地,转化平台将转化后的AI模型文件发送给接收端,该AI模型文件是所述接收端所支持的,换句话说,接收端可以理解并使用该AI模型文件。
该步骤中,并还可包含发送端AI框架信息(如发送端IP地址等支持TensorFlow框架)。该信息用于向接收端指明发送端具体使用的AI框架。
可选地,图16为本申请实施例提供的AI模型传输方法的交互流程示意图之八。在图15的基础上,所述方法还包括:
所述模型发送端设备接收第三方设备发送的第四请求消息,所述第四请求消息用于请求所述模型发送端设备向所述模型转化平台发送第二AI模型,以使得模型转化平台将第二AI模型转化为第一AI模型并将第一AI模型下发给模型接收端设备。
本实施例中的其他步骤可以参考图15,在此不再赘述。
本申请实施例提供的AI模型传输方法,执行主体可以为AI模型传输装置。本申请实施例中以AI模型传输装置执行AI模型传输方法为例,说明本申请实施例提供的AI模型传输装置。
图17为本申请实施例提供的AI模型传输装置的结构示意图之一。如图17所示,该AI模型传输装置1700包括:
第一发送单元1710,用于发送第一请求消息,所述第一请求消息用于请求第一AI模 型,所述第一请求消息包括AI模型描述信息;
第一接收单元1720,用于接收第一响应消息;
其中,所述第一响应消息包括以下其中之一:
所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
在本申请实施例中,通过发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息,接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,从而使得不同平台框架的AI模型传递互通成为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
可选地,所述第一发送单元,用于:
向模型发送端设备发送所述第一请求消息;或者,
向所述模型转化平台发送所述第一请求消息。
可选地,所述第一接收单元,用于:
从所述模型转化平台获取所述第一响应消息,所述第一响应消息包括所述第一AI模型或所述第一指示信息或所述第二指示信息。
可选地,所述第一接收单元,用于:
从所述模型发送端设备获取所述第一响应消息,所述第一响应消息包括所述第一指示信息或所述第二指示信息。
可选地,所述AI模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
可选地,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
所述模型转化平台的属性信息;
所述模型接收端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,所述装置还包括:
第一上报单元,用于向所述模型转化平台注册上报所述模型接收端设备的属性信息,所述属性信息包括所述模型接收端设备支持的AI框架信息。
可选地,所述装置还包括:
第二发送单元,用于在接收到所述模型转化平台发送的第一AI框架请求消息的情况下,向所述模型转化平台发送所述模型接收端设备支持的AI框架信息,其中,所述第一AI框架请求消息用于请求所述模型接收端设备支持的AI框架信息。
可选地,在向所述模型转化平台发送所述第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
所述模型发送端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,所述模型转化平台的属性信息包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息包括以下至少一项:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,从 所述模型转化平台获取所述第二指示信息。
可选地,在满足第一条件的情况下,所述模型接收端设备从所述模型发送端设备获取所述第二指示信息;
其中,所述第一条件包括以下至少一项:
所述模型发送端设备根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
所述模型发送端设备接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
可选地,在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,从所述模型发送端设备或模型转化平台获取所述第一指示信息。
在本申请实施例中,通过发送第一请求消息,并接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,其中,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
本申请实施例提供的AI模型传输装置能够实现图2、图6-图12的方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
图18为本申请实施例提供的AI模型传输装置的结构示意图之二。如图18所示,该AI模型传输装置1800包括:
第二接收单元1810,用于接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;
第一执行单元1820,用于根据所述AI模型请求消息,执行第一操作;
其中,所述执行第一操作包括以下至少一项:
选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
在本申请实施例中,通过接收AI模型请求消息,根据所述AI模型请求消息,执行第一操作,包括选择或训练生成第二AI模型,向模型转化平台发送第二AI模型,或者,发送第一指示信息或第二指示信息,从而使得不同平台框架的AI模型传递互通成为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
可选地,所述第二接收单元,用于:
接收所述模型转化平台发送的第二请求消息,其中,所述第二请求信息是所述模型转化平台根据所述模型接收端设备发送给所述模型转化平台的第一请求消息确定;
或者,
接收所述模型接收端设备发送的第一请求消息。
可选地,在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
向所述模型转化平台发送第二响应消息,所述第二响应消息包括所述第二AI模型。
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
向所述模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;所述第三请求消息包括所述第二AI模型。
可选地,所述AI模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
可选地,所述第二响应消息还包括以下至少一项:
所述第二AI模型的框架信息;
所述模型接收端设备的属性信息。
可选地,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
所述模型转化平台的属性信息;
所述模型接收端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,所述装置还包括:
第二上报单元,用于向所述模型转化平台注册上报所述模型发送端设备的属性信息,所述属性信息包括所述模型发送端设备支持的AI框架信息。
可选地,所述装置还包括:
第三发送单元,用于在接收所述模型转化平台发送的第二AI框架请求消息的情况下,向所述模型转化平台发送所述模型发送端设备支持的AI框架信息,其中,所述第二AI框架请求消息用于请求所述模型发送端设备支持的AI框架信息。
可选地,在接收所述模型转化平台发送的第二请求消息的情况下,所述第二请求消息还包括以下至少一项:
所述模型转化平台的属性信息;
其中,所述发送给所述模型转化平台的第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
所述模型发送端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,所述模型转化平台的属性信息包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,所述发送第一指示信息,包括:
在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型接收端设备发送所述第一指示信息。
可选地,所述装置还包括:
第四发送单元,用于在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型转化平台发送第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
可选地,所述装置还包括:
第一确定单元,用于根据所述模型转化平台的转化能力信息确定所述模型转化平台是否可以完成对所述第二AI模型的转化操作。
可选地,所述装置还包括:
第三接收单元,用于接收所述模型转化平台发送的第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
可选地,所述发送第二指示信息,包括:
在满足第一条件的情况下,向所述模型接收端设备发送第二指示信息;
其中,所述第一条件包括以下至少一项:
根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
在本申请实施例中,通过接收AI模型请求消息,根据所述AI模型请求消息,执行第一操作,包括选择或训练生成第二AI模型,向模型转化平台发送第二AI模型,或者,发送第一指示信息或第二指示信息,第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
本申请实施例提供的AI模型传输装置能够实现图3、图6-图12的方法实施例实现的 各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
图19为本申请实施例提供的AI模型传输装置的结构示意图之三。如图19所示,该AI模型传输装置1900包括:
第四接收单元1910,用于接收模型发送端设备发送的第二AI模型;
第一处理单元1920,用于将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;或者,
可选地,模型转化平台包括:
第五发送单元,用于向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
在本申请实施例中,通过接收模型发送端设备发送的第二AI模型;将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备,或者向模型接收端设备发送第一指示信息或第二指示信息,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
可选地,所述第四接收单元,用于:
接收模型发送端设备发送的第二响应消息,所述第二响应消息包括所述第二AI模型。
可选地,所述第四接收单元,用于:
接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型。
可选地,所述装置还包括第五接收单元,用于:
接收所述模型发送端设备上报的所述模型发送端设备的属性信息,所述模型发送端设备的属性信息包括所述模型发送端设备支持的AI框架信息;
接收所述模型接收端设备上报的所述模型接收端设备的属性信息,所述模型接收端设备的属性信息包括所述模型接收端设备支持的AI框架信息。
可选地,所述装置还包括第六发送单元,用于:
向模型接收端设备发送第一AI框架请求消息,接收所述模型接收端设备支持的AI框架信息;
向模型发送端设备发送第二AI框架请求消息,接收所述模型发送端设备支持的AI框架信息。
可选地,所述第二响应消息还包括以下至少一项:
所述模型发送端设备支持的AI框架信息;
所述模型接收端设备的属性信息。
可选地,所述装置还包括:
第六接收单元,用于接收所述模型接收端设备发送的第一请求消息;
第七发送单元,用于所述模型转化平台向模型发送端设备发送第二请求消息;
其中,所述第二请求信息根据所述第一请求消息确定。
可选地,所述第一请求消息或第二请求消息包括:AI模型描述信息,所述模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
可选地,所述第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
所述模型发送端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息;
所述第二请求消息还包括以下至少一项:
所述模型转化平台的属性信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息包括以下至少一项:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,所述模型转化平台的属性信息,包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述装置还包括:
第二确定单元,用于确定所述模型转化平台具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力。
可选地,所述装置还包括:
第三确定单元,用于确定所述模型转化平台不具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力;
第八发送单元,用于向所述模型接收端设备发送第三响应消息,所述第三响应消息用于指示所述第一请求消息被拒绝。
可选地,在所述第一请求消息中未携带所述模型发送端设备的属性信息的情况下,所述装置还包括:
第四确定单元,用于确定所述模型发送端设备。
可选地,所述装置还包括:
第九发送单元,用于向模型发送端设备发送第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
可选地,所述第五发送单元,用于;
在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,向所述模型接收端设备发送所述第二指示信息。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括以下至少一项:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型提供时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确度不能满足所述模型接收端设备请求的模型准确度要求。
可选地,所述装置还包括:
第七接收单元,用于接收模型发送端设备发送的第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
在本申请实施例中,通过接收模型发送端设备发送的第二AI模型;将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备,或者向模型接收端设备发送第一指示信息或第二指示信息,其中,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传 递提供了解决方案。
本申请实施例提供的AI模型传输装置能够实现图4-图12的方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
图20为本申请实施例提供的AI模型传输装置的结构示意图之四。如图20所示,该AI模型传输装置2000包括:
第十发送单元2010,用于向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
在本申请实施例中,在某些场景下,通过向模型转化平台发送发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型,可使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
可选地,所述第三请求消息包括以下至少一项:
所述第二AI模型;
模型发送端设备的属性信息;
模型接收端设备的属性信息;
模型转化要求信息。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与第二模型表示方法对应。
可选地,所述模型接收端设备的属性信息,包括:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型转化要求信息包括:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,所述装置还包括:
第八接收单元,用于接收模型转化平台发送的第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作。
可选地,所述装置还包括:
第九接收单元,用于所述模型发送端设备接收第三方设备发送的第四请求消息,所述第四请求消息用于请求所述模型发送端设备向所述模型转化平台发送第二AI模型。
本申请实施例提供了一种跨平台框架,跨用户/网元的AI模型传递方案,可为模型发送端节省了硬盘空间,减少了后续的更新成本的情况下,增加了转化平台的重复利用率;增加了支持转换的框架种类;使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了方法;还可以使模型群发,而不再限制于一对一的发送。
本申请实施例提供的AI模型传输装置能够实现图13、图15-图16的方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
图21为本申请实施例提供的AI模型传输装置的结构示意图之五。如图21所示,该AI模型传输装置2100包括:
第十接收单元2110,用于接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
第二执行单元2120,用于根据所述第三请求消息,执行第二操作;
其中,所述执行第二操作包括以下至少一项:
确定是否能够完成对所述第二AI模型的转化操作;
向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
将第二AI模型转化为第一AI模型;
向模型接收端设备发送所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
可选地,所述第三请求消息包括以下至少一项:
所述第二AI模型;
模型发送端设备的属性信息;
模型接收端设备的属性信息;
模型转化要求信息。
可选地,所述模型发送端设备的属性信息,包括:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与第二模型表示方法对应。
可选地,所述模型接收端设备的属性信息,包括:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型转化要求信息包括:
模型转化时延要求信息;
模型转化准确率要求信息。
本申请实施例提供了一种跨平台框架,跨用户/网元的AI模型传递方案,可为模型发送端节省了硬盘空间,减少了后续的更新成本的情况下,增加了转化平台的重复利用率;增加了支持转换的框架种类;使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了方法;还可以使模型群发,而不再限制于一对一的发送。
本申请实施例提供的AI模型传输装置能够实现图14-图16的方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例中的AI模型传输装置可以是电子设备,例如具有操作系统的电子设备,也可以是电子设备中的部件,例如集成电路或芯片。该电子设备可以是终端,也可以为除终端之外的其他设备。示例性的,终端可以包括但不限于上述所列举的终端11的类型,其他设备可以为服务器、网络附属存储器(Network Attached Storage,NAS)等,本申请实施例不作具体限定。
可选的,如图22所示,本申请实施例还提供一种通信设备2200,包括处理器2201和存储器2202,存储器2202上存储有可在所述处理器2201上运行的程序或指令,例如,该通信设备2200为终端时,该程序或指令被处理器2201执行时实现上述AI模型传输方法实施例的各个步骤,且能达到相同的技术效果。该通信设备2200为网络侧设备时,该程序或指令被处理器2201执行时实现上述AI模型传输方法实施例的各个步骤,且能达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例还提供一种模型接收端设备,包括处理器和通信接口,其中,所述通信接口用于发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;所述通信接口还用于接收第一响应消息;其中,所述第一响应消息包括以下其中之一:所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
该模型接收端设备实施例与上述模型接收端设备侧方法实施例对应,上述方法实施例的各个实施过程和实现方式均可适用于该模型接收端设备实施例中,且能达到相同的技术效果。
具体地,图23为实现本申请实施例的一种模型接收端设备的硬件结构示意图。
该终端2300包括但不限于:射频单元2301、网络模块2302、音频输出单元2303、输入单元2304、传感器2305、显示单元2306、用户输入单元2307、接口单元2308、存储器2309以及处理器2310等中的至少部分部件。
本领域技术人员可以理解,终端2300还可以包括给各个部件供电的电源(比如电池),电源可以通过电源管理系统与处理器2310逻辑相连,从而通过电源管理系统实现管理充电、放电、以及功耗管理等功能。图23中示出的终端结构并不构成对终端的限定,终端可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件布置,在此不再赘述。
应理解的是,本申请实施例中,输入单元2304可以包括图形处理单元(Graphics Processing Unit,GPU)23041和麦克风23042,图形处理器23041对在视频捕获模式或图像捕获模式中由图像捕获装置(如摄像头)获得的静态图片或视频的图像数据进行处理。显示单元2306可包括显示面板2061,可以采用液晶显示器、有机发光二极管等形式来配置显示面板23061。用户输入单元2307包括触控面板23071以及其他输入设备23072中的至少一种。触控面板23071,也称为触摸屏。触控面板23071可包括触摸检测装置和触摸控制器两个部分。其他输入设备23072可以包括但不限于物理键盘、功能键(比如音量控制按键、开关按键等)、轨迹球、鼠标、操作杆,在此不再赘述。
本申请实施例中,射频单元2301接收来自网络侧设备的下行数据后,可以传输给处理器2310进行处理;另外,射频单元2301可以向网络侧设备发送上行数据。通常,射频单元2301包括但不限于天线、放大器、收发信机、耦合器、低噪声放大器、双工器等。
存储器2309可用于存储软件程序或指令以及各种数据。存储器2309可主要包括存储程序或指令的第一存储区和存储数据的第二存储区,其中,第一存储区可存储操作系统、至少一个功能所需的应用程序或指令(比如声音播放功能、图像播放功能等)等。此外,存储器2309可以包括易失性存储器或非易失性存储器,或者,存储器2309可以包括易失性和非易失性存储器两者。其中,非易失性存储器可以是只读存储器(Read-Only Memory,ROM)、可编程只读存储器(Programmable ROM,PROM)、可擦除可编程只读存储器(Erasable PROM,EPROM)、电可擦除可编程只读存储器(Electrically EPROM,EEPROM)或闪存。易失性存储器可以是随机存取存储器(Random Access Memory,RAM),静态随机存取存储器(Static RAM,SRAM)、动态随机存取存储器(Dynamic RAM,DRAM)、同步动态随机存取存储器(Synchronous DRAM,SDRAM)、双倍数据速率同步动态随机存取存储器(Double Data Rate SDRAM,DDRSDRAM)、增强型同步动态随机存取存储器(Enhanced SDRAM,ESDRAM)、同步连接动态随机存取存储器(Synch link DRAM,SLDRAM)和直接内存总线随机存取存储器(Direct Rambus RAM,DRRAM)。本申请实施例中的存储器2309包括但不限于这些和任意其它适合类型的存储器。
处理器2310可包括一个或多个处理单元;可选的,处理器2310集成应用处理器和调制解调处理器,其中,应用处理器主要处理涉及操作系统、用户界面和应用程序等的操作, 调制解调处理器主要处理无线通信信号,如基带处理器。可以理解的是,上述调制解调处理器也可以不集成到处理器2310中。
其中,射频单元2301,用于发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;
射频单元2301,还用于接收第一响应消息;
其中,所述第一响应消息包括以下其中之一:
所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
在本申请实施例中,模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息,接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,从而使得不同平台框架的AI模型传递互通成为了可能,为不同厂商的设备的AI模型传递提供了解决方法。
可选地,射频单元2301,用于:
向模型发送端设备发送所述第一请求消息;或者,
向所述模型转化平台发送所述第一请求消息。
可选地,射频单元2301,用于:
所述模型接收端设备从所述模型转化平台获取所述第一响应消息,所述第一响应消息包括所述第一AI模型或所述第一指示信息或所述第二指示信息。
可选地,射频单元2301,用于:
从所述模型发送端设备获取所述第一响应消息,所述第一响应消息包括所述第一指示信息或所述第二指示信息。
可选地,所述AI模型描述信息包括以下至少一项:
AI模型类型信息;
AI模型标识信息;
AI模型名称信息;
AI模型算法信息;
AI模型对应的训练对象信息;
AI模型对应的时间信息;
AI模型训练精度要求信息。
可选地,在所述模型转化平台在所述模型接收端设备的后方的情况下,所述向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
所述模型转化平台的属性信息;
所述模型接收端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,在所述模型转化平台在所述模型发送端设备的后方或者所述模型转化平台为公共转化平台的情况下,向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,在所述模型转化平台为公共转化平台的情况下,向模型发送端设备发送的所述第一请求消息还包括:
所述模型接收端设备的地址和/或标识信息。
可选地,射频单元2301,用于:
向所述模型转化平台注册上报所述模型接收端设备的属性信息,所述属性信息包括所述模型接收端设备支持的AI框架信息。
可选地,射频单元2301,用于:
在接收到所述模型转化平台发送的第一AI框架请求消息的情况下,向所述模型转化平台发送所述模型接收端设备支持的AI框架信息,其中,所述第一AI框架请求消息用于请求所述模型接收端设备支持的AI框架信息。
可选地,在向所述模型转化平台发送所述第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
所述模型接收端设备的属性信息;
所述模型发送端设备的属性信息;
模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
可选地,所述模型转化平台的属性信息包括以下至少一项:
所述模型转化平台的地址信息;
所述模型转化平台的标识信息;
所述模型转化平台的转化能力信息。
可选地,所述模型接收端设备的属性信息包括以下至少一项:
所述模型接收端设备的地址信息;
所述模型接收端设备的标识信息;
所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
可选地,所述模型发送端设备的属性信息包括以下至少一项:
模型发送端设备的地址信息;
模型发送端设备的标识信息;
模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
可选地,所述模型转化要求信息包括以下至少一项:
模型转化时延要求信息;
模型转化准确率要求信息。
可选地,在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,从所述模型转化平台获取所述第二指示信息。
可选地,在满足第一条件的情况下,所述模型接收端设备从所述模型发送端设备获取所述第二指示信息;
其中,所述第一条件包括以下至少一项:
所述模型发送端设备根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
所述模型发送端设备接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
可选地,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
可选地,在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,从所述模型发送端设备或模型转化平台获取所述第一指示信息。
在本申请实施例中,模型接收端设备发送第一请求消息,并接收第一响应消息,第一响应消息包括第一AI模型,第一指示信息或第二指示信息,第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的、能够被模型接收端设备使用的模型,其中,模型转化平台可以部署在模型接收端设备的后方,也可以部署在模型发送端设备的后方,或者模型转化平台为公共转化平台,减少了后续的更新成本的情况下,增加了转化平台的重复利用率,增加了支持转换的框架种类,使得不同平台框架的AI模型传递互通成为了可能,也为不同厂商的设备的AI模型传递提供了解决方案。
本申请实施例还提供一种模型发送端设备,包括处理器和通信接口,其中,所述通信接口用于接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;所述处理器用于根据所述AI模型请求消息,执行第一操作;.其中,所述执行第一操作包括以下至少一项:选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
或者,所述通信接口用于向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
该模型发送端设备实施例与上述模型发送端设备侧方法实施例对应,上述方法实施例的各个实施过程和实现方式均可适用于该模型发送端设备实施例中,且能达到相同的技术效果。
具体地,本申请实施例还提供了一种模型发送端设备。如图24所示,该模型发送端设备2400包括:天线2401、射频装置2402、基带装置2403、处理器2404和存储器2405。天线2401与射频装置2402连接。在上行方向上,射频装置2402通过天线2401接收信息,将接收的信息发送给基带装置2403进行处理。在下行方向上,基带装置2403对要发送的信息进行处理,并发送给射频装置2402,射频装置2402对收到的信息进行处理后经过天线2401发送出去。
以上实施例中模型发送端设备执行的方法可以在基带装置2403中实现,该基带装置2403包括基带处理器。
基带装置2403例如可以包括至少一个基带板,该基带板上设置有多个芯片,如图24所示,其中一个芯片例如为基带处理器,通过总线接口与存储器2405连接,以调用存储器2405中的程序,执行以上方法实施例中所示的网络设备操作。
该模型发送端设备还可以包括网络接口2406,该接口例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本发明实施例的模型发送端设备2400还包括:存储在存储器2405上并可在处理器2404上运行的指令或程序,处理器2404调用存储器2405中的指令或程序执行图18或图20所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
本申请实施例还提供一种模型转化平台,包括处理器及通信接口,其中,所述通信接口用于接收模型发送端设备发送的第二AI模型;所述处理器用于将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;所述通信接口还用于:向 所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
或者,所述通信接口用于接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;所述处理器用于根据所述第三请求消息,执行第二操作;其中,所述执行第二操作包括以下至少一项:确定是否能够完成对所述第二AI模型的转化操作;向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;将第二AI模型转化为第一AI模型;向模型接收端设备发送所述第一AI模型。
该模型转化平台实施例与上述模型转化平台侧方法实施例对应,上述方法实施例的各个实施过程和实现方式均可适用于该模型转化平台实施例中,且能达到相同的技术效果。
本申请实施例还提供一种可读存储介质,所述可读存储介质上存储有程序或指令,该程序或指令被处理器执行时实现上述AI模型传输方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
其中,所述处理器为上述实施例中所述的终端中的处理器。所述可读存储介质,包括计算机可读存储介质,如计算机只读存储器ROM、随机存取存储器RAM、磁碟或者光盘等。
本申请实施例另提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和所述处理器耦合,所述处理器用于运行程序或指令,实现上述AI模型传输方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
应理解,本申请实施例提到的芯片还可以称为系统级芯片,系统芯片,芯片系统或片上系统芯片等。
本申请实施例另提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现上述AI模型传输方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例还提供了一种AI模型传输系统,包括:模型接收端设备、模型发送端设备以及模型转化平台,所述模型接收端设备可用于执行如上所述的AI模型传输方法的步骤,所述模型发送端设备可用于执行如上所述的AI模型传输方法的步骤,所述模型转化平台可用于执行如上所述的AI模型传输方法的步骤。
需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除 在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。此外,需要指出的是,本申请实施方式中的方法和装置的范围不限按示出或讨论的顺序来执行功能,还可包括根据所涉及的功能按基本同时的方式或按相反的顺序来执行功能,例如,可以按不同于所描述的次序来执行所描述的方法,并且还可以添加、省去、或组合各种步骤。另外,参照某些示例所描述的特征可在其他示例中被组合。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以计算机软件产品的形式体现出来,该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端(可以是手机,计算机,服务器,空调器,或者网络设备等)执行本申请各个实施例所述的方法。
上面结合附图对本申请的实施例进行了描述,但是本申请并不局限于上述的具体实施方式,上述的具体实施方式仅仅是示意性的,而不是限制性的,本领域的普通技术人员在本申请的启示下,在不脱离本申请宗旨和权利要求所保护的范围情况下,还可做出很多形式,均属于本申请的保护之内。

Claims (78)

  1. 一种AI模型传输方法,包括:
    模型接收端设备发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;
    所述模型接收端设备接收第一响应消息;
    其中,所述第一响应消息包括以下其中之一:
    所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
    第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
    第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  2. 根据权利要求1所述的AI模型传输方法,其中,所述模型接收端设备发送第一请求消息,包括:
    所述模型接收端设备向模型发送端设备发送所述第一请求消息;或者,
    所述模型接收端设备向所述模型转化平台发送所述第一请求消息。
  3. 根据权利要求1所的AI模型传输方法,其中,所述模型接收端设备接收第一响应消息,包括:
    所述模型接收端设备从所述模型转化平台获取所述第一响应消息,所述第一响应消息包括所述第一AI模型或所述第一指示信息或所述第二指示信息。
  4. 根据权利要求1所述的AI模型传输方法,其中,所述模型接收端设备接收第一响应消息,包括:
    所述模型接收端设备从所述模型发送端设备获取所述第一响应消息,所述第一响应消息包括所述第一指示信息或所述第二指示信息。
  5. 根据权利要求1至4任一项所述的AI模型传输方法,其中,所述AI模型描述信息包括以下至少一项:
    AI模型类型信息;
    AI模型标识信息;
    AI模型名称信息;
    AI模型算法信息;
    AI模型对应的训练对象信息;
    AI模型对应的时间信息;
    AI模型训练精度要求信息。
  6. 根据权利要求2所述的AI模型传输方法,其中,所述模型接收端设备向模型发送端设备发送的所述第一请求消息还包括以下至少一项:
    所述模型转化平台的属性信息;
    所述模型接收端设备的属性信息;
    模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
  7. 根据权利要求2所述的AI模型传输方法,其中,所述方法还包括:
    所述模型接收端设备向所述模型转化平台注册上报所述模型接收端设备的属性信息,所述属性信息包括所述模型接收端设备支持的AI框架信息。
  8. 根据权利要求2所述的AI模型传输方法,其中,所述方法还包括:
    在接收到所述模型转化平台发送的第一AI框架请求消息的情况下,所述模型接收端设备向所述模型转化平台发送所述模型接收端设备支持的AI框架信息,其中,所述第一AI框架请求消息用于请求所述模型接收端设备支持的AI框架信息。
  9. 根据权利要求2所述的AI模型传输方法,其中,在所述模型接收端设备向所述模型转化平台发送所述第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
    所述模型接收端设备的属性信息;
    所述模型发送端设备的属性信息;
    模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
  10. 根据权利要求6所述的AI模型传输方法,其中,所述模型转化平台的属性信息包括以下至少一项:
    所述模型转化平台的地址信息;
    所述模型转化平台的标识信息;
    所述模型转化平台的转化能力信息。
  11. 根据权利要求6或9所述的AI模型传输方法,其中,所述模型接收端设备的属性信息包括以下至少一项:
    所述模型接收端设备的地址信息;
    所述模型接收端设备的标识信息;
    所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
  12. 根据权利要求9所述的AI模型传输方法,其中,所述模型发送端设备的属性信息包括以下至少一项:
    模型发送端设备的地址信息;
    模型发送端设备的标识信息;
    模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述 第二模型表示方法对应。
  13. 根据权利要求6或9所述的AI模型传输方法,其中,所述模型转化要求信息包括以下至少一项:
    模型转化时延要求信息;
    模型转化准确率要求信息。
  14. 根据权利要求3所述的AI模型传输方法,其中,在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,所述模型接收端设备从所述模型转化平台获取所述第二指示信息。
  15. 根据权利要求4所述的AI模型传输方法,其中,在满足第一条件的情况下,所述模型接收端设备从所述模型发送端设备获取所述第二指示信息;
    其中,所述第一条件包括以下至少一项:
    所述模型发送端设备根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
    所述模型发送端设备接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
  16. 根据权利要求14或15所述的AI模型传输方法,其中,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
    所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
    所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
    所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
  17. 根据权利要求1所述的AI模型传输方法,其中,在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,所述模型接收端设备从所述模型发送端设备或模型转化平台获取所述第一指示信息。
  18. 一种AI模型传输方法,包括:
    模型发送端设备接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;
    模型发送端设备根据所述AI模型请求消息,执行第一操作;
    其中,所述执行第一操作包括以下至少一项:
    选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
    发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  19. 根据权利要求18所述的AI模型传输方法,其中,所述模型发送端设备接收AI模型请求消息,包括:
    所述模型发送端设备接收所述模型转化平台发送的第二请求消息,其中,所述第二请求信息是所述模型转化平台根据所述模型接收端设备发送给所述模型转化平台的第一请求消息确定;
    或者,
    所述模型发送端设备接收所述模型接收端设备发送的第一请求消息。
  20. 根据权利要求19所述的AI模型传输方法,其中,在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
    所述模型发送端设备向所述模型转化平台发送第二响应消息,所述第二响应消息包括所述第二AI模型。
  21. 根据权利要求19所述的AI模型传输方法,其中,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息的情况下,所述向模型转化平台发送所述第二AI模型,包括:
    所述模型发送端设备向所述模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;所述第三请求消息包括所述第二AI模型。
  22. 根据权利要求18所述的AI模型传输方法,其中,所述AI模型描述信息包括以下至少一项:
    AI模型类型信息;
    AI模型标识信息;
    AI模型名称信息;
    AI模型算法信息;
    AI模型对应的训练对象信息;
    AI模型对应的时间信息;
    AI模型训练精度要求信息。
  23. 根据权利要求20所述的AI模型传输方法,其中,所述第二响应消息还包括以下至少一项:
    所述第二AI模型的框架信息;
    所述模型接收端设备的属性信息。
  24. 根据权利要求19所述的AI模型传输方法,其中,在所述模型发送端设备接收所述模型接收端设备发送的第一请求消息的情况下,所述第一请求消息还包括以下至少一项:
    所述模型转化平台的属性信息;
    所述模型接收端设备的属性信息;
    模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息。
  25. 根据权利要求24所述的AI模型传输方法,其中,所述方法还包括:
    所述模型发送端设备向所述模型转化平台注册上报所述模型发送端设备的属性信息,所述属性信息包括所述模型发送端设备支持的AI框架信息。
  26. 根据权利要求24所述的AI模型传输方法,其中,所述方法还包括:
    在接收所述模型转化平台发送的第二AI框架请求消息的情况下,所述模型发送端设备向所述模型转化平台发送所述模型发送端设备支持的AI框架信息,其中,所述第二AI框架请求消息用于请求所述模型发送端设备支持的AI框架信息。
  27. 根据权利要求19所述的AI模型传输方法,其中,在所述模型发送端设备接收所述模型转化平台发送的第二请求消息的情况下,所述第二请求消息还包括以下至少一项:
    所述模型转化平台的属性信息。
  28. 根据权利要求24或27所述的AI模型传输方法,其中,所述模型转化平台的属性信息包括以下至少一项:
    所述模型转化平台的地址信息;
    所述模型转化平台的标识信息;
    所述模型转化平台的转化能力信息。
  29. 根据权利要求23或24所述的AI模型传输方法,其中,所述模型接收端设备的属性信息包括以下至少一项:
    所述模型接收端设备的地址信息;
    所述模型接收端设备的标识信息;
    所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
  30. 根据权利要求25所述的AI模型传输方法,其中,所述模型发送端设备的属性信息,包括:
    模型发送端设备的地址信息;
    模型发送端设备的标识信息;
    模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
  31. 根据权利要求24所述的AI模型传输方法,其中,所述模型转化要求信息包括以 下至少一项:
    模型转化时延要求信息;
    模型转化准确率要求信息。
  32. 根据权利要求18所述的AI模型传输方法,其中,所述发送第一指示信息,包括:
    在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型接收端设备发送所述第一指示信息。
  33. 根据权利要求18所述的AI模型传输方法,其中,所述方法还包括:
    在所述模型发送端设备没有或无法生成所述第二AI模型的情况下,向所述模型转化平台发送第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
  34. 根据权利要求28所述的AI模型传输方法,其中,所述方法还包括:
    所述模型发送端设备根据所述模型转化平台的转化能力信息确定所述模型转化平台是否可以完成对所述第二AI模型的转化操作。
  35. 根据权利要求18所述的AI模型传输方法,其中,所述方法还包括:
    接收所述模型转化平台发送的第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
  36. 根据权利要求18所述的AI模型传输方法,其中,所述发送第二指示信息,包括:
    在满足第一条件的情况下,向所述模型接收端设备发送第二指示信息;
    其中,所述第一条件包括以下至少一项:
    根据所述模型转化平台的转化能力信息确定所述模型转化平台无法完成对所述第二AI模型的转化操作;
    接收到所述模型转化平台发送的第一模型转化响应消息,且所述第一模型转化响应消息指示所述模型转化平台无法完成对所述第二AI模型的转化操作。
  37. 根据权利要求36所述的AI模型传输方法,其中,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括:
    所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
    所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型转化时延要求;或,
    所述模型转化平台对所述第二AI模型的转化准确率不能满足所述模型接收端设备请求的模型转化准确率要求。
  38. 一种AI模型传输方法,包括:
    模型转化平台接收模型发送端设备发送的第二AI模型;
    所述模型转化平台将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发 送至模型接收端设备;或者,
    所述模型转化平台向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供第一AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  39. 根据权利要求38所述的AI模型传输方法,其中,所述模型转化平台接收模型发送端设备发送的第二AI模型,包括:
    所述模型转化平台接收模型发送端设备发送的第二响应消息,所述第二响应消息包括所述第二AI模型。
  40. 根据权利要求38所述的AI模型传输方法,其中,所述模型转化平台接收模型发送端设备发送的第二AI模型,包括:
    所述模型转化平台接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型。
  41. 根据权利要求39所述的AI模型传输方法,其中,所述方法还包括:
    所述模型转化平台接收所述模型接收端设备发送的第一请求消息;
    所述模型转化平台向模型发送端设备发送第二请求消息;
    其中,所述第二请求信息根据所述第一请求消息确定。
  42. 根据权利要求41所述的AI模型传输方法,其中,所述第一请求消息或第二请求消息包括:AI模型描述信息,所述模型描述信息包括以下至少一项:
    AI模型类型信息;
    AI模型标识信息;
    AI模型名称信息;
    AI模型算法信息;
    AI模型对应的训练对象信息;
    AI模型对应的时间信息;
    AI模型训练精度要求信息。
  43. 根据权利要求42所述的AI模型传输方法,其中,所述第一请求消息还包括以下至少一项:
    所述模型接收端设备的属性信息;
    所述模型发送端设备的属性信息;
    模型转化要求信息,所述模型转化要求信息用于指示对转化得到所述第一AI模型的要求信息;
    所述第二请求消息还包括以下至少一项:
    所述模型转化平台的属性信息。
  44. 根据权利要求38所述的AI模型传输方法,其中,所述方法还包括以下至少一项:
    所述模型转化平台接收所述模型发送端设备上报的所述模型发送端设备的属性信息,所述模型发送端设备的属性信息包括所述模型发送端设备支持的AI框架信息;
    所述模型转化平台接收所述模型接收端设备上报的所述模型接收端设备的属性信息,所述模型接收端设备的属性信息包括所述模型接收端设备支持的AI框架信息。
  45. 根据权利要求38所述的AI模型传输方法,其中,所述方法还包括以下至少一项:
    向模型接收端设备发送第一AI框架请求消息,接收所述模型接收端设备支持的AI框架信息;
    向模型发送端设备发送第二AI框架请求消息,接收所述模型发送端设备支持的AI框架信息。
  46. 根据权利要求39所述的AI模型传输方法,其中,所述第二响应消息还包括以下至少一项:
    所述模型发送端设备支持的AI框架信息;
    所述模型接收端设备的属性信息。
  47. 根据权利要求43或44或46所述的AI模型传输方法,其中,所述模型接收端设备的属性信息包括以下至少一项:
    所述模型接收端设备的地址信息;
    所述模型接收端设备的标识信息;
    所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
  48. 根据权利要求43或44所述的AI模型传输方法,其中,所述模型发送端设备的属性信息包括以下至少一项:
    模型发送端设备的地址信息;
    模型发送端设备的标识信息;
    模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
  49. 根据权利要求43所述的AI模型传输方法,其中,所述模型转化要求信息包括以下至少一项:
    模型转化时延要求信息;
    模型转化准确率要求信息。
  50. 根据权利要求43所述的AI模型传输方法,其中,所述模型转化平台的属性信息,包括以下至少一项:
    所述模型转化平台的地址信息;
    所述模型转化平台的标识信息;
    所述模型转化平台的转化能力信息。
  51. 根据权利要求41所述的AI模型传输方法,其中,所述模型转化平台向模型发送端设备发送第二请求消息之前,所述方法还包括:
    所述模型转化平台确定所述模型转化平台具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力。
  52. 根据权利要求41所述的AI模型传输方法,其中,所述模型转化平台接收所述模型接收端设备发送的第一请求消息之后,所述方法还包括:
    所述模型转化平台确定所述模型转化平台不具有从模型发送端设备支持的AI框架转化为模型接收端设备支持的AI框架的能力,所述模型转化平台向所述模型接收端设备发送第三响应消息,所述第三响应消息用于指示所述第一请求消息被拒绝。
  53. 根据权利要求41所述的AI模型传输方法,其中,在所述第一请求消息中未携带所述模型发送端设备的属性信息的情况下,所述模型转化平台向模型发送端设备发送第二请求消息之前,所述方法还包括:
    所述模型转化平台确定所述模型发送端设备。
  54. 根据权利要求38-40中任一项所述的AI模型传输方法,其中,所述模型转化平台接收模型发送端设备发送的第二AI模型之后,所述方法还包括:
    所述模型转化平台向模型发送端设备发送第一模型转化响应消息,所述第一模型转化响应消息用于指示所述模型转化平台是否完成对所述第二AI模型的转化操作。
  55. 根据权利要求38所述的AI模型传输方法,其中,所述模型转化平台向所述模型接收端设备发送第二指示信息,包括;
    在所述模型转化平台无法完成对所述第二AI模型的转化操作的情况下,所述模型转化平台向所述模型接收端设备发送所述第二指示信息。
  56. 根据权利要求55所述的AI模型传输方法,其中,所述模型转化平台无法完成对所述第二AI模型的转化操作,包括以下至少一项:
    所述模型转化平台不支持将所述第二AI模型从模型发送端设备的AI框架转化成所述模型接收端设备的AI框架;或,
    所述模型转化平台对所述第二AI模型的转化时延不能满足所述模型接收端设备请求的模型提供时延要求;或,
    所述模型转化平台对所述第二AI模型的转化准确度不能满足所述模型接收端设备请求的模型准确度要求。
  57. 根据权利要求38所述的模型传输方法,其中,所述模型转化平台向所述模型接收端设备发送第一指示信息之前,所述方法还包括:
    接收模型发送端设备发送的第四响应消息,所述第四响应消息用于指示所述模型发送端设备无法提供所述第二AI模型。
  58. 一种AI模型传输方法,包括:
    模型发送端设备向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
    其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示方法。
  59. 根据权利要求58所述的AI模型传输方法,其中,所述第三请求消息包括以下至少一项:
    所述第二AI模型;
    模型发送端设备的属性信息;
    模型接收端设备的属性信息;
    模型转化要求信息。
  60. 根据权利要求59所述的AI模型传输方法,其中,所述模型发送端设备的属性信息,包括:
    模型发送端设备的地址信息;
    模型发送端设备的标识信息;
    模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与第二模型表示方法对应。
  61. 根据权利要求59所述的AI模型传输方法,其中,所述模型接收端设备的属性信息,包括:
    所述模型接收端设备的地址信息;
    所述模型接收端设备的标识信息;
    所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
  62. 根据权利要求59所述的AI模型传输方法,其中,所述模型转化要求信息包括:
    模型转化时延要求信息;
    模型转化准确率要求信息。
  63. 根据权利要求58-62中任一项所述的AI模型传输方法,其中,所述方法还包括:
    模型发送端设备接收模型转化平台发送的第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作。
  64. 根据权利要求58-62中任一项所述的AI模型传输方法,其中,所述模型发送端设备向模型转化平台发送第三请求消息之前,所述方法还包括:
    所述模型发送端设备接收第三方设备发送的第四请求消息,所述第四请求消息用于请求所述模型发送端设备向所述模型转化平台发送第二AI模型。
  65. 一种AI模型传输方法,包括:
    模型转化平台接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求 所述模型转化平台将第二AI模型转化为第一AI模型;
    所述模型转化平台根据所述第三请求消息,执行第二操作;
    其中,所述执行第二操作包括以下至少一项:
    确定是否能够完成对所述第二AI模型的转化操作;
    向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
    将第二AI模型转化为第一AI模型;
    向模型接收端设备发送所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  66. 根据权利要求65所述的AI模型传输方法,其中,所述第三请求消息包括以下至少一项:
    所述第二AI模型;
    模型发送端设备的属性信息;
    模型接收端设备的属性信息;
    模型转化要求信息。
  67. 根据权利要求66所述的AI模型传输方法,其中,所述模型发送端设备的属性信息,包括:
    模型发送端设备的地址信息;
    模型发送端设备的标识信息;
    模型发送端设备支持的AI框架信息,所述模型发送端设备支持的AI框架信息与所述第二模型表示方法对应。
  68. 根据权利要求66所述的AI模型传输方法,其中,所述模型接收端设备的属性信息,包括:
    所述模型接收端设备的地址信息;
    所述模型接收端设备的标识信息;
    所述模型接收端设备支持的AI框架信息,所述模型接收端设备支持的AI框架信息与所述第一模型表示方法对应。
  69. 根据权利要求66所述的AI模型传输方法,其中,所述模型转化要求信息包括:
    模型转化时延要求信息;
    模型转化准确率要求信息。
  70. 一种AI模型传输装置,包括:
    第一发送单元,用于发送第一请求消息,所述第一请求消息用于请求第一AI模型,所述第一请求消息包括AI模型描述信息;
    第一接收单元,用于接收第一响应消息;
    其中,所述第一响应消息包括以下其中之一:
    所述第一AI模型,所述第一AI模型是来自模型发送端设备的第二AI模型经过模型转化平台转化后得到的模型,所述第一AI模型能够被所述模型接收端设备使用;
    第一指示信息,用于指示模型发送端设备无法提供所述第二AI模型;
    第二指示信息,用于指示模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  71. 一种AI模型传输装置,包括:
    第二接收单元,用于接收AI模型请求消息,所述AI模型请求消息用于请求第一AI模型,所述AI模型请求消息包括AI模型描述信息;
    第一执行单元,用于根据所述AI模型请求消息,执行第一操作;
    其中,所述执行第一操作包括以下至少一项:
    选择或训练生成第二AI模型,并向模型转化平台发送所述第二AI模型,其中所述第二AI模型经过模型转化平台转化后得到所述第一AI模型,所述第一AI模型能够被模型接收端设备使用;
    发送第一指示信息或第二指示信息,所述第一指示信息用于指示所述模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  72. 一种AI模型传输装置,包括:
    第四接收单元,用于接收模型发送端设备发送的第二AI模型;
    第一处理单元,用于将所述第二AI模型转化为第一AI模型,并将所述第一AI模型发送至模型接收端设备;或者,
    第五发送单元,用于向所述模型接收端设备发送第一指示信息或第二指示信息,所述第一指示信息用于指示模型发送端设备无法提供所述第二AI模型,所述第二指示信息用于指示所述模型转化平台无法将所述第二AI模型转换成所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  73. 一种AI模型传输装置,包括:
    第十发送单元,用于向模型转化平台发送第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
    其中,所述第二AI模型对应第二模型表示方法,所述第一AI模型对应第一模型表示 方法。
  74. 一种AI模型传输装置,包括:
    第十接收单元,用于接收模型发送端设备发送的第三请求消息,所述第三请求消息用于请求所述模型转化平台将第二AI模型转化为第一AI模型;
    第二执行单元,用于根据所述第三请求消息,执行第二操作;
    其中,所述执行第二操作包括以下至少一项:
    确定是否能够完成对所述第二AI模型的转化操作;
    向所述模型发送端设备发送第二模型转化响应消息,所述第二模型转化响应消息用于指示所述模型转化平台是否能够完成对所述第二AI模型的转化操作;
    将第二AI模型转化为第一AI模型;
    向模型接收端设备发送所述第一AI模型;
    其中,所述第一AI模型对应第一模型表示方法,所述第二AI模型对应第二模型表示方法。
  75. 一种模型接收端设备,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如权利要求1至17任一项所述的AI模型传输方法的步骤。
  76. 一种模型发送端设备,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如权利要求18至37任一项所述的AI模型传输方法的步骤,或实现如权利要求58至64任一项所述的AI模型传输方法的步骤。
  77. 一种模型转化平台,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如权利要求38至57任一项所述的AI模型传输方法的步骤,或实现如权利要求65至69任一项所述的AI模型传输方法的步骤。
  78. 一种可读存储介质,所述可读存储介质上存储程序或指令,所述程序或指令被处理器执行时实现如权利要求1至17任一项所述的AI模型传输方法的步骤,或者实现如权利要求18至37任一项所述的AI模型传输方法的步骤,或者实现如权利要求38至57任一项所述的AI模型传输方法的步骤,或实现如权利要求58至64任一项所述的AI模型传输方法的步骤,或实现如权利要求65至69任一项所述的AI模型传输方法的步骤。
PCT/CN2022/142586 2021-12-28 2022-12-28 Ai模型传输方法、装置、设备及存储介质 WO2023125594A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111627453.3 2021-12-28
CN202111627453.3A CN116367187A (zh) 2021-12-28 2021-12-28 Ai模型传输方法、装置、设备及存储介质

Publications (1)

Publication Number Publication Date
WO2023125594A1 true WO2023125594A1 (zh) 2023-07-06

Family

ID=86925438

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/142586 WO2023125594A1 (zh) 2021-12-28 2022-12-28 Ai模型传输方法、装置、设备及存储介质

Country Status (2)

Country Link
CN (1) CN116367187A (zh)
WO (1) WO2023125594A1 (zh)

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569106A (zh) * 2019-08-27 2019-12-13 Oppo广东移动通信有限公司 代码加载方法、装置、电子设备及计算机可读介质
US20200042832A1 (en) * 2019-09-09 2020-02-06 Lg Electronics Inc. Artificial intelligence apparatus and method for updating artificial intelligence model
CN111290778A (zh) * 2020-02-06 2020-06-16 网易(杭州)网络有限公司 Ai模型的包装方法、平台及电子设备
CN112965804A (zh) * 2021-03-25 2021-06-15 深圳市优必选科技股份有限公司 一种处理信息的方法、装置、终端、系统以及存储介质
CN113747462A (zh) * 2020-05-30 2021-12-03 华为技术有限公司 一种信息处理方法及相关设备
CN113849179A (zh) * 2020-06-28 2021-12-28 华为技术有限公司 一种人工智能模型发布方法及装置

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110569106A (zh) * 2019-08-27 2019-12-13 Oppo广东移动通信有限公司 代码加载方法、装置、电子设备及计算机可读介质
US20200042832A1 (en) * 2019-09-09 2020-02-06 Lg Electronics Inc. Artificial intelligence apparatus and method for updating artificial intelligence model
CN111290778A (zh) * 2020-02-06 2020-06-16 网易(杭州)网络有限公司 Ai模型的包装方法、平台及电子设备
CN113747462A (zh) * 2020-05-30 2021-12-03 华为技术有限公司 一种信息处理方法及相关设备
CN113849179A (zh) * 2020-06-28 2021-12-28 华为技术有限公司 一种人工智能模型发布方法及装置
CN112965804A (zh) * 2021-03-25 2021-06-15 深圳市优必选科技股份有限公司 一种处理信息的方法、装置、终端、系统以及存储介质

Also Published As

Publication number Publication date
CN116367187A (zh) 2023-06-30

Similar Documents

Publication Publication Date Title
WO2022022024A1 (zh) 训练样本构建方法、装置、设备及计算机可读存储介质
JP2022554384A (ja) ネットワークデータ分析方法、機能エンティティ及び電子機器
WO2023125594A1 (zh) Ai模型传输方法、装置、设备及存储介质
WO2020192673A1 (zh) 资源配置方法、资源确定方法、网络侧设备和终端
WO2023093609A1 (zh) 物联网设备的会话建立方法及装置
WO2021136497A1 (zh) 多播业务处理方法、多播业务配置方法及通信设备
WO2023125879A1 (zh) 数据处理的方法、装置及通信设备
WO2023125747A1 (zh) 模型训练方法、装置及通信设备
WO2023125869A1 (zh) Ai网络信息传输方法、装置及通信设备
WO2023213270A1 (zh) 模型训练处理方法、装置、终端及网络侧设备
WO2023125934A1 (zh) Ai网络信息传输方法、装置及通信设备
WO2023098535A1 (zh) 信息交互方法、装置及通信设备
WO2023125870A1 (zh) 通信方法、装置、通信设备及可读存储介质
WO2023207980A1 (zh) 模型信息获取方法、发送方法、装置、节点和储存介质
WO2023186099A1 (zh) 信息反馈方法、装置及设备
WO2023125932A1 (zh) Ai网络信息传输方法、装置及通信设备
WO2023246584A1 (zh) 算力处理方法、装置及通信设备
WO2023185822A1 (zh) 网元注册方法、模型确定方法、装置、网元、通信系统及存储介质
WO2023143361A1 (zh) 能力信息上报方法、装置及终端
WO2023246756A1 (zh) 算力服务方法、装置、终端及核心网设备
WO2023098586A1 (zh) 信息交互方法、装置及通信设备
WO2024027582A1 (zh) 模型更新方法、装置、通信设备及可读存储介质
WO2023125760A1 (zh) 模型训练方法、装置及通信设备
WO2023088268A1 (zh) Ai信息的传输方法和设备
WO2024099111A1 (zh) Ai服务的处理方法及设备

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22914846

Country of ref document: EP

Kind code of ref document: A1