WO2024016363A1 - 异构人工智能ai框架的模型交互方法、装置及系统 - Google Patents

异构人工智能ai框架的模型交互方法、装置及系统 Download PDF

Info

Publication number
WO2024016363A1
WO2024016363A1 PCT/CN2022/107536 CN2022107536W WO2024016363A1 WO 2024016363 A1 WO2024016363 A1 WO 2024016363A1 CN 2022107536 W CN2022107536 W CN 2022107536W WO 2024016363 A1 WO2024016363 A1 WO 2024016363A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
framework
information
terminal device
network device
Prior art date
Application number
PCT/CN2022/107536
Other languages
English (en)
French (fr)
Inventor
乔雪梅
牟勤
Original Assignee
北京小米移动软件有限公司
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 北京小米移动软件有限公司 filed Critical 北京小米移动软件有限公司
Priority to CN202280002408.5A priority Critical patent/CN117751563A/zh
Priority to PCT/CN2022/107536 priority patent/WO2024016363A1/zh
Publication of WO2024016363A1 publication Critical patent/WO2024016363A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L27/00Modulated-carrier systems

Definitions

  • the present disclosure relates to the field of communication technology, and in particular to a model interaction method, device and system for a heterogeneous artificial intelligence AI framework.
  • model training may be required (for example, model training on the network device side, or model training on the terminal device side). If the computing energy consumption of the terminal device, or support After the computing power falls below a certain threshold, AI model training may not be completed quickly, so model training can be performed on network devices.
  • the model frameworks supported by the network device and the terminal device may be different, which affects the deployment of the artificial intelligence AI model trained based on the model framework and affects the interaction effect of the artificial intelligence AI model.
  • Embodiments of the present disclosure provide a model interaction method, device and system for a heterogeneous artificial intelligence AI framework, which can be applied in the field of communication technology to effectively avoid the impact of differences in model frameworks between network equipment and terminal equipment on artificial intelligence AI model interaction. , which can effectively ensure the transmission accuracy of the artificial intelligence AI model and improve the interactive effect of the artificial intelligence AI model.
  • embodiments of the present disclosure provide a model interaction method of a heterogeneous artificial intelligence AI framework, which is executed by a network device.
  • the method includes: determining that the terminal device supports first framework information, wherein the first framework information includes the Describe the first AI model framework supported by the terminal device; determine the second AI model framework.
  • the first frame information further includes any of the following:
  • the terminal device supports deployment framework information
  • the terminal device supports converted frame information.
  • the method further includes:
  • determining the second model framework includes:
  • the second model frame is determined based on the first frame information and the plurality of first candidate frame information.
  • determining the second model framework based on the first framework information and the plurality of first candidate framework information includes:
  • the candidate frame to which the first candidate frame information belongs is used as the second model frame; wherein,
  • the first candidate frame information is the same as the first frame information; or,
  • the candidate frame to which the first candidate frame information belongs supports model conversion.
  • the method further includes:
  • the first candidate framework supports model conversion, and the first AI model is generated based on the first candidate framework
  • the first AI model is not sent to the terminal device, the network device does not support the intermediate model representation framework, and the network device does not support model conversion.
  • the method further includes:
  • the network device meets at least one of the following:
  • the first supported candidate framework does not support model conversion
  • the method further includes:
  • the network device satisfies one of the following:
  • the supported first candidate frame information is the same as the first frame information
  • the method further includes:
  • the method further includes:
  • a first AI model is sent to the terminal device according to the first determination result and/or the second determination result.
  • the method further includes:
  • the indicated content includes at least one of the following:
  • the method further includes:
  • selecting one first frame information from multiple first frame information further includes:
  • the selected first frame information select one first frame information from a plurality of first frame information; wherein the selected first frame information includes the selected first model frame, and the selected first frame information Is any of the following:
  • First framework information that is different from the first candidate framework information, and the first candidate framework to which the first candidate framework information belongs supports converting the model into a model supported by the selected first framework information.
  • the method further includes:
  • the first AI model is converted into a fourth AI model based on the selected first model framework, and the second model framework determined by the selected first framework information supports model conversion, and the first AI model is formed by the Network equipment is generated based on the second model framework;
  • sending the first AI model to the terminal device according to the first determination result and/or the second determination result includes:
  • the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device supports model conversion;
  • the first AI model is sent to the terminal device.
  • sending the first AI model to the terminal device according to the first determination result and/or the second determination result includes:
  • the first AI model is processed based on the intermediate model representation framework to obtain an AI model file, the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network The device does not support model conversion;
  • the method further includes:
  • embodiments of the present disclosure provide another model interaction method of a heterogeneous artificial intelligence AI framework, which is executed by a terminal device.
  • the method includes: indicating supported first framework information to a network device, wherein the first framework
  • the information includes the first AI model framework supported by the terminal device.
  • the first frame information includes any of the following:
  • the terminal device supports deployment framework information
  • the terminal device supports converted frame information.
  • the method further includes:
  • determining the first framework information supported by the terminal device includes:
  • the supported first framework information is determined.
  • the method further includes:
  • Receive the second AI model sent by the network device wherein the second AI model is obtained by converting the first AI model by the network device based on the first framework information or the intermediate model representation framework, and the first AI Models are generated by the network device.
  • the method further includes:
  • Receive second indication information sent by the network device where the second indication information is used to indicate that the cell to which the network device belongs does not support AI model deployment based on the first framework information.
  • the method further includes:
  • the terminal device satisfies one of the following:
  • the supported first frame information is the same as the first candidate frame information supported by the network device;
  • the supported first frame information is different from the first candidate frame information supported by the network device.
  • the method further includes:
  • the terminal device meets the following requirements: the number of supported first frame information is multiple.
  • the method further includes:
  • the method further includes:
  • the method further includes:
  • Receive an AI model file sent by the network device wherein the AI model file is obtained by the network device processing a first AI model based on the intermediate model representation framework, and the first AI model is obtained by the network device based on the first AI model.
  • Two model frames are generated, and the second model frame is determined based on the first frame information.
  • the method further includes:
  • embodiments of the present disclosure provide a model interaction method of a heterogeneous artificial intelligence AI framework, which is executed by a network device.
  • the method includes: indicating supported second framework information to a terminal device, wherein the second framework
  • the information includes the second AI model framework supported by the network device.
  • the second frame information includes any of the following:
  • the network device supports deployment framework information
  • the network device supports converted frame information.
  • the method further includes:
  • embodiments of the present disclosure provide another model interaction method of heterogeneous artificial intelligence AI framework, which is executed by a terminal device.
  • the method includes: determining second framework information supported by the network device, wherein the second framework The information includes the second AI model framework supported by the network device.
  • the second frame information includes any of the following:
  • the network device supports deployment framework information
  • the network device supports converted frame information.
  • the method further includes:
  • first indication information indicates first framework information and indicates that the first AI model does not carry model format information
  • a second AI model is sent to the network device, where the second AI model is generated by the terminal device.
  • embodiments of the present disclosure provide a communication device, which has the function of implementing some or all of the network equipment in the method described in the first aspect or the method described in the third aspect, such as the functions of the communication device. It may have the functions of some or all embodiments of the present disclosure, or may have the functions of independently implementing any one of the embodiments of the present disclosure.
  • the functions described can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the structure of the communication device may include a transceiver module and a processing module, and the processing module is configured to support the communication device to perform corresponding functions in the above method.
  • the transceiver module is used to support communication between the communication device and other devices.
  • the communication device may further include a storage module coupled to the transceiver module and the processing module, which stores necessary computer programs and data for the communication device.
  • the processing module may be a processor
  • the transceiver module may be a transceiver or a communication interface
  • the storage module may be a memory
  • embodiments of the present disclosure provide another communication device, which has the ability to implement part or all of the functions of the terminal device in the method example described in the second aspect, or the method example described in the fourth aspect, such as communication
  • the function of the device may have the functions of some or all embodiments of the present disclosure, or may have the function of independently implementing any one of the embodiments of the present disclosure.
  • the functions described can be implemented by hardware, or can be implemented by hardware executing corresponding software.
  • the hardware or software includes one or more units or modules corresponding to the above functions.
  • the structure of the communication device may include a transceiver module and a processing module, and the processing module is configured to support the communication device in performing corresponding functions in the above method.
  • the transceiver module is used to support communication between the communication device and other devices.
  • the communication device may further include a storage module coupled to the transceiver module and the processing module, which stores necessary computer programs and data for the communication device.
  • an embodiment of the present disclosure provides a communication device.
  • the communication device includes a processor.
  • the processor calls a computer program in a memory, it executes the model interaction of the heterogeneous artificial intelligence AI framework described in the first aspect. method, or execute the model interaction method of the heterogeneous artificial intelligence AI framework described in the third aspect above.
  • an embodiment of the present disclosure provides a communication device.
  • the communication device includes a processor.
  • the processor calls a computer program in a memory, it executes the model interaction of the heterogeneous artificial intelligence AI framework described in the second aspect. method, or execute the model interaction method of the heterogeneous artificial intelligence AI framework described in the fourth aspect above.
  • an embodiment of the present disclosure provides a communication device.
  • the communication device includes a processor and a memory, and a computer program is stored in the memory; the processor executes the computer program stored in the memory, so that the communication device executes The model interaction method of the heterogeneous artificial intelligence AI framework described in the above first aspect, or the model interaction method of the heterogeneous artificial intelligence AI framework described in the above third aspect.
  • an embodiment of the present disclosure provides a communication device.
  • the communication device includes a processor and a memory, and a computer program is stored in the memory; the processor executes the computer program stored in the memory, so that the communication device executes The model interaction method of the heterogeneous artificial intelligence AI framework described in the above second aspect, or the model interaction method of the heterogeneous artificial intelligence AI framework described in the above fourth aspect.
  • an embodiment of the present disclosure provides a communication device.
  • the device includes a processor and an interface circuit.
  • the interface circuit is used to receive code instructions and transmit them to the processor.
  • the processor is used to run the code instructions to cause
  • the device executes the model interaction method of the heterogeneous artificial intelligence AI framework described in the first aspect, or executes the model interaction method of the heterogeneous artificial intelligence AI framework described in the third aspect.
  • an embodiment of the present disclosure provides a communication device.
  • the device includes a processor and an interface circuit.
  • the interface circuit is used to receive code instructions and transmit them to the processor.
  • the processor is used to run the code instructions to cause
  • the device executes the model interaction method of the heterogeneous artificial intelligence AI framework described in the above second aspect, or executes the model interaction method of the heterogeneous artificial intelligence AI framework described in the above fourth aspect.
  • embodiments of the present disclosure provide a communication system, which includes the communication device described in the fifth aspect and the communication device described in the sixth aspect, or the system includes the communication device described in the seventh aspect and The communication device according to the eighth aspect, or the system includes the communication device according to the ninth aspect and the communication device according to the tenth aspect, or the system includes the communication device according to the eleventh aspect and the twelfth aspect The communication device described in this aspect.
  • embodiments of the present disclosure provide a computer-readable storage medium for storing instructions used by the above-mentioned network device.
  • the network device is caused to execute the above-mentioned first aspect. method, or perform the method described in the third aspect above.
  • embodiments of the present disclosure provide a readable storage medium for storing instructions used by the above-mentioned terminal device. When the instructions are executed, the terminal device is caused to perform the method described in the second aspect. , or perform the method described in the fourth aspect above.
  • the present disclosure also provides a computer program product including a computer program, which, when run on a computer, causes the computer to perform the method described in the first aspect, or to perform the method described in the third aspect.
  • the present disclosure also provides a computer program product including a computer program, which, when run on a computer, causes the computer to perform the method described in the second aspect, or to perform the method described in the fourth aspect.
  • the present disclosure provides a chip system.
  • the chip system includes at least one processor and an interface, and is used to support a network device to implement the functions involved in the first aspect or the third aspect, for example, determining or processing the above method. At least one of the data and information involved.
  • the chip system further includes a memory, and the memory is used to store necessary computer programs and data for the network device.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • the present disclosure provides a chip system, which includes at least one processor and an interface for supporting a terminal device to implement the functions involved in the second aspect or the fourth aspect, for example, determining or processing the above method. At least one of the data and information involved.
  • the chip system further includes a memory, and the memory is used to store necessary computer programs and data for the terminal device.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • the present disclosure provides a computer program that, when run on a computer, causes the computer to perform the method described in the first aspect, or to perform the method described in the third aspect.
  • the present disclosure provides a computer program that, when run on a computer, causes the computer to perform the method described in the second aspect, or to perform the method described in the fourth aspect.
  • the second AI model framework is determined by determining that the terminal device supports the first framework information, where the first framework information includes the first AI model framework supported by the terminal device.
  • Figure 1 is a schematic architectural diagram of a communication system provided by an embodiment of the present disclosure
  • Figure 2 is a schematic flowchart of a model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 3 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 4 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 5 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 6 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 7 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 8 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 9 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • Figure 10 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 11 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 12 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 13 is a schematic flow chart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure
  • Figure 14 is a schematic structural diagram of a communication device provided by an embodiment of the present disclosure.
  • Figure 15 is a schematic structural diagram of another communication device provided by an embodiment of the present disclosure.
  • Figure 16 is a schematic structural diagram of a chip provided by an embodiment of the present disclosure.
  • first, second, third, etc. may be used to describe various information in the embodiments of the present disclosure, the information should not be limited to these terms. These terms are only used to distinguish information of the same type from each other.
  • first information may also be called second information, and similarly, the second information may also be called first information.
  • the words "if” and “if” as used herein may be interpreted as “when” or “when” or “in response to determining.”
  • Artificial intelligence refers to a new technical science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence. Artificial intelligence is a branch of computer science that attempts to understand the essence of intelligence and produce a new intelligent machine that can respond in a manner similar to human intelligence. Research in this field includes robotics, language recognition, image recognition, Natural language processing and expert systems, etc.
  • the model framework can provide developers with an interface (mathematical operation) for building neural networks, automatically train the neural network (perform reverse derivation, and approximate the optimal value), and obtain a neural network model (approximation function) for solving classification problems. , regression, and fitting problems to achieve application scenarios such as target classification and speech recognition.
  • FIG. 1 is a schematic architectural diagram of a communication system provided by an embodiment of the present disclosure.
  • the communication system may include but is not limited to one network device and one terminal device.
  • the number and form of devices shown in Figure 1 are only for examples and do not constitute a limitation on the embodiments of the present disclosure. In actual applications, two or more devices may be included.
  • the communication system shown in Figure 1 includes a network device 101 and a terminal device 102 as an example.
  • LTE long term evolution
  • 5th generation fifth generation
  • 5G new radio (NR) system 5th generation new radio
  • the network device 101 in the embodiment of the present disclosure is an entity on the network side that is used to transmit or receive signals.
  • the network device 101 can be an evolved base station (evolved NodeB, eNB), a transmission point (transmission reception point, TRP), a next generation base station (next generation NodeB, gNB) in an NR system, or other base stations in future mobile communication systems. Or access nodes in wireless fidelity (WiFi) systems, etc.
  • eNB evolved NodeB
  • TRP transmission reception point
  • gNB next generation base station
  • WiFi wireless fidelity
  • the embodiments of the present disclosure do not limit the specific technologies and specific equipment forms used by network equipment.
  • the network equipment provided by the embodiments of the present disclosure may be composed of a centralized unit (CU) and a distributed unit (DU).
  • the CU may also be called a control unit (control unit).
  • CU-DU is used.
  • the structure can separate the protocol layers of network equipment, such as base stations, and place some protocol layer functions under centralized control on the CU. The remaining part or all protocol layer functions are distributed in the DU, and the CU centrally controls the DU.
  • the terminal device 102 in the embodiment of the present disclosure is an entity on the user side that is used to receive or transmit signals, such as a mobile phone.
  • Terminal equipment can also be called terminal equipment (terminal), user equipment (user equipment, UE), mobile station (mobile station, MS), mobile terminal equipment (mobile terminal, MT), etc.
  • the terminal device can be a car with communication functions, a smart car, a mobile phone, a wearable device, a tablet computer (Pad), a computer with wireless transceiver functions, a virtual reality (VR) terminal device, an augmented reality (augmented reality (AR) terminal equipment, wireless terminal equipment in industrial control, wireless terminal equipment in self-driving, wireless terminal equipment in remote medical surgery, smart grid ( Wireless terminal equipment in smart grid, wireless terminal equipment in transportation safety, wireless terminal equipment in smart city, wireless terminal equipment in smart home, etc.
  • VR virtual reality
  • AR augmented reality
  • the embodiments of the present disclosure do not limit the specific technology and specific equipment form used by the terminal equipment.
  • FIG. 2 is a schematic flowchart of a model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the method is executed by a network device.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied to network equipment, such as evolved base stations, transmission points, next-generation base stations in NR systems, base stations in other future mobile communication systems, or wireless fidelity There are no restrictions on access nodes in the system, etc.
  • the method may include but is not limited to the following steps:
  • S102 Determine that the terminal device supports the first framework information, where the first framework information includes the first AI model framework supported by the terminal device.
  • AI Artificial Intelligence
  • the first artificial intelligence AI model refers to the artificial intelligence AI model generated by the network device side based on the second model framework.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks.
  • the first framework information also includes any of the following: framework information that the terminal device supports deployment, and framework information that the terminal device supports conversion. Thus, , which can effectively improve the comprehensiveness of the representation of the first frame information for the frame information supported by the terminal device.
  • the framework can automatically train the initial model (for example, perform reverse derivation and approximate the optimal value) to obtain an AI model that meets expected needs.
  • the framework information may be information used to describe the framework, such as framework type, framework name, framework version, etc.
  • the first framework information refers to relevant information describing the framework supported by the terminal device, for example, the framework supported by the terminal device.
  • the framework type, framework name, framework version, etc. of the supported framework is information used to describe the framework, such as framework type, framework name, framework version, etc.
  • the terminal device can indicate to the network device which model frameworks the terminal device supports parsing.
  • the framework of the terminal device is A, but it indicates that it can parse the model generated by the network device based on the B framework. Then the network device can directly parse the model based on the B framework. The generated model is sent to the terminal device.
  • the terminal device can also report that it supports parsing multiple frames B, C, D, E, and F.
  • the network device matches the frame C, it can further indicate to the terminal device that the supported frame type is C.
  • the first AI model framework may refer to an AI model framework supported by the terminal device side.
  • the second AI model framework may refer to the AI model framework supported by the network device side.
  • the framework may correspond to the above-mentioned first framework information, and the correspondence may be the same type, the same version, or matching.
  • the types may be different, but framework conversion is supported, or the types may be the same but versions are different, but version compatibility is supported, etc. There is no restriction on this.
  • the number of first frame information supported by the terminal device may be multiple, and the number of frame information supported by the network device may also be multiple, and there may be one or more first frame information.
  • the corresponding second model framework may be multiple, and the number of frame information supported by the network device may also be multiple, and there may be one or more first frame information.
  • the framework information supported by the terminal device side and the network device may be different, and the model types supported by different frameworks may also be different. Therefore, in the embodiment of the present disclosure, by determining that the terminal device supports the first Framework information, wherein the first framework information includes a first AI model framework supported by the terminal device, and determining the second AI model framework can effectively improve the adaptability of the obtained second AI model framework and the framework supported by the terminal device.
  • the embodiment of the present disclosure also provides a model interaction method for heterogeneous artificial intelligence AI framework, which determines the first framework information supported by the terminal device. It can also determine the first framework information supported by the terminal device based on a predefined protocol to achieve Quickly obtain the first framework information supported by the terminal device based on the predefined protocol to improve the efficiency of interactive matching.
  • the predefined protocol may refer to a protocol that is set in advance based on application scenarios for the interaction process between the network device and the terminal device.
  • the predefined protocol can stipulate that the network device side and the terminal device side uniformly use the same AI framework, such as uniformly using the convolution architecture for feature extraction (Convolution Architecture For Feature Extraction, Caffe) specified in multiple versions. Version, such as caffe or caffe2. Alternatively, it can also be specified that the network device side and the terminal device side can use different AI frameworks.
  • the predefined protocol can stipulate that the base station side and the terminal side use a predefined matching AI framework.
  • the base station side uses pytorch (an open source Python machine learning library for natural language processing and other applications), and the terminal side uses pytorch.
  • Use a unified framework such as NCNN (a high-performance neural network forward computing framework optimized for terminal devices), etc.; or use a unified framework such as pytorch for model training, use a unified framework such as NCNN for model inference deployment, and model training is completed in When deploying, it is stipulated to support the conversion of models trained based on pytorch into models supported by NCNN.
  • model translation can be performed through a unified intermediate model representation framework, such as a predefined protocol stipulating the unified use of Open Neural Network Exchange (ONNX) or the unified use of other institutions such as the Institute of Electrical and Electronics Engineers (Institute of Electrical and Electronics Engineers) Engineers, IEEE) standardized AI model representation architecture.
  • a unified intermediate model representation framework such as a predefined protocol stipulating the unified use of Open Neural Network Exchange (ONNX) or the unified use of other institutions such as the Institute of Electrical and Electronics Engineers (Institute of Electrical and Electronics Engineers) Engineers, IEEE) standardized AI model representation architecture.
  • the model framework between the network device and the terminal device is effectively avoided.
  • the difference affects the interaction of the artificial intelligence AI model, which can effectively ensure the transmission accuracy of the artificial intelligence AI model and improve the interaction effect of the artificial intelligence AI model.
  • FIG 3 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of a heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 3, the method may include but is not limited to the following steps:
  • S103 Receive the first indication information sent by the terminal device.
  • the first indication information is used to instruct the terminal device to support the first framework information.
  • the instruction message may refer to a carrier for information exchange between the terminal device and the network device, and the instruction message may be an electrical signal or electromagnetic wave.
  • the first indication information is generated by the terminal device and can be used to indicate at least the first framework information supported by the terminal device.
  • the first indication information is used to instruct the terminal device to support the first framework information, which can provide a reliable reference basis for the network device to determine that the terminal device supports the first framework information.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks.
  • Determining the second model framework includes: determining that the network device supports multiple first candidate framework information, based on the first framework information and multiple third candidate frameworks.
  • One candidate frame information determines the second model frame, whereby the second model frame corresponding to the first frame information can be accurately and flexibly determined based on whether the first candidate frame information is the same as the first frame information.
  • the candidate frame refers to the model framework supported by the network device
  • the first candidate frame information refers to the frame information corresponding to the candidate frame.
  • the number of the first candidate frame information can be one or more, and there is no need to do this. limit.
  • the first candidate framework information supported by the network device may be determined, and the second candidate framework information corresponding to the first framework information may be determined.
  • the model framework provides reliable reference data.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks.
  • determining the second model framework includes: assigning the first candidate framework information to The candidate frame serves as the second model frame, where the first candidate frame information is the same as the first frame information, or the candidate frame to which the first candidate frame information belongs supports model conversion, which can effectively improve the reliability of the second model frame determination process.
  • the embodiment of the present disclosure also provides a model interaction method of heterogeneous artificial intelligence AI framework, which also includes: generating a first AI model based on the second model framework, sending the first AI model to the terminal device, and the third AI model supported by the network device.
  • the candidate frame information is the same as the first frame information to realize the model interaction process from the network device side to the terminal device side and ensure the adaptability between the first AI model and the terminal device.
  • the first AI model can be correctly parsed and deployed by the first framework. For example, it can be converted into an expression form supported by the first framework in the second model framework of the network device, or a model representation form supported by the second model framework can be directly generated. Then it is delivered to the terminal device, and then the model is converted by the corresponding model conversion function of the first model framework of the terminal device, and then deployed in the terminal device.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: converting the first AI model into a second AI model based on the first framework information, and sending the second AI model to the terminal device.
  • the first candidate framework supports model conversion
  • the first AI model is generated based on the first candidate framework, so that the second AI model can be adapted to the first framework information supported by the terminal device, and it is convenient for the terminal device to accurately perform the second AI model parse.
  • the second AI model may be an AI model obtained by converting the first AI model based on the first framework information.
  • the first candidate frame information may be generated based on the first frame information supported by the terminal device.
  • the first AI model is model converted to obtain a second AI model, so that the second AI model can be adapted to the first framework information supported by the terminal device, and facilitate the terminal device to accurately analyze the second AI model.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which also includes: converting the first AI model into a second AI model based on the intermediate model representation framework, and sending the second AI model to the terminal device.
  • both terminal devices and network devices support the intermediate model representation framework. Therefore, both terminal devices and network devices can support the intermediate model representation framework, and the deployment effect is effectively improved based on the intermediate model representation framework.
  • the intermediate model represents the framework, which can uniformly convert the models trained by each framework into the same format for storage, so as to facilitate the cross-platform and cross-framework transmission and deployment of AI models.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which also includes: not sending the first AI model to the terminal device, the network device does not support the intermediate model representation framework, and the network device does not support model conversion .
  • the first AI model may not be applicable to the terminal device, and the first AI model is not sent to the terminal device at this time model, which can effectively reduce resource consumption and effectively prevent the first AI model from causing interference to the terminal device.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: sending second indication information to the terminal device, where the second indication information is used to indicate that the cell to which the network device belongs does not support the first framework.
  • AI model deployment of information wherein the network device meets at least one of the following: the supported first candidate framework does not support model conversion, does not support model conversion, does not support the intermediate model representation framework, and can promptly indicate to the terminal device based on the second indication information
  • the cell to which the network device belongs does not support AI model deployment based on the first frame information, so that the terminal device can quickly take countermeasures.
  • the second indication information may be a message indicating that the cell to which the network device belongs does not support AI model inference based on the first framework information.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: sending second instruction information to a terminal device, where the second instruction information is used to instruct the terminal device to report the model to the network device, wherein, The network device meets one of the following conditions: the supported first candidate framework information is the same as the first framework information and supports model conversion. Therefore, the terminal device can be instructed to report the model to the network device in a timely manner based on the second instruction message.
  • Figure 4 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of a heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 4, the method may include but is not limited to the following steps:
  • S104 Send second instruction information to the terminal device.
  • the second instruction information is used to instruct the terminal device to select one first frame information from multiple first frame information.
  • the selected first frame information may refer to the frame information used for this interaction process among the plurality of first frame information.
  • the selected first frame information may be the one determined by the network device and suitable for the terminal device.
  • the frame information of model parsing is not limited.
  • the same first frame information can be directly used as the selected first frame information. If there is more than one identical first frame information, the base station can select one from the more than one identical first frame information and indicate it to the terminal device, or the base station can also select one from more than one identical first frame information based on predefined rules. Select one of the first frame information and indicate it to the terminal device. There is no restriction on this.
  • the second instruction information is used to instruct the terminal device to select one first frame information from multiple first frame information, and the terminal device can be accurately instructed based on the second instruction information.
  • the determination process of the first frame information effectively improves the reliability of the selected first frame information.
  • FIG. 5 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of a heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 5, the method may include but is not limited to the following steps:
  • S105 Determine whether the network device supports the intermediate model representation framework, and obtain the first determination result.
  • the first determination result may be, for example, that the network device supports the intermediate model representation framework, or that the network device does not support the intermediate model representation framework.
  • S205 Determine whether the network device supports model conversion, and obtain a second determination result.
  • the frame information used by the specific AI model does not need to be indicated to the terminal device; if the terminal device supports multiple intermediate model representation frameworks, the terminal device can be instructed again.
  • the network device is the frame information used; if the network device supports both conversion and the intermediate model representation framework supported by the terminal, the base station can decide on its own which method to deliver, or the priority can be determined by the pre-protocol definition; If the network device decides on its own, the network device can also indicate to the terminal device the specific framework information used by the AI model.
  • the second determination result may be, for example, that the network device supports model conversion, or that the network device does not support model conversion.
  • S305 Send the first AI model to the terminal device according to the first determination result and/or the second determination result.
  • the reporting method of the first AI model can be any of the following:
  • a message instructing the terminal device to generate a model and report it is sent to the terminal device.
  • This message may be called fourth instruction information.
  • the fourth instruction information instructs the terminal device to generate a model and report it.
  • the terminal device can generate an AI model based on the instructions of the fourth instruction information and based on the framework information that supports deployment, and report the AI model to the network device, without limitation.
  • the device generates a model and reports it.
  • the terminal device can generate an AI model based on the instruction of the fourth instruction information and based on the framework information that supports deployment, and report the AI model to the network device, and the network device performs model conversion on the received AI model, without limitation.
  • the terminal device can learn based on the second indication information that the network device cell does not support AI model inference based on the first framework information.
  • a message instructing the terminal device to convert the model and report it is sent to the terminal device. This message may also be indicated by the fourth instruction information. .
  • the terminal device can convert the AI model based on the instruction of the fourth instruction information and based on the framework information that supports conversion, obtain the converted AI model, and report the converted AI model to the network device, without limitation.
  • the fourth instruction information instructs the terminal device to report the model, which may include: instructing the terminal device to generate a model and report it, or instructing the terminal device to convert the model and report it, without limitation.
  • the terminal device can generate an AI model based on the instruction of the fourth instruction information, and report the AI model to the network device, and the network device performs model conversion on the received AI model, without limitation.
  • the terminal device can directly select one from multiple framework information that supports deployment to generate a model based on instructions from the network device, and there is no limit to this.
  • the network device can decide on its own to select one from multiple framework information that supports deployment and indicate it to the terminal device, or the network device can also select one from multiple framework information that supports deployment based on predefined rules and indicate it to the terminal device.
  • Terminal equipment there is no restriction on this.
  • the terminal device can directly select one from multiple frame information that supports conversion to convert the model based on instructions from the network device, and there is no restriction on this.
  • the network device can decide on its own to select one from multiple frame information that supports conversion and indicate it to the terminal device, or the network device can also select one from multiple frame information that supports conversion based on predefined rules and indicate it to the terminal device. Terminal equipment, there is no restriction on this.
  • Second indication information is sent to the terminal device, where the second indication information indicates that the network device cell does not support conversion based on the first AI model reasoning with frame information.
  • the network device can also make a decision on whether to send the AI model from the terminal device, and the network device can send the decision to the terminal device.
  • the terminal device performs corresponding operations based on the decision-making of the network device, or it can also support any other possible implementation manner, and there is no limit to this.
  • the first determination result is obtained by determining whether the network device supports the intermediate model representation framework
  • the second determination result is obtained by determining whether the network device supports model conversion, and according to the first determination result and/or the second determination result
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving a third AI model sent by a terminal device, where the third AI model is generated by the terminal device based on content indicated by the network device. Generated, wherein the indicated content includes at least one of the following: content indicated by the second indication information, whereby the AI model can be generated by the terminal device when the network device cell does not support AI model inference based on the first framework information, and sent to the network device side.
  • the third AI model refers to the AI model generated by the terminal device based on content instructed by the network device.
  • Figure 6 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of a heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 6, the method may include but is not limited to the following steps:
  • S106 Select one first frame information from multiple first frame information.
  • S206 Send third indication information to the terminal device, where the third indication information is used to indicate the selected first frame information to the terminal device.
  • the third indication message may be used to indicate the selected first frame information to the terminal device.
  • S306 Generate the first AI model based on the selected first framework information.
  • S406 Send the first AI model to the terminal device.
  • third indication information is sent to the terminal device by selecting one first frame information from multiple first frame information, where the third indication information is used to indicate the selected first frame information to the terminal device, based on The selected first frame information generates a first AI model, the first AI model is sent to the terminal device, and the selection process of the selected first frame information is accurately indicated based on the third instruction message to ensure the reliability of the generated first AI model.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks, selecting one first framework information from multiple first framework information, and further including: determining the first candidate framework information supported by the network device , select one first frame information from multiple first frame information according to the first candidate frame information, wherein the selected first frame information includes the selected first model frame, and the selected first frame information is any of the following : First frame information that is the same as the first candidate frame information, or first frame information that is different from the first candidate frame information, and the first candidate frame to which the first candidate frame information belongs supports converting the model into the selected first framework
  • the model supported by the information can effectively improve the practicality of the selected first frame information.
  • FIG. 7 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 7, the method may include but is not limited to the following steps:
  • S107 Convert the first AI model to a fourth AI model based on the selected first model frame.
  • the second model frame determined by the selected first frame information supports model conversion.
  • the first AI model is generated by the network device based on the second model. Frame generation.
  • the fourth AI model may be an AI model obtained by converting the first AI model based on the selected first model framework.
  • the degree of matching between the first AI model and the terminal device may be low, and converting the first AI model based on the selected first model framework can greatly improve the matching between the fourth AI model and the terminal device. Adaptability between devices.
  • S207 Send the fourth AI model to the terminal device.
  • the first AI model if the first AI model is not suitable for the terminal device, when the second model framework supports the model conversion function, the first AI model can be converted based on the selected first model framework. , to obtain a fourth AI model suitable for the terminal device, and send the obtained fourth AI model to the terminal device to complete model interaction between the network device and the terminal device.
  • the second model framework determined by the selected first framework information supports model conversion
  • the first AI model is configured by the network device Based on the generation of the second model framework
  • the fourth AI model is sent to the terminal device.
  • Mutual conversion between different AI models can be realized based on the selected first model framework, which can effectively improve model interaction and matching between network devices and terminal devices. efficiency.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which sends the first AI model to the terminal device according to the first determination result and/or the second determination result, including: determining model delivery method, the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device supports model conversion.
  • the first AI model is sent to the terminal device, whereby the network device can When multiple model delivery methods are met, the matching AI model delivery method is accurately indicated to ensure the reliability of the AI model delivery process.
  • network equipment can support the following processing logic:
  • Second instruction information when it is determined that the first candidate framework information supported by the network device is the same as the framework information supported by the terminal device, a message instructing the terminal device to generate a model and report it is sent to the terminal device.
  • This message may be called second instruction information.
  • the second instruction information may instruct the terminal device to generate a model and report it.
  • the device generates a model and reports it.
  • a message instructing the terminal device to convert the model and report it is sent to the terminal device. This message may also be indicated by the second instruction information.
  • the second instruction information instructs the terminal device to report the model, which may include: instructing the terminal device to generate a model and report it, or instructing the terminal device to convert the model and report it, without limitation.
  • the terminal device can directly select one from multiple framework information that supports deployment to generate a model based on instructions from the network device, and there is no limit to this.
  • the network device can decide on its own to select one from multiple framework information that supports deployment and indicate it to the terminal device, or the network device can also select one from multiple framework information that supports deployment based on predefined rules and indicate it to the terminal device.
  • Terminal equipment there is no restriction on this.
  • the terminal device can directly select one from multiple frame information that supports conversion to convert the model based on instructions from the network device, and there is no restriction on this.
  • the network device can decide on its own to select one from multiple frame information that supports conversion and indicate it to the terminal device, or the network device can also select one from multiple frame information that supports conversion based on predefined rules and indicate it to the terminal device. Terminal equipment, there is no restriction on this.
  • Second indication information is sent to the terminal device, where the second indication information indicates that the network device cell does not support conversion based on the first AI model reasoning with frame information.
  • the model delivery method refers to the method in which the AI model is sent from the network device side to the terminal device side.
  • the AI model can be delivered based on an intermediate model representation framework, or it can be delivered based on model conversion.
  • the network device can decide whether to interact and deploy the AI model based on the intermediate model representation framework or based on the model conversion. ;
  • the base station can also determine whether to use an intermediate model representation framework or perform AI model conversion based on predefined rules, or it can also use any other possible method without restrictions.
  • the base station instructs the terminal device how to deliver the final decision model.
  • FIG 8 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 8, the method may include but is not limited to the following steps:
  • S108 Process the first AI model based on the intermediate model representation framework to obtain the AI model file.
  • the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device does not support model conversion.
  • the AI model file refers to the file obtained by processing the first AI model through the intermediate model representation framework.
  • the format of the AI model file There are many possibilities for the format of the AI model file.
  • the first AI model is processed based on the intermediate model representation framework in a timely manner to obtain the AI model file, and Send the AI model file to the terminal device.
  • the AI model file is obtained by processing the first AI model based on the intermediate model representation framework.
  • the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device does not support model conversion.
  • the AI The model file is sent to the terminal device.
  • the network device supports the intermediate model representation framework but does not support model conversion
  • the first AI model can be converted in a timely manner to ensure the compatibility between the resulting AI model file and the terminal device.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: sending second instruction information to the terminal device, where the second instruction information indicates the AI model storage format of the AI model file, the AI The model file does not carry model format information.
  • the model format information may refer to the storage format corresponding to the AI model file.
  • Network equipment may support one or more AI frameworks. If it supports multiple AI frameworks, or only supports one but can achieve direct model conversion, such as TF->caffe, TF->NCNN, base station You can select a matching framework to generate an AI model or convert it into a suitable model and send it to the terminal; if the base station only supports one framework and cannot implement model conversion, it will indicate to the terminal that model inference is not supported in the cell.
  • the network device may determine that the terminal device supports the first framework information, where the first framework information includes the first AI model framework supported by the terminal device, and determine the second AI model framework,
  • the first frame information includes any of the following: frame information that the terminal device supports deployment, or frame information that the terminal device supports conversion.
  • the network device may receive first indication information sent by the terminal device, where the first indication information is used to instruct the terminal device to support the first framework information.
  • the network device may determine that the network device supports multiple first candidate framework information, and determine the second model framework based on the first framework information and the multiple first candidate framework information.
  • the network device may use the candidate frame to which the first candidate frame information belongs as the second model frame; wherein the first candidate frame information is the same as the first frame information; or the candidate frame to which the first candidate frame information belongs supports model conversion.
  • the network device can generate the first AI model based on the second model framework, and send the first AI model to the terminal device.
  • the first candidate framework information supported by the network device is the same as the first framework information, or, based on the first framework information,
  • the first AI model is converted into a second AI model, and the second AI model is sent to the terminal device.
  • the first candidate framework supports model conversion, and the first AI model is generated based on the first candidate framework, or the first AI model is generated based on the intermediate model representation framework.
  • the AI model is converted into a second AI model and the second AI model is sent to the terminal device. Both the terminal device and the network device support the intermediate model representation framework.
  • the first AI model is not sent to the terminal device and the network device does not support the intermediate model representation. framework, and the network device does not support model conversion.
  • the network device may send second indication information to the terminal device, and the second indication information is used to indicate that the cell to which the network device belongs does not support AI model deployment based on the first framework information, wherein the network device satisfies at least one of the following: supported third One candidate framework does not support model conversion, does not support model conversion, and does not support intermediate model representation frameworks.
  • the network device may send second instruction information to the terminal device, and the second instruction information is used to instruct the terminal device to report the model to the network device, wherein the network device satisfies one of the following: the supported first candidate framework information and the first framework The information is the same and model conversion is supported.
  • the network device may send second indication information to the terminal device, where the second indication information is used to instruct the terminal device to select one first frame information from multiple first frame information.
  • the network device may determine the second frame information corresponding to the first frame information based on a predefined protocol, and use the model frame to which the second frame information belongs as the second model frame.
  • the predefined protocol includes any of the following: defining the same frame information used by the terminal device, or defining frame information matching the use of the terminal device, or defining frame information corresponding to the model processing stage, or defining a representation frame based on an intermediate model Process model.
  • the predefined protocol can specify, for example, the specific types of AI frameworks that can be adopted:
  • Option 1 Provide that network devices and terminal devices use the same AI framework, such as caffe or caffe2.
  • Network equipment and terminal equipment can use different AI frameworks.
  • Solution 2-1 The predefined protocol stipulates that network equipment and terminal equipment use a predefined matching AI framework.
  • network equipment uses pytorch uniformly, and terminal equipment uses a unified framework such as NCNN; or a unified AI framework is used during model training.
  • Frameworks such as pytorch, the model inference deployment process uses a unified framework such as NCNN.
  • NCNN a unified framework
  • the model training is completed and deployed, it is stipulated to support the conversion of pytorch into an NCNN model.
  • Option 2-2 Carry out model translation through a unified intermediate model representation framework. For example, you can specify the unified use of onnx or the unified use of the AI model representation architecture standardized by other institutions such as IEEE.
  • the first frame information is the same as the second frame information, or the first frame information is different from the second frame information, or the first frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. .
  • model processing stage includes any of the following: model training stage, or model inference deployment stage.
  • the base station can determine and instruct the unique/matching one AI framework to the terminal to enable the AI framework. In addition, the base station needs to convert the AI model into multiple One of the matching AI frameworks.
  • the network device may send second instruction information to the terminal device, where the second instruction information is used to instruct the terminal device to select one first frame information from multiple first frame information.
  • the network device can determine whether the network device supports the intermediate model representation framework, obtain a first determination result, determine whether the network device supports model conversion, obtain a second determination result, and according to the first determination result and/or the second determination result, Send the first AI model to the terminal device.
  • the network device may convert the first AI model into a third AI model based on the selected first model framework when the second model framework determined based on the selected first framework information supports model conversion, where the first AI model The model is generated by the network device based on the second model framework, and the third AI model is sent to the terminal device.
  • the network device may generate the first AI model based on the selected first framework information, and send the first AI model to the terminal device.
  • the network device may receive a third AI model sent by the terminal device, where the third AI model is generated by the terminal device based on content indicated by the network device, where the indicated content includes at least one of the following: second indication information content of instructions.
  • the network device may select one first frame information from multiple first frame information and send third indication information to the terminal device, where the third indication information is used to indicate the selected first frame information to the terminal device, based on The selected first frame information generates a first AI model, and the first AI model is sent to the terminal device.
  • the network device may determine the first candidate frame information supported by the network device, and select one first frame information from multiple first frame information according to the first candidate frame information; wherein the selected first frame information includes the selected first frame information.
  • Select the first model frame, and the selected first frame information is any of the following: first frame information that is the same as the first candidate frame information, or first frame information that is different from the first candidate frame information, and the first candidate
  • the first candidate framework to which the framework information belongs supports converting the model into a model supported by the selected first framework information.
  • the network device can convert the first AI model into a fourth AI model based on the selected first model framework.
  • the second model framework determined by the selected first frame information supports model conversion, and the first AI model is configured by the network device. Based on the second model framework generation, the fourth AI model is sent to the terminal device.
  • network equipment can support the following processing logic:
  • Second instruction information when it is determined that the first candidate framework information supported by the network device is the same as the framework information supported by the terminal device, a message instructing the terminal device to generate a model and report it is sent to the terminal device.
  • This message may be called second instruction information.
  • the second instruction information instructs the terminal device to generate a model and report it.
  • the device generates a model and reports it.
  • a message instructing the terminal device to convert the model and report it is sent to the terminal device. This message may also be indicated by the second instruction information.
  • the second instruction information instructs the terminal device to report the model, which may include: instructing the terminal device to generate a model and report it, or instructing the terminal device to convert the model and report it, without limitation.
  • the terminal device can directly select one from multiple framework information that supports deployment to generate a model based on instructions from the network device, and there is no limit to this.
  • the network device can decide on its own to select one from multiple framework information that supports deployment and indicate it to the terminal device, or the network device can also select one from multiple framework information that supports deployment based on predefined rules and indicate it to the terminal device.
  • Terminal equipment there is no restriction on this.
  • the terminal device can directly select one from multiple frame information that supports conversion to convert the model based on instructions from the network device, and there is no restriction on this.
  • the network device can decide on its own to select one from multiple frame information that supports conversion and indicate it to the terminal device, or the network device can also select one from multiple frame information that supports conversion based on predefined rules and indicate it to the terminal device. Terminal equipment, there is no restriction on this.
  • the network device can receive the third AI model sent by the terminal device, where the third AI model is generated by the terminal device based on content indicated by the network device; wherein the indicated content includes at least one of the following: second indication information content of instructions.
  • the network device may issue a second instruction to the terminal device when it is determined that the first candidate frame information supported by the network device is different from the frame information supported by the terminal device for conversion, and the network device or the terminal device does not support the intermediate model representation framework.
  • Information wherein the second indication information indicates that the network device cell does not support AI model inference based on the first framework information.
  • the base station may need to indicate to the terminal which method to use for model delivery; in addition, if the base station cannot implement model conversion but supports the AI intermediate framework, it may Convert the model into a model format supported by the intermediate framework and deliver it to the terminal. At the same time, if the model file itself does not carry model format information, you may need to indicate the corresponding AI model storage format; if neither is supported, deliver it to the cell. Enabling AI model inference instructions is not supported, or instructions are implicitly provided by not delivering the AI model.
  • the network device can determine the model delivery method, the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device supports model conversion, and based on the model delivery method, the first AI model is sent to the terminal equipment.
  • the network device can process the first AI model based on the intermediate model representation framework to obtain the AI model file, the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device does not support model conversion, and the AI model is The file is sent to the terminal device.
  • the network device may send second indication information to the terminal device, where the second indication information indicates the AI model storage format of the AI model file, and the AI model file does not carry model format information.
  • the network device may determine whether the network device supports the intermediate model representation framework, obtain a first determination result, determine whether the network device supports model conversion, obtain a second determination result, and according to the first determination result and/or the second determination result, The first AI model is sent to the terminal device.
  • the network device may determine the model delivery method, and send the first AI model based on the model delivery method. to the terminal device.
  • the network device may process the first AI model based on the intermediate model representation framework to obtain the AI model file, and The AI model file is sent to the terminal device.
  • the network device may send second indication information to the terminal device, where the second indication information indicates the AI model storage format of the AI model file.
  • the network device may not send the first AI model to the terminal device when the first determination result indicates that the network device does not support the intermediate model representation framework, and the second determination result indicates that the network device does not support model conversion.
  • the network device may send the second indication information to the terminal device, wherein the second indication information indicates The network equipment cell does not support AI model inference based on the first frame information.
  • the network device may convert the first AI model into a fifth AI model based on the intermediate model representation framework, and send the fifth AI model to the terminal device.
  • the network device may determine that the network device supports the intermediate model representation framework when determining that the terminal device supports the intermediate model representation framework.
  • FIG. 9 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied in terminal devices, as shown in Figure As shown in 9, the method may include but is not limited to the following steps:
  • S109 Indicate the supported first framework information to the network device, where the first framework information includes the first AI model framework supported by the terminal device.
  • the terminal device may send the first framework information to the network device to indicate to the network device the first framework information supported by the terminal device.
  • a model interaction method of heterogeneous artificial intelligence AI framework is also provided, which can determine supported first framework information based on a predefined protocol.
  • the embodiment of the present disclosure also provides a model interaction method for heterogeneous artificial intelligence AI framework.
  • the predefined protocol includes any of the following: defining the same framework information used by the network device, or defining framework information matching the use of the network device. , or define the framework information corresponding to the model processing stage, or define the framework processing model based on the intermediate model representation.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework.
  • the network device When determining the supported first framework information based on the predefined protocol and/or the second framework information, it may be that the network device supports The second frame information is used as the first frame information.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks.
  • the supported first framework information based on the predefined protocol and/or the second framework information, it can also be used with the second framework information.
  • the frame information matching the frame information is used as the first frame information.
  • the embodiment of the present disclosure also provides a model interaction method for heterogeneous artificial intelligence AI framework.
  • determining the supported first framework information based on the predefined protocol and/or the second framework information it can also be used with the terminal device.
  • the frame information corresponding to the current model processing stage is used as the first frame information.
  • the network device by indicating the supported first framework information to the network device, where the first framework information includes the first AI model framework supported by the terminal device, it can effectively ensure that the AI model generated by the network device side is suitable for the terminal device side. applicability.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks.
  • the first framework information includes any of the following: framework information that the terminal device supports deployment, and framework information that the terminal device supports conversion.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: sending first instruction information to a network device, where the first instruction information is used to instruct the terminal device to support the first framework information, and the terminal device The first framework information supported by the terminal device may be indicated to the network device based on the first indication information.
  • FIG 10 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied to terminal devices, as shown in Figure As shown in 10, the method may include but is not limited to the following steps:
  • S110 Determine second framework information supported by the network device, where the second framework information includes a second model framework supported by the network device.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks, in which the first framework information and the second framework information are the same, or the first framework information and the second framework information are different, or the first framework information is different from the second framework information.
  • the frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. Therefore, the interaction process can be adapted to personalized application scenarios to effectively expand the model of the heterogeneous artificial intelligence AI framework.
  • the scope of application of interactive methods are provided.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, in which the model processing stage may include any of the following: a model training stage, or a model inference deployment stage.
  • S210 Determine the supported first framework information based on the predefined protocol and/or the second framework information.
  • the supported first framework is determined based on the predefined protocol and/or the second framework information. information, effectively improving the flexibility of the first framework information determination process.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving a first AI model sent by a network device, where the first AI model is generated by the network device, or receiving the first AI model from the network device.
  • the second AI model is sent, where the second AI model is converted by the network device based on the first frame information or the intermediate model representation framework.
  • the first AI model is generated by the network device, which can effectively improve the reception of the terminal device. Applicability of AI models.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving second indication information sent by a network device, where the second indication information is used to indicate that the cell to which the network device belongs does not support the first based on
  • the AI model deployment of the framework information enables the terminal device to timely obtain relevant information that the cell to which the network device belongs does not support AI model inference based on the first framework information based on the second indication information.
  • An embodiment of the present disclosure also provides a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving second instruction information sent by a network device, and the second instruction information is used to instruct the terminal device to report the model to the network device, wherein , the terminal device satisfies the following one: the supported first frame information is the same as the first candidate frame information supported by the network device, and the supported first frame information is different from the first candidate frame information supported by the network device. Therefore, it can
  • the second instruction information refers to the model reporting process of the terminal device.
  • the first candidate framework information is the same as the first framework information, for example, the framework type indicated by the first candidate framework information is the same as the framework type indicated by the first framework information, or the framework type and version number indicated by the first candidate framework information are the same as The framework type and version number indicated by the first framework information are the same, or the framework type indicated by the first candidate framework information is different from the framework type indicated by the first framework information, but both support conversion and expression based on the intermediate model framework, and no action is taken on this limit.
  • the first candidate framework information is different from the first framework information, for example, the framework type indicated by the first candidate framework information is different from the framework type indicated by the first framework information, or the framework type or version number indicated by the first candidate framework information, It is different from the framework type or version number indicated by the first framework information, or the framework type indicated by the first candidate framework information is different from the framework type indicated by the first framework information, but neither of them supports conversion expression based on the intermediate model framework. There is no restriction on this.
  • the embodiment of the present disclosure can accurately and flexibly determine the second model frame corresponding to the first frame information based on whether the first candidate frame information is the same as the first frame information.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving second instruction information sent by a network device, where the second instruction information is used to instruct the terminal device to select from multiple first Select one first frame information from the frame information; wherein the terminal device satisfies: the number of supported first frame information is multiple, thus the first frame information selection process of the terminal device can be accurately instructed based on the second indication information.
  • FIG 11 is a schematic flowchart of another model interaction method of a heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of a heterogeneous artificial intelligence AI framework in this embodiment can be applied to terminal devices, as shown in Figure As shown in 11, the method may include but is not limited to the following steps:
  • S111 Determine the frame information indicated by the network device according to the content indicated by the second indication information.
  • the third AI model may be an AI model generated by the terminal device based on the indicated frame information.
  • S311 Send the third AI model to the network device.
  • the third AI model can be effectively improved. Adaptability between AI models and network devices.
  • An embodiment of the present disclosure also provides a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving a fourth AI model sent by a network device.
  • the fourth AI model may refer to an AI model generated by the network device side.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving an AI model file sent by a network device, wherein the AI model file is processed by the network device based on the intermediate model representation framework for the first AI The model is obtained.
  • the first AI model is generated by the network device based on the second model framework.
  • the second model framework is determined based on the first framework information. Therefore, cross-frame transmission of the first AI model can be implemented based on the intermediate model representation framework.
  • An embodiment of the present disclosure also provides a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving second instruction information sent by the network device, where the second instruction information indicates the AI model storage format of the AI model file, The AI model file does not carry model format information. Therefore, the terminal device can accurately obtain the AI model storage format of the AI model file based on the second instruction information, which can effectively prevent the unknown storage format from affecting the storage effect of the AI model file.
  • the terminal reports the AI framework it supports to the base station, which can be one type or multiple types.
  • the terminal device may indicate supported first framework information to the network device, where the first framework information includes a first AI model framework supported by the terminal device.
  • the first frame information includes any of the following: frame information that the terminal device supports deployment, or frame information that the terminal device supports conversion.
  • the terminal device may send first indication information to the network device, where the first indication information is used to instruct the terminal device to support the first framework information.
  • the terminal device may receive the first artificial intelligence AI model sent by the network device, where the first AI model is generated by the network device.
  • the terminal device may receive the second AI model sent by the network device, where the second AI model is obtained by the network device converting the first AI model based on the first framework information, and the first AI model is generated by the network device.
  • the terminal device may receive second indication information sent by the network device, where the second indication information indicates that the cell to which the network device belongs does not support AI model inference based on the first framework information.
  • the terminal device may send first indication information to the network device, where the first indication information is used to instruct the terminal device to support the first framework information.
  • the terminal device may determine the second framework information supported by the network device, where the second framework information includes a second model framework supported by the network device, and determine the supported first framework information based on the predefined protocol and/or the second framework information.
  • the terminal device may receive a first AI model sent by the network device, where the first AI model is generated by the network device, or receive a second AI model sent by the network device, where the second AI model is generated by the network device based on the first
  • the frame information or intermediate model indicates that the frame is obtained by converting the first AI model, and the first AI model is generated by the network device.
  • the terminal device may receive second indication information sent by the network device, and the second indication information is used to indicate that the cell to which the network device belongs does not support AI model deployment based on the first framework information.
  • the terminal device can receive the second instruction information sent by the network device, and the second instruction information is used to instruct the terminal device to report the model to the network device, wherein the terminal device satisfies the following one: the supported first framework information is consistent with the supported first framework information by the network device.
  • the first candidate frame information is the same, but the supported first frame information is different from the first candidate frame information supported by the network device.
  • the terminal device may receive second instruction information sent by the network device, wherein the second instruction information is used to instruct the terminal device to select one first frame information from multiple first frame information, wherein the terminal device meets: the supported third frame information.
  • the number of information in a frame is multiple.
  • the terminal device can determine the frame information indicated by the network device according to the content indicated by the second indication information; generate a third AI model based on the indicated frame information, and send the third AI model to the network device.
  • the terminal device may receive the fourth AI model sent by the network device.
  • the terminal can report whether it supports the use of intermediate model representation framework, and which/several intermediate model representation frameworks are supported.
  • the terminal device can determine to support the intermediate model representation framework based on the predefined protocol.
  • the terminal device may determine that the intermediate model representation framework is supported, and send second indication information to the network device, where the second indication information indicates that the terminal device supports the intermediate model representation framework.
  • the terminal device can receive the AI model file sent by the network device, wherein the AI model file is obtained by the network device processing the first AI model based on the intermediate model representation framework, the first AI model is generated by the network device based on the second model framework, and the second AI model file is generated by the network device based on the second model framework.
  • the model frame is determined based on the first frame information.
  • the terminal reports the version information of the AI framework it supports, etc.
  • the terminal device may receive second indication information sent by the network device, where the second indication information indicates the AI model storage format of the AI model file, and the AI model file does not carry model format information.
  • the terminal device may determine the second framework information supported by the network device, and determine the supported first framework information based on the predefined protocol and/or the second framework information.
  • the predefined protocol includes any of the following: defining the same frame information used by the network device, or defining frame information matching the use of the network device, or defining frame information corresponding to the model processing stage, or defining a representation frame based on an intermediate model Process model.
  • the specific example of the predefined protocol can stipulate the specific type of AI framework that can be adopted:
  • Option 1 Provide that network devices and terminal devices use the same AI framework, such as caffe or caffe2.
  • Network equipment and terminal equipment can use different AI frameworks.
  • Solution 2-1 The predefined protocol stipulates that network equipment and terminal equipment use a predefined matching AI framework.
  • network equipment uses pytorch uniformly, and terminal equipment uses a unified framework such as NCNN; or a unified AI framework is used during model training.
  • Frameworks such as pytorch, the model inference deployment process uses a unified framework such as NCNN.
  • NCNN a unified framework
  • the model training is completed and deployed, it is stipulated to support the conversion of pytorch into an NCNN model.
  • Option 2-2 Carry out model translation through a unified intermediate model representation framework. For example, you can specify the unified use of onnx or the unified use of the AI model representation architecture standardized by other institutions such as IEEE.
  • the first frame information is the same as the second frame information, or the first frame information is different from the second frame information, or the first frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. .
  • the terminal device may use the second frame information supported by the network device as the first frame information, or use the frame information matching the second frame information as the first frame information, or use the frame information corresponding to the model processing stage in which the terminal device is.
  • the frame information is used as the first frame information.
  • model processing stage includes any of the following: model training stage and model inference deployment stage.
  • the terminal device can receive the AI model file sent by the network device, wherein the AI model file is obtained by the network device processing the first AI model based on the intermediate model representation framework, the first AI model is generated by the network device based on the second model framework, and the second AI model file is generated by the network device based on the second model framework.
  • the model frame is determined based on the first frame information.
  • FIG 12 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied to network devices, as shown in Figure As shown in 12, the method may include but is not limited to the following steps:
  • S112 Indicate supported second framework information to the terminal device, where the second framework information includes a second AI model framework supported by the network device.
  • the terminal device by indicating the supported second framework information to the terminal device, where the second framework information includes the second AI model framework supported by the network device, the terminal device can accurately obtain the second framework information supported by the network device.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework.
  • the second framework information includes any of the following: framework information that the network device supports deployment, and framework information that the network device supports conversion.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework, which further includes: receiving a first AI model sent by a terminal device, wherein the first AI model is generated by the terminal device based on the first framework information supported by the terminal device.
  • the included first model framework is generated, the first framework information is determined based on the second framework information, or the first indication information sent by the terminal device is received, wherein the first indication information indicates the first framework information and indicates the first AI model.
  • No model format information is carried, or second indication information is sent to the terminal device, where the second indication information indicates that the network device supports the intermediate model representation framework, or the second AI model sent by the terminal device is received, where the second AI model Generated by the terminal device.
  • the network device can indicate to the terminal all the AI frameworks it supports.
  • the network device may indicate the supported second framework information to the terminal device, where the second framework information includes a second AI model framework supported by the network device.
  • the second frame information includes any of the following: frame information that the network device supports deployment, and frame information that the network device supports conversion.
  • the base station matches the AI framework, and/or makes a decision on whether to deliver the model, and/or makes a decision on which AI framework to use.
  • a set of terminal-side and base-station-side AI framework conversion rules can be predefined.
  • the network device may receive the first indication information sent by the terminal device, where the first indication information indicates the first framework information and indicates that the first AI model does not carry model format information.
  • the first frame information includes any of the following: frame information that the terminal device supports deployment, or frame information that the terminal device supports conversion.
  • the predefined protocol can specify the specific AI framework types that can be adopted:
  • Option 1 Provide that network devices and terminal devices use the same AI framework, such as caffe or caffe2.
  • Network equipment and terminal equipment can use different AI frameworks.
  • Solution 2-1 The predefined protocol stipulates that network equipment and terminal equipment use a predefined matching AI framework.
  • network equipment uses pytorch uniformly, and terminal equipment uses a unified framework such as NCNN; or a unified AI framework is used during model training.
  • Frameworks such as pytorch, the model inference deployment process uses a unified framework such as NCNN.
  • NCNN a unified framework
  • the model training is completed and deployed, it is stipulated to support the conversion of pytorch into an NCNN model.
  • Option 2-2 Carry out model translation through a unified intermediate model representation framework. For example, you can specify the unified use of onnx or the unified use of the AI model representation architecture standardized by other institutions such as IEEE.
  • the first frame information is the same as the second frame information, or the first frame information is different from the second frame information, or the first frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. .
  • Model conversion can be completed on the terminal device side.
  • the network device may receive the first AI model sent by the terminal device, wherein the first AI model is generated by the terminal device based on the first model framework included in the supported first framework information, and the first framework information is based on the second framework information.
  • Network devices can indicate the intermediate model expression framework they support.
  • the network device may send second indication information to the terminal device, where the second indication information indicates that the network device supports the intermediate model representation framework.
  • the network device may receive the second AI model sent by the terminal device, where the second AI model is generated by the terminal device.
  • FIG 13 is a schematic flowchart of another model interaction method of the heterogeneous artificial intelligence AI framework provided by an embodiment of the present disclosure.
  • the model interaction method of the heterogeneous artificial intelligence AI framework in this embodiment can be applied in terminal devices, as shown in Figure As shown in 18, the method may include but is not limited to the following steps:
  • S113 Determine the second framework information supported by the network device, where the second framework information includes the second AI model framework supported by the network device.
  • a reliable reference basis can be provided for the terminal device to determine the AI model framework to be used.
  • Embodiments of the present disclosure also provide a model interaction method for a heterogeneous artificial intelligence AI framework.
  • the second framework information includes any of the following: framework information that the network device supports deployment, and framework information that the network device supports conversion.
  • Embodiments of the present disclosure also provide a model interaction method for heterogeneous artificial intelligence AI frameworks, which further includes: determining a first model framework corresponding to the second framework information, and generating a first AI model based on the first model framework, and Send the first AI model to the network device, or send first indication information to the network device, where the first indication information indicates the first framework information and indicates that the first AI model does not carry model format information, or receive the second indication information, wherein the second indication information indicates that the network device supports the intermediate model representation framework, or determines whether the terminal device supports the intermediate model representation framework, and obtains a third determination result, wherein the network device supports the intermediate model representation framework, and determines Whether the terminal device supports model conversion, a fourth determination result is obtained, and the second AI model is sent to the network device according to the third determination result and/or the fourth determination result, where the second AI model is generated by the terminal device.
  • the terminal device can report the AI framework it supports or the base station side AI framework that can support conversion.
  • the terminal device may determine the first model frame corresponding to the second frame information, and generate the first AI model based on the first model frame.
  • the second frame information includes any of the following: frame information that the network device supports deployment, and frame information that the network device supports conversion.
  • the terminal device determines whether it can complete the conversion of models between different AI frameworks. If the conversion can be completed, the converted or pre-converted model will be sent to the base station. Optionally, if the transmitted model itself does not contain any file format and other related information, the AI framework type corresponding to the model also needs to be indicated to the base station to facilitate analysis on the base station side.
  • the terminal device may send first indication information to the network device, where the first indication information indicates the first framework information and indicates that the first AI model does not carry model format information.
  • Model conversion can be completed on the terminal device side or on the network device side.
  • the terminal device may determine the first model frame corresponding to the second frame information, generate the first AI model based on the first model frame, and send the first AI model to the network device.
  • the terminal device may determine the first model frame information corresponding to the second frame information based on a predefined protocol, and use the model frame to which the first model frame information belongs as the first model frame.
  • the specific example of the predefined protocol can stipulate the specific type of AI framework that can be adopted:
  • Option 1 Provide that network devices and terminal devices use the same AI framework, such as caffe or caffe2.
  • Network equipment and terminal equipment can use different AI frameworks.
  • Solution 2-1 The predefined protocol stipulates that network equipment and terminal equipment use a predefined matching AI framework.
  • network equipment uses pytorch uniformly, and terminal equipment uses a unified framework such as NCNN; or a unified AI framework is used during model training.
  • Frameworks such as pytorch, the model inference deployment process uses a unified framework such as NCNN.
  • NCNN a unified framework
  • the model training is completed and deployed, it is stipulated to support the conversion of pytorch into an NCNN model.
  • Option 2-2 Carry out model translation through a unified intermediate model representation framework. For example, you can specify the unified use of onnx or the unified use of the AI model representation architecture standardized by other institutions such as IEEE.
  • the first frame information is the same as the second frame information, or the first frame information is different from the second frame information, or the first frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. .
  • the terminal device can indicate the intermediate model expression framework it supports.
  • the terminal device may receive second indication information sent by the network device, where the second indication information indicates that the network device supports the intermediate model representation framework.
  • the terminal device can determine whether the terminal device supports the intermediate model representation framework, and obtain a third determination result, wherein the network device supports the intermediate model representation framework, and determine whether the terminal device supports model conversion, and obtain a fourth determination result, and according to the third determination result
  • the determination result and/or the fourth determination result is to send the second AI model to the network device, where the second AI model is generated by the terminal device.
  • the terminal device may determine the second candidate framework information supported by the terminal device, and determine the first model framework corresponding to the second framework information based on the second candidate framework information.
  • the terminal device may use the candidate frame to which the second candidate frame information belongs as the first model frame, and when determining that the second candidate frame information is different from the second frame information, and When the candidate frame to which it belongs supports model conversion, the candidate frame to which the second candidate frame information belongs is used as the first model frame.
  • the predefined protocol includes any of the following: defining the same frame information used by the network device, or defining frame information matching the use of the network device, or defining frame information corresponding to the model processing stage, or defining a representation frame based on an intermediate model Process model.
  • the first frame information is the same as the second frame information, or the first frame information is different from the second frame information, or the first frame information is different from the second frame information, and the second frame information is determined based on the model processing stage. .
  • model processing stage includes any of the following: model training stage and model inference deployment stage.
  • FIG. 14 is a schematic structural diagram of a communication device provided by an embodiment of the present disclosure.
  • the communication device 140 shown in FIG. 14 may include a transceiver module 1401 and a processing module 1402.
  • the transceiving module 1401 may include a sending module and/or a receiving module.
  • the sending module is used to implement the sending function
  • the receiving module is used to implement the receiving function.
  • the transceiving module 1401 may implement the sending function and/or the receiving function.
  • the communication device 140 may be a terminal device (such as the terminal device in the foregoing method embodiment), a device in the terminal device, or a device that can be used in conjunction with the terminal device.
  • the communication device 140 may be a network device (such as the network device in the foregoing method embodiment), a device in the network device, or a device that can be used in conjunction with the network device.
  • Communication device 140 on the network device side, the device includes:
  • the processing module 1402 is used to: determine that the terminal device supports the first framework information, where the first framework information includes the first AI model framework supported by the terminal device; determine the second AI model framework.
  • the first frame information also includes any of the following:
  • Terminal device supports deployment framework information
  • the terminal device supports converted frame information.
  • the device should include:
  • the transceiver module 1401 is configured to receive first indication information sent by the terminal device, where the first indication information is used to instruct the terminal device to support the first framework information.
  • processing module 1402 is also used for:
  • a second model frame is determined based on the first frame information and the plurality of first candidate frame information.
  • processing module 1402 is also used for:
  • the candidate frame to which the first candidate frame information belongs is used as the second model frame;
  • the first candidate frame information is the same as the first frame information; or,
  • the candidate frame to which the first candidate frame information belongs supports model conversion.
  • processing module 1402 is also used for:
  • the first AI model is generated based on the second model framework.
  • the transceiver module 1401 is also used for:
  • the first AI model is sent to the terminal device, and the first candidate framework information supported by the network device is the same as the first framework information.
  • processing module 1402 is also used for:
  • the transceiver module 1401 is also used for:
  • the second AI model is sent to the terminal device, the first candidate framework supports model conversion, and the first AI model is generated based on the first candidate framework.
  • processing module 1402 is also used for:
  • the first AI model is converted into a second AI model based on the intermediate model representation framework.
  • the transceiver module 1401 is also used for:
  • the transceiver module 1401 is also used for:.
  • the first AI model is not sent to the terminal device, the network device does not support the intermediate model representation framework, and the network device does not support model conversion.
  • the transceiver module 1401 is also used for:
  • the network equipment meets at least one of the following:
  • the first supported candidate framework does not support model conversion
  • the transceiver module 1401 is also used for:
  • the network equipment meets one of the following:
  • the supported first candidate frame information is the same as the first frame information
  • the transceiver module 1401 is also used for:
  • Second instruction information is sent to the terminal device, and the second instruction information is used to instruct the terminal device to select one first frame information from multiple first frame information.
  • processing module 1402 is also used for:
  • the transceiver module 1401 is also used for:
  • the first AI model is sent to the terminal device according to the first determination result and/or the second determination result.
  • the transceiver module 1401 is also used for:
  • the indicated content includes at least one of the following:
  • processing module 1402 is also used for:
  • the transceiver module 1401 is also used for:
  • processing module 1402 is also used for:
  • the transceiver module 1401 is also used for:
  • processing module 1402 is also used for:
  • First framework information that is different from the first candidate framework information, and the first candidate framework to which the first candidate framework information belongs supports converting the model into a model supported by the selected first framework information.
  • processing module 1402 is also used for:
  • the first AI model is converted into a fourth AI model based on the selected first model framework.
  • the second model framework determined by the selected first model framework supports model conversion.
  • the first AI model is generated by the network device based on the second model framework. .
  • the transceiver module 1401 is also used for:
  • processing module 1402 is also used for:
  • the first determination result indicates that the network device supports the intermediate model representation framework
  • the second determination result indicates that the network device supports model conversion
  • the transceiver module 1401 is also used for:
  • the first AI model is sent to the terminal device.
  • processing module 1402 is also used for:
  • the first AI model is processed based on the intermediate model representation framework to obtain the AI model file.
  • the first determination result indicates that the network device supports the intermediate model representation framework, and the second determination result indicates that the network device does not support model conversion.
  • the transceiver module 1401 is also used for:
  • the transceiver module 1401 is also used for:
  • the network device can determine that the terminal device supports the first framework information, where the first framework information includes the first AI model framework supported by the terminal device, and determine the second AI model framework.
  • the impact of differences in model frameworks between network devices and terminal devices on artificial intelligence AI model interaction can be effectively avoided, thereby effectively ensuring the transmission accuracy of the artificial intelligence AI model and improving the interaction effect of the artificial intelligence AI model.
  • the device includes:
  • the transceiver module 1401 is configured to indicate supported first framework information to the network device, where the first framework information includes the first AI model framework supported by the terminal device.
  • the first frame information includes any of the following:
  • Terminal device supports deployment framework information
  • the terminal device supports converted frame information.
  • the transceiver module 1401 is also used for:
  • the device also includes:
  • Processing module 1402 configured to determine second framework information supported by the network device, where the second framework information includes a second model framework supported by the network device; determine the supported first framework based on the predefined protocol and/or the second framework information information.
  • the transceiver module 1401 is also used for:
  • Receive the second AI model sent by the network device where the second AI model is obtained by the network device converting the first AI model based on the first framework information or the intermediate model representation framework, and the first AI model is generated by the network device.
  • the transceiver module 1401 is also used for:
  • Receive second indication information sent by the network device where the second indication information is used to indicate that the cell to which the network device belongs does not support AI model deployment based on the first framework information.
  • the transceiver module 1401 is also used for:
  • the terminal equipment meets one of the following:
  • the supported first frame information is the same as the first candidate frame information supported by the network device;
  • the supported first frame information is different from the first candidate frame information supported by the network device.
  • the transceiver module 1401 is also used for:
  • the terminal device satisfies: the number of supported first frame information is multiple.
  • processing module 1402 is also used for:
  • the sending and receiving module 1401 is also used to send the third AI model to the network device.
  • the transceiver module 1401 is also used for:
  • the transceiver module 1401 is also used for:
  • Receive the AI model file sent by the network device where the AI model file is obtained by the network device processing the first AI model based on the intermediate model representation framework, the first AI model is generated by the network device based on the second model framework, and the second model framework is based on the second model framework.
  • a frame information is determined.
  • the transceiver module 1401 is also used for:
  • the terminal device can indicate the supported first framework information to the network device, where the first framework information includes the first AI model framework supported by the terminal device.
  • the applicability of the AI model generated by the network device side to the terminal device side can be effectively guaranteed.
  • Communication device 140 on the network device side, the device includes:
  • the transceiver module 1401 is configured to indicate supported second framework information to the terminal device, where the second framework information includes a second AI model framework supported by the network device.
  • the second frame information includes any of the following:
  • Network equipment supports deployment framework information
  • the network device supports converted frame information.
  • the transceiver module 1401 is also used for:
  • the network device can indicate the supported second framework information to the terminal device, where the second framework information includes the second AI model framework supported by the network device.
  • the terminal device can accurately obtain the second framework information supported by the network device.
  • the device includes:
  • the processing module 1402 is configured to determine second framework information supported by the network device, where the second framework information includes a second AI model framework supported by the network device.
  • the second frame information includes any of the following:
  • Network equipment supports deployment framework information
  • the network device supports converted frame information.
  • processing module 1402 is also used for:
  • the sending and receiving module 1401 is also used to send the first AI model to the network device.
  • the transceiving module 1401 is also configured to send first indication information to the network device, where the first indication information indicates the first framework information and indicates that the first AI model does not carry model format information.
  • the transceiver module 1401 is also configured to receive second indication information sent by the network device, where the second indication information indicates that the network device supports the intermediate model representation framework; or,
  • the processing module 1402 is also used to determine whether the terminal device supports the intermediate model representation framework, and obtain a third determination result, wherein the network device supports the intermediate model representation framework, and determine whether the terminal device supports model conversion, and obtain a fourth determination result. result.
  • the transceiving module 1401 is also configured to send the second AI model to the network device according to the third determination result and/or the fourth determination result, where the second AI model is generated by the terminal device.
  • the terminal device can determine the second framework information supported by the network device, where the second framework information includes the second AI model framework supported by the network device.
  • the terminal device can determine the second framework information supported by the network device, where the second framework information includes the second AI model framework supported by the network device.
  • FIG 15 is a schematic structural diagram of another communication device provided by an embodiment of the present disclosure.
  • the communication device 150 may be a network device (such as the network device in the foregoing method embodiment), a terminal device (such as the terminal device in the foregoing method embodiment), or a chip or chip system that supports the network device to implement the above method. , or processor, etc., or it can also be a chip, chip system, or processor, etc. that supports the terminal device to implement the above method.
  • the device can be used to implement the method described in the above method embodiment. For details, please refer to the description in the above method embodiment.
  • Communication device 150 may include one or more processors 1501.
  • the processor 1501 may be a general-purpose processor or a special-purpose processor, or the like.
  • it can be a baseband processor or a central processing unit.
  • the baseband processor can be used to process communication protocols and communication data.
  • the central processor can be used to control communication devices (such as base stations, baseband chips, terminal equipment, terminal equipment chips, DU or CU, etc.) and execute computer programs. , processing data for computer programs.
  • the communication device 150 may also include one or more memories 1502, on which a computer program 1504 may be stored.
  • the processor 1501 may store a computer program 1503, and the processor 1501 may execute the computer program 1504 and/or Computer program 1503, so that the communication device 150 executes the method described in the above method embodiment.
  • the memory 1502 may also store data.
  • the communication device 150 and the memory 1502 can be provided separately or integrated together.
  • the communication device 150 may also include a transceiver 1505 and an antenna 1506.
  • the transceiver 1505 may be called a transceiver unit, a transceiver, a transceiver circuit, etc., and is used to implement transceiver functions.
  • the transceiver 1505 may include a receiver and a transmitter.
  • the receiver may be called a receiver or a receiving circuit, etc., used to implement the receiving function;
  • the transmitter may be called a transmitter, a transmitting circuit, etc., used to implement the transmitting function.
  • the communication device 150 may also include one or more interface circuits 1507.
  • the interface circuit 1507 is used to receive code instructions and transmit them to the processor 1501 .
  • the processor 1501 executes the code instructions to cause the communication device 150 to perform the method described in the above method embodiment.
  • the processor 1501 may include a transceiver for implementing receiving and transmitting functions.
  • the transceiver may be a transceiver circuit, an interface, or an interface circuit.
  • the transceiver circuits, interfaces or interface circuits used to implement the receiving and transmitting functions can be separate or integrated together.
  • the above-mentioned transceiver circuit, interface or interface circuit can be used for reading and writing codes/data, or the above-mentioned transceiver circuit, interface or interface circuit can be used for signal transmission or transfer.
  • the processor 1501 may store a computer program 1503, and the computer program 1503 runs on the processor 1501, causing the communication device 150 to perform the method described in the above method embodiment.
  • the computer program 1503 may be solidified in the processor 1501, in which case the processor 1501 may be implemented by hardware.
  • the communication device 150 may include a circuit, which may implement the functions of sending or receiving or communicating in the foregoing method embodiments.
  • the processors and transceivers described in this disclosure may be implemented on integrated circuits (ICs), analog ICs, radio frequency integrated circuits (RFICs), mixed signal ICs, application specific integrated circuits (ASICs), printed circuit boards ( printed circuit board (PCB), electronic equipment, etc.
  • the processor and transceiver can also be manufactured using various IC process technologies, such as complementary metal oxide semiconductor (CMOS), n-type metal oxide-semiconductor (NMOS), P-type Metal oxide semiconductor (positive channel metal oxide semiconductor, PMOS), bipolar junction transistor (BJT), bipolar CMOS (BiCMOS), silicon germanium (SiGe), gallium arsenide (GaAs), etc.
  • CMOS complementary metal oxide semiconductor
  • NMOS n-type metal oxide-semiconductor
  • PMOS P-type Metal oxide semiconductor
  • BJT bipolar junction transistor
  • BiCMOS bipolar CMOS
  • SiGe silicon germanium
  • GaAs gallium arsenide
  • the communication device described in the above embodiments may be a network device (such as the network device in the aforementioned method embodiment) or a terminal device (such as the terminal device in the aforementioned method embodiment), but the scope of the communication device described in the present disclosure is not limited to Furthermore, the structure of the communication device may not be limited by FIG. 15 .
  • the communication device may be a stand-alone device or may be part of a larger device.
  • the communication device may be:
  • the IC collection may also include storage components for storing data and computer programs;
  • the communication device may be a chip or a chip system
  • the schematic structural diagram of the chip shown in FIG. 16 refer to the schematic structural diagram of the chip shown in FIG. 16 .
  • the chip shown in Figure 16 includes a processor 1601 and an interface 1602.
  • the number of processors 1601 may be one or more, and the number of interfaces 1602 may be multiple.
  • the processor 1601 is used to implement S102 in Figure 2, or to implement S105 and S205 in Figure 5, or to implement S106 and S306 in Figure 6, etc.
  • the interface 1602 is used to implement S103 in Figure 3, or to implement S104 in Figure 4, or to implement S206 and S406 in Figure 6, etc.
  • the interface 1602 is used to implement S109 in Figure 9, or to implement S311 in Figure 11, etc.
  • the processor 1601 is used to implement S110 and S210 in Figure 10, or is used to implement S111 and S211 in Figure 11, etc.
  • the chip also includes a memory 1603, which is used to store necessary computer programs and data.
  • Embodiments of the present disclosure also provide a communication system, which includes a communication device as a network device (such as a network device in the aforementioned method embodiment) in the embodiment of FIG. 14 and a terminal device (such as a terminal in the aforementioned method embodiment). equipment), or the system includes the communication device as a network device (such as the network device in the previous method embodiment) in the embodiment of FIG. 15 and the communication device as a terminal device (such as the terminal device in the previous method embodiment). Communication device.
  • a communication device as a network device (such as a network device in the aforementioned method embodiment) in the embodiment of FIG. 14 and a terminal device (such as a terminal in the aforementioned method embodiment). equipment
  • the communication device includes the communication device as a network device (such as the network device in the previous method embodiment) in the embodiment of FIG. 15 and the communication device as a terminal device (such as the terminal device in the previous method embodiment).
  • the present disclosure also provides a readable storage medium on which instructions are stored, and when the instructions are executed by a computer, the functions of any of the above method embodiments are implemented.
  • the present disclosure also provides a computer program product, which, when executed by a computer, implements the functions of any of the above method embodiments.
  • the above embodiments it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof.
  • software it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer programs.
  • the computer program When the computer program is loaded and executed on a computer, the processes or functions described in accordance with the embodiments of the present disclosure are generated in whole or in part.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device.
  • the computer program may be stored in or transferred from one computer-readable storage medium to another, for example, the computer program may be transferred from a website, computer, server, or data center Transmission to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means.
  • the computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains one or more available media integrated therein.
  • the available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., high-density digital video discs (DVD)), or semiconductor media (e.g., solid state disks, SSD)) etc.
  • magnetic media e.g., floppy disks, hard disks, magnetic tapes
  • optical media e.g., high-density digital video discs (DVD)
  • DVD digital video discs
  • semiconductor media e.g., solid state disks, SSD
  • At least one in the present disclosure can also be described as one or more, and the plurality can be two, three, four or more, and the present disclosure is not limited.
  • the technical feature is distinguished by “first”, “second”, “third”, “A”, “B”, “C” and “D” etc.
  • the technical features described in “first”, “second”, “third”, “A”, “B”, “C” and “D” are in no particular order or order.
  • each table in this disclosure can be configured or predefined.
  • the values of the information in each table are only examples and can be configured as other values, which is not limited by this disclosure.
  • it is not necessarily required to configure all the correspondences shown in each table.
  • the corresponding relationships shown in some rows may not be configured.
  • appropriate deformation adjustments can be made based on the above table, such as splitting, merging, etc.
  • the names of the parameters shown in the titles of the above tables may also be other names understandable by the communication device, and the values or expressions of the parameters may also be other values or expressions understandable by the communication device.
  • other data structures can also be used, such as arrays, queues, containers, stacks, linear lists, pointers, linked lists, trees, graphs, structures, classes, heaps, hash tables or hash tables. wait.
  • Predefinition in this disclosure may be understood as definition, pre-definition, storage, pre-storage, pre-negotiation, pre-configuration, solidification, or pre-burning.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

本申请公开了一种异构人工智能AI框架的模型交互方法、装置及系统,可以应用于通信系统中,该方法由网络设备执行时包括:确定终端设备支持第一框架信息,其中,第一框架信息包括终瑞设备支持的第一AI模型框架,确定第二AI模型框架。通过实施本申请的方法,有效避免网络设备与终端设备间模型框架的差异对人工智能AI模型交互造成影响,从而能够有效保障人工智能AI模型的传输准确率,提升人工智能AI模型的交互效果。

Description

异构人工智能AI框架的模型交互方法、装置及系统 技术领域
本公开涉及通信技术领域,尤其涉及一种异构人工智能AI框架的模型交互方法、装置及系统。
背景技术
在无线人工智能(Artificial Intelligence,AI)应用领域,可能会需要进行模型训练(例如,在网络设备侧进行模型训练,或者在终端设备侧进行模型训练),如果终端设备的运算能耗,或者支持的计算能力低于一定的门限之后,可能无法快速完成AI模型训练,则可以在网络设备进行模型训练。
相关技术中,在AI模型交互过程中,由于网络设备和终端设备所支持的模型框架可能存在差异,影响基于模型框架训练所得人工智能AI模型的部署,影响人工智能AI模型的交互效果。
发明内容
本公开实施例提供一种异构人工智能AI框架的模型交互方法、装置及系统,可应用于通信技术领域中,有效避免网络设备与终端设备间模型框架的差异对人工智能AI模型交互造成影响,从而能够有效保障人工智能AI模型的传输准确率,提升人工智能AI模型的交互效果。
第一方面,本公开实施例提供一种异构人工智能AI框架的模型交互方法,由网络设备执行,该方法包括:确定终端设备支持第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架;确定第二AI模型框架。
在一种实施方式中,所述第一框架信息还包括以下任一项:
所述终端设备支持部署的框架信息;
所述终端设备支持转换的框架信息。
在一种实施方式中,所述方法还包括:
接收所述终端设备发送的第一指示信息,所述第一指示信息用于指示所述终端设备支持第一框架信息。
在一种实施方式中,所述确定第二模型框架,包括:
确定所述网络设备支持多个第一候选框架信息;
基于所述第一框架信息和所述多个第一候选框架信息,确定所述第二模型框架。
在一种实施方式中,所述基于所述第一框架信息和所述多个第一候选框架信息,确定所述第二模型框架,包括:
将所述第一候选框架信息所属候选框架作为所述第二模型框架;其中,
所述第一候选框架信息与所述第一框架信息相同;或者,
所述第一候选框架信息所属候选框架支持模型转换。
在一种实施方式中,所述方法还包括:
基于所述第二模型框架生成第一AI模型;
将第一AI模型发送至所述终端设备,所述网络设备支持的第一候选框架信息与所述第一框架信息相同;或者,
基于所述第一框架信息将第一AI模型转换为第二AI模型;
将所述第二AI模型发送至所述终端设备,所述第一候选框架支持模型转换,所述第一AI模型基于所述第一候选框架生成;或者,
基于中间模型表示框架将第一AI模型转换为第二AI模型;
将所述第二AI模型发送至所述终端设备,所述终端设备和所述网络设备均支持所述中间模型表示框架;或者,
不向所述终端设备发送所述第一AI模型,所述网络设备不支持所述中间模型表示框架,且所述网络设备不支持模型转换。
在一种实施方式中,所述方法还包括:
向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述网络设备所属小区不支持基于所述第一框架信息的AI模型部署;
其中,所述网络设备满足以下至少一种:
支持的第一候选框架不支持模型转换;
不支持模型转换;
不支持中间模型表示框架。
在一种实施方式中,所述方法还包括:
向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述终端设备向所述网络设备上报模型;
其中,所述网络设备满足以下一种:
所支持的第一候选框架信息与所述第一框架信息相同;
支持模型转换。
在一种实施方式中,所述方法还包括:
向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述终端设备从多个所述第一框架信息中选择一个第一框架信息。
在一种实施方式中,所述方法还包括:
确定所述网络设备是否支持所述中间模型表示框架,得到第一确定结果;
确定所述网络设备是否支持模型转换,得到第二确定结果,以及
根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备。
在一种实施方式中,所述方法还包括:
接收所述终端设备发送的第三AI模型,其中,所述第三AI模型是由所述终端设备基于所述网络设备所指示内容生成的;
其中,所述所指示内容包括以下至少一项:
所述第二指示信息指示的内容。
在一种实施方式中,所述方法还包括:
从多个所述第一框架信息中选择一个第一框架信息;
向所述终端设备发送第三指示信息,其中,所述第三指示信息用于向所述终端设备指示所选择第一框架信息;
基于所述所选择第一框架信息生成第一AI模型,
将所述第一AI模型发送至所述终端设备。
在一种实施方式中,所述从多个第一框架信息中选择一个第一框架信息,还包括:
确定所述网络设备所支持的第一候选框架信息;
根据所述第一候选框架信息,从多个第一框架信息中选择一个第一框架信息;其中,所述所选择第一框架信息包括所选择第一模型框架,所述所选择第一框架信息是以下任一项:
与所述第一候选框架信息相同的第一框架信息;或者,
与所述第一候选框架信息不同的第一框架信息,且所述第一候选框架信息所属第一候选框架支持将模型转换为所述所选择第一框架信息所支持的模型。
在一种实施方式中,所述方法还包括:
基于所述所选择第一模型框架将第一AI模型转换为第四AI模型,由所述所选择第一框架信息所确定的第二模型框架支持模型转换,所述第一AI模型由所述网络设备基于第二模型框架生成;
将所述第四AI模型发送至所述终端设备。
在一种实施方式中,所述根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备,包括:
确定模型下发方式,所述第一确定结果指示所述网络设备支持所述中间模型表示框架,且所述第二确定结果表示所述网络设备支持模型转换;
基于所述模型下发方式,将所述第一AI模型发送至所述终端设备。
在一种实施方式中,所述根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备,包括:
基于所述中间模型表示框架处理所述第一AI模型,得到AI模型文件,所述第一确定结果指示所述网络设备支持所述中间模型表示框架,且所述第二确定结果表示所述网络设备不支持模型转换;
将所述AI模型文件发送至所述终端设备。
在一种实施方式中,所述方法还包括:
向所述终端设备发送第二指示信息,其中,所述第二指示信息指示所述AI模型文件的AI模型存储格式,所述AI模型文件未携带模型格式信息。第二方面,本公开实施例提供另一种异构人工智能AI框架的模型交互方法,由终端设备执行,该方法包括:向网络设备指示支持的第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架。
在一种实施方式中,所述第一框架信息包括以下任一项:
所述终端设备支持部署的框架信息;
所述终端设备支持转换的框架信息。
在一种实施方式中,所述方法还包括:
向所述网络设备发送第一指示信息,所述第一指示信息用于指示所述终端设备支持第一框架信息。
在一种实施方式中,确定所述终端设备支持的第一框架信息,包括:
确定所述网络设备支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二模型框架;
基于预定义协议和/或所述第二框架信息,确定支持的第一框架信息。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的第一AI模型,其中,第一AI模型由所述网络设备生成;或者,
接收所述网络设备发送的第二AI模型,其中,所述第二AI模型由所述网络设备基于所述第一框架信息或中间模型表示框架对第一AI模型转换得到,所述第一AI模型由所述网络设备生成。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的第二指示信息,所述第二指示信息用于指示所述网络设备所属小区不支持基于所述第一框架信息的AI模型部署。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的第二指示信息,所述第二指示信息用于指示所述终端设备向所述网络设备上报模型;
其中,所述终端设备满足以下一种:
支持的第一框架信息与所述网络设备所支持的第一候选框架信息相同;
支持的第一框架信息与所述网络设备所支持的第一候选框架信息不同。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的第二指示信息,其中,所述第二指示信息用于指示所述终端设备从多个所述第一框架信息中选择一个第一框架信息;
其中,所述终端设备满足:支持的第一框架信息的数量是多个。
在一种实施方式中,所述方法还包括:
根据所述第二指示信息指示的内容,确定所述网络设备所指示框架信息;
基于所述所指示框架信息,生成第三AI模型;
将所述第三AI模型发送至所述网络设备。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的第四AI模型。
在一种实施方式中,所述方法还包括:
接收所述网络设备发送的AI模型文件,其中,所述AI模型文件由所述网络设备基于所述中间模型表示框架处理第一AI模型得到,所述第一AI模型由所述网络设备基于第二模型框架生成,所述第二模型框架是基于所述第一框架信息确定。
在一种实施方式中,该方法,还包括:
接收所述网络设备发送的第二指示信息,其中,所述第二指示信息指示所述AI模型文件的AI模型存储格式,所述AI模型文件未携带模型格式信息。
第三方面,本公开实施例提供一种异构人工智能AI框架的模型交互方法,由网络设备执行,该方法包括:向终端设备指示所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
在一种实施方式中,所述第二框架信息包括以下任一项:
所述网络设备支持部署的框架信息;
所述网络设备支持转换的框架信息。
在一种实施方式中,该方法,还包括:
接收所述终端设备发送的第一AI模型,其中,所述第一AI模型由所述终端设备基于支持的第一框架信息所包括的第一模型框架生成,所述第一框架信息是基于所述第二框架信息确定;或者,
接收所述终端设备发送的第一指示信息,其中,所述第一指示信息指示第一框架信息以及指示所述第一AI模型未携带模型格式信息;或者,
向所述终端设备发送第二指示信息,其中,所述第二指示信息指示所述网络设备支持中间模型表示框架;或者,
接收所述终端设备发送的第二AI模型,其中,所述第二AI模型由所述终端设备生成。
第四方面,本公开实施例提供另一种异构人工智能AI框架的模型交互方法,由终端设备执行,该方法包括:确定网络设备所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
在一种实施方式中,所述第二框架信息包括以下任一项:
所述网络设备支持部署的框架信息;
所述网络设备支持转换的框架信息。
在一种实施方式中,所述方法还包括:
确定与所述第二框架信息对应的第一模型框架,并基于所述第一模型框架生成第一AI模型,以及向所述网络设备发送所述第一AI模型;或者,
向所述网络设备发送第一指示信息,其中,所述第一指示信息指示第一框架信息以及指示所述第一AI模型未携带模型格式信息;或者,
接收所述网络设备发送的第二指示信息,其中,所述第二指示信息指示所述网络设备支持中间模型表示框架;或者,
确定所述终端设备是否支持所述中间模型表示框架,得到第三确定结果,其中,所述网络设备支持中间模型表示框架,并确定所述终端设备是否支持模型转换,得到第四确定结果,以及根据所述第三确定结果和/或所述第四确定结果,将第二AI模型发送至所述网络设备,其中,所述第二AI模型由所述终端设备生成。
第五方面,本公开实施例提供一种通信装置,该通信装置具有实现上述第一方面所述的方法,或者第三方面所述的方法中网络设备的部分或全部功能,比如通信装置的功能可具备本公开中的部分或全部实施例中的功能,也可以具备单独实施本公开中的任一个实施例的功能。所述功能可以通过硬件实现,也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的单元或模块。
在一种实施方式中,该通信装置的结构中可包括收发模块和处理模块,所述处理模块被配置为支持通信装置执行上述方法中相应的功能。所述收发模块用于支持通信装置与其他设备之间的通信。所述通信装置还可以包括存储模块,所述存储模块用于与收发模块和处理模块耦合,其保存通信装置必要的计算机程序和数据。
作为示例,处理模块可以为处理器,收发模块可以为收发器或通信接口,存储模块可以为存储器。
第六方面,本公开实施例提供另一种通信装置,该通信装置具有实现上述第二方面所述的方法示例,或者第四方面所述的方法示例中终端设备的部分或全部功能,比如通信装置的功能可具备本公开中的部分或全部实施例中的功能,也可以具备单独实施本公开中的任一个实施例的功能。所述功能可以通过硬 件实现,也可以通过硬件执行相应的软件实现。所述硬件或软件包括一个或多个与上述功能相对应的单元或模块。
可选的,在本公开的一个实施例之中,该通信装置的结构中可包括收发模块和处理模块,该处理模块被配置为支持通信装置执行上述方法中相应的功能。收发模块用于支持通信装置与其他设备之间的通信。所述通信装置还可以包括存储模块,所述存储模块用于与收发模块和处理模块耦合,其保存通信装置必要的计算机程序和数据。
第七方面,本公开实施例提供一种通信装置,该通信装置包括处理器,当该处理器调用存储器中的计算机程序时,执行上述第一方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第三方面所述的异构人工智能AI框架的模型交互方法。
第八方面,本公开实施例提供一种通信装置,该通信装置包括处理器,当该处理器调用存储器中的计算机程序时,执行上述第二方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第四方面所述的异构人工智能AI框架的模型交互方法。
第九方面,本公开实施例提供一种通信装置,该通信装置包括处理器和存储器,该存储器中存储有计算机程序;所述处理器执行该存储器所存储的计算机程序,以使该通信装置执行上述第一方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第三方面所述的异构人工智能AI框架的模型交互方法。
第十方面,本公开实施例提供一种通信装置,该通信装置包括处理器和存储器,该存储器中存储有计算机程序;所述处理器执行该存储器所存储的计算机程序,以使该通信装置执行上述第二方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第四方面所述的异构人工智能AI框架的模型交互方法。
第十一方面,本公开实施例提供一种通信装置,该装置包括处理器和接口电路,该接口电路用于接收代码指令并传输至该处理器,该处理器用于运行所述代码指令以使该装置执行上述第一方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第三方面所述的异构人工智能AI框架的模型交互方法。
第十二方面,本公开实施例提供一种通信装置,该装置包括处理器和接口电路,该接口电路用于接收代码指令并传输至该处理器,该处理器用于运行所述代码指令以使该装置执行上述第二方面所述的异构人工智能AI框架的模型交互方法,或者执行上述第四方面所述的异构人工智能AI框架的模型交互方法。
第十三方面,本公开实施例提供一种通信系统,该系统包括第五方面所述的通信装置以及第六方面所述的通信装置,或者,该系统包括第七方面所述的通信装置以及第八方面所述的通信装置,或者,该系统包括第九方面所述的通信装置以及第十方面所述的通信装置,或者,该系统包括第十一方面所述的通信装置以及第十二方面所述的通信装置。
第十四方面,本公开实施例提供一种计算机可读存储介质,用于储存为上述网络设备所用的指令,当所述指令被执行时,使所述网络设备执行上述第一方面所述的方法,或者执行上述第三方面所述的方法。
第十五方面,本公开实施例提供一种可读存储介质,用于储存为上述终端设备所用的指令,当所述指令被执行时,使所述终端设备执行上述第二方面所述的方法,或者执行上述第四方面所述的方法。
第十六方面,本公开还提供一种包括计算机程序的计算机程序产品,当其在计算机上运行时,使得计算机执行上述第一方面所述的方法,或者执行上述第三方面所述的方法。
第十七方面,本公开还提供一种包括计算机程序的计算机程序产品,当其在计算机上运行时,使得计算机执行上述第二方面所述的方法,或者执行上述第四方面所述的方法。
第十八方面,本公开提供一种芯片系统,该芯片系统包括至少一个处理器和接口,用于支持网络设备实现第一方面、或者第三方面所涉及的功能,例如,确定或处理上述方法中所涉及的数据和信息中的至少一种。
在一种可能的设计中,所述芯片系统还包括存储器,所述存储器,用于保存网络设备必要的计算机程序和数据。该芯片系统,可以由芯片构成,也可以包括芯片和其他分立器件。
第十九方面,本公开提供一种芯片系统,该芯片系统包括至少一个处理器和接口,用于支持终端设备实现第二方面、或者第四方面所涉及的功能,例如,确定或处理上述方法中所涉及的数据和信息中的至少一种。
在一种可能的设计中,所述芯片系统还包括存储器,所述存储器,用于保存终端设备必要的计算机程序和数据。该芯片系统,可以由芯片构成,也可以包括芯片和其他分立器件。
第二十方面,本公开提供一种计算机程序,当其在计算机上运行时,使得计算机执行上述第一方面所述的方法,或者执行上述第三方面所述的方法。
第二十一方面,本公开提供一种计算机程序,当其在计算机上运行时,使得计算机执行上述第二方面所述的方法,或者执行上述第四方面所述的方法。
综上所述,在本公开实施例提供的方法、装置、设备、芯片系统、存储介质、计算机程序及计算机程序产品,可以实现以下技术效果:
通过确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,确定第二AI模型框架。通过实施本公开的方法,有效避免网络设备与终端设备间模型框架的差异对人工智能AI模型交互造成影响,从而能够有效保障人工智能AI模型的传输准确率,提升人工智能AI模型的交互效果。
附图说明
为了更清楚地说明本公开实施例或背景技术中的技术方案,下面将对本公开实施例或背景技术中所需要使用的附图进行说明。
图1为本公开实施例提供的一种通信系统的架构示意图;
图2是本公开实施例提供的一种异构人工智能AI框架的模型交互方法的流程示意图;
图3是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图4是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图5是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图6是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图7是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图8是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图9是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图10是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图11是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图12是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图13是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图;
图14为本公开实施例提供的一种通信装置的结构示意图;
图15是本公开实施例提供的另一种通信装置的结构示意图;
图16是本公开实施例提供的一种芯片的结构示意图。
具体实施方式
这里将详细地对示例性实施例进行说明,其示例表示在附图中。下面的描述涉及附图时,除非另有表示,不同附图中的相同数字表示相同或相似的要素。以下示例性实施例中所描述的实施方式并不代表与本公开实施例相一致的所有实施方式。相反,它们仅是与如所附权利要求书中所详述的、本公开实施例的一些方面相一致的装置和方法的例子。
在本公开实施例使用的术语是仅仅出于描述特定实施例的目的,而非旨在限制本公开实施例。在本公开实施例和所附权利要求书中所使用的单数形式的“一种”和“该”也旨在包括多数形式,除非上下文清楚地表示其他含义。还应当理解,本文中使用的术语“和/或”是指并包含一个或多个相关联的列出项目的任何或所有可能组合。
应当理解,尽管在本公开实施例可能采用术语第一、第二、第三等来描述各种信息,但这些信息不应限于这些术语。这些术语仅用来将同一类型的信息彼此区分开。例如,在不脱离本公开实施例范围的情况下,第一信息也可以被称为第二信息,类似地,第二信息也可以被称为第一信息。取决于语境,如在此所使用的词语“如果”及“若”可以被解释成为“在……时”或“当……时”或“响应于确定”。
为了便于理解,首先介绍本公开涉及的术语。
1、人工智能(Artificial Intelligence,AI)
人工智能,是指研究、开发用于模拟、延伸和扩展人的智能的理论、方法、技术及应用系统的一门新的技术科学。人工智能是计算机科学的一个分支,它企图了解智能的实质,并生产出一种新的能以人类智能相似的方式做出反应的智能机器,该领域的研究包括机器人、语言识别、图像识别、自然语言处理和专家系统等。
2、模型框架
模型框架,可以提供开发者构建神经网络的接口(数学操作),自动对神经网络训练(进行反向求导,逼近地求解最优值),得到一个神经网络模型(逼近函数)用于解决分类、回归、拟合的问题,实现目标分类、语音识别等应用场景。
为了更好的理解本公开实施例公开的一种异构人工智能AI框架的模型交互方法,下面首先对本公开实施例适用的通信系统进行描述。
请参见图1,图1为本公开实施例提供的一种通信系统的架构示意图。该通信系统可包括但不限于一个网络设备和一个终端设备,图1所示的设备数量和形态仅用于举例并不构成对本公开实施例的限定,实际应用中可以包括两个或两个以上的网络设备,两个或两个以上的终端设备。图1所示的通信系统以包括一个网络设备101和一个终端设备102为例。
需要说明的是,本公开实施例的技术方案可以应用于各种通信系统。例如:长期演进(long term evolution,LTE)系统、第五代(5th generation,5G)移动通信系统、5G新空口(new radio,NR)系统,或者其他未来的新型移动通信系统等。
本公开实施例中的网络设备101是网络侧的一种用于发射或接收信号的实体。例如,网络设备101可以为演进型基站(evolved NodeB,eNB)、传输点(transmission reception point,TRP)、NR系统中的下一代基站(next generation NodeB,gNB)、其他未来移动通信系统中的基站或无线保真(wireless fidelity,WiFi)系统中的接入节点等。本公开的实施例对网络设备所采用的具体技术和具体设备形态不 做限定。
本公开实施例提供的网络设备可以是由集中单元(central unit,CU)与分布式单元(distributed unit,DU)组成的,其中,CU也可以称为控制单元(control unit),采用CU-DU的结构可以将网络设备,例如基站的协议层拆分开,部分协议层的功能放在CU集中控制,剩下部分或全部协议层的功能分布在DU中,由CU集中控制DU。
本公开实施例中的终端设备102是用户侧的一种用于接收或发射信号的实体,如手机。终端设备也可以称为终端设备(terminal)、用户设备(user equipment,UE)、移动台(mobile station,MS)、移动终端设备(mobile terminal,MT)等。终端设备可以是具备通信功能的汽车、智能汽车、手机(mobile phone)、穿戴式设备、平板电脑(Pad)、带无线收发功能的电脑、虚拟现实(virtual reality,VR)终端设备、增强现实(augmented reality,AR)终端设备、工业控制(industrial control)中的无线终端设备、无人驾驶(self-driving)中的无线终端设备、远程手术(remote medical surgery)中的无线终端设备、智能电网(smart grid)中的无线终端设备、运输安全(transportation safety)中的无线终端设备、智慧城市(smart city)中的无线终端设备、智慧家庭(smart home)中的无线终端设备等等。
本公开的实施例对终端设备所采用的具体技术和具体设备形态不做限定。
可以理解的是,本公开实施例描述的通信系统是为了更加清楚的说明本公开实施例的技术方案,并不构成对于本公开实施例提供的技术方案的限定,本领域普通技术人员可知,随着系统架构的演变和新业务场景的出现,本公开实施例提供的技术方案对于类似的技术问题,同样适用。
下面结合附图对本公开所提供的异构人工智能AI框架的模型交互方法及其装置进行详细地介绍。图2是本公开实施例提供的一种异构人工智能AI框架的模型交互方法的流程示意图,该方法由网络设备执行。本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,例如演进型基站、传输点、NR系统中的下一代基站、其他未来移动通信系统中的基站或无线保真系统中的接入节点等,对此不做限制。
如图2所示,该方法可以包括但不限于如下步骤:
S102:确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架。
其中,人工智能(Artificial Intelligence,AI),是指研究、开发用于模拟、延伸和扩展人的智能的理论、方法、技术及应用系统的一门新的技术科学。AI模型,是指人工智能所采用的一种算法。而第一人工智能AI模型,则是指网络设备侧基于第二模型框架所生成的人工智能AI模型。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,第一框架信息还包括以下任一项:终端设备支持部署的框架信息、终端设备支持转换的框架信息,由此,可以有效提升第一框架信息对于终端设备所支持框架信息的表征全面性。
S202:确定第二AI模型框架。
其中,框架,可以自动对初始模型训练(例如进行反向求导,逼近地求解最优值),以得到一个满足预期需求的AI模型。
框架信息,可以是用于描述框架的信息,框架信息例如框架类型、框架名称、框架版本等,而第一框架信息,则是指描述终端设备所支持的框架的相关信息,例如,终端设备所支持框架的框架类型、框架名称、框架版本等。
举例而言,终端设备可以向网络设备指示该终端设备支持解析哪些模型框架,比如终端设备的框架是A,但是指示能够解析网络设备基于B框架生成的模型,则网络设备可以直接将基于B框架生成的模型发送给终端设备。
终端设备也可以上报支持解析多个框架B、C、D、E、F,网络设备在匹配到框架为C时,还可以向终端设备进一步指示所支持的框架类型为C。
其中,第一AI模型框架,可以是指终端设备侧所支持的AI模型框架。
其中,第二AI模型框架,可以是指网络设备侧所支持的AI模型框架,该框架可以与上述第一框架信息相对应,而相对应可以是类型相同,版本相同,或者相匹配,相匹配可以例如类型不相同,但是支持框架转换,或者类型相同,版本不相同,但是支持版本兼容等,对此不做限制。
本公开实施例中,终端设备所支持的第一框架信息的数量可能是多个,网络设备侧所支持的框架信息的数量也可能是多个,且可能存在一个或多个与第一框架信息对应的第二模型框架。
可以理解的是,终端设备侧与网络设备各自所支持的框架信息可能存在差异,而不同框架所支持的 模型类型也可能存在差异,由此,本公开实施例中,通过确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,确定第二AI模型框架,可以有效提升所得第二AI模型框架与终端设备所支持框架的适配性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,确定终端设备支持的第一框架信息,还可以是基于预定义协议,确定终端设备支持的第一框架信息,实现基于预定义协议快速地获知终端设备所支持的第一框架信息,提升交互匹配效率。
其中,预定义协议,可以是指预先基于应用场景针对网络设备与终端设备之间的交互过程所设定的协议。
举例而言,该预定义协议可以规定网络设备侧与终端设备侧统一使用相同的AI框架,比如统一使用用于特征抽取的卷积框架(Convolution Architecture For Feature Extraction,Caffe)多个版本中的指定版本,例如caffe或者caffe2。或者,也可以规定网络设备侧与终端设备侧可以使用不同的AI框架。
例如,预定义协议,可以规定基站侧与终端侧使用预定义的能够相匹配的AI框架,比如基站侧统一使用pytorch(一个开源的Python机器学习库,用于自然语言处理等应用程序),终端侧使用统一的框架如NCNN(一个为终端设备提供优化的高性能神经网络前向计算框架)等;或者模型训练使用统一的框架如pytorch,模型推理部署使用统一的框架如NCNN,模型训练完成在进行部署时,规定支持将基于pytorch训练好的模型转换为NCNN支持的模型。
或者,可以通过统一的中间模型表示框架进行模型翻译,比如预定义协议规定统一使用开放神经网络交换(Open Neural Network Exchange,ONNX)或者统一使用其他机构如电气与电子工程师协会(Institute of Electrical and Electronics Engineers,IEEE)标准化的AI模型表示架构。
本实施例中,通过确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,确定第二AI模型框架,有效避免网络设备与终端设备间模型框架的差异对人工智能AI模型交互造成影响,从而能够有效保障人工智能AI模型的传输准确率,提升人工智能AI模型的交互效果。
图3是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图3所示,该方法可以包括但不限于如下步骤:
S103:接收终端设备发送的第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
其中,指示消息,可以是指终端设备与网络设备之间进行信息交流的载体,该指示消息可以是电信号或者电磁波等。而第一指示信息,则是由终端设备生成,可以被用于至少指示终端设备所支持的第一框架信息。
本实施例中,通过接收终端设备发送的第一指示信息,第一指示信息用于指示终端设备支持第一框架信息,可以为网络设备确定终端设备支持第一框架信息提供可靠的参考依据。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,确定第二模型框架,包括:确定网络设备支持多个第一候选框架信息,基于第一框架信息和多个第一候选框架信息,确定第二模型框架,由此,可以基于第一候选框架信息与第一框架信息是否相同的情况,准确地、灵活地确定出与第一框架信息对应的第二模型框架。
其中,候选框架,是指网络设备所支持的模型框架,而第一候选框架信息,是指候选框架对应的框架信息,该第一候选框架信息的数量可以是一个或多个,对此不做限制。
可以理解的是,不同网络设备所支持的模型框架类型可能存在差异,由此,本实施例中可以确定网络设备所支持的第一候选框架信息,可以为确定与第一框架信息对应的第二模型框架提供可靠的参考数据。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,基于第一框架信息和多个第一候选框架信息,确定第二模型框架,包括:将第一候选框架信息所属候选框架作为第二模型框架,其中,第一候选框架信息与第一框架信息相同,或者,第一候选框架信息所属候选框架支持模型转换,可以有效提升第二模型框架确定过程的可靠性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:基于第二模型框架生成第一AI模型,将第一AI模型发送至终端设备,网络设备支持的第一候选框架信息与第一框架信息相同,以实现从网络设备侧至终端设备侧的模型交互过程,保证第一AI模型与终端设备之间的适配性。
其中,第一AI模型能够被第一框架正确解析部署,比如可以在网络设备的第二模型框架中转换成第一框架所支持的表达形式,或者,直接生成第二模型框架支持的模型表示形式后下发给终端设备,然后由终端设备的第一模型框架的相应模型转换函数进行模型转换后,在终端设备中进行部署。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:基于第一框架信息将第一AI模型转换为第二AI模型,将第二AI模型发送至终端设备,第一候选框架支持模型转换,第一AI模型基于第一候选框架生成,使得第二AI模型能够与终端设备所支持的第一框架信息相适配,便于终端设备对第二AI模型进行准确解析。
其中,第二AI模型,可以是基于第一框架信息对第一AI模型进行转换处理所得到的AI模型。
也即是说,如果确定第一候选框架信息与第一框架信息不同,但是确定第一候选框架可以支持模型转换,则可以基于终端设备所支持的第一框架信息将由第一候选框架信息所生成的第一AI模型进行模型转换得到第二AI模型,使得第二AI模型能够与终端设备所支持的第一框架信息相适配,便于终端设备对第二AI模型进行准确解析。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:基于中间模型表示框架将第一AI模型转换为第二AI模型,将第二AI模型发送至终端设备,终端设备和网络设备均支持中间模型表示框架,由此,可以在终端设备和网络设备均支持中间模型表示框架,基于中间模型表示框架有效提升部署效果。
其中,中间模型表示框架,可以将各个框架训练好的模型统一转换为同一格式进行存储,以便于AI模型的跨平台和跨框架传输和部署。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:不向终端设备发送第一AI模型,网络设备不支持中间模型表示框架,且网络设备不支持模型转换。
可以理解的是,当网络设备不支持中间模型表示框架,且第二确定结果指示网络设备不支持模型转换时,第一AI模型可能不适用于终端设备,此时不向终端设备发送第一AI模型,可以有效降低资源消耗,同时有效避免第一AI模型对终端设备带入干扰。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:向终端设备发送第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署,其中,网络设备满足以下至少一种:支持的第一候选框架不支持模型转换、不支持模型转换、不支持中间模型表示框架,可以基于第二指示信息及时向终端设备指示网络设备所属小区不支持基于第一框架信息的AI模型部署,以便于终端设备快速采取应对措施。
其中,第二指示信息,可以是指示网络设备所属小区不支持基于第一框架信息的AI模型推理的消息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:向终端设备发送第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型,其中,网络设备满足以下一种:所支持的第一候选框架信息与第一框架信息相同、支持模型转换,由此,可以基于第二指示消息及时指示终端设备向网络设备上报模型。
图4是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图4所示,该方法可以包括但不限于如下步骤:
S104:向终端设备发送第二指示信息,第二指示信息用于指示终端设备从多个第一框架信息中选择 一个第一框架信息。
其中,所选择第一框架信息,可以是指多个第一框架信息中被用于此次交互过程的框架信息,该所选择第一框架信息,可以是由网络设备所确定的适用于终端设备模型解析的框架信息,对此不做限制。
如果多个第一框架信息中,与第一候选框架信息相同的第一框架信息只有一个,则可以直接将该相同的第一框架信息作为所选择第一框架信息,如果与第一候选框架信息相同的第一框架信息有一个以上,则可以由基站从一个以上的相同的第一框架信息中选择一个,并向终端设备指示,或者,也可以由基站基于预定义规则从一个以上的相同的第一框架信息中选择一个,并向终端设备指示,对此不做限制。
本实施例中,通过向终端设备发送第二指示信息,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息,可以基于第二指示信息准确指示终端设备中第一框架信息的确定过程,从而有效提升所选择第一框架信息的可靠性。
图5是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图5所示,该方法可以包括但不限于如下步骤:
S105:确定网络设备是否支持中间模型表示框架,得到第一确定结果。
其中,第一确定结果,例如可以是网络设备支持中间模型表示框架,或者,网络设备不支持中间模型表示框架。
S205:确定网络设备是否支持模型转换,得到第二确定结果。
举例而言,当中间模型表示框架是唯一的一个类型时,则可以不向终端设备指示具体AI模型所用框架信息;如果终端设备支持的中间模型表示框架是多个,则可以再向终端设备指示网络设备是所用框架信息;如果网络设备既支持转换,也支持终端所支持的中间模型表示框架,则可以由基站自行决定具体采用哪种方式下发,也可以是由预协议定义确定优先级;如果是网络设备自行决定,那么,网络设备同样可以向终端设备指示AI模型具体使用的框架信息。
其中,第二确定结果,例如可以是网络设备支持模型转换,或者,网络设备不支持模型转换。
S305:根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备。
举例而言,第一AI模型的上报方式可以是以下任一项:
第一、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息相同时,向终端设备发送指示终端设备生成模型并上报的消息,该消息可以被称为第四指示信息。
即,第四指示信息指示终端设备生成模型并上报。
则终端设备可以基于第四指示信息的指示,基于支持部署的框架信息生成AI模型,并将AI模型上报至网络设备,对此不做限制。
第二、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备支持模型转换时,向终端设备发送第四指示信息,其中,第四指示信息指示终端设备生成模型并上报。
则终端设备可以基于第四指示信息的指示,基于支持部署的框架信息生成AI模型,并将AI模型上报至网络设备,由网络设备对所接收AI模型进行模型转换,对此不做限制。
第三、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
则终端设备可以基于该第二指示信息获知网络设备小区不支持基于第一框架信息的AI模型推理。
第四、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息相同时,向终端设备发送指示终端设备转换模型并上报的消息,该消息也可以由第四指示信息指示。
则终端设备可以基于第四指示信息的指示,基于支持转换的框架信息转换AI模型,得到转换所得 AI模型,并将转换所得AI模型上报至网络设备,对此不做限制。
即第四指示信息指示终端设备上报模型,可以包括:指示终端设备生成模型并上报,或者指示终端设备转换模型并上报,对此不做限制。
第五、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息。
第六、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备支持模型转换时,向终端设备发送第四指示信息,其中,第四指示信息指示终端设备生成模型并上报。
则终端设备可以基于第四指示信息的指示,生成AI模型,并将AI模型上报至网络设备,由网络设备对所接收AI模型进行模型转换,对此不做限制。
第七、在确定终端设备支持部署的框架信息的数量是多个时,向终端设备发送第五指示信息,其中,第五指示信息指示终端设备从多个支持部署的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持部署的框架信息选择一个来生成模型,对此不做限制。
其中,网络设备可以自行决定从多个支持部署的框架信息中选择一个,并指示给终端设备,或者,网络设备还可以基于预定义规则从多个支持部署的框架信息中选择一个,并指示给终端设备,对此不做限制。
第八、在确定终端设备支持转换的框架信息的数量是多个时,向终端设备发送第五指示信息,其中,第五指示信息还可以指示终端设备从多个支持转换的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持转换的框架信息选择一个来转换模型,对此不做限制。
其中,网络设备可以自行决定从多个支持转换的框架信息中选择一个,并指示给终端设备,或者,网络设备还可以基于预定义规则从多个支持转换的框架信息中选择一个,并指示给终端设备,对此不做限制。
第九、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
上述提供了终端设备向网络设备上报AI模型的实施方式,在其他任意可能的实施方式中,也可以由网络设备作出是否由终端设备发送AI模型的决策,网络设备可以将该决策发送给终端设备,由终端设备基于网络设备的决策来进行相应的操作,或者也可以支持其他任意可能的实施方式,对此不做限制。
本实施例中,通过确定网络设备是否支持中间模型表示框架,得到第一确定结果,确定网络设备是否支持模型转换,得到第二确定结果,以及根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备,可以基于第一确定结果和/或第二确定结果灵活制定对应的发送策略,从而有效提升第一AI模型发送策略与个性化应用场景的适用性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收终端设备发送的第三AI模型,其中,第三AI模型是由终端设备基于网络设备所指示内容生成的,其中,所指示内容包括以下至少一项:第二指示信息指示的内容,由此,可以在网络设备小区不支持基于第一框架信息的AI模型推理时,由终端设备生成AI模型,并发送至网络设备侧。
其中,第三AI模型,是指终端设备基于网络设备所指示内容生成的AI模型。
图6是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图6所示,该方法可以包括但不限于如下步骤:
S106:从多个第一框架信息中选择一个第一框架信息。
S206:向终端设备发送第三指示信息,其中,第三指示信息用于向终端设备指示所选择第一框架信息。
其中,第三指示消息,可以被用于向终端设备指示所选择第一框架信息。
S306:基于所选择第一框架信息生成第一AI模型。
S406:将第一AI模型发送至终端设备。
本实施例中,通过从多个第一框架信息中选择一个第一框架信息,向终端设备发送第三指示信息,其中,第三指示信息用于向终端设备指示所选择第一框架信息,基于所选择第一框架信息生成第一AI模型,将第一AI模型发送至终端设备,基于第三指示消息准确指示所选择第一框架信息的选择过程,保证所生成第一AI模型的可靠性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,从多个第一框架信息中选择一个第一框架信息,还包括:确定网络设备所支持的第一候选框架信息,根据第一候选框架信息,从多个第一框架信息中选择一个第一框架信息,其中,所选择第一框架信息包括所选择第一模型框架,所选择第一框架信息是以下任一项:与第一候选框架信息相同的第一框架信息,或者,与第一候选框架信息不同的第一框架信息,且第一候选框架信息所属第一候选框架支持将模型转换为所选择第一框架信息所支持的模型,可以有效提升所选择第一框架信息的实用性。
图7是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图7所示,该方法可以包括但不限于如下步骤:
S107:基于所选择第一模型框架将第一AI模型转换为第四AI模型,由所选择第一框架信息所确定的第二模型框架支持模型转换,第一AI模型由网络设备基于第二模型框架生成。
其中,第四AI模型,可以是基于所选择第一模型框架对第一AI模型进行转换处理所得到的AI模型。
可以理解的是,第一AI模型与终端设备之间的匹配程度可能较低,而基于所选择第一模型框架将第一AI模型转换处理,可以较大程度地提升所得第四AI模型与终端设备之间的适配性。
S207:将第四AI模型发送至终端设备。
也即是说,本公开实施例中,如果第一AI模型不适用于终端设备,则可以在第二模型框架支持模型转换功能时,基于所选择第一模型框架对第一AI模型进行转换处理,以得到适用于终端设备的第四AI模型,并将所得第四AI模型发送至终端设备,以完成网络设备与终端设备之间的模型交互。
本实施例中,通过基于所选择第一模型框架将第一AI模型转换为第四AI模型,由所选择第一框架信息所确定的第二模型框架支持模型转换,第一AI模型由网络设备基于第二模型框架生成,将第四AI模型发送至终端设备,可以基于所选择第一模型框架实现不同AI模型之间的互相转换,可以有效提升网络设备与终端设备之间的模型交互与匹配效率。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备,包括:确定模型下发方式,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备支持模型转换,基于模型下发方式,将第一AI模型发送至终端设备,由此,可以在网络设备满足多种模型下发方式时,准确指示相匹配的AI模型下发 方式,以保障AI模型下发过程的可靠性。
举例而言,网络设备可以支持以下若干种处理逻辑:
第一、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息相同时,向终端设备发送指示终端设备生成模型并上报的消息,该消息可以被称为第二指示信息。
即,第二指示信息可以指示终端设备生成模型并上报。
第二、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示终端设备生成模型并上报。
第三、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
第四、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息相同时,向终端设备发送指示终端设备转换模型并上报的消息,该消息也可以由第二指示信息指示。
即第二指示信息指示终端设备上报模型,可以包括:指示终端设备生成模型并上报,或者指示终端设备转换模型并上报,对此不做限制。
第五、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息。
第六、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示终端设备生成模型并上报;
第七、在确定终端设备支持部署的框架信息的数量是多个时,向终端设备发送第二指示信息,其中,第二指示信息指示终端设备从多个支持部署的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持部署的框架信息选择一个来生成模型,对此不做限制。
其中,网络设备可以自行决定从多个支持部署的框架信息中选择一个,并指示给终端设备,或者,网络设备还可以基于预定义规则从多个支持部署的框架信息中选择一个,并指示给终端设备,对此不做限制。
第八、在确定终端设备支持转换的框架信息的数量是多个时,向终端设备发送第二指示信息,其中,第二指示信息还可以指示终端设备从多个支持转换的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持转换的框架信息选择一个来转换模型,对此不做限制。
其中,网络设备可以自行决定从多个支持转换的框架信息中选择一个,并指示给终端设备,或者,网络设备还可以基于预定义规则从多个支持转换的框架信息中选择一个,并指示给终端设备,对此不做限制。
第九、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
其中,模型下发方式,是指AI模型由网络设备侧发送至终端设备侧的方式。例如,可以是基于中间模型表示框架进行下发,或者,还可以是基于模型转换的方式下发。
其中,如果网络设备既支持中间模型表示框架,又支持模型转换,则可以由网络设备侧自行决定是基于中间模型表示框架进行AI模型的交互与部署,还是基于模型转换进行AI模型的交互与部署;当然,基站也可以基于预定义规则确定使用中间模型表示框架还是进行AI模型转换,或者,还可以采用其他任意可能的方式,对此不做限制。当然,一种可能的方式是,基站将最终决策的模型下发方式,指示给终端设备。
图8是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图8所示,该方法可以包括但不限于如下步骤:
S108:基于中间模型表示框架处理第一AI模型,得到AI模型文件,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换。
其中,AI模型文件,是指第一AI模型经由中间模型表示框架处理所得到的文件。该AI模型文件的格式存在多种可能。
S208:将AI模型文件发送至终端设备。
也即是说,本公开实施例在网络设备支持中间模型表示框架,且第二确定结果指示网络设备不支持模型转换时,及时基于中间模型表示框架处理第一AI模型,得到AI模型文件,并将AI模型文件发送至终端设备。
本实施例中,通过基于中间模型表示框架处理第一AI模型,得到AI模型文件,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换,将AI模型文件发送至终端设备,可以在网络设备支持中间模型表示框架但不支持模型转换时,及时对第一AI模型进行转换处理,以保障所得AI模型文件与终端设备之间的适配性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:向终端设备发送第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息。
其中,模型格式信息,可以是是指AI模型文件对应的存储格式。
针对网络设备的执行步骤可以举例说明如下:
1.网络设备(例如可以是基站)可能支持一种或多种AI框架,若支持多种,或者只支持一种但是可以实现直接的模型转化,比如TF->caffe,TF->NCNN,基站可以选择相匹配的框架生成AI模型或者转化成合适的模型下发给终端;如果基站只支持一种框架,并且不能实现模型转换,则向终端指示该小区下不支持启用模型推理。
其中,网络设备,可以确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,确定第二AI模型框架,
其中,第一框架信息包括以下任一项:终端设备支持部署的框架信息,或者终端设备支持转换的框架信息。
其中,网络设备,可以接收终端设备发送的第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
其中,网络设备,可以确定网络设备支持多个第一候选框架信息,基于第一框架信息和多个第一候选框架信息,确定第二模型框架。
其中,网络设备可以将第一候选框架信息所属候选框架作为第二模型框架;其中,第一候选框架信息与第一框架信息相同;或者,第一候选框架信息所属候选框架支持模型转换。
其中,网络设备可以基于第二模型框架生成第一AI模型,将第一AI模型发送至终端设备,网络设备支持的第一候选框架信息与第一框架信息相同,或者,基于第一框架信息将第一AI模型转换为第二AI模型,将第二AI模型发送至终端设备,第一候选框架支持模型转换,第一AI模型基于第一候选框架生成,或者,基于中间模型表示框架将第一AI模型转换为第二AI模型,将第二AI模型发送至终端 设备,终端设备和网络设备均支持中间模型表示框架,或者,不向终端设备发送第一AI模型,网络设备不支持中间模型表示框架,且网络设备不支持模型转换。
其中,网络设备可以向终端设备发送第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署,其中,网络设备满足以下至少一种:支持的第一候选框架不支持模型转换、不支持模型转换、不支持中间模型表示框架。
其中,网络设备可以向终端设备发送第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型,其中,网络设备满足以下一种:所支持的第一候选框架信息与第一框架信息相同,支持模型转换。
其中,网络设备可以向终端设备发送第二指示信息,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息。
其中,网络设备可以基于预定义协议,确定与第一框架信息对应的第二框架信息,将第二框架信息所属模型框架作为第二模型框架。
其中,预定义协议包括以下任一项:定义与终端设备使用相同框架信息,或者定义与终端设备使用相匹配的框架信息,或者定义与模型处理阶段对应的框架信息,或者定义基于中间模型表示框架处理模型。
其中,预定义协议具体例如可以规定具体可采用的AI框架类型:
方案一:规定网络设备与终端设备统一使用相同的AI框架,比如统一使用caffe或者caffe2。
方案二:网络设备与终端设备可以使用不同的AI框架。
方案2-1:预定义协议规定网络设备与终端设备使用预定义的能够相匹配的AI框架,比如网络设备统一使用pytorch,终端设备使用统一的框架如NCNN等;或者模型训练过程中使用统一的框架如pytorch,模型推理部署过程使用统一的框架如NCNN,模型训练完成在进行部署时,规定支持将pytorch转换为NCNN模型。
方案2-2:通过统一的中间模型表示框架进行模型翻译,例如可以规定统一使用onnx或者统一使用其他机构如IEEE标准化的AI模型表示架构。
其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定。
其中,模型处理阶段包括以下任一项:模型训练阶段,或者模型推理部署阶段。
2.如果终端上报了多个所支持的AI框架,可以由基站确定并指示唯一/其中相匹配的一个AI框架给到终端,实现AI框架的使能,此外,基站需要将AI模型转换成多个匹配的AI框架的其中一个。
其中,网络设备,可以向终端设备发送第二指示信息,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息。
其中,网络设备,可以确定网络设备是否支持中间模型表示框架,得到第一确定结果,确定网络设备是否支持模型转换,得到第二确定结果,以及根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备。
其中,网络设备,可以在基于所选择第一框架信息所确定的第二模型框架支持模型转换时,基于所选择第一模型框架将第一AI模型转换为第三AI模型,其中,第一AI模型由网络设备基于第二模型框架生成,将第三AI模型发送至终端设备。
其中,网络设备,可以基于所选择第一框架信息生成第一AI模型,将第一AI模型发送至终端设备。
其中,网络设备,可以接收终端设备发送的第三AI模型,其中,第三AI模型是由终端设备基于网络设备所指示内容生成的,其中,所指示内容包括以下至少一项:第二指示信息指示的内容。
其中,网络设备,可以从多个第一框架信息中选择一个第一框架信息,向终端设备发送第三指示信息,其中,第三指示信息用于向终端设备指示所选择第一框架信息,基于所选择第一框架信息生成第一AI模型,将第一AI模型发送至终端设备。
其中,网络设备,可以确定网络设备所支持的第一候选框架信息,根据第一候选框架信息,从多个第一框架信息中选择一个第一框架信息;其中,所选择第一框架信息包括所选择第一模型框架,所选择第一框架信息是以下任一项:与第一候选框架信息相同的第一框架信息,或者,与第一候选框架信息不同的第一框架信息,且第一候选框架信息所属第一候选框架支持将模型转换为所选择第一框架信息所支持的模型。
其中,网络设备,可以基于所选择第一模型框架将第一AI模型转换为第四AI模型,由所选择第一框架信息所确定的第二模型框架支持模型转换,第一AI模型由网络设备基于第二模型框架生成,将第四AI模型发送至终端设备。
其中,举例而言,网络设备可以支持以下若干种处理逻辑:
第一、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息相同时,向终端设备发送指示终端设备生成模型并上报的消息,该消息可以被称为第二指示信息。
即,第二指示信息指示终端设备生成模型并上报。
第二、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示终端设备生成模型并上报。
第三、在确定网络设备所支持的第一候选框架信息与终端设备支持部署的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
第四、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息相同时,向终端设备发送指示终端设备转换模型并上报的消息,该消息也可以由第二指示信息指示。
即第二指示信息指示终端设备上报模型,可以包括:指示终端设备生成模型并上报,或者指示终端设备转换模型并上报,对此不做限制。
第五、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备不支持模型转换时,向终端设备发送第二指示信息。
第六、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示终端设备生成模型并上报;
第七、在确定终端设备支持部署的框架信息的数量是多个时,向终端设备发送第三指示信息,其中,第三指示信息指示终端设备从多个支持部署的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持部署的框架信息选择一个来生成模型,对此不做限制。
其中,网络设备可以自行决定从多个支持部署的框架信息中选择一个,并指示给终端设备,或者,网络设备还可以基于预定义规则从多个支持部署的框架信息中选择一个,并指示给终端设备,对此不做限制。
第八、在确定终端设备支持转换的框架信息的数量是多个时,向终端设备发送第二指示信息,其中,第二指示信息还可以指示终端设备从多个支持转换的框架信息中选择一个。
其中,终端设备可以直接基于网络设备的指示从多个支持转换的框架信息选择一个来转换模型,对此不做限制。
其中,网络设备可以自行决定从多个支持转换的框架信息中选择一个,并指示给终端设备,或者, 网络设备还可以基于预定义规则从多个支持转换的框架信息中选择一个,并指示给终端设备,对此不做限制。
第九、在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同时,向终端设备发送第二指示信息。
其中,网络设备,可以接收终端设备发送的第三AI模型,其中,第三AI模型是由终端设备基于网络设备所指示内容生成的;其中,所指示内容包括以下至少一项:第二指示信息指示的内容。
其中,网络设备,可以在确定网络设备所支持的第一候选框架信息与终端设备支持转换的框架信息不同,且网络设备或终端设备不支持中间模型表示框架时,向终端设备下发第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
3.基站若同时支持AI模型转化和该中间模型表示框架,则可能还需要向终端指示具体采用哪种方式进行模型下发;此外,基站若不能实现模型转换但支持该AI中间框架,则可将模型转换成该中间框架所支持的模型格式下发给终端,同时如果模型文件本身不携带模型格式信息,可能需要指示对应的AI模型存储格式;若两者均不支持,则下发该小区不支持启用AI模型推理指示,或者通过不下发AI模型这一方式来隐式指示。
其中,网络设备可以确定模型下发方式,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备支持模型转换,基于模型下发方式,将第一AI模型发送至终端设备。
其中,网络设备可以基于中间模型表示框架处理第一AI模型,得到AI模型文件,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换,将AI模型文件发送至终端设备。
其中,网络设备可以向终端设备发送第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息。
其中,网络设备可以确定网络设备是否支持中间模型表示框架,得到第一确定结果,确定网络设备是否支持模型转换,得到第二确定结果,以及根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备。
其中,网络设备可以在第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备支持模型转换时,确定模型下发方式,基于模型下发方式,将第一AI模型发送至终端设备。
其中,网络设备可以在第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换时,基于中间模型表示框架处理第一AI模型,得到AI模型文件,将AI模型文件发送至终端设备。
其中,网络设备可以在确定AI模型文件未携带模型格式信息时,向终端设备发送第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式。
其中,网络设备可以在第一确定结果指示网络设备不支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换时,不向终端设备发送第一AI模型。
其中,网络设备可以在第一确定结果指示网络设备不支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换时,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备小区不支持基于第一框架信息的AI模型推理。
其中,网络设备可以在确定终端设备支持中间模型表示框架时,基于中间模型表示框架将第一AI模型转换处理为第五AI模型,将第五AI模型发送至终端设备。
其中,网络设备可以在确定终端设备支持中间模型表示框架时,确定网络设备支持中间模型表示框 架。
图9是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在终端设备中,如图9所示,该方法可以包括但不限于如下步骤:
S109:向网络设备指示支持的第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架。
也即是说,本公开实施例中,终端设备可以向网络设备发送第一框架信息,以向网络设备指示终端设备支持的第一框架信息。
本实施例中,还提供了一种异构人工智能AI框架的模型交互方法,可以于预定义协议,确定支持的第一框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,预定义协议包括以下任一项:定义与网络设备使用相同框架信息,或者定义与网络设备使用相匹配的框架信息,或者定义与模型处理阶段对应的框架信息,或者定义基于中间模型表示框架处理模型。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,在基于预定义协议和/或第二框架信息,确定支持的第一框架信息时,可以是将与网络设备支持的第二框架信息作为第一框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,在基于预定义协议和/或第二框架信息,确定支持的第一框架信息时,还可以是将与第二框架信息相匹配的框架信息作为第一框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,在基于预定义协议和/或第二框架信息,确定支持的第一框架信息时,还可以是将与终端设备所处模型处理阶段对应的框架信息作为第一框架信息。
本实施例中,通过向网络设备指示支持的第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,能够有效保障网络设备侧所生成的AI模型对于终端设备侧的适用性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,第一框架信息包括以下任一项:终端设备支持部署的框架信息、终端设备支持转换的框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:向网络设备发送第一指示信息,第一指示信息用于指示终端设备支持第一框架信息,终端设备可以基于第一指示信息向网络设备指示终端设备支持的第一框架信息。
图10是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在终端设备中,如图10所示,该方法可以包括但不限于如下步骤:
S110:确定网络设备支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二模型框架。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定,由此,可以使交互过程适配于个性化的应用场景,以有效扩展该异构人工智能AI框架的模型交互方法的适用范围。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,其中,模型处理阶段可以包括以下任一项:模型训练阶段,或者模型推理部署阶段。
S210:基于预定义协议和/或第二框架信息,确定支持的第一框架信息。
本实施例中,通过确定网络设备支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二模型框架,基于预定义协议和/或第二框架信息,确定支持的第一框架信息,有效提升第一框架信息确定过程的灵活性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的第一AI模型,其中,第一AI模型由网络设备生成,或者,接收网络设备发送的第二AI模型,其中,第二AI模型由网络设备基于第一框架信息或中间模型表示框架对第一AI模型转换得到,第一AI模型 由网络设备生成,可以有效提升终端设备所接收AI模型的适用性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署,可以使终端设备基于第二指示信息,及时获取网络设备所属小区不支持基于第一框架信息的AI模型推理的相关信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型,其中,终端设备满足以下一种:支持的第一框架信息与网络设备所支持的第一候选框架信息相同、支持的第一框架信息与网络设备所支持的第一候选框架信息不同,由此,可以第二指示信息准确是指终端设备的模型上报过程。
其中,第一候选框架信息与第一框架信息相同,例如第一候选框架信息指示的框架类型与第一框架信息指示的框架类型相同,或者第一候选框架信息指示的框架类型和版本号,与第一框架信息指示的框架类型和版本号相同,或者第一候选框架信息指示的框架类型与第一框架信息指示的框架类型不相同,但是均支持基于中间模型框架进行转换表达,对此不做限制。
其中,第一候选框架信息与第一框架信息不同,例如第一候选框架信息指示的框架类型与第一框架信息指示的框架类型不相同,或者第一候选框架信息指示的框架类型或版本号,与第一框架信息指示的框架类型或版本号不相同,或者第一候选框架信息指示的框架类型与第一框架信息指示的框架类型不相同,但是均不支持基于中间模型框架进行转换表达,对此不做限制。
由此,本公开实施例在交互过程中,可以基于第一候选框架信息与第一框架信息是否相同的情况,准确地、灵活地确定出与第一框架信息对应的第二模型框架。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的第二指示信息,其中,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息;其中,终端设备满足:支持的第一框架信息的数量是多个,由此,可以基于第二指示信息准确指示终端设备的第一框架信息选择过程。
图11是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在终端设备中,如图11所示,该方法可以包括但不限于如下步骤:
S111:根据第二指示信息指示的内容,确定网络设备所指示框架信息。
S211:基于所指示框架信息,生成第三AI模型。
其中,第三AI模型,可以是终端设备基于所指示框架信息所生成得AI模型。
S311:将第三AI模型发送至网络设备。
本实施例中,通过根据第二指示信息指示的内容,确定网络设备所指示框架信息,基于所指示框架信息,生成第三AI模型,将第三AI模型发送至网络设备,可以有效提升第三AI模型与网络设备之间的适配性。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的第四AI模型。
其中,第四AI模型,可以是指网络设备侧所生成的AI模型。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送的AI模型文件,其中,AI模型文件由网络设备基于中间模型表示框架处理第一AI模型得到,第一AI模型由网络设备基于第二模型框架生成,第二模型框架是基于第一框架信息确定,由此,可以基于中间模型表示框架实现第一AI模型的跨框架传输。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收网络设备发送 的第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息,由此,终端设备可以基于第二指示信息准确获取AI模型文件的AI模型存储格式,能够有效避免存储格式未知而影响AI模型文件的存储效果。
针对终端设备的执行步骤可以举例说明如下:
1.终端向基站上报自己所支持的AI框架,可以为一种,也可以为多种。
其中,终端设备可以向网络设备指示支持的第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架。
其中,第一框架信息包括以下任一项:终端设备支持部署的框架信息,或者终端设备支持转换的框架信息。
其中,终端设备可以向网络设备发送第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
其中,终端设备可以接收网络设备发送的第一人工智能AI模型,其中,第一AI模型由网络设备生成。
其中,终端设备可以接收网络设备发送的第二AI模型,其中,第二AI模型由网络设备基于第一框架信息对第一AI模型转换处理得到,第一AI模型由网络设备生成。
其中,终端设备可以接收网络设备发送的第二指示信息,其中,第二指示信息指示网络设备所属小区不支持基于第一框架信息的AI模型推理。
其中,终端设备可以向网络设备发送第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
其中,终端设备可以确定网络设备支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二模型框架,基于预定义协议和/或第二框架信息,确定支持的第一框架信息。
其中,终端设备可以接收网络设备发送的第一AI模型,其中,第一AI模型由网络设备生成,或者,接收网络设备发送的第二AI模型,其中,第二AI模型由网络设备基于第一框架信息或中间模型表示框架对第一AI模型转换得到,第一AI模型由网络设备生成。
其中,终端设备可以接收网络设备发送的第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署。
其中,终端设备可以接收网络设备发送的第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型,其中,终端设备满足以下一种:支持的第一框架信息与网络设备所支持的第一候选框架信息相同、支持的第一框架信息与网络设备所支持的第一候选框架信息不同。
其中,终端设备可以接收网络设备发送的第二指示信息,其中,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息,其中,终端设备满足:支持的第一框架信息的数量是多个。
其中,终端设备可以根据第二指示信息指示的内容,确定网络设备所指示框架信息;基于所指示框架信息,生成第三AI模型,将第三AI模型发送至网络设备。
其中,终端设备可以接收网络设备发送的第四AI模型。
2.终端可以上报是否支持使用中间模型表示框架,以及支持采用哪种/哪几种中间模型表示框架。
其中,终端设备可以基于预定义协议,确定支持中间模型表示框架。
其中,终端设备可以确定支持中间模型表示框架,向网络设备发送第二指示信息,其中,第二指示信息指示终端设备支持中间模型表示框架。
其中,终端设备可以接收网络设备发送的AI模型文件,其中,AI模型文件由网络设备基于中间模型表示框架处理第一AI模型得到,第一AI模型由网络设备基于第二模型框架生成,第二模型框架是基于第一框架信息确定。
3.终端上报自己所支持的AI框架的版本信息等。
其中,终端设备可以接收网络设备发送的第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息。
其中,终端设备可以确定网络设备支持的第二框架信息,基于预定义协议和/或第二框架信息,确定支持的第一框架信息。
其中,预定义协议包括以下任一项:定义与网络设备使用相同框架信息,或者定义与网络设备使用相匹配的框架信息,或者定义与模型处理阶段对应的框架信息,或者定义基于中间模型表示框架处理模型。
其中,预定义协议具体例如,可以规定具体可采用的AI框架类型:
方案一:规定网络设备与终端设备统一使用相同的AI框架,比如统一使用caffe或者caffe2。
方案二:网络设备与终端设备可以使用不同的AI框架。
方案2-1:预定义协议规定网络设备与终端设备使用预定义的能够相匹配的AI框架,比如网络设备统一使用pytorch,终端设备使用统一的框架如NCNN等;或者模型训练过程中使用统一的框架如pytorch,模型推理部署过程使用统一的框架如NCNN,模型训练完成在进行部署时,规定支持将pytorch转换为NCNN模型。
方案2-2:通过统一的中间模型表示框架进行模型翻译,例如可以规定统一使用onnx或者统一使用其他机构如IEEE标准化的AI模型表示架构。
其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定。
其中,终端设备可以将与网络设备支持的第二框架信息作为第一框架信息,或者将与第二框架信息相匹配的框架信息作为第一框架信息,或者将与终端设备所处模型处理阶段对应的框架信息作为第一框架信息。
其中,模型处理阶段包括以下任一项:模型训练阶段、模型推理部署阶段。
其中,终端设备可以接收网络设备发送的AI模型文件,其中,AI模型文件由网络设备基于中间模型表示框架处理第一AI模型得到,第一AI模型由网络设备基于第二模型框架生成,第二模型框架是基于第一框架信息确定。
图12是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在网络设备中,如图12所示,该方法可以包括但不限于如下步骤:
S112:向终端设备指示所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。
本实施例中,通过向终端设备指示所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架,可以使终端设备准确获取网络设备所支持的第二框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,第二框架信息包括以下任一项:网络设备支持部署的框架信息、网络设备支持转换的框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:接收终端设备发送的第一AI模型,其中,第一AI模型由终端设备基于支持的第一框架信息所包括的第一模型框架生成,第一框架信息是基于第二框架信息确定,或者,接收终端设备发送的第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息,或者,向终端设备发送第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架,或者,接收终端设备发送的第二AI模型,其中,第二AI模型由终端设备生成。
针对网络设备的执行步骤可以举例说明如下:
1.网络设备可以向终端指示自己所支持的全部AI框架。
其中,网络设备,可以向终端设备指示所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。
其中,第二框架信息包括以下任一项:网络设备支持部署的框架信息,网络设备支持转换的框架信 息。
2.由基站进行AI框架的匹配,和/或做出是否进行模型下发的决策,和/或做出具体采用哪种AI框架的决策。此外,还可以预定义一组终端侧和基站侧AI框架转换规则。
其中,网络设备可以接收终端设备发送的第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息。
其中,第一框架信息包括以下任一项:终端设备支持部署的框架信息,或者终端设备支持转换的框架信息。
可以基于预定义协议的方式,其中,预定义协议具体例如,可以规定具体可采用的AI框架类型:
方案一:规定网络设备与终端设备统一使用相同的AI框架,比如统一使用caffe或者caffe2。
方案二:网络设备与终端设备可以使用不同的AI框架。
方案2-1:预定义协议规定网络设备与终端设备使用预定义的能够相匹配的AI框架,比如网络设备统一使用pytorch,终端设备使用统一的框架如NCNN等;或者模型训练过程中使用统一的框架如pytorch,模型推理部署过程使用统一的框架如NCNN,模型训练完成在进行部署时,规定支持将pytorch转换为NCNN模型。
方案2-2:通过统一的中间模型表示框架进行模型翻译,例如可以规定统一使用onnx或者统一使用其他机构如IEEE标准化的AI模型表示架构。
其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定。
3.模型转换可以在终端设备侧完成。
其中,网络设备可以接收终端设备发送的第一AI模型,其中,第一AI模型由终端设备基于支持的第一框架信息所包括的第一模型框架生成,第一框架信息是基于第二框架信息确定。
4.网络设备可以指示自己所支持的中间模型表达框架。
其中,网络设备可以向终端设备发送第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架。
其中,网络设备可以接收终端设备发送的第二AI模型,其中,第二AI模型由终端设备生成。
图13是本公开实施例提供的另一种异构人工智能AI框架的模型交互方法的流程示意图,本实施例中的异构人工智能AI框架的模型交互方法可以应用在终端设备中,如图18所示,该方法可以包括但不限于如下步骤:
S113:确定网络设备所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。
本实施例中,通过确定网络设备所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架,可以为终端设备确定所使用的AI模型框架提供可靠的参考依据。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,第二框架信息包括以下任一项:网络设备支持部署的框架信息、网络设备支持转换的框架信息。
本公开实施例中还提供了一种异构人工智能AI框架的模型交互方法,还包括:确定与第二框架信息对应的第一模型框架,并基于第一模型框架生成第一AI模型,以及向网络设备发送第一AI模型,或者,向网络设备发送第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息,或者,接收网络设备发送的第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架,或者,确定终端设备是否支持中间模型表示框架,得到第三确定结果,其中,网络设备支持中间模型表示框架,并确定终端设备是否支持模型转换,得到第四确定结果,以及根据第三确定结果和/或第四确定结果,将第二AI模型发送至网络设备,其中,第二AI模型由终端设备生成。
针对终端设备的执行步骤可以举例说明如下:
1.终端设备可以上报自己所支持的AI框架或可支持转换的基站侧AI框架。
其中,终端设备可以确定与第二框架信息对应的第一模型框架,基于第一模型框架生成第一AI模 型。
其中,第二框架信息包括以下任一项:网络设备支持部署的框架信息,网络设备支持转换的框架信息。
2.由终端设备自行确定是否能够完成不同AI框架间模型的转换,如果能够完成转换,则将转换后或者转换前的模型发送给基站。可选的,如果传输的模型本身中如果没有任何文件格式等相关信息,则还需要将模型对应的AI框架类型指示给基站,以便于基站侧进行解析。
其中,终端设备可以向网络设备发送第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息。
3.模型转换可以在终端设备侧完成,也可以在网络设备侧完成。
其中,终端设备可以确定与第二框架信息对应的第一模型框架,并基于第一模型框架生成第一AI模型,以及向网络设备发送第一AI模型。
其中,终端设备可以基于预定义协议,确定与第二框架信息对应的第一模型框架信息,将第一模型框架信息所属模型框架作为第一模型框架。
其中,预定义协议具体例如,可以规定具体可采用的AI框架类型:
方案一:规定网络设备与终端设备统一使用相同的AI框架,比如统一使用caffe或者caffe2。
方案二:网络设备与终端设备可以使用不同的AI框架。
方案2-1:预定义协议规定网络设备与终端设备使用预定义的能够相匹配的AI框架,比如网络设备统一使用pytorch,终端设备使用统一的框架如NCNN等;或者模型训练过程中使用统一的框架如pytorch,模型推理部署过程使用统一的框架如NCNN,模型训练完成在进行部署时,规定支持将pytorch转换为NCNN模型。
方案2-2:通过统一的中间模型表示框架进行模型翻译,例如可以规定统一使用onnx或者统一使用其他机构如IEEE标准化的AI模型表示架构。
其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定。
4.终端设备可以指示自己所支持的中间模型表达框架。
其中,终端设备可以接收网络设备发送的第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架。
其中,终端设备可以确定终端设备是否支持中间模型表示框架,得到第三确定结果,其中,网络设备支持中间模型表示框架,并确定终端设备是否支持模型转换,得到第四确定结果,以及根据第三确定结果和/或第四确定结果,将第二AI模型发送至网络设备,其中,第二AI模型由终端设备生成。
其中,终端设备可以确定终端设备所支持的第二候选框架信息,基于第二候选框架信息,确定与第二框架信息对应的第一模型框架。
其中,终端设备可以在确定第二候选框架信息与第二框架信息相同时,将第二候选框架信息所属候选框架作为第一模型框架,在确定第二候选框架信息与第二框架信息不同,且所属候选框架支持模型转换时,将第二候选框架信息所属候选框架作为第一模型框架。
其中,预定义协议包括以下任一项:定义与网络设备使用相同框架信息,或者定义与网络设备使用相匹配的框架信息,或者定义与模型处理阶段对应的框架信息,或者定义基于中间模型表示框架处理模型。
其中,第一框架信息与第二框架信息相同,或者第一框架信息与第二框架信息不相同,或者第一框架信息与第二框架信息不相同,且第二框架信息是基于模型处理阶段确定。
其中,模型处理阶段包括以下任一项:模型训练阶段、模型推理部署阶段。
图14为本公开实施例提供的一种通信装置的结构示意图。图14所示的通信装置140可包括收发模块1401和处理模块1402。收发模块1401可包括发送模块和/或接收模块,发送模块用于实现发送功能,接收模块用于实现接收功能,收发模块1401可以实现发送功能和/或接收功能。
通信装置140可以是终端设备(如前述方法实施例中的终端设备),也可以是终端设备中的装置,还可以是能够与终端设备匹配使用的装置。或者,通信装置140可以是网络设备(如前述方法实施例中的网络设备),也可以是网络设备中的装置,还可以是能够与网络设备匹配使用的装置。
通信装置140,在网络设备侧,该装置包括:
处理模块1402,用于:确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持 的第一AI模型框架;确定第二AI模型框架。
可选的,第一框架信息还包括以下任一项:
终端设备支持部署的框架信息;
终端设备支持转换的框架信息。
可选的,该装置,该包括:
收发模块1401,用于接收终端设备发送的第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
可选的,处理模块1402,还用于:
确定网络设备支持多个第一候选框架信息;
基于第一框架信息和多个第一候选框架信息,确定第二模型框架。
可选的,处理模块1402,还用于:
将第一候选框架信息所属候选框架作为第二模型框架;其中,
第一候选框架信息与第一框架信息相同;或者,
第一候选框架信息所属候选框架支持模型转换。
可选的,处理模块1402,还用于:
基于第二模型框架生成第一AI模型。
可选的,收发模块1401,还用于:
将第一AI模型发送至终端设备,网络设备支持的第一候选框架信息与第一框架信息相同。
可选的,处理模块1402,还用于:
基于第一框架信息将第一AI模型转换为第二AI模型。
可选的,收发模块1401,还用于:
将第二AI模型发送至终端设备,第一候选框架支持模型转换,第一AI模型基于第一候选框架生成。
可选的,处理模块1402,还用于:
基于中间模型表示框架将第一AI模型转换为第二AI模型。
可选的,收发模块1401,还用于:
将第二AI模型发送至终端设备,终端设备和网络设备均支持中间模型表示框架
可选的,收发模块1401,还用于:。
不向终端设备发送第一AI模型,网络设备不支持中间模型表示框架,且网络设备不支持模型转换。
可选的,收发模块1401,还用于:
向终端设备发送第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署;
其中,网络设备满足以下至少一种:
支持的第一候选框架不支持模型转换;
不支持模型转换;
不支持中间模型表示框架。
可选的,收发模块1401,还用于:
向终端设备发送第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型;
其中,网络设备满足以下一种:
所支持的第一候选框架信息与第一框架信息相同;
支持模型转换。
可选的,收发模块1401,还用于:
向终端设备发送第二指示信息,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息。
可选的,处理模块1402,还用于:
确定网络设备是否支持中间模型表示框架,得到第一确定结果;
确定网络设备是否支持模型转换,得到第二确定结果。
可选的,收发模块1401,还用于:
根据第一确定结果和/或第二确定结果,将第一AI模型发送至终端设备。
可选的,收发模块1401,还用于:
接收终端设备发送的第三AI模型,其中,第三AI模型是由终端设备基于网络设备所指示内容生成的;
其中,所指示内容包括以下至少一项:
第二指示信息指示的内容。
可选的,处理模块1402,还用于:
从多个第一框架信息中选择一个第一框架信息;
可选的,收发模块1401,还用于:
向终端设备发送第三指示信息,其中,第三指示信息用于向终端设备指示所选择第一框架信息。
可选的,处理模块1402,还用于:
基于所选择第一框架信息生成第一AI模型。
可选的,收发模块1401,还用于:
将第一AI模型发送至终端设备。
可选的,处理模块1402,还用于:
确定网络设备所支持的第一候选框架信息;
根据第一候选框架信息,从多个第一框架信息中选择一个第一框架信息;其中,所选择第一框架信息包括所选择第一模型框架,所选择第一框架信息是以下任一项:
与第一候选框架信息相同的第一框架信息;或者,
与第一候选框架信息不同的第一框架信息,且第一候选框架信息所属第一候选框架支持将模型转换为所选择第一框架信息所支持的模型。
可选的,处理模块1402,还用于:
基于所选择第一模型框架将第一AI模型转换为第四AI模型,由所选择第一框架信息所确定的第二模型框架支持模型转换,第一AI模型由网络设备基于第二模型框架生成。
可选的,收发模块1401,还用于:
将第四AI模型发送至终端设备。
可选的,处理模块1402,还用于:
确定模型下发方式,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备支持模型转换;
可选的,收发模块1401,还用于:
基于模型下发方式,将第一AI模型发送至终端设备。
可选的,处理模块1402,还用于:
基于中间模型表示框架处理第一AI模型,得到AI模型文件,第一确定结果指示网络设备支持中间模型表示框架,且第二确定结果表示网络设备不支持模型转换。
可选的,收发模块1401,还用于:
将AI模型文件发送至终端设备。
可选的,收发模块1401,还用于:
向终端设备发送第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息。
通过实施本公开的方法,网络设备可以确定终端设备支持第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架,确定第二AI模型框架。通过实施本公开的方法,有效避免网络设备与终端设备间模型框架的差异对人工智能AI模型交互造成影响,从而能够有效保障人工智能AI模型的传输准确率,提升人工智能AI模型的交互效果。
通信装置140,在终端设备侧,该装置包括:
收发模块1401,用于向网络设备指示支持的第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架。
可选的,第一框架信息包括以下任一项:
终端设备支持部署的框架信息;
终端设备支持转换的框架信息。
可选的,收发模块1401,还用于:
向网络设备发送第一指示信息,第一指示信息用于指示终端设备支持第一框架信息。
可选的,该装置,还包括:
处理模块1402,用于确定网络设备支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二模型框架;基于预定义协议和/或第二框架信息,确定支持的第一框架信息。
可选的,收发模块1401,还用于:
接收网络设备发送的第一AI模型,其中,第一AI模型由网络设备生成;或者,
接收网络设备发送的第二AI模型,其中,第二AI模型由网络设备基于第一框架信息或中间模型表示框架对第一AI模型转换得到,第一AI模型由网络设备生成。
可选的,收发模块1401,还用于:
接收网络设备发送的第二指示信息,第二指示信息用于指示网络设备所属小区不支持基于第一框架信息的AI模型部署。
可选的,收发模块1401,还用于:
接收网络设备发送的第二指示信息,第二指示信息用于指示终端设备向网络设备上报模型;
其中,终端设备满足以下一种:
支持的第一框架信息与网络设备所支持的第一候选框架信息相同;
支持的第一框架信息与网络设备所支持的第一候选框架信息不同。
可选的,收发模块1401,还用于:
接收网络设备发送的第二指示信息,其中,第二指示信息用于指示终端设备从多个第一框架信息中选择一个第一框架信息;
其中,终端设备满足:支持的第一框架信息的数量是多个。
可选的,处理模块1402,还用于:
根据第二指示信息指示的内容,确定网络设备所指示框架信息;
基于所指示框架信息,生成第三AI模型;
可选的,收发模块1401,还用于将第三AI模型发送至网络设备。
可选的,收发模块1401,还用于:
接收网络设备发送的第四AI模型。
可选的,收发模块1401,还用于:
接收网络设备发送的AI模型文件,其中,AI模型文件由网络设备基于中间模型表示框架处理第一AI模型得到,第一AI模型由网络设备基于第二模型框架生成,第二模型框架是基于第一框架信息确定。
可选的,收发模块1401,还用于:
接收网络设备发送的第二指示信息,其中,第二指示信息指示AI模型文件的AI模型存储格式,AI模型文件未携带模型格式信息。
通过实施本公开的方法,终端设备可以向网络设备指示支持的第一框架信息,其中,第一框架信息包括终端设备支持的第一AI模型框架。通过实施本公开的方法,能够有效保障网络设备侧所生成的AI模型对于终端设备侧的适用性。
通信装置140,在网络设备侧,该装置包括:
收发模块1401,用于向终端设备指示所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。
可选的,第二框架信息包括以下任一项:
网络设备支持部署的框架信息;
网络设备支持转换的框架信息。
可选的,收发模块1401,还用于:
接收终端设备发送的第一AI模型,其中,第一AI模型由终端设备基于支持的第一框架信息所包括的第一模型框架生成,第一框架信息是基于第二框架信息确定;或者,
接收终端设备发送的第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息;或者,
向终端设备发送第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架;或者,
接收终端设备发送的第二AI模型,其中,第二AI模型由终端设备生成。
通过实施本公开的方法,网络设备可以向终端设备指示所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。通过实施本公开的方法,可以使终端设备准确获取网络设备所支持的第二框架信息。
通信装置140,在终端设备侧,该装置包括:
处理模块1402,用于确定网络设备所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。
可选的,第二框架信息包括以下任一项:
网络设备支持部署的框架信息;
网络设备支持转换的框架信息。
可选的,处理模块1402,还用于:
确定与第二框架信息对应的第一模型框架,并基于第一模型框架生成第一AI模型。
可选的,收发模块1401,还用于向网络设备发送第一AI模型。
可选的,收发模块1401,还用于向网络设备发送第一指示信息,其中,第一指示信息指示第一框架信息以及指示第一AI模型未携带模型格式信息。
可选的,收发模块1401,还用于接收网络设备发送的第二指示信息,其中,第二指示信息指示网络设备支持中间模型表示框架;或者,
可选的,处理模块1402,还用于确定终端设备是否支持中间模型表示框架,得到第三确定结果,其中,网络设备支持中间模型表示框架,并确定终端设备是否支持模型转换,得到第四确定结果。
可选的,收发模块1401,还用于根据第三确定结果和/或第四确定结果,将第二AI模型发送至网络设备,其中,第二AI模型由终端设备生成。
通过实施本公开的方法,终端设备可以确定网络设备所支持的第二框架信息,其中,第二框架信息包括网络设备支持的第二AI模型框架。通过实施本公开的方法,可以为终端设备确定所使用的AI模型框架提供可靠的参考依据。
图15是本公开实施例提供的另一种通信装置的结构示意图。通信装置150可以是网络设备(如前述方法实施例中的网络设备),也可以是终端设备(如前述方法实施例中的终端设备),也可以是支持网络设备实现上述方法的芯片、芯片系统、或处理器等,还可以是支持终端设备实现上述方法的芯片、芯片系统、或处理器等。该装置可用于实现上述方法实施例中描述的方法,具体可以参见上述方法实施例中的说明。
通信装置150可以包括一个或多个处理器1501。处理器1501可以是通用处理器或者专用处理器等。例如可以是基带处理器或中央处理器。基带处理器可以用于对通信协议以及通信数据进行处理,中央处理器可以用于对通信装置(如,基站、基带芯片,终端设备、终端设备芯片,DU或CU等)进行控制,执行计算机程序,处理计算机程序的数据。
可选的,通信装置150中还可以包括一个或多个存储器1502,其上可以存有计算机程序1504,处理器1501中可以存有计算机程序1503,处理器1501执行所述计算机程序1504和/或计算机程序1503,以使得通信装置150执行上述方法实施例中描述的方法。可选的,所述存储器1502中还可以存储有数据。通信装置150和存储器1502可以单独设置,也可以集成在一起。
可选的,通信装置150还可以包括收发器1505、天线1506。收发器1505可以称为收发单元、收发机、或收发电路等,用于实现收发功能。收发器1505可以包括接收器和发送器,接收器可以称为接收机或接收电路等,用于实现接收功能;发送器可以称为发送机或发送电路等,用于实现发送功能。
可选的,通信装置150中还可以包括一个或多个接口电路1507。接口电路1507用于接收代码指令并传输至处理器1501。处理器1501运行所述代码指令以使通信装置150执行上述方法实施例中描述的方法。
在一种实现方式中,处理器1501中可以包括用于实现接收和发送功能的收发器。例如该收发器可以是收发电路,或者是接口,或者是接口电路。用于实现接收和发送功能的收发电路、接口或接口电路可以是分开的,也可以集成在一起。上述收发电路、接口或接口电路可以用于代码/数据的读写,或者,上述收发电路、接口或接口电路可以用于信号的传输或传递。
在一种实现方式中,处理器1501可以存有计算机程序1503,计算机程序1503在处理器1501上运行,可使得通信装置150执行上述方法实施例中描述的方法。计算机程序1503可能固化在处理器1501中,该种情况下,处理器1501可能由硬件实现。
在一种实现方式中,通信装置150可以包括电路,所述电路可以实现前述方法实施例中发送或接收或者通信的功能。本公开中描述的处理器和收发器可实现在集成电路(integrated circuit,IC)、模拟IC、射频集成电路RFIC、混合信号IC、专用集成电路(application specific integrated circuit,ASIC)、印刷电路板(printed circuit board,PCB)、电子设备等上。该处理器和收发器也可以用各种IC工艺技术来制造,例如互补金属氧化物半导体(complementary metal oxide semiconductor,CMOS)、N型金属氧化物半导体(nMetal-oxide-semiconductor,NMOS)、P型金属氧化物半导体(positive channel metal oxide semiconductor,PMOS)、双极结型晶体管(bipolar junction transistor,BJT)、双极CMOS(BiCMOS)、硅锗(SiGe)、砷化镓(GaAs)等。
以上实施例描述中的通信装置可以是网络设备如前述方法实施例中的网络设备)或者终端设备(如前述方法实施例中的终端设备),但本公开中描述的通信装置的范围并不限于此,而且通信装置的结构可以不受图15的限制。通信装置可以是独立的设备或者可以是较大设备的一部分。例如所述通信装置可以是:
(1)独立的集成电路IC,或芯片,或,芯片系统或子系统;
(2)具有一个或多个IC的集合,可选的,该IC集合也可以包括用于存储数据,计算机程序的存储部件;
(3)ASIC,例如调制解调器(Modem);
(4)可嵌入在其他设备内的模块;
(5)接收机、终端设备、智能终端设备、蜂窝电话、无线设备、手持机、移动单元、车载设备、网络设备、云设备、人工智能设备等等;
(6)其他等等。
对于通信装置可以是芯片或芯片系统的情况,可参见图16所示的芯片的结构示意图。图16所示的芯片包括处理器1601和接口1602。其中,处理器1601的数量可以是一个或多个,接口1602的数量可以是多个。
对于芯片用于实现本申请实施例中网络设备的功能的情况:
处理器1601,用于实现图2中的S102,或者用于实现图5中的S105和S205,或者,用于实现图6中的S106和S306等。
接口1602,用于实现图3中的S103,或者用于实现图4中的S104,或者用于实现图6中的S206和S406等。
对于芯片用于实现本申请实施例中终端设备的功能的情况:
接口1602,用于实现图9中的S109,或者用于实现图11中的S311等。
处理器1601,用于实现图10中的S110和S210,或者用于实现图11中的S111和S211等。
可选的,芯片还包括存储器1603,存储器1603用于存储必要的计算机程序和数据。
本领域技术人员还可以了解到本公开实施例列出的各种说明性逻辑块(illustrative logical block)和步骤(step)可以通过电子硬件、电脑软件,或两者的结合进行实现。这样的功能是通过硬件还是软件来实现取决于特定的应用和整个系统的设计要求。本领域技术人员可以对于每种特定的应用,可以使用各种方法实现所述的功能,但这种实现不应被理解为超出本公开实施例保护的范围。
本公开实施例还提供一种通信系统,该系统包括前述图14实施例中作为网络设备(如前述方法实施例中的网络设备)的通信装置和作为终端设备(如前述方法实施例中的终端设备)的通信装置,或者,该系统包括前述图15实施例中作为网络设备(如前述方法实施例中的网络设备)的通信装置和作为终端设备(如前述方法实施例中的终端设备)的通信装置。
本公开还提供一种可读存储介质,其上存储有指令,该指令被计算机执行时实现上述任一方法实施例的功能。
本公开还提供一种计算机程序产品,该计算机程序产品被计算机执行时实现上述任一方法实施例的功能。
在上述实施例中,可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时,可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机程序。在计算机上加载和执行所述计算机程序时,全部或部分地产生按照本公开实施例所述的流程或功能。 所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机程序可以存储在计算机可读存储介质中,或者从一个计算机可读存储介质向另一个计算机可读存储介质传输,例如,所述计算机程序可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(digital subscriber line,DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如,软盘、硬盘、磁带)、光介质(例如,高密度数字视频光盘(digital video disc,DVD))、或者半导体介质(例如,固态硬盘(solid state disk,SSD))等。
本领域普通技术人员可以理解:本公开中涉及的第一、第二等各种数字编号仅为描述方便进行的区分,并不用来限制本公开实施例的范围,也表示先后顺序。
本公开中的至少一个还可以描述为一个或多个,多个可以是两个、三个、四个或者更多个,本公开不做限制。在本公开实施例中,对于一种技术特征,通过“第一”、“第二”、“第三”、“A”、“B”、“C”和“D”等区分该种技术特征中的技术特征,该“第一”、“第二”、“第三”、“A”、“B”、“C”和“D”描述的技术特征间无先后顺序或者大小顺序。
本公开中各表所示的对应关系可以被配置,也可以是预定义的。各表中的信息的取值仅仅是举例,可以配置为其他值,本公开并不限定。在配置信息与各参数的对应关系时,并不一定要求必须配置各表中示意出的所有对应关系。例如,本公开中的表格中,某些行示出的对应关系也可以不配置。又例如,可以基于上述表格做适当的变形调整,例如,拆分,合并等等。上述各表中标题示出参数的名称也可以采用通信装置可理解的其他名称,其参数的取值或表示方式也可以通信装置可理解的其他取值或表示方式。上述各表在实现时,也可以采用其他的数据结构,例如可以采用数组、队列、容器、栈、线性表、指针、链表、树、图、结构体、类、堆、散列表或哈希表等。
本公开中的预定义可以理解为定义、预先定义、存储、预存储、预协商、预配置、固化、或预烧制。
本领域普通技术人员可以意识到,结合本文中所公开的实施例描述的各示例的单元及算法步骤,能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行,取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能,但是这种实现不应认为超出本公开的范围。
所属领域的技术人员可以清楚地了解到,为描述的方便和简洁,上述描述的系统、装置和单元的具体工作过程,可以参考前述方法实施例中的对应过程,在此不再赘述。
以上所述,仅为本公开的具体实施方式,但本公开的保护范围并不局限于此,任何熟悉本技术领域的技术人员在本公开揭露的技术范围内,可轻易想到变化或替换,都应涵盖在本公开的保护范围之内。因此,本公开的保护范围应以所述权利要求的保护范围为准。

Claims (40)

  1. 一种异构人工智能AI框架的模型交互方法,其特征在于,由网络设备执行,所述方法包括:
    确定终端设备支持第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架;
    确定第二AI模型框架。
  2. 如权利要求1所述的方法,其特征在于,所述第一框架信息还包括以下任一项:
    所述终端设备支持部署的框架信息;
    所述终端设备支持转换的框架信息。
  3. 如权利要求1或2所述的方法,其特征在于,所述方法还包括:
    接收所述终端设备发送的第一指示信息,所述第一指示信息用于指示所述终端设备支持第一框架信息。
  4. 如权利要求2所述的方法,其特征在于,所述确定第二模型框架,包括:
    确定所述网络设备支持多个第一候选框架信息;
    基于所述第一框架信息和所述多个第一候选框架信息,确定所述第二模型框架。
  5. 如权利要求4所述的方法,其特征在于,所述基于所述第一框架信息和所述多个第一候选框架信息,确定所述第二模型框架,包括:
    将所述第一候选框架信息所属候选框架作为所述第二模型框架;其中,
    所述第一候选框架信息与所述第一框架信息相同;或者,
    所述第一候选框架信息所属候选框架支持模型转换。
  6. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    基于所述第二模型框架生成第一AI模型;
    将第一AI模型发送至所述终端设备,所述网络设备支持的第一候选框架信息与所述第一框架信息相同;或者,
    基于所述第一框架信息将第一AI模型转换为第二AI模型;
    将所述第二AI模型发送至所述终端设备,所述第一候选框架支持模型转换,所述第一AI模型基于所述第一候选框架生成;或者,
    基于中间模型表示框架将第一AI模型转换为第二AI模型;
    将所述第二AI模型发送至所述终端设备,所述终端设备和所述网络设备均支持所述中间模型表示框架;或者,
    不向所述终端设备发送所述第一AI模型,所述网络设备不支持所述中间模型表示框架,且所述网络设备不支持模型转换。
  7. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述网络设备所属小区不支持基于所述第一框架信息的AI模型部署;
    其中,所述网络设备满足以下至少一种:
    支持的第一候选框架不支持模型转换;
    不支持模型转换;
    不支持中间模型表示框架。
  8. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述终端设备向所述网络设备上报模型;
    其中,所述网络设备满足以下一种:
    所支持的第一候选框架信息与所述第一框架信息相同;
    支持模型转换。
  9. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    向所述终端设备发送第二指示信息,所述第二指示信息用于指示所述终端设备从多个所述第一框架信息中选择一个第一框架信息。
  10. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    确定所述网络设备是否支持所述中间模型表示框架,得到第一确定结果;
    确定所述网络设备是否支持模型转换,得到第二确定结果,以及
    根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备。
  11. 如权利要求8所述的方法,其特征在于,所述方法还包括:
    接收所述终端设备发送的第三AI模型,其中,所述第三AI模型是由所述终端设备基于所述网络设备所指示内容生成的;
    其中,所述所指示内容包括以下至少一项:
    所述第二指示信息指示的内容。
  12. 如权利要求2所述的方法,其特征在于,所述方法还包括:
    从多个所述第一框架信息中选择一个第一框架信息;
    向所述终端设备发送第三指示信息,其中,所述第三指示信息用于向所述终端设备指示所选择第一框架信息;
    基于所述所选择第一框架信息生成第一AI模型,
    将所述第一AI模型发送至所述终端设备。
  13. 如权利要求12所述的方法,其特征在于,所述从多个第一框架信息中选择一个第一框架信息,还包括:
    确定所述网络设备所支持的第一候选框架信息;
    根据所述第一候选框架信息,从多个第一框架信息中选择一个第一框架信息;其中,所述所选择第一框架信息包括所选择第一模型框架,所述所选择第一框架信息是以下任一项:
    与所述第一候选框架信息相同的第一框架信息;或者,
    与所述第一候选框架信息不同的第一框架信息,且所述第一候选框架信息所属第一候选框架支持将模型转换为所述所选择第一框架信息所支持的模型。
  14. 如权利要求13所述的方法,其特征在于,所述方法还包括:
    基于所述所选择第一模型框架将第一AI模型转换为第四AI模型,由所述所选择第一框架信息所确定的第二模型框架支持模型转换,所述第一AI模型由所述网络设备基于第二模型框架生成;
    将所述第四AI模型发送至所述终端设备。
  15. 如权利要求10所述的方法,其特征在于,所述根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备,包括:
    确定模型下发方式,所述第一确定结果指示所述网络设备支持所述中间模型表示框架,且所述第二确定结果表示所述网络设备支持模型转换;
    基于所述模型下发方式,将所述第一AI模型发送至所述终端设备。
  16. 如权利要求10所述的方法,其特征在于,所述根据所述第一确定结果和/或所述第二确定结果,将第一AI模型发送至所述终端设备,包括:
    基于所述中间模型表示框架处理所述第一AI模型,得到AI模型文件,所述第一确定结果指示所述网络设备支持所述中间模型表示框架,且所述第二确定结果表示所述网络设备不支持模型转换;
    将所述AI模型文件发送至所述终端设备。
  17. 如权利要求16所述的方法,其特征在于,还包括:
    向所述终端设备发送第二指示信息,其中,所述第二指示信息指示所述AI模型文件的AI模型存储格式,所述AI模型文件未携带模型格式信息。
  18. 一种异构人工智能AI框架的模型交互方法,其特征在于,由终端设备执行,所述方法包括:
    向网络设备指示支持的第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架。
  19. 如权利要求18所述的方法,其特征在于,所述第一框架信息包括以下任一项:
    所述终端设备支持部署的框架信息;
    所述终端设备支持转换的框架信息。
  20. 如权利要求18-19任一项所述的方法,其特征在于,所述方法还包括:
    向所述网络设备发送第一指示信息,所述第一指示信息用于指示所述终端设备支持第一框架信息。
  21. 如权利要求19所述的方法,其特征在于,确定所述终端设备支持的第一框架信息,包括:
    确定所述网络设备支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二模型框架;
    基于预定义协议和/或所述第二框架信息,确定支持的第一框架信息。
  22. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的第一AI模型,其中,第一AI模型由所述网络设备生成;或者,
    接收所述网络设备发送的第二AI模型,其中,所述第二AI模型由所述网络设备基于所述第一框架信息或中间模型表示框架对第一AI模型转换得到,所述第一AI模型由所述网络设备生成。
  23. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的第二指示信息,所述第二指示信息用于指示所述网络设备所属小区不支持基于所述第一框架信息的AI模型部署。
  24. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的第二指示信息,所述第二指示信息用于指示所述终端设备向所述网络设备上报模型;
    其中,所述终端设备满足以下一种:
    支持的第一框架信息与所述网络设备所支持的第一候选框架信息相同;
    支持的第一框架信息与所述网络设备所支持的第一候选框架信息不同。
  25. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的第二指示信息,其中,所述第二指示信息用于指示所述终端设备从多个所述第一框架信息中选择一个第一框架信息;
    其中,所述终端设备满足:支持的第一框架信息的数量是多个。
  26. 如权利要求24所述的方法,其特征在于,所述方法还包括:
    根据所述第二指示信息指示的内容,确定所述网络设备所指示框架信息;
    基于所述所指示框架信息,生成第三AI模型;
    将所述第三AI模型发送至所述网络设备。
  27. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的第四AI模型。
  28. 如权利要求19所述的方法,其特征在于,所述方法还包括:
    接收所述网络设备发送的AI模型文件,其中,所述AI模型文件由所述网络设备基于所述中间模型表示框架处理第一AI模型得到,所述第一AI模型由所述网络设备基于第二模型框架生成,所述第二模型框架是基于所述第一框架信息确定。
  29. 如权利要求28所述的方法,其特征在于,还包括:
    接收所述网络设备发送的第二指示信息,其中,所述第二指示信息指示所述AI模型文件的AI模型存储格式,所述AI模型文件未携带模型格式信息。
  30. 一种异构人工智能AI框架的模型交互方法,其特征在于,由网络设备执行,所述方法包括:
    向终端设备指示所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
  31. 如权利要求30所述的方法,其特征在于,所述第二框架信息包括以下任一项:
    所述网络设备支持部署的框架信息;
    所述网络设备支持转换的框架信息。
  32. 如权利要求29所述的方法,其特征在于,还包括:
    接收所述终端设备发送的第一AI模型,其中,所述第一AI模型由所述终端设备基于支持的第一框架信息所包括的第一模型框架生成,所述第一框架信息是基于所述第二框架信息确定;或者,
    接收所述终端设备发送的第一指示信息,其中,所述第一指示信息指示第一框架信息以及指示所述第一AI模型未携带模型格式信息;或者,
    向所述终端设备发送第二指示信息,其中,所述第二指示信息指示所述网络设备支持中间模型表示框架;或者,
    接收所述终端设备发送的第二AI模型,其中,所述第二AI模型由所述终端设备生成。
  33. 一种异构人工智能AI框架的模型交互方法,其特征在于,由终端设备执行,所述方法包括:
    确定网络设备所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
  34. 如权利要求33所述的方法,其特征在于,所述第二框架信息包括以下任一项:
    所述网络设备支持部署的框架信息;
    所述网络设备支持转换的框架信息。
  35. 如权利要求34所述的方法,其特征在于,所述方法还包括:
    确定与所述第二框架信息对应的第一模型框架,并基于所述第一模型框架生成第一AI模型,以及向所述网络设备发送所述第一AI模型;或者,
    向所述网络设备发送第一指示信息,其中,所述第一指示信息指示第一框架信息以及指示所述第一AI模型未携带模型格式信息;或者,
    接收所述网络设备发送的第二指示信息,其中,所述第二指示信息指示所述网络设备支持中间模型表示框架;或者,
    确定所述终端设备是否支持所述中间模型表示框架,得到第三确定结果,其中,所述网络设备支持中间模型表示框架,并确定所述终端设备是否支持模型转换,得到第四确定结果,以及根据所述第三确定结果和/或所述第四确定结果,将第二AI模型发送至所述网络设备,其中,所述第二AI模型由所述终端设备生成。
  36. 一种通信装置,其特征在于,所述装置包括:
    处理模块,用于确定终端设备支持第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架,并确定与所述第一框架信息对应的第二AI模型框架;
    收发模块,用于向终端设备指示所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
  37. 一种通信装置,其特征在于,所述装置包括:
    收发模块,用于向网络设备指示支持的第一框架信息,其中,所述第一框架信息包括所述终端设备支持的第一AI模型框架;
    处理模块,用于确定网络设备所支持的第二框架信息,其中,所述第二框架信息包括所述网络设备支持的第二AI模型框架。
  38. 一种通信系统,其特征在于,所述通信系统包括网络设备和终端设备,所述网络设备执行如权利要求1-17中任一项所述的方法,所述终端设备执行如权利要求18-29中任一项所述的方法。
  39. 一种通信系统,其特征在于,所述通信系统包括网络设备和终端设备,所述网络设备执行如权利要求30-32中任一项所述的方法,所述终端设备执行如权利要求33-35中任一项所述的方法。
  40. 一种计算机可读存储介质,用于存储有指令,当所述指令被执行时,使如权利要求1-35中任一项所述的方法被实现。
PCT/CN2022/107536 2022-07-22 2022-07-22 异构人工智能ai框架的模型交互方法、装置及系统 WO2024016363A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN202280002408.5A CN117751563A (zh) 2022-07-22 2022-07-22 异构人工智能ai框架的模型交互方法、装置及系统
PCT/CN2022/107536 WO2024016363A1 (zh) 2022-07-22 2022-07-22 异构人工智能ai框架的模型交互方法、装置及系统

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/107536 WO2024016363A1 (zh) 2022-07-22 2022-07-22 异构人工智能ai框架的模型交互方法、装置及系统

Publications (1)

Publication Number Publication Date
WO2024016363A1 true WO2024016363A1 (zh) 2024-01-25

Family

ID=89616864

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/107536 WO2024016363A1 (zh) 2022-07-22 2022-07-22 异构人工智能ai框架的模型交互方法、装置及系统

Country Status (2)

Country Link
CN (1) CN117751563A (zh)
WO (1) WO2024016363A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220036236A1 (en) * 2018-09-19 2022-02-03 Huawei Technologies Co., Ltd. Ai model development method and apparatus
CN114189889A (zh) * 2021-12-03 2022-03-15 中国信息通信研究院 一种无线通信人工智能处理方法和设备
CN114330696A (zh) * 2021-12-31 2022-04-12 中国联合网络通信集团有限公司 多框架的深度学习模型处理方法、装置及电子设备
CN114697984A (zh) * 2020-12-28 2022-07-01 中国移动通信有限公司研究院 信息传输方法、终端及网络设备
CN114745712A (zh) * 2021-01-07 2022-07-12 中国移动通信有限公司研究院 基于终端能力的处理方法、装置、终端及网络设备

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20220036236A1 (en) * 2018-09-19 2022-02-03 Huawei Technologies Co., Ltd. Ai model development method and apparatus
CN114697984A (zh) * 2020-12-28 2022-07-01 中国移动通信有限公司研究院 信息传输方法、终端及网络设备
CN114745712A (zh) * 2021-01-07 2022-07-12 中国移动通信有限公司研究院 基于终端能力的处理方法、装置、终端及网络设备
CN114189889A (zh) * 2021-12-03 2022-03-15 中国信息通信研究院 一种无线通信人工智能处理方法和设备
CN114330696A (zh) * 2021-12-31 2022-04-12 中国联合网络通信集团有限公司 多框架的深度学习模型处理方法、装置及电子设备

Also Published As

Publication number Publication date
CN117751563A (zh) 2024-03-22

Similar Documents

Publication Publication Date Title
WO2024026801A1 (zh) 一种侧行链路sl波束配置方法、装置、设备及存储介质
WO2024016363A1 (zh) 异构人工智能ai框架的模型交互方法、装置及系统
WO2024060143A1 (zh) 一种上报方法/装置/设备及存储介质
WO2023197187A1 (zh) 一种信道状态信息的处理方法及装置
WO2023010499A1 (zh) 一种无线资源管理测量方法及其装置
US20240244566A1 (en) Positioning method and apparatus
WO2023010428A1 (zh) 准共址配置方法、准共址qcl信息确定方法及其装置
WO2024026889A1 (zh) 一种数据类型确定方法/装置/设备及存储介质
WO2024007172A1 (zh) 信道估计方法及装置
WO2024168917A1 (zh) 一种ai模型的注册方法/装置/设备及存储介质
WO2024130519A1 (zh) 人工智能ai服务的调度方法及其装置
WO2023115279A1 (zh) 数据传输方法及装置
WO2023184452A1 (zh) 终端设备使用的模型的确定方法和装置
WO2024020747A1 (zh) 一种模型的生成方法及装置
WO2023236124A1 (zh) 一种人工智能ai模型训练方法/装置/设备及存储介质
WO2024020752A1 (zh) 一种基于人工智能ai提供服务的方法、装置、设备及存储介质
WO2022266963A1 (zh) 资源分配方法及其装置
WO2024011545A1 (zh) 切换方法及装置
WO2024026885A1 (zh) 一种路径切换的方法及装置
WO2024207366A1 (zh) 基于人工智能的定位功能的信息同步方法和装置
WO2024197481A1 (zh) 信息处理方法及装置
WO2024036456A1 (zh) 一种基于人工智能ai提供服务的方法、装置、设备及存储介质
WO2023230794A1 (zh) 一种定位方法及装置
WO2023201502A1 (zh) 测量配置的处理方法和装置
WO2023236123A1 (zh) 一种系统消息传输方法/装置/设备及存储介质

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 202280002408.5

Country of ref document: CN

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

Ref document number: 22951632

Country of ref document: EP

Kind code of ref document: A1