Model parameter transmission method and apparatus, and communication device

Abstract

The present application belongs to the technical field of communications. Disclosed are a model parameter transmission method and apparatus, and a communication device. The model parameter transmission method of the embodiments of the present application comprises: a first communication device determining, on the basis of a target artificial intelligence (AI) model, a model parameter to be sent; and the first communication device sending the model parameter to a second communication device.

Description

Model parameter transmission method and device and communication equipment

[0001] Cross-reference to Related Applications

[0002] This application claims priority to Chinese Patent Application No. 202410984664.X, filed on July 22, 2024, the contents of which are incorporated herein by reference in its entirety. TECHNICAL FIELD

[0003] The present application belongs to the field of communication technology, and specifically relates to a model parameter transmission method, device and communication equipment. BACKGROUND

[0004] Artificial intelligence (AI) is currently widely used in various fields. Integrating artificial intelligence into wireless communication networks significantly improves technical indicators such as throughput, latency, and user capacity, which is an important task for future wireless communication networks. AI modules have various implementation methods, such as neural networks, decision trees, support vector machines, and Bayesian classifiers.

[0005] How to implement the transmission of AI models and ensure that AI models can be used between different devices is a problem to be solved. SUMMARY

[0006] Embodiments of the present application provide a model parameter transmission method, device and communication equipment, which can implement the transmission of models between different communication devices.

[0007] In a first aspect, a model parameter transmission method is provided, comprising:

[0008] The first communication device determines the model parameters to be sent according to a target artificial intelligence (AI) model;

[0009] The first communication device sends the model parameters to a second communication device.

[0010] In a second aspect, a model parameter transmission method is provided, comprising:

[0011] The second communication device receives the model parameters of a target artificial intelligence (AI) model sent by a first communication device;

[0012] The second communication device applies the target AI model according to the model parameters.

[0013] In a third aspect, a model parameter transmission device is provided, applied to a first communication device, comprising:

[0014] A first processing module is configured to determine the model parameters to be sent according to a target artificial intelligence (AI) model;

[0015] The first sending module is configured to send the model parameters to the second communication device.

[0016] In a fourth aspect, a model parameter transmission apparatus is provided, which is applied to a second communication device and includes:

[0017] The first receiving module is configured to receive model parameters of a target artificial intelligence (AI) model sent by a first communication device.

[0018] The second processing module is configured to apply the target AI model according to the model parameters.

[0019] In a fifth aspect, a model parameter transmission apparatus is provided, which is configured to perform the steps of the method according to the first aspect or implement the steps of the method according to the second aspect.

[0020] In a sixth aspect, a communication device is provided, which is a first communication device and includes a processor and a memory, the memory storing programs or instructions executable on the processor, and the programs or instructions, when executed by the processor, implement the steps of the method according to the first aspect.

[0021] In a seventh aspect, a communication device is provided, which is a first communication device and includes a processor and a communication interface, wherein the processor is configured to determine model parameters to be sent according to a target artificial intelligence (AI) model, and the communication interface is configured to send the model parameters to a second communication device.

[0022] In an eighth aspect, a communication device is provided, which is a second communication device and includes a processor and a memory, the memory storing programs or instructions executable on the processor, and the programs or instructions, when executed by the processor, implement the steps of the method according to the second aspect.

[0023] In a ninth aspect, a communication device is provided, which is a second communication device and includes a processor and a communication interface, wherein the communication interface is configured to receive model parameters of a target artificial intelligence (AI) model sent by a first communication device, and the processor is configured to apply the target AI model according to the model parameters.

[0024] In a tenth aspect, a readable storage medium is provided, which stores programs or instructions, and the programs or instructions, when executed by a processor, implement the steps of the method according to the first aspect or implement the steps of the method according to the second aspect.

[0025] In an eleventh aspect, a wireless communication system is provided, comprising a first communication device configured to perform the steps of the method according to the first aspect, and a second communication device configured to perform the steps of the method according to the second aspect.

[0026] In a twelfth aspect, a chip is provided, comprising a processor and a communication interface coupled to the processor, the processor configured to execute a program or an instruction to implement the method according to the first aspect or the method according to the second aspect.

[0027] In a thirteenth aspect, a computer program / program product is provided, stored in a storage medium, and executed by at least one processor to implement the steps of the model parameter transmission method according to the first aspect or the second aspect.

[0028] In the embodiments of the present application, the model parameters to be transmitted are determined according to the target AI model, and the model parameters are transmitted to the second communication device, so as to realize the transmission of the AI model between different communication devices and guarantee the reliability of model application. BRIEF DESCRIPTION OF DRAWINGS

[0029] FIG. 1 is a block diagram of a wireless communication system to which embodiments of the present application can be applied;

[0030] FIG. 2 is a structural diagram of a neuron;

[0031] FIG. 3 is a structural diagram of a layer;

[0032] FIG. 4 is a flowchart of a model parameter transmission method according to an embodiment of the present application;

[0033] FIG. 5 is an example of an ELU function formula;

[0034] FIG. 6 is a flowchart of a model parameter transmission method according to another embodiment of the present application;

[0035] FIG. 7 is a block diagram of a model parameter transmission apparatus according to an embodiment of the present application;

[0036] FIG. 8 is a block diagram of a model parameter transmission apparatus according to another embodiment of the present application;

[0037] FIG. 9 is a structural diagram of a communication device according to an embodiment of the present application;

[0038] FIG. 10 is a structural diagram of a terminal according to an embodiment of the present application;

[0039] FIG. 11 is a structural diagram of an access network device according to an embodiment of the present application. DETAILED DESCRIPTION

[0040] The technical solutions in the embodiments of the present application will be clearly described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art are within the scope of protection of the present application.

[0041] The terms "first", "second", and the like in the present application are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, so that the embodiments of the present application can be implemented in an order other than those illustrated or described herein, and the objects distinguished by "first", "second" are generally a class, and are not limited to the number of objects, for example, the first object can be one or more. In addition, "or" in the present application means at least one of the connected objects. For example, the protection scope of "A or B" at least covers three schemes, namely, scheme one: including A and not including B; scheme two: including B and not including A; scheme three: including A and including B. In addition, the terms "A and / or B", "at least one of A and B", "at least one of A or B" also at least cover the above three schemes, respectively. The character " / " generally represents that the objects before and after are in an "or" relationship.

[0042] The term "indication" in the present application can be a direct indication (or explicit indication) or an indirect indication (or implicit indication). Among them, the direct indication can be understood as that the sender explicitly informs the receiver of the specific information, the operation to be performed or the request result, etc. in the sent indication; the indirect indication can be understood as that the receiver determines the corresponding information according to the indication sent by the sender, or judges and determines the operation to be performed or the request result, etc. according to the judgment result.

[0043] It is worth noting that the technology described in the embodiments of the present application is not limited to Long Term Evolution (LTE) / LTE-Advanced (LTE-A) systems, but can also be used in other wireless communication systems, such as Code Division Multiple Access (CDMA), Time Division Multiple Access (TDMA), Frequency Division Multiple Access (FDMA), Orthogonal Frequency Division Multiple Access (OFDMA), Single-carrier Frequency-Division Multiple Access (SC-FDMA) or other systems. The terms "system" and "network" in the embodiments of the present application are often used interchangeably, and the described technology can be used in the above-mentioned systems and radio technologies, as well as in other systems and radio technologies. The following description describes a New Radio (NR) system for the purpose of example, and NR terminology is used in most of the following description, but these technologies can also be applied to systems other than NR systems, such as 6th Generation (6G) communication systems.

[0044] FIG. 1 shows a block diagram of a wireless communication system to which embodiments of the present application can be applied. The wireless communication system includes a terminal 11 and a network-side device 12. The terminal 11 can be a terminal-side device such as a mobile phone, a Tablet Personal Computer, a Laptop Computer, a notebook computer, a Personal Digital Assistant (PDA), a palmtop computer, a netbook, an Ultra-mobile Personal Computer (UMPC), a Mobile Internet Device (MID), an Augmented Reality (AR) device, a Virtual Reality (VR) device, a robot, a wearable device, a flight vehicle, a Vehicle User Equipment (VUE), a shipboard device, a Pedestrian User Equipment (PUE), a smart home (a home device with a wireless communication function such as a refrigerator, a television, a washing machine, or furniture), a game console, a Personal Computer (PC), a kiosk, or a self-service machine. The wearable device includes a smart watch, a smart bracelet, a smart earphone, smart glasses, smart jewelry (a smart bracelet, a smart necklace, a smart ring, a smart necklace, a smart anklet, a smart necklace, etc.), a smart wristband, smart clothes, etc. The vehicle-mounted device can also be referred to as a vehicle-mounted terminal, a vehicle-mounted controller, a vehicle-mounted module, a vehicle-mounted component, a vehicle-mounted chip, or a vehicle-mounted unit, etc. It should be noted that the specific type of the terminal 11 is not limited in the embodiments of the present application. The network-side device 12 can include an access network device or a core network device. The access network device can also be referred to as a Radio Access Network (RAN) device, a radio access network function, or a radio access network unit. The access network device can include a base station, a Wireless Local Area Network (WLAN) Access Point (AP), or a Wireless Fidelity (WiFi) node, etc.The base station can be referred to as a Node B (NB), an evolved Node B (eNB), a next generation Node B (gNB), a New Radio Node B (NR Node B), an access point, a relay station (RBS), a serving base station (SBS), a base transceiver station (BTS), a radio base station, a radio transceiver, a basic service set (BSS), an extended service set (ESS), a home Node B (HNB), a home evolved Node B, a transmit / receive point (TRP), or some other suitable terminology in the art, and is not limited to a particular technical terminology, provided that the same technical effect is achieved. It should be noted that in the embodiments of the present application, only the base station in the NR system is taken as an example for introduction, and the specific type of the base station is not limited.

[0045] The core network device can also be referred to as a core network node, a core network function, or a core network network element, etc., which includes but is not limited to at least one of the following: a mobility management entity (MME), an access and mobility management function (AMF), a session management function (SMF), a user plane function (UPF), a policy control function (PCF), a policy and charging rules function (PCRF), an edge application server discovery function (EASDF), a unified data management (UDM), a unified data repository (UDR), a home subscriber server (HSS), a centralized network configuration (CNC), a network repository function (NRF), a network exposure function (NEF), a local NEF (L-NEF), a binding support function (BSF), an application function (AF), a location management function (LMF), a gateway mobile location center (GMLC), a network data analytics function (NWDAF), etc. It should be noted that only the core network device in the NR system is taken as an example for introduction in the embodiments of the present application, and the specific type of the core network device is not limited. If the name of the core network device mentioned in the embodiments of the present application changes in the subsequent protocol version (for example, 6G), it is also within the protection scope of the present application.

[0046] Optionally, the core network device can be implemented by one or more function modules in one device, or can be implemented by multiple devices together, and the embodiments of the present application do not make a specific limitation here. It can be understood that the above function modules can be network elements in a hardware device, can be software function modules running on a special hardware, or can be virtualized function modules instantiated on a platform (for example, a cloud platform).

[0047] The related technologies of the embodiments of the present application are described as follows.

[0048] A neural network is composed of neurons, and a schematic of a neuron is shown in FIG. 2. Among them, a1, a2, … a K are inputs, w is a weight (multiplicative coefficient), b is a bias (additive coefficient), and σ(.) is an activation function. Common activation functions include Sigmoid, tanh, ReLU (Rectified Linear Unit, linear rectification function, modified linear unit), etc.

[0049] The parameters of a neural network are optimized by a gradient optimization algorithm. The gradient optimization algorithm is a class of algorithms for minimizing or maximizing an objective function (sometimes also called a loss function), and the objective function is often a mathematical combination of model parameters and data. For example, given data X and its corresponding label Y, we construct a neural network model f(.). After having the model, we can get the predicted output f(x) according to the input x, and can calculate the gap between the predicted value and the true value (f(x)-Y), which is the loss function. Our goal is to find appropriate W, b to make the value of the above loss function reach the minimum, and the smaller the loss value is, the closer our model is to the true situation.

[0050] At present, the common optimization algorithm is basically based on the error back propagation (BP) algorithm. The basic idea of the BP algorithm is that the learning process consists of two processes of forward propagation of signals and backward propagation of errors. When the forward propagation is performed, the input sample is transmitted from the input layer, processed layer by layer through each hidden layer, and then transmitted to the output layer. If the actual output of the output layer does not match the expected output, the backward propagation of errors is entered. The error back propagation is to transmit the output error to the input layer through the hidden layer in a certain form, and allocate the error to all units of each layer, so as to obtain the error signal of each unit, which is used as the basis for correcting the weight of each unit. The weight adjustment process of each layer in the signal forward propagation and error backward propagation is repeatedly performed. The process of continuously adjusting the weight is the learning and training process of the network. This process is performed until the error of the network output is reduced to an acceptable level, or until the preset learning times are reached.

[0051] Common optimization algorithms include Gradient Descent, Stochastic Gradient Descent (SGD), mini-batch gradient descent, Momentum, Nesterov with Momentum, ADAptive GRADient descent (Adagrad), Adadelta, root mean square prop (RMSprop), Adaptive Moment Estimation (Adam), and the like.

[0052] These optimization algorithms, when error backpropagation, are based on the error / loss obtained from the loss function, the derivative / partial derivative of the current neuron, the learning rate, the previous gradient / derivative / partial derivative, and the like, to obtain the gradient, and pass the gradient to the previous layer.

[0053] The AI model mentioned in the embodiments of the present application can also be referred to as an AI unit, a machine learning (ML) model, an ML unit, an AI structure, an AI function, an AI feature, a machine learning model, a neural network, a neural network function, a neural network function, or the like. Alternatively, the AI model can refer to a processing unit that can implement a specific algorithm, formula, feature, processing flow, capability, or the like related to AI, or the AI model can be a processing method, algorithm, function, feature, module or unit for a specific data set, or the AI model can be a processing method, algorithm, function, feature, module or unit running on a graphics processing unit (GPU), neural processing unit (NPU), tensor processing unit (TPU) or application-specific integrated circuit (ASIC) and the like AI / ML related hardware, which is not specifically limited in the embodiments of the present application. Alternatively, the specific data set includes the input and / or output of the AI model.

[0054] Alternatively, the identification of the AI model can be the identification of the AI model, the identification of the AI structure, the identification of the AI algorithm, or the identification of the specific data set associated with the AI model, or the identification of the specific scene, environment, area, cell, channel feature, device related to AI / ML, or the identification of the function, feature, capability or module related to AI / ML, which is not specifically limited in the embodiments of the present application.

[0055] Explanation of some terms mentioned in the embodiments of the present application:

[0056] A neuron can refer to an element or a unit;

[0057] A layer can refer to a collection of multiple elements, or a collection of multiple units;

[0058] An AI model can refer to a collection of mathematical operations containing multiple neurons, elements or units.

[0059] The vectors or matrices in the embodiments of the present application can also be indicated, sent or stored in a split manner. For example, a three-dimensional matrix is split into multiple two-dimensional matrices or multiple vectors for indication, sending or storage. Not listed one by one here.

[0060] The vectors or matrices in the embodiments of the present application can also be indicated, sent or stored in a combined manner. For example, multiple two-dimensional matrices or multiple vectors are combined into a high-dimensional matrix for indication, sending or storage. Not listed one by one here.

[0061] The modules, models, networks in the embodiments of the present application can refer to complete modules, models, networks, or sub-modules, sub-models, sub-networks in the entire modules, models, networks.

[0062] The layer or current layer in the embodiments of the present application can refer to a complete layer of an AI model, or a layer of part of sub-modules, sub-models, sub-networks in the AI model (possibly there are multiple parallel sub-modules in a certain part of the AI model, and these parallel sub-modules all have their own current layers).

[0063] For example, in FIG. 3, the current layer in the embodiments of the present application can be layer 1, layer 4 or layer 5, or sub-module 1-layer 2, sub-module 1-layer 3, sub-module 2-layer 2, sub-module 2-layer 3, sub-module 3-layer 2, sub-module 3-layer 3, etc.

[0064] For sub-module 1-layer 2, the input is the output of layer 1, and the output is the input of sub-module 1-layer 3; for sub-module 1-layer 3, the input is the output of sub-module 1-layer 2, and the output is the input of layer 4.

[0065] The input or output of one layer in the embodiments of the present application can be a vector (mainly for a fully connected model or a fully connected layer), a three-dimensional matrix, a multi-channel two-dimensional matrix, or a plurality of feature maps (mainly for a convolution model or a convolution layer, a transformer model or a transformer layer, an attention module, an attention model or an attention layer, etc.) (including a special case where the channel is 1, in which case the three-dimensional matrix degenerates into a two-dimensional matrix, and the feature map has only one).

[0066] The model parameter transmission method, device and communication device provided by the embodiments of the present application will be described in detail in combination with the accompanying drawings and some embodiments and application scenarios.

[0067] As shown in FIG. 4, the embodiments of the present application provide a model parameter transmission method, comprising:

[0068] Step 401, a first communication device determines to-be-sent model parameters according to a target artificial intelligence (AI) model;

[0069] Step 402, the first communication device sends the model parameters to a second communication device.

[0070] It should be noted that the embodiments of the present application determine to-be-sent model parameters according to a target AI model, and send the model parameters to a second communication device, thereby realizing the transmission of an AI model between different communication devices and ensuring the reliability of model application.

[0071] Optionally, the first communication device in the embodiments of the present application can be a terminal, an access network device, or a core network device, and the second communication device can be a terminal, an access network device, or a core network device; for example, the first communication device is an access network device, and the second communication device is a terminal, or for example, the first communication device is a core network device, and the second communication device is an access network device; it should be noted that the combination of the first communication device and the second communication device in the embodiments of the present application is only an example and does not constitute a limitation on the protection scope of the present application.

[0072] Optionally, in an implementation manner, the target AI model comprises at least one of the following:

[0073] A11, a first AI model used by the first communication device and the second communication device;

[0074] A12, a reference AI model of the first AI model;

[0075] A13, a second AI model used by the first communication device, the second communication device, or a test device in testing;

[0076] A14, a reference AI model of the second AI model;

[0077] A15, the first communication device, the second communication device or the test device is configured to match a third AI model of the second AI model used in the test;

[0078] It should be noted that this feature is mainly for the dual-end communication use case (such as channel state information (CSI) compression and recovery, encoding and decoding, etc.).

[0079] A16, a reference AI model of the third AI model.

[0080] Optionally, in an implementation, the target AI model is configured to implement at least one of the following:

[0081] A101, reference signal processing;

[0082] Optionally, the reference signal processing can include signal detection, filtering, equalization, etc.

[0083] Optionally, the reference signal can include, but is not limited to, at least one of the following: demodulation reference signal (DMRS), sounding reference signal (SRS), synchronization signal / physical broadcast channel signal block (or synchronization signal block) (SSB), channel state information reference signal (CSI-RS), tracking reference signal (TRS), positioning reference signal (PRS), phase tracking reference signal (PTRS), etc.

[0084] A102, channel signal transmission;

[0085] A103, channel signal reception;

[0086] A104, channel signal demodulation;

[0087] A105, channel signal transmission;

[0088] Optionally, the channel signal includes, but is not limited to, at least one of the following: a physical downlink control channel (PDCCH), a physical downlink shared channel (PDSCH), a physical uplink control channel (PUCCH), a physical uplink shared channel (PUSCH), a physical random-access channel (PRACH), a physical broadcast channel (PBCH), and the like.

[0089] A106, channel state information acquisition;

[0090] Optionally, such a case can include at least one of the following:

[0091] a) Channel state information feedback, the main content of the feedback can include channel related information, channel matrix related information, channel feature information, channel matrix feature information, precoding matrix indication (PMI), rank indication (RI), CSI-RS resource indication (CRI), channel quality indication (CQI), layer indication (LI), and the like.

[0092] b) Frequency division duplex (FDD) uplink-downlink partial reciprocity. For the FDD system, according to the partial reciprocity, the access network device obtains the angle and time delay information of the uplink channel, can inform the terminal of the angle and time delay information through the method of CSI-RS precoding or direct indication, and the terminal reports according to the indication of the access network device or selects and reports within the indication range of the access network device, thereby reducing the calculation amount of the terminal and the overhead of CSI reporting.

[0093] A107, beam management;

[0094] Optionally, the beam management can include, but is not limited to, at least one of the following: beam measurement, beam reporting, beam prediction (spatial domain prediction or frequency domain prediction), beam failure detection, beam failure recovery, new beam indication in beam failure recovery.

[0095] A108, channel prediction;

[0096] Optionally, the channel prediction can include, but is not limited to, at least one of the following: prediction of channel state information, beam prediction.

[0097] A109, channel coding and decoding;

[0098] Optionally, the channel coding and decoding can include: channel coding, channel decoding, joint source-channel coding, joint source-channel decoding.

[0099] A110, source coding and decoding;

[0100] Optionally, the source coding and decoding can include: source coding, source decoding, joint source-channel coding, joint source-channel decoding.

[0101] A111, interference suppression;

[0102] Optionally, the interference mentioned in this case can include, but is not limited to, at least one of the following: intra-cell interference, inter-cell interference, out-of-band interference, intermodulation interference.

[0103] A112, positioning;

[0104] Optionally, the positioning in this case can be understood as: the specific position (including horizontal position and / or vertical position) of the terminal estimated by the reference signal (such as SRS) or the future possible trajectory, or the information assisting the position estimation or trajectory estimation, such as Timing of Arrival (TOA), line of sight or non-line of sight, or Reference Signal Time Difference (RSTD).

[0105] A113, prediction and management of high-level services and parameters;

[0106] Optionally, the high-level services and parameters can include, but are not limited to, at least one of the following: throughput, required packet size, service demand, moving speed, noise information.

[0107] A114, analysis of control signaling;

[0108] Optionally, the control signaling can include power control related signaling, beam management related signaling, etc.

[0109] It should be noted that the above A101-A114 can be understood as the functions that the target AI model can achieve.

[0110] Optionally, after receiving the model parameters, the second communication device needs to apply the target AI model according to the model parameters.

[0111] Optionally, the application target AI model in the embodiments of the present application can be understood as using the target AI model to perform a function operation related to the target AI model, for example, using the target AI model to perform positioning, or using the target AI model to perform reference signal processing, etc.

[0112] Optionally, in an implementation, the target AI model is described by at least one of the following:

[0113] B11, a model inference formula;

[0114] That is, in this case, the model inference formula of the target AI model is given, that is, the target AI model is composed of the model inference formula.

[0115] B12, a model structure description;

[0116] Optionally, the model structure description can be described by a model structure diagram or a model structure text, etc.

[0117] B13, code;

[0118] Optionally, the code can be real executable code, or core code, or illustrative pseudo code based on tensorflow or pytorch.

[0119] Optionally, in an implementation, the target object of the target AI model is obtained by at least one of the following:

[0120] B21, protocol agreement;

[0121] That is, in this case, the target object of the target AI model is a protocol agreement, which can be obtained by each communication device accessing the network.

[0122] For example, in the case where the target AI model is described by a model inference formula, the model inference formula is agreed by the protocol, and the names of the parameters in the formula are labeled.

[0123] For example, in the case where the target AI model is described by a model structure description (such as a model structure diagram or a model structure text description), the model structure diagram or the model structure is agreed by the protocol, and the names of the parameters in the neurons are labeled.

[0124] For example, in the case where the target AI model is described by code, the code is agreed by the protocol, and the names of the parameters in the code are contained or labeled in the code.

[0125] B22, offline negotiation;

[0126] It should be noted that offline negotiation is another agreement mode different from agreement, that is, only the devices participating in negotiation can know how the model is, for example, the model inference formula, model structure diagram, code of the model can be negotiated offline, and the name and dimension of the parameter are labeled.

[0127] Optionally, in an implementation, the target object includes at least one of the following:

[0128] B31, model structure;

[0129] B32, name of the model parameter;

[0130] B33, type of the model parameter;

[0131] B34, description of the model parameter;

[0132] B35, format of the model parameter;

[0133] Optionally, the format can be a transmission format of the model parameter, a protocol format of the model parameter, etc.

[0134] B36, operation corresponding to the model parameter.

[0135] For example, the operation can be a multiplication operation, an addition operation, a convolution operation, an activation function related operation, a normalization operation, etc. corresponding to the model parameter.

[0136] Optionally, different target objects in the embodiments of the present application can adopt different acquisition modes, for example, the model structure can adopt a protocol predetermined mode, and the format of the model parameter can adopt an offline negotiation mode.

[0137] Optionally, in an implementation, the type of the model parameter includes at least one of the following:

[0138] C11, first coefficient, the first coefficient being used to represent a coefficient that an input of a target layer needs to be multiplied by;

[0139] It should be noted that the target layer mentioned in the embodiments of the present application can be understood as an arbitrary layer of the target AI model; or can be understood as a current layer being processed by the target AI model.

[0140] For example, the first coefficient represents a coefficient that an input of a target layer needs to be multiplied by, and the input can be an input vector, which can be understood as an input combination of elements of multiple neurons of the target layer.

[0141] Optionally, the first coefficient can be understood as a multiplicative coefficient, that is, a weight.

[0142] C12, a second coefficient, the second coefficient is used to represent a coefficient that needs to be added to the input of the target layer after the target operation;

[0143] Optionally, the target operation can include, but is not limited to, at least one of the following: multiplication by a multiplicative coefficient, convolution, attention, transposition, dimension conversion.

[0144] That is, the second coefficient represents a coefficient that needs to be added to the input of the target layer before entering the activation function after the target operation. The input can be an input vector, which can be understood as an input combination of elements of multiple neurons of the target layer.

[0145] Optionally, the second coefficient can be understood as an additive coefficient, that is, representing a bias.

[0146] C13, a third coefficient, the third coefficient is used to represent a coefficient that the input of the target layer needs to be convolved by;

[0147] Optionally, the input can be an input vector, which can be understood as an input combination of multiple neurons or feature maps of the target layer.

[0148] Optionally, the third coefficient can be understood as a convolution kernel parameter or a convolution parameter.

[0149] C14, a fourth coefficient, the fourth coefficient is used to represent an activation function configuration parameter of the target layer;

[0150] Optionally, the activation function configuration parameter can include at least one of the following: the type of the activation function, the parameter in the activation function.

[0151] Optionally, the parameter in the activation function can include, but is not limited to, at least one of the following: the Leaky coefficient of LeakyReLu (for example, the formula of LeakyReLu is y = max (ax, x), where a is a small slope (usually close to zero), and a is the Leaky coefficient); the coefficient in the Exponential Linear Unit (ELU) function, for example, as shown in the formula in FIG. 5, where a in the formula is the coefficient.

[0152] It should be noted here that multiple layers or all layers can share a set of activation function configuration parameters. In this case, the first communication device does not need to repeatedly send the activation function configuration parameters of each layer.

[0153] C15, a fifth coefficient, the fifth coefficient is used to represent a normalization type or parameter.

[0154] Optionally, in an implementation, the type of the model parameter satisfies at least one of the following:

[0155] C21. For a target AI model containing fully connected components, the type of the model parameters includes at least one of the following: first coefficient, second coefficient, fourth coefficient, and fifth coefficient;

[0156] Optionally, the target AI model may contain fully connected components in its model, module, network, submodule, submodel, or subnetwork.

[0157] For example, a fully connected layer is: y1=F1(W1*x+b1), where x is the input information, which is a vector; W1 is the weight or multiplicative coefficient, which is a matrix; b1 is the bias or additive coefficient, which is a vector; and F1 is the activation function of the layer.

[0158] For example, a fully connected system consisting of three layers can be fully described by the following inference formula:

[0159] Formula 1: y1 = F1(W1*x + b1);

[0160] Formula 2, y2=F2(W2*y1+b2);

[0161] Formula 3: y3 = F3(W3*y2 + b3), or y3 = W3*y2 + b3;

[0162] Where x is the input information vector, W1 is the weight or multiplicative coefficient matrix of the first layer, b1 is the bias or additive coefficient vector of the first layer, F1 is the activation function of the first layer, and y1 is the result vector of the first layer; W2 is the weight or multiplicative coefficient matrix of the second layer, b2 is the bias or additive coefficient vector of the second layer, F2 is the activation function of the first layer, and y2 is the result vector of the second layer; W3 is the weight or multiplicative coefficient matrix of the third layer, b3 is the bias or additive coefficient vector of the third layer, F3 is the activation function of the third layer (the third layer is the output layer, and may not have an activation function), and y3 is the result vector of the third layer, which is also the final output vector of the model. At least some of W1, b1, W2, b2, W3, and b3 are model parameters to be sent (it may only be necessary to update some parameters). Optionally, F1, F2, and F3 can also be model parameters to be sent. Optionally, normalization operations can be performed on x, y1, y2, and y3.

[0163] C22. For a target AI model that includes convolution, the type of the model parameters includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, and fifth coefficient;

[0164] Optionally, the target AI model may contain convolutions in its model, module, network, submodule, submodel, or subnetwork.

[0165] Optionally, for the target AI model containing convolution, the convolution operation needs to be defined, and the convolution kernel parameters and additive coefficients are the model parameters that need to be sent.

[0166] Optionally, for the target AI model containing convolution, the activation function and normalization need to be configured.

[0167] Optionally, the convolution model generally also contains a fully connected sub-module, and the multiplicative coefficients and additive coefficients of the fully connected sub-module are the model parameters that need to be sent. Of course, there are also convolution models or convolution layers that do not include a fully connected sub-module.

[0168] C23, for the target AI model containing attention, the type of model parameters includes attention parameters, and the type of attention parameters includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, fifth coefficient;

[0169] Optionally, the model, module, network, sub-module, sub-model or sub-network of the target AI model can contain attention. For such a target AI model, the attention operation needs to be defined, and at this time, the multiplicative coefficients and additive coefficients in the fully connected sub-module used in the attention operation, and / or the convolution kernel parameters and additive coefficients in the convolution sub-module used in the attention operation are the model parameters that need to be sent.

[0170] It should be noted here that the attention mechanism generally contains feature extraction, factor prediction, and feature recalibration, specifically:

[0171] Feature extraction generally needs to use a fully connected sub-module, and / or a convolution sub-module, and / or pooling. The model parameters that need to be sent are the multiplicative coefficients and additive coefficients in the fully connected sub-module, the convolution kernel parameters and additive coefficients in the convolution sub-module. For the parameters of the pooling, the general protocol does not need to be sent.

[0172] Factor prediction generally needs to use a fully connected sub-module and / or a convolution sub-module. The model parameters that need to be sent are the multiplicative coefficients and additive coefficients in the fully connected sub-module, and the convolution kernel parameters and additive coefficients in the convolution sub-module.

[0173] Feature recalibration is generally matrix point multiplication or matrix multiplication by elements, and the relevant parameters can not need to be sent.

[0174] Need to be explained, for the transformer, using multi-head self-attention mechanism, a block (block) or sub-module (such as a Transformer block, or attention block, or multi-head self-attention block), mainly includes: input to the query (Query), key (Key) and value (Value) transformation or mapping; Query, Key and Value itself operation; After multi-head self-attention, full connection or convolution module, specifically:

[0175] For the transformation or mapping of input to Query, Key and Value, generally using full connection sub-module and / or convolution sub-module, the model parameters to be sent are the multiplicative coefficients, additive coefficients in the full connection sub-module, and the convolution kernel parameters, additive coefficients in the convolution sub-module.

[0176] For the operation of Query, Key and Value itself, no related parameters need to be issued.

[0177] For the full connection or convolution module after multi-head self-attention, generally using full connection sub-module and / or convolution sub-module, the model parameters to be sent are the multiplicative coefficients, additive coefficients in the full connection sub-module, and the convolution kernel parameters, additive coefficients in the convolution sub-module.

[0178] Optionally, for the target AI model containing attention, the activation function and normalization need to be configured.

[0179] C24, for the target AI model containing multi-layer perceptron mixer (MLP-Mixer) or mixer (Mixer), the type of model parameters includes at least one of the following: first coefficient, second coefficient, fourth coefficient, fifth coefficient;

[0180] Optionally, the target AI model can contain multi-layer perceptron mixer or mixer in the model, module, network, sub-module, sub-model or sub-network.

[0181] Need to be explained, MLP-Mixer mainly includes per-patch fully-connected, token-mixing MLP, channel-mixing MLP, patch to channel conversion, channel to patch conversion and some full connection sub-modules; Specifically:

[0182] Per-patch Fully-connected is a fully connected sub-module.

[0183] token-mixing MLP is token mixing, generally composed of a number of parallel fully connected sub-modules.

[0184] Channel-mixing MLP is Channel mixing, generally composed of a number of parallel fully connected sub-modules.

[0185] Channel-to-patch conversion is generally a transpose operation of two dimensions in a high-dimensional matrix, which may not require the delivery of related parameters.

[0186] Optionally, for target AI models containing MLP-Mixer or Mixer, an activation function and normalization need to be configured.

[0187] It should be noted that the model parameters required to be delivered by defining target AI models meeting different requirements can ensure accurate model parameters for different types of target AI models and ensure accurate transmission of model parameters.

[0188] Optionally, in an implementation manner, the sending of the model parameters to the second communication device comprises:

[0189] The model parameters are sent to the second communication device in the form of a matrix and / or a vector.

[0190] Further optionally, the sending of the model parameters to the second communication device in the form of a matrix or a vector comprises at least one of the following:

[0191] D11, in the case where the model parameters comprise first coefficients, the first coefficients of at least one layer in the target AI model are represented in the form of a vector or a matrix, the first coefficients are sent to the second communication device, all first coefficients of at least one layer are combined into a matrix, or all first coefficients of at least one layer are combined by a plurality of first coefficient vectors, the first coefficient vector is a vector composed of first coefficients associated with a target input or a target output;

[0192] Optionally, the target input can be any input, and the target output can be any output.

[0193] It should be noted that when the specific model parameter transmission is performed, not all layers of the target AI model need to transmit parameters, and the layer in the embodiment of the present application refers to the layer that needs to transmit the model parameter. Alternatively, the first coefficients of these layers can constitute a vector or a matrix alone, or the first coefficients of multiple layers can constitute a vector or a matrix. For example, the first coefficient of a layer is a matrix, and the first coefficients of multiple layers are combined into a three-dimensional matrix.

[0194] It is assumed that the target layer input is an N1 vector, the target layer has N2 neurons, or the target output is an N2 vector.

[0195] Alternatively, if represented by a matrix, all multiplicative coefficients of the target layer are combined into a matrix (for example, the matrix can have dimensions of N2*N1 or N1*N2).

[0196] Alternatively, if represented by a vector, the multiplicative coefficients associated with a certain input or a certain output constitute a vector (the vector has a dimension of N1 vector (all inputs associated with a certain output) or a dimension of N2 vector (all outputs associated with a certain input)), and all multiplicative coefficient vectors of the target layer (N1 N2-dimensional vectors or N2 N1-dimensional vectors) are combined to obtain all multiplicative coefficients of the target layer.

[0197] D12, in the case where the model parameters include second coefficients, the second coefficients of at least one layer in the target AI model are represented by a vector or a matrix, the second coefficients are sent to the second communication device, and all second coefficients of at least one layer are combined into at least one matrix or at least one vector;

[0198] It should be noted that when the specific model parameter transmission is performed, not all layers of the target AI model need to transmit parameters, and the layer in the embodiment of the present application refers to the layer that needs to transmit the model parameter. Alternatively, the second coefficients of these layers can constitute a vector or a matrix alone, or the second coefficients of multiple layers can constitute a vector or a matrix. For example, the second coefficient of a layer is a matrix, and the second coefficients of multiple layers are combined into a three-dimensional matrix.

[0199] Alternatively, if represented by a vector (when the target layer input or output is a vector, mainly for a fully connected model), all additive coefficients of the target layer are combined into a vector. It is assumed that the target layer has N2 neurons or the target layer output is an N2 vector, and the additive coefficient vector is an N2-dimensional vector.

[0200] Optionally, if represented by a three-dimensional matrix, a two-dimensional matrix or a vector (when the current layer input or output is a matrix or a feature map, mainly for convolution models, transformer models, attention modules / models, etc.), all additive coefficient combinations of the target layer are combined into a matrix or a vector. Assuming that the target layer has N two-dimensional feature maps or two-dimensional matrices of M1*M2, or the target layer output is N two-dimensional feature maps or two-dimensional matrices of M1*M2, where M1 is one dimension of the two-dimensional input dimension or two-dimensional output dimension of the feature map or neuron of the target layer, and M2 is the other dimension of the two-dimensional input dimension or two-dimensional output dimension of the feature map or neuron of the target layer; optionally, the second coefficient of at least one layer in the target AI model is represented by a vector or a matrix, including at least one of the following:

[0201] a. The second coefficient of at least one layer in the target AI model is represented by a three-dimensional matrix, and the three dimensions of the three-dimensional matrix are represented by N, M1 and M2.

[0202] Optionally, the dimensions of the three-dimensional matrix can be represented by N*M1*M2.

[0203] Where N is the number of all feature maps or all neurons or all outputs of the target layer.

[0204] It should be noted that the three-dimensional matrix is the additive coefficient of all feature maps or all neurons or all outputs of the target layer; and one three-dimensional matrix is needed to describe the additive coefficient of all feature maps or all neurons or all outputs of the target layer.

[0205] b. The second coefficient of at least one layer in the target AI model is represented by N two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by M1 and M2.

[0206] Optionally, the dimensions of the two-dimensional matrix can be represented by M1*M2.

[0207] It should be noted that the two-dimensional matrix is the additive coefficient of the current feature map or neuron or output; and N two-dimensional matrices are needed to describe the additive coefficient of all feature maps or all neurons or all outputs of the target layer.

[0208] c. The second coefficient of at least one layer in the target AI model is represented by M1*M2 vectors, and the vector includes N elements.

[0209] It should be noted that the vector is the additive coefficient of all single or specified feature map positions, or positions on neurons, or positions on outputs of the target layer. For example, the specified position is the coordinate [x1, y1] on the two-dimensional feature map of M1*M2, and the vector is the additive coefficient corresponding to the coordinate [x1, y1] on N2 two-dimensional feature maps. In this case, M1*M2 are required to describe the additive coefficients of all feature maps or all neurons or all outputs of the target layer.

[0210] D13, in the case that the model parameters include third coefficients, representing the third coefficients in the target case by a matrix, sending the third coefficients to the second communication device, and combining all third coefficients of at least one layer into at least one matrix, the target case being all inputs to all outputs of the target layer in the target AI model, or the target layer to the next layer of the target layer in the target AI model;

[0211] It should be noted that when performing specific model parameter transmission, not all parameters of all layers of the target AI model need to be transmitted, and the layer in the embodiment of the application is the layer that needs to perform model parameter transmission. Alternatively, the third coefficients of these layers can form a vector or a matrix individually, or the third coefficients of multiple layers can form a vector or a matrix; for example, the third coefficients of one layer form a matrix, and the third coefficients of multiple layers form a three-dimensional matrix.

[0212] For example, it can be represented by a four-dimensional matrix, multiple three-dimensional matrices, or multiple two-dimensional matrices. Assuming that the convolution kernel is K1*K2, the current layer input has N1 channels (also called dimensions, feature map quantities, neuron quantities, element quantities, unit quantities, etc.), and the output has N2 channels. Alternatively, the third coefficients in the target case are represented by a matrix, including at least one of the following:

[0213] a. representing the third coefficients in the target case by a four-dimensional matrix, and the four dimensions of the four-dimensional matrix are represented by K1, K2, N1, and N2;

[0214] It should be noted that N1 is the number of input channels of the target layer, N2 is the number of output channels of the target layer, K1 is one dimension of the two-dimensional dimension of the convolution kernel, and K2 is the other dimension of the two-dimensional dimension of the convolution kernel.

[0215] It should be noted that the four-dimensional matrix is all the input of the target layer to all the output of the target layer, or all the convolution parameters of the target layer to the next layer of the target layer, and the dimension order of the four-dimensional matrix can be changed, which can be represented by N1*N2*K1*K2, K2*K1*N1*N2, and the like, which will not be listed one by one. It should be noted that in this case, only one four-dimensional matrix is needed to describe all the input of the target layer to all the output of the target layer or all the convolution parameters of the target layer to the next layer of the target layer.

[0216] b. The third coefficient in the target case is represented by N2 three-dimensional matrices, and the three dimensions of the three-dimensional matrix are represented by N1, K1 and K2.

[0217] It should be noted that the three-dimensional matrix is the convolution parameter of all input channels associated with a single output channel, and the order of the three dimensions can be changed, which can be represented by N1*K1*K2, K2*K1*N1, and the like, which will not be listed one by one; in this case, N2 three-dimensional matrices are needed to describe all the input of the target layer to all the output of the target layer or all the convolution parameters of the target layer to the next layer of the target layer.

[0218] c. The third coefficient in the target case is represented by N1 three-dimensional matrices, and the three dimensions of the three-dimensional matrix are represented by N2, K1 and K2.

[0219] It should be noted that the three-dimensional matrix is the convolution parameter of all input channels associated with a single input channel, and the order of the three dimensions can be changed, which can be represented by N2*K1*K2, K2*K1*N2, and the like, which will not be listed one by one; in this case, N1 three-dimensional matrices are needed to describe all the input of the target layer to all the output of the target layer or all the convolution parameters of the target layer to the next layer of the target layer.

[0220] d. The third coefficient in the target case is represented by N1*N2 two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by K1 and K2.

[0221] Alternatively, the dimensions of the two-dimensional matrix can be represented by K2*K1 or K1*K2.

[0222] It should be noted that in this case, N1*N2 two-dimensional matrices are needed to describe all the input of the target layer to all the output of the target layer or all the convolution parameters of the target layer to the next layer of the target layer.

[0223] e. The third coefficient in the target case is represented by K1*K2 two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by N1 and N2.

[0224] Alternatively, the dimensions of the two-dimensional matrix can be represented by N1*N2 or N2*N1.

[0225] Optionally, in an implementation, the model parameter includes: a name of the parameter; or, the model parameter includes: the name of the parameter and a value of the parameter.

[0226] Optionally, the value of the parameter can be a numerical value, content, data, etc.

[0227] Optionally, in an implementation, the sending of the model parameter to the second communication device includes:

[0228] The first communication device sends the model parameter to the second communication device in a target order;

[0229] The target order includes at least one of E11-E13.

[0230] E11, the order of the model parameter name in the model structure description;

[0231] Optionally, the order of the model parameter name in the model structure description includes at least one of:

[0232] E111, in the order of layers in the target AI model, in the order of neurons in the layer;

[0233] Optionally, for a target AI model containing full connection, for a full connection layer, the multiplicative coefficient is first multiplied in the neuron, and then the additive coefficient is added; or, the additive coefficient is first added, and then the multiplicative coefficient is multiplied.

[0234] Optionally, for a target AI model containing convolution, for a convolution layer, the convolution parameter is first multiplied in the neuron, and then the additive coefficient is added; or, the additive coefficient is first added, and then the convolution parameter is multiplied.

[0235] E112, in the order of layers in the target AI model, in the order of the same parameter combination in the layer;

[0236] Optionally, for a target AI model containing full connection, for a full connection layer, the multiplicative coefficient combination in a layer is a two-dimensional matrix.

[0237] Optionally, for a target AI model containing full connection, for a full connection layer, the additive coefficient combination in a layer is a vector.

[0238] Optionally, for a target AI model containing convolution, for a convolution layer, the convolution parameter combination in a layer is a four-dimensional matrix.

[0239] Optionally, for a target AI model containing convolution, for a convolution layer, the additive coefficient combination in a layer is a three-dimensional matrix.

[0240] E12, the order of the model parameter name in the model inference formula;

[0241] Optionally, the model parameter name order of the model inference formula includes at least one of the following:

[0242] E121、the order of appearance of the model parameter name in the model inference formula;

[0243] For example, according to the order of appearance of the model parameter name in the model inference formula (such as from left to right, from top to bottom), the model parameter name that appears first is sent earlier than the model parameter name that appears later; or vice versa, the model parameter name that appears first is sent later than the model parameter name that appears later. For example, a model corresponds to multiple inference formulas, the model parameter name in the formula that appears first is sent earlier than the model parameter name in the formula that appears later; or vice versa, the model parameter name in the formula that appears first is sent later than the model parameter name in the formula that appears later. For example, in the same formula, the model parameter name on the left is sent earlier than the model parameter name on the right; or the model parameter name on the left is sent later than the model parameter name on the right.

[0244] For example, for the three-layer fully connected model of the above-mentioned formula one to formula three, according to the order of the inference formula, W1, b1, W2, b2, W3, b3 are sent in order, or b1, W1, b2, W2, b3, W3 are sent in order, or vice versa, b3, W3, b2, W2, b1, W1 are sent in order, or W3, b3, W2, b2, W1, b1 are sent in order.

[0245] E122、the order of calculation of the model parameter name corresponding to the variable in the model inference formula.

[0246] For example, according to the order of calculation of the model parameter name in the model inference formula, the model parameter name that is calculated first is sent earlier than the model parameter name that is calculated later; or vice versa, the model parameter name that is calculated first is sent later than the model parameter name that is calculated later.

[0247] E13、the order of the model parameter name in the code;

[0248] Optionally, the order of the model parameter name in the code includes:

[0249] E131、the order of appearance of the model parameter name in the code;

[0250] For example, according to the order of appearance, the model parameter name that appears first is sent earlier than the model parameter name that appears later; or vice versa, the model parameter name that appears first is sent later than the model parameter name that appears later.

[0251] E132、In the code, the order of calculation of the model parameter name corresponding variable.

[0252] For example, according to the order of calculation in the code, the model parameter name calculated first is sent earlier than the model parameter name calculated later; or vice versa, the model parameter name calculated first is sent later than the model parameter name calculated later.

[0253] Of course, in the embodiments of the present application, the model parameters can also be sent in any order.

[0254] Optionally, in the case where the model parameters include the name of the parameter and the value of the parameter, the first communication device can send the model parameters in any order; in the case where the model parameters only include the value of the parameter, the first communication device needs to send the model parameters in the order of at least one of E11-E13.

[0255] It should be noted that when the sending of the model parameters of the target AI model is first performed, the sending of all model parameters corresponding to the target AI model is usually required, and when the sending of the model parameters of the target AI model is performed except for the first time, it can be understood as an update of the model parameters. Optionally, in an implementation manner, the model parameters are at least part of the model parameters corresponding to the target AI model.

[0256] That is, when updating, all model parameters corresponding to the target AI model can be updated, or only part of the model parameters corresponding to the target AI model can be updated.

[0257] Optionally, in an implementation manner, the sending of the model parameters to the second communication device comprises:

[0258] The first communication device sends the model parameters to the second communication device according to at least one of the model parameter list and the reporting indication information;

[0259] The reporting indication information is used to indicate at least one of the following: supporting partial update, supporting full update, and not supporting partial update.

[0260] Optionally, the partial update can be understood as updating part of the model parameters, and can also be referred to as partial update of the model parameters; and the full update can be understood as updating all of the model parameters, and can also be referred to as full update of the model parameters.

[0261] Optionally, the reporting indication information is usually sent by the second communication device to the first communication device.

[0262] Optionally, in one implementation, before the first communication device sends the model parameters to the second communication device according to the model parameter list and the reporting indication information, the method further comprises at least one of the following:

[0263] The first communication device sends a model parameter list to the second communication device.

[0264] The first communication device receives a model parameter list to be updated sent by the second communication device.

[0265] Optionally, in the first implementation, the first communication device sends the model parameter list to the second communication device, and then the first communication device sends the model parameters indicated by the model parameter list to the second communication device based on the model parameter list. Optionally, in the second implementation, the second communication device sends a model parameter list to be updated to the first communication device, and then the first communication device sends the model parameters indicated by the model parameter list to be updated to the second communication device based on the model parameter list to be updated. Optionally, in the third implementation, the first communication device sends a model parameter list to the second communication device, and then the second communication device reports a model parameter list to be updated, and the first communication device sends the model parameters indicated by the model parameter list to be updated to the second communication device based on the model parameter list to be updated.

[0266] Optionally, in one implementation, after the first communication device receives the model parameter list to be updated sent by the second communication device, the method further comprises at least one of the following:

[0267] The first communication device sends indication information to the second communication device, and the indication information is used to indicate whether the first communication device supports the model parameter list.

[0268] It should be noted that after the second communication device sends the model parameter list to be updated to the first communication device, if the indication information sent by the first communication device indicates that the first communication device supports the model parameter list, the second communication device can continue to receive the model parameters corresponding to the model parameter list, and if the indication information sent by the first communication device indicates that the first communication device does not support the model parameter list, the second communication device no longer receives the model parameters corresponding to the model parameter list. This case can avoid the transmission of invalid model parameters and ensure the reliability of information transmission.

[0269] Optionally, the first communication device can also send the model parameters meeting the capability of the second communication device to the second communication device according to the reporting indication information.

[0270] The specific application of the update of the model parameters is illustrated as follows.

[0271] Application case one,

[0272] Step S11, the second communication device sends an indication information that the second communication device supports partial update and supports full update, or the second communication device reports an indication information that the second communication device supports partial update;

[0273] Step S12, the first communication device sends all model parameters or partial model parameters of the target AI model;

[0274] Step S13, the second communication device updates the target AI model using the parameters sent by the first communication device.

[0275] It should be noted that step S11 is an optional step or an unnecessary step.

[0276] Application case two,

[0277] Step S21, the second communication device sends an indication information that the second communication device supports full update, or the second communication device reports an indication information that the second communication device does not support partial update;

[0278] Step S22, the first communication device can only send all model parameters of the target AI model;

[0279] Step S23, the second communication device updates the target AI model using the parameters sent by the first communication device.

[0280] It should be noted that step S21 is an optional step or an unnecessary step.

[0281] Application case three,

[0282] Step S31, the second communication device sends an indication information that the second communication device supports partial update and supports full update, or the second communication device reports an indication information that the second communication device supports partial update;

[0283] Step S32, the first communication device sends a list of optional model parameters.

[0284] For example, the target AI model has 100 parameter names, the first communication device sends 3 partial parameter name lists, table 1 is 40 of the 100 parameter names, table 2 is 30 of the 100 parameter names, and table 3 is 20 of the 100 parameter names.

[0285] Step S33, the second communication device reports a list of model parameters that need to be updated;

[0286] It should be noted that in the case of step S32, the second communication device can feed back the list ID of the list of optional model parameters sent by the first communication device, so as to indicate that the parameters that the second communication device needs to update are the parameters in the model parameter list corresponding to the list ID.

[0287] Step S34, the first communication device feeds back indication information of whether to support the model parameter list reported by the second communication device.

[0288] Step S35, the first communication device issues all or part of the model parameters in the model parameter list reported by the second communication device.

[0289] Step S36, the second communication device updates the target AI model using the parameters issued by the first communication device.

[0290] It should be noted that steps S31, S32 and S34 are optional steps or non-essential steps.

[0291] It should be noted that, in the embodiments of the present application, by defining the name and value of the model parameters to be sent, the meaning of the model structure, the transmission mode and format of the model parameters, only the model parameters are transmitted to enable different communication devices to complete the transmission of the model, thereby reducing the delay and overhead of model transmission.

[0292] As shown in FIG. 6, the embodiments of the present application provide a model parameter transmission method, comprising:

[0293] Step 601, the second communication device receives the model parameters of the target artificial intelligence (AI) model sent by the first communication device;

[0294] Step 602, the second communication device applies the target AI model according to the model parameters.

[0295] Optionally, in an implementation, the target AI model is described by at least one of the following:

[0296] a model inference formula;

[0297] a model structure description;

[0298] code.

[0299] Optionally, in an implementation, the target object of the target AI model is obtained by at least one of the following:

[0300] protocol agreement;

[0301] offline negotiation;

[0302] The target object includes at least one of the following:

[0303] a model structure;

[0304] the name of a model parameter;

[0305] the type of a model parameter;

[0306] A description of the model parameter;

[0307] A format of the model parameter;

[0308] An operation corresponding to the model parameter.

[0309] Optionally, in an implementation, the type of the model parameter includes at least one of the following:

[0310] A first coefficient, the first coefficient being used to represent a coefficient that an input of a target layer needs to be multiplied by;

[0311] A second coefficient, the second coefficient being used to represent a coefficient that an input of a target layer needs to be added after a target operation;

[0312] A third coefficient, the third coefficient being used to represent a coefficient that an input of a target layer needs to be convolved by;

[0313] A fourth coefficient, the fourth coefficient being used to represent an activation function configuration parameter of a target layer;

[0314] A fifth coefficient, the fifth coefficient being used to represent a normalization type or parameter.

[0315] Optionally, in an implementation, the type of the model parameter satisfies at least one of the following:

[0316] For a target AI model containing a full connection, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient;

[0317] For a target AI model containing a convolution, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient;

[0318] For a target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient;

[0319] For a target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient.

[0320] Optionally, in an implementation, the model parameter includes a name of the parameter; or the model parameter includes the name of the parameter and a value of the parameter.

[0321] Optionally, in an implementation, the model parameter is at least part of all model parameters corresponding to a target AI model.

[0322] Optionally, in an implementation, before receiving the model parameters of the target artificial intelligence (AI) model sent by the first communication device, the method further includes at least one of the following:

[0323] The second communication device sends reporting indication information to the first communication device, and the reporting indication information is used to indicate at least one of the following: supporting partial update, supporting full update, and not supporting partial update.

[0324] The second communication device receives the model parameter list sent by the first communication device.

[0325] The second communication device sends the model parameter list that needs to be updated to the first communication device.

[0326] Optionally, in an implementation, after the model parameter list that needs to be updated is sent to the first communication device, the method further includes:

[0327] The second communication device receives the indication information sent by the first communication device, and the indication information is used to indicate whether the first communication device supports the model parameter list.

[0328] Optionally, in an implementation, the target AI model includes at least one of the following:

[0329] A first AI model used by the first communication device and the second communication device;

[0330] A reference AI model of the first AI model;

[0331] A second AI model used by the first communication device, the second communication device, or a test device in a test;

[0332] A reference AI model of the second AI model;

[0333] A third AI model used by the first communication device, the second communication device, or the test device to match the second AI model used in the test;

[0334] A reference AI model of the third AI model.

[0335] Optionally, in an implementation, the target AI model is used to implement at least one of the following:

[0336] Reference signal processing;

[0337] Channel signal transmission;

[0338] Channel signal reception;

[0339] Channel signal demodulation;

[0340] Channel signaling;

[0341] Channel state information acquisition;

[0342] Beam management;

[0343] Channel prediction;

[0344] Channel coding and decoding;

[0345] Source coding and decoding;

[0346] Interference suppression;

[0347] Positioning;

[0348] Prediction and management of higher layer services and parameters;

[0349] Analysis of control signaling.

[0350] It should be noted that all the descriptions about the second communication device side in the above embodiments are applicable to the embodiments of the model parameter transmission method applied to the second communication device side, and the same technical effects can be achieved, which will not be repeated here.

[0351] The model parameter transmission device provided in the embodiments of the present application can be a communication device or a component in the communication device, for example, a chip. The communication device can be a terminal, a network side device, a server, etc. For example, the terminal can include but is not limited to the types of the terminal 11 listed above, the network side device can include but is not limited to the types of the network side device 12 listed above, and the embodiments of the present application are not limited specifically.

[0352] The model parameter transmission apparatus comprises a receiving module, a sending module and a processing module. The receiving module, the sending module and the processing module can be implemented by software or by hardware. When implemented by hardware, the processing module can be implemented by a processor, which can include a general-purpose processor, a special-purpose processor, etc., such as a central processing unit (CPU), a microprocessor, a digital signal processor (DSP), an artificial intelligent (AI) processor, a graphics processing unit (GPU), an application specific integrated circuit (ASIC), a network processor (NP), a field programmable gate array (FPGA) or other programmable logic devices, a gate circuit, a transistor, a discrete hardware component, etc. The receiving module and the sending module can be implemented by a communication interface, which can include one or more of a transceiver, a pin, a circuit, a bus, a radio frequency unit, etc.

[0353] Specifically, referring to FIG. 7, when the model parameter transmission apparatus is a first communication device or a component in the first communication device, the model parameter transmission apparatus 700 comprises:

[0354] a first processing module 701 configured to determine model parameters to be sent according to a target artificial intelligent (AI) model;

[0355] a first sending module 702 configured to send the model parameters to a second communication device.

[0356] Optionally, the target AI model is described by at least one of the following:

[0357] a model inference formula;

[0358] a model structure description;

[0359] code.

[0360] Optionally, a target object of the target AI model is obtained by at least one of the following:

[0361] an agreement;

[0362] offline negotiation;

[0363] The target object comprises at least one of the following:

[0364] a model structure.

[0365] a name of the model parameter;

[0366] a type of the model parameter;

[0367] a description of the model parameter;

[0368] a format of the model parameter;

[0369] an operation corresponding to the model parameter.

[0370] Optionally, the type of the model parameter comprises at least one of the following:

[0371] a first coefficient used to represent a coefficient that an input of a target layer needs to be multiplied by;

[0372] a second coefficient used to represent a coefficient that needs to be added after the input of the target layer undergoes a target operation;

[0373] a third coefficient used to represent a coefficient that the input of the target layer needs to undergo convolution;

[0374] a fourth coefficient used to represent an activation function configuration parameter of the target layer;

[0375] a fifth coefficient used to represent a normalization type or parameter.

[0376] Optionally, the type of the model parameter satisfies at least one of the following:

[0377] For a target AI model containing a full connection, the type of the model parameter comprises at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient;

[0378] For a target AI model containing a convolution, the type of the model parameter comprises at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient;

[0379] For a target AI model containing attention, the type of the model parameter comprises an attention parameter, and the type of the attention parameter comprises at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient;

[0380] For a target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter comprises at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient.

[0381] Optionally, the first sending module 702 is configured to:

[0382] send the model parameter to the second communication device in a manner of a matrix and / or a vector.

[0383] Optionally, the first sending module 702 comprises at least one of the following:

[0384] In the case where the model parameters comprise first coefficients, the first coefficients of at least one layer in the target AI model are represented by vectors or matrices, and the first coefficients are sent to the second communication device, all first coefficients of at least one layer are combined into one matrix or all first coefficients of at least one layer are combined by a plurality of first coefficient vectors, and the first coefficient vector is a vector composed of first coefficients associated with a target input or a target output;

[0385] In the case where the model parameters comprise second coefficients, the second coefficients of at least one layer in the target AI model are represented by vectors or matrices, and the second coefficients are sent to the second communication device, all second coefficients of at least one layer are combined into at least one matrix or all second coefficients of at least one layer are combined into at least one vector;

[0386] In the case where the model parameters comprise third coefficients, the third coefficients in a target condition are represented by matrices, and the third coefficients are sent to the second communication device, all third coefficients of at least one layer are combined into at least one matrix, and the target condition is all inputs to all outputs of a target layer in the target AI model or a target layer to a next layer of the target layer in the target AI model.

[0387] Optionally, the second coefficients of at least one layer in the target AI model are represented by vectors or matrices, comprising at least one of the following:

[0388] The second coefficients of at least one layer in the target AI model are represented by a three-dimensional matrix, and the three dimensions of the three-dimensional matrix are represented by N, M1 and M2;

[0389] The second coefficients of at least one layer in the target AI model are represented by N two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by M1 and M2;

[0390] The second coefficients of at least one layer in the target AI model are represented by M1*M2 vectors, and the vector comprises N elements;

[0391] Wherein, N is the number of all feature maps or all neurons or all outputs of the target layer, M1 is one dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer, and M2 is another dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer.

[0392] Optionally, the third coefficients in the target condition are represented by matrices, comprising at least one of the following:

[0393] the third coefficient in the target case is represented by a four-dimensional matrix, four dimensions of the four-dimensional matrix are represented by K1, K2, N1 and N2;

[0394] the third coefficient in the target case is represented by N2 three-dimensional matrices, three dimensions of the three-dimensional matrix are represented by N1, K1 and K2;

[0395] the third coefficient in the target case is represented by N1 three-dimensional matrices, three dimensions of the three-dimensional matrix are represented by N2, K1 and K2;

[0396] the third coefficient in the target case is represented by N1*N2 two-dimensional matrices, two dimensions of the two-dimensional matrix are represented by K1 and K2;

[0397] the third coefficient in the target case is represented by K1*K2 two-dimensional matrices, two dimensions of the two-dimensional matrix are represented by N1 and N2;

[0398] wherein, N1 is the number of input channels of the target layer, N2 is the number of output channels of the target layer, K1 is one dimension of the two-dimensional dimension of the convolution kernel, and K2 is another dimension of the two-dimensional dimension of the convolution kernel.

[0399] Optionally, the model parameter comprises: a name of the parameter; or the model parameter comprises: the name of the parameter and a value of the parameter.

[0400] Optionally, the first sending module 702 is configured to:

[0401] send the model parameter to the second communication device according to a target order;

[0402] wherein, the target order comprises at least one of the following:

[0403] a model parameter name order in a model structure description;

[0404] a model parameter name order in a model inference formula;

[0405] a model parameter name order in code.

[0406] Optionally, the model parameter name order in the model structure description comprises at least one of the following:

[0407] according to an order of layers in the target AI model, according to an order of neurons in the layers;

[0408] according to an order of layers in the target AI model, according to an order of similar parameter combinations in the layers.

[0409] Optionally, the model parameter name order in the model inference formula comprises at least one of the following:

[0410] an order of appearance of the model parameter names in the model inference formula;

[0411] an order of calculation of the variables corresponding to the model parameter names in the model inference formula.

[0412] Optionally, the order of the model parameter names in the code comprises:

[0413] an order of appearance of the model parameter names in the code;

[0414] an order of calculation of the variables corresponding to the model parameter names in the code.

[0415] Optionally, the model parameters are at least part of all model parameters corresponding to the target AI model.

[0416] Optionally, the first sending module 702 is configured to:

[0417] send the model parameters to the second communication device according to at least one of the model parameter list and the reporting indication information;

[0418] The reporting indication information is used to indicate at least one of the following: supporting partial update, supporting full update, and not supporting partial update.

[0419] Optionally, before the first sending module 702 sends the model parameters to the second communication device according to at least one of the model parameter list and the reporting indication information, the apparatus further comprises at least one of the following:

[0420] a second sending module configured to send the model parameter list to the second communication device;

[0421] a second receiving module configured to receive the model parameter list to be updated sent by the second communication device.

[0422] Optionally, after the second receiving module receives the model parameter list to be updated sent by the second communication device, the apparatus further comprises:

[0423] a third sending module configured to send indication information to the second communication device, the indication information being used to indicate whether the first communication device supports the model parameter list.

[0424] Optionally, the target AI model comprises at least one of the following:

[0425] a first AI model used by the first communication device and the second communication device;

[0426] a reference AI model of the first AI model;

[0427] a second AI model used in the test by the first communication device, the second communication device, or a test device;

[0428] a reference AI model of the second AI model;

[0429] a third AI model used by the first communication device, the second communication device, or a test device to match the second AI model used in the test;

[0430] a reference AI model of the third AI model.

[0431] Optionally, the target AI model is used to implement at least one of the following:

[0432] reference signal processing;

[0433] channel signal transmission;

[0434] channel signal reception;

[0435] channel signal demodulation;

[0436] channel signal sending;

[0437] channel state information acquisition;

[0438] beam management;

[0439] channel prediction;

[0440] channel coding and decoding;

[0441] source coding and decoding;

[0442] interference suppression;

[0443] positioning;

[0444] prediction and management of higher layer services and parameters;

[0445] analysis of control signaling.

[0446] The model parameter transmission apparatus provided by the embodiments of the present application can implement each process implemented by the method embodiment of FIG. 4 and achieve the same technical effects. To avoid repetition, the same will not be described here.

[0447] Referring to FIG. 8, when the model parameter transmission apparatus is a second communication device or a component in the second communication device, the model parameter transmission apparatus 800 includes:

[0448] a first receiving module 801 configured to receive model parameters of a target artificial intelligence (AI) model sent by a first communication device;

[0449] a second processing module 802 configured to apply the target AI model according to the model parameters.

[0450] Optionally, the target AI model is described by at least one of:

[0451] a model inference formula;

[0452] a model structure description;

[0453] code.

[0454] Optionally, the target object of the target AI model is obtained by at least one of:

[0455] a protocol agreement;

[0456] offline negotiation;

[0457] wherein the target object includes at least one of:

[0458] a model structure;

[0459] a name of a model parameter;

[0460] a type of a model parameter;

[0461] a description of a model parameter;

[0462] a format of a model parameter;

[0463] an operation corresponding to a model parameter.

[0464] Optionally, the type of the model parameter includes at least one of:

[0465] a first coefficient, the first coefficient being used to represent a coefficient that an input of a target layer needs to be multiplied by;

[0466] a second coefficient, the second coefficient being used to represent a coefficient that an input of a target layer needs to be added after a target operation;

[0467] a third coefficient, the third coefficient being used to represent a coefficient that an input of a target layer needs to be convolved by;

[0468] a fourth coefficient, the fourth coefficient being used to represent an activation function configuration parameter of a target layer;

[0469] a fifth coefficient, the fifth coefficient being used to represent a normalization type or parameter.

[0470] Optionally, the type of the model parameter satisfies at least one of:

[0471] for a target AI model containing a full connection, the type of the model parameter includes at least one of: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient;

[0472] For the target AI model containing convolution, the type of the model parameter includes at least one of the following: a first coefficient, a second coefficient, a third coefficient, a fourth coefficient, and a fifth coefficient.

[0473] For the target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: a first coefficient, a second coefficient, a third coefficient, a fourth coefficient, and a fifth coefficient.

[0474] For the target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: a first coefficient, a second coefficient, a fourth coefficient, and a fifth coefficient.

[0475] Optionally, the model parameter includes a name of the parameter, or the model parameter includes the name of the parameter and a value of the parameter.

[0476] Optionally, the model parameter is at least part of all model parameters corresponding to the target AI model.

[0477] Optionally, before the first receiving module 801 receives the model parameter of the target AI model sent by the first communication device, the apparatus further includes at least one of the following:

[0478] The fourth sending module is configured to send reporting indication information to the first communication device, and the reporting indication information is used to indicate at least one of the following: supporting partial update, supporting full update, and not supporting partial update.

[0479] The third receiving module is configured to receive a model parameter list sent by the first communication device.

[0480] The fifth sending module is configured to send a model parameter list that needs to be updated to the first communication device.

[0481] Optionally, after the fifth sending module sends the model parameter list that needs to be updated to the first communication device, the apparatus further includes:

[0482] The fourth receiving module is further configured to receive indication information sent by the first communication device, and the indication information is used to indicate whether the first communication device supports the model parameter list.

[0483] Optionally, the target AI model includes at least one of the following:

[0484] A first AI model used by the first communication device and the second communication device;

[0485] A reference AI model of the first AI model;

[0486] a second AI model used in the test by the first communication device, the second communication device, or a test device;

[0487] a reference AI model of the second AI model;

[0488] a third AI model used by the first communication device, the second communication device, or a test device to match the second AI model used in the test;

[0489] a reference AI model of the third AI model.

[0490] Optionally, the target AI model is used to implement at least one of the following:

[0491] reference signal processing;

[0492] channel signal transmission;

[0493] channel signal reception;

[0494] channel signal demodulation;

[0495] channel signal sending;

[0496] channel state information acquisition;

[0497] beam management;

[0498] channel prediction;

[0499] channel coding and decoding;

[0500] source coding and decoding;

[0501] interference suppression;

[0502] positioning;

[0503] prediction and management of higher layer services and parameters;

[0504] analysis of control signaling.

[0505] The model parameter transmission apparatus provided by the embodiments of the present application can implement each process implemented by the method embodiment of FIG. 6 and achieve the same technical effects. To avoid repetition, details are not described herein.

[0506] Optionally, as shown in FIG. 9, the embodiment of the present application further provides a communication device 900, comprising a processor 901 and a memory 902, wherein the memory 902 stores programs or instructions executable on the processor 901. For example, when the communication device 900 is a first communication device, the programs or instructions, when executed by the processor 901, implement each step of the model parameter transmission method embodiment described above and achieve the same technical effects. When the communication device 900 is a second communication device, the programs or instructions, when executed by the processor 901, implement each step of the model parameter transmission method embodiment described above and achieve the same technical effects. To avoid repetition, details are not described herein.

[0507] Preferably, the embodiment of the present application further provides a communication device, which is a terminal, comprising a processor, a memory, programs or instructions stored on the memory and executable on the processor, wherein the programs or instructions, when executed by the processor, implement each process of the communication processing method embodiment described above and achieve the same technical effects. The terminal can be the model parameter transmission apparatus shown in FIG. 7. Specifically, FIG. 10 is a schematic diagram of a hardware structure of a terminal implementing the embodiment of the present application.

[0508] The terminal 1000 includes, but is not limited to, at least part of the components such as a radio frequency unit 1001, a network module 1002, an audio output unit 1003, an input unit 1004, a sensor 1005, a display unit 1006, a user input unit 1007, an interface unit 1008, a memory 1009, and a processor 1010.

[0509] Those skilled in the art can understand that the terminal 1000 can further include a power supply (such as a battery) for supplying power to each component, and the power supply can be logically connected to the processor 1010 through a power management system, so as to realize functions such as management of charging, discharging, and power consumption management through the power management system. The terminal structure shown in FIG. 10 does not constitute a limitation on the terminal, and the terminal can include more or fewer components than those shown, or combine certain components, or different component arrangements, which are not described herein.

[0510] It should be understood that in the embodiments of the present application, the input unit 1004 can include a graphics processor 10041 and a microphone 10042, and the graphics processor 10041 processes image data of a still picture or a video obtained by an image capture device (such as a camera) in a video capture mode or an image capture mode. The display unit 1006 can include a display panel 10061, which can be configured in the form of a liquid crystal display, an organic light-emitting diode, etc. The user input unit 1007 includes at least one of a touch panel 10071 and other input devices 10072. The touch panel 10071 is also called a touch screen. The touch panel 10071 can include two parts of a touch detection device and a touch controller. The other input devices 10072 can include, but are not limited to, a physical keyboard, function keys (such as volume control keys, on-off keys, etc.), a trackball, a mouse, a joystick, and the like, which will not be described here.

[0511] In the embodiments of the present application, after the radio frequency unit 1001 receives the downlink data from the access network device, it can be transmitted to the processor 1010 for processing. In addition, the radio frequency unit 1001 can send uplink data to the network side device. Generally, the radio frequency unit 1001 includes, but is not limited to, an antenna, an amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, etc.

[0512] The memory 1009 can be used to store software programs or instructions and various data. The memory 1009 can mainly include a first storage area storing programs or instructions and a second storage area storing data, wherein the first storage area can store an operating system, application programs or instructions required by at least one function (such as a sound playing function, an image playing function, etc.), and the like. In addition, the memory 1009 can include a volatile memory or a non-volatile memory, or the memory 1009 can include both volatile and non-volatile memories. The non-volatile memory can be a Read-Only Memory (ROM), a Programmable ROM (PROM), an Erasable PROM (EPROM), an Electrically EPROM (EEPROM), or a flash memory. The volatile memory can be a Random Access Memory (RAM), a Static RAM (SRAM), a Dynamic RAM (DRAM), a Synchronous DRAM (SDRAM), a Double Data Rate SDRAM (DDR SDRAM), an Enhanced SDRAM (ESDRAM), a Synch link DRAM (SLDRAM), and a Direct Rambus RAM (DRRAM). The memory 1009 in the embodiments of the present application includes but is not limited to these and any other suitable types of memories.

[0513] The processor 1010 can include one or more processing units; optionally, the processor 1010 integrates an application processor and a modem processor, wherein the application processor mainly processes operations related to an operating system, a user interface, and an application program, and the modem processor mainly processes wireless communication signals, such as a baseband processor. It can be understood that the above-mentioned modem processor can also not be integrated into the processor 1010.

[0514] The processor 1010 is configured to determine the model parameters to be sent according to a target artificial intelligence (AI) model.

[0515] The radio frequency unit 1001 is configured to send the model parameters to a second communication device.

[0516] Optionally, the target AI model is described by at least one of the following:

[0517] Model inference formula

[0518] Model structure description

[0519] Code

[0520] Optionally, the target object of the target AI model is obtained through at least one of the following:

[0521] Protocol agreement

[0522] Offline negotiation

[0523] Among them, the target object includes at least one of the following:

[0524] Model structure

[0525] Name of model parameter

[0526] Type of model parameter

[0527] Description of model parameter

[0528] Format of model parameter

[0529] Operation corresponding to the model parameter

[0530] Optionally, the type of the model parameter includes at least one of the following:

[0531] First coefficient, the first coefficient is used to represent the coefficient that the input of the target layer needs to multiply;

[0532] Second coefficient, the second coefficient is used to represent the coefficient that the input of the target layer needs to add after the target operation;

[0533] Third coefficient, the third coefficient is used to represent the coefficient that the input of the target layer needs to pass through convolution;

[0534] Fourth coefficient, the fourth coefficient is used to represent the activation function configuration parameter of the target layer;

[0535] Fifth coefficient, the fifth coefficient is used to represent the normalization type or parameter.

[0536] Optionally, the type of the model parameter satisfies at least one of the following:

[0537] For the target AI model containing full connection, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, fourth coefficient, fifth coefficient;

[0538] For the target AI model containing convolution, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, fifth coefficient;

[0539] For the target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: a first coefficient, a second coefficient, a third coefficient, a fourth coefficient, and a fifth coefficient.

[0540] For the target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: a first coefficient, a second coefficient, a fourth coefficient, and a fifth coefficient.

[0541] Optionally, the radio frequency unit 1001 is configured to:

[0542] The model parameter is transmitted to the second communication device in the form of a matrix and / or a vector.

[0543] Optionally, the radio frequency unit 1001 is configured to implement at least one of the following:

[0544] In a case where the model parameter includes a first coefficient, the first coefficient of at least one layer in the target AI model is represented in the form of a vector or a matrix, the first coefficient is transmitted to the second communication device, all first coefficients of at least one layer are combined into a matrix, or all first coefficients of at least one layer are combined by a plurality of first coefficient vectors, and the first coefficient vector is a vector composed of first coefficients associated with a target input or a target output;

[0545] In a case where the model parameter includes a second coefficient, the second coefficient of at least one layer in the target AI model is represented in the form of a vector or a matrix, the second coefficient is transmitted to the second communication device, all second coefficients of at least one layer are combined into at least one matrix or at least one vector;

[0546] In a case where the model parameter includes a third coefficient, the third coefficient in a target condition is represented in the form of a matrix, the third coefficient is transmitted to the second communication device, and all third coefficients of at least one layer are combined into at least one matrix, the target condition being all inputs to all outputs of a target layer in the target AI model or a target layer to a next layer of the target layer in the target AI model.

[0547] Optionally, the second coefficient of at least one layer in the target AI model is represented in the form of a vector or a matrix, including at least one of the following:

[0548] The second coefficient of at least one layer in the target AI model is represented by a three-dimensional matrix, and three dimensions of the three-dimensional matrix are represented by N, M1, and M2.

[0549] The second coefficient of at least one layer in the target AI model is represented by N two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by M1 and M2.

[0550] The second coefficient of at least one layer in the target AI model is represented by M1*M2 vectors, the vectors include N elements.

[0551] Wherein, N is the number of all feature maps or all neurons or all outputs of the target layer, M1 is one dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer, and M2 is another dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer.

[0552] Optionally, the third coefficient in the case of the target is represented by a matrix in the following at least one way:

[0553] The third coefficient in the case of the target is represented by a four-dimensional matrix, four dimensions of the four-dimensional matrix are represented by K1, K2, N1 and N2.

[0554] The third coefficient in the case of the target is represented by N2 three-dimensional matrices, three dimensions of the three-dimensional matrices are represented by N1, K1 and K2.

[0555] The third coefficient in the case of the target is represented by N1 three-dimensional matrices, three dimensions of the three-dimensional matrices are represented by N2, K1 and K2.

[0556] The third coefficient in the case of the target is represented by N1*N2 two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by K1 and K2.

[0557] The third coefficient in the case of the target is represented by K1*K2 two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by N1 and N2.

[0558] Wherein, N1 is the number of input channels of the target layer, N2 is the number of output channels of the target layer, K1 is one dimension of the two-dimensional dimension of the convolution kernel, and K2 is another dimension of the two-dimensional dimension of the convolution kernel.

[0559] Optionally, the model parameter includes: the name of the parameter; or the model parameter includes: the name of the parameter and the value of the parameter.

[0560] Optionally, the radio frequency unit 1001 is used for:

[0561] The model parameter is sent to the second communication device in a target order.

[0562] Wherein, the target order includes at least one of the following:

[0563] An order of model parameter names in a model structure description;

[0564] An order of model parameter names in a model inference formula;

[0565] An order of model parameter names in code.

[0566] Optionally, the order of model parameter names in the model structure description comprises at least one of the following:

[0567] In an order of layers in the target AI model, and in an order of neurons in a layer;

[0568] In an order of layers in the target AI model, and in an order of similar parameter combinations in a layer.

[0569] Optionally, the order of model parameter names in the model inference formula comprises at least one of the following:

[0570] An order of appearance of model parameter names in the model inference formula;

[0571] An order of calculation of variables corresponding to model parameter names in the model inference formula.

[0572] Optionally, the order of model parameter names in the code comprises:

[0573] An order of appearance of model parameter names in the code;

[0574] An order of calculation of variables corresponding to model parameter names in the code.

[0575] Optionally, the model parameters are at least part of all model parameters corresponding to the target AI model.

[0576] Optionally, the radio frequency unit 1001 is configured to:

[0577] According to at least one of the model parameter list and the reporting indication information, the model parameters are sent to the second communication device;

[0578] The reporting indication information is used to indicate at least one of the following: support partial update, support full update, and do not support partial update.

[0579] Optionally, before the radio frequency unit 1001 is configured to send the model parameters to the second communication device according to at least one of the model parameter list and the reporting indication information, it is further configured to implement at least one of the following:

[0580] The model parameter list is sent to the second communication device;

[0581] The model parameter list that needs to be updated is received and sent by the second communication device.

[0582] Optionally, after the radio frequency unit 1001 receives the model parameter list that needs to be updated sent by the second communication device, the radio frequency unit 1001 is further configured to:

[0583] send indication information to the second communication device, the indication information being used to indicate whether the first communication device supports the model parameter list.

[0584] Optionally, the target AI model comprises at least one of the following:

[0585] a first AI model used by the first communication device and the second communication device;

[0586] a reference AI model of the first AI model;

[0587] a second AI model used by the first communication device, the second communication device or a test device in a test;

[0588] a reference AI model of the second AI model;

[0589] a third AI model used by the first communication device, the second communication device or the test device to match the second AI model used in the test;

[0590] a reference AI model of the third AI model.

[0591] Optionally, the target AI model is used to implement at least one of the following:

[0592] reference signal processing;

[0593] channel signal transmission;

[0594] channel signal reception;

[0595] channel signal demodulation;

[0596] channel signal sending;

[0597] channel state information acquisition;

[0598] beam management;

[0599] channel prediction;

[0600] channel coding and decoding;

[0601] source coding and decoding;

[0602] interference suppression;

[0603] positioning;

[0604] prediction and management of higher layer services and parameters;

[0605] analysis of control signaling.

[0606] The embodiment of the application further provides a communication device, which is a first communication device, comprising a processor and a communication interface, the processor is used to determine model parameters to be sent according to a target artificial intelligence (AI) model; and the communication interface is used to send the model parameters to a second communication device.

[0607] Optionally, the target AI model is described by at least one of the following:

[0608] a model inference formula;

[0609] a model structure description;

[0610] code.

[0611] Optionally, a target object of the target AI model is obtained by at least one of the following:

[0612] an agreement;

[0613] offline negotiation;

[0614] The target object includes at least one of the following:

[0615] a model structure;

[0616] a name of a model parameter;

[0617] a type of a model parameter;

[0618] a description of a model parameter;

[0619] a format of a model parameter;

[0620] an operation corresponding to a model parameter.

[0621] Optionally, the type of the model parameter includes at least one of the following:

[0622] a first coefficient, the first coefficient being used to indicate a coefficient that an input of a target layer needs to be multiplied by;

[0623] a second coefficient, the second coefficient being used to indicate a coefficient that needs to be added after the input of the target layer is subjected to a target operation;

[0624] a third coefficient, the third coefficient being used to indicate a coefficient that the input of the target layer needs to be subjected to convolution;

[0625] a fourth coefficient, the fourth coefficient being used to indicate an activation function configuration parameter of the target layer;

[0626] a fifth coefficient, the fifth coefficient being used to indicate a normalization type or parameter.

[0627] Optionally, the type of the model parameter satisfies at least one of the following:

[0628] For the target AI model containing full connection, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, fourth coefficient, fifth coefficient;

[0629] For the target AI model containing convolution, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, fifth coefficient;

[0630] For the target AI model containing attention, the type of the model parameter includes attention parameter, and the type of the attention parameter includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, fifth coefficient;

[0631] For the target AI model containing multi-layer perceptron mixer or mixer, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, fourth coefficient, fifth coefficient.

[0632] Optionally, the communication interface is configured to:

[0633] The model parameter is sent to the second communication device in the form of a matrix and / or a vector.

[0634] Optionally, the communication interface is configured to implement at least one of the following:

[0635] In the case where the model parameter includes the first coefficient, the first coefficient of at least one layer in the target AI model is represented in the form of a vector or a matrix, the first coefficient is sent to the second communication device, all first coefficients of at least one layer are combined into a matrix or all first coefficients of at least one layer are combined by a plurality of first coefficient vectors, and the first coefficient vector is a vector composed of first coefficients associated with a target input or a target output;

[0636] In the case where the model parameter includes the second coefficient, the second coefficient of at least one layer in the target AI model is represented in the form of a vector or a matrix, the second coefficient is sent to the second communication device, all second coefficients of at least one layer are combined into at least one matrix or at least one vector;

[0637] In the case where the model parameter includes the third coefficient, the third coefficient in a target case is represented in the form of a matrix, the third coefficient is sent to the second communication device, and all third coefficients of at least one layer are combined into at least one matrix, the target case being all inputs to all outputs of a target layer in the target AI model or a target layer to a next layer of the target layer in the target AI model.

[0638] Optionally, the second coefficients of at least one layer in the target AI model are represented by vectors or matrices, including at least one of the following:

[0639] The second coefficients of at least one layer in the target AI model are represented by a three-dimensional matrix, three dimensions of the three-dimensional matrix are represented by N, M1 and M2;

[0640] The second coefficients of at least one layer in the target AI model are represented by N two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by M1 and M2;

[0641] The second coefficients of at least one layer in the target AI model are represented by M1*M2 vectors, the vectors include N elements;

[0642] Wherein, N is the number of all feature maps or all neurons or all outputs of the target layer, M1 is one dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer, and M2 is another dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer.

[0643] Optionally, the third coefficients in the case of the matrix representation of the target include at least one of the following:

[0644] The third coefficients in the case of the target are represented by a four-dimensional matrix, four dimensions of the four-dimensional matrix are represented by K1, K2, N1 and N2;

[0645] The third coefficients in the case of the target are represented by N2 three-dimensional matrices, three dimensions of the three-dimensional matrices are represented by N1, K1 and K2;

[0646] The third coefficients in the case of the target are represented by N1 three-dimensional matrices, three dimensions of the three-dimensional matrices are represented by N2, K1 and K2;

[0647] The third coefficients in the case of the target are represented by N1*N2 two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by K1 and K2;

[0648] The third coefficients in the case of the target are represented by K1*K2 two-dimensional matrices, two dimensions of the two-dimensional matrices are represented by N1 and N2;

[0649] Wherein, N1 is the number of input channels of the target layer, N2 is the number of output channels of the target layer, K1 is one dimension of the two-dimensional dimension of the convolution kernel, and K2 is another dimension of the two-dimensional dimension of the convolution kernel.

[0650] Optionally, the model parameters include the name of the parameter; or the model parameters include the name of the parameter and the value of the parameter.

[0651] Optionally, the communication interface is configured to:

[0652] send the model parameters to the second communication device according to a target order;

[0653] The target order comprises at least one of:

[0654] a model parameter name order in a model structure description;

[0655] a model parameter name order in a model inference formula; and

[0656] a model parameter name order in code.

[0657] Optionally, the model parameter name order in the model structure description comprises at least one of:

[0658] an order of layers in a target AI model and an order of neurons in a layer;

[0659] an order of layers in a target AI model and an order of same-type parameter combinations in a layer.

[0660] Optionally, the model parameter name order in the model inference formula comprises at least one of:

[0661] an order of appearance of model parameter names in the model inference formula;

[0662] an order of calculation of variables corresponding to the model parameter names in the model inference formula.

[0663] Optionally, the model parameter name order in the code comprises:

[0664] an order of appearance of model parameter names in the code;

[0665] an order of calculation of variables corresponding to the model parameter names in the code.

[0666] Optionally, the model parameters are at least part of all model parameters corresponding to the target AI model.

[0667] Optionally, the communication interface is configured to:

[0668] send the model parameters to the second communication device according to at least one of a model parameter list and report indication information;

[0669] The report indication information is used to indicate at least one of: supporting partial update, supporting full update, and not supporting partial update.

[0670] Optionally, before the communication interface sends the model parameter list to the second communication device according to at least one of the model parameter list and the reporting indication information, the communication interface is further configured to implement at least one of the following:

[0671] sending the model parameter list to the second communication device;

[0672] receiving the model parameter list that needs to be updated sent by the second communication device.

[0673] Optionally, after the communication interface receives the model parameter list that needs to be updated sent by the second communication device, the communication interface is further configured to:

[0674] sending indication information to the second communication device, the indication information being used to indicate whether the first communication device supports the model parameter list.

[0675] Optionally, the target AI model includes at least one of the following:

[0676] a first AI model used by the first communication device and the second communication device;

[0677] a reference AI model of the first AI model;

[0678] a second AI model used by the first communication device, the second communication device or a test device in testing;

[0679] a reference AI model of the second AI model;

[0680] a third AI model used by the first communication device, the second communication device or the test device to match the second AI model used in testing;

[0681] a reference AI model of the third AI model.

[0682] Optionally, the target AI model is used to implement at least one of the following:

[0683] reference signal processing;

[0684] channel signal transmission;

[0685] channel signal reception;

[0686] channel signal demodulation;

[0687] channel signal sending;

[0688] channel state information acquisition;

[0689] beam management;

[0690] channel prediction;

[0691] channel coding and decoding;

[0692] Source coding;

[0693] Interference suppression;

[0694] Positioning;

[0695] Prediction and management of higher layer services and parameters;

[0696] Parsing of control signaling.

[0697] Specifically, the embodiment of the present application further provides a communication device, which is an access network device, and the access network device can be the model parameter transmission apparatus shown in FIG. 8. As shown in FIG. 11, the access network device 1100 includes an antenna 1101, a radio frequency device 1102, a baseband device 1103, a processor 1104 and a memory 1105. The antenna 1101 is connected with the radio frequency device 1102. In the uplink direction, the radio frequency device 1102 receives information through the antenna 1101, and sends the received information to the baseband device 1103 for processing. In the downlink direction, the baseband device 1103 processes information to be sent, and sends the information to the radio frequency device 1102, and the radio frequency device 1102 processes the received information and sends the information out through the antenna 1101.

[0698] The method performed by the network side device in the above embodiment can be implemented in the baseband device 1103, and the baseband device 1103 includes a baseband processor.

[0699] The baseband device 1103 can include at least one baseband board, and a plurality of chips are arranged on the baseband board, as shown in FIG. 11, one of the chips is a baseband processor, and the baseband processor is connected with the memory 1105 through a bus interface to call programs in the memory 1105 and perform the operations of the network device shown in the above method embodiment.

[0700] The communication device can further include a network interface 1106, for example, a common public radio interface (CPRI).

[0701] Specifically, the access network device 1100 of the embodiment of the present application further includes instructions or programs stored in the memory 1105 and executable on the processor 1104, the processor 1104 calls the instructions or programs in the memory 1105 to perform the method performed by each module shown in FIG. 8, and achieves the same technical effect, and thus the description is omitted here.

[0702] The embodiment of the application further provides a communication device, which is a second communication device, comprising a processor and a communication interface, the communication interface is used for receiving model parameters of a target artificial intelligence (AI) model sent by a first communication device; and the processor is used for applying the target AI model according to the model parameters.

[0703] Optionally, the target AI model is described by at least one of the following:

[0704] a model inference formula;

[0705] a model structure description;

[0706] code.

[0707] Optionally, a target object of the target AI model is obtained by at least one of the following:

[0708] protocol agreement;

[0709] offline negotiation;

[0710] The target object includes at least one of the following:

[0711] a model structure;

[0712] a name of a model parameter;

[0713] a type of a model parameter;

[0714] a description of a model parameter;

[0715] a format of a model parameter;

[0716] an operation corresponding to a model parameter.

[0717] Optionally, the type of the model parameter includes at least one of the following:

[0718] a first coefficient, the first coefficient being used for indicating a coefficient that needs to be multiplied by an input of a target layer;

[0719] a second coefficient, the second coefficient being used for indicating a coefficient that needs to be added after an input of a target layer is subjected to a target operation;

[0720] a third coefficient, the third coefficient being used for indicating a coefficient that needs to be subjected to convolution by an input of a target layer;

[0721] a fourth coefficient, the fourth coefficient being used for indicating an activation function configuration parameter of a target layer;

[0722] a fifth coefficient, the fifth coefficient being used for indicating a normalization type or parameter.

[0723] Optionally, the type of the model parameter satisfies at least one of the following:

[0724] For the target AI model containing full connection, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, the fifth coefficient;

[0725] For the target AI model containing convolution, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, the fifth coefficient;

[0726] For the target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, the fifth coefficient;

[0727] For the target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, the fifth coefficient.

[0728] Optionally, the model parameter includes a name of the parameter, or the model parameter includes the name of the parameter and a value of the parameter.

[0729] Optionally, the model parameter is at least part of all model parameters corresponding to the target AI model.

[0730] Optionally, before receiving the model parameter of the target artificial intelligence AI model sent by the first communication device, the communication interface is further configured to implement at least one of the following:

[0731] Send the reporting indication information to the first communication device, and the reporting indication information is used to indicate at least one of the following: support partial update, support full update, and do not support partial update;

[0732] Receive the model parameter list sent by the first communication device;

[0733] Send the model parameter list that needs to be updated to the first communication device.

[0734] Optionally, after the communication interface sends the model parameter list that needs to be updated to the first communication device, the communication interface is further configured to:

[0735] Receive the indication information sent by the first communication device, and the indication information is used to indicate whether the first communication device supports the model parameter list.

[0736] Optionally, the target AI model includes at least one of the following:

[0737] The first AI model used by the first communication device and the second communication device;

[0738] a reference AI model of the first AI model;

[0739] a second AI model used by the first communication device, the second communication device or the test device in the test;

[0740] a reference AI model of the second AI model;

[0741] a third AI model used by the first communication device, the second communication device or the test device to match the second AI model used in the test;

[0742] a reference AI model of the third AI model.

[0743] Optionally, the target AI model is used to implement at least one of the following:

[0744] reference signal processing;

[0745] channel signal transmission;

[0746] channel signal reception;

[0747] channel signal demodulation;

[0748] channel signal sending;

[0749] channel state information acquisition;

[0750] beam management;

[0751] channel prediction;

[0752] channel coding and decoding;

[0753] source coding and decoding;

[0754] interference suppression;

[0755] positioning;

[0756] prediction and management of higher layer services and parameters;

[0757] analysis of control signaling.

[0758] The embodiments of the present application also provide a readable storage medium, the readable storage medium stores a program or instructions, the program or instructions are executed by a processor to implement various processes of the above-mentioned model parameter transmission method embodiments, and the same technical effects can be achieved. To avoid repetition, details are not described here.

[0759] The processor is the processor in the terminal or the network side device in the above-mentioned embodiments. The readable storage medium can be non-volatile and non-transient. The readable storage medium can include a computer readable storage medium, such as a computer readable only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0760] The embodiment of the present application further provides a chip, which comprises a processor and a communication interface, wherein the communication interface is coupled with the processor, the processor is used to run programs or instructions, and realizes each process of the model parameter transmission method embodiment and achieves the same technical effects. To avoid repetition, details are not described herein.

[0761] It should be understood that the chip mentioned in the embodiment of the present application can also be referred to as a system chip, a system chip, a chip system or a system on chip, etc.

[0762] The embodiment of the present application further provides a computer program / program product, which is stored in a storage medium, and is executed by at least one processor to realize each process of the model parameter transmission method embodiment and achieve the same technical effects. To avoid repetition, details are not described herein.

[0763] The embodiment of the present application further provides a communication system, which comprises a first communication device and a second communication device, wherein the first communication device can be used to execute the steps of the model parameter transmission method, and the second communication device can be used to execute the steps of the model parameter transmission method.

[0764] It should be noted that in this document, the term "comprising" or "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not explicitly listed or inherent to such a process, method, article or device. Without more limitations, the element defined by the statement "comprising a" does not exclude the presence of additional identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present application is not limited to the order of performing the functions shown or discussed, but can also include performing the functions in a substantially simultaneous manner or in a reverse order, for example, the described method can be performed in an order different from that described, and various steps can be added, omitted or combined. In addition, the features described with reference to certain examples can be combined in other examples.

[0765] Through the description of the above embodiments, those skilled in the art can clearly understand that the above-mentioned example methods can be realized by means of a computer software product and a general hardware platform as necessary, and of course can also be realized by hardware. The computer software product is stored in a storage medium (such as a ROM, a RAM, a magnetic disc, an optical disc, etc.), and includes a plurality of instructions for enabling a terminal or a network side device to execute the method described in each embodiment of the present application.

[0766] The embodiments of the present application are described above in combination with the drawings, but the present application is not limited to the specific embodiments described above, and the specific embodiments described above are merely illustrative rather than limiting. Those skilled in the art can make many forms of embodiments under the inspiration of the present application without departing from the scope of the present application and the scope protected by the claims, and these embodiments all belong to the protection of the present application.

Claims

1. A model parameter transmission method, comprising: determining, by a first communication device, model parameters to be sent according to a target artificial intelligence (AI) model; sending, by the first communication device, the model parameters to a second communication device.

2. The method of claim 1, wherein, The target AI model is described by at least one of the following: a model inference formula; a model structure description; code.

3. The method of claim 1 or 2, wherein, The target object of the target AI model is obtained by at least one of the following: protocol agreement; offline negotiation; The target object includes at least one of the following: model structure; name of the model parameter; type of the model parameter; description of the model parameter; format of the model parameter; operation corresponding to the model parameter.

4. The method of claim 1 or 2, wherein, The type of the model parameter includes at least one of the following: a first coefficient, the first coefficient being used to indicate a coefficient that an input of a target layer needs to be multiplied by; a second coefficient, the second coefficient being used to indicate a coefficient that an input of a target layer needs to be added after a target operation; a third coefficient, the third coefficient being used to indicate a coefficient that an input of a target layer needs to be convolved by; a fourth coefficient, the fourth coefficient being used to indicate an activation function configuration parameter of a target layer; a fifth coefficient, the fifth coefficient being used to indicate a normalization type or parameter.

5. The method of claim 4, wherein, The type of the model parameter satisfies at least one of the following: For a target AI model containing a full connection, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, fourth coefficient, and fifth coefficient; For a target AI model containing a convolution, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, and fifth coefficient; For a target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: first coefficient, second coefficient, third coefficient, fourth coefficient, and fifth coefficient; For a target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: first coefficient, second coefficient, fourth coefficient, and fifth coefficient.

6. The method according to any one of claims 1 to 5, wherein, The sending of the model parameters to the second communication device includes: sending the model parameters to the second communication device in the form of a matrix and / or a vector.

7. The method of claim 6, wherein, The sending of the model parameters to the second communication device in the form of a matrix or a vector includes at least one of the following: In a case where the model parameters include first coefficients, the first coefficients of at least one layer in the target AI model are represented in the form of a vector or a matrix, the first coefficients are sent to the second communication device, all first coefficients of at least one layer are combined into a matrix, or all first coefficients of at least one layer are combined into a plurality of first coefficient vectors, the first coefficient vector being a vector composed of first coefficients associated with a target input or a target output; In a case where the model parameters include second coefficients, the second coefficients of at least one layer in the target AI model are represented in the form of a vector or a matrix, the second coefficients are sent to the second communication device, all second coefficients of at least one layer are combined into at least one matrix or at least one vector. In the case that the model parameters include third coefficients, the third coefficients in a target case are represented by a matrix, the third coefficients are sent to the second communication device, all third coefficients of at least one layer are combined into at least one matrix, and the target case is all inputs to all outputs of a target layer in the target AI model or a target layer to a next layer of the target layer in the target AI model.

8. The method of claim 7, wherein, The second coefficients of at least one layer in the target AI model are represented by a vector or a matrix, and the representation includes at least one of the following: The second coefficients of at least one layer in the target AI model are represented by a three-dimensional matrix, and the three dimensions of the three-dimensional matrix are represented by N, M1 and M2. The second coefficients of at least one layer in the target AI model are represented by N two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by M1 and M2. The second coefficients of at least one layer in the target AI model are represented by M1*M2 vectors, and the vector includes N elements. Wherein, N is the number of all feature maps or all neurons or all outputs of the target layer, M1 is one dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer, and M2 is another dimension of the two-dimensional input dimension or the two-dimensional output dimension of the feature map or the neuron of the target layer.

9. The method of claim 7, wherein, The third coefficients in the target case are represented by a matrix, and the representation includes at least one of the following: The third coefficients in the target case are represented by a four-dimensional matrix, and the four dimensions of the four-dimensional matrix are represented by K1, K2, N1 and N2. The third coefficients in the target case are represented by N2 three-dimensional matrices, and the three dimensions of the three-dimensional matrix are represented by N1, K1 and K2. The third coefficients in the target case are represented by N1 three-dimensional matrices, and the three dimensions of the three-dimensional matrix are represented by N2, K1 and K2. The third coefficients in the target case are represented by N1*N2 two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by K1 and K2. The third coefficients in the target case are represented by K1*K2 two-dimensional matrices, and the two dimensions of the two-dimensional matrix are represented by N1 and N2. Wherein, N1 is the number of input channels of the target layer, N2 is the number of output channels of the target layer, K1 is one dimension of the two-dimensional dimension of the convolution kernel, and K2 is another dimension of the two-dimensional dimension of the convolution kernel.

10. The method of any one of claims 1-9, wherein, The model parameters include: the name of the parameter; or, the model parameters include: the name of the parameter and the value of the parameter.

11. The method according to any one of claims 1-9, wherein, The model parameters are sent to the second communication device, including: The first communication device sends the model parameters to the second communication device in a target order. Wherein, the target order includes at least one of the following: The order of the model parameter names in the model structure description; The order of the model parameter names in the model inference formula; The order of the model parameter names in the code.

12. The method of claim 11, wherein, The order of the model parameter names in the model structure description includes at least one of the following: In the order of layers in the target AI model, in the order of neurons in the layer; In the order of layers in the target AI model, in the order of the same parameter combination in the layer.

13. The method of claim 11, wherein, The model parameter name order of the model inference formula includes at least one of the following: The order of appearance of the model parameter name in the model inference formula; The order of calculation of the variable corresponding to the model parameter name in the model inference formula.

14. The method of claim 11, wherein, The model parameter name order in the code includes: The order of appearance of the model parameter name in the code; The order of calculation of the variable corresponding to the model parameter name in the code.

15. The method of claim 1, wherein, The model parameter is at least part of the model parameters corresponding to the target AI model.

16. The method of claim 1, wherein, The method of sending the model parameter to the second communication device includes: The first communication device sends the model parameter to the second communication device according to at least one of the model parameter list and the reporting indication information; Wherein, the reporting indication information is used to indicate at least one of the following: support partial update, support full update, do not support partial update.

17. The method of claim 16, wherein, Before the first communication device sends the model parameter to the second communication device according to at least one of the model parameter list and the reporting indication information, the method further includes at least one of the following: The first communication device sends the model parameter list to the second communication device; The first communication device receives the model parameter list sent by the second communication device.

18. The method of claim 17, wherein, After the first communication device receives the model parameter list sent by the second communication device, the method further includes: The first communication device sends the indication information to the second communication device, which is used to indicate whether the first communication device supports the model parameter list.

19. The method of any one of claims 1-18, wherein, The target AI model includes at least one of the following: The first AI model used by the first communication device and the second communication device; The reference AI model of the first AI model; The second AI model used by the first communication device, the second communication device or the test device in the test; The reference AI model of the second AI model; The third AI model used by the first communication device, the second communication device or the test device to match the second AI model used in the test; The reference AI model of the third AI model.

20. The method of any one of claims 1-19, wherein, The target AI model is used to implement at least one of the following: Reference signal processing; Channel signal transmission; Channel signal reception; Channel signal demodulation; Channel signal sending; Channel state information acquisition; Beam management; Channel prediction; Channel coding and decoding; Source coding and decoding; Interference suppression; Positioning; Prediction and management of high layer service and parameters; Analysis of control signaling.

21. A model parameter transmission method, comprising: The second communication device receives the model parameter of the target artificial intelligence AI model sent by the first communication device; The second communication device applies the target AI model according to the model parameter.

22. The method of claim 21, wherein, The target AI model is described by at least one of the following: Model inference formula; Model structure description; Code.

23. The method of claim 21 or 22, wherein, The target object of the target AI model is obtained by at least one of the following: Protocol agreement; Offline negotiation; Wherein, the target object includes at least one of the following: Model structure; The name of the model parameter; The type of the model parameter; The description of the model parameter; The format of the model parameter; The operation corresponding to the model parameter.

24. The method of claim 21 or 22, wherein, The type of the model parameter includes at least one of the following: A first coefficient, which is used to represent a coefficient that an input of a target layer needs to be multiplied by; A second coefficient, which is used to represent a coefficient that an input of a target layer needs to be added after a target operation; A third coefficient, which is used to represent a coefficient that an input of a target layer needs to be convolved by; A fourth coefficient, which is used to represent an activation function configuration parameter of a target layer; A fifth coefficient, which is used to represent a normalization type or parameter.

25. The method of claim 24, wherein, The type of the model parameter satisfies at least one of the following: For a target AI model containing a full connection, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient; For a target AI model containing a convolution, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient; For a target AI model containing attention, the type of the model parameter includes an attention parameter, and the type of the attention parameter includes at least one of the following: the first coefficient, the second coefficient, the third coefficient, the fourth coefficient, and the fifth coefficient; For a target AI model containing a multi-layer perceptron mixer or a mixer, the type of the model parameter includes at least one of the following: the first coefficient, the second coefficient, the fourth coefficient, and the fifth coefficient.

26. The method of any one of claims 21-25, wherein, The model parameter includes a name of the parameter, or the model parameter includes the name of the parameter and a value of the parameter.

27. The method of claim 21, wherein, The model parameter is at least part of all model parameters corresponding to the target AI model.

28. The method of claim 21, wherein, Before receiving the model parameter of the target AI model sent by the first communication device, the method further includes at least one of the following: The second communication device sends reporting indication information to the first communication device, and the reporting indication information is used to indicate at least one of the following: supporting partial update, supporting full update, and not supporting partial update; The second communication device receives a model parameter list sent by the first communication device; The second communication device sends a model parameter list that needs to be updated to the first communication device.

29. The method of claim 28, wherein, After the model parameter list that needs to be updated is sent to the first communication device, the method further includes: The second communication device receives indication information sent by the first communication device, and the indication information is used to indicate whether the first communication device supports the model parameter list.

30. The method of any one of claims 21-29, wherein, The target AI model includes at least one of the following: A first AI model used by the first communication device and the second communication device; A reference AI model of the first AI model; A second AI model used by the first communication device, the second communication device, or a test device in testing; A reference AI model of the second AI model; A third AI model used by the first communication device, the second communication device, or a test device to match the second AI model used in testing; A reference AI model of the third AI model.

31. The method of any one of claims 21-29, wherein, The target AI model is used to implement at least one of the following: Reference signal processing; Channel signal transmission; Channel signal reception; Channel signal demodulation; Channel signal sending; Channel state information acquisition; Beam management; Channel prediction; Channel coding and decoding; Source coding and decoding; Interference suppression; Positioning; Prediction and management of higher layer services and parameters; Parsing of control signaling.

32. A model parameter transmission apparatus applied to a first communication device, comprising: a first processing module configured to determine model parameters to be sent according to a target artificial intelligence (AI) model; a first sending module configured to send the model parameters to a second communication device.

33. A communication device, which is a first communication device, comprising a processor and a memory, the memory storing programs or instructions executable on the processor, and the programs or instructions, when executed by the processor, implement the steps of the model parameter transmission method according to any one of claims 1 to 20.

34. A model parameter transmission apparatus applied to a second communication device, comprising: a first receiving module configured to receive model parameters of a target artificial intelligence (AI) model sent by a first communication device; a second processing module configured to apply the target AI model according to the model parameters.

35. A communication device, which is a second communication device, comprising a processor and a memory, the memory storing programs or instructions executable on the processor, and the programs or instructions, when executed by the processor, implement the steps of the model parameter transmission method according to any one of claims 21 to 31.

36. A readable storage medium, the readable storage medium storing programs or instructions, and the programs or instructions, when executed by a processor, implement the steps of the model parameter transmission method according to any one of claims 1 to 31.

37. A computer program product comprising computer instructions, and the computer instructions, when executed by a processor, implement the steps of the method according to any one of claims 1 to 31.

Images