WO2023125869A1 - Ai网络信息传输方法、装置及通信设备 - Google Patents

Ai网络信息传输方法、装置及通信设备 Download PDF

Info

Publication number
WO2023125869A1
WO2023125869A1 PCT/CN2022/143640 CN2022143640W WO2023125869A1 WO 2023125869 A1 WO2023125869 A1 WO 2023125869A1 CN 2022143640 W CN2022143640 W CN 2022143640W WO 2023125869 A1 WO2023125869 A1 WO 2023125869A1
Authority
WO
WIPO (PCT)
Prior art keywords
protocol class
network
onnx
target
file structure
Prior art date
Application number
PCT/CN2022/143640
Other languages
English (en)
French (fr)
Inventor
任千尧
孙鹏
Original Assignee
维沃移动通信有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 维沃移动通信有限公司 filed Critical 维沃移动通信有限公司
Publication of WO2023125869A1 publication Critical patent/WO2023125869A1/zh

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04LTRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L67/00Network arrangements or protocols for supporting network services or applications
    • H04L67/01Protocols
    • H04L67/06Protocols specially adapted for file transfer, e.g. file transfer protocol [FTP]
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04WWIRELESS COMMUNICATION NETWORKS
    • H04W28/00Network traffic management; Network resource management
    • H04W28/02Traffic management, e.g. flow control or congestion control
    • H04W28/06Optimizing the usage of the radio link, e.g. header compression, information sizing, discarding information
    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04WWIRELESS COMMUNICATION NETWORKS
    • H04W28/00Network traffic management; Network resource management
    • H04W28/02Traffic management, e.g. flow control or congestion control
    • H04W28/06Optimizing the usage of the radio link, e.g. header compression, information sizing, discarding information
    • H04W28/065Optimizing the usage of the radio link, e.g. header compression, information sizing, discarding information using assembly or disassembly of packets

Definitions

  • the present application belongs to the technical field of communication, and in particular relates to an AI network information transmission method, device and communication equipment.
  • AI Artificial Intelligence
  • communication data can be transmitted between network-side devices and terminals through AI networks.
  • the neural network frameworks used by different communication devices are different, and the file structures of AI network information saved under different neural network frameworks are different, so that the communication devices may not be able to read the data transmitted by other communication devices with different neural network frameworks.
  • AI network information resulting in limited transmission of AI network information between communication devices.
  • Embodiments of the present application provide an AI network information transmission method, device, and communication device, which can solve the problem of limited transmission of AI network information between communication devices in the related art.
  • an AI network information transmission method including:
  • the first end converts the AI network information of the target AI network into an open neural network exchange ONNX file structure
  • the first end sends the ONNX file structure to the second end.
  • an AI network information transmission method including:
  • the second end receives the ONNX file structure sent by the first end, and the ONNX file structure is obtained by converting the AI network information of the target AI network by the first end.
  • an AI network information transmission device including:
  • the conversion module is used to convert the AI network information of the target AI network into an open neural network exchange ONNX file structure
  • a sending module configured to send the ONNX file structure to the second end.
  • an AI network information transmission device including:
  • the receiving module is configured to receive the ONNX file structure sent by the first end, and the ONNX file structure is obtained by converting the AI network information of the target AI network by the first end.
  • a communication device including a processor and a memory, the memory stores programs or instructions that can run on the processor, and when the programs or instructions are executed by the processor, the first The steps of the AI network information transmission method described in the aspect, or the steps of implementing the AI network information transmission method described in the second aspect.
  • a sixth aspect provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the AI network information transmission method as described in the first aspect are implemented , or implement the steps of the AI network information transmission method as described in the second aspect.
  • a chip in the seventh aspect, there is provided a chip, the chip includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the AI described in the first aspect The steps of the network information transmission method, or the steps of realizing the AI network information transmission method described in the second aspect.
  • a computer program/program product is provided, the computer program/program product is stored in a storage medium, and the computer program/program product is executed by at least one processor to implement the method described in the first aspect The steps of the AI network information transmission method, or the steps of realizing the AI network information transmission method as described in the second aspect.
  • the first end converts the AI network information of the target AI network based on its own neural network framework into an ONNX file structure, and then sends the ONNX file structure to the second end, and then the second end can convert the ONNX file structure to the ONNX file structure.
  • the file structure is converted into an AI network under its own neural network framework. In this way, even two communication devices with different neural network frameworks can transmit AI network information based on the ONNX file structure, avoiding the blockage of the transmission of AI network information between communication devices.
  • FIG. 1 is a block diagram of a wireless communication system to which an embodiment of the present application is applicable;
  • FIG. 2 is a flow chart of an AI network information transmission method provided by an embodiment of the present application.
  • Fig. 3 is a flow chart of another AI network information transmission method provided by the embodiment of the present application.
  • FIG. 4 is a structural diagram of an AI network information transmission device provided by an embodiment of the present application.
  • Fig. 5 is a structural diagram of another AI network information transmission device provided by the embodiment of the present application.
  • FIG. 6 is a structural diagram of a communication device provided by an embodiment of the present application.
  • FIG. 7 is a structural diagram of a terminal provided in an embodiment of the present application.
  • FIG. 8 is a structural diagram of a network-side device provided by an embodiment of the present application.
  • FIG. 9 is a structural diagram of another network-side device provided by an embodiment of the present application.
  • first, second and the like in the specification and claims of the present application are used to distinguish similar objects, and are not used to describe a specific sequence or sequence. It is to be understood that the terms so used are interchangeable under appropriate circumstances such that the embodiments of the application are capable of operation in sequences other than those illustrated or described herein and that "first" and “second” distinguish objects. It is usually one category, and the number of objects is not limited. For example, there may be one or more first objects.
  • “and/or” in the description and claims means at least one of the connected objects, and the character “/” generally means that the related objects are an "or” relationship.
  • LTE Long Term Evolution
  • LTE-Advanced LTE-Advanced
  • LTE-A Long Term Evolution-Advanced
  • CDMA Code Division Multiple Access
  • TDMA Time Division Multiple Access
  • FDMA Frequency Division Multiple Access
  • OFDMA Orthogonal Frequency Division Multiple Access
  • SC-FDMA Single-carrier Frequency Division Multiple Access
  • system and “network” in the embodiments of the present application are often used interchangeably, and the described technologies can be used for the above-mentioned systems and radio technologies as well as other systems and radio technologies.
  • NR New Radio
  • the following description describes the New Radio (NR) system for illustrative purposes, and uses NR terminology in most of the following descriptions, but these techniques can also be applied to applications other than NR system applications, such as the 6th generation (6 th Generation, 6G) communication system.
  • 6G 6th Generation
  • Fig. 1 shows a block diagram of a wireless communication system to which the embodiment of the present application is applicable.
  • the wireless communication system includes a terminal 11 and a network side device 12 .
  • the terminal 11 can be a mobile phone, a tablet computer (Tablet Personal Computer), a laptop computer (Laptop Computer) or a notebook computer, a personal digital assistant (Personal Digital Assistant, PDA), a palmtop computer, a netbook, a super mobile personal computer (ultra-mobile personal computer, UMPC), mobile Internet device (Mobile Internet Device, MID), augmented reality (augmented reality, AR) / virtual reality (virtual reality, VR) equipment, robot, wearable device (Wearable Device) , Vehicle User Equipment (VUE), Pedestrian User Equipment (PUE), smart home (home equipment with wireless communication functions, such as refrigerators, TVs, washing machines or furniture, etc.), game consoles, personal computers (personal computer, PC), teller machine or self-service machine and other terminal side devices, wearable devices include: smart watches, smart bracelet
  • the network side device 12 may include an access network device or a core network device, where the access network device may also be called a radio access network device, a radio access network (Radio Access Network, RAN), a radio access network function, or a wireless network. access network unit.
  • RAN Radio Access Network
  • RAN Radio Access Network
  • the access network device may include a base station, a wireless local area network (Wireless Local Area Networks, WLAN) access point or a WiFi node, etc., and the base station may be called a node B, an evolved node B (eNB), an access point, a base transceiver station ( Base Transceiver Station, BTS), radio base station, radio transceiver, Basic Service Set (BSS), Extended Service Set (Extended Service Set, ESS), Home Node B, Home Evolved Node B, sending and receiving point ( Transmitting Receiving Point, TRP) or some other appropriate term in the field, as long as the same technical effect is achieved, the base station is not limited to specific technical vocabulary.
  • Core network equipment may include but not limited to at least one of the following: core network nodes, core network functions, mobility management entities (Mobility Management Entity, MME), access mobility management functions (Access and Mobility Management Function, AMF), session management functions (Session Management Function, SMF), User Plane Function (UPF), Policy Control Function (Policy Control Function, PCF), Policy and Charging Rules Function (PCRF), edge application service Discovery function (Edge Application Server Discovery Function, EASDF), unified data management (Unified Data Management, UDM), unified data storage (Unified Data Repository, UDR), home subscriber server (Home Subscriber Server, HSS), centralized network configuration ( Centralized network configuration, CNC), network storage function (Network Repository Function, NRF), network exposure function (Network Exposure Function, NEF), local NEF (Local NEF, or L-NEF), binding support function (Bind
  • Each framework has a different focus.
  • Caffe2 and Keras are high-level deep learning frameworks that can quickly verify models
  • TensorFlow and PyTorch are low-level deep learning frameworks that can modify the underlying details of neural networks.
  • PyTorch focuses on supporting dynamic graph models
  • TensorFlow focuses on supporting multiple hardware and running fast
  • Caffe2 focuses on lightweight.
  • Each implementation framework will use its own method to describe the neural network and complete operations such as network construction, training, and inference.
  • ONNX is an AI interactive network.
  • ONNX itself is just a data structure used to describe an AI network, not including the implementation plan.
  • ONNX saves the entire AI network in the calculation graph protocol class (GraphProto), including the network structure and network parameters, and completes the basic AI network through the node protocol class (NodeProto), parameter information protocol class (ValueInfoProto), and tensor protocol class (TensorProto). describe. Among them, NodeProto is used to describe the network structure.
  • ONNX describes each operator of the AI network as a node. The name of the input and output of each node is globally unique, and the entire AI is described by the matching relationship between the input and output names. The network structure of the network.
  • All network parameters are regarded as input or output, and are also retrieved by name, so ONNX can express the structure of the entire network through a series of NodeProto entities.
  • ValueInfoProto is used to describe all input and output information, including dimensions and element types, indicating the size of the corresponding element, the name of each input and output and the correspondence recorded in NodeProto.
  • TensorProto is used to store the value of specific network parameters, and obtain the corresponding parameters according to the name of the input and output of each node to the storage location.
  • ONNX records the functions of nodes through the attribute protocol class (AttributeProto), such as convolution layer, multiplication layer, etc., and gives corresponding node functions, and the values of hyperparameters required by these functions are stored in TensorProto, and the dimension of hyperparameters All stored in ValueInfoProto.
  • attribute Protocol class such as convolution layer, multiplication layer, etc.
  • the protocol class can also be called a structure, such as a node structure (NodeProto), a parameter information structure (ValueInfoProto), a tensor structure (TensorProto), an attribute structure (AttributeProto) wait.
  • Proto can be considered as a kind of data, files, etc. formed based on a specific protocol.
  • Figure 2 is a flow chart of an AI network information transmission method provided in the embodiment of the present application, as shown in Figure 2, the method includes the following steps:
  • Step 201 the first end converts the AI network information of the target AI network into an ONNX file structure
  • Step 202 the first end sends the ONNX file structure to the second end.
  • the first end converts the AI network information of the target AI network into an ONNX file structure.
  • the AI network information includes the complete network structure and all network parameters of the target AI network
  • the first end converts the network
  • the structure and network parameters are expressed as ONNX file structure based on the structure of ONNX.
  • the network structure is described by the node protocol class (NodeProto) and the parameter information protocol class (ValueInfoProto), and the network parameters are described by the tensor protocol class (TensorProto), etc. .
  • the first end sends the ONNX file structure to the second end
  • the second end converts the ONNX file structure into an AI network under its own neural network framework, so as to realize the training and application of the AI network by the second end .
  • the first end converts the AI network information of the target AI network applicable to its own neural network framework into an ONNX file structure, and then sends the ONNX file structure to the second end, and then the second end can convert the ONNX file structure to the ONNX file structure.
  • the file structure is converted into an AI network under its own neural network framework. In this way, two communication devices with different neural network frameworks can transmit AI network information based on the ONNX file structure, avoiding the blockage of the transmission of AI network information between communication devices.
  • the AI network information includes at least one of the network structure and network parameters of the target AI network.
  • the AI network information includes the complete network structure and all network parameters of the target AI network, or the AI network information includes the complete network structure of the target AI network, or only includes the updated AI network structure in the target AI network, or Only include some network parameters of the target AI network, or only include some values of the network parameters of the target AI network, and so on.
  • the network structure and network parameters of the AI network can be sent separately, and then there is no need to transmit all the AI networks including the entire network structure and network parameters together during the communication process, which can effectively reduce the transmission overhead during the communication process .
  • the ONNX file structure includes a target protocol class
  • the target protocol class includes at least one of the following: node protocol class (NodeProto), parameter information protocol class (ValueInfoProto), tensor protocol class (TensorProto), Attribute protocol class (AttributeProto).
  • NodeProto is used to describe the network structure
  • ValueInfoProto is used to describe the information of all input and output and network parameters, including dimensions and element types, indicating the size of the corresponding element, the name or index of each input and output and network parameters and the records in NodeProto Correspondence
  • TensorProto is used to store the value of specific network parameters, and obtain the corresponding parameters according to the name of the input and output of each node
  • AttributeProto is used to record the function of the node, such as convolution layer, multiplication layer, etc., and assign it to the corresponding node functions, and the values of the hyperparameters required by these functions are stored in TensorProto, and the dimensions of the hyperparameters are stored in ValueInfoProto.
  • the target protocol class includes at least one of the above-mentioned node protocol class, parameter information protocol class, tensor protocol class and attribute protocol class
  • the ONNX file structure also includes the above-mentioned node protocol class, parameter information protocol class, At least one of tensor protocol classes and attribute protocol classes.
  • the node protocol class can form a complete onnx.proto file to generate the ONNX file structure; if the target protocol class includes a node protocol class, a parameter information protocol class and a tensor protocol class, The onnx.proto file is formed based on the node protocol class, parameter information protocol class and tensor protocol class to generate the ONNX file structure.
  • the protocol classes included in the ONNX file structure may also be other possible situations, which will not be listed here.
  • the ONNX file structure includes at least one computation graph protocol class (GraphProto), and the computation graph protocol class includes the target protocol class. That is to say, when the target protocol class includes at least one of the above-mentioned node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, the content included in the target protocol class can be written into the calculation In the graph protocol class.
  • GraphProto computation graph protocol class
  • the target protocol class includes node protocol class, parameter information protocol class and tensor protocol class, then write the node protocol class, parameter information protocol class and tensor protocol class into the calculation graph protocol class, based on the The calculation graph protocol class generates the ONNX file structure, and the first end sends the ONNX file structure to the second end. In this way, the node protocol class, parameter information protocol class and tensor protocol class are combined for transmission.
  • the ONNX file structure may include multiple image protocol classes, and the target protocol classes included in each calculation graph protocol class may be different.
  • the ONNX file structure includes two calculation graph protocol classes, one of which includes the node protocol class and parameter information protocol class, and the other includes the tensor protocol class.
  • each calculation graph protocol class can also include multiple node protocol classes, parameter information protocol classes, tensor protocol classes, and attribute protocol classes.
  • the ONNX file structure can also include a model protocol class (ModelProto).
  • the model protocol class includes a calculation graph protocol class.
  • the calculation graph protocol class includes several node protocol classes, parameter information protocol classes, and tensor protocol classes. Node A protocol class can contain several attribute protocol classes.
  • the ONNX file structure may include multiple calculation graph protocol classes, regardless of whether there is a model protocol class; that is, the ONNX file structure in the embodiment of the present application may not include the model protocol class, and there is no need to transfer the model Protocol class, which can save the transmission overhead of communication equipment, and is more conducive to the application of ONNX file structure in air interface transmission.
  • the target protocol class may be any one of a node protocol class (NodeProto), a parameter information protocol class (ValueInfoProto), a tensor protocol class (TensorProto) and an attribute protocol class (AttributeProto).
  • NodeProto node protocol class
  • ValueInfoProto parameter information protocol class
  • TorsorProto tensor protocol class
  • attribute Protocol class AttributeProto
  • the first end merges and sends the at least two ONNX file structures to the second end;
  • the first end sends the at least two ONNX file structures to the second end respectively.
  • each proto can be independently formed into an ONNX file structure, and an ONNX file structure corresponds to a type of proto.
  • the AI network information of the target AI network includes network structure and weight parameters, and the AI network information generates a node protocol class, a parameter information protocol class, and a tensor protocol class based on the structure of ONNX, then the node protocol class corresponds to generate an ONNX File structure, parameter information protocol class corresponds to generate an ONNX file structure, tensor protocol class corresponds to generate an ONNX file structure, and three ONNX file structures will be obtained.
  • the generation of the ONNX file structure is more flexible, and it is also beneficial for the second end to distinguish the ONNX file structure.
  • the ONNX file structure can be transmitted in different time slots and time-frequency positions, and the time-frequency resources can be used more effectively. , to facilitate scheduling.
  • the first end may send the three ONNX file structures to the second end respectively, for example, send the three ONNX file structures successively; or, the first end may also merge the three ONNX file structures, And send the merged three ONNX file structures to the second end at one time.
  • the second end can clarify that an ONNX file structure corresponds to a proto type, which makes it easier for the second end to convert the ONNX file structure into AI network information under its own neural network framework.
  • the method when the ONNX file structure includes at least one of the node protocol class and the parameter information protocol class, after the first end sends the ONNX file structure to the second end, the method Also includes:
  • the first end sends the value of the network parameter of the target AI network to the second end.
  • the first end can also send the value of the network parameter of the target AI network to the second end. Understandably, since the first end has sent the ONNX file structure containing the node protocol class and the parameter information protocol class to the second end, and the second end is also known about the number and overhead of network parameters, the first end can The specific values of the network parameters are directly sent in the order of the parameter information protocol class without converting into the ONNX file structure, and the second end receives the specific values of each parameter in order.
  • the node protocol class and the parameter information protocol class may be included in an ONNX file structure, or each may correspond to an ONNX file structure.
  • the first end sends the value of the network parameter of the target AI network to the second end, including:
  • the first end sends the value of the network parameter of the target AI network to the second end through a data channel.
  • the ONNX file structure includes a first digital index and a second digital index of an integer type, the first digital index is used to identify the target protocol class, and one target protocol class corresponds to a first A numerical index, the second numerical index is used to identify at least one of network parameters, input and output of the target AI network.
  • the protocol classes included in the ONNX file structure may be identified by the first numerical index.
  • the target protocol class includes at least one of the node protocol class, parameter information protocol class, tensor protocol class and attribute protocol class, and a protocol class may correspond to a first digital index, such as ONNX file structure
  • a protocol class may correspond to a first digital index, such as ONNX file structure
  • a total of 5 protocol classes are included, and these 5 protocol classes can be represented by 5 first numerical indexes (for example, 1, 2, 3, 4, 5).
  • one protocol class type may correspond to one numerical index, for example, there are 5 protocol classes in total, and 4 protocol class types, then the 5 protocol classes may be represented by 4 first numerical indexes.
  • each protocol class type can also correspond to a numerical index sequence, for example, it includes a total of 3 node protocol classes, 4 parameter information protocol classes and 1 tensor protocol class, and 3 can be represented by digital indexes 0, 1, and 2 1 node protocol class, 0, 1, 2, 3 represent 4 parameter information protocol classes, 0 represents 1 tensor protocol class, and the sequences of the three protocol classes are distinguished by specific identifiers.
  • the content described for the target protocol class may be identified by a second digital index.
  • the node protocol class is used to describe the network structure of the target AI network. The name of the input and output of each node of the network structure is unique.
  • a node protocol class can be used to describe the input and output of a node in the network structure.
  • related network parameters, and the input, output and network parameters corresponding to the node protocol class can be identified by the second digital index, for example, the input, output and network parameters respectively correspond to three different second digital indexes.
  • the content described by other protocol classes may also be identified by the second digital index.
  • the first digital index is used to identify the protocol class included in the ONNX file structure
  • the second digital index is used to identify the content described by the protocol class.
  • the first digital index and the second digital index are both integer (int) type digital indexes, compared to using character strings to identify, using integer type digital indexes to identify can more effectively save transmission overhead .
  • the node protocol class includes the second numerical index used to characterize the input and output of the target AI network.
  • protocol classes such as parameter information protocol class, tensor protocol class, and attribute protocol class, can also use the second digital index to identify the content described by the protocol class, such as the network parameters of the target AI network, input , output, etc.
  • the node protocol class when the number of inputs and outputs of the target AI network is multiple and continuous, includes a target number index, and includes the number of inputs and the number of outputs of the target AI network At least one of , wherein the target numerical index is a numerical index in the second numerical index used to identify the first input and the first output of the plurality of inputs and outputs of the target AI network.
  • the network structure of the target AI network is described by the node protocol class in the ONNX file structure.
  • the network structure includes multiple nodes, and one node corresponds to at least one input and at least one input.
  • the number of nodes included in the network structure is multiple and In the continuous case, the input and output of a node correspond to a second digital index, and then should correspond to multiple consecutive second digital indexes.
  • the node protocol class in the ONNX file structure can only include the Identifying the second numerical index (ie, the target numerical index) of the input and output of the first node of the plurality of nodes and the total number of nodes.
  • the network structure of the target AI network includes 3 nodes, and the input and output of the same node correspond to a second digital index, which in turn corresponds to 3 second digital indexes, such as 11, 12, and 13; in this case,
  • the node protocol class includes the corresponding second digital index and the total number of nodes (that is, the number of inputs and outputs) for identifying the input and output of the first node, so the node protocol class of the ONNX file structure also includes 11 and 3, so that there is no need to include all second numeric indices.
  • the second end can obtain the second digital index corresponding to each input and output based on the target digital index and the total input and output quantity.
  • one parameter information protocol class is used to describe at least one of the input, output and network parameters corresponding to the network nodes in the target AI network, and the parameter information protocol class includes the first element type and dimension information.
  • the parameter information protocol class is used to describe all input and output information of the target AI network.
  • the parameter information protocol class includes dimension information and the first element type, indicating the size of the corresponding element, the name and node protocol of each input and output The input and output correspondence recorded in the class.
  • the first element type is characterized by an integer length of quantization bits. That is to say, the first element type is replaced by the number of quantization bits, so that the first element type can be omitted, and the default type is used, such as an integer (int) length, that is, the number of quantization bits is an integer length, and then the first Element types can be characterized by specifying an integer length.
  • the first element type corresponds to network nodes or network parameters of the target AI network. It can be through a predetermined correspondence, for example, a certain type of network node or network parameter uses a corresponding integer length, for example, the weight of a convolution node uses an integer with a length of 3 bits, and the offset uses a length of 1 Integer of bits.
  • int1 represents the index of 1 bit, that is, only the values of 0 and 1
  • int2 corresponds to 00 01
  • the index of 10 11 or 0123, the quantization table corresponding to the specific index is stipulated by the agreement according to the function of the AI network, or the base station configuration.
  • the first element type and the subsequent second element type can be characterized by such an integer length.
  • a network node of the target AI network corresponds to a parameter information protocol class
  • a parameter information protocol class represents at least one of the input, output, and network parameters of the corresponding network node through dimension information
  • the dimension information includes a dimension number value and a dimension size value.
  • the dimension information includes (3,2,6,14), or (3,(2,6,14)).
  • the dimension size value is the number of numerical values included in the corresponding dimension, and the numerical values are normalized numerical values.
  • the dimension size value is 168, that is, the dimension includes 168 values, and these values are normalized values.
  • the dimension information in the parameter information protocol class includes corresponding dimension quantity values and dimension size values, the dimension quantity values in the parameter information protocol class are arranged based on a preset order, and the dimension information in the parameter information protocol class The dimension size values are arranged based on the preset order, and the preset order is a preset arrangement order of the network parameters of the target AI network.
  • a network parameter of the target AI network corresponds to a parameter information protocol class
  • the parameter information protocol class includes dimension information
  • the dimension information includes dimension quantity value and dimension size value, through dimension quantity value and dimension size value
  • the ONNX file structure also includes multiple corresponding parameter information protocol classes, and each parameter information protocol class includes dimension value and dimension value, then the ONNX file structure It also correspondingly includes multiple dimension quantity values and multiple dimension size values; in this case, these dimension quantity values and dimension size values can be arranged according to the preset arrangement order of the network parameters in the target AI network, for example, it can be according to The default sorting order of the network parameters is to arrange and place all the dimension quantity values first, and then arrange and place all the dimension size values according to the preset sorting order of the network parameters. In this way, the dimension value uses bits of the same length. After parsing the dimension value, the second end also knows the bit length corresponding to the subsequent dimension value, and can omit the identifier of the network parameter to save transmission overhead.
  • the target parameter information protocol class includes the numerical position of the maximum value, and the maximum value is the maximum value among the values of the target network parameters
  • the target tensor protocol class corresponding to the target network parameter includes the quantized value of the ratio of each remaining value in the target network parameter to the maximum value
  • the remaining value is the target parameter information protocol class corresponding to A value other than the stated maximum value in the dimension.
  • the target network parameter is any network parameter of the target AI network
  • the target parameter information protocol class is a parameter information protocol class used to characterize the target network parameter in the ONNX file structure
  • the target tensor protocol class is A tensor protocol class corresponding to the target network parameter represented by the target parameter information protocol class. For example, if the target network parameter includes 168 values, the target parameter information protocol class corresponding to the target network parameter includes the value position of the largest value among the 168 values, and the remaining 167 values are passed through the corresponding target network parameter.
  • the target tensor protocol class is represented by the target tensor protocol class, and the target tensor protocol class includes the quantized value of the ratio of these 167 values to the maximum value.
  • the parameter information protocol class further includes an indication parameter for indicating a position of a non-zero value in the network parameter.
  • the quantized value of 0 in the tensor protocol class corresponding to the network parameter, or the value smaller than the quantized minimum threshold may not be included.
  • Typeproto can be modified to int, which is used to record the element type, that is, the number of digits of quantization. When there is no record, it can be the same as the previous parameter information protocol class; or delete type, and save Typeproto in the calculation graph protocol In the class, it means that all the parameter information protocol classes use the same quantization bit.
  • the tensor protocol class includes the second element type and a value, or, the tensor protocol class may only include a value.
  • the tensor protocol class is used to represent the network parameters of the target AI network, and the name, value and other information of the network parameters are represented by the second element type and value.
  • the second element type is characterized by an integer length of quantization bits, that is, the second element type only needs to indicate the length.
  • the second element type has a corresponding relationship with the first element type in the parameter information protocol class, for example, the second element type and the first element type are used to describe the same network parameter, and the corresponding relationship can be It is sorted according to the index order of the network parameters of the target AI network.
  • the arrangement order of the second element type in the tensor protocol class is the same as the arrangement order of the first element type in the parameter information protocol class. It should be noted that the name and/or index of the network parameter may be omitted in the tensor protocol class.
  • the ONNX file structure includes a protocol class list, and one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, and tensor protocol class and attribute protocol classes. That is to say, proto can be stored by category. A type of proto corresponds to a list of protocol classes.
  • the ONNX file structure includes node protocol classes, parameter information protocol classes, and tensor protocol classes, so the ONNX file structure also includes Three lists, namely node protocol column list, parameter information protocol class list and tensor protocol class list.
  • the protocol class list may include multiple corresponding protocol classes.
  • the node protocol class list stores multiple node protocol classes, and the multiple protocol classes are arranged in a certain order.
  • the protocol class stored in the protocol class list includes a corresponding list indication, and the position of the protocol class in the list is indicated by the list indication.
  • each node protocol class can be found through the offset.
  • the parameter information protocol class and the tensor protocol class which can be arranged into a corresponding list according to the order recorded in the node protocol class.
  • the method may also include:
  • the first end sends indication information to the second end; or,
  • the first end sends the ONNX file structure to the second end, including:
  • the first end sends the ONNX file structure carrying indication information to the second end;
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • each protocol class list can form an ONNX file structure independently, and the second end can determine the start position and/or end position of the protocol class list in the ONNX file structure based on the indication information.
  • the second end may add some content to the ONNX file structure.
  • the first end is the core network device
  • the second end is the base station.
  • the ONNX file structure sent by the core network device to the base station only includes node protocol information.
  • the base station can supplement parameter information protocol information and Zhang Quantity protocol information.
  • the ONNX file structure sent by the core network device to the base station includes all node protocol information, all parameter information protocol information, and part of the tensor protocol information corresponding to the target AI network. The start position and end position of , supplement other tensor protocol class information at relevant positions.
  • protocol classes may be proto groupings, for example, grouping or segmenting according to proto functions or requirements.
  • the ONNX file structure includes a predefined index, and the predefined index is used to represent a preset network node.
  • the first end can pre-define some network nodes with commonly used functions or historical functions, use these network nodes as preset network nodes, and characterize these preset network nodes through predefined indexes, and the second end can Obtain information such as the function of the corresponding preset network node, the dimension of input and output, etc.
  • the predefined index there is no need to represent all the network nodes used by the node protocol class, so that the number of node protocol classes in the ONNX file structure can be reduced, thereby saving the transmission overhead of the ONNX file structure.
  • the network nodes include a discrete Fourier transform (Discrete Fourier Transform, DFT) node, an inverse discrete Fourier transform (Inverse Discrete Fourier Transform, IDFT) node, a filter node, and the like.
  • DFT discrete Fourier Transform
  • IDFT inverse discrete Fourier transform
  • filter node a filter node
  • the preset network node may be an unupdated network node in the target AI network.
  • non-updated network nodes can be used as preset network nodes, and then these non-updated network nodes can be represented by predefined indexes in the ONNX file structure.
  • the unupdated network nodes may be merged into one node, which is indicated by a predefined index, for example, may be indicated by an index corresponding to a historical node.
  • the first end converts the AI network information of the target AI network into an ONNX file structure, including:
  • the first end converts the first AI network information of the target AI network into a first ONNX file structure, and converts the second AI network information of the target AI network into a second ONNX file structure, wherein the first The target protocol class that the ONNX file structure includes is different from the target protocol class that the second ONNX file structure includes;
  • the first end sends the ONNX file structure to the second end, including:
  • the first end sends the first ONNX file structure and the second ONNX file structure to the second end.
  • the first end when the first end generates the ONNX file structure, it can generate different ONNX file structures based on the type of the protocol class.
  • Each ONNX file structure includes different protocol classes, and then sends these ONNX file structures respectively.
  • the first ONNX file structure includes the node protocol class
  • the second ONNX file structure includes the parameter information protocol class
  • the first end sends the two ONNX file structures to the second end respectively.
  • the first end can generate multiple ONNX file structures. For example, if the AI network information of the target AI network corresponds to the node protocol class, parameter information protocol class, and tensor protocol class, based on this Three different protocol class types, the first end can generate three ONNX file structures respectively, these three ONNX file structures correspond to node protocol class, parameter information protocol class and tensor protocol class respectively, and send these three ONNX file structures respectively file structure.
  • the first end can also include all protocol classes in an ONNX file structure.
  • the protocol classes corresponding to the first ONNX file structure and the second ONNX file structure may have multiple situations.
  • the first ONNX file structure includes a node protocol class, and the second ONNX file structure includes a parameter information protocol class; or, the first ONNX file structure includes a node protocol class and a parameter information protocol class, and the second ONNX file structure includes a parameter information protocol class;
  • the ONNX file structure includes a tensor protocol class; or, the first ONNX file structure includes a node protocol class, and the second ONNX file structure includes a parameter information protocol class and a tensor protocol class.
  • the terminal and the network side equipment use the joint AI network to perform channel state information (Channel State Information, CSI) feedback, that is, the terminal converts the channel information into several bits of CSI feedback information through the AI network, and reports it to the base station, and the base station receives The bit information fed back by the terminal recovers the channel information through the AI network on the base station side.
  • CSI Channel State Information
  • the AI network of the base station and the terminal needs to be jointly trained, and the channel conditions of different cells are different, new network parameters may also be required. Therefore, when the terminal accesses the network, the base station needs to send the network parameters used by the terminal to the terminal.
  • the CSI feedback network can be divided into two parts, the terminal coding part and the base station decoding part, and the terminal only needs to obtain the network structure of the coding part.
  • the core network device sends the network structure of the encoded part to the terminal through Non-access Stratum (NAS) information.
  • the specific information can be an ONNX file structure including only NodeProto information, or a file structure including a NodeProto list.
  • the base station forwards it to the terminal.
  • the base station After the base station receives the NAS information of the core network equipment, if the NAS information is transparent to the base station, the base station forwards the NAS information directly, and then saves the weight parameters into an ONNX file structure with only ValueInfoProto and TensorProto, and passes the radio resource control (Radio Resource Control) , RRC) signaling is sent to the terminal.
  • Radio Resource Control Radio Resource Control
  • the base station can interpret this NAS signaling, the base station can send the ONNX file structure of NodeProto and its own ONNX file structure including ValueInfoProto and TensorProto to the terminal through RRC signaling, or combine the two into one ONNX file structure and send it to the terminal. terminal.
  • the base station can add the ValueInfoProto and TensorProto information to the ONNX file structure of the NAS information, and still forward the information to the user according to the NAS information.
  • the core network device may also send an ONNX file structure including NodeProto, ValueInfoProto and TensorProto, where ValueInfoProto and TensorProto may be part or all.
  • the base station After the base station receives the NAS information, if it is transparently and directly forwarded, it will send its own ValueInfoProto and TensorProto ONNX file structure through RRC signaling. If it is non-transparent, the base station can send its own ONNX file structure and the ONNX file structure of the core network device to the terminal through RRC, or send it to the terminal after being combined.
  • Figure 3 is a flowchart of another AI network information transmission method provided by the embodiment of the present application, as shown in Figure 3, the method includes the following steps:
  • Step 301 the second end receives the ONNX file structure sent by the first end, and the ONNX file structure is obtained by converting the AI network information of the target AI network by the first end.
  • the first end converts the AI network information of the target AI network into an ONNX file structure.
  • the AI network information includes the complete network structure and all network parameters of the target AI network
  • the first end converts the network
  • the structure and network parameters are expressed as ONNX file structure based on the structure of ONNX.
  • the network structure is described by the node protocol class (NodeProto) and the parameter information protocol class (ValueInfoProto), and the network parameters are described by the tensor protocol class (TensorProto), etc. .
  • NodeProto node protocol class
  • ValueInfoProto parameter information protocol class
  • TensorProto tensor protocol class
  • the second end receives the ONNX file structure sent by the first end, and the second end converts the ONNX file structure into an AI network under its own neural network framework, so as to realize the training and training of the AI network by the second end. application.
  • two communication devices with different neural network frameworks can transmit AI network information based on the ONNX file structure, avoiding the blockage of the transmission of AI network information between communication devices.
  • the AI network information includes at least one of the network structure and network parameters of the target AI network.
  • the AI network information includes the complete network structure and all network parameters of the target AI network, or the AI network information includes the complete network structure of the target AI network, or only includes the updated AI network structure in the target AI network, or Only include some network parameters of the target AI network, or only include some values of the network parameters of the target AI network, and so on.
  • the network structure and network parameters of the AI network can be sent separately, and then there is no need to transmit all the AI networks including the entire network structure and network parameters together during the communication process, which can effectively reduce the transmission overhead during the communication process .
  • the ONNX file structure includes a target protocol class
  • the target protocol class includes at least one of the following: node protocol class (NodeProto), parameter information protocol class (ValueInfoProto), tensor protocol class (TensorProto), Attribute protocol class (AttributeProto).
  • the ONNX file structure includes at least one computation graph protocol class (GraphProto), and the computation graph protocol class includes the target protocol class. That is to say, when the target protocol class includes at least one of the above-mentioned node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, the content included in the target protocol class can be written into the calculation In the graph protocol class.
  • GraphProto computation graph protocol class
  • the target protocol class includes any one of the node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, when the number of the ONNX file structure is at least two , a described ONNX file structure corresponds to a target protocol class;
  • the second end receives the ONNX file structure sent by the first end, including any of the following:
  • the second end receives at least two ONNX file structures combined and sent by the first end;
  • the second end receives at least two ONNX file structures respectively sent by the first end.
  • each proto can be independently formed into an ONNX file structure, and an ONNX file structure corresponds to a type of proto.
  • the AI network information of the target AI network includes network structure and weight parameters, and the AI network information generates a node protocol class, a parameter information protocol class, and a tensor protocol class based on the structure of ONNX, then the node protocol class corresponds to generate an ONNX File structure, parameter information protocol class corresponds to generate an ONNX file structure, tensor protocol class corresponds to generate an ONNX file structure, and three ONNX file structures will be obtained.
  • the first end may send the three ONNX file structures to the second end respectively, for example, send the three ONNX file structures successively; or, the first end may also merge the three ONNX file structures, And send the merged three ONNX file structures to the second end at one time.
  • the second end can perform corresponding receiving actions.
  • an ONNX file structure corresponds to a proto type, which is more convenient for the second end to convert the ONNX file structure into its own neural network framework. AI network information.
  • the method further includes:
  • the second end receives the value of the network parameter of the target AI network sent by the first end.
  • the first end can also send the value of the network parameter of the target AI network to the second end. Understandably, since the first end has sent the ONNX file structure including the node protocol class and the parameter information protocol class to the second end, the second end also knows the number and overhead of network parameters.
  • the node protocol class and the parameter information protocol class may be included in an ONNX file structure, or each may correspond to an ONNX file structure.
  • the second end receives the value of the network parameter of the target AI network sent by the first end through a data channel.
  • the ONNX file structure includes a first digital index and a second digital index of an integer type, the first digital index is used to identify the target protocol class, and one target protocol class corresponds to a first A numerical index, the second numerical index is used to identify at least one of network parameters, input and output of the target AI network.
  • the node protocol class includes the second numerical index used to characterize the input and output of the target AI network.
  • the node protocol class when the number of inputs and outputs of the target AI network is multiple and continuous, includes a target number index, and includes the number of inputs and the number of outputs of the target AI network at least one of the
  • the target digital index is a digital index used to identify the first input and the first output among the multiple inputs and outputs of the target AI network in the second digital index.
  • one parameter information protocol class is used to characterize at least one of the input, output and network parameters corresponding to a network node in the target AI network, and the parameter information protocol class includes the first element type and dimension information .
  • the first element type is characterized by an integer length of quantization bits.
  • the first element type corresponds to network nodes or network parameters of the target AI network.
  • the dimension information in the parameter information protocol class includes corresponding dimension quantity values and dimension size values, the dimension quantity values in the parameter information protocol class are arranged based on a preset order, and the dimension information in the parameter information protocol class The dimension size values are arranged based on the preset order, and the preset order is a preset arrangement order of the network parameters of the target AI network.
  • the dimension size value is the number of numerical values included in the corresponding dimension, and the numerical values are normalized numerical values.
  • the target parameter information protocol class includes the numerical position of the maximum value, and the maximum value is the maximum value among the values of the target network parameters
  • the target tensor protocol class corresponding to the target network parameter includes the quantized value of the ratio of each remaining value in the target network parameter to the maximum value
  • the remaining value is the target parameter information protocol class corresponding to A value other than the stated maximum value in the dimension.
  • the parameter information protocol class includes an indication parameter for indicating a position of a non-zero value.
  • the tensor protocol class includes a second element type and a value, or, the tensor protocol class includes a value.
  • the second element type is characterized by an integer length of quantization bits.
  • the second element type is in a corresponding relationship with the first element type.
  • the arrangement order of the second element type in the tensor protocol class is the same as the arrangement order of the first element type in the parameter information protocol class.
  • the ONNX file structure includes a protocol class list, one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, tensor protocol class and attribute Protocol class.
  • the method also includes:
  • the second end receives the indication information sent by the first end; or,
  • the second end receives the ONNX file structure sent by the first end, including:
  • the second end receives the ONNX file structure carrying the indication information sent by the first end;
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • each protocol class list can form an ONNX file structure independently, and the second end can determine the start position and/or end position of the protocol class list in the ONNX file structure based on the indication information.
  • the second end may add some content to the ONNX file structure.
  • the first end is the core network device
  • the second end is the base station.
  • the ONNX file structure sent by the core network device to the base station only includes node protocol information.
  • the base station can supplement parameter information protocol information and Zhang Quantity protocol information.
  • the ONNX file structure sent by the core network device to the base station includes all node protocol information, all parameter information protocol information, and part of the tensor protocol information corresponding to the target AI network. The start position and end position of , supplement other tensor protocol class information at relevant positions.
  • the ONNX file structure includes a predefined index, and the predefined index is used to represent a preset network node.
  • the preset network node is a network node in the target AI network that has not been updated.
  • the AI network information transmission method provided by the embodiment of the present application is applied to the second end, corresponding to the AI network information transmission method applied to the first end provided in the embodiment of FIG. 2 above.
  • the specific implementation process of the relevant steps and the relevant concepts involved in the ONNX file structure can refer to the description in the method embodiment described in FIG. 2 above. To avoid repetition, details are not repeated here.
  • the second end receives the ONNX file structure sent by the first end, and then the second end can convert the ONNX file structure into an AI network under its own neural network framework, so as to realize the second end's AI network training and application.
  • two communication devices with different neural network frameworks can transmit AI network information based on the ONNX file structure, avoiding the blockage of the transmission of AI network information between communication devices.
  • the AI network information transmission method provided in the embodiment of the present application may be executed by an AI network information transmission device.
  • the AI network information transmission device provided in the embodiment of the present application is described by taking the AI network information transmission device executing the AI network information transmission method as an example.
  • FIG. 4 is a structural diagram of an AI network information transmission device provided in an embodiment of the present application. As shown in FIG. 4, the AI network information transmission device 400 includes:
  • the conversion module 401 is used to convert the AI network information of the target AI network into an open neural network exchange ONNX file structure
  • the ONNX file structure includes a target protocol class, and the target protocol class includes at least one of the following: node protocol class;
  • the ONNX file structure includes at least one computation graph protocol class, and the computation graph protocol class includes the target protocol class.
  • the target protocol class includes any one of the node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, when the number of the ONNX file structure is at least two , a described ONNX file structure corresponds to a target protocol class;
  • the sending module 402 is also configured to perform any of the following:
  • the sending module 402 is also used for:
  • the sending module 402 is also configured to:
  • the ONNX file structure includes a first number index and a second number index of an integer type, the first number index is used to identify the target protocol class, and one target protocol class corresponds to a first number An index, the second numerical index is used to identify at least one of network parameters, input and output of the target AI network.
  • the node protocol class includes the second numerical index used to characterize the input and output of the target AI network.
  • the node protocol class when the number of inputs and outputs of the target AI network is multiple and continuous, includes a target number index, and includes the number of inputs and the number of outputs of the target AI network at least one of the
  • the target digital index is a digital index used to identify the first input and the first output among the multiple inputs and outputs of the target AI network in the second digital index.
  • one parameter information protocol class is used to characterize at least one of the input, output and network parameters corresponding to a network node in the target AI network, and the parameter information protocol class includes the first element type and dimension information .
  • the first element type is characterized by an integer length of quantization bits.
  • the first element type corresponds to network nodes or network parameters of the target AI network.
  • the dimension information in the parameter information protocol class includes corresponding dimension quantity values and dimension size values, the dimension quantity values in the parameter information protocol class are arranged based on a preset order, and the dimension information in the parameter information protocol class The dimension size values are arranged based on the preset order, and the preset order is a preset arrangement order of the network parameters of the target AI network.
  • the dimension size value is the number of numerical values included in the corresponding dimension, and the numerical values are normalized numerical values.
  • the target parameter information protocol class includes the numerical position of the maximum value, and the maximum value is the maximum value among the values of the target network parameters
  • the target tensor protocol class corresponding to the target network parameter includes the quantized value of the ratio of each remaining value in the target network parameter to the maximum value
  • the remaining value is the target parameter information protocol class corresponding to A value other than the stated maximum value in the dimension.
  • the parameter information protocol class includes an indication parameter for indicating a position of a non-zero value.
  • the tensor protocol class includes a second element type and a value, or, the tensor protocol class includes a value.
  • the second element type is characterized by an integer length of quantization bits.
  • the second element type is in a corresponding relationship with the first element type.
  • the arrangement order of the second element type in the tensor protocol class is the same as the arrangement order of the first element type in the parameter information protocol class.
  • the ONNX file structure includes a protocol class list, one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, tensor protocol class and attribute Protocol class.
  • the sending module is also used for:
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • the ONNX file structure includes a predefined index, and the predefined index is used to represent a preset network node.
  • the preset network node is a network node in the target AI network that has not been updated.
  • the conversion module 401 is also used for:
  • the sending module 402 is also used for:
  • the first ONNX file structure includes a node protocol class
  • the second ONNX file structure includes a parameter information protocol class
  • the first ONNX file structure includes node protocol class and parameter information protocol class, and the second ONNX file structure includes tensor protocol class; or,
  • the first ONNX file structure includes a node protocol class
  • the second ONNX file structure includes a parameter information protocol class and a tensor protocol class.
  • the device converts the AI network information of the target AI network applicable to its own neural network framework into an ONNX file structure, and then sends the ONNX file structure to the second end, and then the second end can convert the ONNX file structure
  • the file structure is converted into an AI network under its own neural network framework.
  • the AI network information transmission apparatus 400 in the embodiment of the present application may be an electronic device, such as an electronic device with an operating system, or a component of the electronic device, such as an integrated circuit or a chip.
  • the electronic device may be a terminal, or other devices other than the terminal.
  • the terminal may include, but not limited to, the types of terminal 11 listed above, and other devices may be servers, Network Attached Storage (NAS), etc., which are not specifically limited in this embodiment of the present application.
  • NAS Network Attached Storage
  • the AI network information transmission device 400 provided in the embodiment of the present application can realize each process implemented in the method embodiment shown in FIG. 2 and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • FIG. 5 is a structural diagram of another AI network information transmission device provided in the embodiment of the present application. As shown in FIG. 5, the AI network information transmission device 500 includes:
  • the receiving module 501 is configured to receive the ONNX file structure sent by the first end, and the ONNX file structure is obtained by converting the AI network information of the target AI network by the first end.
  • the ONNX file structure includes a target protocol class, and the target protocol class includes at least one of the following: node protocol class;
  • the ONNX file structure includes at least one computation graph protocol class, and the computation graph protocol class includes the target protocol class.
  • the target protocol class includes any one of the node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, when the number of the ONNX file structure is at least two , a described ONNX file structure corresponds to a target protocol class;
  • the receiving module 501 is also configured to perform any of the following:
  • the receiving module 501 is also used for:
  • the receiving module 501 is also used for:
  • the ONNX file structure includes a first number index and a second number index of an integer type, the first number index is used to identify the target protocol class, and one target protocol class corresponds to a first number An index, the second numerical index is used to identify at least one of network parameters, input and output of the target AI network.
  • the node protocol class includes the second numerical index used to characterize the input and output of the target AI network.
  • the node protocol class when the number of inputs and outputs of the target AI network is multiple and continuous, includes a target number index, and includes the number of inputs and the number of outputs of the target AI network at least one of the
  • the target digital index is a digital index used to identify the first input and the first output among the multiple inputs and outputs of the target AI network in the second digital index.
  • one parameter information protocol class is used to characterize at least one of the input, output and network parameters corresponding to a network node in the target AI network, and the parameter information protocol class includes the first element type and dimension information .
  • the first element type is characterized by an integer length of quantization bits.
  • the first element type corresponds to network nodes or network parameters of the target AI network.
  • the dimension information in the parameter information protocol class includes corresponding dimension quantity values and dimension size values, the dimension quantity values in the parameter information protocol class are arranged based on a preset order, and the dimension information in the parameter information protocol class The dimension size values are arranged based on the preset order, and the preset order is a preset arrangement order of the network parameters of the target AI network.
  • the dimension size value is the number of numerical values included in the corresponding dimension, and the numerical values are normalized numerical values.
  • the target parameter information protocol class includes the numerical position of the maximum value, and the maximum value is the maximum value among the values of the target network parameters
  • the target tensor protocol class corresponding to the target network parameter includes the quantized value of the ratio of each remaining value in the target network parameter to the maximum value
  • the remaining value is the target parameter information protocol class corresponding to A value other than the stated maximum value in the dimension.
  • the parameter information protocol class includes an indication parameter for indicating a position of a non-zero value.
  • the tensor protocol class includes a second element type and a value
  • the tensor protocol class includes a value
  • the second element type is in a corresponding relationship with the first element type.
  • the arrangement order of the second element type in the tensor protocol class is the same as the arrangement order of the first element type in the parameter information protocol class.
  • the ONNX file structure includes a protocol class list, one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, tensor protocol class and attribute Protocol class.
  • the receiving module 501 is also used for:
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • the ONNX file structure includes a predefined index, and the predefined index is used to represent a preset network node.
  • the preset network node is a network node in the target AI network that has not been updated.
  • the device can receive the ONNX file structure sent by the first end, and then the device can convert the ONNX file structure into an AI network under its own neural network framework, so as to realize the training and training of the AI network. application. In this way, even two communication devices with different neural network frameworks can transmit AI network information based on the ONNX file structure, avoiding the blockage of the transmission of AI network information between communication devices.
  • the AI network information transmission apparatus 500 in the embodiment of the present application may be an electronic device, such as an electronic device with an operating system, or a component of the electronic device, such as an integrated circuit or a chip.
  • the electronic device may be a terminal, or other devices other than the terminal.
  • the terminal may include, but not limited to, the types of terminal 11 listed above, and other devices may be servers, Network Attached Storage (NAS), etc., which are not specifically limited in this embodiment of the present application.
  • NAS Network Attached Storage
  • the AI network information transmission device 500 provided in the embodiment of the present application can realize each process implemented in the method embodiment shown in FIG. 3 and achieve the same technical effect. To avoid repetition, details are not repeated here.
  • this embodiment of the present application also provides a communication device 600, including a processor 601 and a memory 602, and the memory 602 stores programs or instructions that can run on the processor 601.
  • the programs or instructions are executed by the processor 601, the various steps of the embodiment of the AI network information transmission method described above in FIG. 2 or FIG. 3 can be realized, and the same technical effect can be achieved. To avoid repetition, details are not repeated here.
  • FIG. 7 is a schematic diagram of a hardware structure of a terminal implementing an embodiment of the present application.
  • the terminal 700 includes, but is not limited to: a radio frequency unit 701, a network module 702, an audio output unit 703, an input unit 704, a sensor 705, a display unit 706, a user input unit 707, an interface unit 708, a memory 709, and a processor 710. At least some parts.
  • the terminal 700 may also include a power supply (such as a battery) for supplying power to various components, and the power supply may be logically connected to the processor 710 through the power management system, so as to manage charging, discharging, and power consumption through the power management system. Management and other functions.
  • a power supply such as a battery
  • the terminal structure shown in FIG. 7 does not constitute a limitation on the terminal, and the terminal may include more or fewer components than shown in the figure, or combine some components, or arrange different components, which will not be repeated here.
  • the input unit 704 may include a graphics processing unit (Graphics Processing Unit, GPU) 7041 and a microphone 7042, and the graphics processor 7041 is used by the image capture device (such as the image data of the still picture or video obtained by the camera) for processing.
  • the display unit 706 may include a display panel 7061, and the display panel 7061 may be configured in the form of a liquid crystal display, an organic light emitting diode, or the like.
  • the user input unit 707 includes at least one of a touch panel 7071 and other input devices 7072 .
  • the touch panel 7071 is also called a touch screen.
  • the touch panel 7071 may include two parts, a touch detection device and a touch controller.
  • Other input devices 7072 may include, but are not limited to, physical keyboards, function keys (such as volume control buttons, switch buttons, etc.), trackballs, mice, and joysticks, which will not be described in detail here.
  • the radio frequency unit 701 may transmit the downlink data from the network side device to the processor 710 for processing after receiving the downlink data; in addition, the radio frequency unit 701 may send uplink data to the network side device.
  • the radio frequency unit 701 includes, but is not limited to, an antenna, an amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, and the like.
  • the memory 709 can be used to store software programs or instructions as well as various data.
  • the memory 709 may mainly include a first storage area for storing programs or instructions and a second storage area for storing data, wherein the first storage area may store an operating system, an application program or instructions required by at least one function (such as a sound playing function, image playback function, etc.), etc.
  • memory 709 may include volatile memory or nonvolatile memory, or, memory 709 may include both volatile and nonvolatile memory.
  • the non-volatile memory can be read-only memory (Read-Only Memory, ROM), programmable read-only memory (Programmable ROM, PROM), erasable programmable read-only memory (Erasable PROM, EPROM), electronically programmable Erase Programmable Read-Only Memory (Electrically EPROM, EEPROM) or Flash.
  • ROM Read-Only Memory
  • PROM programmable read-only memory
  • Erasable PROM Erasable PROM
  • EPROM erasable programmable read-only memory
  • Electrical EPROM Electrical EPROM
  • EEPROM electronically programmable Erase Programmable Read-Only Memory
  • Volatile memory can be random access memory (Random Access Memory, RAM), static random access memory (Static RAM, SRAM), dynamic random access memory (Dynamic RAM, DRAM), synchronous dynamic random access memory (Synchronous DRAM, SDRAM), double data rate synchronous dynamic random access memory (Double Data Rate SDRAM, DDRSDRAM), enhanced synchronous dynamic random access memory (Enhanced SDRAM, ESDRAM), synchronous connection dynamic random access memory (Synch link DRAM , SLDRAM) and Direct Memory Bus Random Access Memory (Direct Rambus RAM, DRRAM).
  • RAM Random Access Memory
  • SRAM static random access memory
  • DRAM dynamic random access memory
  • DRAM synchronous dynamic random access memory
  • SDRAM double data rate synchronous dynamic random access memory
  • Double Data Rate SDRAM Double Data Rate SDRAM
  • DDRSDRAM double data rate synchronous dynamic random access memory
  • Enhanced SDRAM, ESDRAM enhanced synchronous dynamic random access memory
  • Synch link DRAM , SLDRAM
  • Direct Memory Bus Random Access Memory Direct Rambus
  • the processor 710 may include one or more processing units; optionally, the processor 710 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to the operating system, user interface, and application programs, etc., Modem processors mainly process wireless communication signals, such as baseband processors. It can be understood that the foregoing modem processor may not be integrated into the processor 710 .
  • the terminal 700 is the first end.
  • the processor 710 is used to convert the AI network information of the target AI network into an open neural network exchange ONNX file structure;
  • the radio frequency unit 701 is configured to send the ONNX file structure to the second end.
  • the ONNX file structure includes a target protocol class, and the target protocol class includes at least one of the following:
  • the ONNX file structure includes at least one computation graph protocol class, and the computation graph protocol class includes the target protocol class.
  • the target protocol class includes any one of the node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, when the number of the ONNX file structure is at least two , a described ONNX file structure corresponds to a target protocol class;
  • the radio frequency unit 701 is configured to perform any of the following:
  • the radio frequency unit 701 is configured to:
  • the radio frequency unit 701 is configured to:
  • the ONNX file structure includes a first number index and a second number index of an integer type, the first number index is used to identify the target protocol class, and one target protocol class corresponds to a first number An index, the second numerical index is used to identify at least one of network parameters, input and output of the target AI network.
  • the node protocol class includes the second numerical index used to characterize the input and output of the target AI network.
  • the node protocol class when the number of inputs and outputs of the target AI network is multiple and continuous, includes a target number index, and includes the number of inputs and the number of outputs of the target AI network at least one of the
  • the target digital index is a digital index used to identify the first input and the first output among the multiple inputs and outputs of the target AI network in the second digital index.
  • one parameter information protocol class is used to characterize at least one of the input, output and network parameters corresponding to a network node in the target AI network, and the parameter information protocol class includes the first element type and dimension information .
  • the first element type is characterized by an integer length of quantization bits.
  • the first element type corresponds to network nodes or network parameters of the target AI network.
  • the dimension information in the parameter information protocol class includes corresponding dimension quantity values and dimension size values, the dimension quantity values in the parameter information protocol class are arranged based on a preset order, and the dimension information in the parameter information protocol class The dimension size values are arranged based on the preset order, and the preset order is a preset arrangement order of the network parameters of the target AI network.
  • the dimension size value is the number of numerical values included in the corresponding dimension, and the numerical values are normalized numerical values.
  • the target parameter information protocol class includes the numerical position of the maximum value, and the maximum value is the maximum value among the values of the target network parameters
  • the target tensor protocol class corresponding to the target network parameter includes the quantized value of the ratio of each remaining value in the target network parameter to the maximum value
  • the remaining value is the target parameter information protocol class corresponding to A value other than the stated maximum value in the dimension.
  • the parameter information protocol class includes an indication parameter for indicating a position of a non-zero value.
  • the tensor protocol class includes a second element type and a value, or, the tensor protocol class includes a value.
  • the second element type is characterized by an integer length of quantization bits.
  • the second element type is in a corresponding relationship with the first element type.
  • the arrangement order of the second element type in the tensor protocol class is the same as the arrangement order of the first element type in the parameter information protocol class.
  • the ONNX file structure includes a protocol class list, one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, tensor protocol class and attribute Protocol class.
  • the radio frequency unit 701 is also used for:
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • the ONNX file structure includes a predefined index, and the predefined index is used to represent a preset network node.
  • the preset network node is a network node in the target AI network that has not been updated.
  • processor 710 is further configured to:
  • the radio frequency unit 701 is also used for:
  • the first ONNX file structure includes a node protocol class
  • the second ONNX file structure includes a parameter information protocol class
  • the first ONNX file structure includes a node protocol class and a parameter information protocol class
  • the second ONNX file structure includes a tensor protocol class
  • the first ONNX file structure includes a node protocol class
  • the second ONNX file structure includes a parameter information protocol class and a tensor protocol class.
  • the terminal 700 may serve as the second terminal.
  • the radio frequency unit 701 is configured to receive the ONNX file structure sent by the first end, and the ONNX file structure is obtained by converting the AI network information of the target AI network by the first end.
  • the ONNX file structure includes a target protocol class, and the target protocol class includes at least one of the following:
  • the target protocol class includes any one of the node protocol class, parameter information protocol class, tensor protocol class, and attribute protocol class, when the number of the ONNX file structure is at least two , a described ONNX file structure corresponds to a target protocol class;
  • the radio frequency unit 701 is configured to perform any of the following:
  • the radio frequency unit 701 is also used for:
  • the ONNX file structure includes a protocol class list, one protocol class list corresponds to a protocol class type, and the protocol class type includes the node protocol class, parameter information protocol class, tensor protocol class and attribute Protocol class.
  • the radio frequency unit 701 is also used for:
  • the indication information is used to indicate at least one of a start position and an end position of the protocol class list in the ONNX file structure.
  • the terminal 700 provided by the embodiment of the present application can be used as the first terminal or the second terminal to execute the AI network information transmission method described above in FIG. 2 or FIG. 3 , and can achieve the same technical effect, which will not be repeated here.
  • the embodiment of the present application also provides a network-side device.
  • the various implementation processes and implementation methods of the above-mentioned method embodiments shown in FIG. 2 and FIG. 3 can be applied to the network-side device embodiment, and can achieve the same technical effect.
  • the embodiment of the present application also provides a network side device.
  • the network side device 800 includes: an antenna 81 , a radio frequency device 82 , a baseband device 83 , a processor 84 and a memory 85 .
  • the antenna 81 is connected to a radio frequency device 82 .
  • the radio frequency device 82 receives information through the antenna 81, and sends the received information to the baseband device 83 for processing.
  • the baseband device 83 processes the information to be sent and sends it to the radio frequency device 82
  • the radio frequency device 82 processes the received information and sends it out through the antenna 81 .
  • the method performed by the network side device in the above embodiments may be implemented in the baseband device 83, where the baseband device 83 includes a baseband processor.
  • the baseband device 83 can include at least one baseband board, for example, a plurality of chips are arranged on the baseband board, as shown in FIG.
  • the program executes the network device operations shown in the above method embodiments.
  • the network side device may also include a network interface 86, such as a common public radio interface (common public radio interface, CPRI).
  • a network interface 86 such as a common public radio interface (common public radio interface, CPRI).
  • the network-side device 800 in this embodiment of the present invention further includes: instructions or programs stored in the memory 85 and operable on the processor 84, and the processor 84 calls the instructions or programs in the memory 85 to execute FIG. 4 or FIG. 5
  • the methods executed by each module shown in the figure achieve the same technical effect, so in order to avoid repetition, they are not repeated here.
  • the embodiment of the present application also provides another network side device.
  • the network side device 900 includes: a processor 901 , a network interface 902 and a memory 903 .
  • the network interface 902 is, for example, a common public radio interface (common public radio interface, CPRI).
  • the network-side device 900 in this embodiment of the present invention further includes: instructions or programs stored in the memory 903 and operable on the processor 901, and the processor 901 calls the instructions or programs in the memory 903 to execute FIG. 4 or FIG. 5
  • the methods executed by each module shown in the figure achieve the same technical effect, so in order to avoid repetition, they are not repeated here.
  • the embodiment of the present application also provides a readable storage medium, on which a program or instruction is stored, and when the program or instruction is executed by the processor, each process of the method embodiment described above in FIG. 2 or FIG. 3 is implemented. , and can achieve the same technical effect, in order to avoid repetition, it will not be repeated here.
  • the processor is the processor in the terminal described in the foregoing embodiments.
  • the readable storage medium includes a computer-readable storage medium, such as a computer read-only memory ROM, a random access memory RAM, a magnetic disk or an optical disk, and the like.
  • the embodiment of the present application further provides a chip, the chip includes a processor and a communication interface, the communication interface is coupled to the processor, and the processor is used to run programs or instructions to implement the above-mentioned Figure 2 or Figure 3.
  • the chip includes a processor and a communication interface
  • the communication interface is coupled to the processor
  • the processor is used to run programs or instructions to implement the above-mentioned Figure 2 or Figure 3.
  • the chip mentioned in the embodiment of the present application may also be called a system-on-chip, a system-on-chip, a system-on-a-chip, or a system-on-a-chip.
  • An embodiment of the present application further provides a computer program/program product, the computer program/program product is stored in a non-volatile storage medium, and the computer program/program product is executed by at least one processor to implement the above-mentioned 2 or the various processes of the method embodiment described in FIG. 3 can achieve the same technical effect, so in order to avoid repetition, details are not repeated here.
  • the embodiment of the present application also provides a communication system, including: a terminal and a network-side device, the terminal can be used to perform the steps of the method described in Figure 2 above, and the network-side device can be used to perform the method described in Figure 3 above Alternatively, the terminal may be used to perform the steps of the method described in FIG. 3 above, and the network side device may be used to perform the steps of the method described in FIG. 2 above.
  • the term “comprising”, “comprising” or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, article or apparatus comprising a set of elements includes not only those elements, It also includes other elements not expressly listed, or elements inherent in the process, method, article, or device. Without further limitations, an element defined by the phrase “comprising a " does not preclude the presence of additional identical elements in the process, method, article, or apparatus comprising that element.
  • the scope of the methods and devices in the embodiments of the present application is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in reverse order according to the functions involved. Functions are performed, for example, the described methods may be performed in an order different from that described, and various steps may also be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.
  • the methods of the above embodiments can be implemented by means of software plus a necessary general-purpose hardware platform, and of course also by hardware, but in many cases the former is better implementation.
  • the technical solution of the present application can be embodied in the form of computer software products, which are stored in a storage medium (such as ROM/RAM, magnetic disk, etc.) , CD-ROM), including several instructions to make a terminal (which may be a mobile phone, a computer, a server, an air conditioner, or a network device, etc.) execute the methods described in the various embodiments of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Computing Systems (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Artificial Intelligence (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Communication Control (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

本申请公开了一种AI网络信息传输方法、装置及通信设备,属于通信技术领域,本申请实施例的AI网络信息传输方法包括:第一端将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;所述第一端向第二端发送所述ONNX文件结构。

Description

AI网络信息传输方法、装置及通信设备
相关申请的交叉引用
本申请主张在2021年12月31日在中国提交的中国专利申请No.202111666991.3的优先权,其全部内容通过引用包含于此。
技术领域
本申请属于通信技术领域,具体涉及一种AI网络信息传输方法、装置及通信设备。
背景技术
人工智能(Artificial Intelligence,AI)是研究和开发用于模拟、延伸、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,受到人们的广泛关注,针对AI的应用也越来越广泛。目前,人们已经开始研究将AI网络应用在通信系统中,例如网络侧设备和终端之间可以通过AI网络来传输通信数据。在通信系统中,不同的通信设备应用的神经网络框架不同,而不同的神经网络框架下的AI网络信息保存的文件结构不同,导致通信设备可能不能读取与其神经网络框架不同的其他通信设备传输的AI网络信息,造成通信设备之间AI网络信息的传输受限。
发明内容
本申请实施例提供一种AI网络信息传输方法、装置及通信设备,能够解决相关技术中通信设备之间AI网络信息的传输受限的问题。
第一方面,提供了一种AI网络信息传输方法,包括:
第一端将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
所述第一端向第二端发送所述ONNX文件结构。
第二方面,提供了一种AI网络信息传输方法,包括:
第二端接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
第三方面,提供了一种AI网络信息传输装置,包括:
转换模块,用于将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
发送模块,用于向第二端发送所述ONNX文件结构。
第四方面,提供了一种AI网络信息传输装置,包括:
接收模块,用于接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
第五方面,提供了一种通信设备,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,所述程序或指令被所述处理器执行时实现如第一方面所述的AI网络信息传输方法的步骤,或者实现如第二方面所述的AI网络信息传输方法的步骤。
第六方面,提供了一种可读存储介质,所述可读存储介质上存储程序或指令,所述程序或指令被处理器执行时实现如第一方面所述的AI网络信息传输方法的步骤,或者实现如第二方面所述的AI网络信息传输方法的步骤。
第七方面,提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和所述处理器耦合,所述处理器用于运行程序或指令,实现如第一方面所述的AI网络信息传输方法的步骤,或者实现如第二方面所述的AI网络信息传输方法的步骤。
第八方面,提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现如第一方面所述的AI网络信息传输方法的步骤,或者实现如第二方面所述的AI网络信息传输方法的步骤。
在本申请实施例中,第一端将基于自身神经网络框架的目标AI网络的AI网络信息转换成ONNX文件结构,然后将所述ONNX文件结构发送给第二端,进而第二端能够将ONNX文件结构转换成自身神经网络框架下的AI网络。这样,也就使得即使神经网络框架不同的两个通信设备,也能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
附图说明
图1是本申请实施例可应用的一种无线通信系统的框图;
图2是本申请实施例提供的一种AI网络信息传输方法的流程图;
图3是本申请实施例提供的另一种AI网络信息传输方法的流程图;
图4是本申请实施例提供的一种AI网络信息传输装置的结构图;
图5是本申请实施例提供的另一种AI网络信息传输装置的结构图;
图6是本申请实施例提供的一种通信设备的结构图;
图7是本申请实施例提供的一种终端的结构图;
图8是本申请实施例提供的一种网络侧设备的结构图;
图9是本申请实施例提供的另一种网络侧设备的结构图。
具体实施方式
下面将结合本申请实施例中的附图,对本申请实施例中的技术方案进行清楚描述,显然,所描述的实施例是本申请一部分实施例,而不是全部的实施例。基于本申请中的实施例,本领域普通技术人员所获得的所有其他实施例,都属于本申请保护的范围。
本申请的说明书和权利要求书中的术语“第一”、“第二”等是用于区别类似的对象,而不用于描述特定的顺序或先后次序。应该理解这样使用的术语在适当情况下可以互换,以便本申请的实施例能够以除了在这里图示或描述的那些以外的顺序实施,且“第一”、“第二”所区别的对象通常为一类,并不限定对象的个数,例如第一对象可以是一个,也可以是多个。此外,说明书以及权利要求中“和/或”表示所连接对象的至少其中之一,字符“/”一般表示前后关联对象是一种“或”的关系。
值得指出的是,本申请实施例所描述的技术不限于长期演进型(Long Term Evolution,LTE)/LTE的演进(LTE-Advanced,LTE-A)系统,还可用于其他无线通信系统,诸如码分多址(Code Division Multiple Access,CDMA)、时分多址(Time Division Multiple Access,TDMA)、频分多址(Frequency Division Multiple Access,FDMA)、正交频分多址(Orthogonal Frequency Division Multiple Access,OFDMA)、单载波频分多址(Single-carrier Frequency Division Multiple Access,SC-FDMA)和其他系统。本申请实施例中的术语“系统”和“网络”常被可互换地使用,所描述的技术既可用于以上提及的系统和无线电技术,也可用于其他系统和无线电技术。以下描述出于示例目的描述了新空口(New Radio,NR)系统,并且在以下大部分描述中使用NR术语,但是这些技术也可应用于NR系统应用以外的应用,如第6代(6 th Generation,6G)通信系统。
图1示出本申请实施例可应用的一种无线通信系统的框图。无线通信系统包括终端11和网络侧设备12。其中,终端11可以是手机、平板电脑(Tablet Personal Computer)、膝上型电脑(Laptop Computer)或称为笔记本电脑、个人数字助理(Personal Digital Assistant,PDA)、掌上电脑、上网本、超级移动个人计算机(ultra-mobile personal computer,UMPC)、移动上网装置(Mobile Internet Device,MID)、增强现实(augmented reality,AR)/虚拟现实(virtual reality,VR)设备、机器人、可穿戴式设备(Wearable Device)、车载设备(Vehicle User Equipment,VUE)、行人终端(Pedestrian User Equipment,PUE)、智能家居(具有无线通信功能的家居设备,如冰箱、电视、洗衣机或者家具等)、 游戏机、个人计算机(personal computer,PC)、柜员机或者自助机等终端侧设备,可穿戴式设备包括:智能手表、智能手环、智能耳机、智能眼镜、智能首饰(智能手镯、智能手链、智能戒指、智能项链、智能脚镯、智能脚链等)、智能腕带、智能服装等。需要说明的是,在本申请实施例并不限定终端11的具体类型。网络侧设备12可以包括接入网设备或核心网设备,其中,接入网设备也可以称为无线接入网设备、无线接入网(Radio Access Network,RAN)、无线接入网功能或无线接入网单元。接入网设备可以包括基站、无线局域网(Wireless Local Area Networks,WLAN)接入点或WiFi节点等,基站可被称为节点B、演进节点B(eNB)、接入点、基收发机站(Base Transceiver Station,BTS)、无线电基站、无线电收发机、基本服务集(Basic Service Set,BSS)、扩展服务集(Extended Service Set,ESS)、家用B节点、家用演进型B节点、发送接收点(Transmitting Receiving Point,TRP)或所述领域中其他某个合适的术语,只要达到相同的技术效果,所述基站不限于特定技术词汇,需要说明的是,在本申请实施例中仅以NR系统中的基站为例进行介绍,并不限定基站的具体类型。核心网设备可以包含但不限于如下至少一项:核心网节点、核心网功能、移动管理实体(Mobility Management Entity,MME)、接入移动管理功能(Access and Mobility Management Function,AMF)、会话管理功能(Session Management Function,SMF)、用户平面功能(User Plane Function,UPF)、策略控制功能(Policy Control Function,PCF)、策略与计费规则功能单元(Policy and Charging Rules Function,PCRF)、边缘应用服务发现功能(Edge Application Server Discovery Function,EASDF)、统一数据管理(Unified Data Management,UDM),统一数据仓储(Unified Data Repository,UDR)、归属用户服务器(Home Subscriber Server,HSS)、集中式网络配置(Centralized network configuration,CNC)、网络存储功能(Network Repository Function,NRF),网络开放功能(Network Exposure Function,NEF)、本地NEF(Local NEF,或L-NEF)、绑定支持功能(Binding Support Function,BSF)、应用功能(Application Function,AF)等。需要说明的是,在本申请实施例中仅以NR系统中的核心网设备为例进行介绍,并不限定核心网设备的具体类型。
为更好地理解,以下对本申请实施例中可能涉及的相关概念进行解释说明。
神经网络框架
神经网络有很多种实现框架,包括TensorFlow,PyTorch,Keras,MXNet,Caffe2等,每种框架的侧重点各不相同。例如,Caffe2与Keras是高层的深度学习框架可以快速地验证模型,TensorFlow与PyTorch是底层的深度学习 框架可以实现对神经网络底层细节的修改。又例如,PyTorch的重点在于支持动态图模型,TensorFlow重点在于支持多种硬件,运行速度快,Caffe2在于轻量级等。每一个实现框架都会使用自己的方法对神经网络进行描述,完成网络的搭建、训练、推断等操作。
通常,两个不同的神经网络框架下的网络信息保存的文件结构不同,无法直接读取,需要通过其他的标准结构进行交互。
开放式神经网络交互(Open Neural Network Exchange,ONNX)
ONNX是一种AI交互网络,ONNX本身只是一种数据结构,用于描述一个AI网络,不包括实现方案。ONNX在计算图协议类(GraphProto)中保存整个AI网络,包括网络结构和网络参数,通过节点协议类(NodeProto)、参数信息协议类(ValueInfoProto)、张量协议类(TensorProto)完成AI网络的基本描述。其中,NodeProto用于描述网络结构,ONNX将AI网络的每个算子都描述成一个节点,每个节点的输入和输出的名称都是全局唯一的,通过输入和输出名称的匹配关系描述整个AI网络的网络结构。所有的网络参数都被视为输入或输出,也是通过名称检索,因此ONNX可以通过一系列的NodeProto实体表述整个网络的结构。ValueInfoProto用于描述所有输入输出的信息,包括维度和元素类型,表明了对应元素的大小,每一个输入输出的名称和NodeProto中记录的对应。TensorProto用于存储具体网络参数的数值,根据每个节点输入输出的名称去存储位置获取对应的参数。
另外,ONNX通过属性协议类(AttributeProto)记录节点的功能,例如卷积层、乘法层等,赋予对应的节点功能,而这些功能所需要的超参数的数值都保存在TensorProto中,超参数的维度都保存在ValueInfoProto中。
在一些实施例中,协议类(Proto,Protocol)也可以称为结构体,如节点结构体(NodeProto)、参数信息结构体(ValueInfoProto)、张量结构体(TensorProto)、属性结构体(AttributeProto)等。Proto可以认为是一种基于特定协议形成的数据、文件等。
下面结合附图,通过一些实施例及其应用场景对本申请实施例提供的AI网络信息传输方法进行详细地说明。
请参照图2,图2是本申请实施例提供的一种AI网络信息传输方法的流程图,如图2所示,所述方法包括以下步骤:
步骤201、第一端将目标AI网络的AI网络信息转换成ONNX文件结构;
步骤202、所述第一端向第二端发送所述ONNX文件结构。
本申请实施例中,第一端将目标AI网络的AI网络信息转换成ONNX文件结构,例如若AI网络信息包括目标AI网络完整的网络结构和全部的网络参数,则第一端将所述网络结构和网络参数基于ONNX的结构形式表述成 ONNX文件结构,例如通过节点协议类(NodeProto)和参数信息协议类(ValueInfoProto)来描述网络结构,通过张量协议类(TensorProto)来描述网络参数,等。
进一步地,第一端将所述ONNX文件结构发送给第二端,第二端将所述ONNX文件结构转换成自身神经网络框架下的AI网络,以实现第二端对AI网络的训练和应用。
本申请实施例中,第一端将适用于自身神经网络框架的目标AI网络的AI网络信息转换成ONNX文件结构,然后将所述ONNX文件结构发送给第二端,进而第二端能够将ONNX文件结构转换成自身神经网络框架下的AI网络。这样,也就使得神经网络框架不同的两个通信设备,能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
需要说明地,所述AI网络信息包括目标AI网络的网络结构和网络参数中的至少一项。例如,所述AI网络信息包括目标AI网络完整的网络结构和全部的网络参数,或者所述AI网络信息包括目标AI网络完整的网络结构,或者仅包括目标AI网络中更新的AI网络结构,或者仅包括目标AI网络的部分网络参数,或者仅包括目标AI网络的网络参数的部分数值,等等。这样,也就使得AI网络的网络结构和网络参数可以分开发送,进而在通信过程中也就无需将包括整个网络结构和网络参数的全部AI网络一起进行传输,能够有效降低通信过程中的传输开销。
本申请实施例中,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:节点协议类(NodeProto)、参数信息协议类(ValueInfoProto)、张量协议类(TensorProto)、属性协议类(AttributeProto)。
其中,NodeProto用于描述网络结构;ValueInfoProto用于描述所有输入输出及网络参数的信息,包括维度和元素类型,表明了对应元素的大小,每一个输入输出及网络参数的名称或索引和NodeProto中记录的对应;TensorProto用于存储具体网络参数的数值,根据每个节点输入输出的名称去存储位置获取对应的参数;AttributeProto用于记录节点的功能,例如卷积层、乘法层等,赋予对应的节点功能,而这些功能所需要的超参数的数值都保存在TensorProto中,超参数的维度都保存在ValueInfoProto中。
需要说明地,目标协议类包括上述节点协议类、参数信息协议类、张量协议类和属性协议类中的至少一项,所述ONNX文件结构也就包括上述节点协议类、参数信息协议类、张量协议类和属性协议类中的至少一项。例如,若目标协议类包括节点协议类,所述节点协议类可以组成完整的onnx.proto文件进行ONNX文件结构的生成;若目标协议类包括节点协议类、参数信息协议类和张量协议类,则基于所述节点协议类、参数信息协议类和张量协议 类组成onnx.proto文件进行ONNX文件结构的生成。当然,所述ONNX文件结构中包括的协议类还可以是其他的可能情况,此处不做过多列举。
可选地,所述ONNX文件结构包括至少一个计算图协议类(GraphProto),所述计算图协议类包括所述目标协议类。也就是说,在所述目标协议类包括上述节点协议类、参数信息协议类、张量协议类和属性协议类中的至少一项的情况下,所述目标协议类包括的内容可以写进计算图协议类中。
例如,所述目标协议类包括节点协议类、参数信息协议类和张量协议类,则将所述节点协议类、参数信息协议类和张量协议类写进计算图协议类中,基于所述计算图协议类生成ONNX文件结构,第一端向第二端发送所述ONNX文件结构。这样也就将节点协议类、参数信息协议类和张量协议类组合在了一起进行发送。
可选地,ONNX文件结构可以是包括多个图像协议类,每个计算图协议类中包括的目标协议类可以不一样。例如,ONNX文件结构包括两个计算图协议类,其中一个计算图协议类包括节点协议类、参数信息协议类,另一个计算图协议类中包括张量协议类。另外,每一个计算图协议类中也可以包括多个节点协议类、参数信息协议类、张量协议类和属性协议类。
需要说明地,ONNX文件结构还可以包括模型协议类(ModelProto),模型协议类包括一个计算图协议类,计算图协议类中包括若干个节点协议类、参数信息协议类和张量协议类,节点协议类中可以包含若干个属性协议类。本申请实施例中,所述ONNX文件结构可以包括多个计算图协议类,无论有没有模型协议类;也即本申请实施例中的ONNX文件结构可以不包括模型协议类,也就无需传输模型协议类,这样能够节省通信设备的传输开销,更有利于ONNX文件结构在空口传输中的应用。
可选地,所述目标协议类可以是包括节点协议类(NodeProto)、参数信息协议类(ValueInfoProto)、张量协议类(TensorProto)和属性协议类(AttributeProto)中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类。这种情况下,所述第一端向第二端发送所述ONNX文件结构,包括如下任意一项:
所述第一端向第二端合并发送所述至少两个ONNX文件结构;
所述第一端向第二端分别发送所述至少两个ONNX文件结构。
本申请实施例中,每一个proto都可以独立成ONNX文件结构,进而一个ONNX文件结构也就对应一种类型的proto。例如,若目标AI网络的AI网络信息包括网络结构和权重参数,所述AI网络信息基于ONNX的结构形式生成节点协议类、参数信息协议类和张量协议类,则节点协议类对应生成一个ONNX文件结构,参数信息协议类对应生成一个ONNX文件结构,张 量协议类对应生成一个ONNX文件结构,也就会得到三个ONNX文件结构。这样,也就使得ONNX文件结构的生成更为灵活,也有利于第二端对ONNX文件结构的区分,同时可以在不同的时隙、时频位置传输ONNX文件结构,更有效的利用时频资源,方便调度。
可选地,第一端可以是向第二端分别发送这三个ONNX文件结构,例如先后发送这三个ONNX文件结构;或者,第一端也可以是将这三个ONNX文件结构进行合并,并向第二端一次发送合并后的这三个ONNX文件结构。进而,第二端对于接收到的ONNX文件结构,也就能够明确一个ONNX文件结构对应一种proto类型,更便于第二端将ONNX文件结构转换成自身神经网络框架下的AI网络信息。
本申请实施例中,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述第一端向第二端发送所述ONNX文件结构之后,所述方法还包括:
所述第一端向第二端发送所述目标AI网络的网络参数的数值。
具体地,第一端在向第二传递包含节点协议类和参数信息协议类的ONNX文件结构后,第一端还可以向第二端发送目标AI网络的网络参数的数值。可以理解地,由于第一端已经向第二端发送了包含节点协议类和参数信息协议类的ONNX文件结构,进而第二端对于网络参数的数量和开销也就是已知的,第一端可以按照参数信息协议类的顺序直接发送网络参数的具体数值,而不用转换成ONNX文件结构,第二端按照顺序接收每个参数的具体数值。其中,所述节点协议类和参数信息协议类可以是包含在一个ONNX文件结构中,也可以是各自对应一个ONNX文件结构。
可选地,所述第一端向第二端发送所述目标AI网络的网络参数的数值,包括:
所述第一端通过数据信道向第二端发送所述目标AI网络的网络参数的数值。
本申请实施例中,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参数、输入和输出中的至少一项。
可以理解地,ONNX文件结构中所包括的协议类可以是通过第一数字索引来标识。本申请实施例中,所述目标协议类包括节点协议类、参数信息协议类、张量协议类和属性协议类中的至少一项,一个协议类可以对应一个第一数字索引,例如ONNX文件结构总共包括5个协议类,也就可以通过5个第一数字索引(例如1、2、3、4、5)来表征这5个协议类。
或者,也可以是一种协议类类型对应一个数字索引,例如总共包括5个协议类,4种协议类类型,则可以是通过4个第一数字索引来表征这5个协议类。
或者,也可以是每种协议类类型对应一个数字索引序列,例如总共包括3个节点协议类,4个参数信息协议类和1个张量协议类,可以用数字索引0、1、2表示3个节点协议类,0、1、2、3表示4个参数信息协议类,0表示1个张量协议类,三种协议类的序列由特定标识符区分。
本申请实施例中,对于目标协议类描述的内容可以通过第二数字索引来标识。例如,节点协议类用于描述目标AI网络的网络结构,网络结构的每个节点的输入和输出的名称都是唯一的,一个节点协议类可以是用于描述网络结构中一个节点的输入和输入以及相关的网络参数,进而节点协议类对应的输入、输出和网络参数可以是通过第二数字索引来标识,例如输入、输出和网络参数分别对应三个不同的第二数字索引。同样地,对于其他协议类描述的内容,也可以是通过第二数字索引来进行标识。
这样,对于ONNX文件结构,通过第一数字索引来标识ONNX文件结构包括的协议类,通过第二数字索引来标识协议类描述的内容。其中,所述第一数字索引和第二数字索引均为整型(int)类型的数字索引,相比于通过字符串进行标识,通过整型类型的数字索引进行标识能够更有效地节省传输开销。
可选地,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输出的所述第二数字索引。需要说明地,其他协议类,例如参数信息协议类、张量协议类、属性协议类,也可以是通过第二数字索引来标识该协议类所描述的内容,如目标AI网络的网络参数、输入、输出等。
可选地,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项,其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
需要说明地,ONNX文件结构中通过节点协议类来描述目标AI网络的网络结构,网络结构包括多个节点,一个节点对应至少一个输入和至少一个输入,在网络结构包括的节点数量为多个且连续的情况下,一个节点的输入和输出对应一个第二数字索引,进而也就应该对应多个连续的第二数字索引,这种情况下,ONNX文件结构中的节点协议类可以只包括用于标识多个节点中第一个节点的输入和输出的第二数字索引(也即目标数字索引)和总共的 节点的数量。
例如,目标AI网络的网络结构包括3个节点,同一个节点的输入和输出对应一个第二数字索引,进而也就对应3个第二数字索引,例如11、12、13;这种情况下,节点协议类包括用于标识第一个节点的输入和输出对应的第二数字索引和总共的节点的数量(也即输入和输出的数量),那么ONNX文件结构的节点协议类也就包括11和3,这样也就无需包括所有的第二数字索引。在目标AI网络的输入和输出数量较多的情况下,无需包括用于标识每一个输入和输出的第二数字索引,节点协议类仅需要包括目标数字索引和总的输入输出数量即可,这样也就能够节省ONNX文件结构的内容,有效节省ONNX文件结构的传输开销。而第二端在接收到所述ONNX文件结构后,基于所述目标数字索引和总的输入输出数量,能够得出每一个输入和输出对应的第二数字索引。
本申请实施例中,一个所述参数信息协议类用于描述所述目标AI网络中的网络节点对应的输入、输出以及网络参数中的至少一项,所述参数信息协议类包括第一元素类型和维度信息。具体地,参数信息协议类用于描述目标AI网络所有输入输出的信息,所述参数信息协议类包括维度信息和第一元素类型,表明了对应元素的大小,每一个输入输出的名称和节点协议类中记录的输入输出对应。
可选地,所述第一元素类型通过量化位数的整型长度进行表征。也就是说,通过量化位数代替第一元素类型,这样第一元素类型可以省略,使用缺省的类型,例如整型(int)长度,也即量化位数是一个整型长度,进而第一元素类型可以通过指定整型长度来进行表征。
可选地,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。可以是通过预先预定好的对应关系,例如某一种类网络节点或者网络参数使用对应的整型长度,比如卷积节点的权重使用长度为3比特(bit)的整型,偏置使用长度为1比特的整型。
需要说明地,目前ONNX文件结构中元素类型定义为:
enum DataType{UNDEFINED=0;FLOAT=1;UINT8=2;...INT32=6;INT64=7;STRING=8;BOOL=9;...COMPLEX128=15;...}
本申请实施例中,可以修改为元素类型为int1=0,int2=1,int3=2,int4=3..,其中int1表示1bit的索引,即只有0和1的取值,int2对应00 01 10 11或0123的索引,具体索引对应的量化表格根据AI网络的功能由协议约定,或者基站配置。可选地,第一元素类型及后续的第二元素类型可以通过这样的整型长度来进行表征。
本申请实施例中,目标AI网络的一个网络节点对应一个参数信息协议类, 一个所述参数信息协议类通过维度信息来表征对应的所述网络节点的输入、输出和网络参数中的至少一项,所述维度信息包括维度数量值和维度大小值例如,对于一个3维矩阵(2×6×14),维度数量值为3,也即维度为3,维度大小值为2×6×14=168,则维度信息包括(3,2,6,14),或者是(3,(2,6,14))。
可选地,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。例如,对于一个3维矩阵(2×6×14),维度大小值也即168,也即该维度包括168个数值,且这些数值为归一化处理后的数值。
可选地,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
本申请实施例中,目标AI网络的一个网络参数对应一个参数信息协议类,所述参数信息协议类包括维度信息,而维度信息包括维度数量值和维度大小值,通过维度数量值和维度大小值来描述参数信息协议类对应的网络参数,进而每一个网络参数也就对应有一个维度数量值和一个维度大小值。在目标AI网络包括的网络参数为多个的情况下,ONNX文件结构也就包括对应的多个参数信息协议类,而每一个参数信息协议类包括维度数量值和维度大小值,则ONNX文件结构也就相应地包括多个维度数量值和多个维度大小值;这种情况下,这些维度数量值和维度大小值可以按照目标AI网络中网络参数的预设排列顺序进行排列,例如可以是按照网络参数的预设排列顺序先排列放置所有的维度数量值,再按照网络参数的预设排列顺序排列放置所有的维度大小值。这样维度数量值采用相同长度的比特,第二端在对维度数量值解析之后也就知道后面维度大小值对应的比特长度,进而可以省略网络参数的标识,以节省传输开销。
可选地,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
其中,所述目标网络参数为目标AI网络的任一个网络参数,所述目标参数信息协议类为ONNX文件结构中用于表征该目标网络参数的参数信息协议类,所述目标张量协议类为与所述目标参数信息协议类表征的目标网络参数对应的张量协议类。例如,目标网络参数包括168个数值,则该目标网络参数对应的目标参数信息协议类中包括这168个数值中的最大数值的数值位置, 而剩下的167个数值通过该目标网络参数对应的目标张量协议类来表征,所述目标张量协议类中包括这167个数值与最大数值的比值的量化值。
本申请实施例中,所述参数信息协议类还包括用于指示网络参数中非零数值的位置的指示参数。相应地,该网络参数对应的张量协议类中量化后为0的数值,或者小于量化最小阈值的数值都可以不用包含。
需要说明地,目前的ONNX文件结构中参数信息协议类的定义为:
{optional string name=1;
optional Typeproto type=2;
optional string doc_string=3;}
本申请实施例中,可以将Typeproto修改为int,用于记录元素类型也即量化位数,没有记录的时候可以是与前一个参数信息协议类相同;或者删除type,将Typeproto保存在计算图协议类中,表示全部的参数信息协议类都是使用相同的量化位数。
本申请实施例中,所述张量协议类包括第二元素类型和数值,或者,所述张量协议类也可以只包括数值。所述张量协议类用于表征目标AI网络的网络参数,通过第二元素类型和数值来表征网络参数的名称、数值等信息。
可选地,所述第二元素类型通过量化位数的整型长度进行表征,也即第二元素类型只需要指示长度即可。
可选地,所述第二元素类型与所述参数信息协议类中的第一元素类型呈对应关系,例如第二元素类型和第一元素类型用于描述同一个网络参数,所述对应关系可以是按照目标AI网络的网络参数的索引顺序排序。
可选地,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。需要说明地,所述张量协议类中可以省略网络参数的名称和/或索引。
本申请实施例中,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。也就是说,可以通过类别的方式来存储proto,一种类型的proto对应一个协议类列表,例如ONNX文件结构包括节点协议类、参数信息协议类和张量协议类,则ONNX文件结构也就包括三个列表,分别为节点协议列列表、参数信息协议类列表和张量协议类列表。需要说明地,协议类列表中可以包括多个对应的协议类,例如节点协议类列表存储多个节点协议类,这多个协议类按照一定顺序排列。协议类列表中存储的协议类包括对应的列表指示,通过所述列表指示来指示协议类在列表中的位置。
需要说明地,对于一个AI网络,可以实现约定网络节点的读取顺序,例 如从内到外或者从上到下,例如将所有的节点协议类按照顺序排列,即每个节点的输入一定是自己之前的某个或某些节点,从而也就不需要每个节点的列表指示。
可选地,如果每个节点协议类的长度也是协议约定好的,或者基站配置的,那么可以通过偏移找到对应的节点协议类,当删除、插入、新增节点协议类的时候,可以直接找到对应的位置。同样,对于参数信息协议类和张量协议类也一样,可以按照节点协议类中记录的顺序排列成对应的列表。
可选地,所述方法还可以包括:
所述第一端向第二端发送指示信息;或者,
所述第一端向第二端发送所述ONNX文件结构,包括:
所述第一端向第二端发送携带指示信息的所述ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
也就是说,所述指示信息可以单独发送,也可以是携带在ONNX文件结构中发送。可选地,每一个协议类列表可以是单独形成一个ONNX文件结构,第二端基于所述指示信息,也就能够确定ONNX文件结构中这个协议类列表的起始位置和/或结束位置。
可选地,第二端在接收到所述ONNX文件结构之后,可以对所述ONNX文件结构增加部分内容。例如,第一端为核心网设备,第二端为基站,核心网设备发送给基站的ONNX文件结构只包括节点协议类信息,基站接收到该ONNX文件结构之后可以补充参数信息协议类信息和张量协议类信息。又如,核心网设备发送给基站的ONNX文件结构包括目标AI网络对应的全部的节点类协议信息、全部的参数信息协议类信息和部分张量协议类信息,基站根据张量协议类对应的列表的起始位置和结束位置,在相关位置补充其他的张量协议类信息。
需要说明地,所述协议类列表可以是proto的分组,例如按照proto的功能或者需求进行分组或分段。
本申请实施例中,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。例如,第一端可以预先定义一些常用功能或历史功能的网络节点,将这些网络节点作为预设网络节点,通过预定义索引来表征这些预设网络节点,第二端基于所述预定义索引能够获取对应的预设网络节点的功能、输入输出的维度等信息。通过预定义索引的设置,也就无需将所用的网络节点都通过节点协议类进行表征,这样也就能够减少ONNX文件结构中节点协议类的数量,进而以节省ONNX文件结构的传输开销。
可选地,所述网络节点包括离散傅里叶变换(Discrete Fourier Transform, DFT)节点、离散傅里叶逆变换(Inverse Discrete Fourier Transform,IDFT)节点、滤波节点等。
在所述目标AI网络发生更新的情况下,所述预设网络节点可以为所述目标AI网络中未更新的网络节点。可选地,在更新目标AI网络的时候,可以将未更新的网络节点作为预设网络节点,进而这些未更新的网络节点在ONNX文件结构中能够通过预定义索引来进行表征。可选地,可以是将未更新的网络节点合并为一个节点,通过一个预定义索引来指示,例如可以是历史节点对应的索引进行指示。
本申请实施例中,所述第一端将目标AI网络的AI网络信息转换成ONNX文件结构,包括:
所述第一端将目标AI网络的第一AI网络信息转换成第一ONNX文件结构,以及将所述目标AI网络的第二AI网络信息转换成第二ONNX文件结构,其中,所述第一ONNX文件结构包括的所述目标协议类不同于所述第二ONNX文件结构包括的所述目标协议类;
所述第一端向第二端发送所述ONNX文件结构,包括:
所述第一端向第二端发送所述第一ONNX文件结构和第二ONNX文件结构。
也就是说,第一端在生成ONNX文件结构的时候,能够基于协议类的类型生成不同的ONNX文件结构,每一个ONNX文件结构所包括的协议类不同,然后分别发送这些ONNX文件结构。例如,第一ONNX文件结构包括节点协议类,第二ONNX文件结构包括参数信息协议类,第一端向第二端分别发送这两个ONNX文件结构。
需要说明地,基于协议类的类型数量,第一端可以是生成多个ONNX文件结构,例如若目标AI网络的AI网络信息分别对应节点协议类、参数信息协议类和张量协议类,基于这三种不同的协议类类型,第一端可以是分别生成三个ONNX文件结构,这三个ONNX文件结构分别对应节点协议类、参数信息协议类和张量协议类,并分别发送这三个ONNX文件结构。
当然,第一端也可以是将所有的协议类包含在一个ONNX文件结构中。
可选地,第一ONNX文件结构和第二ONNX文件结构对应的协议类可以是有多种情况。例如,所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类;或者,所述第一ONNX文件结构包括节点协议类和参数信息协议类,所述第二ONNX文件结构包括张量协议类;或者,所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类和张量协议类。
为更好地理解,以下通过具体的实施例来对本申请实施例的技术方案进 行说明。
终端和网络侧设备使用联合的AI网络进行信道状态信息(Channel State Information,CSI)反馈,即终端通过AI网络将信道信息转换成若干比特(bit)的CSI反馈信息,并上报给基站,基站接收终端反馈的bit信息,通过基站侧的AI网络将信道信息恢复出来。
由于基站和终端的AI网络需要进行联合训练,不同的小区信道情况不同,也可能需要新的网络参数,因此当终端接入网络的时候,基站需要将终端使用的网络参数发送给终端。
CSI反馈的网络可以分成两个部分,终端编码部分和基站解码部分,终端只需要获得编码部分的网络结构。
核心网设备通过非接入服务(Non-access Stratum,NAS)信息将编码部分的网络结构发送给终端,具体信息可以是只包括NodeProto信息的ONNX文件结构,或者是一个包括NodeProto列表的文件结构。核心网设备将该ONNX文件结构发送到基站侧后,由基站转发给终端。
基站收到核心网设备的NAS信息之后,如果这个NAS信息对基站是透明的,基站直接转发NAS信息,然后将权重参数保存成只有ValueInfoProto和TensorProto的ONNX文件结构,通过无线资源控制(Radio Resource Control,RRC)信令发送给终端。
如果这个NAS信令基站可以解读,基站可以将NodeProto的ONNX文件结构和自己的包括ValueInfoProto和TensorProto的ONNX文件结构一起通过RRC信令发送给终端,或者是把二者合并为一个ONNX文件结构发送给终端。
可选的,基站可以把ValueInfoProto和TensorProto信息补充到NAS信息的ONNX文件结构中,依然按照NAS信息转发给用户。
另外,核心网设备发送的也可以是包括NodeProto、ValueInfoProto和TensorProto的ONNX文件结构,其中的ValueInfoProto和TensorProto可以是部分,也可以是全部。基站接收到NAS信息之后,如果是透明的直接转发,再通过RRC信令发送自己的ValueInfoProto和TensorProto的ONNX文件结构。如果是非透明的,基站可以将自己的ONNX文件结构与核心网设备的ONNX文件结构一起通过RRC发送给终端,或者合并后发送给终端。
请参照图3,图3是本申请实施例提供的另一种AI网络信息传输方法的流程图,如图3所示,所述方法包括以下步骤:
步骤301、第二端接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
本申请实施例中,第一端将目标AI网络的AI网络信息转换成ONNX文 件结构,例如若AI网络信息包括目标AI网络完整的网络结构和全部的网络参数,则第一端将所述网络结构和网络参数基于ONNX的结构形式表述成ONNX文件结构,例如通过节点协议类(NodeProto)和参数信息协议类(ValueInfoProto)来描述网络结构,通过张量协议类(TensorProto)来描述网络参数,等。所述第一端将AI网络信息转换成ONNX文件结构的具体实现方式可参照图2所述实施例中的描述,本实施例不再赘述。
本申请实施例中,第二端接收第一端发送的ONNX文件结构,第二端将所述ONNX文件结构转换成自身神经网络框架下的AI网络,以实现第二端对AI网络的训练和应用。这样,也就使得神经网络框架不同的两个通信设备,能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
需要说明地,所述AI网络信息包括目标AI网络的网络结构和网络参数中的至少一项。例如,所述AI网络信息包括目标AI网络完整的网络结构和全部的网络参数,或者所述AI网络信息包括目标AI网络完整的网络结构,或者仅包括目标AI网络中更新的AI网络结构,或者仅包括目标AI网络的部分网络参数,或者仅包括目标AI网络的网络参数的部分数值,等等。这样,也就使得AI网络的网络结构和网络参数可以分开发送,进而在通信过程中也就无需将包括整个网络结构和网络参数的全部AI网络一起进行传输,能够有效降低通信过程中的传输开销。
本申请实施例中,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:节点协议类(NodeProto)、参数信息协议类(ValueInfoProto)、张量协议类(TensorProto)、属性协议类(AttributeProto)。
可选地,所述ONNX文件结构包括至少一个计算图协议类(GraphProto),所述计算图协议类包括所述目标协议类。也就是说,在所述目标协议类包括上述节点协议类、参数信息协议类、张量协议类和属性协议类中的至少一项的情况下,所述目标协议类包括的内容可以写进计算图协议类中。
可选地,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
所述第二端接收第一端发送的ONNX文件结构,包括如下任意一项:
所述第二端接收第一端合并发送的至少两个所述ONNX文件结构;
所述第二端接收第一端分别发送的至少两个所述ONNX文件结构。
本申请实施例中,每一个proto都可以独立成ONNX文件结构,进而一个ONNX文件结构也就对应一种类型的proto。例如,若目标AI网络的AI网络信息包括网络结构和权重参数,所述AI网络信息基于ONNX的结构形 式生成节点协议类、参数信息协议类和张量协议类,则节点协议类对应生成一个ONNX文件结构,参数信息协议类对应生成一个ONNX文件结构,张量协议类对应生成一个ONNX文件结构,也就会得到三个ONNX文件结构。
可选地,第一端可以是向第二端分别发送这三个ONNX文件结构,例如先后发送这三个ONNX文件结构;或者,第一端也可以是将这三个ONNX文件结构进行合并,并向第二端一次发送合并后的这三个ONNX文件结构。进而,第二端能够执行对应的接收动作,对于接收到的ONNX文件结构,也就能够明确一个ONNX文件结构对应一种proto类型,更便于第二端将ONNX文件结构转换成自身神经网络框架下的AI网络信息。
可选地,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述第二端接收第一端发送的ONNX文件结构之后,所述方法还包括:
所述第二端接收所述第一端发送的所述目标AI网络的网络参数的数值。
具体地,第一端在向第二传递包含节点协议类和参数信息协议类的ONNX文件结构后,第一端还可以向第二端发送目标AI网络的网络参数的数值。可以理解地,由于第一端已经向第二端发送了包含节点协议类和参数信息协议类的ONNX文件结构,进而第二端对于网络参数的数量和开销也就是已知的。其中,所述节点协议类和参数信息协议类可以是包含在一个ONNX文件结构中,也可以是各自对应一个ONNX文件结构。
可选地,第二端接收所述第一端通过数据信道发送的所述目标AI网络的网络参数的数值。
本申请实施例中,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参数、输入和输出中的至少一项。
可选地,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输出的所述第二数字索引。
可选地,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项;
其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
可选地,一个所述参数信息协议类用于表征所述目标AI网络中的一个网络节点对应的输入、输出以及网络参数中至少一个,所述参数信息协议类包括第一元素类型和维度信息。
可选地,所述第一元素类型通过量化位数的整型长度进行表征。
可选地,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。
可选地,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
可选地,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。
可选地,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
可选地,所述参数信息协议类包括用于指示非零数值的位置的指示参数。
可选地,所述张量协议类包括第二元素类型和数值,或者,所述张量协议类包括数值。
可选地,所述第二元素类型通过量化位数的整型长度进行表征。
可选地,所述第二元素类型与所述第一元素类型呈对应关系。
可选地,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。
可选地,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
可选地,所述方法还包括:
所述第二端接收所述第一端发送的指示信息;或者,
所述第二端接收第一端发送的ONNX文件结构,包括:
所述第二端接收第一端发送的携带有指示信息的ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
也就是说,所述指示信息可以单独发送,也可以是携带在ONNX文件结构中发送。可选地,每一个协议类列表可以是单独形成一个ONNX文件结构,第二端基于所述指示信息,也就能够确定ONNX文件结构中这个协议类列表的起始位置和/或结束位置。
可选地,第二端在接收到所述ONNX文件结构之后,可以对所述ONNX文件结构增加部分内容。例如,第一端为核心网设备,第二端为基站,核心网设备发送给基站的ONNX文件结构只包括节点协议类信息,基站接收到该ONNX文件结构之后可以补充参数信息协议类信息和张量协议类信息。又如,核心网设备发送给基站的ONNX文件结构包括目标AI网络对应的全部的节点类协议信息、全部的参数信息协议类信息和部分张量协议类信息,基站根据张量协议类对应的列表的起始位置和结束位置,在相关位置补充其他的张量协议类信息。
本申请实施例中,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。
可选地,在所述目标AI网络发生更新的情况下,所述预设网络节点为所述目标AI网络中未更新的网络节点。
需要说明地,本申请实施例所提供的AI网络信息传输方法应用于第二端,与上述图2实施例中所提供的应用于第一端的AI网络信息传输方法相对应,本申请实施例中相关步骤的具体实现过程以及ONNX文件结构涉及的相关概念可以参照上述图2所述方法实施例中的描述,为避免重复,此处不再赘述。
本申请实施例中,第二端接收第一端发送的ONNX文件结构,进而第二端能够将所述ONNX文件结构转换成自身神经网络框架下的AI网络,以实现第二端对AI网络的训练和应用。这样,也就使得神经网络框架不同的两个通信设备,能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
本申请实施例提供的AI网络信息传输方法,执行主体可以为AI网络信息传输装置。本申请实施例中以AI网络信息传输装置执行AI网络信息传输方法为例,说明本申请实施例提供的AI网络信息传输装置。
请参照图4,图4是本申请实施例提供的一种AI网络信息传输装置的结构图,如图4所示,AI网络信息传输装置400包括:
转换模块401,用于将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
发送模块402,用于向第二端发送所述ONNX文件结构。
可选地,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:节点协议类;
参数信息协议类;
张量协议类;
属性协议类。
可选地,所述ONNX文件结构包括至少一个计算图协议类,所述计算图 协议类包括所述目标协议类。
可选地,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
所述发送模块402还用于执行如下任意一项:
向第二端合并发送至少两个所述ONNX文件结构;
向第二端分别发送至少两个所述ONNX文件结构。
可选地,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述发送模块402还用于:
向第二端发送所述目标AI网络的网络参数的数值。
可选地,所述发送模块402还用于:
通过数据信道向第二端发送所述目标AI网络的网络参数的数值。
可选地,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参数、输入和输出中的至少一项。
可选地,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输出的所述第二数字索引。
可选地,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项;
其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
可选地,一个所述参数信息协议类用于表征所述目标AI网络中的一个网络节点对应的输入、输出以及网络参数中至少一个,所述参数信息协议类包括第一元素类型和维度信息。
可选地,所述第一元素类型通过量化位数的整型长度进行表征。
可选地,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。
可选地,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
可选地,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。
可选地,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
可选地,所述参数信息协议类包括用于指示非零数值的位置的指示参数。
可选地,所述张量协议类包括第二元素类型和数值,或者,所述张量协议类包括数值。
可选地,所述第二元素类型通过量化位数的整型长度进行表征。
可选地,所述第二元素类型与所述第一元素类型呈对应关系。
可选地,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。
可选地,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
可选地,所述发送模块还用于:
向所述第二端发送指示信息;或者,
向第二端发送携带指示信息的所述ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
可选地,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。
可选地,在所述目标AI网络发生更新的情况下,所述预设网络节点为所述目标AI网络中未更新的网络节点。
可选地,所述转换模块401还用于:
将目标AI网络的第一AI网络信息转换成第一ONNX文件结构,以及将所述目标AI网络的第二AI网络信息转换成第二ONNX文件结构,其中,所述第一ONNX文件结构包括的所述目标协议类不同于所述第二ONNX文件结构包括的所述目标协议类;
所述发送模块402还用于:
向第二端发送所述第一ONNX文件结构和第二ONNX文件结构;
可选地,所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类;或者,
所述第一ONNX文件结构包括节点协议类和参数信息协议类,所述第二 ONNX文件结构包括张量协议类;或者,
所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类和张量协议类。
本申请实施例中,所述装置将适用于自身神经网络框架的目标AI网络的AI网络信息转换成ONNX文件结构,然后将所述ONNX文件结构发送给第二端,进而第二端能够将ONNX文件结构转换成自身神经网络框架下的AI网络。这样,也就使得神经网络框架不同的两个通信设备,能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
本申请实施例中的AI网络信息传输装置400可以是电子设备,例如具有操作系统的电子设备,也可以是电子设备中的部件,例如集成电路或芯片。该电子设备可以是终端,也可以为除终端之外的其他设备。示例性的,终端可以包括但不限于上述所列举的终端11的类型,其他设备可以为服务器、网络附属存储器(Network Attached Storage,NAS)等,本申请实施例不作具体限定。
本申请实施例提供的AI网络信息传输装置400能够实现图2所述方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
请参照图5,图5是本申请实施例提供的另一种AI网络信息传输装置的结构图,如图5所示,所述AI网络信息传输装置500包括:
接收模块501,用于接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
可选地,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:节点协议类;
参数信息协议类;
张量协议类;
属性协议类。
可选地,所述ONNX文件结构包括至少一个计算图协议类,所述计算图协议类包括所述目标协议类。
可选地,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
所述接收模块501还用于执行如下任意一项:
接收第一端合并发送的至少两个所述ONNX文件结构;
接收第一端分别发送的至少两个所述ONNX文件结构。
可选地,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述接收模块501还用于:
接收所述第一端发送的所述目标AI网络的网络参数的数值。
可选地,所述接收模块501还用于:
通过数据信道接收所述第一端发送的所述目标AI网络的网络参数的数值。
可选地,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参数、输入和输出中的至少一项。
可选地,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输出的所述第二数字索引。
可选地,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项;
其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
可选地,一个所述参数信息协议类用于表征所述目标AI网络中的一个网络节点对应的输入、输出以及网络参数中至少一个,所述参数信息协议类包括第一元素类型和维度信息。
可选地,所述第一元素类型通过量化位数的整型长度进行表征。
可选地,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。
可选地,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
可选地,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。
可选地,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
可选地,所述参数信息协议类包括用于指示非零数值的位置的指示参数。
可选地,所述张量协议类包括第二元素类型和数值,或者,所述张量协 议类包括数值。
可选地,所述第二元素类型通过量化位数的整型长度进行表征。
可选地,所述第二元素类型与所述第一元素类型呈对应关系。
可选地,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。
可选地,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
可选地,所述接收模块501还用于:
接收所述第一端发送的指示信息;或者,
接收第一端发送的携带有指示信息的ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
可选地,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。
可选地,在所述目标AI网络发生更新的情况下,所述预设网络节点为所述目标AI网络中未更新的网络节点。
本申请实施例中,所述装置能够接收第一端发送的ONNX文件结构,进而所述装置能够将所述ONNX文件结构转换成自身神经网络框架下的AI网络,以实现对AI网络的训练和应用。这样,也就使得即使神经网络框架不同的两个通信设备,也能够基于ONNX文件结构来传输AI网络信息,避免通信设备之间AI网络信息的传输受阻。
本申请实施例中的AI网络信息传输装置500可以是电子设备,例如具有操作系统的电子设备,也可以是电子设备中的部件,例如集成电路或芯片。该电子设备可以是终端,也可以为除终端之外的其他设备。示例性的,终端可以包括但不限于上述所列举的终端11的类型,其他设备可以为服务器、网络附属存储器(Network Attached Storage,NAS)等,本申请实施例不作具体限定。
本申请实施例提供的AI网络信息传输装置500能够实现图3所述方法实施例实现的各个过程,并达到相同的技术效果,为避免重复,这里不再赘述。
可选的,如图6所示,本申请实施例还提供一种通信设备600,包括处理器601和存储器602,存储器602上存储有可在所述处理器601上运行的程序或指令,该程序或指令被处理器601执行时实现上述图2或图3所述的AI网络信息传输方法实施例的各个步骤,且能达到相同的技术效果。为避免重复,这里不再赘述。
本申请实施例还提供一种终端,上述图2或图3方法实施例的各个实施过程和实现方式均可适用于该终端实施例中,且能达到相同的技术效果。具体地,图7为实现本申请实施例的一种终端的硬件结构示意图。
该终端700包括但不限于:射频单元701、网络模块702、音频输出单元703、输入单元704、传感器705、显示单元706、用户输入单元707、接口单元708、存储器709以及处理器710等中的至少部分部件。
本领域技术人员可以理解,终端700还可以包括给各个部件供电的电源(比如电池),电源可以通过电源管理系统与处理器710逻辑相连,从而通过电源管理系统实现管理充电、放电、以及功耗管理等功能。图7中示出的终端结构并不构成对终端的限定,终端可以包括比图示更多或更少的部件,或者组合某些部件,或者不同的部件布置,在此不再赘述。
应理解的是,本申请实施例中,输入单元704可以包括图形处理单元(Graphics Processing Unit,GPU)7041和麦克风7042,图形处理器7041对在视频捕获模式或图像捕获模式中由图像捕获装置(如摄像头)获得的静态图片或视频的图像数据进行处理。显示单元706可包括显示面板7061,可以采用液晶显示器、有机发光二极管等形式来配置显示面板7061。用户输入单元707包括触控面板7071以及其他输入设备7072中的至少一种。触控面板7071,也称为触摸屏。触控面板7071可包括触摸检测装置和触摸控制器两个部分。其他输入设备7072可以包括但不限于物理键盘、功能键(比如音量控制按键、开关按键等)、轨迹球、鼠标、操作杆,在此不再赘述。
本申请实施例中,射频单元701接收来自网络侧设备的下行数据后,可以传输给处理器710进行处理;另外,射频单元701可以向网络侧设备发送上行数据。通常,射频单元701包括但不限于天线、放大器、收发信机、耦合器、低噪声放大器、双工器等。
存储器709可用于存储软件程序或指令以及各种数据。存储器709可主要包括存储程序或指令的第一存储区和存储数据的第二存储区,其中,第一存储区可存储操作系统、至少一个功能所需的应用程序或指令(比如声音播放功能、图像播放功能等)等。此外,存储器709可以包括易失性存储器或非易失性存储器,或者,存储器709可以包括易失性和非易失性存储器两者。其中,非易失性存储器可以是只读存储器(Read-Only Memory,ROM)、可编程只读存储器(Programmable ROM,PROM)、可擦除可编程只读存储器(Erasable PROM,EPROM)、电可擦除可编程只读存储器(Electrically EPROM,EEPROM)或闪存。易失性存储器可以是随机存取存储器(Random Access Memory,RAM),静态随机存取存储器(Static RAM,SRAM)、动态随机存取存储器(Dynamic RAM,DRAM)、同步动态随机存取存储器(Synchronous  DRAM,SDRAM)、双倍数据速率同步动态随机存取存储器(Double Data Rate SDRAM,DDRSDRAM)、增强型同步动态随机存取存储器(Enhanced SDRAM,ESDRAM)、同步连接动态随机存取存储器(Synch link DRAM,SLDRAM)和直接内存总线随机存取存储器(Direct Rambus RAM,DRRAM)。本申请实施例中的存储器709包括但不限于这些和任意其它适合类型的存储器。
处理器710可包括一个或多个处理单元;可选的,处理器710集成应用处理器和调制解调处理器,其中,应用处理器主要处理涉及操作系统、用户界面和应用程序等的操作,调制解调处理器主要处理无线通信信号,如基带处理器。可以理解的是,上述调制解调处理器也可以不集成到处理器710中。
在本申请实施例的一种实施方式中,终端700为第一端。其中,处理器710,用于将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
射频单元701,用于向第二端发送所述ONNX文件结构。
可选地,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:
节点协议类;
参数信息协议类;
张量协议类;
属性协议类。
可选地,所述ONNX文件结构包括至少一个计算图协议类,所述计算图协议类包括所述目标协议类。
可选地,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
所述射频单元701,用于执行如下任意一项:
向第二端合并发送至少两个所述ONNX文件结构;
向第二端分别发送至少两个所述ONNX文件结构。
可选地,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述射频单元701,用于:
向第二端发送所述目标AI网络的网络参数的数值。
可选地,所述射频单元701,用于:
通过数据信道向第二端发送所述目标AI网络的网络参数的数值。
可选地,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参 数、输入和输出中的至少一项。
可选地,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输出的所述第二数字索引。
可选地,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项;
其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
可选地,一个所述参数信息协议类用于表征所述目标AI网络中的一个网络节点对应的输入、输出以及网络参数中至少一个,所述参数信息协议类包括第一元素类型和维度信息。
可选地,所述第一元素类型通过量化位数的整型长度进行表征。
可选地,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。
可选地,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
可选地,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。
可选地,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
可选地,所述参数信息协议类包括用于指示非零数值的位置的指示参数。
可选地,所述张量协议类包括第二元素类型和数值,或者,所述张量协议类包括数值。
可选地,所述第二元素类型通过量化位数的整型长度进行表征。
可选地,所述第二元素类型与所述第一元素类型呈对应关系。
可选地,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。
可选地,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、 张量协议类和属性协议类。
可选地,所述射频单元701还用于:
向所述第二端发送指示信息;或者,
向第二端发送携带指示信息的所述ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
可选地,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。
可选地,在所述目标AI网络发生更新的情况下,所述预设网络节点为所述目标AI网络中未更新的网络节点。
可选地,所述处理器710还用于:
将目标AI网络的第一AI网络信息转换成第一ONNX文件结构,以及将所述目标AI网络的第二AI网络信息转换成第二ONNX文件结构,其中,所述第一ONNX文件结构包括的所述目标协议类不同于所述第二ONNX文件结构包括的所述目标协议类;
所述射频单元701还用于:
向第二端发送所述第一ONNX文件结构和第二ONNX文件结构;
可选地,所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类;或者,
所述第一ONNX文件结构包括节点协议类和参数信息协议类,所述第二ONNX文件结构包括张量协议类;或者,
所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类和张量协议类。
在本申请实施例的另一种实施方式中,终端700可以作为第二端。其中,射频单元701,用于接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
可选地,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:
节点协议类;
参数信息协议类;
张量协议类;
属性协议类。
可选地,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
所述射频单元701用于执行如下任意一项:
接收第一端合并发送的至少两个所述ONNX文件结构;
接收第一端分别发送的至少两个所述ONNX文件结构。
可选地,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述射频单元701还用于:
接收所述第一端发送的所述目标AI网络的网络参数的数值。
可选地,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
可选地,所述射频单元701还用于:
接收所述第一端发送的指示信息;或者,
接收第一端发送的携带有指示信息的ONNX文件结构;
其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
本申请实施例提供的终端700能够作为第一端或第二端执行上述图2或图3所述的AI网络信息传输方法,并能达到相同的技术效果,此处不再赘述。
本申请实施例还提供一种网络侧设备,上述图2和图3所述方法实施例的各个实施过程和实现方式均可适用于该网络侧设备实施例中,且能达到相同的技术效果。
具体地,本申请实施例还提供了一种网络侧设备。如图8所示,该网络侧设备800包括:天线81、射频装置82、基带装置83、处理器84和存储器85。天线81与射频装置82连接。在上行方向上,射频装置82通过天线81接收信息,将接收的信息发送给基带装置83进行处理。在下行方向上,基带装置83对要发送的信息进行处理,并发送给射频装置82,射频装置82对收到的信息进行处理后经过天线81发送出去。
以上实施例中网络侧设备执行的方法可以在基带装置83中实现,该基带装置83包括基带处理器。
基带装置83例如可以包括至少一个基带板,该基带板上设置有多个芯片,如图8所示,其中一个芯片例如为基带处理器,通过总线接口与存储器85连接,以调用存储器85中的程序,执行以上方法实施例中所示的网络设备操作。
该网络侧设备还可以包括网络接口86,该接口例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本发明实施例的网络侧设备800还包括:存储在存储器85上并可在处理器84上运行的指令或程序,处理器84调用存储器85中的指令或程序执行图4或图5所示各模块执行的方法,并达到相同的技术效果,为避免 重复,故不在此赘述。
具体地,本申请实施例还提供了另一种网络侧设备。如图9所示,该网络侧设备900包括:处理器901、网络接口902和存储器903。其中,网络接口902例如为通用公共无线接口(common public radio interface,CPRI)。
具体地,本发明实施例的网络侧设备900还包括:存储在存储器903上并可在处理器901上运行的指令或程序,处理器901调用存储器903中的指令或程序执行图4或图5所示各模块执行的方法,并达到相同的技术效果,为避免重复,故不在此赘述。
本申请实施例还提供一种可读存储介质,所述可读存储介质上存储有程序或指令,该程序或指令被处理器执行时实现上述图2或图3所述方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
其中,所述处理器为上述实施例中所述的终端中的处理器。所述可读存储介质,包括计算机可读存储介质,如计算机只读存储器ROM、随机存取存储器RAM、磁碟或者光盘等。
本申请实施例另提供了一种芯片,所述芯片包括处理器和通信接口,所述通信接口和所述处理器耦合,所述处理器用于运行程序或指令,实现上述图2或图3所述方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
应理解,本申请实施例提到的芯片还可以称为系统级芯片,系统芯片,芯片系统或片上系统芯片等。
本申请实施例另提供了一种计算机程序/程序产品,所述计算机程序/程序产品被存储在非易失的存储介质中,所述计算机程序/程序产品被至少一个处理器执行以实现上述图2或图3所述方法实施例的各个过程,且能达到相同的技术效果,为避免重复,这里不再赘述。
本申请实施例还提供了一种通信系统,包括:终端及网络侧设备,所述终端可用于执行如上述图2所述方法的步骤,所述网络侧设备可用于执行如上图3所述方法的步骤,或者,所述终端可用于执行如上述图3所述方法的步骤,所述网络侧设备可用于执行如上图2所述方法的步骤。
需要说明的是,在本文中,术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含,从而使得包括一系列要素的过程、方法、物品或者装置不仅包括那些要素,而且还包括没有明确列出的其他要素,或者是还包括为这种过程、方法、物品或者装置所固有的要素。在没有更多限制的情况下,由语句“包括一个……”限定的要素,并不排除在包括该要素的过程、方法、物品或者装置中还存在另外的相同要素。此外,需要指出的是,本申请实施方式中的方法和装置的范围不限按示出或讨论的顺序来执行功能,还 可包括根据所涉及的功能按基本同时的方式或按相反的顺序来执行功能,例如,可以按不同于所描述的次序来执行所描述的方法,并且还可以添加、省去、或组合各种步骤。另外,参照某些示例所描述的特征可在其他示例中被组合。
通过以上的实施方式的描述,本领域的技术人员可以清楚地了解到上述实施例方法可借助软件加必需的通用硬件平台的方式来实现,当然也可以通过硬件,但很多情况下前者是更佳的实施方式。基于这样的理解,本申请的技术方案本质上或者说对现有技术做出贡献的部分可以以计算机软件产品的形式体现出来,该计算机软件产品存储在一个存储介质(如ROM/RAM、磁碟、光盘)中,包括若干指令用以使得一台终端(可以是手机,计算机,服务器,空调器,或者网络设备等)执行本申请各个实施例所述的方法。
上面结合附图对本申请的实施例进行了描述,但是本申请并不局限于上述的具体实施方式,上述的具体实施方式仅仅是示意性的,而不是限制性的,本领域的普通技术人员在本申请的启示下,在不脱离本申请宗旨和权利要求所保护的范围情况下,还可做出很多形式,均属于本申请的保护之内。

Claims (37)

  1. 一种人工智能AI网络信息传输方法,包括:
    第一端将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
    所述第一端向第二端发送所述ONNX文件结构。
  2. 根据权利要求1所述的方法,其中,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:
    节点协议类;
    参数信息协议类;
    张量协议类;
    属性协议类。
  3. 根据权利要求2所述的方法,其中,所述ONNX文件结构包括至少一个计算图协议类,所述计算图协议类包括所述目标协议类。
  4. 根据权利要求2所述的方法,其中,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
    所述第一端向第二端发送所述ONNX文件结构,包括如下任意一项:
    所述第一端向第二端合并发送至少两个所述ONNX文件结构;
    所述第一端向第二端分别发送至少两个所述ONNX文件结构。
  5. 根据权利要求2-4中任一项所述的方法,其中,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述第一端向第二端发送所述ONNX文件结构之后,所述方法还包括:
    所述第一端向第二端发送所述目标AI网络的网络参数的数值。
  6. 根据权利要求5所述的方法,其中,所述第一端向第二端发送所述目标AI网络的网络参数的数值,包括:
    所述第一端通过数据信道向第二端发送所述目标AI网络的网络参数的数值。
  7. 根据权利要求2-4中任一项所述的方法,其中,所述ONNX文件结构包括整型类型的第一数字索引和第二数字索引,所述第一数字索引用于标识所述目标协议类,一个所述目标协议类对应一个第一数字索引,所述第二数字索引用于标识所述目标AI网络的网络参数、输入和输出中的至少一项。
  8. 根据权利要求7所述的方法,其中,在所述ONNX文件结构包括节点协议类的情况下,所述节点协议类包括用于表征所述目标AI网络的输入和输 出的所述第二数字索引。
  9. 根据权利要求8所述的方法,其中,在所述目标AI网络的输入和输出的数量为多个且连续的情况下,所述节点协议类包括目标数字索引,以及包括所述目标AI网络的输入的数量和输出的数量中的至少一项;
    其中,所述目标数字索引为所述第二数字索引中用于标识所述目标AI网络的多个输入和输出中的第一个输入和第一个输出的数字索引。
  10. 根据权利要求2-4中任一项所述的方法,其中,一个所述参数信息协议类用于表征所述目标AI网络中的一个网络节点对应的输入、输出以及网络参数中至少一个,所述参数信息协议类包括第一元素类型和维度信息。
  11. 根据权利要求10所述的方法,其中,所述第一元素类型通过量化位数的整型长度进行表征。
  12. 根据权利要求10所述的方法,其中,所述第一元素类型与所述目标AI网络的网络节点或网络参数对应。
  13. 根据权利要求10所述的方法,其中,所述参数信息协议类中的维度信息包括对应的维度数量值和维度大小值,所述参数信息协议类中的维度数量值基于预设顺序排列,所述参数信息协议类中的维度大小值基于所述预设顺序排列,所述预设顺序为所述目标AI网络的网络参数的预设排列顺序。
  14. 根据权利要求13所述的方法,其中,所述维度大小值为对应的维度包括的数值数量,所述数值为归一化处理后的数值。
  15. 根据权利要求10所述的方法,其中,在目标参数信息协议类用于表征目标网络参数的情况下,所述目标参数信息协议类包括最大数值的数值位置,所述最大数值为所述目标网络参数的数值中的最大值,所述目标网络参数对应的目标张量协议类中包括所述目标网络参数中每一个剩余数值与所述最大数值的比值的量化值,所述剩余数值为所述目标参数信息协议类对应的维度中除所述最大数值以外的其他数值。
  16. 根据权利要求10所述的方法,其中,所述参数信息协议类包括用于指示非零数值的位置的指示参数。
  17. 根据权利要求10所述的方法,其中,所述张量协议类包括第二元素类型和数值,或者,所述张量协议类包括数值。
  18. 根据权利要求17所述的方法,其中,所述第二元素类型通过量化位数的整型长度进行表征。
  19. 根据权利要求17所述的方法,其中,所述第二元素类型与所述第一元素类型呈对应关系。
  20. 根据权利要求17所述的方法,其中,所述张量协议类中第二元素类型的排列顺序与所述参数信息协议类中第一元素类型的排列顺序相同。
  21. 根据权利要求2-4中任一项所述的方法,其中,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
  22. 根据权利要求21所述的方法,其中,所述方法还包括:
    所述第一端向所述第二端发送指示信息;或者,
    所述第一端向第二端发送所述ONNX文件结构,包括:
    所述第一端向第二端发送携带指示信息的所述ONNX文件结构;
    其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
  23. 根据权利要求1-4中任一项所述的方法,其中,所述ONNX文件结构包括预定义索引,所述预定义索引用于表征预设网络节点。
  24. 根据权利要求23所述的方法,其中,在所述目标AI网络发生更新的情况下,所述预设网络节点为所述目标AI网络中未更新的网络节点。
  25. 根据权利要求2-4中任一项所述的方法,其中,所述第一端将目标AI网络的AI网络信息转换成ONNX文件结构,包括:
    所述第一端将目标AI网络的第一AI网络信息转换成第一ONNX文件结构,以及将所述目标AI网络的第二AI网络信息转换成第二ONNX文件结构,其中,所述第一ONNX文件结构包括的所述目标协议类不同于所述第二ONNX文件结构包括的所述目标协议类;
    所述第一端向第二端发送所述ONNX文件结构,包括:
    所述第一端向第二端发送所述第一ONNX文件结构和第二ONNX文件结构。
  26. 根据权利要求25所述的方法,其中,所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类;或者,
    所述第一ONNX文件结构包括节点协议类和参数信息协议类,所述第二ONNX文件结构包括张量协议类;或者,
    所述第一ONNX文件结构包括节点协议类,所述第二ONNX文件结构包括参数信息协议类和张量协议类。
  27. 一种AI网络信息传输方法,包括:
    第二端接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
  28. 根据权利要求27所述的方法,其中,所述ONNX文件结构包括目标协议类,所述目标协议类包括如下至少一项:
    节点协议类;
    参数信息协议类;
    张量协议类;
    属性协议类。
  29. 根据权利要求28所述的方法,其中,所述ONNX文件结构包括至少一个计算图协议类,所述计算图协议类包括所述目标协议类。
  30. 根据权利要求28所述的方法,其中,所述目标协议类包括所述节点协议类、参数信息协议类、张量协议类和属性协议类中的任意一项,在所述ONNX文件结构的数量为至少两个的情况下,一个所述ONNX文件结构对应一种目标协议类;
    所述第二端接收第一端发送的ONNX文件结构,包括如下任意一项:
    所述第二端接收第一端合并发送的至少两个所述ONNX文件结构;
    所述第二端接收第一端分别发送的至少两个所述ONNX文件结构。
  31. 根据权利要求28-30中任一项所述的方法,其中,在所述ONNX文件结构包括节点协议类和参数信息协议类中的至少一项的情况下,所述第二端接收第一端发送的ONNX文件结构之后,所述方法还包括:
    所述第二端接收所述第一端发送的所述目标AI网络的网络参数的数值。
  32. 根据权利要求28-30中任一项所述的方法,其中,所述ONNX文件结构包括协议类列表,一个所述协议类列表对应一种协议类类型,所述协议类类型包括所述节点协议类、参数信息协议类、张量协议类和属性协议类。
  33. 根据权利要求32所述的方法,其中,所述方法还包括:
    所述第二端接收所述第一端发送的指示信息;或者,
    所述第二端接收第一端发送的ONNX文件结构,包括:
    所述第二端接收第一端发送的携带有指示信息的ONNX文件结构;
    其中,所述指示信息用于指示所述ONNX文件结构中所述协议类列表的起始位置和结束位置中的至少一项。
  34. 一种AI网络信息传输装置,包括:
    转换模块,用于将目标AI网络的AI网络信息转换成开放式神经网络交换ONNX文件结构;
    发送模块,用于向第二端发送所述ONNX文件结构。
  35. 一种AI网络信息传输装置,包括:
    接收模块,用于接收第一端发送的ONNX文件结构,所述ONNX文件结构为所述第一端将目标AI网络的AI网络信息转换得到。
  36. 一种通信设备,包括处理器和存储器,所述存储器存储可在所述处理器上运行的程序或指令,其中,所述程序或指令被所述处理器执行时实现如权利要求1-26中任一项所述的AI网络信息传输方法的步骤,或者实现如权利要求27-33中任一项所述的AI网络信息传输方法的步骤。
  37. 一种可读存储介质,所述可读存储介质上存储程序或指令,其中,所述程序或指令被处理器执行时实现如权利要求1-26中任一项所述的AI网络信息传输方法的步骤,或者实现如权利要求27-33中任一项所述的AI网络信息传输方法的步骤。
PCT/CN2022/143640 2021-12-31 2022-12-29 Ai网络信息传输方法、装置及通信设备 WO2023125869A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111666991.3 2021-12-31
CN202111666991.3A CN116418797A (zh) 2021-12-31 2021-12-31 Ai网络信息传输方法、装置及通信设备

Publications (1)

Publication Number Publication Date
WO2023125869A1 true WO2023125869A1 (zh) 2023-07-06

Family

ID=86998077

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/143640 WO2023125869A1 (zh) 2021-12-31 2022-12-29 Ai网络信息传输方法、装置及通信设备

Country Status (2)

Country Link
CN (1) CN116418797A (zh)
WO (1) WO2023125869A1 (zh)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111768004A (zh) * 2020-06-10 2020-10-13 中国人民解放军军事科学院国防科技创新研究院 一种基于智能计算框架的模型自适应方法及系统
WO2021029797A1 (en) * 2019-08-15 2021-02-18 Telefonaktiebolaget Lm Ericsson (Publ) Network nodes and methods for handling machine learning models in a communications network
CN112541577A (zh) * 2020-12-16 2021-03-23 上海商汤智能科技有限公司 神经网络生成方法及装置、电子设备和存储介质

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021029797A1 (en) * 2019-08-15 2021-02-18 Telefonaktiebolaget Lm Ericsson (Publ) Network nodes and methods for handling machine learning models in a communications network
CN111768004A (zh) * 2020-06-10 2020-10-13 中国人民解放军军事科学院国防科技创新研究院 一种基于智能计算框架的模型自适应方法及系统
CN112541577A (zh) * 2020-12-16 2021-03-23 上海商汤智能科技有限公司 神经网络生成方法及装置、电子设备和存储介质

Also Published As

Publication number Publication date
CN116418797A (zh) 2023-07-11

Similar Documents

Publication Publication Date Title
RU2689203C2 (ru) Гибкая схема для настройки языковой модели
CN103702297B (zh) 短信增强方法、装置及系统
US20170293681A1 (en) Techniques to respond to user requests using natural-language machine learning based on example conversations
US20120078595A1 (en) Method and apparatus for ontology matching
CN107273528A (zh) 一种前后端分离方法、移动终端以及计算机可读存储介质
CN108880867A (zh) 一种网络设备数据采集方法及系统
CN105701122A (zh) 一种日志收集方法、装置及系统
CN110795528A (zh) 一种数据查询方法、装置、电子设备及存储介质
CN109543195A (zh) 一种文本翻译的方法、信息处理的方法以及装置
US20170249934A1 (en) Electronic device and method for operating the same
CN102939579A (zh) 绑定用户接口元素和粒度反映处理的方法和装置
US10599402B2 (en) Techniques to configure a web-based application for bot configuration
CN110069715A (zh) 一种信息推荐模型训练的方法、信息推荐的方法及装置
CN105468793B (zh) 一种仿真模型数据的自动化管理方法
CN110069458A (zh) 一种删除安卓系统中无效资源的方法、设备及存储介质
WO2023125869A1 (zh) Ai网络信息传输方法、装置及通信设备
CN110019648B (zh) 一种训练数据的方法、装置及存储介质
CN104967598B (zh) 一种用户多媒体权限资讯的获取方法和装置
KR20210065629A (ko) 챗봇 채널연계 통합을 위한 챗봇 통합 에이전트 플랫폼 시스템 및 그 서비스 방법
CN109190064A (zh) 清单记录方法、终端及计算机可读存储介质
CN115114927A (zh) 一种模型训练方法和相关装置
WO2023125934A1 (zh) Ai网络信息传输方法、装置及通信设备
WO2023098535A1 (zh) 信息交互方法、装置及通信设备
WO2023125594A1 (zh) Ai模型传输方法、装置、设备及存储介质
WO2023088268A1 (zh) Ai信息的传输方法和设备

Legal Events

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

Ref document number: 22915118

Country of ref document: EP

Kind code of ref document: A1