WO2023231635A1 - 一种模型传输的方法及装置 - Google Patents

一种模型传输的方法及装置 Download PDF

Info

Publication number
WO2023231635A1
WO2023231635A1 PCT/CN2023/089764 CN2023089764W WO2023231635A1 WO 2023231635 A1 WO2023231635 A1 WO 2023231635A1 CN 2023089764 W CN2023089764 W CN 2023089764W WO 2023231635 A1 WO2023231635 A1 WO 2023231635A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
signals
models
information
training
Prior art date
Application number
PCT/CN2023/089764
Other languages
English (en)
French (fr)
Inventor
胡斌
张公正
王坚
李榕
Original Assignee
华为技术有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 华为技术有限公司 filed Critical 华为技术有限公司
Publication of WO2023231635A1 publication Critical patent/WO2023231635A1/zh

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/0413—MIMO systems
    • H04B7/0456—Selection of precoding matrices or codebooks, e.g. using matrices antenna weighting
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04B—TRANSMISSION
    • H04B7/00—Radio transmission systems, i.e. using radiation field
    • H04B7/02—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas
    • H04B7/04—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas
    • H04B7/06—Diversity systems; Multi-antenna system, i.e. transmission or reception using multiple antennas using two or more spaced independent antennas at the transmitting station

Definitions

  • the embodiments of the present application relate to the field of communication technology, and specifically relate to a method and device for model transmission.
  • ANN Artificial neural networks
  • ANN is a nonlinear statistical data modeling tool that simulates biological neural networks for information processing.
  • artificial neural networks can change the internal structure based on external information and are an adaptive system with learning capabilities.
  • This type of network relies on the complexity of the system to process information by adjusting the interconnected relationships between a large number of internal nodes. It has been widely used in fields such as image classification, face recognition, and speech recognition.
  • AI artificial intelligence
  • in-vehicle mobile devices such as in-vehicle mobile devices, drones, smartphones, personal digital assistants
  • neural network models are also becoming more and more popular. The more it follows, the more likely it will be necessary to transfer multiple neural network models for the same client.
  • Figure 1 there is a known communication technology, as shown in Figure 1, which regards a neural network model as a data to be transmitted, encodes the data to generate a corresponding signal, and sends it through a wireless link to implement the neural network Model transfer.
  • this transmission process does not combine the communication process with the training process of artificial intelligence applications, and multiple neural network models are transmitted independently of each other, which results in low transmission performance of the neural network models and is difficult to meet the needs of current business.
  • Embodiments of the present application provide a method and device for model transmission, which can realize joint transmission of multiple models, reduce the resource overhead and delay of the multiple models in the transmission process, and at the same time reduce the need for the target neural network model to be processed by the receiving end device. of storage space.
  • a method of model transmission is provided, applied to a sending end device or a chip in the sending end device.
  • the method includes: acquiring information of N first models, where the N first models correspond to Nth A task, where N is an integer greater than or equal to 2; obtaining fusion auxiliary information, the fusion auxiliary information includes external feature values of the target model, the target model corresponds to the second task, and the second task is different from the first task; according to The information of N first models and the fusion auxiliary information determine N first signals; and send the N first signals.
  • the method can be applied to the sending end device or the chip or chip system on the sending end device side.
  • the sending end device obtains the information and fusion auxiliary information of N first models, and processes the information and fusion auxiliary information of the N first models according to the first module to obtain N first signal, and then sends the N first signals. Then, the receiving end device can receive the second signal corresponding to the target model. The second signal is obtained by superposing the N first signals in the channel.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the N first signals are obtained by processing the information of the N first models and the fusion auxiliary information by the first module in the sending end device.
  • the first module is trained based on training data
  • the training data includes M first training signals and a target model
  • the M first training signals are M training models correspond one to one
  • the target model can satisfy the tasks corresponding to the M training models, where M is an integer greater than or equal to 2.
  • the first module is determined according to parameters of a channel that sends N first signals.
  • the parameters of the above channel include one or more of the following: channel state information (channel state information, CSI), signal-to-noise ratio (signal-to-noise ratio, SNR), and link quality.
  • the channel status information may include: channel quality indicator (channel quality indicator, CQI), rank indicator (rank indicator, RI), and precoding matrix (precoder matrix, PMI).
  • the first module corresponding to the i-th model among the N first models is based on the i-th model among the N first models except the first model. Determined by information other than at least one first model.
  • the method of transmitting the model further includes: acquiring M original models, the M original models corresponding to the M first training signals one-to-one, and M being greater than or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal, Among them, the third training signal corresponds to the intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the original adjusted models. model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the first module is obtained from the model library.
  • the first model can be obtained through model training, which is beneficial to realizing the superposition of N first signals in the channel, and thereby obtaining the second signal corresponding to the target model.
  • the method of model transmission further includes: sending the fusion auxiliary information.
  • the fusion auxiliary information comes from the receiving end devices of the N first signals, or the fusion auxiliary information comes from the receiving end devices of the N first signals. service server.
  • the information of the N first models is carried in N signals A, and the N first models and the N signals A corresponds one to one.
  • the process of the N first modules processing the N signals A corresponding to the N first models may include one or more of the following methods.
  • the N signals A may be analog signals or digital signals.
  • the N signals A are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A undergo corresponding N processes are those that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing. Signal.
  • the N first signals may be processed through subsequent source coding, channel coding, modulation, and frequency upconversion or downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding processing.
  • the N first signals obtained after the N first signals are processed by the corresponding N first modules are signals obtained by source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals can be processed through subsequent up-conversion or down-conversion and transmitted through the same channel. Send the processed N first signals.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals processed by source coding and channel coding.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals are the signals of A after source coding, channel coding and modulation processing.
  • the N second signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. get the signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • a method of model transmission is provided, which is applied to a transmitting end device or a chip in the transmitting end device.
  • the method includes: acquiring N first signals, and the N first signals are related to the N first signals.
  • the M first training signals correspond to M training models.
  • Each first training signal is a signal of the corresponding training model.
  • the target model can satisfy the corresponding requirements of the M training models.
  • M is an integer greater than or equal to 2
  • the M services corresponding to the M first training models include the N services corresponding to the N first models, and the first model and the second model corresponding to each other correspond to the same Business; process the N first signals according to the N second models to obtain N second signals, wherein the n-th second signal among the N second signals is processed according to the n-th second model.
  • the signal obtained by processing the n-th first signal which is the signal of the first model corresponding to the n-th second model, n ⁇ [1, N]; sending the N through the same channel Second signal.
  • the sending end device obtains N first signals corresponding to N first models, processes the N first signals according to the N second models, and obtains N second signals, and then The N second signals are sent through the first channel. Then, the receiving end device can receive the third signal corresponding to the target model. The third signal is obtained by superimposing N second signals in the first channel.
  • the external feature values of the above model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • obtaining the N second models includes: obtaining the N second models according to parameters of the channel that sends the N second signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • obtaining the N second models further includes: based on the information of at least one first model other than the i-th first model among the N first models. , determine the second model corresponding to the i-th first model, i ⁇ [1, N].
  • the method of transmitting the model also includes: obtaining M original models, the M original models correspond to the M first training signals one-to-one, and M is greater than Or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal; The third training signal is analyzed to obtain an intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the adjusted models. The original model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the N second models are obtained from the model library.
  • the second model that meets the task requirements can be obtained through model training, thereby realizing the superposition of N first signals in the channel, and thereby obtaining the second signal corresponding to the target model.
  • parameters of the above model may include weight matrices, weight vectors, bias matrices, and bias vectors.
  • the process of processing the N first signals corresponding to the N first models by the N second models may include one or more of the following methods.
  • the N first signals may be analog signals or digital signals.
  • the N first signals are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are not subjected to source coding, channel coding, modulation, frequency upconversion or downconversion. Frequency conversion processed signal.
  • the N second signals may be processed through subsequent source coding, channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals processed by source coding and channel coding.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals can be processed through subsequent modulation and up-conversion or down-conversion, and passed The processed N second signals are sent on the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding and modulation processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N second signals through the same channel.
  • the method for model transmission further includes: receiving first information, the first information being used to determine the N first models.
  • the first information comes from the receiving end device of the second signal, or the first information comes from a server that provides services to the receiving end device of the second signal.
  • the first information includes one or more of the following: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a method of model transmission is provided, which is applied to a system in which at least two sending end devices transmit models to one receiving end device.
  • the method includes: the first sending end device obtains N first models. Information of S first models, the S first models among the N first models correspond to S first tasks; obtain the first fusion auxiliary information, the first fusion auxiliary information includes the external feature values of the target model, The target model corresponds to a second task, and the second task is different from the S first tasks; the second sending end device obtains information of K first models among the N first models, and the K first models among the N first models K first models correspond to K first tasks; second fusion auxiliary information is obtained.
  • the second fusion auxiliary information includes the external feature values of the target model.
  • the target model corresponds to the second task.
  • the second task is related to the K-th One task is different, the intersection of the S first models and the K first models is an empty set; the S first signals and the K first signals are sent. It can be understood that if there are only two sending end devices, then the N first models are the union of the S first models and the K first models.
  • the device may be a device or a chip in the device.
  • the information of the S first models and the first fusion auxiliary information corresponding to the end device are processed by the corresponding first module into S first signals, and the information of the K first models corresponding to the second sending end device and the first fusion information are processed by the corresponding first module.
  • the two fused auxiliary information are processed by the corresponding first module into K first signals), and then the N first signals are sent. Then, the receiving end device can receive the second signal corresponding to the target model. The second signal is obtained by superposing the N first signals in the channel.
  • the N first signals are obtained by the first module after processing the information of the N first models and the fusion auxiliary information.
  • the above-mentioned S first signals are the first sending end device transmitting the S first models through the corresponding first module.
  • the K first signals are obtained by processing the information of the S first models and the first fusion auxiliary information.
  • the K first signals are obtained by the second sending end device through the corresponding first module. It is obtained after processing the information of the first model and the second fusion auxiliary information.
  • the first fusion auxiliary information and the second fusion auxiliary information may be the same or different.
  • both the first fusion auxiliary information and the second fusion auxiliary information can be used to represent the external feature values of the target model.
  • the first fusion auxiliary information is different from the second fusion auxiliary information
  • the first fusion auxiliary information represents a first subset of external feature values of the target model
  • the second fusion auxiliary information represents a second subset of external feature values of the target model. subset, the first subset and the second subset are different, and the external feature value of the target model is the union of the first subset and the second subset.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the first module corresponding to the first sending end device or the first module corresponding to the second sending end device is trained based on training data, and the training The data includes M first training signals and target models.
  • the M first training signals correspond to the M training models one-to-one.
  • the target model can satisfy the tasks corresponding to the M training models.
  • M is an integer greater than or equal to 2.
  • the first module corresponding to the first sending end device or the first module corresponding to the second sending end device is based on sending the corresponding first signal or the Nth The parameters of a signal channel are determined.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the first module corresponding to the i-th model among the S first models is based on the i-th model among the S first models except the first model. Letter of at least one first model other than The information is determined; the first module corresponding to the j-th model among the K first models is determined based on the information of at least one first model among the K first models except the j-th model among the first models. of.
  • the method of transmitting the model also includes: obtaining M original models, the M original models correspond one-to-one with the M first training signals, and M is greater than or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal, Among them, the third training signal corresponds to the intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the original adjusted models. model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the first module is obtained from the model library.
  • the method of transmitting the model further includes: sending the fusion auxiliary information.
  • the fusion auxiliary information comes from the receiving end devices of the N first signals, or the fusion auxiliary information comes from the receiving end devices of the N first signals. service server.
  • the information of the N first models is carried in N signals A, and the N first models correspond to the N signals A one-to-one.
  • the process of the N first modules processing the N signals A corresponding to the N first models may include one or more of the following methods.
  • the N signals A may be analog signals or digital signals.
  • the N signals A are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A undergo corresponding N processes are those that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing. Signal.
  • the N first signals may be processed through subsequent source coding, channel coding, modulation, and frequency upconversion or downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N signals A are processed by the corresponding N first modules to obtain
  • the obtained N first signals are signals obtained through source coding, channel coding, modulation, and up-conversion or down-conversion processing. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding processing.
  • the N first signals obtained after the N first signals are processed by the corresponding N first modules are signals obtained by source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals processed by source coding and channel coding.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals are the signals of A after source coding, channel coding and modulation processing.
  • the N second signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N signals A are processed by the corresponding N first modules to obtain
  • the obtained N first signals are signals obtained through source coding, channel coding, modulation, and up-conversion or down-conversion processing. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. get the signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • a method of model transmission is provided, which is applied to a system in which at least two transmitting end devices transmit models to one receiving end device.
  • the method includes: the first transmitting end device acquires S first signals, The S first signals correspond to the S first models one-to-one, and each first signal is used to carry information of the corresponding first model, where the information of the first model includes the external characteristic values of the model and /or parameters of the model, S is an integer greater than or equal to 1; obtain S second models, which correspond to the S first models one-to-one, where the S second models are based on the training data Obtained by training, the training data includes M first training signals and target models.
  • the M first training signals correspond to M training models.
  • Each first training signal is the signal of the corresponding training model.
  • the target model can The services corresponding to the M training models are satisfied, and M is an integer greater than or equal to 2.
  • the M services corresponding to the M first training models include the N services corresponding to the S first models, and the first services corresponding to each other are The model and the second model correspond to the same service; the N first signals are processed according to the S second models to obtain S second signals, wherein the s-th second signal among the S second signals is based on The signal obtained by processing the s-th first signal by the s-th second model.
  • the s-th first signal is the signal of the first model corresponding to the s-th second model, s ⁇ [1, S];
  • the second sending end device acquires K first signals, which correspond to the K first models one-to-one, and each first signal is used to carry information of the corresponding first model, wherein, the The information of the first model includes the external eigenvalues of the model and/or the parameters of the model, and K is an integer greater than or equal to 1; obtain K second models, and the K second models correspond to the K first models one-to-one, Wherein, the K second models are trained based on training data.
  • the training data includes the M first training signals and the target model.
  • the M first training signals correspond to M training models. Each first training signal is the signal of the corresponding training model.
  • the target model can satisfy the services corresponding to the M training models.
  • M is an integer greater than or equal to 2.
  • the M services corresponding to the M first training models include the K first N services corresponding to the model, and the first model and the second model corresponding to each other correspond to the same service; the N first signals are processed according to the K second models to obtain K second signals, where the K The k-th second signal among the second signals is a signal obtained by processing the k-th first signal according to the k-th second model, and the k-th first signal is the k-th second signal corresponding to the k-th second model.
  • the intersection of the K first models and the S first models is the empty set;
  • the S second signals and the K second signals are sent.
  • the second sending end device can obtain K first signals corresponding to K first models among the N first models, and the K first models are related to the S The intersection of the first models is the empty set)
  • the first signal is processed by the corresponding second model and becomes S second signals
  • the K first signals corresponding to the second sending end device are processed by the corresponding second model and becomes K second signals
  • the first channel transmits the N second signals.
  • the receiving end device can receive the third signal corresponding to the target model.
  • the third signal is obtained by superimposing N second signals in the first channel.
  • the external feature values of the above-mentioned model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • obtaining the N second models includes: obtaining the N second models according to parameters of the channel that sends the N second signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • obtaining the N second models further includes: based on the information of at least one first model other than the i-th first model among the N first models. , determine the second model corresponding to the i-th first model, i ⁇ [1, N].
  • the method of transmitting the model also includes: obtaining M original models, the M original models corresponding to the M first training signals one-to-one, and M is greater than Or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal; The third training signal is analyzed to obtain an intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the adjusted models. The original model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the N second models are obtained from the model library.
  • the second model that meets the task requirements is obtained through model training, and N second signals can be superimposed in the channel, and a third signal corresponding to the target model can be further obtained.
  • parameters of the above model may include weight matrices, weight vectors, bias matrices, and bias vectors.
  • the process of processing the N first signals corresponding to the N first models by the N second models may include one or more of the following methods.
  • the N first signals may be analog signals or digital signals.
  • the N first signals are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are not subjected to source coding, channel coding, modulation, frequency upconversion or downconversion. Frequency conversion processed signal.
  • the N second signals may be processed through subsequent source coding, channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals processed by source coding and channel coding.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N first signals are processed by the corresponding N second models
  • the obtained N second signals are signals obtained through source coding, channel coding, modulation, and up-conversion or down-conversion processing. Subsequently, the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding and modulation processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N second signals through the same channel.
  • the method for model transmission further includes: receiving first information, the first information being used to determine the N first models.
  • the first information comes from the receiving end device of the second signal, or the first information comes from a server that provides services to the receiving end device of the second signal.
  • the first information includes one or more of the following: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a method of model transmission is provided, applied to a sending end device or a chip in the sending end device.
  • the method includes: obtaining information of S first models among N first models, and the S first models are The first model corresponds to S first tasks, where N is an integer greater than or equal to 1, and S is an integer less than or equal to N; the first fusion auxiliary information is obtained, and the first fusion auxiliary information includes the external feature value of the target model , the target model corresponds to the second task, and the second task is different from the S first tasks; determine the S first signals according to the information of the S first models and the first fusion auxiliary information; send the S first signals A signal.
  • the sending end device obtains the information of S first models and the first fusion auxiliary information among the N first models, and combines the information of the S first models with the first fusion auxiliary information according to the first module.
  • the auxiliary information is fused and processed to obtain S first signals, and then the S first signals are sent.
  • S is less than N
  • one or more other sending ends will also send the first signal among the N-S first signals, and then the receiving end device can receive the second signal corresponding to the target model.
  • the second signal It is obtained by superposing N first signals including S first signals in the channel.
  • the joint transmission of multiple different neural network models can be achieved, which is also suitable for scenarios where the sending end device is distributed, and can reduce the resource overhead and delay in the transmission process of multiple neural network models, and To a certain extent, the storage space of the receiving end device for the neural network model is reduced.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, model parameters include one or more of the following parameters: weight matrix, weight vector, Bias matrix, bias vector.
  • the first fusion auxiliary information may be the external feature value of the target model, or may be a subset of the external feature value of the target model.
  • the S first signals are obtained by processing the information of the S first models and the first fusion auxiliary information by the first module in the sending end device. of.
  • the first module is trained based on training data
  • the training data includes M first training signals and a target model
  • the M first training signals are M training models correspond one to one
  • the target model can satisfy the tasks corresponding to the M training models, where M is an integer greater than or equal to 2.
  • the first module is determined based on parameters of a channel that sends S first signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the first module corresponding to the i-th model among the S first models is based on the i-th model among the S first models except the first model. Information about at least one first model other than the model is determined.
  • the method for transmitting the model further includes: obtaining M original models, the M original models corresponding to the M first training signals one-to-one, and M being greater than or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal, Among them, the third training signal corresponds to the intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the original adjusted models. model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the first module is obtained from the model library.
  • the method of model transmission further includes: sending the first fusion auxiliary information.
  • the first fusion auxiliary information comes from the receiving end devices of the S first signals, or the first fusion auxiliary information comes from the receiving end devices of the S first signals.
  • the information of the S first models is carried in S signals A, and the S first models correspond to the S signals A one-to-one.
  • the process of the S first modules processing the S signals A corresponding to the S first models may include one or more of the following methods.
  • the S signals A may be analog signals or digital signals.
  • the S signals A are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the S first signals obtained after the S signals A undergo corresponding S processes are not subjected to source coding, channel coding, modulation, up-conversion or down-conversion processing. Signal.
  • the S first signals may be processed through subsequent source coding, channel coding, modulation, and frequency upconversion or downconversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained through source coding processing.
  • the S first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained by source coding and channel coding.
  • the S first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained by source coding, channel coding and modulation processing.
  • the S first signals may be processed through subsequent up-conversion or down-conversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the S first signals through the same channel.
  • the S signals A are signals that have been processed by source coding.
  • the S first signals obtained after the S first signals are processed by the corresponding S first modules are signals obtained by source coding processing.
  • the S first signals may be processed through subsequent channel coding, modulation, and frequency upconversion or downconversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained by source coding and channel coding.
  • the S first signals may be processed through subsequent modulation and frequency upconversion or downconversion, and the processed N first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained by source coding, channel coding and modulation processing.
  • the S first signals may be processed through subsequent up-conversion or down-conversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the S first signals through the same channel.
  • the S signals A are signals processed by source coding and channel coding.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained by source coding and channel coding.
  • the S first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed S first signals may be sent through the same channel.
  • the S signals A are processed by the corresponding S first modules to obtain
  • the S first signals obtained are signals obtained through source coding, channel coding and modulation processing.
  • the S first signals may be processed through subsequent up-conversion or down-conversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after being processed by the corresponding S first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the S first signals through the same channel.
  • the S signals are the signals of A after source coding, channel coding and modulation processing.
  • the S second signals obtained after the S signals A are processed by the corresponding S first modules are signals obtained through source coding, channel coding and modulation processing.
  • the S first signals may be processed through subsequent up-conversion or down-conversion, and the processed S first signals may be sent through the same channel.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the S first signals through the same channel.
  • the S signals A are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the S first signals obtained after the S signals A are processed by the corresponding S first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. get the signal. Subsequently, the sending end device sends the S first signals through the same channel.
  • a sixth aspect provides a method of model transmission, applied to a receiving end device or a chip in the receiving end device.
  • the method includes: receiving a second signal, the second signal is associated with N first models, and the N first models are The first model corresponds to N first tasks; according to the second signal, a target model is obtained, and the target model corresponds to a second task, and the second task is different from the first task.
  • the receiving end device receives the second signal sent by the transmitting end device to obtain the target model.
  • the second signal is obtained by superimposing N first signals in the channel.
  • the method of model transmission further includes: sending fusion auxiliary information, where the fusion auxiliary information includes external feature values of the target model.
  • the external feature values of the model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a method of model transmission is provided, which is applied in a system in which a sending end device transmits a model to a receiving end device.
  • the method includes: the sending end device obtains information of N first models and fusion auxiliary information, Determine N first signals based on the information of the N first models and the fusion auxiliary information, and send the N first signals, where the N first models correspond to N first tasks, and N is greater than or equal to 2 An integer of The tasks are different; the receiving end device receives the second signal sent by the transmitting end device and analyzes the second signal to obtain the target model, where the second signal is obtained from the N first signals.
  • the sending end device obtains the information and fusion auxiliary information of N first models, and processes the information and fusion auxiliary information of the N first models according to the first module to obtain N first signal, and then sends the N first signals. Then, the receiving end device can receive the second signal corresponding to the target model. The second signal is obtained by superposing the N first signals in the channel.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the N first signals are obtained by processing the information of the N first models and the fusion auxiliary information by the first module in the sending end device.
  • the second signal is obtained by superposing the N first signals in the channel.
  • the first module is trained based on training data
  • the training data includes M first training signals and a target model
  • the M first training signals are M training models correspond one to one
  • the target model can satisfy the tasks corresponding to the M training models, where M is an integer greater than or equal to 2.
  • the first module is determined according to parameters of a channel that sends N first signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the first module corresponding to the i-th model among the N first models is based on the i-th model among the N first models except the first model. Determined by information other than at least one first model.
  • the method of model transmission further includes: the sending end device acquires M original models, and inputs each first training signal into the corresponding original model to obtain M second training signals are superimposed in the same channel to obtain a third training signal.
  • the third training signal corresponds to the intermediate model, and according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, and the first module is obtained from the model library.
  • M is an integer greater than or equal to 2.
  • the model library includes the adjusted original model, and the deviation between the adjusted intermediate model and the target model. within the preset range.
  • the first model can be obtained through model training, which is beneficial to realizing the superposition of N first signals in the channel, and thereby obtaining the second signal corresponding to the target model.
  • the method for transmitting the model further includes: sending the fusion auxiliary information.
  • the fusion auxiliary information comes from the receiving end devices of the N first signals, or the fusion auxiliary information comes from the receiving end devices provided by the N first signals. service of service device.
  • the information of the N first models is carried in N signals A, and the N first models correspond to the N signals A one-to-one.
  • the process of the N first modules processing the N signals A corresponding to the N first models may include one or more of the following methods.
  • the N signals A may be analog signals or digital signals.
  • the N signals A are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A undergo corresponding N processes are those that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing. Signal.
  • the N first signals may be processed through subsequent source coding, channel coding, modulation, and frequency upconversion or downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding processing.
  • the N first signals obtained after the N first signals are processed by the corresponding N first modules are signals obtained by source coding processing.
  • the N first signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N signals A are processed by the corresponding N first modules to obtain
  • the obtained N first signals are signals obtained through source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals processed by source coding and channel coding.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding and channel coding.
  • the N first signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained by source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals are the signals of A after source coding, channel coding and modulation processing.
  • the N second signals obtained after the N signals A are processed by the corresponding N first modules are signals obtained through source coding, channel coding and modulation processing.
  • the N first signals may be processed through subsequent up-conversion or down-conversion, and the processed N first signals may be sent through the same channel.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • the N signals A are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N first signals obtained after the N signals A are processed by the corresponding N first modules are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. get the signal. Subsequently, the sending end device sends the N first signals through the same channel.
  • a method of model transmission is provided, applied in a system in which at least two transmitting end devices transmit models to one receiving end device.
  • the method includes: the first transmitting end device acquires S first signals. , the S first signals correspond to the S first models one-to-one, and each first signal is used to carry the information of the corresponding first model, where the information of the first model includes the external characteristic values of the model. and/or model parameters, S is an integer greater than or equal to 1; obtain S second models, the S second models correspond to the S first models one-to-one, where, the S second models is obtained by training based on training data.
  • the training data includes M first training signals and target models.
  • the M first training signals correspond to M training models.
  • Each first training signal is the signal of the corresponding training model.
  • the target model can satisfy the services corresponding to the M training models.
  • M is an integer greater than or equal to 2.
  • the M services corresponding to the M first training models include the N services corresponding to the S first models, and they are mutually exclusive.
  • the corresponding first model and the second model correspond to the same service;
  • the N first signals are processed according to the S second models to obtain S second signals, where the sth of the S second signals is
  • the second signal is a signal obtained by processing the s-th first signal according to the s-th second model.
  • the s-th first signal is the signal of the first model corresponding to the s-th second model, s ⁇ [1 ,S];
  • the second sending end device acquires K first signals, which correspond to the K first models one-to-one, and each first signal is used to carry information of the corresponding first model, wherein, the The information of the first model includes the external eigenvalues of the model and/or the parameters of the model, and K is an integer greater than or equal to 1; obtain K second models, and the K second models correspond to the K first models one-to-one, Wherein, the K second models are trained based on training data.
  • the training data includes the M first training signals and the target model.
  • the M first training signals correspond to M training models. Each first training signal is the signal of the corresponding training model.
  • the target model can satisfy the services corresponding to the M training models.
  • M is an integer greater than or equal to 2.
  • the M services corresponding to the M first training models include the K first N services corresponding to the model, and the first model and the second model corresponding to each other correspond to the same service; the N first signals are processed according to the K second models to obtain K second signals, where the K The k-th second signal among the second signals is a signal obtained by processing the k-th first signal according to the k-th second model, and the k-th first signal is the k-th second signal corresponding to the k-th second model.
  • the intersection of the K first models and the S first models is the empty set;
  • the receiving end device acquires a third signal, which is obtained from N second signals.
  • the second sending end device can obtain K first signals corresponding to K first models among the N first models, and the K first models are related to the S The intersection of the first models is the empty set)
  • the first signals are processed into S second signals after being processed by the corresponding second model, and the K first signals corresponding to the second sending end device are processed into K second signals after being processed by the corresponding second model), and then
  • the N second signals are sent through the first channel.
  • the receiving end device can receive the third signal corresponding to the target model.
  • the third signal is obtained by superimposing N second signals in the first channel.
  • the external feature values of the above-mentioned model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • obtaining the N second models includes: obtaining the N second models according to parameters of a channel that sends the N second signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • obtaining the N second models further includes: according to the The information of at least one first model among the N first models except the i-th first model determines the second model corresponding to the i-th first model, i ⁇ [1, N].
  • the third signal is obtained by superposing N second signals in the channel.
  • the method for transmitting the model further includes: the transmitting end device acquires M original models, and the M original models correspond to the M first training signals one-to-one, M is an integer greater than or equal to 2; the sending end device inputs each first training signal into the corresponding original model to obtain M second training signals; and superimposes the M second training signals in the same channel.
  • a third training signal analyze the third training signal to obtain an intermediate model; adjust the parameters of the M original models according to the deviation between the intermediate model and the target model to obtain a model library, which model The library includes the adjusted original model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the N second models are obtained from the model library.
  • the second model that meets the task requirements is obtained through model training, and N second signals can be superimposed in the channel, and a third signal corresponding to the target model can be further obtained.
  • parameters of the above model may include weight matrices, weight vectors, bias matrices, and bias vectors.
  • the process of processing the N first signals corresponding to the N first models by the N second models may include one or more of the following methods.
  • the N first signals may be analog signals or digital signals.
  • the N first signals are signals that have not undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are not subjected to source coding, channel coding, modulation, frequency upconversion or downconversion. Frequency conversion processed signal.
  • the N second signals may be processed through subsequent source coding, channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Obtained by frequency conversion processing Signal. Subsequently, the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding processing.
  • the N second signals may be processed through subsequent channel coding, modulation, and up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals processed by source coding and channel coding.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding and channel coding.
  • the N second signals may be processed through subsequent modulation and frequency upconversion or frequency downconversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding and modulation processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are signals obtained by source coding, channel coding and modulation processing.
  • the N second signals may be processed through subsequent up-conversion or down-conversion, and the processed N second signals may be sent through the same channel.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Obtained by frequency conversion processing Signal. Subsequently, the sending end device sends the N second signals through the same channel.
  • the N first signals are signals that have undergone source coding, channel coding, modulation, and up-conversion or down-conversion processing.
  • the N second signals obtained after the N first signals are processed by the corresponding N second models are processed by source coding, channel coding, modulation, and up-conversion or down-conversion. Process the resulting signal. Subsequently, the sending end device sends the N second signals through the same channel.
  • the method of model transmission further includes: receiving first information, the first information being used to determine the N first models.
  • the first information comes from the receiving end device of the second signal, or the first information comes from a server that provides services to the receiving end device of the second signal.
  • the first information includes one or more of the following: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a system for model transmission including: a sending end device, configured to obtain information and fusion auxiliary information of N first models, and determine N based on the information and fusion auxiliary information of the N first models.
  • first signal and send the N first signals, where the N first models correspond to N first tasks, N is an integer greater than or equal to 2, and the fusion auxiliary information of the N first models includes the target model
  • the external characteristic value of , the target model corresponds to the second task, and the second task is different from the first task;
  • the receiving end device is configured to receive a second signal and analyze the received second signal to obtain a target model, where the second signal is obtained from the N first signals.
  • a sending end device obtains N first models and fusion auxiliary information, and processes the information of the N first models and the fusion auxiliary information according to the first module to obtain N first signals. , and then send the N first signals. Then, the receiving end device can receive the second signal corresponding to the target model. The second signal is obtained by superposing the N first signals in the channel.
  • a system for model transmission including: a first sending end device, configured to obtain information and first fusion auxiliary information of S first models among N first models, according to the N first models The information of the S first models in the model and the first fusion auxiliary information determine the S first signals, and send the S first signals, where the S models among the N first models correspond to the S first signals.
  • a task, S is an integer greater than or equal to 1
  • the first fusion auxiliary information includes the external feature value of the target model, the target model corresponds to the second task, and the second task is different from the N first tasks;
  • the second sending end device is used to obtain K first models among the N first models and the second fusion auxiliary information, based on the information of the K first models among the N first models and the second fusion auxiliary information.
  • the information determines K first signals and sends the K first signals, where K models among the N first models correspond to K first tasks, K is an integer greater than or equal to 1, and the second fusion
  • the auxiliary information includes external feature values of the target model.
  • the target model corresponds to a second task.
  • the second task is different from the N first tasks.
  • the intersection of the S first models and the K first models is empty set;
  • the receiving end device is configured to receive a second signal and analyze the received second signal to obtain the target model, wherein the second signal is composed of the S first signals and the Kth first signal. A signal is obtained.
  • the receiving end device can receive the second signal corresponding to the target model, and the second signal is obtained from the N first signals.
  • the target model corresponds to a second task, and the second task is different from the N first tasks.
  • the N first signals are obtained by superimposing in the channel.
  • the first fusion auxiliary information and the second fusion auxiliary information may be the same or different.
  • both the first fusion auxiliary information and the second fusion auxiliary information can be used to represent the external feature values of the target model.
  • the first fusion auxiliary information is different from the second fusion auxiliary information
  • the first fusion auxiliary information represents a first subset of external feature values of the target model
  • the second fusion auxiliary information represents a second subset of external feature values of the target model. subset, the first subset and the second subset are different, and the external feature value of the target model is the union of the first subset and the second subset.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the N first signals are obtained by the first module after processing the information of the N first models and the fusion auxiliary information.
  • the above-mentioned S first signals are the responses of the first module to the S first models among the S first models.
  • the K first signals are obtained by the first module after processing the information of the K first models among the K first models and the second fusion auxiliary information.
  • the first module is trained based on training data, and the training data includes M first training signals and the target model,
  • the M first training signals correspond to the M training models one-to-one, and the target model can meet the tasks corresponding to the M training models, and M is an integer greater than or equal to 2.
  • the first module is determined based on the parameters of the channel for transmitting the N first signals.
  • the parameters of the above channel include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, Link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the external feature values of the above-mentioned model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • the first module corresponding to the i-th model among the N first models is based on the Determined by the information of at least one first model other than the i-th model in a model.
  • the sending end device is also used to obtain M original models, and input each first training signal into the corresponding original model. model to obtain M second training signals, and superimpose the M second training signals in the same channel to obtain a third training signal.
  • the third training signal corresponds to the intermediate model, and according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, and the first module is obtained from the model library.
  • M is an integer greater than or equal to 2.
  • the model library includes the adjusted original model, and the deviation between the adjusted intermediate model and the target model. within the preset range.
  • the first model can be obtained through model training, which is beneficial to realizing the superposition of N first signals in the channel, and thereby obtaining the second signal corresponding to the target model.
  • the sending end device is also configured to send the fusion auxiliary information to the receiving end device.
  • the fusion assistance information comes from the receiving end device, or the fusion assistance information comes from a server that provides services for the receiving end device.
  • a system for model transmission including: a sending end device for acquiring N first signals, and processing the N first signals according to N second models to obtain Nth Two signals, and the N second signals are sent through the first channel, wherein the N first signals correspond to the N first models one-to-one, and the N second models correspond to the N first models one-to-one.
  • each first signal is used to carry information of the corresponding first model, where the information of the first model includes external feature values of the model and/or parameters of the model, and the N second models are based on the training data.
  • the training data includes M first training signals and target models.
  • the M first training signals correspond to M training models.
  • Each first training signal is the signal of the corresponding training model.
  • the target model can Satisfying the services corresponding to the M training models, the M services corresponding to the M first training models include the N services corresponding to the N first models, and the n-th second of the N second signals
  • the signal is a signal obtained by processing the n-th first signal according to the n-th second model.
  • the n-th first signal is the signal of the first model corresponding to the n-th second model.
  • the N first signals There is a one-to-one correspondence with the N first models, and the first model and the second model corresponding to each other correspond to the same business, n ⁇ [1, N], N is an integer greater than or equal to 2, and M is an integer greater than or equal to 2. integer;
  • the receiving end device is configured to receive signals through the first channel and analyze the received signals to obtain the third model.
  • a transmitting end device obtains N first signals corresponding to N first models, and processes the N first signals according to N second models to obtain N second signals, The N second signals are then sent over the first channel. Furthermore, a receiving end device corresponding to a transmitting end device can receive a third signal corresponding to the target model. The third signal is obtained by superposing N second signals in the first channel.
  • a system for model transmission including: a first sending end device for acquiring S first signals corresponding to S first models among N first models, and transmitting them according to the Nth first models.
  • the S second models in the two models process the S first signals to obtain S second signals, and send the S second signals through the first channel, where the S first signals are related to the S
  • Each first signal is used to carry the information of the corresponding first model, where the information of the first model It includes external characteristic values of the model and/or parameters of the model, and the first model and the second model corresponding to each other correspond to the same business, and S is an integer greater than or equal to 1;
  • the second transmitting end device is configured to obtain K first signals corresponding to the K first models among the N first models, and calculate the K first signals based on the K second models among the N second models.
  • a signal is processed to obtain K second signals, and the K second signals are sent through the first channel, where the K first signals correspond to the K first models one-to-one, and each first signal is In order to carry the information of the corresponding first model, where the information of the first model includes the external characteristic values of the model and/or the parameters of the model, the K second models correspond to the K first models one-to-one, And the first model and the second model corresponding to each other correspond to the same business, K is an integer greater than or equal to 1, and the intersection of the K first models and S first models is an empty set;
  • the N second models correspond to the N first models one-to-one, and the N second models are trained according to training data.
  • the training data includes M first training signals and target models, and the M The first training signal corresponds to M training models. Each first training signal is the signal of the corresponding training model.
  • the target model can meet the services corresponding to the M training models.
  • the M corresponding to the M first training models The services include N services corresponding to the N first models, and the n-th second signal among the N second signals is a signal obtained by processing the n-th first signal according to the n-th second model.
  • the n-th first signal is the signal of the first model corresponding to the n-th second model
  • the N first signals correspond to the N first models one-to-one
  • the first models and the corresponding N-th models correspond to each other.
  • the two models correspond to the same business, n ⁇ [1, N], N is an integer greater than or equal to 2, and M is an integer greater than or equal to 2;
  • the receiving end device is configured to receive signals through the first channel and analyze the received signals to obtain the third model.
  • the second sending end device can obtain K first signals corresponding to K first models among the N first models, and the K first models are related to the S The intersection of the first models is the empty set)
  • the first signals are processed into S second signals after being processed by the corresponding second model, and the K first signals corresponding to the second sending end device are processed into K second signals after being processed by the corresponding second model), and then
  • the N second signals are sent through the first channel.
  • the receiving end device can receive the third signal corresponding to the target model.
  • the third signal is obtained by superimposing N second signals in the first channel.
  • the obtained Obtaining the N second models includes: obtaining the N second models according to parameters of the channel that sends the N second signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the external feature values of the above-mentioned model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • obtaining the N second models further includes: excluding the i-th one from the N first models.
  • the parameters of at least one first model other than the first model determine the second model corresponding to the i-th first model, i ⁇ [1, N].
  • the method of model transmission also includes: obtaining M original models, and the M original models are related to the M There is a one-to-one correspondence between the first training signals, M is an integer greater than or equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; convert the M second training signals Superimpose in the same channel to obtain a third training signal; analyze the third training signal to obtain an intermediate model; adjust the parameters of the M original models according to the deviation between the intermediate model and the target model to obtain
  • the model library includes the original model after adjustment, and the deviation between the intermediate model after adjustment and the target model is within a preset range; the N second models are obtained from the model library.
  • parameters of the above model may include weight matrices, weight vectors, bias matrices, and bias vectors.
  • the method of model transmission further includes: receiving first information, the first information being used to determine the N a first model.
  • the first information comes from the receiving end device of the second signal, or the first information comes from a server that provides services to the receiving end device of the second signal.
  • the first information includes one or more of the following: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a device for model transmission including:
  • An acquisition unit used to acquire information of N first models and fusion auxiliary information.
  • the fusion auxiliary information includes external feature values of the target model, and the target model corresponds to the second task;
  • a processing unit configured to determine N first signals based on the information of the N first models and the fusion auxiliary information, and send the N first signals, where the N first models correspond to N first tasks, N is an integer greater than or equal to 2, and the second task is different from the first task;
  • the sending unit is configured to send the N first signals to the receiving end device.
  • the sending end device obtains the information of the N first models and the fusion auxiliary information, and processes the information of the N first models and the N fusion auxiliary information according to the first module to obtain the Nth A signal, and then the N first signals are sent. Then, the receiving end device can receive the second signal corresponding to the target model. The second signal is obtained by superposing the N first signals in the channel.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the processing unit determines the N first signals through the first module.
  • the first module is determined based on parameters of a channel that sends N first signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the first module corresponding to the i-th model among the N first models is based on the i-th model among the N first models except the i-th model among the N first models. Determined by information on at least one first model other than the first model.
  • the processing unit is also used to obtain M original models, the M original models correspond to the M first training signals one-to-one, and M is greater than Or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal; The third training signal is analyzed to obtain an intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the adjusted models. The original model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the N first modules are obtained from the model library.
  • the sending unit is also configured to send the fusion auxiliary information to the receiving end device.
  • the fusion auxiliary information comes from the receiving end device, or the fusion auxiliary information comes from a server that provides services for the receiving end device.
  • the acquisition unit is a receiving unit or a transceiver unit; if the information of the N first models and the fusion auxiliary information are transmitted by the model If the device obtains it locally (for example, determined), then the acquisition unit can be the processing unit; if one type of information among the N first model information and the fusion auxiliary information is received from other devices, the other type of information is If acquired locally, the acquisition unit can be a functional integration unit of the receiving unit and the processing unit.
  • a fourteenth aspect provides a device for model transmission, including:
  • a processing unit configured to acquire N first signals, process the N first signals according to N second models, obtain N second signals, and send the N second signals through the first channel, where , the N first signals correspond to the N first models one-to-one, the N second models correspond to the N first models one-to-one, and each first signal is used to carry the information of the corresponding first model.
  • the information of the first model includes external feature values of the model and/or parameters of the model, and the N second models are trained according to the training data.
  • the training data includes M first training signals and the target model.
  • the M first training signals correspond to M training models
  • each first training signal is the signal of the corresponding training model
  • the target model can meet the services corresponding to the M training models
  • the M first training models correspond to
  • the M services include N services corresponding to N first models
  • the n-th second signal among the N second signals is obtained by processing the n-th first signal according to the n-th second model.
  • the nth first signal is the signal of the first model corresponding to the nth second model
  • the N first signals correspond to the N first models one-to-one
  • the two models correspond to the same business, n ⁇ [1, N], N is an integer greater than or equal to 2, and M is an integer greater than or equal to 2;
  • the sending unit is configured to send the N second signals to the receiving end device through the same channel.
  • the first module in the sending end device can obtain N first signals corresponding to N first models, and process the N first signals according to N second models to obtain N
  • the second signal is then sent by the sending module through the first channel.
  • the receiving end device can receive the third signal corresponding to the target model.
  • the third signal is obtained by superimposing N second signals in the first channel.
  • the external feature values of the above-mentioned model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • the N second models are determined based on parameters of the first channel.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the second model corresponding to the i-th first model among the N first models is based on the i-th one of the N first models. Determined by the information of at least one first model other than the first model, i ⁇ [1, N].
  • the processing unit is also used to obtain M original models, and input each first training signal into the corresponding original model to obtain the Mth Two training signals, and the M second training signals are superimposed in the same channel to obtain a third training signal, and the third training signal is analyzed to obtain an intermediate model, and according to the deviation between the intermediate model and the target model, Adjust the parameters of the M original models to obtain a model library, and obtain N second models from the model library.
  • the M original models correspond to the M first training signals one-to-one, and M is greater than or equal to An integer of 2
  • the model library includes the original model after adjustment, and the deviation between the intermediate model after adjustment and the target model is within a preset range.
  • the acquisition unit is further configured to: receive first information, the first information being used to determine the N first models, wherein the first information from the receiving end device, or the first information comes from a server that provides services to the receiving end device.
  • a device for model transmission which device includes:
  • An acquisition unit configured to acquire information of S first models among the N first models and first fusion auxiliary information, where the first fusion auxiliary information includes external feature values of the target model, and the target model corresponds to the second task;
  • a processing unit configured to determine S first signals based on the information of the S first models and the first fusion auxiliary information, wherein the S first models correspond to S first tasks, and the second tasks are related to the S The first task is different;
  • the sending unit is used to send the S first signals to the receiving end device.
  • the sending end device obtains the information of S first models and the first fusion auxiliary information among the N first models, and combines the information of the S first models with the first fusion auxiliary information according to the first module.
  • the auxiliary information is fused and processed to obtain S first signals, and then the S first signals are sent.
  • S is less than N
  • one or more other sending ends will also send the first signal among NS first signals, and then the receiving end device can receive the second signal corresponding to the target model.
  • the second signal is composed of S It is obtained by superposing N first signals of N first signals in the channel.
  • the joint transmission of multiple different neural network models can be achieved, which is also suitable for scenarios where the sending end device is distributed, and can reduce the resource overhead and delay in the transmission process of multiple neural network models, and To a certain extent, the storage space of the receiving end device for the neural network model is reduced.
  • the information of the first model or the information of the target model includes external feature values of the model and/or parameters of the model, where the external feature values of the model include one or more of the following information: neural network calculation graph information, Optimizer information, hyperparameter information, and model parameters include one or more of the following parameters: weight matrix, weight vector, bias matrix, and bias vector.
  • the first fusion auxiliary information may be the external feature value of the target model, or may be a subset of the external feature value of the target model.
  • the processing unit determines the S first signals through the first module.
  • the first module is trained based on training data
  • the training data includes M first training signals and target models
  • the M first training signals corresponds one-to-one to M training models.
  • the target model can satisfy the tasks corresponding to the M training models.
  • M is an integer greater than or equal to 2.
  • the first module is determined based on parameters of the channel for sending the S first signals.
  • channel parameters include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the first module corresponding to the i-th model among the S first models is based on the S first models except the i-th model among the first models.
  • the information of at least one first model other than i models is determined.
  • the processing unit is also used to obtain M original models, the M original models correspond to the M first training signals one-to-one, and M is greater than or an integer equal to 2; input each first training signal into the corresponding original model to obtain M second training signals; superimpose the M second training signals in the same channel to obtain a third training signal, Among them, the third training signal corresponds to the intermediate model; according to the deviation between the intermediate model and the target model, the parameters of the M original models are adjusted to obtain a model library, which includes the original adjusted models. model, and the deviation between the adjusted intermediate model and the target model is within a preset range; the first module is obtained from the model library.
  • the sending unit is further configured to send the first fusion auxiliary information to the receiving end device.
  • the first fusion auxiliary information comes from the receiving end devices of the S first signals, or the first fusion auxiliary information comes from the S first signals.
  • the receiving end device of the signal provides the server with the service.
  • a sixteenth aspect provides a device for model reception, which device includes:
  • a receiving unit configured to receive a second signal, the second signal is associated with the first model, and the N first models correspond to the N first tasks;
  • the processing unit obtains a target model according to the second signal.
  • the target model corresponds to a second task, and the second task is different from the first task.
  • the receiving end device receives the second signal sent by the transmitting end device to obtain the target model.
  • the second signal is obtained by superimposing N first signals in the channel.
  • the device of the model further includes a sending unit, the sending unit is used to send the fusion auxiliary information to the receiving end device, the fusion auxiliary information includes external components of the target model Eigenvalues.
  • the external feature values of the model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the The processor is configured to execute the first aspect and the method in any implementation manner of the first aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the The processor is configured to execute the second aspect and the method in any implementation manner of the second aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the The processor is configured to execute the third aspect and the method in any implementation manner of the third aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the The processor is configured to execute the fourth aspect and the method in any one implementation manner of the fourth aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the processor is configured to execute the fifth aspect and the method in any one implementation manner of the fifth aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the processor is configured to execute the sixth aspect and the method in any one implementation manner of the sixth aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the processor is configured to execute the seventh aspect and the method in any one implementation manner of the seventh aspect.
  • a device for model transmission includes: a memory for storing a program; a processor for executing the program stored in the memory.
  • the processor is configured to execute the eighth aspect and the method in any one implementation manner of the eighth aspect.
  • the processor in the above-mentioned seventeenth aspect to the twenty-fourth aspect can be either a central processing unit (CPU) or a combination of a CPU and a neural network operation processor.
  • the neural network operation processor here can be Including graphics processing unit (GPU), neural network processing unit (NPU), tensor processing unit (TPU), etc.
  • GPU graphics processing unit
  • NPU neural network processing unit
  • TPU tensor processing unit
  • TPU is an artificial intelligence accelerator special integrated circuit fully customized by Google for machine learning.
  • a computer-readable medium stores a program code for device execution.
  • the program code includes a program code for executing any one of the implementations of the first aspect and the eighth aspect. Methods.
  • a twenty-sixth aspect provides a computer program product containing instructions, which when the computer program product is run on a computer, causes the computer to execute the method in any one of the above-mentioned implementations of the first aspect and the eighth aspect.
  • a twenty-seventh aspect provides a chip.
  • the chip includes a processor and a data interface.
  • the processor reads instructions stored in the memory through the data interface and executes any of the above first and eighth aspects.
  • the chip may further include a memory, in which instructions are stored, and the processor is configured to execute the instructions stored in the memory.
  • the processor is configured to execute the method in any one implementation manner of the first aspect and the eighth aspect.
  • the above-mentioned chip can specifically be a field-programmable gate array (FPGA) or an application-specific integrated circuit (ASIC).
  • FPGA field-programmable gate array
  • ASIC application-specific integrated circuit
  • Figure 1 is a schematic diagram of an end-to-end communication system based on neural networks.
  • FIG. 2 is a schematic diagram of an applicable communication system architecture provided by an embodiment of the present application.
  • FIG. 3 is a schematic diagram of another communication system architecture provided by an embodiment of the present application.
  • FIG. 4 is a schematic diagram of another communication system architecture provided by an embodiment of the present application.
  • Figure 5 shows a schematic diagram of an example of the artificial intelligence main framework according to the embodiment of the present application.
  • FIG. 6 shows a schematic diagram of an example system architecture 600 provided by an embodiment of the present application.
  • Figure 7 is a schematic structural diagram of a convolutional neural network model provided by an embodiment of the present application.
  • Figure 8 is a schematic structural diagram of another convolutional neural network model provided by an embodiment of the present application.
  • FIG. 9 is a schematic diagram of the hardware structure of a chip provided by an embodiment of the present application.
  • Figure 10 is a schematic diagram of a system architecture 1000 provided by an embodiment of the present application.
  • Figure 11 is a schematic diagram of device deployment of a training model provided by an embodiment of the present application.
  • Figure 12 is a schematic flow chart of a model transmission method provided by an embodiment of the present application.
  • Figure 13 is a schematic flow chart of a model transmission method provided by an embodiment of the present application.
  • Figure 14 is a schematic flow chart of another model transmission method provided by an embodiment of the present application.
  • Figure 15 is a schematic diagram of a directed graph structure of a neural network model provided by an embodiment of the present application.
  • Figure 16 is a schematic structural diagram of a calculation graph of a neural network model provided by an embodiment of the present application.
  • Figure 17 is a schematic diagram of neural network model transmission at the computational graph level provided by the embodiment of the present application.
  • Figure 18 is a schematic flow chart of yet another model transmission method provided by an embodiment of the present application.
  • Figure 19 is a schematic block diagram of an example of a model transmission device provided by an embodiment of the present application.
  • Figure 20 is a schematic block diagram of another example of a model transmission device provided by an embodiment of the present application.
  • NB-IoT narrowband-internet of things
  • GSM global system for mobile communications
  • EDGE enhanced Data rate GSM evolution system
  • WCDMA wideband code division multiple access system
  • CDMA2000 code division multiple access 2000 system
  • TD-SCDMA time division synchronization code Time division-synchronization code division multiple access
  • LTE long term evolution
  • FDD frequency division duplex
  • TDD fifth generation
  • 5G fifth generation
  • WiFi wireless fidelity
  • 5G mobile communication system Large application scenarios: enhanced mobile broadband (eMBB), ultra-reliable low-latency communication (uRLLC), enhanced machine type communication (LTE enhanced MTO, eMTC), supporting multiple Communication systems integrating wireless technologies and the sixth generation (6th generation, 6G) communication systems evolved after 5G, etc.
  • eMBB enhanced mobile broadband
  • uRLLC ultra-reliable low-latency communication
  • eMTC enhanced machine type communication
  • 6G sixth generation
  • wireless cellular communication systems such as wireless cellular communication systems, wireless mesh (Mesh) networks, and satellite communication systems.
  • a wireless cellular communication system includes a cellular base station and multiple terminal devices
  • the multiple terminal devices can assist the cellular base station in performing artificial intelligence (AI) model calculations.
  • AI artificial intelligence
  • a wireless cellular communication system includes multiple cellular base stations and a terminal device
  • multiple cellular base stations can assist the terminal device in AI model calculations.
  • Wireless Mesh networks include cellular macro base stations, micro base stations and terminal equipment. Cellular macro base stations transmit downlink data to terminal devices through multiple micro base station relays.
  • Satellite communication systems include satellite base stations and terminal equipment.
  • the terminal in the embodiment of this application may be a device with wireless transceiver function, which may specifically refer to user equipment (UE), access terminal, subscriber unit (subscriber unit), user station, or mobile station (mobile station). , remote station, remote terminal, mobile device, user terminal, wireless communication device, user agent or user device.
  • UE user equipment
  • access terminal subscriber unit (subscriber unit)
  • subscriber unit subscriber unit
  • user station or mobile station (mobile station).
  • remote station remote terminal, mobile device, user terminal, wireless communication device, user agent or user device.
  • the terminal device may also be a satellite phone, a cellular phone, a smartphone, a wireless data card, a wireless modem, a machine type communications device, a cordless phone, a session initiation protocol (SIP) phone, a wireless local loop (wireless local) loop (WLL) station, personal digital assistant (PDA), handheld device with wireless communication capabilities, computing device or other processing device connected to a wireless modem, vehicle-mounted equipment, communication equipment carried on high-altitude aircraft, wearable Equipment, drones, robots, terminals in device-to-device (D2D) communication, terminals in vehicle outreach (vehicle to everything, V2X), virtual reality (VR) terminal equipment, Augmented reality (AR) terminal equipment, wireless terminals in industrial control, wireless terminals in self-driving, wireless terminals in remote medical, smart grid Wireless terminals in transportation safety (transportation safety), wireless terminals in smart cities (smart cities), wireless terminals in smart homes (smart homes), or communication networks evolved after 5G (such as 6G) Terminal equipment
  • the device used to implement the functions of the terminal device may be a terminal device; it may also be a device that can support the terminal device to implement the function, such as a chip system.
  • the device can be installed in a terminal device or used in conjunction with the terminal device.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • the network device in the embodiment of the present application has a wireless transceiver function and is used to communicate with the terminal device.
  • the access network device may be a node in a radio access network (radio access network, RAN), which may also be called a base station or a RAN node. It can be an evolved base station (evolved Node B, eNB or eNodeB) in LTE; or a base station in a 5G network such as gNodeB (gNB) or a base station in a public land mobile network (PLMN) evolved after 5G.
  • 3GPP non-3rd generation partnership project
  • the network equipment in the embodiment of the present application may include various forms of base stations, such as: macro base stations, micro base stations (also called small stations), relay stations, access nodes in the WiFi system, and transmitting and receiving points. and receiving point, TRP), transmitting point (TP), mobile switching center and device-to-device (D2D), vehicle outreach (vehicle-to-everything, V2X), machine-to-machine (machine- Equipment that undertakes base station functions in to-machine (M2M) communications, etc., can also include centralized units (centralized units, CUs) and distributed units (distributed units) in cloud radio access network (cloud radio access network, C-RAN) systems. unit, DU), and network equipment in the NTN communication system, which are not specifically limited in the embodiments of this application.
  • base stations such as: macro base stations, micro base stations (also called small stations), relay stations, access nodes in the WiFi system, and transmitting and receiving points. and receiving point, TRP), transmitting point (TP), mobile switching center and
  • the device used to implement the function of the network device may be a network device; it may also be a device that can support the network device to implement the function, such as a chip system.
  • the device can be installed in a network device or used in conjunction with a network device.
  • the chip system may be composed of chips, or may include chips and other discrete devices.
  • sending end device and the receiving end device involved in the embodiments of this application may be one or more of the above-mentioned terminal devices, or may be one or more of the above-mentioned network devices, which this application does not limited.
  • the cloud device in the embodiment of this application may be a server or server cluster serving the receiving device or the sending device.
  • the cloud device may also be called a computing node or a cloud computing cluster.
  • FIG. 2 is a schematic diagram of an example communication system architecture provided by an embodiment of the present application. As shown in Figure 2, the communication system architecture is a satellite communication system.
  • the satellite communication system shown in (a) of Figure 2 includes equipment such as a satellite base station, terminal equipment #1, and terminal equipment #2, where the satellite base station may also be called a satellite.
  • the satellite base station can provide communication services for terminal equipment #1 and terminal equipment #2.
  • the satellite base station can transmit downlink data to terminal equipment #1 and terminal equipment #2, where the downlink data can be transmitted to terminal equipment #1 and terminal equipment #2 after channel coding, modulation mapping;
  • terminal equipment #1 and terminal equipment #1 Device #2 can also transmit uplink data to the satellite base station, where the uplink data can also be transmitted to the satellite base station after channel coding and modulation mapping.
  • the satellite communication system shown in (b) of Figure 2 may also include a ground base station, and the satellite base station can communicate with the base station on the ground.
  • the satellite base station can be a drone, a hot air balloon, a low-orbit satellite, a medium-orbit satellite or a high-orbit satellite, etc.
  • a satellite base station may also refer to a non-terrestrial base station or non-terrestrial equipment. Satellite base stations can be used as both network equipment and terminal equipment.
  • the satellite base station may not have the functions of a base station, or may have some or all of the functions of a base station, which is not limited in this application.
  • FIG. 3 is a schematic diagram of another communication system architecture provided by an embodiment of the present application.
  • the communication system is a satellite inter-satellite link communication system, which can also be called an inter-satellite communication system.
  • the satellite in the communication system may include a communication system and a space beam acquisition, tracking and alignment technology (acquisition, pointing and tracking, APT) system.
  • the communication system may include a communication module and a transceiver antenna, which is responsible for the transmission of inter-satellite information and is the main body of the inter-satellite communication system.
  • the APT system is responsible for the acquisition, alignment and tracking of satellites. Among them, capture refers to the judgment and identification of targets in an uncertain area. The capture process includes early data guidance and initial pointing of the visual axis, scanning in the uncertain area, until the determination of the beacon light spot in the capture field of view.
  • alignment refers to adjusting the transmitting wave to aim at the receiving direction, so that the communication transmitting boresight and tracking boresight need to maintain very precise coaxiality during assembly and calibration, or the launch needs to be corrected through the advance servo unit
  • the deviation between the boresight and the receiving boresight; tracking refers to the continuous adjustment of alignment and acquisition throughout the communication process.
  • the essence of tracking in the space laser communication system is to achieve active photoelectric tracking of the laser beam. tracking, thereby overcoming the relative motion and platform vibration caused by the shaking of the boresight axis.
  • the purpose of tracking is to stabilize the tracking boresight within the allowable error range.
  • APT In the satellite inter-satellite link communication system, in order to minimize the attenuation and interference effects in the channel, while requiring high confidentiality and transmission rate, APT must be adjusted in real time to continuously adapt to changes.
  • Existing APT systems are all optical systems. Most of the existing communication subsystems are optical communication systems, and some systems in the microwave band mostly use a single high-gain antenna. The existing APT system and communication subsystem are independent systems.
  • FIG. 4 is a schematic diagram of another communication system architecture provided by an embodiment of the present application.
  • the communication system shown is a wireless cellular communication system, which is usually composed of cells.
  • Each cell contains a base station (BS), which can provide communication services to multiple terminal devices.
  • BS base station
  • the cellular communication system shown in Figure 4 includes a base station, terminal device #1, terminal device #2 and other equipment.
  • the base station can send downlink data to the terminal device, where the downlink data can be encoded using channel coding; the terminal device can also send uplink data to the base station, and the uplink data can also be encoded using channel coding.
  • the base station includes a baseband unit (BBU) and a remote radio unit (RRU).
  • BBU baseband unit
  • RRU remote radio unit
  • the BBU and RRU can be placed in different places.
  • the RRU can be placed in a high-traffic area
  • the BBU can be placed in the central computer room
  • the BBU and the RRU can be placed in the same computer room
  • the BBU and the RRU can be in different locations under the same rack. part.
  • the wireless cellular communication system shown in Figure 4 may include multiple base stations and other numbers of terminal devices may be included within the coverage of each base station, which is not limited in this embodiment of the present application.
  • Figure 5 shows a schematic diagram of an artificial intelligence main framework according to an embodiment of the present application.
  • the main framework describes the overall workflow of the artificial intelligence system and is suitable for general requirements in the field of artificial intelligence.
  • Intelligent information chain reflects a series of processes from data acquisition to processing. For example, it can be the general process of intelligent information perception, intelligent information representation and formation, intelligent reasoning, intelligent decision-making, intelligent execution and output. In this process, the data has gone through the condensation process of "data-information-knowledge-wisdom".
  • the "IT value chain” reflects the value that artificial intelligence brings to the information technology industry, from the underlying infrastructure of human intelligence and information (providing and processing technology implementation) to the systematic industrial ecological process.
  • the infrastructure provides computing power support for the artificial intelligence system, realizes communication with the external real world, and achieves support through the basic platform.
  • the infrastructure can communicate with the outside through sensors, and the computing power of the infrastructure can be provided by smart chips.
  • the smart chip here can be a central processing unit (CPU), a neural network model processor (neural-network processing unit, NPU), a graphics processor (graphics processing unit, GPU), or an integrated circuit for specialized applications. (application specific integrated circuit, ASIC) and field programmable gate array (field programmable gate array, FPGA) and other hardware acceleration chips.
  • CPU central processing unit
  • NPU neural network model processor
  • GPU graphics processor
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • the basic platform of infrastructure can include distributed computing framework and network related platform guarantees and support, and can include cloud storage and computing, interconnection networks, etc.
  • data can be obtained through communication between sensors and the outside world, and then provided to smart chips in the distributed computing system provided by the basic platform for calculation.
  • Data from the upper layer of the infrastructure is used to represent data sources in the field of artificial intelligence.
  • This data involves graphics, images, voice, and text, as well as IoT data of the device, including business data of existing systems and sensory data such as force, displacement, liquid level, temperature, and humidity.
  • the above data processing usually includes data training, machine learning, deep learning, search, reasoning, decision-making and other processing methods.
  • machine learning and deep learning can perform symbolic and formal intelligent information modeling, extraction, preprocessing, training, etc. on data.
  • Training is a computationally intensive operation. Models generally need to use a large amount of data for training, and the parameters of the model are continuously optimized through backpropagation so that the model can acquire certain capabilities. During the training process, we often give top priority to the fit of the model on the data set.
  • Inference refers to the process of simulating human intelligent reasoning in computers or intelligent systems, using formalized information to perform machine thinking and problem solving based on reasoning control strategies. Typical functions are search and matching.
  • Decision-making refers to the process of decision-making after intelligent information is reasoned, which usually provides functions such as classification, sorting, and prediction.
  • some general capabilities can be formed based on the results of data processing, such as algorithms or a general system, such as translation, text analysis, computer vision processing, speech recognition, and image recognition. etc.
  • Intelligent products and industry applications refer to the products and applications of artificial intelligence systems in various fields. They are the encapsulation of overall artificial intelligence solutions, productizing intelligent information decision-making and realizing practical applications. Its application fields mainly include: intelligent manufacturing, intelligent transportation, Smart home, smart medical care, smart security, autonomous driving, safe city, smart terminal, etc.
  • Embodiments of the present application can be applied to many fields in artificial intelligence, such as smart manufacturing, smart transportation, smart home, smart medical care, smart security, autonomous driving, safe cities and other fields.
  • the method of obtaining a neural network model in the embodiment of the present application can be specifically applied to autonomous driving, image classification, image retrieval, image semantic segmentation, image quality enhancement, image super-resolution and natural language processing, etc. ( The field of deep) neural network models.
  • Artificial intelligence is a theory, method, technology and application system that uses digital computers or machines controlled by digital computers to simulate, extend and expand human intelligence, perceive the environment, acquire knowledge and use knowledge to obtain the best results.
  • artificial intelligence is a branch of computer science that attempts to understand the nature of intelligence and produce a new class of intelligent machines that can respond in a manner similar to human intelligence.
  • Artificial intelligence is the study of the design principles and implementation methods of various intelligent machines, so that the machines have the functions of perception, reasoning and decision-making.
  • Research in the field of artificial intelligence includes robotics, natural language processing, computer vision, decision-making and reasoning, human-computer interaction, recommendation and search, basic AI theory, etc.
  • Edge AI is one of the new fields in the field of artificial intelligence. Its purpose is to allow users to perform artificial intelligence processes without worrying about privacy leaks or the impact of slower data transfer rates.
  • Edge artificial intelligence combines artificial intelligence technology with edge computing, allowing artificial intelligence algorithms to run on devices capable of edge computing, enabling real-time processing of data.
  • Edge artificial intelligence can make artificial intelligence technology more widely used, allowing smart devices to respond quickly to inputs without accessing the cloud platform.
  • self-driving cars need to constantly scan the surrounding environment and evaluate the current driving conditions while driving, and need to continuously correct their trajectory based on emergencies.
  • the processing of real-time data of the car is crucial, and its on-board artificial intelligence system will be responsible for the storage, analysis and processing of data.
  • edge artificial intelligence technology real-time analysis and processing of data can be achieved without relying on cloud computing connections.
  • Neural network can also be called artificial neural network.
  • the neural network can be composed of neural units.
  • the neural unit can refer to an arithmetic unit that takes x s and intercept 1 as input.
  • the output of the arithmetic unit can be:
  • s 1, 2,...n, n is a natural number greater than 1
  • W s is the weight of x s
  • b is the bias of the neural unit.
  • f is the activation function of the neural unit, which is used to introduce nonlinear characteristics into the neural network to convert the input signal in the neural unit into an output signal.
  • the output signal of this activation function can be used as the input of the next convolutional layer, and the activation function can be a sigmoid function.
  • a neural network is a network formed by connecting multiple above-mentioned single neural units together, that is, the output of one neural unit can be the input of another neural unit.
  • the input of each neural unit can be connected to the local receptive field of the previous layer to extract the features of the local receptive field.
  • the local receptive field can be an area composed of several neural units.
  • Deep neural network also known as multi-layer neural network
  • DNN can be understood as a neural network with multiple hidden layers.
  • DNN is divided according to the position of different layers.
  • the neural network inside DNN can be divided into three categories: input layer, hidden layer and output layer.
  • the first layer is the input layer
  • the last layer is the output layer
  • the middle layers are hidden layers.
  • the layers can be fully connected, that is, any neuron in the i-th layer is connected to any neuron in the i+1-th layer. Simply put, it can be expressed by the following linear relationship:
  • the coefficient from the k-th neuron in layer L-1 to the j-th neuron in layer L is defined as
  • the input layer has no parameters.
  • DNN can fit any function with arbitrary accuracy.
  • more hidden layers can make the network more capable of depicting complex situations in the real world.
  • the more parameters the higher the complexity of the model and the greater the "capacity", which means that it can complete more complex learning tasks.
  • Training a deep neural network is the process of learning a weight matrix. The ultimate goal is to train the weight matrix of all layers of the deep neural network (a weight matrix formed by vectors of many layers).
  • Convolutional neural network is a neural network with a convolutional structure.
  • the convolutional neural network contains a feature extractor consisting of a convolutional layer and a subsampling layer, which can be regarded as a filter.
  • the convolutional layer refers to the neuron layer in the convolutional neural network that convolves the input signal.
  • a neuron can be connected to only some of the neighboring layer neurons.
  • a convolutional layer usually contains several feature planes, and each feature plane can be composed of some rectangularly arranged neural units. Neural units in the same feature plane share weights, and the shared weights here are convolution kernels. Shared weights can be understood as a way to extract image information independent of position.
  • the convolution kernel can obtain reasonable weights through learning.
  • the direct benefit of sharing weights is to reduce the connections between the layers of the convolutional neural network, while reducing the risk of overfitting.
  • the neural network can use the gradient back propagation (BP) algorithm to correct the size of the parameters in the neural network model during the training process, making the reconstruction error loss of the neural network model smaller and smaller. Specifically, forward propagation of the input signal until the output will produce an error loss, and the parameters in the neural network model are updated by backpropagating the error loss information, so that the error loss converges.
  • the backpropagation algorithm is a backpropagation movement dominated by error loss, aiming to obtain the optimal parameters of the neural network model, such as the weight matrix.
  • the loss value generated by each training of the neural network model is passed layer by layer from back to front in the neural network model.
  • the update amount (partial derivative operation) of the layer parameters is also calculated. This update amount is related to the gradient.
  • Deep learning is a new research direction in the field of machine learning (ML). Deep learning is the process of learning the inherent patterns and representation levels of sample data. The information obtained during this learning process is of great help in the interpretation of data such as text, images, and sounds. Its goal is to enable machines to have the ability to analyze and learn like humans, and to recognize data such as text, images, and sounds. Deep learning is a complex machine learning algorithm that achieves results in speech and image recognition that far exceed those of previous related technologies.
  • Depth in deep learning refers to a series of continuous representation layers, and the number of layers contained in a data model is called the depth of the model.
  • the model feature of deep learning is a neural network with multiple hidden layers, which combines low-level features to form more abstract high-level features.
  • Deep learning has achieved many results in search technology, data mining, machine learning, machine translation, natural language processing, multimedia learning, voice, recommendation and personalization technology, and other related fields. Deep learning enables machines to imitate human activities such as audio-visual and thinking, solves many complex pattern recognition problems, and makes great progress in artificial intelligence-related technologies.
  • Federated learning on the premise of fully ensuring the privacy and security of user data, prompts various distributed nodes and central nodes to cooperate to efficiently complete the learning tasks of the model.
  • data sets exist at distributed nodes, that is, distributed nodes collect local data sets, conduct local training, and report the local results (model or gradient) obtained by training to the central node.
  • the central node itself does not have a data set, and is only responsible for fusion processing of the training results of distributed nodes, obtaining the global model, and issuing it to distributed nodes.
  • Transfer learning is a term in machine learning, which refers to the impact of one type of learning on another type of learning, or the impact of acquired experience on the completion of other activities.
  • transfer learning is a learning idea and model.
  • Transfer learning refers to a learning process that uses the similarities between data, tasks or models to apply models learned in old fields to new fields. Its core It is to find the similarities between new problems and old problems and successfully transfer knowledge. Transfer widely exists in the learning of various knowledge, skills and social norms.
  • Transfer learning focuses on storing solving models for existing problems and leveraging them on other different but related problems.
  • the knowledge (or model) used to identify cars can also be used to improve the ability to identify trucks.
  • Transfer learning in the computer field has a certain conceptual relationship with learning transfer often mentioned in psychology, but the academic relationship between the two fields is very limited. Deep learning allows machines to autonomously acquire knowledge from data and apply it to solve new problems; while transfer learning focuses on transferring learned knowledge to solve new problems.
  • Neural network pruning can be divided into two categories. One is constraints during training, adding constraints to the loss function to induce network sparseness without the need for retraining; the other is post-training pruning to eliminate relatively redundant and inconsequential features in the network. The important part. According to the pruning granularity, there are inter-layer pruning and feature pruning. Both reduce network depth and width respectively, and conduct fine-grained in-kernel pruning.
  • neural network pruning Through neural network pruning, the training time is reduced, and as the amount of calculation is reduced, the speed of each iteration of the connection in the network is improved, and the neural network model can converge to the optimal solution faster. It can also make the running speed faster.
  • the number of sparse network convolution layers and the number of convolution kernels in the convolution layer are smaller.
  • a simpler and lighter model means more efficient and faster weight updates.
  • the advantage of neural network pruning also lies in more feasible embedded deployment, that is, the pruned network provides wider possibilities for application on mobile devices and other embedded devices.
  • Model quantization refers to compressing the original network by reducing the number of bits required for weights, and mainly includes two methods: low precision and re-encoding.
  • low precision means using lower digits of floating point numbers or integers to represent weight parameters for training, testing or storage
  • recoding means recoding the original data and using fewer digits to represent the original data, representing There are binarized weights/binarized neural networks.
  • Model quantization is the process of approximating floating-point model weights with continuous values (or a large number of possible discrete values) or fixed-point tensor data flowing through the model to a finite number of discrete values at a low loss of inference accuracy.
  • Knowledge distillation also known as network distillation, is a commonly used method for model compression. Different from pruning and quantification in the model compression process, knowledge distillation builds a lightweight small model and takes advantage of higher performance. Use the best supervision information of the large model to train the small model in order to achieve better performance and accuracy.
  • the large model can be called the teacher model
  • the small model is called the student model
  • the supervision information output from the teacher model is called knowledge.
  • the process of the student model transferring and learning the supervision information from the teacher model is called knowledge distillation.
  • FIG. 6 shows a schematic diagram of an example system architecture 600 provided by an embodiment of the present application.
  • data collection device 660 is used to collect training data.
  • the training data may include training images and classification results corresponding to the training images, where the results of the training images may be manually pre-annotated results.
  • the data collection device 660 After collecting the training data, the data collection device 660 stores the training data in the database 630, and the training device 620 trains to obtain the target model/rule 601 based on the training data in the database 630.
  • the training device 620 processes the input original image and compares the output image with the original image until the image output by the training device 620 is consistent with the original image.
  • the difference between the original images is less than a certain preset threshold, thereby completing the training process of the target model/rule 601.
  • the training device 620 can be used to obtain a pre-trained network model, migrate the pre-trained network model based on the target data set, and search for sub-network models in the migrated network model to obtain the target model/ Rule 601.
  • the target data set may be stored in database 630.
  • the training device 620 can also be used to pre-train the network model.
  • the network model is trained based on the source data set, which can also be stored in the database 630 .
  • the target model/rule 601 in the embodiment of this application may specifically be a neural network model.
  • a neural network model For example, convolutional neural networks, recurrent neural networks, or deep neural networks.
  • the training data maintained in the database 630 does not necessarily come from the collection of the data collection device 660, and may also be received from other devices, such as the target input by the client device 640. data set.
  • the training device 620 does not necessarily perform training of the target model/rules 601 based entirely on the training data maintained by the database 630. It is also possible to obtain training data from the cloud or other places for model training.
  • the above description should not be regarded as a limitation of this application. Limitations of Examples.
  • the target model/rules 601 trained according to the training device 620 can be applied to different systems or devices, for example, to the execution device 610 shown in Figure 6.
  • the execution device 610 can be a terminal, such as a mobile phone terminal, a tablet computer, Laptops, augmented reality (AR), AR/virtual reality (VR), vehicle-mounted terminals, etc., or servers or clouds, etc.
  • the execution device 610 is configured with an input/output (I/O) interface 612 for data interaction with external devices.
  • the user can input data to the I/O interface 612 through the client device 640, for example , the input data may include: images to be processed input by the client device.
  • the execution device 610 may also include a preprocessing module 613, which is used to preprocess the input data (for example: picture data) received by the I/O interface 612, if the input data does not need to be For preprocessing, the input data received from the I/O interface 612 (for example, the image data directly collected) is directly input to the execution device 610 for processing.
  • a preprocessing module 613 which is used to preprocess the input data (for example: picture data) received by the I/O interface 612, if the input data does not need to be
  • the input data received from the I/O interface 612 for example, the image data directly collected
  • the execution device 610 When the execution device 610 preprocesses the input data, or when the calculation module 611 of the execution device 610 performs calculations and other related processes, the execution device 610 can call the data, code, etc. in the data storage system 650 to perform corresponding processing. Processing, the data, instructions, etc. obtained by corresponding processing can also be stored in the data storage system 650.
  • the I/O interface 612 provides the processed results to the user device 640, for example, the above obtained image classification results are returned to the client device 640 for use by the client.
  • the training device 620 can generate corresponding target models/rules 601 based on different training data for different goals or different task requirements, and the corresponding target models/rules 601 can be used to achieve the above goals. Or complete the above tasks to provide users with the desired processing results.
  • the user can manually provide input data, and the manually provided data can be operated through the interface provided by the I/O interface 612 .
  • the client device 640 can automatically send input data to the I/O interface 612. If requiring the client device 640 to automatically send the input data requires the user's authorization, the user can set the settings in the client device 640. Corresponding permissions. The user can view the results output by the execution device 610 on the client device 640, and the specific presentation form can be specific ways such as display, sound, action, etc.
  • the client device 640 can also be used as a data collection terminal to collect the input data of the input I/O interface 612 and the output results of the output I/O interface 612 as new sample data, and store them in the database 630.
  • the I/O interface 612 directly uses the input data of the input I/O interface 612 and the output result of the output I/O interface 612 as shown in Figure 6 as new data.
  • the sample data is stored in database 630.
  • FIG. 6 shows a schematic diagram of a system architecture provided by an embodiment of the present application.
  • the present application does not place any restrictions on the positional relationships between devices, devices, modules, etc. in the system architecture.
  • the data storage system 650 is an external memory relative to the execution device 610 . In other possible situations, the data storage system 650 can also be placed in the execution device 610 .
  • a target model/rule 601 can be obtained by training according to the training device 620.
  • the target model/rule 601 can be the neural network model in the embodiment of the present application.
  • the neural network model constructed in the embodiment of the present application can be Including CNN, deep neural network models (deep neural networks, DNN), deep convolutional neural networks models (deep convolutional neural networks, DCNN), recurrent neural network models (recurrent neural network, RNN), etc., the embodiments of this application are No restrictions.
  • FIG. 7 is a schematic structural diagram of a convolutional neural network model related to an embodiment of the present application.
  • the convolutional neural network model (CNN) 700 may include an input layer 710 , a convolutional layer/pooling layer 720 (where the pooling layer is optional), and a neural network model layer 730 .
  • the input layer 710 can obtain an image to be processed, and pass the obtained image to be processed to the convolution layer/pooling layer 720 and the subsequent neural network model layer 730 for processing.
  • the following is a detailed introduction to the internal layer structure of CNN 700 in Figure 7.
  • Convolutional layer/pooling layer 720
  • the convolution layer/pooling layer 720 may include layers 721-726 as an example.
  • layer 721 is a convolution layer
  • layer 722 is a pooling layer
  • layer 723 is a convolution layer
  • layer 724 is a pooling layer
  • 725 is a convolution layer
  • 726 is a pooling layer
  • layers 721 and 722 are convolution layers
  • 723 is a pooling layer
  • 724 And 725 is the convolution layer
  • 726 is the pooling layer. That is, the output of the convolutional layer can be used as the input of the subsequent pooling layer, or can be used as the input of another convolutional layer to continue the convolution operation.
  • the convolution layer 721 may include multiple convolution operators.
  • the convolution operator may also be called a convolution kernel. Its role in image processing is equivalent to a filter that extracts specific information from the input image matrix.
  • the convolution operator In essence, it can be a weight matrix. This weight matrix is usually predefined. During the convolution operation on the image, the weight matrix is usually one pixel after another pixel (or two pixels next to each other) along the horizontal direction on the input image. Two pixels (depending on the value of stride) are processed to complete the work of extracting specific features from the image.
  • the size of the weight matrix should be related to the size of the image. It should be noted that the depth dimension of the weight matrix is the same as the depth dimension of the input image.
  • the weight matrix will extend to the input image Entire depth. Therefore, convolution with a single weight matrix will produce a convolved output with a single depth dimension, but in most cases, instead of using a single weight matrix, multiple weight matrices of the same size (rows ⁇ columns) are applied, That is, multiple matrices of the same type.
  • the output of each weight matrix is stacked to form the depth dimension of the convolution image.
  • the dimension here can be understood as being determined by the "multiple" mentioned above. Different weight matrices can be used to extract different features in the image.
  • one weight matrix is used to extract the edge information of the image
  • another weight matrix is used to extract the specific color of the image
  • another weight matrix is used to extract the noise information in the image. Extraction etc.
  • the multiple weight matrices have the same size (row ⁇ column), and the convolution feature maps extracted by the multiple weight matrices with the same size are also the same size. Then, the multiple extracted convolution feature maps with the same size are merged to form The output of the convolution operation.
  • weight values in these weight matrices require a lot of training in practical applications.
  • Each weight matrix formed by the weight values obtained through training can be used to extract information from the input image, thereby allowing the convolutional neural network model 700 to perform correct operations. predict.
  • the initial convolutional layer for example, layer 721
  • the features extracted by subsequent convolutional layers become more and more complex, such as high-level semantic features. The higher the semantics, the more suitable the features to be solved. The problem.
  • each layer 721-726 as shown at 720 in Figure 7 can be a convolutional layer followed by a
  • a layer pooling layer can also be a multi-layer convolution layer followed by one or more pooling layers.
  • the purpose of the pooling layer is to reduce the spatial size of the image.
  • the pooling layer may include an average pooling operator or a maximum pooling operator to sample the input image to obtain a smaller size image.
  • the average pooling operator can calculate the pixel values in the image within a specific range to generate an average value as the result of average pooling.
  • the max pooling operator can take the pixel with the largest value in a specific range as the result of max pooling.
  • the operators in the pooling layer should also be related to the size of the image.
  • the size of the image output after processing by the pooling layer can be smaller than the size of the image input to the pooling layer.
  • Each pixel in the image output by the pooling layer represents the average or maximum value of the corresponding sub-region of the image input to the pooling layer.
  • Neural network model layer 730 is a Neural network model layer 730
  • the convolutional neural network model 700 After being processed by the convolutional layer/pooling layer 720, the convolutional neural network model 700 is not enough to output the required output information. As mentioned before, the convolutional layer/pooling layer 720 only extracts features and reduces the parameters brought by the input image. In order to generate the final output information (required class information or other related information), the convolutional neural network model 700 needs to use the neural network model layer 730 to generate an output or a set of required number of classes. Therefore, the neural network model layer 730 may include multiple hidden layers (731, 732 to 73n as shown in Figure 7) and an output layer 740. The parameters included in the multiple hidden layers may be based on specific tasks. The relevant training data of the type can be pre-trained. For example, the task type can include image recognition, image classification, super-resolution reconstruction of images, etc.
  • the output layer 740 has a loss function similar to classification cross entropy and is specifically used to calculate predictions. Error, once the forward propagation of the entire convolutional neural network model 700 (the propagation from the direction 710 to 740 in Figure 7 is forward propagation) is completed, the back propagation (the propagation from the direction 740 to 710 in Figure 7 is back propagation) ) will start to update the weight values and biases of each layer mentioned above to reduce the loss of the convolutional neural network model 700, and the convolutional neural network The error between the result output by the model 700 through the output layer and the target result.
  • the schematic structural diagram of a convolutional neural network model provided by the embodiment of the present application can be shown in Figure 8.
  • the convolutional neural network model (CNN) 800 shown in FIG. 8 may include an input layer 810, a convolutional layer/pooling layer 820 (where the pooling layer is optional), and a neural network model layer 830.
  • CNN convolutional neural network model
  • FIG. 8 may include an input layer 810, a convolutional layer/pooling layer 820 (where the pooling layer is optional), and a neural network model layer 830.
  • multiple convolution layers/pooling layers in the convolution layer/pooling layer 820 in Figure 8 are parallel, and the respectively extracted features are input to the neural network model layer 230 for processing.
  • the convolutional neural network model shown in Figures 7 and 8 is only a structural example of a neural network model.
  • the convolutional neural network model may also exist in the form of other network models, which is not limited in the embodiments of the present application.
  • Figure 9 shows the hardware structure of a chip provided by an embodiment of the present application.
  • the chip includes a neural network model processor 900.
  • the chip can be disposed in the execution device 610 as shown in FIG. 6 to complete the calculation work of the calculation module 611.
  • the chip can also be installed in the training device 620 as shown in FIG. 6 to complete the training work of the training device 620 and output the target model/rule 601.
  • the algorithms of each layer in the convolutional neural network model shown in Figures 7 and 8 can be implemented in the chip shown in Figure 9.
  • the neural network model processor NPU 50 is mounted on the main central processing unit (CPU) as a co-processor, and the main CPU allocates tasks.
  • the core part of the NPU is the arithmetic circuit 903.
  • the controller 904 controls the arithmetic circuit 903 to extract data in the memory (weight memory or input memory) and perform operations.
  • the computing circuit 903 internally includes multiple processing units (process engines, PEs).
  • the arithmetic circuit 903 may be a two-dimensional systolic array.
  • the arithmetic circuit 903 may also be a one-dimensional systolic array or other electronic circuit capable of performing mathematical operations such as multiplication and addition.
  • the operation circuit 903 is a general matrix processor.
  • the operation circuit obtains the corresponding data of matrix B from the weight memory 902 and caches it on each PE in the operation circuit.
  • the operation circuit obtains matrix A and matrix B from the input memory 901 and performs matrix operations, and the partial result or final result of the obtained matrix is stored in an accumulator (accumulator) 908 .
  • accumulator accumulator
  • the vector calculation unit 907 can further process the output of the operation circuit, such as vector multiplication, vector addition, exponential operation, logarithmic operation, size comparison, etc.
  • the vector calculation unit 907 can be used for network calculations of non-convolutional or non-FC layers in neural network models, such as pooling, batch normalization, and local response normalization. wait.
  • the vector calculation unit 907 can store the processed output vector into the unified buffer 906 .
  • the vector calculation unit 907 may apply a nonlinear function to the output of the arithmetic circuit 903, such as a vector of accumulated values, which may be used to generate an activation value.
  • vector calculation unit 907 generates normalized values, merged values, or both.
  • the processed output vector can be used as an activation input to the arithmetic circuit 903, for example for use in subsequent layers in a neural network model.
  • the unified memory 906 is used to store input data and output data.
  • the storage unit access controller 905 (direct memory access controller, DMAC) is used to transfer the input data in the external memory to the input memory 901 and/or the unified memory 906, store the weight data in the external memory into the weight memory 902, and store the weight data in the external memory into the weight memory 902.
  • the data in the unified memory 906 is stored in the external memory.
  • the bus interface unit (bus interface unit, BIU) 910 is used to realize the interaction between the main CPU, the DMAC and the fetch memory 909 through the bus.
  • An instruction fetch buffer 909 connected to the controller 904 is used to store instructions used by the controller 904;
  • the controller 904 is used to call instructions cached in the fetch memory 909 to control the working process of the computing accelerator.
  • the unified memory 906, the input memory 901, the weight memory 902 and the instruction memory 909 are all on-chip memories, and the external memory is a memory external to the NPU.
  • the external memory can be double data rate synchronous dynamic random access. Memory (double data rate synchronous dynamic random access memory, DDR SDRAM for short), high bandwidth memory (high bandwidth memory, HBM) or other readable and writable memory.
  • the training device 610 in Figure 6 introduced above can perform various steps of the method for obtaining a neural network model in the embodiment of the present application.
  • the chip shown in Figure 9 can also be used to perform various steps of obtaining the neural network model in the embodiment of the present application. step.
  • this embodiment of the present application provides a system architecture 1000.
  • the system architecture includes a local device 1001, a local device 1002, an execution device 1010 and a data storage system 1050, where the local device 1001 and the local device 1002 are connected to the execution device 1010 through a communication network.
  • Execution device 1010 may be implemented by one or more servers.
  • the execution device 1010 can be used in conjunction with other computing devices, such as data storage, routers, load balancers and other devices.
  • Execution device 1010 may be deployed on one physical site, or distributed across multiple physical sites.
  • the execution device 1010 can use the data in the data storage system 1050, or call the program code in the data storage system 1050 to implement the method of obtaining the neural network model or the method of training the neural network model in the embodiment of the present application.
  • a target neural network model can be obtained, and the target neural network model can be used for image classification or image processing, etc.
  • Each local device may represent any computing device, such as a personal computer, computer workstation, smartphone, tablet, smart camera, smart car or other type of cellular phone, media consumption device, wearable device, set-top box, game console, etc.
  • Each user's local device can interact with the execution device 1010 through a communication network of any communication mechanism/communication standard.
  • the communication network can be a wide area network, a local area network, a point-to-point connection, etc., or any combination thereof.
  • the local device 1001 and the local device 1002 obtain the relevant parameters of the target neural network model from the execution device 1010, deploy the target neural network model on the local device 1001 and the local device 1002, and use the target neural network model Perform image classification or image processing, etc.
  • the target neural network model can be directly deployed on the execution device 1010.
  • the execution device 1010 obtains the image to be processed from the local device 1001 and the local device 1002, and uses the target neural network model to classify the image to be processed or other types. image processing.
  • the above execution device 1010 can also be a cloud device, in which case the execution device 1010 can be deployed in the cloud; or the above execution device 1010 can also be a terminal device, in which case the execution device 1010 can be deployed on the user terminal side. Embodiments of the present application There is no limit to this.
  • FIG 11 is a schematic diagram of device deployment of a training model provided by an embodiment of the present application.
  • the training device 1110 can be deployed in a cloud environment.
  • the cloud environment refers to the utilization of cloud computing mode.
  • Basic resources The entity that provides cloud services to users.
  • the cloud environment includes a cloud data center and a cloud service platform.
  • the cloud data center includes a large number of basic resources (including computing resources, storage resources and network resources) owned by the cloud service provider.
  • the cloud data center may include a large number of computing resources.
  • Computing equipment such as servers).
  • the training device 1110 may be a server that performs neural network model training in cloud data, or may also be a virtual machine that trains a neural network model.
  • the training device 1110 can also be a software device deployed on a server or virtual machine in a cloud data center.
  • the software device is used to train a neural network model.
  • the software device can be distributed on multiple servers or distributed. Deployed on multiple virtual machines locally, or deployed on virtual machines and servers in a distributed manner.
  • the training device 1110 can be abstracted by the cloud service provider into a cloud service for training neural network models on the cloud service platform and provided to the user. After the user purchases the cloud service on the cloud service platform, the cloud environment uses the cloud service. Provide users with cloud services for training neural networks.
  • users can upload the neural network model to be trained (or they can also upload the original training set) through the application program interface (API) or through the web interface provided by the cloud service platform. ) to the cloud environment, the training device 1110 receives the neural network model to be trained and the training set, and trains the neural network model to be trained in the training device 1110. The finally trained target neural network is returned to the user's location by the training device 1110. Edge devices.
  • API application program interface
  • edge devices refer to any device with computing resources and network resources between the data generation source and the cloud center.
  • mobile phones are the edge devices between people and the cloud center
  • gateways are the edge devices between smart homes and the cloud center.
  • edge devices are devices that analyze or process data close to the source of the data. Since there is no data flow, network traffic and response time are reduced.
  • the edge device in the embodiment of the present application may be a mobile phone, a tablet personal computer (TPC), a media player, a smart home, a laptop computer (LC), or a personal digital assistant with computing capabilities.
  • assistant PDA
  • personal computer PC
  • camera camcorder
  • smart watch wearable device
  • self-driving vehicle etc. It should be understood that the embodiments of this application do not limit the specific form of the edge device.
  • the user can upload the target task type to the cloud environment through the application program interface or the web interface provided by the cloud service platform. Furthermore, the user can also upload the training set, and the training device receives the target task type and the training set through the training device 1110
  • the neural network model corresponding to the type of the target task for example, the image recognition model to be trained
  • the finally trained target neural network model for example, the target image recognition model
  • the above-mentioned training device 1110 can be deployed in a cloud environment as shown in (a) of Figure 11; or the above-mentioned training device 1110 can also be a terminal device. In this case, the execution device 1110 can be deployed on the user terminal side.
  • This application implements This example is not limiting.
  • a common neural network model transmission scheme is to treat the neural network model as a source for processing, and obtain the neural network data that meets the channel bandwidth requirements through processes such as source coding and channel coding, and then use the wireless channel to It is sent to the receiving device. After receiving the data, the receiving device performs processes such as channel decoding and source decoding to complete the transmission of the neural network model.
  • some AI technologies such as network pruning, network parameter quantification and knowledge distillation, can be used to obtain a neural network model with a relatively small size, thus reducing the number of tasks without significantly affecting the accuracy of inference. Storage space and inference delay at the receiving end.
  • the above transmission scheme still combines the traditional As a transmission channel for neural network models, the communication system requires the same network at the sending and receiving ends. It is still essentially a traditional end-to-end communication system. Multiple neural network models transmit independently of each other and do not involve the communication process in AI applications. training and inference, which leads to low transmission performance of neural network models, making it difficult to meet the current transmission needs of multiple neural network models.
  • this application proposes a method for joint transmission of neural network models in order to meet the high-performance transmission requirements of multiple neural network models in future communication scenarios. It combines neural networks with traditional communication systems and transmits different data according to task requirements. Multiple neural network models on the terminal are combined through the channel to generate a new neural network model, thereby realizing the transmission requirements of the receiving terminal device for multiple neural network models, which is conducive to controlling the diversity and effectiveness of AI applications and avoiding excessive training.
  • the transmission of a single neural network reduces the resource overhead and delay of multiple neural network models under multi-task requirements and multi-modality, and saves the storage space of the neural network models in the receiving and sending end devices to a certain extent.
  • the neural network model refers to the program and data used to perform cognitive calculations obtained through a large amount of labeled training data.
  • the neural network model includes neural network architecture components and neural network parameter components.
  • the neural network architecture component refers to the network and its hierarchical structure related to the neural network algorithm in the neural network model, which is used to perform cognitive computing in the above neural network model. program of.
  • the neural network parameter component refers to a large number of parameters obtained when training the neural network model, as the values of the neurons in the neural network architecture, that is, the data used to perform cognitive calculations in the above-mentioned neural network model.
  • the model transmission method according to the embodiment of the present application will be introduced in detail below with reference to Figure 12 .
  • Both the sending device and the terminal device in the method shown in Figure 12 can be executed by a network device or a terminal device, which is not limited in this embodiment of the present application.
  • the sending device is a network device
  • the receiving device may be a terminal device or a network device.
  • the method 1200 includes steps S1210 to S1240. Steps S1210 to S1240 will be described in detail below.
  • the system suitable for the method shown in Figure 12 may include at least one sending end device and one receiving end device.
  • the method of the embodiment of the present application can be used in a scenario where multiple neural network models in one sender perform joint transmission, or in a scenario where at least one neural network model in multiple different senders perform joint transmission.
  • the receiving end device may also be called a receiving end device
  • the sending end device may also be called a sending end device, which is not limited in this application.
  • the following uses a joint transmission scenario of multiple neural network models in a sending end device as an example to describe the model transmission method of this application in detail.
  • S1210 The sending device obtains information of N first models.
  • the N first models can correspond to N first tasks, and N is an integer greater than or equal to 2.
  • the N first models can be determined based on task requirement information.
  • the sending end device can determine the task requirement information by receiving the task requirement information.
  • Determine the multiple sender neural network models that meet the business requirements, or the sender neural network model that meets the business requirements can also be determined by the cloud device or cloud service platform involved in Figure 11 above based on the business requirement information.
  • the above-mentioned device can be understood as a server that provides services to the sending device.
  • the N first models can be understood as pre-trained neural network models.
  • the sending end can be trained according to the source data set (such as the Image net data set) to meet the business requirements.
  • the neural network model is used by the sending end.
  • the pre-trained neural network model can be completed in the offline stage, also This can be done online.
  • the sending device obtains the neural network model based on the task requirement information, it can directly obtain the trained neural network model without performing pre-training online. Training operations.
  • N first models are any neural network models that meet the task requirements, and the embodiments of the present application do not limit this.
  • the above task requirement information may include external feature values of the target neural network model, and the external feature values of the neural network model may be understood as values that will not change during the training process during the training of the neural network model.
  • the external feature values may also be external feature values that have been determined artificially or based on experience before training the neural network model.
  • the external feature values of the target neural network model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • the above neural network calculation graph information can be used to indicate the structure of the neural network, where the structure of the neural network includes one or more of the following: the type of neural network, the number of layers of the neural network, and the number of nodes in each layer of the neural network. , node connection method of neural network.
  • the types of neural networks involved in the embodiments of this application may be fully connected neural networks, convolutional neural networks, deep neural networks, recurrent neural networks, etc., which are not limited in the embodiments of this application.
  • Hyperparameters in neural networks can include three categories, namely network parameters, optimization parameters, and regularization parameters.
  • network parameters refer to the interaction mode between network layers (for example, addition, multiplication or concatenation, etc.), the number and size of convolution kernels, the number of network layers (or depth) and activation functions etc.
  • Optimization parameters refer to the learning rate (learning rate), the number of batch samples (batch size), the parameters of different optimizers (for example, the momentum parameter momentum) and the adjustable parameters of some loss functions.
  • Regularization parameters refer to the weight attenuation coefficient and the dropout parameters of the neural network model, etc.
  • the hyperparameters of the neural network model include parameters that remain unchanged during the training process of the neural network model, that is, the hyperparameters are not obtained through the training of the neural network model, but are usually determined before the training of the neural network model.
  • the hyperparameter information can be manually set based on experience or task requirements, and the embodiment of the present application does not limit this.
  • the N first models may correspond to N signals A.
  • N signals A are used to carry information of the corresponding first model.
  • the sending end edge device can encode the information of N first models through its own encoder and generate N signals A.
  • the sending device obtains the corresponding N signals A after sequentially processing the information of the N first models through source coding, channel coding, modulation, and up-conversion or down-conversion.
  • the information of the model may include the external characteristic values of the neural network model and/or the parameters of the neural network model.
  • the information of the model of the present application is not limited to the first model or the target model, as long as it is of the present application. Emerging neural network models can contain this information.
  • the external feature values of the neural network model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information;
  • the parameters of the neural network model include one or more of the following: : Weight matrix, weight vector, bias matrix, bias vector, activation function.
  • the N signals A corresponding to the N first models may be analog signals or digital signals. No., the embodiments of this application do not limit this.
  • N signals A correspond one-to-one to the above-mentioned N first models, where N is an integer greater than or equal to 2.
  • the sending device obtains the fusion auxiliary information.
  • the fusion auxiliary information includes external feature values of the target model.
  • the external feature value of the target model includes one or more of the following information: neural network calculation graph information of the target model, optimizer information of the target model, and hyperparameter information of the target model.
  • the target model may correspond to a second task.
  • the second task may be the same as or different from the N first tasks corresponding to the N first models.
  • the first first model in the sending end can be used to detect trees in the image
  • the second first model can be used to detect birds in the image.
  • S1230 Determine N first signals based on the information of N first models and the fusion auxiliary information.
  • the N first signals are obtained by N first modules processing the information of the above-mentioned N first models and the fusion auxiliary information.
  • the first model may be a module that implements a specific function, or a neural network model that meets task requirements. As an example, the embodiment of the present application does not limit this.
  • the N first modules are trained neural network models trained based on training data, where the N first modules can correspond one-to-one to the above-mentioned N first models.
  • the N first modules can be understood as N neural network models, and the neural network models correspond to the above-mentioned N first models one-to-one.
  • the one-to-one correspondence between the N first modules and the above-mentioned N first models can be understood as the one-to-one correspondence between the N business requirements corresponding to the N first modules and the N business requirements corresponding to the above-mentioned N first models.
  • the first model and the first module that correspond to each other one-to-one correspond to the same business requirement.
  • the business requirements corresponding to the N first modules and N first models may all be image classification services.
  • the type of training data is related to the task requirements of the neural network model.
  • the training data may be images.
  • the training data can be text.
  • the training data may be speech data.
  • the embodiments of the present application do not limit this.
  • the above training data includes M first training signals and data of the target model.
  • the data of the target model can be understood as the external characteristic values of the target model, that is, the neural network calculation graph information and optimizer of the target model. information and hyperparameter information.
  • the above-mentioned M first training signals correspond to M training models one-to-one, and the M training models correspond to M business requirements.
  • N services corresponding to the N first models include M services corresponding to the M first training models.
  • the sending device may determine the first module corresponding to the i-th first model based on information about at least one first model among the N first models except the i-th first model.
  • the information of the at least one first model refers to the parameters and external characteristic values of the neural network model, where the parameters of the neural network model refer to the neural network model.
  • the parameters of the neural network model refer to the neural network model.
  • the parameters of the above model refer to one or more of the following: weight matrix, weight vector, bias matrix, bias vector, and activation function.
  • the services corresponding to the two first models are both target detection services in the image processing service
  • the first first model can be used to detect trees in the image
  • the second first model can be used to detect the image.
  • the first first model can obtain the first first module corresponding to the first first model based on the parameters (for example, weight matrix, weight vector, etc.) in the second first model , wherein the services of the first first module correspond to the services of the first first model, and the first first module can be used to detect birds and trees in the image at the same time.
  • the sending end device can obtain M original models, which correspond to the M first training signals one-to-one, where M is an integer greater than or equal to 2; the sending end sequentially inputs the M first training signals with one of them. One corresponding M original models are used to obtain M second training signals, where the M second training signals correspond one-to-one to the above-mentioned M first training signals.
  • the transmitting end superimposes the above-mentioned M second training signals in the same channel to obtain a third training signal; the receiving end device obtains the intermediate model by analyzing the third training signal. It should be understood that the receiving end device can The intermediate model corresponding to the third training signal is obtained by decoding the third training signal; the sending end device adjusts the parameters of the above M original models according to the deviation between the intermediate model and the target model to obtain the training The final model library.
  • the trained model library includes the original model after the above adjustment process, and the deviation between the adjusted intermediate model and the target model is within a preset range, or it can also be understood as the output value of the intermediate model. Compare with the target value of the target model until the difference between the output value of the intermediate model and the target value is less than a certain preset threshold, that is, when the value of the loss function reaches the preset threshold, the training of the intermediate model is stopped.
  • the sending device can obtain the N first modules from the model library obtained after training. Specifically, the sending device can search the trained model library for a neural network model that meets the current task requirements as the first module according to the task requirements.
  • the above original model can be determined based on task requirement information, that is, the original model can be trained based on a source data set (for example, ImageNet, Labelme, etc. data set) that meets the task requirement information.
  • a source data set for example, ImageNet, Labelme, etc. data set
  • training stop conditions can be flexibly set according to actual application requirements and application scenarios.
  • training stop conditions may include but are not limited to, the training time reaches a preset duration, the number of training times reaches a preset number, the value of the loss function or the attenuation amplitude reaches a preset threshold, etc.
  • N first models can be processed to obtain N signals A that correspond to them one-to-one. After the N signals A are processed by the N first modules after the above training, N signals A can be obtained. The first signal, the N first signals are superimposed in the same channel to obtain a second signal corresponding to the target model.
  • the trained neural network model can perform the target task.
  • the target task can be an image processing task, including: target detection, image segmentation, instance segmentation, image denoising, image super-resolution, etc.
  • the target task may also be a speech processing task, such as speech recognition.
  • the target task may also be a text processing task, such as text recognition or text translation.
  • the training process for the N first modules can be executed by the training device of the neural network model.
  • the training device can be executed by a cloud device or a terminal device, for example, an edge device.
  • Devices, computers, servers and other devices with sufficient computing power to execute the training method of the neural network model can also be executed by a system composed of cloud devices and terminal devices.
  • the training process of the above model can be performed by the training device 620 in Figure 6, the neural network processor 900 in Figure 9, the execution device 1010 in Figure 10 or the local device 1001 or 1002, and the training device 1110 in Figure 11 Execution, the embodiment of this application does not limit this.
  • the training process of the N first modules can be completed in an offline stage or can be completed online.
  • the sending device can directly obtain the trained neural network model without performing pre-training operations online.
  • the sending end device processes the above-mentioned N signals A according to the N first modules to obtain N first signals.
  • the sending device processes N signals A corresponding to N first models to obtain N first signals.
  • the sending end device processes N first models to obtain N signals A, and processes the N signals A according to the acquired N first modules to obtain Nth signals A. A signal.
  • the N signals A correspond to the N first models in a one-to-one correspondence
  • the N first modules correspond to the N first models in a one-to-one correspondence.
  • the N first modules are trained based on training data, and the training data includes M first training signals and target models.
  • the M first training signals correspond to M training models, and each first training signal is a The signal of the corresponding training model, the target model can meet the business requirements corresponding to the M training models, and the M services corresponding to the M first training models include N services corresponding to the N first models.
  • the n-th first signal among the N first signals is a signal obtained by processing the n-th signal A according to the n-th first module.
  • the n-th signal A is the signal corresponding to the n-th first module.
  • the N signals A correspond to the N first models one-to-one, and the first models and the first modules corresponding to each other correspond to the same service, n ⁇ [1, N], N is greater than or An integer equal to 2, M is an integer greater than or equal to 2;
  • the first first signal among the five first signals generated by the transmitting end is a signal obtained by processing the first signal A according to the first first module, where the first first signal is the signal obtained by processing the first signal A according to the first first module.
  • One signal A corresponds to the first first model, and the first first model and the first first module correspond to the same service.
  • the N first signals can be superimposed into a second signal in the channel, and the second signal corresponds to the target model.
  • the second signal can be decoded by the receiving end device to obtain the target model data required for the task, and further obtain the target neural network model.
  • the sending end device can also determine the above-mentioned N first modules through parameters of the channels for sending the N first signals, wherein the N first modules are the same as the N first signals.
  • the N first modules are the same as the N first signals.
  • the parameters of the above channel include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the sending device can obtain the configuration information of the neural network model based on the channel state information and signal-to-noise ratio in the channel parameters, and further determine N first modules that correspond one-to-one to the above N first models.
  • the sending end device can send reference information to the receiving end device.
  • the reference information is information provided by the sending end to the receiving end device for channel estimation or channel detection.
  • the sending end device can send the reference information.
  • the channel information is measured in real time, and then the receiving device will determine based on the obtained channel parameters.
  • Configuration parameters of the neural network model of the sending end device thereby achieving the purpose of obtaining N first modules.
  • the configuration parameters of the neural network model include external eigenvalues of the neural network model and/or parameters of the neural network model.
  • the external eigenvalues of the neural network model include one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information; the parameters of the neural network model include one or more of the following: weight matrix , weight vector, bias matrix, bias vector, activation function.
  • the sending end device sends the generated N first signals to the receiving end device, wherein the N first signals are superimposed to generate a second signal, and the receiving end device receives the second signal.
  • the data of the target model that meets the task requirements can be obtained. And further obtain a target neural network model, where the target model corresponds to the second signal.
  • model transmission method 1200 can also be represented by the flow chart of model transmission shown in Figure 13. It should be understood that the steps involved in the sending end have been described in detail in the above-mentioned Figure 12 and will not be described again here.
  • N first signals obtained by the sending end device can be superimposed in the channel to obtain a second signal, where the second signal corresponds to the target model.
  • the receiving device can process the second signal through the analysis module to obtain the target model information, and finally obtain the target neural network model that meets the task requirements.
  • the analysis method corresponding to the analysis module is related to the external eigenvalues of the model.
  • the method for joint transmission of neural network models in multiple sending end devices is similar to the above method, and will not be described again here for the sake of brevity.
  • joint transmission of multiple different neural network models can be achieved, and the resource overhead and delay of multiple neural network models in the transmission process can be reduced, and to a certain extent, the receiving end device can reduce the impact of neural network models on the neural network.
  • Figure 14 is a schematic flowchart of an example of a model transmission method provided by an embodiment of the present application. The method includes steps S1410 to S1460. Steps S1410 to S1460 will be described in detail below.
  • each sending device there may be one or more sending devices.
  • multiple neural networks in a sending device jointly transmit models to a receiving device, or three sending devices jointly transmit models to a receiving device.
  • the receiving end sends task requirement information to the sending end.
  • the sending end obtains the task requirement information sent by the receiving end.
  • the receiving end device can send task requirement information to the sending end device, and the sending end device selects a model that meets the task requirements from the receiving end neural network model library obtained through offline training based on the received task information.
  • a group or multiple groups of neural network models serve as the neural network model of the sending end. It should be understood that one or more sets of neural network models determined here can be understood as the N first models in the above method 1200.
  • the task requirement information may also include the accuracy of the task requirement and the performance of the receiving end device.
  • the above task requirement information may also include a decoding method of the target neural network model corresponding to the task requirement information.
  • the decoding method of the target neural network model is consistent with the target neural network model that can be expressed by the receiving end knowledge base.
  • the neural network after decoding by the receiving end should be the neural network model that can be expressed in the knowledge base.
  • the sending end determines the external characteristic value of the target neural network model based on the task requirement information.
  • the sending end after the sending end receives the task requirement information sent by the receiving end, it can select multiple neural network models that meet the task requirements as the sending end neural network model.
  • the combination scheme of the transmitter neural network model may be one or more combinations of the transmitter neural network model in the implementable manner obtained after the above channel communication operation method.
  • the external characteristic value of the target neural network model includes one or more of the following information: neural network calculation graph information, optimizer information, and hyperparameter information.
  • the external feature values of the neural network model in the embodiments of the present application can be understood as feature values that will not change due to training during the training of the neural network model, wherein the external feature values can also be in the neural network model. Feature values have been determined during model training.
  • the above neural network calculation graph information can be used to indicate the structure of the neural network, where the structure of the neural network includes one or more of the following: the type of neural network, the number of layers of the neural network, and the number of nodes in each layer of the neural network. , node connection method of neural network.
  • neural networks involved in the embodiments of this application may be fully connected neural networks, convolutional neural networks, deep neural networks, recurrent neural networks, etc., which are used as examples rather than limitations, and are not limited in the embodiments of this application.
  • Hyperparameters in neural networks can include three categories, namely network parameters, optimization parameters, and regularization parameters.
  • network parameters refer to the interaction mode between network layers (for example, addition, multiplication or concatenation, etc.), the number and size of convolution kernels, the number of network layers (or depth) and activation functions etc.
  • Optimization parameters refer to the learning rate (learning rate), the number of batch samples (batch size), the parameters of different optimizers (for example, the momentum parameter momentum) and the adjustable parameters of some loss functions.
  • Regularization parameters refer to the weight attenuation coefficient and the dropout parameters of the neural network model, etc.
  • the hyperparameters of the neural network model include parameters that remain unchanged during the training process of the neural network model, that is, the hyperparameters are not obtained through the training of the neural network model, but are usually determined before the training of the neural network model.
  • the information of the hyperparameter can be manually set based on experience or task requirements, and the embodiment of the present application does not limit this.
  • the sending device sends the fusion auxiliary information to the receiving device.
  • the receiving end device receives the fusion auxiliary information sent by the transmitting end.
  • the fusion auxiliary information is used to indicate the external feature values of the target neural network model that meet the task requirements.
  • the receiving end device after receiving the first information, performs the calculation according to the external characteristic value of the target neural network model in the fusion auxiliary information (including one or more of the following information: neural network calculation graph information, optimization information, hyperparameter information) to reserve the storage space of the target neural network model.
  • the fusion auxiliary information including one or more of the following information: neural network calculation graph information, optimization information, hyperparameter information
  • the sending end device and the receiving end device can achieve information synchronization of the external feature values of the target neural network model between the sending and receiving end devices by maintaining information on the external feature values of multiple target neural network models. .
  • the list of external feature values of the target neural network model in the knowledge base can be identified and distinguished using identification information.
  • the list of external feature values of the target neural network model in the synchronization knowledge base can be represented by the following table:
  • the first columns of Tables 1 to 3 respectively represent the external eigenvalues of the neural network model, and the second column represents the ID information corresponding to the external eigenvalues of the neural network model.
  • the transceiver device can synchronize the target neural network model information of the transceiver device through the first information, that is, the ID combination in the above list.
  • the target network model structure includes a convolution layer with a convolution kernel size of 3*3
  • the selected network model optimizer is the root mean square propagation (root mean square prop, RMSProp) optimization algorithm and the network model
  • the hyperparameters are: learning rate 0.01, and the value of the momentum hyperparameter is 0.9.
  • the network model optimizer in Table 2 above is related to the hardware capabilities of the receiving end device. For example, in some cases, due to the limited computing power of the terminal device, only a simpler network model structure or optimizer can be run. .
  • the fusion auxiliary information also includes reference information, which is used to provide the receiving end with channel estimation or channel detection to obtain channel parameters.
  • the sending end can perform real-time measurement of channel information by sending reference information, and then the receiving end will determine the configuration parameters of the sending end's neural network model based on the obtained channel parameters, and then obtain the trained sending end neural network model. the goal of.
  • the configuration parameters of the neural network model include external eigenvalues of the neural network model and/or parameters of the neural network model.
  • the external eigenvalues of the neural network model include one or more of the following: neural network calculation graph information, optimizer information, and hyperparameter information; the parameters of the neural network model include one or more of the following: weight matrix, Weight vector, bias matrix, bias vector, activation function.
  • the above steps S1410 to S1430 may be executed by a cloud device that provides services for the sending device.
  • the receiving end device sends the fusion auxiliary information to the sending end device.
  • the sending end device receives the fusion auxiliary information sent by the receiving end device to realize the synchronization of the external feature values of the target neural network model by the sending and receiving ends.
  • S1450 The sending end performs joint transmission of the neural network model.
  • this step can be performed in the manner of the above-mentioned method 1200, which will not be described in detail here.
  • the embodiment of this application provides a channel communication operation method, which can be expressed as
  • the network model obtained by the sending end device is processed by the encoding module E i and the first module to generate a neural network model E i (M i ), and the neural network model obtained after channel superposition is:
  • h i represents the channel passed by the i-th transmitter neural network model
  • n represents the noise interference signal
  • the above formula In formula (3) The symbol represents the superposition operation process of the neural network model.
  • the superimposed neural network model generated by the channel can be expressed as:
  • the neural network model generated by the transmitting device through the channel is processed by the receiving decoding module D and the resulting neural network model can be expressed as:
  • the neural network model finally obtained by the decoding module D at the receiving end can be expressed as:
  • M 4 is the neural network model finally obtained by the receiving end that meets the task requirements.
  • the neural network model involved in the embodiment of the present application can be regarded as a classic directed graph structure, where the directed graph is a directional graph consisting of a set of vertices and a set of directional It consists of edges, each edge connects a pair of ordered vertices.
  • the vertex can be regarded as a neuron node or a neural unit.
  • the connections between different neuron nodes each correspond to a weight value, which can be adjusted through the training process.
  • the neural network model in Figure 15 can be divided into input layers, hidden layers and output layers, and the neural network model can be stored and expressed through a directed graph data structure. Therefore, the channel communication operation at the basic graph level can be expressed as a combination of directed graph structures corresponding to different neural network models, that is, the same parts of the directed graph data structures of the neural network models corresponding to multiple senders are superimposed. .
  • the joint transmission process of the neural network model represented by the basic graph level in the channel communication operation can be divided into the following steps:
  • Neural network model m 1 is a 3-layer neural network, including a hidden layer (also called layer 1), in which the input layer (also called layer 0) has 2 neurons and the hidden layer has 1 neuron. Yuan, the output layer (also called layer 2) has 2 neurons.
  • the neural network model m 2 is a 3-layer neural network, which also includes a hidden layer. The input layer of the neural network model m 2 has 2 neurons, the hidden layer has 2 neurons, and the output layer has 2 neurons.
  • the above neural network model m 1 can be further expressed as:
  • the neural network model m 2 can be expressed as:
  • l represents the network node in the directed graph structure corresponding to the neural network model.
  • the network node can also be called a neuron or network unit.
  • the superscript of l represents the number of layers where the neurons in the neural network model are located, and the subscript represents The position number of the neuron of the neural network model in the directed graph in the layer to which it belongs.
  • the sending end sends the neural network calculation graph information, optimizer, hyperparameters and other characteristic values corresponding to the neural network model to the receiving end through the synchronized knowledge base information.
  • the receiving end reserves the information that meets the task requirements based on the received synchronized knowledge base network information.
  • each sending end neural network model passes through the encoding module and is processed by the trained neural network model, and the signal obtained is sent to the receiving end through the same channel.
  • the signals of the neural network model sent by the transmitter can be superimposed in the channel to obtain the signal corresponding to the target neural network model.
  • the combined neural network model m 3 obtained by the receiving end is a 3-layer neural network.
  • the neural network model m 3 contains a hidden layer, in which the input layer contains 3 neurons and the hidden layer contains 2 Neurons, the output layer contains 3 neurons.
  • the combined neural network model M 3 can be expressed as:
  • neural network model m 3 and The other parts of the neural network model m 3 are the parts that need to be superimposed with the help of channel communication operations, and the other parts of the neural network model m 3 are parts that each sending end can transmit independently.
  • the neural network model sent by the transmitting end can send parameters of the neural network model that need to be superimposed on the same channel, and the parameters of the remaining neural network models can be sent on orthogonal channels.
  • f ref f 2 -f 1 .
  • a 1 and A 2 represent the amplitude values corresponding to different clock frequencies, and t represents the time spent transmitting the reference signal.
  • the signal received by the receiving end is:
  • B 1 and B 2 represent the reference signal passing through the channel, and the amplitude values corresponding to different clock frequencies obtained at the receiving end. and Respectively represent the reference signal passing through the channel, and the phase values corresponding to different clock frequencies obtained at the receiving end.
  • a channel communication operation provided by the embodiment of the present application can also be represented at the computational graph level as shown in Figure 16.
  • the computational graph is a language used to describe calculations and is a method of formalizing calculations.
  • calculations can be represented as directed graphs.
  • Each node or module in the graph represents a variable, where the variable can be a scalar, vector, matrix, tensor, etc.
  • the calculation graph may include a backbone network, convolutional layers, pooling layers and fully connected layers. Among them, the function of the convolutional layer is to extract features from the input data. It contains multiple convolution kernels. Each element making up the convolution kernel corresponds to a weight coefficient and an offset.
  • the role of the pooling layer is to compress the features in the convolutional layer and extract the main features.
  • common pooling operations include maximum pooling and average pooling.
  • the fully connected layer is used to connect all feature values and send the output values to the classifier.
  • the fully connected layer is generally located after the pooling layer and converts all feature matrices of the pooling layer into one-dimensional feature vectors. At the same time, the fully connected layer is generally located at the last layer of the convolutional neural network structure.
  • the neural network model can be expressed as the data output by the backbone network at the computational graph level shown in Figure 16.
  • the convolution layer module of the neural network model m 1 includes convolution The convolutional layer with a kernel size of 1*1 and the convolutional layer with a convolutional kernel size of 3*3.
  • the neural network model m 2 can be expressed as the data of the neural network model output after the data output by the backbone network passes through the convolution layer, the pooling layer and the fully connected layer in sequence, where the convolution layer includes a size of 5*5 convolution kernel.
  • the neural network model m 3 obtained through the combination of channel communication operations can be represented by a calculation diagram as shown in Figure 16.
  • the data output through the backbone network (for example, the data can be a picture of 200*200 pixels) successively passes through a convolution layer including a convolution kernel size of 1*1 and a convolution layer including
  • the data output after the convolution layer with a kernel size of 3*3 and the data output through the backbone network after passing through the convolution layer with a convolution kernel size of 5*5 arrive at the pooling layer at the same time, and are processed in the pooling layer.
  • the layers are overlaid.
  • neuron nodes can be regarded as specific network modules.
  • the cascade between these network modules can be realized through the communication process, and the communication paths can be switched to meet task requirements.
  • the backbone network in the calculation diagram can be located in the base station, and the functions of the remaining modules (including the convolution layer, the pooling layer and the fully connected layer) can be executed by one terminal device or multiple terminal devices respectively.
  • This application implements This example does not limit this.
  • the extraction results of different features of the data can be achieved through communication routing, thereby improving the expression ability and performance of the neural network.
  • the process of joint transmission of the neural network model can be regarded as the reverse process of the multicast routing algorithm, as shown in Figure 17, which is the neural network model transmission at the computational graph level.
  • Figure 17 is the neural network model transmission at the computational graph level.
  • the shunt point can be used as a node that the neural network needs to perform communication operations, and The neural network models are superimposed and combined at this branching point to obtain the final target neural network model.
  • the N first modules obtained after training are processed according to the information of the N first models and the fusion auxiliary information of the N first models to obtain N first signals, and the sending end device
  • the processed N first signals are sent to the receiving end device, and the N first signals are superimposed in the channel to obtain a second signal.
  • the second signal corresponds to a target model that meets the task requirements.
  • the receiving end receives the target neural network model.
  • the receiving end device receives the second signal.
  • the receiving end device can decode the second signal through the decoding module of the receiving end, and further perform analytical processing to obtain the target neural network model.
  • the second signal is obtained by superposing N first signals sent by the sending end device in the channel.
  • one or more sending end devices can determine multiple neural network models based on the task requirement information sent by the receiving end device, and send the combined neural network models to the receiving end device through joint transmission. , compared with the traditional independent model transmission method, it can significantly reduce the resource overhead and delay during the transmission process, and effectively reduce the storage space reserved by the receiving end device for the target neural network model.
  • FIG. 18 is a schematic flowchart of another example of a model transmission method provided by an embodiment of the present application.
  • the method includes steps S1810 to S1860. Steps S1810 to S1860 are described in detail below. Different from the model transmission method shown in Figure 14, steps S1810 to S1830 and step S1850 of the method shown in Figure 18 can be executed by a server (eg, cloud device) that provides services for the receiving device. It should be understood that steps S1810 to S1860 are similar to steps S1410 to S1460 in FIG. 14 , and detailed description thereof is omitted here to avoid redundancy.
  • a server eg, cloud device
  • cloud devices can be introduced to process the task requirement information sent by the receiving end device.
  • the information determines multiple neural network models, and the cloud device sends the combined neural network models to the receiving device through joint transmission.
  • the cloud device can significantly reduce the resources in the transmission process. Overhead and delay are beneficial to control the diversity and effectiveness of AI applications, and effectively reduce the storage space reserved by the receiving end device for the target neural network model.
  • Figure 19 is a schematic block diagram of a model transmission device provided by an embodiment of the present application.
  • the device for model transmission shown in Figure 19 includes an acquisition unit 1910, a processing unit 1920 and a sending unit 1930.
  • the acquisition unit 1910, the processing unit 1920, and the sending unit 1930 may be used to perform the model transmission method in the embodiment of the present application, for example, may be used to perform the method 1200 in Figure 12 above.
  • the acquisition unit 1910 is configured to receive information of N first models and fusion auxiliary information.
  • the fusion auxiliary information includes external feature values of the target model, and the target model corresponds to the second task.
  • the processing unit 1920 determines N first signals based on the information of the N first models and the fusion auxiliary information, and sends the N first signals, where the N first models correspond to N first tasks, N is an integer greater than or equal to 2, and the second task is different from the first task.
  • the sending unit 1930 is configured to send the N first signals to the receiving end device.
  • the acquisition unit 1910 may also be a receiving unit, a processing unit, or a unit with receiving and processing functions.
  • the sending device when the obtaining unit 1910 is a receiving unit, can receive the information of the first model sent by the cloud device or the receiving device; when the obtaining unit 1910 is a processing unit, the sending device The device can be used to determine the information of the first model locally; when the acquisition unit 1910 is a unit with receiving and processing functions, the sending device can receive the information of the first model sent by the cloud device or the receiving device, or can Information about the first model is determined locally.
  • the processing unit 1920 determines the N first signals through the first module.
  • the first module is determined based on parameters of a channel that sends N first signals.
  • the parameters of the above channel include one or more of the following: channel state information CSI, signal-to-noise ratio SNR, and link quality.
  • the channel state information may include: channel quality indicator CQI, rank indicator RI, and precoding matrix PMI.
  • the sending device can obtain the configuration information of the neural network model based on the channel state information and signal-to-noise ratio in the channel parameters, and further determine N first modules that correspond one-to-one to the above N first models.
  • the device for model transmission provided by the embodiment of the present application further includes a first module configured to determine N first signals based on the N first models and the fusion auxiliary information.
  • the first module corresponding to the i-th first model among the N first models is based on at least one first module other than the i-th first model among the N first models.
  • the model information is determined, i ⁇ [1, N].
  • the N first modules are trained based on training data.
  • the training data includes M first training signals and target models.
  • the M first training signals are consistent with the M training models.
  • the target model can satisfy the tasks corresponding to the M training models, where M is an integer greater than or equal to 2.
  • the processing unit 1920 is also configured to obtain M original models, and input each first training signal into the corresponding original model to obtain M second training signals, and convert the M The second training signals are superposed in the same channel to obtain the third training signal, the third training signal is analyzed to obtain the intermediate model, and the parameters of the M original models are adjusted according to the deviation between the intermediate model and the target model.
  • M original models correspond to M first training signals one-to-one, M is an integer greater than or equal to 2
  • the model library includes the original model after adjustment, and after adjustment
  • the deviation between the processed intermediate model and the target model is within the preset range.
  • the sending unit 1930 is also configured to send the fusion auxiliary information of the N first models to the receiving end device.
  • unit here can be implemented in the form of software and/or hardware, and is not specifically limited.
  • a "unit” may be a software program, a hardware circuit, or a combination of both that implements the above functions.
  • the hardware circuit may include an application specific integrated circuit (ASIC), an electronic circuit, a processor (such as a shared processor, a dedicated processor, or a group processor) for executing one or more software or firmware programs. etc.) and memory, merged logic circuitry, and/or other suitable components to support the described functionality.
  • ASIC application specific integrated circuit
  • processor such as a shared processor, a dedicated processor, or a group processor for executing one or more software or firmware programs. etc.
  • memory merged logic circuitry, and/or other suitable components to support the described functionality.
  • the units of each example described in the embodiments of the present application can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each specific application, but such implementations should not be considered beyond the scope of this application.
  • FIG. 20 is a schematic diagram of the hardware structure of a training device for a neural network model provided by an embodiment of the present application.
  • the neural network model training device 2000 shown in FIG. 20 includes a memory 2010, a processor 2020, a communication interface 2030, and a bus 2040.
  • the memory 2010, the processor 2020, and the communication interface 2030 implement communication connections between each other through the bus 2040.
  • the memory 2010 may be a read only memory (ROM), a static storage device, a dynamic storage device or a random access memory (RAM).
  • the memory 2010 can store programs.
  • the processor 2020 is used to execute various steps of the training method of the neural network model in the embodiment of the present application. Specifically, the processor 2020 may execute method 1200 shown in FIG. 12 above.
  • the processor 2020 may adopt a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), a graphics processing unit (GPU), or one or more
  • the integrated circuit is used to execute relevant programs to implement the training method of the neural network model in the method embodiment of the present application.
  • the processor 2020 may also be an integrated circuit chip with signal processing capabilities, for example, it may be the chip shown in FIG. 9 .
  • each step of the training method of the neural network model of the present application can be completed by instructions in the form of hardware integrated logic circuits or software in the processor 2020 .
  • the above-mentioned processor 2020 can also be a general-purpose processor, a digital signal processing (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, Discrete gate or transistor logic devices, discrete hardware components.
  • DSP digital signal processing
  • ASIC application-specific integrated circuit
  • FPGA field programmable gate array
  • a general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.
  • the steps of the method involved in the embodiments of the present application can be directly implemented by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, registers and other mature storage media in this field.
  • the storage medium is located in the memory 2010 and the processor 2020 Read the information in the memory 2010, and combine its hardware to complete the functions required to be performed by the units included in the training device shown in Figure 19, or perform the transmission method of the neural network model shown in Figure 12 of the method embodiment of the present application.
  • the communication interface 2030 uses a transceiver device such as but not limited to a transceiver to implement communication between the device 2000 and other devices or communication networks. For example, training data can be obtained through communication interface 2030.
  • Bus 2040 may include a path that carries information between various components of device 2000 (eg, memory 2010, processor 2020, communication interface 2030).
  • the device 2000 may also include other devices necessary for normal operation. At the same time, based on specific needs, those skilled in the art should understand that the device 2000 may also include hardware devices that implement other additional functions. In addition, those skilled in the art should understand that the device 2000 may only include components necessary to implement the embodiments of the present application, and does not necessarily include all components shown in FIG. 20 .
  • the processor in the embodiment of the present application can be a central processing unit (CPU).
  • the processor can also be other general-purpose processors, digital signal processors (DSP), or application-specific integrated circuits. (application specific integrated circuit, ASIC), off-the-shelf programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
  • a general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.
  • non-volatile memory may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.
  • non-volatile memory can be read-only memory (ROM), programmable ROM (PROM), erasable programmable read-only memory (erasable PROM, EPROM), electrically removable memory. Erase electrically programmable read-only memory (EPROM, EEPROM) or flash memory.
  • Volatile memory can be random access memory (RAM), which is used as an external cache.
  • RAM random access memory
  • static random access memory static random access memory
  • DRAM dynamic random access memory
  • RAM synchronous dynamic random access memory
  • SDRAM synchronous dynamic random access memory
  • double data rate SDRAM double data rate SDRAM
  • DDR SDRAM enhanced synchronous dynamic random access memory
  • ESDRAM enhanced synchronous dynamic random access memory
  • serial DRAM serial DRAM
  • direct memory bus random access memory direct rambus RAM, DRRAM
  • the above embodiments may be implemented in whole or in part by software, hardware, firmware, or any other combination.
  • the above-described embodiments may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on the computer, the processes or functions described in the embodiments of the present application are generated in whole or in part.
  • the computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another, e.g., the computer instructions may be transferred from a website, computer, server, or data center Transmit to another website, computer, server or data center through wired (such as infrared, wireless, microwave, etc.) means.
  • the computer-readable storage medium may be any available medium that a computer can access, or a data storage device such as a server or a data center that contains one or more sets of available media.
  • the usable media may be magnetic media (eg, floppy disk, hard disk, tape), optical media (eg, DVD), or semiconductor media.
  • Semiconductor medium It can be a solid state drive.
  • At least one refers to one or more, and “plurality” refers to two or more.
  • At least one of the following” or similar expressions thereof refers to any combination of these items, including any combination of a single item (items) or a plurality of items (items).
  • at least one of a, b, or c can mean: a, b, c, a-b, a-c, b-c, or a-b-c, where a, b, c can be single or multiple .
  • the size of the sequence numbers of the above-mentioned processes does not mean the order of execution.
  • the execution order of each process should be determined by its functions and internal logic, and should not be used in the embodiments of the present application.
  • the implementation process constitutes any limitation.
  • the disclosed systems, devices and methods can be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components may be combined or can be integrated into another system, or some features can be ignored, or not implemented.
  • the coupling or direct coupling or communication connection between each other shown or discussed may be through some interfaces, and the indirect coupling or communication connection of the devices or units may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present application can be integrated into one processing unit, each unit can exist physically alone, or two or more units can be integrated into one unit.
  • the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium.
  • the technical solution of the present application is essentially or the part that contributes to the existing technology or the part of the technical solution can be embodied in the form of a software product.
  • the computer software product is stored in a storage medium, including Several instructions are used to cause a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in various embodiments of this application.
  • the aforementioned storage media include: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk and other media that can store program code. .

Landscapes

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

Abstract

本申请实施例提供了一种模型传输的方法及装置，涉及通信技术领域，该方法包括：获取N个第一模型的信息，该N个第一模型对应N个第一任务，其中，N为大于或等于2的整数；获取融合辅助信息，该融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与该N个第一任务不同；根据N个第一模型的信息和该融合辅助信息确定N个第一信号；发送该N个第一信号。本申请的模型传输的方法能够实现多个模型的联合传输，同时可有效降低多个模型在传输过程中的资源开销与时延，并可减少接收端装置对于模型预留的存储空间。

Description

一种模型传输的方法及装置
本申请要求于2022年05月31日提交中国专利局、申请号为202210609813.5、申请名称为“一种模型传输的方法及装置”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请实施例涉及通信技术领域，具体涉及一种模型传输的方法及装置。
背景技术
人工神经网络(artificial neural networks,ANN)是一种模拟生物神经网络进行信息处理的非线性统计性数据建模工具。在多数情况下人工神经网络能在外界信息的基础上改变内部结构，是一种具备学习能力的自适应系统。这类网络依靠系统的复杂程度，通过调整内部大量节点之间相互连接的关系以达到处理信息的目的，已广泛应用于图像分类、人脸识别和语音识别等领域。
随着人工神经网络技术的普及，以及边缘人工智能(artificial intelligence，AI)设备(例如车载移动装置、无人机、智能手机、个人数字助理)的广泛应用，使用神经网络模型的业务也越来越多，随之而来的，可能需要为同一客户端传输多个神经网络模型。
目前，已知一种通信技术，如图1所示，将一个神经网络模型视作一个待传输的数据，对该数据进行编码后生成对应的信号，并通过无线链路进行发送以实现神经网络模型的传输。但这一传输过程并未将通信过程与人工智能应用的训练过程进行结合，且多个神经网络模型之间相互独立传输，进而导致神经网络模型的传输性能较低，难以满足当前业务的需求。
因此，如何实现多个神经网络模型的高效传输以及减小多个神经网络模型在传输过程中的资源开销和时延，已成为业界亟需解决的技术问题。
发明内容
本申请实施例提供一种模型传输的方法及装置，可实现多个模型的联合传输，降低该多个模型在传输过程中的资源开销和时延，并同时减少接收端设备对于目标神经网络模型的存储空间。
第一方面，提供了一种模型传输的方法，应用于发送端装置或发送端装置中的芯片上，该方法包括：获取N个第一模型的信息，该N个第一模型对应N个第一任务，其中，N为大于或等于2的整数；获取融合辅助信息，该融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与第一任务不同；根据N个第一模型的信息和该融合辅助信息确定N个第一信号；发送该N个第一信号。根据本申请实施例的方案，该方法可以应用于发送端装置或者发送端装置侧的芯片或芯片系统。
根据本申请实施例的方案，发送端装置获取N个第一模型的信息与融合辅助信息，并根据第一模块对该N个第一模型的信息与融合辅助信息进行处理，得到N个第一信号，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端对于神经网络模型的存储空间。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
结合第一方面，在第一方面的某些实现方式中，该N个第一信号是通过发送端装置内的第一模块对N个第一模型的信息和该融合辅助信息处理后得到的。
结合第一方面，在第一方面的某些实现方式中，该第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第一方面，在第一方面的某些实现方式中，第一模块是根据发送N个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息(channel state information,CSI)、信噪比(signal-to-noise ratio，SNR)、链路质量。其中，信道状态信息可包括：信道质量指示(channel quality indicator，CQI)、秩指示(rank indicator，RI)、预编码矩阵(precoder matrix，PMI)。
结合第一方面，在第一方面的某些实现方式中，N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第一方面，在第一方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，该第三训练信号与中间模型对应；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该第一模块。
根据本申请实施例的方案，该第一模型可通过模型训练的方式得到，从而有利于实现N个第一信号在信道中完成叠加，进而得到与目标模型对应的第二信号。
结合第一方面，在第一方面的某些实现方式中，该模型传输的方法还包括：发送该融合辅助信息。
结合第一方面，在第一方面的某些实现方式中，该融合辅助信息来自该N个第一信号的接收端装置，或者该融合辅助信息来自为该N个第一信号的接收端装置提供服务的服务器。
可选地，该N个第一模型的信息承载于N个信号A中，该N个第一模型与N个信号 A一一对应。
具体地，在本申请实施例的模型传输的方法中，该N个第一模块处理N个第一模型对应的N个信号A的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个信号A可为模拟信号或者数字信号。
方式一，
在方式一中，该N个信号A为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个信号A经过所对应的N个处理处理后得到的N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第一信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式二，
在方式二中，该N个信号A为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发 送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式三，
在方式三中，该N个信号A为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式四，
在方式四中，该N个信号为A经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式四的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式五，
在方式五中，该N个信号A为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
第二方面，提供了一种模型传输的方法，应用于发送端装置或发送端装置中的芯片上，该方法包括：获取N个第一信号，该N个第一信号与该N个第一模型一一对应，每个第一信号用于承载所对应的该第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，N为大于或等于2的整数；获取N个第二模型，该N个第二模型与该N个第一模型一一对应，其中，该N个第二模型是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应 的业务，M为大于或等于2的整数，该M个第一训练模型对应的M个业务包括该N个第一模型对应的N个业务，且彼此对应的第一模型和第二模型对应同一业务；根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，其中，该N个第二信号中的第n个第二信号是根据第n个第二模型对第n个第一信号进行处理得到的信号，该第n个第一信号是该第n个第二模型对应的第一模型的信号，n∈[1，N]；通过同一信道发送该N个第二信号。
根据本申请实施例的方案，发送端装置获取N个第一模型对应的N个第一信号，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，随后通过第一信道发送该N个第二信号。进而接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端对于神经网络模型的存储空间。
其中，上述模型的外部特征值包括以下信息中的一个或多个：神经网络计算图信息、优化器信息、超参数信息。
结合第二方面，在第二方面的某些实现方式中，该获取N个第二模型包括：根据发送该N个第二信号的信道的参数，获取该N个第二模型。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第二方面，在第二方面的某些实现方式中，该获取N个第二模型还包括：根据该N个第一模型中除第i个第一模型以外的至少一个第一模型的信息，确定第i个第一模型对应的第二模型，i∈[1，N]。
结合第二方面，在第二方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号；对该第三训练信号进行解析以获得中间模型；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该N个第二模型。
根据本申请实施例的方案，可通过模型训练的方式获得满足任务需求的第二模型，从而实现N个第一信号在信道中的叠加，进而得到与目标模型对应的第二信号。
应理解，上述模型的参数可包括权重矩阵、权重向量、偏置矩阵、偏置向量。
具体地，在本申请实施例的模型传输的方法中，该N个第二模型处理N个第一模型对应的N个第一信号的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个第一信号可为模拟信号或者数字信号。
方式一，
在方式一中，该N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第二信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式二，
在方式二中，该N个第一信号为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式三，
在方式三中，该N个第一信号为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过 同一信道发送该处理后的N个第二信号。
在方式三的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式三的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式四，
在方式四中，该N个第一信号为经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式四的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式五，
在方式五中，该N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
结合第二方面，在第二方面的某些实现方式中，该模型传输的方法还包括：接收第一信息，该第一信息用于确定该N个第一模型。
其中，该第一信息来自该第二信号的接收端装置，或者该第一信息来自为该第二信号的接收端装置提供服务的服务器。该第一信息包括以下中的一项或者多项：神经网络计算图信息、优化器信息和超参数信息。
第三方面，提供了一种模型传输的方法，应用于至少两个发送端装置向一个接收端装置进行模型的传输的系统中，该方法包括：第一发送端装置获取N个第一模型中的S个第一模型的信息，该N个第一模型中的S个第一模型对应S个第一任务；获取第一融合辅助信息，该第一融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与该S个第一任务不同；第二发送端装置获取N个第一模型中的K个第一模型的信息，该N个第一模型中的K个第一模型对应K个第一任务；获取第二融合辅助信息，该第二融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与该K个第一任务不同，该S个第一模型与该K个第一模型的交集为空集；发送该S个第一信号与K个第一信号。可以理解，如果只有两个发送端装置，那么所述N个第一模型是所述S个第一模型与所述K个第一模型的并集。可选的，所述装置可以是装置或者装置中的芯片。
根据本申请实施例的方案，多个发送端装置可获取对N个第一模型的信息以及该融合辅助信息(例如，当N＝2时，第一发送端装置可获取N个第一模型中的S个第一模型和第一融合辅助信息，第二发送端装置可获取N个第一模型中的K个第一模型和第二融合辅助信息，该K个第一模型与该S个第一模型的交集为空集)，并根据N个第一模块对该N个第一模型的信息和该融合辅助信息进行处理，得到N个第一信号(例如，N＝2时，第一发送端装置对应的S个第一模型的信息和第一融合辅助信息经过所对应的第一模块处理后为S个第一信号，第二发送端装置对应的K个第一模型的信息和该第二融合辅助信息经过所对应的第一模块处理后为K个第一信号)，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
具体地，对于一个发送端装置向一个接收端装置进行模型传输的过程中，该N个第一信号是第一模块对该N个第一模型的信息和该融合辅助信息处理后得到的。对于多个发送端装置向一个接收端装置进行模型传输的过程中，例如两个发送端装置，上述S个第一信号是第一发送端装置通过对应的第一模块对该S个第一模型中的S个第一模型的信息和第一融合辅助信息处理后得到的，上述K个第一信号是第二发送端装置通过对应的第一模块对该K个第一模型中的K个第一模型的信息和第二融合辅助信息处理后得到的。
应理解，在本申请实施例中，第一融合辅助信息和第二融合辅助信息可以相同也可以不同。当第一融合辅助信息与第二融合辅助信息相同时，该第一融合辅助信息和第二融合辅助信息均可用来表示目标模型的外部特征值。当第一融合辅助信息与第二融合辅助信息不同时，该第一融合辅助信息表示目标模型的外部特征值的第一子集，该第二融合辅助信息表示目标模型的外部特征值的第二子集，该第一子集和该第二子集不同，且目标模型的外部特征值为该第一子集和第二子集的并集。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
结合第三方面，在第三方面的某些实现方式中，该第一发送端装置对应的第一模块或所述第二发送端装置对应的第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第三方面，在第三方面的某些实现方式中，第一发送端装置对应的第一模块或所述第二发送端装置对应的第一模块是根据发送对应第一信号或N个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第三方面，在第三方面的某些实现方式中，S个第一模型中的第i个模型对应的第一模块是根据S个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信 息确定的；所述K个第一模型中的第j个模型对应的第一模块是根据K个第一模型中除第一模型中的第j个模型以外的至少一个第一模型的信息确定的。
结合第三方面，在第三方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，该第三训练信号与中间模型对应；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该第一模块。
结合第三方面，在第三方面的某些实现方式中，该模型传输的方法还包括：发送该融合辅助信息。
结合第三方面，在第三方面的某些实现方式中，该融合辅助信息来自该N个第一信号的接收端装置，或者该融合辅助信息来自为该N个第一信号的接收端装置提供服务的服务器。
可选地，该N个第一模型的信息承载于N个信号A中，该N个第一模型与N个信号A一一对应。
具体地，在本申请实施例的模型传输的方法中，该N个第一模块处理N个第一模型对应的N个信号A的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个信号A可为模拟信号或者数字信号。
方式一，
在方式一中，该N个信号A为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个信号A经过所对应的N个处理处理后得到的N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第一信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得 到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式二，
在方式二中，该N个信号A为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式三，
在方式三中，该N个信号A为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式四，
在方式四中，该N个信号为A经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式四的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得 到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式五，
在方式五中，该N个信号A为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
第四方面，提供了一种模型传输的方法，应用于至少两个发送端装置向一个接收端装置进行模型的传输的系统中，该方法包括：第一发送端装置获取S个第一信号，该S个第一信号与该S个第一模型一一对应，每个第一信号用于承载所对应的该第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，S为大于或等于1整数；获取S个第二模型，该S个第二模型与该S个第一模型一一对应，其中，该S个第二模型是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，M为大于或等于2的整数，该M个第一训练模型对应的M个业务包括该S个第一模型对应的N个业务，且彼此对应的第一模型和第二模型对应同一业务；根据S个第二模型对该N个第一信号进行处理，得到S个第二信号，其中，该S个第二信号中的第s个第二信号是根据第s个第二模型对第s个第一信号进行处理得到的信号，该第s个第一信号是该第s个第二模型对应的第一模型的信号，s∈[1，S]；
第二发送端装置获取K个第一信号，该K个第一信号与该K个第一模型一一对应，每个第一信号用于承载所对应的该第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，K为大于或等于1整数；获取K个第二模型，该K个第二模型与该K个第一模型一一对应，其中，该K个第二模型是根据训练数据训练得到的，该训练数据包括该M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，M为大于或等于2的整数，该M个第一训练模型对应的M个业务包括该K个第一模型对应的N个业务，且彼此对应的第一模型和第二模型对应同一业务；根据K个第二模型对该N个第一信号进行处理，得到K个第二信号，其中，该K个第二信号中的第k个第二信号是根据第k个第二模型对第k个第一信号进行处理得到的信号，该第k个第一信号是该第k个第二模型对应的第一模型的信号，k∈[1，K]，该K个第一模型与S个第一模型的交集为空集；
发送该S个第二信号和K个第二信号。
根据本申请实施例的方案，多个发送端装置可获取对N个第一模型对应的N个第一信号(例如，当N＝2时，第一发送端装置可获取N个第一模型中的S个第一模型对应的S个第一信号，第二发送端装置可获取N个第一模型中的K个第一模型对应的K个第一信号，该K个第一模型与该S个第一模型的交集为空集)，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号(例如，N＝2时，第一发送端装置对应的S个 第一信号经过所对应的第二模型处理后为S个第二信号，第二发送端装置对应的K个第一信号经过所对应的第二模型处理后为K个第二信号)，随后通过第一信道发送该N个第二信号。进而接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
其中，上述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息。
结合第四方面，在第四方面的某些实现方式中，该获取N个第二模型包括：根据发送该N个第二信号的信道的参数，获取该N个第二模型。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第四方面，在第四方面的某些实现方式中，该获取N个第二模型还包括：根据该N个第一模型中除第i个第一模型以外的至少一个第一模型的信息，确定第i个第一模型对应的第二模型，i∈[1，N]。
结合第四方面，在第四方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号；对该第三训练信号进行解析以获得中间模型；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该N个第二模型。
根据本申请实施例的方案，通过模型训练的方式获得满足任务需求的第二模型，可实现N个第二信号在信道中进行叠加，并进一步的得到与目标模型对应的第三信号。
应理解，上述模型的参数可包括权重矩阵、权重向量、偏置矩阵、偏置向量。
具体地，在本申请实施例的模型传输的方法中，该N个第二模型处理N个第一模型对应的N个第一信号的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个第一信号可为模拟信号或者数字信号。
方式一，
在方式一中，该N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第二信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式二，
在方式二中，该N个第一信号为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式三，
在方式三中，该N个第一信号为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式三的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式三的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后 得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式四，
在方式四中，该N个第一信号为经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式四的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式五，
在方式五中，该N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
结合第四方面，在第四方面的某些实现方式中，该模型传输的方法还包括：接收第一信息，该第一信息用于确定该N个第一模型。
其中，该第一信息来自该第二信号的接收端装置，或者该第一信息来自为该第二信号的接收端装置提供服务的服务器。该第一信息包括以下中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息。
第五方面，提供了一种模型传输的方法，应用于发送端装置或发送端装置中的芯片上，该方法包括：获取N个第一模型中的S个第一模型的信息，该S个第一模型对应S个第一任务，其中，N为大于或等于1的整数，S为小于或等于N的整数；获取第一融合辅助信息，该第一融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与该S个第一任务不同；根据该S个第一模型的信息和该第一融合辅助信息确定S个第一信号；发送该S个第一信号。
根据本申请实施例的方案，发送端装置获取N个第一模型中的S个第一模型的信息与第一融合辅助信息，并根据第一模块对该S个第一模型的信息与第一融合辅助信息进行处理，得到S个第一信号，随后发送该S个第一信号。当S小于N时，其他一个或多个发送端也会发送N-S个第一信号中的第一信号，进而接收端装置可接收到目标模型对应的第二信号，可选的，该第二信号是由包括S个第一信号的N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，也适用于发送端装置分布式的场景，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、 偏置矩阵、偏置向量。
应理解，在本申请实施例中，该第一融合辅助信息可以为目标模型的外部特征值，也可以为该目标模型外部特征值的子集。
结合第五方面，在第五方面的某些实现方式中，该S个第一信号是通过发送端装置内的第一模块对S个第一模型的信息和第一该融合辅助信息处理后得到的。
结合第五方面，在第五方面的某些实现方式中，该第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第五方面，在第五方面的某些实现方式中，第一模块是根据发送S个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第五方面，在第五方面的某些实现方式中，该S个第一模型中的第i个模型对应的第一模块是根据S个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第五方面，在第五方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，该第三训练信号与中间模型对应；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该第一模块。
结合第五方面，在第五方面的某些实现方式中，该模型传输的方法还包括：发送该第一融合辅助信息。
结合第五方面，在第五方面的某些实现方式中，该第一融合辅助信息来自该S个第一信号的接收端装置，或者该第一融合辅助信息来自为该S个第一信号的接收端装置提供服务的服务器。
可选地，该S个第一模型的信息承载于S个信号A中，该S个第一模型与S个信号A一一对应。
具体地，在本申请实施例的模型传输的方法中，该S个第一模块处理S个第一模型对应的S个信号A的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该S个信号A可为模拟信号或者数字信号。
方式一，
在方式一中，该S个信号A为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该S个信号A经过所对应的S个处理处理后得到的S个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该S个第一信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式一的另一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码处理得到的信号。
可选地，该S个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该S个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式一的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该S个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式一的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该S个第一信号。
方式二，
在方式二中，该S个信号A为经过信源编码处理后的信号。
在方式二的一种实现方式中，该S个第一信号经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码处理得到的信号。
可选地，该S个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式二的另一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该S个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该S个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式二的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该S个第一信号。
方式三，
在方式三中，该S个信号A为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该S个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式三的另一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得 到的S个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该S个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式三的又一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该S个第一信号。
方式四，
在方式四中，该S个信号为A经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该S个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的S个第一信号。
在方式四的另一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该S个第一信号。
方式五，
在方式五中，该S个信号A为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该S个信号A经过所对应的S个第一模块处理后得到的S个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该S个第一信号。
第六方面，提供了一种模型传输的方法，应用于接收端装置或接收端装置中的芯片上，该方法包括：接收第二信号，该第二信号关联N个第一模型，该N个第一模型对应N个第一任务；根据该第二信号，获得目标模型，该目标模型对应第二任务，该第二任务与该第一任务不同。
根据本申请实施例的方案，接收端装置接收发送端装置发送的第二信号，以获得目标模型。其中，可选的，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
结合第六方面，在第六方面的某些实现方式中，该模型传输的方法还包括：发送融合辅助信息，该融合辅助信息包括目标模型的外部特征值。
具体地，模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息。
第七方面，提供了一种模型传输的方法，应用于发送端装置向接收端装置进行模型的传输的系统中，该方法包括：发送端装置获取N个第一模型的信息和融合辅助信息，根据该N个第一模型的信息和融合辅助信息确定N个第一信号，并发送该N个第一信号，其中，该N个第一模型对应N个第一任务，N为大于或等于2的整数，该N个第一模型的融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与第一 任务不同；接收端装置接收发送端装置发送的第二信号，并对该第二信号进行解析，以获取目标模型，其中，该第二信号是由该N个第一信号得到的。
根据本申请实施例的方案，发送端装置获取N个第一模型的信息与融合辅助信息，并根据第一模块对该N个第一模型的信息与融合辅助信息进行处理，得到N个第一信号，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
结合第七方面，在第七方面的某些实现方式中，该N个第一信号是通过发送端装置内的第一模块对N个第一模型的信息和该融合辅助信息处理后得到的。
结合第七方面，在第七方面的某些实现方式中，该第二信号是该N个第一信号在信道中叠加得到的。
结合第七方面，在第七方面的某些实现方式中，该第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第七方面，在第七方面的某些实现方式中，第一模块是根据发送N个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第七方面，在第七方面的某些实现方式中，N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第七方面，在第七方面的某些实现方式中，该模型传输的方法还包括：发送端装置获取M个原始模型，并将每个第一训练信号输入所对应的原始模型，以获得M个第二训练信号，该M个第二训练信号在同一信道中叠加以获得第三训练信号。其中，第三训练信号与中间模型对应，并根据中间模型与目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，从该模型库中获取该第一模块，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数，该模型库包括经过调节处理后的原始模型，并且经过调节处理后的中间模型与目标模型的偏差在预设范围内。
根据本申请实施例的方案，该第一模型可通过模型训练的方式得到，从而有利于实现N个第一信号在信道中完成叠加，进而得到与目标模型对应的第二信号。
结合第七方面，在第七方面的某些实现方式中，该模型传输的方法还包括：发送该融合辅助信息。
结合第七方面，在第七方面的某些实现方式中，该融合辅助信息来自该N个第一信号的接收端装置，或者该融合辅助信息来自为该N个第一信号的接收端装置提供服务的服务 器。
可选地，该N个第一模型的信息承载于N个信号A中，该N个第一模型与N个信号A一一对应。
具体地，在本申请实施例的模型传输的方法中，该N个第一模块处理N个第一模型对应的N个信号A的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个信号A可为模拟信号或者数字信号。
方式一，
在方式一中，该N个信号A为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个信号A经过所对应的N个处理处理后得到的N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第一信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式一的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式二，
在方式二中，该N个信号A为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码处理得到的信号。
可选地，该N个第一信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得 到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式二的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式三，
在方式三中，该N个信号A为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第一信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式三的又一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式四，
在方式四中，该N个信号为A经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第一信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第一信号。
在方式四的另一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
方式五，
在方式五中，该N个信号A为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个信号A经过所对应的N个第一模块处理后得到的N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第一信号。
第八方面，提供了一种模型传输的方法，应用于在至少两个发送端装置向一个接收端装置进行模型的传输的系统中，该方法包括：第一发送端装置获取S个第一信号，该S个第一信号与该S个第一模型一一对应，每个第一信号用于承载所对应的该第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，S为大于或等于1整数；获取S个第二模型，该S个第二模型与该S个第一模型一一对应，其中，该S个第二模型 是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，M为大于或等于2的整数，该M个第一训练模型对应的M个业务包括该S个第一模型对应的N个业务，且彼此对应的第一模型和第二模型对应同一业务；根据S个第二模型对该N个第一信号进行处理，得到S个第二信号，其中，该S个第二信号中的第s个第二信号是根据第s个第二模型对第s个第一信号进行处理得到的信号，该第s个第一信号是该第s个第二模型对应的第一模型的信号，s∈[1，S]；
第二发送端装置获取K个第一信号，该K个第一信号与该K个第一模型一一对应，每个第一信号用于承载所对应的该第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，K为大于或等于1整数；获取K个第二模型，该K个第二模型与该K个第一模型一一对应，其中，该K个第二模型是根据训练数据训练得到的，该训练数据包括该M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，M为大于或等于2的整数，该M个第一训练模型对应的M个业务包括该K个第一模型对应的N个业务，且彼此对应的第一模型和第二模型对应同一业务；根据K个第二模型对该N个第一信号进行处理，得到K个第二信号，其中，该K个第二信号中的第k个第二信号是根据第k个第二模型对第k个第一信号进行处理得到的信号，该第k个第一信号是该第k个第二模型对应的第一模型的信号，k∈[1，K]，该K个第一模型与S个第一模型的交集为空集；
接收端装置获取第三信号，该第三信号是由N个第二信号得到的。
根据本申请实施例的方案，多个发送端装置可获取对N个第一模型对应的N个第一信号(例如，当N＝2时，第一发送端装置可获取N个第一模型中的S个第一模型对应的S个第一信号，第二发送端装置可获取N个第一模型中的K个第一模型对应的K个第一信号，该K个第一模型与该S个第一模型的交集为空集)，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号(例如，N＝2时，第一发送端装置对应的S个第一信号经过所对应的第二模型处理后为S个第二信号，第二发送端装置对应的K个第一信号经过所对应的第二模型处理后为K个第二信号)，随后通过第一信道发送该N个第二信号。进而接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
其中，上述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息。
结合第八方面，在第八方面的某些实现方式中，该获取N个第二模型包括：根据发送该N个第二信号的信道的参数，获取该N个第二模型。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第八方面，在第八方面的某些实现方式中，该获取N个第二模型还包括：根据该 N个第一模型中除第i个第一模型以外的至少一个第一模型的信息，确定第i个第一模型对应的第二模型，i∈[1，N]。
结合第八方面，在第八方面的某些实现方式中，第三信号是由N个第二信号在信道中叠加得到的。
结合第八方面，在第八方面的某些实现方式中，该模型传输的方法还包括：发送端装置获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；该发送端装置将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号；对该第三训练信号进行解析以获得中间模型；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该N个第二模型。
根据本申请实施例的方案，通过模型训练的方式获得满足任务需求的第二模型，可实现N个第二信号在信道中进行叠加，并进一步的得到与目标模型对应的第三信号。
应理解，上述模型的参数可包括权重矩阵、权重向量、偏置矩阵、偏置向量。
具体地，在本申请实施例的模型传输的方法中，该N个第二模型处理N个第一模型对应的N个第一信号的过程可包括以下方式中的一种或者多种。
其中，在本申请实施例中，该N个第一信号可为模拟信号或者数字信号。
方式一，
在方式一中，该N个第一信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
在方式一的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为未经过信源编码、信道编码、调制以及上变频或者下变频处理的信号。
可选地，该N个第二信号可通过后续的信源编码、信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式一的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的 信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式二，
在方式二中，该N个第一信号为经过信源编码处理后的信号。
在方式二的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码处理得到的信号。
可选地，该N个第二信号可通过后续的信道编码、调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式二的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式三，
在方式三中，该N个第一信号为经过信源编码及信道编码处理后的信号。
在方式三的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码及信道编码处理得到的信号。
可选地，该N个第二信号可通过后续的调制以及上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式三的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式三的又一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式四，
在方式四中，该N个第一信号为经过信源编码、信道编码以及调制处理后的信号。
在方式四的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码以及调制处理得到的信号。
可选地，该N个第二信号可通过后续上变频或者下变频进行处理，并通过同一信道发送该处理后的N个第二信号。
在方式四的另一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的 信号。随后，发送端装置通过同一信道发送该N个第二信号。
方式五，
在方式五中，该N个第一信号为经过信源编码、信道编码、调制以及上变频或者下变频处理后的信号。
在方式五的一种实现方式中，该N个第一信号经过所对应的N个第二模型处理后得到的N个第二信号为经过信源编码、信道编码、调制以及上变频或者下变频处理得到的信号。随后，发送端装置通过同一信道发送该N个第二信号。
结合第八方面，在第八方面的某些实现方式中，该模型传输的方法还包括：接收第一信息，该第一信息用于确定该N个第一模型。
其中，该第一信息来自该第二信号的接收端装置，或者该第一信息来自为该第二信号的接收端装置提供服务的服务器。该第一信息包括以下中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息。
第九方面，提供了一种模型传输的系统，包括：发送端装置，用于获取N个第一模型的信息和融合辅助信息，根据该N个第一模型的信息和融合辅助信息确定N个第一信号，并发送该N个第一信号，其中，该N个第一模型对应N个第一任务，N为大于或等于2的整数，该N个第一模型的融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务，该第二任务与第一任务不同；
接收端装置，用于接收第二信号，并对接收到的第二信号进行解析，以获取目标模型，其中，该第二信号是由该N个第一信号得到的。
根据本申请实施例的方案，一个发送端装置获取N个第一模型与融合辅助信息，并根据第一模块对该N个第一模型的信息与融合辅助信息进行处理，得到N个第一信号，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
第十方面，提供了一种模型传输的系统，包括：第一发送端装置，用于获取N个第一模型中的S个第一模型的信息和第一融合辅助信息，根据N个第一模型中的S个第一模型的信息和该第一融合辅助信息确定S个第一信号，并发送该S个第一信号，其中，该N个第一模型中的S个模型对应S个第一任务，S为大于或等于1的整数，该第一融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同；
第二发送端装置，用于获取N个第一模型中的K个第一模型和第二合辅助信息，根据该N个第一模型中的K个第一模型的信息和该第二融合辅助信息确定K个第一信号，并发送该K个第一信号，其中，该N个第一模型中的K个模型对应K个第一任务，K为大于或等于1的整数，该第二融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同，该S个第一模型与该K个第一模型的交集为空集；
接收端装置，用于接收第二信号，并对接收到的该第二信号进行解析，以获取目标模型，其中，所述第二信号是由所述S个第一信号和所述K个第一信号得到的。
根据本申请实施例的方案，多个发送端装置可获取对N个第一模型的信息以及该融合辅助信息(例如，当N＝2时，第一发送端装置可获取N个第一模型中的S个第一模型和第一融合辅助信息，第二发送端装置可获取N个第一模型中的K个第一模型和该第二融合辅助信息，该K个第一模型与该S个第一模型的交集为空集)，并根据N个第一模块对该N个第一模型的信息和该融合辅助信息进行处理，得到N个第一信号(例如，N＝2时，第一发送端装置对应的S个第一模型的信息和该第一融合辅助信息经过所对应的第一模块处理后为S个第一信号，第二发送端装置对应的K个第一模型的信息和该第二融合辅助信息经过所对应的第一模块处理后为K个第一信号)，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，该目标模型对应第二任务，该第二任务与N个第一任务不同。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，该N个第一信号是在信道中叠加得到的。
应理解，在本申请实施例中，第一融合辅助信息和第二融合辅助信息可以相同也可以不同。当第一融合辅助信息与第二融合辅助信息相同时，该第一融合辅助信息和第二融合辅助信息均可用来表示目标模型的外部特征值。当第一融合辅助信息与第二融合辅助信息不同时，该第一融合辅助信息表示目标模型的外部特征值的第一子集，该第二融合辅助信息表示目标模型的外部特征值的第二子集，该第一子集和该第二子集不同，且目标模型的外部特征值为该第一子集和第二子集的并集。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
具体地，对于一个发送端装置向一个接收端装置进行模型传输的过程中，该N个第一信号是第一模块对该N个第一模型的信息和该融合辅助信息处理后得到的。对于多个发送端装置向一个接收端装置进行模型传输的过程中，例如两个发送端装置，上述S个第一信号是第一模块对该S个第一模型中的S个第一模型的信息和第一融合辅助信息处理后得到的，上述K个第一信号是第一模块对该K个第一模型中的K个第一模型的信息和第二融合辅助信息处理后得到的。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，该第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，第一模块是基于发送N个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、 链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
其中，上述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，该发送端装置还用于获取M个原始模型，并将每个第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将M个第二训练信号在同一信道中叠加以获得第三训练信号。其中，第三训练信号与中间模型对应，并根据中间模型与目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，从该模型库中获取该第一模块，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数，该模型库包括经过调节处理后的原始模型，并且经过调节处理后的中间模型与目标模型的偏差在预设范围内。
根据本申请实施例的方案，该第一模型可通过模型训练的方式得到，从而有利于实现N个第一信号在信道中完成叠加，进而得到与目标模型对应的第二信号。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，发送端装置还用于向接收端装置发送融合辅助信息。
结合第九方面和第十方面，在第九方面和第十方面的某些实现方式中，该融合辅助信息来自接收端装置，或者该融合辅助信息来自为该接收端装置提供服务的服务器。
第十一方面，提供了一种模型传输的系统，包括：发送端装置，用于获取N个第一信号，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，并通过第一信道发送该N个第二信号，其中，该N个第一信号与该N个第一模型一一对应，该N个第二模型与该N个第一模型一一对应，每个第一信号用于承载所对应的第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，该N个第二模型是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，该M个第一训练模型对应的M个业务包括该N个第一模型对应的N个业务，并且，该N个第二信号中的第n个第二信号是根据第n个第二模型对第n个第一信号进行处理得到的信号，该第n个第一信号是该第n个第二模型对应的第一模型的信号该N个第一信号与该N个第一模型一一对应，且彼此对应的第一模型和第二模型对应同一业务，n∈[1，N]，N为大于或等于2的整数，M为大于或等于2的整数；
接收端装置，用于通过该第一信道接收信号，并对所接收到的信号进行解析，以获取第三模型。
根据本申请实施例的方案，一个发送端装置获取N个第一模型对应的N个第一信号，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，随后通过第一信道发送该N个第二信号。进而一个发送端装置对应的一个接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在该第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在 传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
第十二方面，提供了一种模型传输的系统，包括：第一发送端装置，用于获取N个第一模型中的S个第一模型对应的S个第一信号，并根据N个第二模型中的S个第二模型对该S个第一信号进行处理，得到S个第二信号，并通过第一信道发送该S个第二信号，其中，该S个第一信号与该S个第一模型一一对应，该S个第二模型与该S个第一模型一一对应，每个第一信号用于承载所对应的第一模型的信息，其中，该第一模型的信息包括模型的外部特征值和/或模型的参数，且彼此对应的第一模型和第二模型对应同一业务，S为大于或等于1的整数；
第二发送端装置，用于获取该N个第一模型中的K个第一模型对应的K个第一信号，并根据该N个第二模型中的K个第二模型对该K个第一信号进行处理，得到K个第二信号，并通过第一信道发送该K个第二信号，其中，该K个第一信号与该K个第一模型一一对应，每个第一信号用于承载所对应的第一模型的信息，其中，所述第一模型的信息包括模型的外部特征值和/或模型的参数，该K个第二模型与该K个第一模型一一对应，且彼此对应的第一模型和第二模型对应同一业务，K为大于或等于1的整数，该K个第一模型与S个第一模型的交集为空集；
其中，该N个第二模型与该N个第一模型一一对应，该N个第二模型是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务，该M个第一训练模型对应的M个业务包括该N个第一模型对应的N个业务，并且，该N个第二信号中的第n个第二信号是根据第n个第二模型对第n个第一信号进行处理得到的信号，该第n个第一信号是该第n个第二模型对应的第一模型的信号，该N个第一信号与该N个第一模型一一对应，且彼此对应的第一模型和第二模型对应同一业务，n∈[1，N]，N为大于或等于2的整数，M为大于或等于2的整数；
接收端装置，用于通过该第一信道接收信号，并对所接收到的信号进行解析，以获取第三模型。
根据本申请实施例的方案，多个发送端装置可获取对N个第一模型对应的N个第一信号(例如，当N＝2时，第一发送端装置可获取N个第一模型中的S个第一模型对应的S个第一信号，第二发送端装置可获取N个第一模型中的K个第一模型对应的K个第一信号，该K个第一模型与该S个第一模型的交集为空集)，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号(例如，N＝2时，第一发送端装置对应的S个第一信号经过所对应的第二模型处理后为S个第二信号，第二发送端装置对应的K个第一信号经过所对应的第二模型处理后为K个第二信号)，随后通过第一信道发送该N个第二信号。进而接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端对于神经网络模型的存储空间。
结合第十一方面和第十二方面，在第十一方面和第十二方面的某些实现方式中，该获 取N个第二模型包括：根据发送该N个第二信号的信道的参数，获取该N个第二模型。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
其中，上述模型的外部特征值包括以下信息中的一个或者多个信息：神经网络计算图信息、优化器信息、超参数信息。
结合第十一方面和第十二方面，在第十一方面和第十二方面的某些实现方式中，该获取N个第二模型还包括：根据该N个第一模型中除第i个第一模型以外的至少一个第一模型的参数，确定第i个第一模型对应的第二模型，i∈[1，N]。
结合第十一方面和第十二方面，在第十一方面和第十二方面的某些实现方式中，该模型传输的方法还包括：获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号；对该第三训练信号进行解析以获得中间模型；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该N个第二模型。
应理解，上述模型的参数可包括权重矩阵、权重向量、偏置矩阵、偏置向量。
结合第十一方面和第十二方面，在第十一方面和第十二方面的某些实现方式中，该模型传输的方法还包括：接收第一信息，该第一信息用于确定该N个第一模型。
其中，该第一信息来自该第二信号的接收端装置，或者该第一信息来自为该第二信号的接收端装置提供服务的服务器。该第一信息包括一下中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息。
第十三方面，提供了一种模型传输的装置，包括：
获取单元，用于获取N个第一模型的信息和融合辅助信息，该融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务；
处理单元，用于根据该N个第一模型的信息和该融合辅助信息确定N个第一信号，并发送该N个第一信号，其中，该N个第一模型对应N个第一任务，N为大于或等于2的整数，该第二任务与该第一任务不同；
发送单元，用于向接收端装置发送该N个第一信号。
根据本申请实施例的方案，发送端装置获取N个第一模型的信息与融合辅助信息，并根据第一模块对该N个第一模型的信息与N融合辅助信息进行处理，得到N个第一信号，随后发送该N个第一信号。进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
结合第十三方面，在第十三方面的某些实现方式中，该处理单元通过第一模块确定该N个第一信号。
结合第十三方面，在第十三方面的某些实现方式中，该第一模块是基于发送N个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第十三方面，在第十三方面的某些实现方式中，N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第十三方面，在第十三方面的某些实现方式中，该处理单元还用于获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号；对该第三训练信号进行解析以获得中间模型；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该N个第一模块。
结合第十三方面，在第十三方面的某些实现方式中，发送单元还用于向接收端装置发送融合辅助信息。
结合第十三方面，在第十三方面的某些实现方式中，该融合辅助信息来自接收端装置，或者该融合辅助信息来自为该接收端装置提供服务的服务器。
应理解，如果N个第一模型的信息和融合辅助信息是从其他装置接收来的，那么获取单元是接收单元或收发单元；如果N个第一模型的信息和融合辅助信息是该模型传输的装置本地获取的(例如，确定出来的)，那么获取单元可以就是处理单元；如果N个第一模型的信息和融合辅助信息中的一类信息是从其他装置接收来的，另一类信息是本地获取的，那么获取单元可以是接收单元和处理单元的功能集成单元。
第十四方面，提供了一种模型传输的装置，包括：
处理单元，用于获取N个第一信号，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，并通过第一信道发送该N个第二信号，其中，该N个第一信号与N个第一模型一一对应，该N个第二模型与该N个第一模型一一对应，每个第一信号用于承载所对应的第一模型的信息，其中，所述第一模型的信息包括模型的外部特征值和/或模型的参数，该N个第二模型是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足M个训练模型对应的业务，该M个第一训练模型对应的M个业务包括N个第一模型对应的N个业务，并且，该N个第二信号中的第n个第二信号是根据第n个第二模型对第n个第一信号进行处理得到的信号，该第n个第一信号是该第n个第二模型对应的第一模型的信号，N个第一信号与N个第一模型一一对应，且彼此对应的第一模型和第二模型对应同一业务，n∈[1，N]，N为大于或等于2的整数，M为大于或等于2的整数；
发送单元，用于通过同一信道向接收端装置发送该N个第二信号。
根据本申请实施例的方案，发送端装置中第一模块可获取N个第一模型对应的N个第一信号，并根据N个第二模型对该N个第一信号进行处理，得到N个第二信号，随后通过发送模块经过第一信道发送该N个第二信号。进而接收端装置可接收到目标模型对应的第三信号，该第三信号是由N个第二信号在第一信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端对于神经网络模型的存储空间。
其中，上述模型的外部特征值包括以下信息中的一个或者多个信息：神经网络计算图信息、优化器信息、超参数信息。
结合第十四方面，在第十四方面的某些实现方式中，该N个第二模型是基于该第一信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第十四方面，在第十四方面的某些实现方式中，该N个第一模型中的第i个第一模型对应的第二模型是根据该N个第一模型中除第i个第一模型以外的至少一个第一模型的信息确定的，i∈[1，N]。
结合第十四方面，在第十四方面的某些实现方式中，该处理单元还用于获取M个原始模型，并将每个第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将该M个第二训练信号在同一信道中叠加以获得第三训练信号，并对该第三训练信号进行解析以获得中间模型，并根据该中间模型与目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，从该模型库中获取N个第二模型，该M个原始模型与M个第一训练信号一一对应，M为大于或等于2的整数，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与目标模型的偏差在预设范围内。
结合第十四方面，在第十四方面的某些实现方式中，该获取单元还用于：接收第一信息，该第一信息用于确定该N个第一模型，其中，该第一信息来自接收端装置，或者该第一信息来自为该接收端装置提供服务的服务器。
第十五方面，提供了一种模型传输的装置，该装置包括：
获取单元，用于获取N个第一模型中的S个第一模型的信息和第一融合辅助信息，该第一融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务；
处理单元，用于根据该S个第一模型的信息和该第一融合辅助信息确定S个第一信号，其中，S个第一模型对应S个第一任务，该第二任务与该S个第一任务不同；
发送单元，用于向接收端装置发送该S个第一信号。
根据本申请实施例的方案，发送端装置获取N个第一模型中的S个第一模型的信息与第一融合辅助信息，并根据第一模块对该S个第一模型的信息与第一融合辅助信息进行处理，得到S个第一信号，随后发送该S个第一信号。当S小于N时，其他一个或多个发送端也会发送N-S个第一信号中的第一信号，进而接收端装置可接收到目标模型对应的第二信号，该第二信号是由包括S个第一信号的N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，也适用于发送端装置分布式的场景，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
具体地，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，该模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
应理解，在本申请实施例中，该第一融合辅助信息可以为目标模型的外部特征值，也可以为该目标模型外部特征值的子集。
结合第十五方面，在第第十五方面的某些实现方式中，该处理单元通过第一模块确定该S个第一信号。
结合第十五方面，在第十五方面的某些实现方式中，该第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，该目标模型能够满足M个训练模型对应的任务，M为大于或等于2的整数。
结合第十五方面，在第十五方面的某些实现方式中，第一模块是根据发送该S个第一信号的信道的参数确定的。
应理解，上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
结合第十五方面，在第十五方面的某些实现方式中，该S个第一模型中的第i个模型对应的第一模块是根据S个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
结合第十五方面，在第十五方面的某些实现方式中，该处理单元还用于获取M个原始模型，该M个原始模型与该M个第一训练信号一一对应，M为大于或等于2的整数；将每个该第一训练信号输入所对应的原始模型，以获得M个第二训练信号；将该M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，该第三训练信号与中间模型对应；根据该中间模型与该目标模型的偏差，对该M个原始模型的参数进行调节处理，以获得模型库，该模型库包括经过调节处理后的原始模型，并且经过该调节处理后的中间模型与该目标模型的偏差在预设范围内；从该模型库中获取该第一模块。
结合第十五方面，在第十五方面的某些实现方式中，发送单元还用于向接收端装置发送第一融合辅助信息。
结合第十五方面，在第十五方面的某些实现方式中，该第一融合辅助信息来自该S个第一信号的接收端装置，或者该第一融合辅助信息来自为该S个第一信号的接收端装置提供服务的服务器。
第十六方面，提供了一种模型接收的装置，该装置包括：
接收单元，用于接收第二信号，该第二信号关联第一模型，该N个第一模型对应N个第一任务；
处理单元，根据该第二信号，获得目标模型，该目标模型对应第二任务，该第二任务与该第一任务不同。
根据本申请实施例的方案，接收端装置接收发送端装置发送的第二信号，以获得目标模型。其中，该第二信号是由N个第一信号在信道中叠加得到的。通过上述模型传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程 中的资源开销和时延，并在一定程度上降低接收端装置对于神经网络模型的存储空间。
结合第十六方面，在第十六方面的某些实现方式中，该模型的装置还包括发送单元，该发送单元用于向接收端装置发送融合辅助信息，该融合辅助信息包括目标模型的外部特征值。
具体地，模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息。
第十七方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第一方面以及第一方面中的任意一种实现方式中的方法。
第十八方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第二方面以及第二方面中的任意一种实现方式中的方法。
第十九方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第三方面以及第三方面中的任意一种实现方式中的方法。
第二十方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第四方面以及第四方面中的任意一种实现方式中的方法。
第二十一方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第五方面以及第五方面中的任意一种实现方式中的方法。
第二十二方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第六方面以及第六方面中的任意一种实现方式中的方法。
第二十三方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第七方面以及第七方面中的任意一种实现方式中的方法。
第二十四方面，提供了一种模型传输的装置，该装置包括：存储器，用于存储程序；处理器，用于执行所述存储器存储的程序，当所述存储器存储的程序被执行时，所述处理器用于执行第八方面以及第八方面中的任意一种实现方式中的方法。
上述第十七方面至第二十四方面中的处理器既可以是中央处理器(central processing unit，CPU)，也可以是CPU与神经网络运算处理器的组合，这里的神经网络运算处理器可以包括图形处理器(graphics processing unit，GPU)、神经网络处理器(neural-network processing unit，NPU)和张量处理器(tensor processing unit，TPU)等等。其中，TPU是谷歌(google)为机器学习全定制的人工智能加速器专用集成电路。
第二十五方面，提供一种计算机可读介质，该计算机可读介质存储用于设备执行的程序代码，该程序代码包括用于执行第一方面以及第八方面中的任意一种实现方式中的方法。
第二十六方面，提供一种包含指令的计算机程序产品，当该计算机程序产品在计算机上运行时，使得计算机执行上述第一方面及第八方面中的任意一种实现方式中的方法。
第二十七方面，提供一种芯片，所述芯片包括处理器与数据接口，所述处理器通过所述数据接口读取存储器上存储的指令，执行上述第一方面及第八方面中的任意一种实现方式中的方法。
可选地，作为一种实现方式，所述芯片还可以包括存储器，所述存储器中存储有指令，所述处理器用于执行所述存储器上存储的指令，当所述指令被执行时，所述处理器用于执行第一方面及第八方面中的任意一种实现方式中的方法。
上述芯片具体可以是现场可编程门阵列(field-programmable gate array，FPGA)或者专用集成电路(application-specific integrated circuit，ASIC)。
附图说明
图1是基于神经网络实现的端到端通信系统示意图。
图2是本申请实施例提供的适用的通信系统架构的示意图。
图3是本申请实施例提供的另一例通信系统架构的示意图。
图4是本申请实施例提供的又一例通信系统架构的示意图。
图5示出了本申请实施例的一例人工智能主体框架的示意图。
图6示出了本申请实施例提供的一例系统架构600的示意图。
图7为本申请实施例提供的一种卷积神经网络模型的示意性结构图。
图8为本申请实施例提供的另一种卷积神经网络模型的示意性结构图。
图9为本申请实施例提供的一种芯片的硬件结构的示意图。
图10为本申请实施例提供的一种系统架构1000的示意图。
图11为本申请实施例提供的一种训练模型的装置部署示意图。
图12为本申请实施例提供的一种模型传输的方法的示意图流程图。
图13为本申请实施例提供的一种模型传输的方法的示意图流程框图。
图14为本申请实施例提供的另一种模型传输的方法的示意图流程图。
图15为本申请实施例提供的一种神经网络模型的有向图结构示意图。
图16为本申请实施例提供的一种神经网络模型的计算图的结构示意图。
图17为本申请实施例提供的计算图层面下的神经网络模型传输的示意图。
图18为本申请实施例提供的又一种模型传输的方法的示意图流程图。
图19为本申请实施例提供的一例模型传输的装置的示意性框图。
图20为本申请实施例提供的另一例模型传输的装置的示意性框图。
具体实施方式
本申请实施例的技术方案可以应用于各种通信系统，例如：窄带物联网系统(narrow band-internet of things，NB-IoT)、全球移动通信系统(global system for mobile communications，GSM)、增强型数据速率GSM演进系统(enhanced data rate for GSM evolution，EDGE)、宽带码分多址系统(wideband code division multiple access，WCDMA)、码分多址2000系统(code division multiple access，CDMA2000)、时分同步码分多址系统(time division-synchronization code division multiple access，TD-SCDMA)、长期演进(long term evolution，LTE)系统、LTE频分双工(frequency division duplex，FDD)系统、LTE 时分双工(time division duplex，TDD)、第五代(5th generation，5G)通信系统或新无线(new radio，NR)、无线保真(wireless fidelity，WiFi)系统、以及5G移动通信系统的三大应用场景增强移动宽带(enhanced mobile broadband，eMBB)、超高可靠性超低时延通信(ultra-reliable low-latency communication，uRLLC)、增强机器类通信(LTE enhanced MTO，eMTC)、支持多种无线技术融合的通信系统以及5G之后演进的第六代(6th generation，6G)通信系统等。
本申请实施例的技术方案还可以应用于无线蜂窝通信系统、无线网格(Mesh)网络和卫星通信系统等通信系统。无线蜂窝通信系统包括一个蜂窝基站和多个终端设备时，多个终端设备能够辅助蜂窝基站进行人工智能(artificial intelligence，AI)模型计算。无线蜂窝通信系统包括多个蜂窝基站和一个终端设备时，多个蜂窝基站能够辅助终端设备进行AI模型计算。无线Mesh网络包括蜂窝宏基站、微基站以及终端设备。蜂窝宏基站通过多个微基站中继向终端设备传输下行数据。卫星通信系统包括卫星基站和终端设备。
本申请实施例中的终端可以是一种具有无线收发功能的设备，具体可以指用户设备(user equipment，UE)、接入终端、用户单元(subscriber unit)、用户站、移动台(mobile station)、远方站、远程终端、移动设备、用户终端、无线通信设备、用户代理或用户装置。终端设备还可以是卫星电话、蜂窝电话、智能手机、无线数据卡、无线调制解调器、机器类型通信设备、可以是无绳电话、会话启动协议(session initiation protocol，SIP)电话、无线本地环路(wireless local loop，WLL)站、个人数字处理(personal digital assistant，PDA)、具有无线通信功能的手持设备、计算设备或连接到无线调制解调器的其它处理设备、车载设备、高空飞机上搭载的通信设备、可穿戴设备、无人机、机器人、设备到设备通信(device-to-device，D2D)中的终端、车辆外联(vehicle to everything，V2X)中的终端、虚拟现实(virtual reality，VR)终端设备、增强现实(augmented reality，AR)终端设备、工业控制(industrial control)中的无线终端、无人驾驶(selfdriving)中的无线终端、远程医疗(remote medical)中的无线终端、智能电网(smart grid)中的无线终端、运输安全(transportation safety)中的无线终端、智慧城市(smart city)中的无线终端、智慧家庭(smart home)中的无线终端或者5G之后演进的通信网络(如，6G)中的终端设备等，本申请不作限制。
在本申请实施例中，用于实现终端设备的功能的装置可以是终端设备；也可以是能够支持终端设备实现该功能的装置，例如芯片系统。该装置可以被安装在终端设备中或者和终端设备匹配使用。本申请实施例中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。
本申请实施例中的网络设备具有无线收发功能的设备，用于与终端设备进行通信。接入网设备可以为无线接入网(radio access network，RAN)中的节点，又可以称为基站，还可以称为RAN节点。可以是LTE中的演进型基站(evolved Node B，eNB或eNodeB)；或者gNodeB(gNB)等5G网络中的基站或者5G之后演进的公共陆地移动网络(public land mobile network，PLMN)中的基站，宽带网络业务网关(broadband network gateway，BNG)，汇聚交换机或者非第三代合作伙伴项目(3rd generation partnership project，3GPP)接入设备等。可选的，本申请实施例中的网络设备可以包括各种形式的基站，例如：宏基站、微基站(也称为小站)、中继站、WiFi系统中的接入节点，传输接收点(transmitting and receiving  point，TRP)、发射点(transmitting point，TP)、移动交换中心以及设备到设备(device-to-device，D2D)、车辆外联(vehicle-to-everything，V2X)、机器到机器(machine-to-machine，M2M)通信中承担基站功能的设备等，还可以包括云接入网(cloud radio access network，C-RAN)系统中的集中式单元(centralized unit，CU)和分布式单元(distributed unit，DU)、NTN通信系统中的网络设备，本申请实施例对此不作具体限定。
本申请实施例中，用于实现网络设备的功能的装置可以是网络设备；也可以是能够支持网络设备实现该功能的装置，例如芯片系统。该装置可以被安装在网络设备中或者和网络设备匹配使用。本申请实施例中，芯片系统可以由芯片构成，也可以包括芯片和其他分立器件。
应理解，本申请实施例所涉及的发送端设备及接收端设备可以是上述终端设备中的一种或多种，也可以是上述网络设备中的一种或多种，本申请对此不做限定。
本申请实施例中的云端设备可以是服务于接收端设备或者发送端设备的服务器或者服务器集群，该云端设备也可以称为计算节点或者云端计算集群。
图2是本申请实施例提供的一例通信系统架构的示意图。如图2所示，该通信系统架构为卫星通信系统。
如图2中的(a)所示的卫星通信系统包括卫星基站、终端设备#1和终端设备#2等设备，其中，卫星基站也可以称为卫星。该卫星基站可为终端设备#1与终端设备#2提供通信服务。具体地，卫星基站可以向终端设备#1和终端设备#2传输下行数据，其中，下行数据可以经过信道编码，调制映射后传输给终端设备#1和终端设备#2；终端设备#1和终端设备#2也可以向卫星基站传输上行数据，其中，上行数据也可以经过信道编码、调制映射后传输给卫星基站。
与上述图2中的(a)示出通信系统相比，图2中的(b)所示的卫星通信系统还可以包括地面基站，卫星基站可以与地面上的基站进行通信。
其中，卫星基站可以是无人机、热气球、低轨卫星、中轨卫星或高轨卫星等。或者，卫星基站也可以是指非地面的基站或非地面的设备等。卫星基站既可以作为网络设备，也可以作为终端设备。卫星基站可以不具备基站的功能，也可以具备部分或者全部基站的功能，本申请对此不做限定。
应理解，图2中的卫星基站、地面基站和终端设备的数量为示意性的，根据实际场景的需求，可以具有任意数量的卫星基站和终端设备。
图3是本申请实施例提供的另一例通信系统架构的示意图。该通信系统为卫星星间链路通信系统，也可以称为星间通信系统。该通信系统中的卫星可包括通信系统与空间光束捕获跟踪对准技术(acquisition,pointing and tracking,APT)系统。
通信系统可包括通信模块和收发天线，用于负责星间信息的传输，是星间通信系统的主体。APT系统用于负责卫星之间的捕获、对准和跟踪。其中，捕获是指在不确定区域内对目标进行判断和识别，捕获过程包括前期的数据导引和视轴初始指向，在不确定区域内的扫描，直至在捕获视场内判断信标光斑的存在，为后续的跟踪奠定基础；对准是指调整发射波瞄准接收方向，使通信发射视轴与跟踪视轴在装校中需要保持非常精密的同轴性或者需要通过提前量伺服单元修正发射视轴与接收视轴间的偏差；跟踪是指在整个通信过程中不断调整对准和捕获。空间激光通信系统中的跟踪的实质是对激光光束实现主动光电跟 踪，进而克服相对运动和平台振动对视轴晃动，跟踪的目的是将跟踪视轴稳定在允许的误差范围之内。
在卫星星间链路通信系统中，为了尽量减少信道中的衰减和干扰影响，同时要求具有较高的保密性和传输率，须实时的调整APT来不断适应变化。现有的APT系统均为光学系统。现有的通信子系统，多数为光通信系统，也有部分微波波段的系统，多采用单个高增益天线。现有的APT系统和通信子系统为独立的系统。
图4是本申请实施例提供的又一例通信系统架构的示意图。所示的通信系统为无线蜂窝通信系统，该无线蜂窝通信系统通常由小区组成，每个小区包含一个基站(base station，BS)，该基站可向多个终端设备提供通信服务。示例性地，图4所示的蜂窝通信系统包括基站、终端设备#1和终端设备#2等设备。具体地，基站可以向终端设备发送下行数据，其中下行数据可以采用信道编码进行编码；终端设备也可以向基站发送上行数据，该上行数据也可以采用信道编码进行编码。
其中，基站包含基带单元(baseband unit，BBU)和远端射频单元(remote radio unit，RRU)。BBU和RRU可以放置在不同的地方，例如：RRU可放置于高话务量的区域，BBU放置于中心机房，或者BBU与RRU放置在同一机房，或者BBU和RRU可以为一个机架下的不同部件。应理解，图4所示的无线蜂窝通信系统可包括多个基站并且每个基站的覆盖范围内可以包括其他数量的终端设备，本申请实施例对此不做限定。
图5示出了本申请实施例的一例人工智能主体框架的示意图，该主体框架描述了人工智能系统总体工作流程，适用于通用的人工智能领域需求。
接下来从“智能信息链”(水平轴)和“信息技术(information technology，IT)价值链”(垂直轴)两个维度对上述人工智能的主题框架结构进行详细的阐述。
“智能信息链”反映从数据的获取到处理的一系列过程。示例性地，可以是智能信息感知、智能信息表示与形成、智能推理、智能决策、智能执行与输出的一般过程。在这个过程中，数据经历了“数据—信息—知识—智慧”的凝练过程。
“IT价值链”从人智能的底层基础设施、信息(提供和处理技术实现)到系统的产业生态过程，反映人工智能为信息技术产业带来的价值。
下面进一步对该人工智能的框架中的子模块进行详细介绍：
(1)基础设施：
基础设施为人工智能系统提供计算能力支持，实现与外部现实世界的沟通，并通过基础平台实现支撑。
基础设施可以通过传感器与外部进行沟通，基础设施的计算能力可以由智能芯片提供。
其中，这里的智能芯片可以是中央处理器(central processing unit，CPU)、神经网络模型处理器(neural-network processing unit，NPU)、图形处理器(graphics processing unit，GPU)、专门应用的集成电路(application specific integrated circuit，ASIC)以及现场可编程门阵列(field programmable gate array，FPGA)等硬件加速芯片。
基础设施的基础平台可以包括分布式计算框架及网络等相关的平台保障和支持，可以包括云存储和计算、互联互通网络等。
例如，对于基础设施而言，可以通过传感器和外部之间的沟通获取数据，随后将这些数据提供给基础平台提供的分布式计算系统中的智能芯片进行计算。
(2)数据：
基础设施的上一层的数据用于表示人工智能领域的数据来源。该数据涉及到图形、图像、语音、文本，还涉及到设备的物联网数据，包括已有系统的业务数据以及力、位移、液位、温度、湿度等感知数据。
(3)数据处理：
上述数据处理通常包括数据训练，机器学习，深度学习，搜索，推理，决策等处理方式。
其中，机器学习和深度学习可以对数据进行符号化和形式化的智能信息建模、抽取、预处理、训练等。
训练(training)是计算密集型操作，模型一般都需要使用大量的数据来进行训练，通过反向传播来不断的优化模型的参数，以使得模型获取某种能力。在训练的过程中，我们常常是将模型在数据集上面的拟合情况放在首要位置的。
推理(inference)是指在计算机或智能系统中，模拟人类的智能推理方式，依据推理控制策略，利用形式化的信息进行机器思维和求解问题的过程，典型的功能是搜索与匹配。
决策是指智能信息经过推理后进行决策的过程，通常可提供分类、排序、预测等功能。
(4)通用能力：
对数据经过上述的数据处理后，进一步基于数据处理的结果可以形成一些通用的能力，比如可以是算法或者一个通用系统，例如，翻译，文本的分析，计算机视觉的处理，语音识别，图像的识别等等。
(5)智能产品及行业应用：
智能产品及行业应用指人工智能系统在各领域的产品和应用，是对人工智能整体解决方案的封装，将智能信息决策产品化、实现落地应用，其应用领域主要包括：智能制造、智能交通、智能家居、智能医疗、智能安防、自动驾驶，平安城市，智能终端等。
本申请实施例可以应用在人工智能中的很多领域，例如，智能制造、智能交通、智能家居、智能医疗、智能安防、自动驾驶，平安城市等领域。
具体地，本申请实施例中的获取神经网络模型的方法可以具体应用在自动驾驶、图像分类、图像检索、图像语义分割、图像质量增强、图像的超分辨率和自然语言的处理等需要使用(深度)神经网络模型的领域。
由于本申请实施例涉及神经网络领域，为了便于理解，下面先对本申请实施例涉及的神经网络领域的相关术语和概念进行介绍。
(1)人工智能
人工智能是利用数字计算机或者数字计算机控制的机器模拟、延伸和扩展人的智能，感知环境、获取知识并使用知识获得最佳结果的理论、方法、技术及应用系统。换句话说，人工智能是计算机科学的一个分支，它企图了解智能的实质，并生产出一种新的能以人类智能相似的方式作出反应的智能机器。人工智能也就是研究各种智能机器的设计原理与实现方法，使机器具有感知、推理与决策的功能。人工智能领域的研究包括机器人，自然语言处理，计算机视觉，决策与推理，人机交互，推荐与搜索，AI基础理论等。
(2)边缘人工智能
边缘人工智能(edge AI)是人工智能领域的新领域之一，其目的是让用户进行人工智 能流程而不必担心隐私泄露或数据传输速率较慢带来的影响。边缘人工智能将人工智能技术和边缘计算相结合，使人工智能算法运行在能够进行边缘计算的设备上，可以实现对数据的实时处理。边缘人工智能可以使人工智能技术得到更加广泛的应用，使得智能设备无需接入云平台的情况下对输入做出快速反应。
例如，自动驾驶汽车在行驶过程中需要不断地扫描周围的环境并评估当前的行驶状况，并需要根据突发事件对行进轨迹进行不断地校正。在此情况下，对于汽车的实时数据的处理至关重要，其车载的人工智能系统将负责数据的存储、分析和处理。通过引入边缘人工智能技术，可在不依赖云计算连接的情况下，实现对数据的实时分析和处理。
(3)神经网络
神经网络(neural network,NN)也可以称为人工神经网络。神经网络可以是由神经单元组成的，神经单元可以是指以xs和截距1为输入的运算单元，该运算单元的输出可以为：
其中，s＝1、2、……n，n为大于1的自然数，Ws为xs的权重，b为神经单元的偏置。f为神经单元的激活函数(activation functions)，用于将非线性特性引入神经网络中，来将神经单元中的输入信号转换为输出信号。该激活函数的输出信号可以作为下一层卷积层的输入，激活函数可以是sigmoid函数。神经网络是将多个上述单一的神经单元联结在一起形成的网络，即一个神经单元的输出可以是另一个神经单元的输入。每个神经单元的输入可以与前一层的局部接受域相连，来提取局部接受域的特征，局部接受域可以是由若干个神经单元组成的区域。
(4)深度神经网络
深度神经网络(deep neural network，DNN)，也称多层神经网络，可以理解为具有多层隐藏层的神经网络。按照不同层的位置对DNN进行划分，DNN内部的神经网络可以分为三类：输入层，隐藏层和输出层。一般来说，第一层是输入层，最后一层是输出层，中间的层都是隐藏层。层与层之间可以是全连接的，也即是说，第i层的任意一个神经元与第i+1层的任意一个神经元相连。简单来说，可用如下线性关系表示：
其中，是输入向量，是输出向量，是偏移向量，W为权重矩阵(也称系数)，α是激活函数。隐藏层的每一层输入向量经过该操作得到输出向量由于DNN层数多，系数W和偏移向量b的数量也比较多。这些参数在DNN中的定义如下：以系数W为例，假设在一个三层的DNN中，第二层的第4个神经元到第三层的第2个神经元的线性系数定义为上标3代表系数W所在的层数，下标对应的是输出的第三层索引2和输入的第二层索引4。
综上，第L-1层的第k个神经元到第L层的第j个神经元的系数定义为
需要注意的是，输入层是没有参数的。在深度神经网络中，由万能近似定理可知，只要给定足够多的隐藏层，DNN可以以任意精度拟合任意函数。也就是说，更多的隐藏层可以让网络更能够刻画现实世界中的复杂情形。理论上而言，参数越多，模型复杂度越高，“容量”也就越大，也就意味着它能完成更复杂的学习任务。训练深度神经网络也就是学习权重矩阵的过程，其最终目的是训练得到深度神经网络的所有层的权重矩阵(由很多层的向量形成的权重矩阵)。
(5)卷积神经网络
卷积神经网络(convolutional neuron network，CNN)是一种带有卷积结构的神经网络。卷积神经网络包含了一个由卷积层和子采样层构成的特征抽取器，该特征抽取器可以看作是滤波器。卷积层是指卷积神经网络中对输入信号进行卷积处理的神经元层。在卷积神经网络的卷积层中，一个神经元可以只与部分邻层神经元连接。一个卷积层中，通常包含若干个特征平面，每个特征平面可以由一些矩形排列的神经单元组成。同一特征平面的神经单元共享权重，这里共享的权重就是卷积核。共享权重可以理解为提取图像信息的方式与位置无关。在卷积神经网络的训练过程中，卷积核可以通过学习得到合理的权重。另外，共享权重带来的直接好处是减少卷积神经网络各层之间的连接，同时又降低了过拟合的风险。
(7)反向传播算法
神经网络可以采用梯度反向传播(back propagation，BP)算法在训练过程中修正的神经网络模型中参数的大小，使得神经网络模型的重建误差损失越来越小。具体地，前向传递输入信号直至输出会产生误差损失，通过反向传播误差损失信息来更新的神经网络模型中参数，从而使误差损失收敛。反向传播算法是以误差损失为主导的反向传播运动，旨在得到最优的神经网络模型的参数，例如权重矩阵。
例如，神经网络模型每次训练产生的loss值在神经网络模型中从后向前逐层传递。传递到每一层时，同时计算出该层参数的更新量(偏导运算)，这个更新量与梯度(gradient)相关。
(8)深度学习
深度学习(deep learning,DL)是机器学习(machine learning,ML)领域中的一个新的研究方向。深度学习是学习样本数据的内在规律和表示层次，这些学习过程中获得的信息对诸如文字，图像和声音等数据的解释有较大的帮助。它的目标是让机器能够像人一样具有分析学习的能力，能够识别文字、图像和声音等数据。深度学习是一个复杂的机器学习算法，在语音和图像识别方面取得的效果，远超过先前相关技术。
深度学习中的深度是指一系列连续的表示层，数据模型中包含的层数被称为模型的深度。深层学习的模型特征是一种含多隐藏层的神经网络，通过组合底层特征形成更加抽象的高层特征。
深度学习在搜索技术，数据挖掘，机器学习，机器翻译，自然语言处理，多媒体学习，语音，推荐和个性化技术，以及其他相关领域都取得了很多成果。深度学习使机器模仿视听和思考等人类的活动，解决了很多复杂的模式识别难题，使得人工智能相关技术取得了很大进步。
(9)联邦学习
联邦学习(federated learning，FL)在充分保障用户数据隐私和安全的前提下，促使各个分布式节点和中心节点协同合作以高效地完成模型的学习任务。在FL框架中，数据集存在于分布式节点处，即分布式节点收集本地的数据集，并进行本地训练，将训练得到的本地结果(模型或梯度)上报给中心节点。中心节点本身没有数据集，只负责将分布式节点的训练结果进行融合处理，得到全局模型，并下发给分布式节点。然而，由于FL通过联邦平均(federated averaging，FedAvg)算法周期性融合整个模型，收敛速度较慢，收敛性能有一定的缺陷；其次，进行FL的设备需要保存及发送整个模型，对于设备计算、 存储及通信能力要求较高。
(10)迁移学习
迁移学习(transfer learning)是机器学习中的一个名词，是指一种学习对另一种学习的影响，或习得的经验对完成其它活动的影响。具体地，迁移学习是一种学习的思想和模式，迁移学习是指利用数据、任务或模型之间的相似性，将在旧领域学习过的模型应用于新领域的一种学习过程，其核心是找到新问题与旧问题之间的相似性，顺利实现知识的迁移。迁移广泛存在于各种知识、技能与社会规范的学习中。
迁移学习专注于存储已有问题的解决模型，并将其利用在其他不同但相关问题上。比如说，用来辨识汽车的知识(或者是模型)也可以被用来提升识别卡车的能力。计算机领域的迁移学习和心理学常常提到的学习迁移在概念上有一定关系，但是两个领域在学术上的关系非常有限。深度学习是让机器自主地从数据中获取知识，从而应用于解决新的问题；而迁移学习则侧重于将已学习过的知识迁移应用于解决新的问题。
(11)神经网络剪枝
在网络训练过程中寻求一种评判机制，以剔除掉不重要的连接、节点甚至卷积核，达到精简网络结构的目的。神经网络剪枝可分为两类，其一为训练中的约束，在损失函数添加约束项诱导网络稀疏化，无需重训练；其二为训练后剪枝，以剔除网络中相对冗余、不重要的部分。根据剪枝粒度，有层间剪枝和特征剪枝。两者分别降低网络深度和宽度，细粒度的有核内剪枝。
通过神经网络剪枝，使得训练时间减少，且随着计算量的减少，网络中连接每次迭代的速度得到提升，神经网络模型可以更快的收敛到最优解。也可使得运行速度快，稀疏化后的网络卷积层数、卷积层中的卷积核个数都更少，更简单更轻便的模型意味着更高效、快速的权重更新。神经网络剪枝的优势还在于更可行的嵌入式部署，即剪枝后的网络在移动设备及其他嵌入式设备上的应用提供了更广泛的可能。
(12)模型量化
模型量化是指通过降低权重所需的比特数来压缩原始网络，主要包括低精度和重编码两类方法。如低精度就是用更低位数的浮点数或整型数表示权重参数进行训练、测试或存储；重编码则是对原始数据进行重编码，采用更少的位数对原有数据进行表示，代表有二值化权重/二值化神经网络。模型量化是以较低的推理精度损失将连续取值(或者大量可能的离散取值)的浮点模型权重或流经模型的张量数据定点近似为有限多个离散值的过程，他是以更少位数的数据类型用于近似表示32位有限范围浮点型数据的过程，而模型的输入输出依然是浮点型，从而达到减少模型尺寸大小、减少模型内存消耗以及加快模型推理速度等目标。
目前对于图像分类许多量化技术都可达到无损压缩；但进行较为复杂的任务例如分割任务时使用量化通常会对模型精度带来巨大影响。
(13)知识蒸馏
知识蒸馏(knowledge distillation，KD)又称网络蒸馏，是模型压缩的一种常用的方法，不同于模型压缩过程中的剪枝和量化，知识蒸馏是通过构建一个轻量化的小模型，利用性能较优的大模型的监督信息来训练小模型，以期达到更优的性能和精度。具体地，大模型可以称为教师模型，小模型称为学生模型，来自教师模型输出的监督信息称为知识， 而学生模型迁移学习来自教师模型的监督信息的过程称为知识蒸馏。
图6示出了本申请实施例提供的一例系统架构600的示意图。如图6所示，数据采集设备660用于采集训练数据。示例性地，针对用于图像分类的神经网络模型而言，训练数据可以包括训练图像以及训练图像对应的分类结果，其中，训练图像的结果可以是人工预先标注的结果。
在采集到训练数据之后，数据采集设备660将该训练数据存入数据库630中，训练设备620基于数据库630中的训练数据训练得到目标模型/规则601。
下面对训练设备620基于训练数据得到目标模型/规则601的过程进行描述，训练设备620对输入的原始图像进行处理，并将输出的图像与原始图像进行对比，直到训练设备620输出的图像与原始图像的差值小于一定的预设阈值，从而完成目标模型/规则601的训练过程。
在一种可能的实现方式中，训练设备620可用于获取预训练的网络模型，基于目标数据集迁移预训练的网络模型，并在迁移后的网络模型中搜索子网络模型，以得到目标模型/规则601。该目标数据集可以存储于数据库630中。其中，训练设备620也可以用于预训练网络模型。该网络模型是基于源数据集训练得到的，该源数据集也可以存储于数据库630中。
本申请实施例中的目标模型/规则601具体可以为神经网络模型。例如，卷积神经网络、循环神经网络或者深度神经网络。需要说明的是，在实际的应用中，所述数据库630中维护的训练数据不一定都来自于数据采集设备660的采集，也有可能是从其他设备接收得到的，例如由客户设备640输入的目标数据集。另外需要说明的是，训练设备620也不一定完全基于数据库630维护的训练数据进行目标模型/规则601的训练，也有可能从云端或其他地方获取训练数据进行模型训练，上述描述不应该作为对本申请实施例的限定。
根据训练设备620训练得到的目标模型/规则601可以应用于不同的系统或设备中，例如应用于图6所示的执行设备610，所述执行设备610可以是终端，如手机终端，平板电脑，笔记本电脑，增强现实(augmented reality，AR)AR/虚拟现实(virtual reality，VR)，车载终端等，还可以是服务器或者云端等。如图6所示，执行设备610配置输入/输出(input/output，I/O)接口612，用于与外部设备进行数据交互，用户可以通过客户设备640向I/O接口612输入数据，例如，所述输入数据可以包括：客户设备输入的待处理的图像。
可选地，执行设备610中还可以包括预处理模块613，该预处理模块613用于根据I/O接口612接收到的输入数据(例如：图片数据)进行预处理，如果不需要对输入数据进行预处理，则直接将从I/O接口612接收的输入数据(例如：直接采集获得的图片数据)输入到执行设备610中进行处理。
在执行设备610对输入数据进行预处理，或者在执行设备610的计算模块611执行计算等相关的处理过程中，执行设备610可以调用数据存储系统650中的数据、代码等以用于进行相应的处理，也可以将相应处理得到的数据、指令等存入数据存储系统650中。
随后，I/O接口612将处理后的结果提供给用户设备640，例如上述得到的图像的分类结果返回给客户设备640以供客户使用。
应理解，训练设备620可以针对不同的目标或者不同的任务需求，基于不同的训练数据生成相应的目标模型/规则601，该相应的目标模型/规则601即可以用于实现上述目标 或完成上述任务，从而为用户提供所需的处理结果。
在图6中所示出情况下，用户可以手动给定输入数据，该手动给定的数据可以通过I/O接口612提供的界面进行操作。在另一种可能实现的方式中，客户设备640可以自动地向I/O接口612发送输入数据，如果要求客户设备640自动发送输入数据需要获得用户的授权，则用户可以在客户设备640中设置相应权限。用户可以在客户设备640查看执行设备610输出的结果，具体的呈现形式可以是显示、声音、动作等具体的方式。客户设备640也可以作为数据采集端，采集如图所示输入I/O接口612的输入数据及输出I/O接口612的输出结果作为新的样本数据，并存入数据库630。当然，也可以不经过客户设备640进行采集，而是由I/O接口612直接将如图6所示输入I/O接口612的输入数据及输出I/O接口612的输出结果，作为新的样本数据存入数据库630。
还应理解，图6示出的是本申请实施例提供的一种系统架构的示意图，本申请对该系统架构中的设备、器件、模块等之间的位置关系不构成任何限制，例如，在图6示出的系统架构中，数据存储系统650相对执行设备610是外部存储器，在其它可能的情况下，也可以将数据存储系统650置于执行设备610中。
如图6所示，根据训练设备620训练可得到目标模型/规则601，该目标模型/规则601可以是本申请实施例中的神经网络模型，具体的，本申请实施例构建的神经网络模型可以包括CNN、深度神经网络模型(deep neural networks，DNN)、深度卷积神经网络模型(deep convolutional neural networks，DCNN)、循环神经网络模型(recurrent neural network，RNN)等等，本申请实施例对此不做限定。
图7为本申请实施例涉及的一种卷积神经网络模型的示意性结构图。如图7所示，该卷积神经网络模型(CNN)700可以包括输入层710，卷积层/池化层720(其中池化层为可选的)，以及神经网络模型层730。例如，输入层710可以获取待处理的图像，并将获取到的待处理图像交由卷积层/池化层720以及后面的神经网络模型层730进行处理。下面对图7中的CNN 700中内部的层结构进行详细的介绍。
卷积层/池化层720：
卷积层：
如图7所示卷积层/池化层720可以包括如示例的721-726层，例如：在一种可能的实现方式中，721层为卷积层，722层为池化层，723层为卷积层，724层为池化层，725为卷积层，726为池化层；在另一种可能的实现方式中，721与722层为卷积层，723为池化层，724与725为卷积层，726为池化层。即卷积层的输出可以作为随后的池化层的输入，也可以作为另一个卷积层的输入以继续进行卷积操作。
下面将以卷积层721为例结合图像处理的方式，详细介绍一层卷积层的内部工作原理。
卷积层721可以包括多个卷积算子，卷积算子也可称为卷积核，其在图像处理中的作用相当于一个从输入图像矩阵中提取特定信息的过滤器，卷积算子本质上可以是一个权重矩阵，这个权重矩阵通常被预先定义，在对图像进行卷积操作的过程中，权重矩阵通常在输入图像上沿着水平方向一个像素接着一个像素(或者两个像素接着两个像素，这取决于步长stride的取值)的进行处理，从而完成从图像中提取特定特征的工作。该权重矩阵的大小应该与图像的大小相关，需要注意的是，权重矩阵的纵深维度(depth dimension)和输入图像的纵深维度是相同的。在进行卷积运算的过程中，权重矩阵会延伸到输入图像的 整个深度。因此，和一个单一的权重矩阵进行卷积会产生一个单一纵深维度的卷积化输出，但是大多数情况下不使用单一权重矩阵，而是应用多个尺寸(行×列)相同的权重矩阵，即多个同型矩阵。每个权重矩阵的输出被堆叠起来形成卷积图像的纵深维度，这里的维度可以理解为由上面所述的“多个”来决定。不同的权重矩阵可以用来提取图像中不同的特征，例如一个权重矩阵用来提取图像边缘信息，另一个权重矩阵用来提取图像的特定颜色，又一个权重矩阵用来对图像中的噪点信息进行提取等。该多个权重矩阵尺寸(行×列)相同，经过该多个尺寸相同的权重矩阵提取后的卷积特征图的尺寸也相同，再将提取到的多个尺寸相同的卷积特征图合并形成卷积运算的输出。
这些权重矩阵中的权重值在实际应用中需要经过大量的训练得到，通过训练得到的权重值形成的各个权重矩阵可以用来从输入图像中提取信息，从而使得卷积神经网络模型700进行正确的预测。
当卷积神经网络模型700有多个卷积层的时候，初始的卷积层(例如721层)往往提取较多的一般特征，该一般特征也可以称之为低级别的特征；随着卷积神经网络模型700深度的加深，越往后的卷积层(例如726层)提取到的特征越来越复杂，比如高级别的语义之类的特征，语义越高的特征越适用于待解决的问题。
池化层：
由于常常需要减少训练参数的数量，因此在卷积层之后常常需要周期性的引入池化层，即如图7中720所示例的721-726各层，可以是一层卷积层后面跟一层池化层，也可以是多层卷积层后面接一层或多层池化层。在图像处理过程中，池化层的目的就是减少图像的空间大小。池化层可以包括平均池化算子或最大池化算子，以用于对输入图像进行采样得到较小尺寸的图像。平均池化算子可以在特定范围内对图像中的像素值进行计算产生平均值作为平均池化的结果。最大池化算子可以在特定范围内取该范围内值最大的像素作为最大池化的结果。另外，与卷积层中用权重矩阵的大小应该与图像尺寸相关一样，池化层中的运算符也应该与图像的大小相关。通过池化层处理后输出的图像尺寸可以小于输入池化层的图像的尺寸，池化层输出的图像中每个像素点表示输入池化层的图像的对应子区域的平均值或最大值。
神经网络模型层730：
在经过卷积层/池化层720的处理后，卷积神经网络模型700还不足以输出所需要的输出信息。如前所述，卷积层/池化层720只会提取特征，并减少输入图像带来的参数。为了生成最终的输出信息(所需要的类信息或其他相关信息)，卷积神经网络模型700需要利用神经网络模型层730来生成一个或者一组所需要的类的数量的输出。因此，在神经网络模型层730中可以包括多层隐含层(如图7所示的731、732至73n)以及输出层740，该多层隐含层中所包含的参数可以根据具体的任务类型的相关训练数据进行预先训练得到，例如该任务类型可以包括图像识别，图像分类，图像的超分辨率重建等等。
在神经网络模型层730中的多层隐含层之后，也就是整个卷积神经网络模型700的最后层为输出层740，该输出层740具有类似分类交叉熵的损失函数，具体用于计算预测误差，一旦整个卷积神经网络模型700的前向传播(如图7由710至740方向的传播为前向传播)完成，反向传播(如图7由740至710方向的传播为反向传播)就会开始更新前面提到的各层的权重值以及偏差，以减少卷积神经网络模型700的损失，以及卷积神经网络 模型700通过输出层输出的结果和目标结果之间的误差。
本申请实施例提供的一种卷积神经网络模型的结构示意图可以如图8所示。在图8所示的卷积神经网络模型(CNN)800可以包括输入层810，卷积层/池化层820(其中池化层为可选的)，以及神经网络模型层830。与图7相比，图8中的卷积层/池化层820中的多个卷积层/池化层并行，将分别提取的特征均输入给神经网络模型层230进行处理。
应理解，图7和图8所示的卷积神经网络模型仅作为一种神经网络模型的结构性示例。在具体的应用中，卷积神经网络模型还可以以其他网络模型的形式存在，本申请实施例对此不做限定。
图9为本申请实施例提供的一种芯片的硬件结构，该芯片包括神经网络模型处理器900。该芯片可以被设置在如图6所示的执行设备610中，用以完成计算模块611的计算工作。该芯片也可以被设置在如图6所示的训练设备620中，用以完成训练设备620的训练工作并输出目标模型/规则601。如图7和图8所示的卷积神经网络模型中各层的算法均可在如图9所示的芯片中得以实现。
神经网络模型处理器NPU 50作为协处理器挂载到主中央处理器(central processing unit，CPU)上，由主CPU分配任务。NPU的核心部分为运算电路903，控制器904控制运算电路903提取存储器(权重存储器或输入存储器)中的数据并进行运算。
在一种实现方式中，运算电路903内部包括多个处理单元(process engine，PE)。运算电路903可以是二维脉动阵列，运算电路903还可以是一维脉动阵列或者能够执行例如乘法和加法这样的数学运算的其它电子线路。在一种可能实现中，运算电路903是通用的矩阵处理器。
例如，假设有输入矩阵A，权重矩阵B，输出矩阵C。运算电路从权重存储器902中获取矩阵B相应的数据，并缓存在运算电路中每一个PE上。运算电路从输入存储器901中获取矩阵A与矩阵B进行矩阵运算，得到的矩阵的部分结果或最终结果，保存在累加器(accumulator)908中。
向量计算单元907可以对运算电路的输出做进一步的处理，如向量乘，向量加，指数运算，对数运算，大小比较等等。例如，向量计算单元907可以用于神经网络模型中非卷积或非FC层的网络计算，如池化(pooling)，批归一化(batch normalization)，局部响应归一化(local response normalization)等。
在一些实现方式中，向量计算单元907能将经处理的输出的向量存储到统一缓存器906之中。例如，向量计算单元907可以将非线性函数应用到运算电路903的输出，例如累加值的向量，可用以生成激活值。在一些实现中，向量计算单元907生成归一化的值、合并值，或二者均有。在一种可能的实现方式中，处理过的输出的向量能够用作到运算电路903的激活输入，例如用于在神经网络模型中的后续层之中的使用。
统一存储器906用于存放输入数据以及输出数据。
存储单元访问控制器905(direct memory access controller，DMAC)用于将外部存储器中的输入数据搬运到输入存储器901和/或统一存储器906、将外部存储器中的权重数据存入权重存储器902，以及将统一存储器906中的数据存入外部存储器。
总线接口单元(bus interface unit，BIU)910，用于通过总线实现主CPU、DMAC和取指存储器909之间进行交互。
与控制器904连接的取指存储器(instruction fetch buffer)909，用于存储控制器904使用的指令；
控制器904，用于调用取指存储器909中缓存的指令，实现控制该运算加速器的工作过程。
一般地，统一存储器906，输入存储器901，权重存储器902以及取指存储器909均为片上(On-Chip)存储器，外部存储器为该NPU外部的存储器，该外部存储器可以为双倍数据率同步动态随机存储器(double data rate synchronous dynamic random access memory，简称DDR SDRAM)、高带宽存储器(high bandwidth memory，HBM)或其他可读可写的存储器。
上文中介绍的图6中的训练设备610能够执行本申请实施例的获取神经网络模型的方法的各个步骤，图9所示的芯片也可以用于执行本申请实施例的获取神经网络模型的各个步骤。
如图10所示，本申请实施例提供了一种系统架构1000。该系统架构包括本地设备1001、本地设备1002以及执行设备1010和数据存储系统1050，其中，本地设备1001和本地设备1002通过通信网络与执行设备1010连接。
执行设备1010可以由一个或多个服务器实现。可选的，执行设备1010可以与其它计算设备配合使用，例如：数据存储器、路由器、负载均衡器等设备。执行设备1010可以布置在一个物理站点上，或者分布在多个物理站点上。执行设备1010可以使用数据存储系统1050中的数据，或者调用数据存储系统1050中的程序代码来实现本申请实施例的获取神经网络模型方法或训练神经网络模型的方法。
通过上述过程执行设备1010能够获取一个目标神经网络模型，该目标神经网络模型可以用于图像分类或者进行图像处理等等。
用户可以操作各自的用户设备(例如本地设备1001和本地设备1002)与执行设备1010进行交互。每个本地设备可以表示任何计算设备，例如个人计算机、计算机工作站、智能手机、平板电脑、智能摄像头、智能汽车或其他类型蜂窝电话、媒体消费设备、可穿戴设备、机顶盒、游戏机等。
每个用户的本地设备可以通过任何通信机制/通信标准的通信网络与执行设备1010进行交互，通信网络可以是广域网、局域网、点对点连接等方式，或它们的任意组合。
在一种实现方式中，本地设备1001、本地设备1002从执行设备1010获取到目标神经网络模型的相关参数，将目标神经网络模型部署在本地设备1001、本地设备1002上，利用该目标神经网络模型进行图像分类或者图像处理等等。
在另一种实现中，执行设备1010上可以直接部署目标神经网络模型，执行设备1010通过从本地设备1001和本地设备1002获取待处理图像，并采用目标神经网络模型对待处理图像进行分类或者其他类型的图像处理。
上述执行设备1010也可以为云端设备，此时，执行设备1010可以部署在云端；或者，上述执行设备1010也可以为终端设备，此时，执行设备1010可以部署在用户终端侧，本申请实施例对此并不限定。
图11是本申请实施例提供的一种训练模型的装置部署示意图，如图11中的(a)所示，训练装置1110可以部署在在云环境中，该云环境是指云计算模式下利用基础资源向 用户提供云服务的实体。该云环境包括云数据中心和云服务平台，所述云数据中心包括云服务提供商拥有的大量基础资源(包括计算资源、存储资源和网络资源)，云数据中心包括的计算资源可以是大量的计算设备(例如服务器)。
训练装置1110可以是在云数据中进行神经网络模型训练的服务器，或者还可以是对神经网络模型进行训练的虚拟机。
训练装置1110还可以是部署在云数据中心的服务器或者虚拟机上的软件装置，该软件装置用于对神经网络模型进行训练，该软件装置可以分布式地部署在多个服务器上、或者分布式地部署在多个虚拟机上、或者分布式地部署在虚拟机和服务器上。
如图11所示，训练装置1110可以由云服务提供商在云服务平台抽象成一种训练神经网络模型的云服务提供给用户，用户在云服务平台购买该云服务后，云环境利用该云服务向用户提供训练神经网络的云服务。
例如，如图11中的(b)所示，用户可以通过应用程序接口(application program interface，API)或者通过云服务平台提供的网页界面上传待训练的神经网络模型(或者还可以上传原始训练集)至云环境，由训练装置1110接收待训练的神经网络模型以及训练集，通过训练装置1110中对待训练的神经网络模型进行训练，最终训练得到的目标神经网络由训练装置1110返回至用户所在的边缘设备。
其中，边缘设备是指在数据产生源头和云中心之间任一具有计算资源和网络资源的设备。比如，手机就是人与云中心之间的边缘设备，网关是智能家居和云中心之间的边缘设备。在理想环境中，边缘设备指是指在数据产生源附近分析或处理数据的设备。由于没有数据的流转，进而减少网络流量和响应时间。
本申请实施例中的边缘设备可以是具有计算能力的移动电话、平板个人电脑(tablet personal computer，TPC)、媒体播放器、智能家居、笔记本电脑(laptop computer，LC)、个人数字助理(personal digital assistant，PDA)、个人计算机(personal computer，PC)、照相机、摄像机、智能手表、可穿戴式设备(wearable device，WD)或者自动驾驶的车辆等。应理解，本申请实施例对边缘设备的具体形式不作限定。
例如，用户可以通过应用程序接口或者通过云服务平台提供的网页界面上传目标任务的类型至云环境，进一步地，还可以上传训练集，由训练装置接收目标任务类型以及训练集，通过训练装置1110中对目标任务的类型对应的神经网络模型(例如，待训练的图像识别模型)进行训练，最终训练得到的目标神经网络模型(例如，目标图像识别模型)由训练装置1110返回至用户所在的边缘设备。
上述训练装置1110可以为如图11中的(a)所示部署在云环境中；或者，上述训练装置1110也可以为终端设备，此时，执行设备1110可以部署在用户终端侧，本申请实施例对此并不限定。
目前，一种常见的神经网络模型的传输方案是通过将神经网络模型看作信源进行处理，通过信源编码及信道编码等过程以获得满足信道带宽需求的神经网络的数据，后续通过无线信道发送至接收端设备，接收端设备接收到数据后进行信道译码及信源译码等过程以完成神经网络模型的传输。在上述模型传输方案中，可使用一些AI技术，例如网络剪枝、网络参数量化与知识蒸馏等来获得尺寸相对较少的神经网络模型，从而可在不显著影响推理准确性的前提下，减少接收端的存储空间与推理时延。但是上述传输方案仍是将传统的 通信系统作为神经网络模型的传输通道，即需要收发端网络相同，本质上依然属于传统的端对端通信系统，多个神经网络模型之间相互独立传输，且并没有使得通信过程参与AI应用的训练和推理，进而导致神经网络模型的传输性能较低，难以满足当前多神经网络模型传输的需求。
因此，本申请针对未来通信场景中的多个神经网络模型的高性能传输需求，提出了一种神经网络模型联合传输的方法，将神经网络与传统的通信系统进行结合，针对任务需求将不同发送端上的多个神经网络模型经过信道进行组合以产生新的神经网络模型，从而实现接收端设备对多个神经网络模型的传输需求，有利于控制AI应用的多样性和有效性，避免训练多个单一神经网络的传输，减少在多任务需求多模态下的多个神经网络模型的资源开销和时延，并在一定程度上节省接收端及发送端设备对于神经网络模型的存储空间。
在本申请实施例中，神经网络模型是指经过大量有标签的训练数据得到的用于执行认知计算的程序和数据。神经网络模型包括神经网络架构组件与神经网络参数组件，其中，神经网络架构组件是指神经网络模型中与神经网络算法相关的网络及其层次结构，即上述神经网络模型中用于执行认知计算的程序。神经网络参数组件是指训练神经网络模型时得到的大量参数，作为神经网络架构中的神经元的值，即上述神经网络模型中用于执行认知计算的数据。
下面结合图12对本申请实施例的模型传输的方法进行详细的介绍。图12所示的方法中的发送端设备和终端设备均可以由网络设备或者由终端设备执行，本申请实施例对此不做限定。例如，当发送端设备为网络设备时，接收端设备可以为终端设备或者网络设备。该方法1200包括步骤S1210至步骤S1240。下面对步骤S1210至S1240进行详细说明。
应理解，在本申请实施例中，适用于如图12所示的方法的系统中可包括至少一个发送端设备和一个接收端设备。例如，本申请实施例的方法可用于一个发送端内的多个神经网络模型进行联合传输的场景，或者多个不同发送端内的至少一个神经网络网络模型进行联合传输的场景。
应理解，在本申请实施例中，接收端设备也可以称为接收端装置，发送端设备也可以成为发送端装置，本申请对此不做限定。
下面以一个发送端设备内的多个神经网络模型的联合传输场景为例对本申请的模型的传输方法进行详细说明。
S1210，发送端设备获取N个第一模型的信息。
其中，该N个第一模型可对应N个第一任务，N为大于或等于2的整数。
应理解，在本申请实施例中，N个第一模型可根据任务需求信息进行确定，例如，当接收端边缘设备需要进行图像分类处理的业务时，发送端设备可通过接收该任务需求信息来确定该符合该业务需求的多个发送端神经网络模型，或者，该符合业务需求的发送端神经网络模型也可由上图11中的所涉及的云端设备或云服务平台根据业务需求信息进行确定，本申请实施例对此不做限定。其中，上述与设备可以理解成为发送端设备提供服务的服务器。
还应理解，该N个第一模型可理解为预训练的神经网络模型，例如，当任务需求为图像分类时，发送端可根据源数据集(如，Image net数据集)训练得到符合业务需求的神经网络模型供发送端进行使用。应理解，该预训练的神经网络模型可以在离线阶段完成，也 可以在线完成。示例性的，当预训练的神经网络模型是在离线阶段完成的情况下，发送端设备在根据任务需求信息获取神经网络模型时，可以直接获取已经训练过的神经网络模型，而无需在线执行预训练的操作。
还应理解，该N个第一模型为符合任务需求的任意的神经网络模型，本申请实施例对此不做限定。
还应理解，上述任务需求信息可包括目标神经网络模型的外部特征值，该神经网络模型的外部特征值可理解为在训练神经网络模型的过程中不会在训练过程中而改变的值，该外部特征值也可以是在神经网络模型的训练之前已经人为或根据经验确定好的外部特征值。
其中，该目标神经网络模型的外部特征值包括以下信息中的一个或者多个信息：神经网络计算图信息、优化器信息、超参数信息。
还应理解上述神经网络计算图信息可用于指示神经网络的结构，其中，神经网络的结构包括以下一项或者多项：神经网络的类型、神经网络的层数、神经网络中各层的节点数、神经网络的节点连接方式。
在本申请实施例中涉及的神经网络的类型可以是全连接神经网络、卷积神经网络、深度神经网络、循环神经网络等，本申请实施例对此不做限定。
神经网络中的超参数可包括三类，即网络参数、优化参数、正则化参数。其中，网络参数是指网络层与层之间的交互方式(例如，相加、相乘或者串接等)、卷积核的数量和卷积核尺寸、网络层数(或称深度)和激活函数等。优化参数是指学习率(learning rate)、批样本数量(batch size)、不同优化器的参数(例如，动量参数momentum)以及部分损失函数的可调参数。正则化参数是指权重衰减系数与神经网络模型的丢弃(dropout)参数等。
应理解，神经网络模型的超参数包括在神经网络模型训练过程中不变的参数，即超参数不是通过神经网络模型的训练得到的，通常是在神经网络模型的训练之前确定的。在本申请实施例中该超参数信息可以根据经验或任务需求人为地进行设置，本申请实施例对此不做限定。
在一种实现方式中，该N个第一模型可对应N个信号A。
应理解，该N个信号A用于承载所对应的第一模型的信息。
在一种实现方式中，发送端边缘设备可通过自身的编码器对N个第一模型的信息进行编码，并生成N个信号A。
在一种可能的实现方式中，发送端设备对该N个第一模型的信息依次通过信源编码、信道编码、调制以及上变频或者下变频处理后得到对应的N个信号A。
在本申请实施例中，模型的信息可包括神经网络模型的外部特征值和/或神经网络模型的参数，本申请模型的信息，不局限于是第一模型的或者目标模型的，只要是本申请出现的神经网络模型均可包含这些信息。具体地，该神经网络模型的外部特征值包括以下信息中的一个或者多个信息：神经网络计算图信息、优化器信息、超参数信息；神经网络模型的参数包括以下中的一项或者多项：权重矩阵、权重向量、偏置矩阵、偏置向量、激活函数。
在本申请实施例中，该N个第一模型对应的N个信号A可以为模拟信号或者数字信 号，本申请实施例对此不做限定。
应理解，该N个信号A与上述N个第一模型一一对应，其中，N为大于或等于2的整数。
S1220，发送端设备获取融合辅助信息。
具体地，该融合辅助信息包括目标模型的外部特征值。
在本申请实施例中，该目标模型的外部特征值包括以下信息中的一项或者多项：目标模型的神经网络计算图信息、目标模型的优化器信息、目标模型的超参数信息
在一种实现方式中，该目标模型可对应第二任务，该第二任务与上述N个第一模型对应的N个第一任务可以相同，也可以不同，本申请实施例对此不做限定。例如，发送端中的第1个第一模型可用于对图像中的树木进行检测，第2个第一模型可用于对图像中的鸟类进行检测，经过本申请提供的模型联合传输的方法得到的目标模型可用于对图像中的花朵进行分类。
S1230，根据N个第一模型的信息和该融合辅助信息确定N个第一信号。
具体地，该N个第一信号是N个第一模块对上述N个第一模型的信息和该融合辅助信息处理后得到的。
应理解，在本申请实施例中，该第一模型可以为实现特定功能的模块，或者是满足任务需求的神经网络模型，作为示例本申请实施例对此不做限定。
在本申请实施例中，该N个第一模块是根据训练数据训练得到的训练好的神经网络模型，其中，该N个第一模块可与上述N个第一模型一一对应。
在一种实现方式中，该N个第一模块可理解为N个神经网络模型，该神经网络模型与上述的N个第一模型一一对应。
应理解，该N个第一模块与上述N个第一模型一一对应可以理解为N个第一模块所对应的N个业务需求与上述N个第一模型对应的N个业务需求一一对应，且彼此一一对应的第一模型和第一模块对应于同一业务需求。例如，N个第一模块与N个第一模型所对应的业务需求可均为图像分类业务。
其中，训练数据的类型与神经网络模型的任务需求有关。示例性地，当神经网络模型用于图像处理任务时(例如，图像分类、图像检测、图像分割或者图像生成等)，则该训练数据可为图像。当神经网络模型用于文本处理任务时(例如，文本识别或者文本翻译等)，则该训练数据可为文本。当神经网络模型用于语音处理任务时(例如，语音识别)，则该训练数据可为语音数据。作为示例而非限定，本申请实施例对此不做限定。
在一种实现方式中，上述训练数据包括M个第一训练信号和目标模型的数据，该目标模型的数据可以理解为目标模型的外部特征值，即目标模型的神经网络计算图信息、优化器信息和超参数信息。其中，上述M个第一训练信号与M个训练模型一一对应，该M个训练模型对应于M个业务需求。
还应理解，上述N个第一模型对应的N个业务包括上述M个第一训练模型对应的M个业务。
在一种实现方式中，发送端设备可根据N个第一模型中的除了第i个第一模型以外的至少一个第一模型的信息来确定该第i个第一模型所对应的第一模块。其中，该至少一个第一模型的信息是指神经网络模型的参数和外部特征值，其中神经网络模型的参数是指神 经网络模型的内部特征参数，即可在训练模型的过程中进行调整改变的参数。
其中，上述模型的参数是指以下中的一项或者多项：权重矩阵、权重向量、偏置矩阵、偏置向量、激活函数。
示例性地，当2个第一模型对应的业务均为图像处理业务中的目标检测业务时，若第1个第一模型可用于检测图像中的树木，第2个第一模型可用于检测图像中的鸟类，则该第1个第一模型可根据第2个第一模型中的参数(例如，权重矩阵、权重向量等)获得与第1个第一模型对应的第1个第一模块，其中，该第1个第一模块的业务与第1个第一模型的业务对应，且该第1个第一模块可同时用于对图像中的鸟类与树木进行检测。
下面详细介绍该N个第一模块的训练过程。
发送端设备可获取M个原始模型，该M个原始模型与上述M个第一训练信号一一对应，其中，M为大于等于2的整数；发送端将M个第一训练信号依次输入与其一一对应的M个原始模型中，以获得M个第二训练信号，其中，该M个第二训练信号与上述M个第一训练信号一一对应。
随后，发送端将上述M个第二训练信号在同一信道中进行叠加以获得第三训练信号；接收端设备通过对该第三训练信号进行解析后得到中间模型，应理解，该接收端设备可通过对该第三训练信号进行译码后得到与第三训练信号对应的中间模型；发送端设备根据中间模型与该目标模型的偏差，对上述M个原始模型的参数进行调节处理，以获得训练后的模型库。
应理解，该训练后的模型库包括经过上述调节处理后的原始模型，并且经过调节处理后的中间模型与目标模型的偏差在预设范围内，或者，也可以理解为将中间模型的输出值与目标模型的目标值进行比较，直到该中间模型的输出值与目标值的差值小于一定的预设阈值，即损失函数的值达到预设阈值时，停止该中间模型的训练。
最后，发送端设备可从该训练后得到的模型库中获取该N个第一模块。具体地，发送端设备可根据任务需求在该训练后的模型库中搜索符合当前任务需求的神经网络模型作为第一模块。
在一种可能的实现方式中，上述原始模型可根据任务需求信息进行确定，即该原始模型可根据符合任务需求信息的源数据集(例如，ImageNet、Labelme等数据集)训练得到。
应理解，上述训练停止的条件可以根据实际的应用需求和应用场景灵活设置。例如，训练的停止条件可以包括但不限于，训练时间达到预设的时长、训练次数达到预设次数、损失函数的值或者衰减幅度达到预设阈值等。
在一种可能的实现方式中，N个第一模型经过处理后可得到与其一一对应的N个信号A，该N个信号A经过上述训练后的N个第一模块处理后可得到N个第一信号，该N个第一信号在同一信道中进行叠加以得到目标模型对应的第二信号。
应理解，训练好的神经网络模型可以执行目标任务。例如，目标任务可以为图像处理任务，包括：目标检测、图像分割、实例分割、图像去噪、图像超分辨等。或者，该目标任务还可以为语音处理任务，例如，语音识别。或者，该目标任务还可以为文本处理任务，例如，文本识别或者文本翻译。
应理解，在本申请实施例中，对于N个第一模块的训练过程可以由神经网络模型的训练装置来执行，该训练装置可以由云端设备，也可以由终端设备进行执行，例如，边缘设 备、电脑、服务器等运算能力足以用来执行神经网络模型的训练方法的装置，也可以是由云端设备和终端设备构成的系统来执行。具体地，上述模型的训练过程可以由图6中的训练设备620、图9中的神经网络处理器900、图10中的执行设备1010或者本地设备1001或1002、图11中的训练装置1110来执行，本申请实施例对此不做限定。
还应理解，该N个第一模块的训练过程可以在离线阶段完成，也可以在线完成。示例性的，若该N个第一模块是在离线阶段完成的情况下，发送端设备可以直接获取已经训练过的神经网络模型，而无需在线执行预训练的操作。
在一种实现方式中，发送端设备根据N个第一模块对上述N个信号A进行处理，得到N个第一信号。
具体地，发送端设备对N个第一模型对应的N个信号A进行处理后可得到N个第一信号。
在一种可能的实现方式中，发送端设备对N个第一模型进行处理后可得到N个信号A，并根据获取的N个第一模块对该N个信号A进行处理后得到N个第一信号。
其中，该N个信号A与该N个第一模型一一对应，该N个第一模块与该N个第一模型一一对应。应理解，该N个第一模块与该N个第一模型一一对应可以理解为N个第一模型的业务与N个第一模块的业务一一对应。该N个第一模块是根据训练数据训练得到的，且该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号对应M个训练模型，每个第一训练信号是所对应的训练模型的信号，该目标模型能够满足该M个训练模型对应的业务需求，且该M个第一训练模型对应的M个业务包括N个第一模型对应的N个业务。
其次，N个第一信号中的第n个第一信号是根据第n个第一模块对第n个信号A进行处理后得到的信号，该第n个信号A是第n个第一模块对应的第一模型的信号，该N个信号A与该N个第一模型一一对应，且彼此对应的第一模型和第一模块对应同一业务，n∈[1，N]，N为大于或等于2的整数，M为大于或等于2的整数；
例如，N＝5时，该发送端产生的5个第一信号中的第1个第一信号是根据第1个第一模块对第1个信号A进行处理后得到的信号，其中，该第1个信号A与第1个第一模型对应，且第1个第一模型与第1个第一模块对应同一业务。
应理解，该N个第一信号可在信道中叠加为第二信号，该第二信号与目标模型对应。
还应理解，该第二信号可经过接收端设备进行译码后得到任务所需的目标模型的数据，并进一步得到目标神经网络模型。
在一种可能的实现方式中，发送端设备还可通过发送该N个第一信号的信道的参数来确定上述N个第一模块，其中，该N个第一模块与N个第一信号一一对应。
上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
示例性地，发送端设备可根据信道参数中的信道状态信息和信噪比来获得神经网络模型的配置信息，并进一步确定与上述N个第一模型一一对应的N个第一模块。
在一种实现方式中，发送端设备可向接收端设备发送参考信息，该参考信息是发送端供给接收端设备用于进行信道估计或者信道探测的信息，例如，发送端设备可以以发送参考信息的方式进行信道信息的实时测量，随后接收端设备将根据获得的信道的参数来确定 发送端设备的神经网络模型的配置参数，进而实现获得N个第一模块的目的。其中，上述神经网络模型的配置参数包括神经网络模型的外部特征值和/或神经网络模型的参数。该神经网络模型的外部特征值包括以下信息中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息；神经网络模型的参数包括以下中的一项或者多项：权重矩阵、权重向量、偏置矩阵、偏置向量、激活函数。
S1240，发送N个第一信号。
在一种可能的实现方式中，发送端设备将生成的N个第一信号发送至接收端设备，其中，上述N个第一信号进行叠加后可生成第二信号，接收端设备接收该第二信号并对其译码后可得到符合任务需求的目标模型的数据。并进一步得到目标神经网络模型，其中，该目标模型与第二信号对应。
在本申请实施例中，模型传输的方法1200还可以通过图13所示的模型传输的流程图进行表示。应理解，发送端所涉及的步骤已在上述图12中进行详细描述，此处不再赘述。
如图13所示，通过发送端设备得到的N个第一信号可在信道中进行叠加以得到第二信号，其中，该第二信号与目标模型对应。
在一种实现方式中，接收端设备接收到第二信号后，可通过解析模块处理该第二信号后得到目标模型的信息，并最终获得符合任务需求的目标神经网络模型。其中，该解析模块所对应的解析方法与模型的外部特征值相关。
在本申请实施例中，对于多个发送端设备内的神经网络模型的联合传输的方法与上述方法相似，此处为了简略不再赘述。
根据上述模型联合传输的方法，可以实现多个不同神经网络模型的联合传输，以及减小多个神经网络模型在传输过程中的资源开销和时延，并在一定程度上降低接收端设备对于神经网络模型的存储空间。
下面将结合附图14与18对本申请实施例提供的神经网络模型的联合传输的方法进行详细描述。
图14为本申请实施例提供的一例模型传输方法的流程示意图。该方法包括步骤S1410至步骤S1460。下面对步骤S1410至S1460进行详细的说明。
应理解，在本申请所提供的模型传输的应用场景下，发送端设备可以为一个或多个。例如，一个发送端设备内的多个神经网络向一个接收端设备进行模型的联合传输，或者三个发送端设备向一个接收端设备进行模型的联合传输。
S1410，接收端向发送端发送任务需求信息。
相应地，发送端获取接收端发送的任务需求信息。
在一种可能的实现方式中，接收端设备可向发送端设备发送任务需求信息，发送端设备根据接收到的任务信息，在离线训练得到的接收端神经网络模型库中选择符合任务需求的一组或者多组神经网络模型作为发送端的神经网络模型。应理解，此处确定的一组或多组神经网络模型可理解为上述方法1200中的N个第一模型。
在一种可能的实现方式中，该任务需求信息还可包括任务需求的准确度与接收端设备的性能。
应理解，上述任务需求信息中还可包括与任务需求信息对应的目标神经网络模型的译码方法，该目标神经网络模型的译码方法与接收端知识库所能表达出的目标神经网络模型 的结构相关，经过接收端译码之后的神经网络应是知识库中所能表达出的神经网络模型。
S1420，发送端根据任务需求信息确定目标神经网络模型的外部特征值。
具体地，发送端接收到接收端发送的任务需求信息后，可选择符合任务需求的多个神经网络模型作为发送端神经网络模型。其中，该发送端神经网络模型的组合方案可以是经过上述信道通信运算方式后得到的可实现方式中一种或多种组合的发送端神经网络模型。
其中，该目标神经网络模型的外部特征值包括以下信息中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息。
应理解，在本申请实施例中的神经网络模型的外部特征值可理解为在训练神经网络模型的过程中不会因训练而改变的特征值，其中，该外部特征值也可以是在神经网络模型的训练之间已经确定好的特征值。
还应理解上述神经网络计算图信息可用于指示神经网络的结构，其中，神经网络的结构包括以下一项或者多项：神经网络的类型、神经网络的层数、神经网络中各层的节点数、神经网络的节点连接方式。
在本申请实施例中涉及的神经网络的类型可以是全连接神经网络、卷积神经网络、深度神经网络、循环神经网络等，作为示例而非限定，本申请实施例对此不做限定。
神经网络中的超参数可包括三类，即网络参数、优化参数、正则化参数。其中，网络参数是指网络层与层之间的交互方式(例如，相加、相乘或者串接等)、卷积核的数量和卷积核尺寸、网络层数(或称深度)和激活函数等。优化参数是指学习率(learning rate)、批样本数量(batch size)、不同优化器的参数(例如，动量参数momentum)以及部分损失函数的可调参数。正则化参数是指权重衰减系数与神经网络模型的丢弃(dropout)参数等。
应理解，神经网络模型的超参数包括在神经网络模型训练过程中不变的参数，即超参数不是通过神经网络模型的训练得到的，通常是在神经网络模型的训练之前确定的。在本申请实施例中该超参数的信息可以根据经验或任务需求人为地进行设置，本申请实施例对此不做限定。
S1430，发送端设备向接收端设备发送融合辅助信息。
相应地，接收端设备接收发送端发送的融合辅助信息。
该融合辅助信息用于指示符合任务需求的目标神经网络模型的外部特征值。
在一种实现方式中，接收端设备接收到第一信息后，根据该融合辅助信息中的目标神经网络模型的外部特征值(包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息)来预留目标神经网络模型的存储空间。
在本申请实施例中，发送端设备与接收端设备之间可通过维护多个目标神经网络模型的外部特征值的信息来实现收发端设备之间的目标神经网络模型的外部特征值的信息同步。具体地，在知识库中目标神经网络模型的外部特征值的列表可使用标识信息进行识别与区分。
示例性地，同步知识库中的目标神经网络模型的外部特征值的列表可由下表进行表示：
表1

表2
表3
表1至表3的第一列分别为神经网络模型的外部特征值，第二列为与该神经网络模型的外部特征值分别对应的ID信息。收发端设备可通过第一信息即上述列表中的ID组合进行收发端的目标神经网络模型信息的同步。例如，当接收端设备接收到的第一信息对应的ID组合为0001、1002和2001时，通过上述知识库中的对应列表可以得出，接收端接收到的目标神经网络模型的外部特征值的信息分别为：目标网络模型结构中包括卷积核尺寸为3*3的的卷积层，选用的网络模型优化器为均方根传播(root mean square prop，RMSProp)的优化算法以及网络模型的超参数为：学习率0.01，动量超参数的值为0.9。
还应理解，上述表2中的网络模型优化器与接收端设备的硬件能力相关，例如，在某些情况下，因终端设备的算力受限只能运行较为简单的网络模型结构或者优化器。
在一种可能的实现方式中，该融合辅助信息还包括参考信息，该参考信息用于供给接收端进行信道估计或者信道探测，以获得信道的参数。具体地，发送端可以以发送参考信息的方式进行信道信息的实时测量，随后接收端将根据获得的信道的参数来确定发送端的神经网络模型的配置参数，进而获得训练后的发送端神经网络模型的目的。
其中，上述神经网络模型的配置参数包括神经网络模型的外部特征值和/或神经网络模型的参数。该神经网络模型的外部特征值包括以下中的一项或者多项：神经网络计算图信息、优化器信息、超参数信息；神经网络模型的参数包括以下中的一项或者多项：权重矩阵、权重向量、偏置矩阵、偏置向量、激活函数。
在一种可能的实现方式中，上述步骤S1410至步骤可S1430可由为发送端设备提供服务的云端设备执行。
可选地，S1440，接收端设备向发送端设备发送融合辅助信息。
对应的，发送端设备接收接收端设备发送的融合辅助信息，以实现收发端对目标神经网络模型的外部特征值的同步。
S1450，发送端进行神经网络模型的联合传输。
具体地，该步骤可采用上述方法1200的方式执行，此处为简略不再赘述。
在本申请实施例提供了一种信道通信运算方式，可表示为发送端设备获取的网络模型经过编码模块Ei与第一模块处理后产生的神经网络模型Ei(Mi)，并通过信道叠加后得到的神经网络模型为：
其中，hi表示第i个发送端神经网络模型经过的信道，n表示噪声干扰信号，上述公 式(3)中的符号表示神经网络模型的叠加运算过程。
例如，当N＝3时，即发送端设备存在三个神经网络模型即M1、M2、M3，经发送端编码模块E1、E2与E3产生的神经网络模型对应的信号分别为E1(M1)、E2(M2)与E3(M3)，并经过训练后的与其一一对应的神经网络模型m1、m2、m3处理后的网络模型可表示为E1(m1)、E2(m2)与E3(m3)，此时经过信道产生的叠加后的神经网络模型可表示为：
随后发送端设备经过信道传输的神经网络模型经过接收端译码模块D处理后产生的神经网络模型可表示为：
例如，N＝3时，最后经过接收端的译码模块D得到的神经网络模型可表示为：
其中，M4为接收端最终获得的符合任务需求的神经网络模型。
为了进一步理解本申请实施例的信道通信运算的具体实现方式，结合神经网络结构的拓扑分析从基础图层面与计算图两个方面进行详细解释。
1.基础图层面(underlying graph)
如图15所示，本申请实施例涉及的神经网络模型可视为经典的有向图结构，其中，该有向图是一幅具有方向性的图，是由一组顶点和一组有方向的边组成，每条边都连接一对有序的顶点。该顶点可看作神经元节点或是神经单元，不同神经元节点之间的连线各对应一个权重值，该权重值可经过训练过程进行调整。具体地，图15中的神经网络模型可以按输入层、隐藏层及输出层进行划分，将神经网络模型可通过有向图数据结构存储表达。因而在该基础图层面下的信道通信运算可表示为不同神经网络模型对应的有向图结构的组合，即将多个发送端对应的神经网络模型的有向图数据结构中的相同部分进行叠加处理。
具体地，该信道通信运算利用基础图层面表示出的神经网络模型的联合传输过程可分为以下几个步骤：
1.选取部分可训练的神经网络模型，例如通过卷积神经网络中的深层卷积层进行抽象特征提取处理后的神经网络模型，并对其按输入层、隐藏层及输出层进行逐层定义，表示出的有向图数据如图15所示。神经网络模型m1为3层神经网络，包含一个隐藏层(也可称为第1层)，其中，输入层(也可称为第0层)有2个神经元，隐藏层有1个神经元，输出层(也可称为第2层)有2个神经元。神经网络模型m2为3层神经网络，同样包含一个隐藏层，其中，该神经网络模型m2的输入层有2个神经元，隐藏层有2个神经元，输出层有2个神经元。
上述神经网络模型m1可以进一步表达为：神经网络模型m2可以表达为：
其中，l表示神经网络模型对应的有向图结构中的网络节点该网络节点也可以称为神经元或网络单元，l的上标表示神经网络模型内的神经元所在的层数，下标表示有向图中神经网络模型的神经元在所属层的位置序号。
2.发送端将神经网络模型对应的神经网络计算图信息、优化器及超参数等特征值通过同步知识库信息发送给接收端，接收端根据接收到的同步知识库网络信息预留符合任务需求的神经网络模型对应的存储空间。
3.各发送端神经网络模型的数据通过编码模块并经过训练后的神经网络模型处理后得到的信号通过同一信道发送至接收端。
在一种实现方式中，经过发送端发出的神经网络模型的信号可在信道中进行叠加处理得到目标神经网络模型对应的信号。
如图15所示，接收端得到的组合后的神经网络模型m3为3层神经网络，该神经网络模型m3包含一个隐藏层，其中，输入层包含3个神经元，隐藏层包含2个神经元，输出层包含3个神经元。该组合后的神经网络模型M3可表示为：
其中，神经网络模型m3中的与为需要借助信道通信运算进行叠加的部分，神经网络模型m3的其它部分为各发送端可独立传输的部分。
在一种可能的实现方式中，发送端发出的神经网络模型可在在同一信道上发送需要叠加的神经网络模型的参数，其余神经网络模型的参数在正交信道上进行发送。
具体地，在同一信道上传输神经网络模型的叠加部分的技术方案可表示如下：
各发送端通过传输两个由所需时钟频率分隔的单频正弦(或者余弦)信号，其中，两个时钟频率分别为f1和f2(f2>f1)，则期望时钟频率为fref＝f2-f1。例如：当f1＝150MHz，f2＝160MHz时，则fref＝10MHz。发送端发送的参考信号为：
Stx(t)＝A1cos(2πf1t)+A2cos(2πf2t)   (7)
其中，A1与A2表示不同时钟频率对应的振幅值，t表示传输参考信号所耗费的时间。
相应地，接收端接收到的信号为：
其中，B1与B2表示参考信号通过信道，在接收端得到的对应于不同时钟频率的振幅值。与分别表示参考信号通过信道，在接收端得到的对应于不同时钟频率的相位值。
将上述接收端接收到的信号进行平方处理后可得：
其中，分别为时钟频率为f2+f1，f2，f1的高频项，表示接收端不同时钟频率的相位差，当f2-f1的差值较大时，通过滤波器筛选掉高频部分，从而得到的用于同步的时钟频率的参考信号为：
2.计算图层面(computation graph)
本申请实施例提供的一种信道通信运算也可通过如图16所示计算图层面进行表示，计算图是用来描述计算的语言，是一种将计算形式化的方法。在计算图中，计算可被表示成有向图，图中的每一个节点或模块表示一个变量，其中，该变量可以是标量、向量、矩阵、张量等。如图16所示，该计算图可包括骨干网络，卷积层，池化层以及全连接层。其中，卷积层的作用是对输入的数据进行特征提取，其内部包含多个卷积核，组成卷积核的每个元素对应一个权重系数和一个偏置量。池化层的作用是对卷积层中的特征进行压缩并提取主要特征，其中，常见的池化操作有最大池化和平均池化。全连接层用于连接所有特征值，并将输出值发送至分类器。全连接层一般位于池化层后，将池化层的所有特征矩阵转化成一维的特征大向量，同时，全连接层一般也位于卷积神经网络结构的最后层。
示例性的，神经网络模型在图16所示的计算图层面可表示为经过骨干网络输出的数 据依次通过卷积层、池化层和全连接层后输出的神经网络模型，其中，对于图16中所示的神经网络模型m1，该神经网络模型m1的卷积层模块包括卷积核尺寸为1*1的卷积层及卷积核尺寸为3*3的卷积层。同理，神经网络模型m2可表示为骨干网络输出的数据依次通过卷积层、池化层以及全连接层后输出的神经网络模型的数据，其中，该卷积层包括尺寸为5*5的卷积核。在此情况下，经过信道通信运算组合得到的神经网络模型m3可用计算图表示为如图16所示。例如，对于神经网络模型m3，通过骨干网络输出的数据(例如，该数据可以为一张200*200像素的图片)依次经过包括卷积核尺寸为1*1的卷积层及包括卷积核尺寸为3*3卷积层后输出的数据与通过骨干网络输出的数据经过卷积核尺寸为5*5的卷积层后得到的数据在同一时刻到达池化层，并在该池化层进行叠加处理。
在计算图层面，可将神经元节点视作特定的网络模块，通过通信过程以实现这些网络模块之间的级联，并可经过切换通信路径来达到任务需求。示例性的，该计算图中的骨干网络可位于基站中，其余模块(包括卷积层、池化层与全连接层)的功能可由一个终端设备执行或多个终端设备分别执行，本申请实施例对此不做限定。如图16所示，通过采用上述计算图层面的方式进行神经网络模型的联合传输，即可通过通信路由实现对数据的不同特征的提取结果，从而提高神经网络的表达能力和性能。
在一种可能的实现方式中，在该计算图层面，可以将神经网络模型联合传输的过程看作为多播路由算法的逆向过程，如图17所示，为计算图层面下的神经网络模型传输的一例示意图。具体地，对于该神经网络模型的联合传输过程，可在接收端建立一个多播路由，经过分流点后将数据传输至多个发送端，该分流点可作为神经网络需要通过通信运算的节点，并在该分流点处进行神经网络模型的叠加组合，以得到最终的目标神经网络模型。
在一种可能的实现方式中，经过训练后得到的N个第一模块根据N个第一模型的信息和该N个第一模型的融合辅助信息处理后得到N个第一信号，发送端设备将处理后的N个第一信号发送至接收端设备，该N个第一信号在信道中进行叠加以获得第二信号，该第二信号与符合任务需求的目标模型对应。
S1460，接收端接收目标神经网络模型。
具体地，接收端设备接收第二信号。
在一种实现方式中，接收端设备可通过接收端的译码模块对该第二信号进行译码后，并进一步经过解析处理后得到目标神经网络模型。其中，该第二信号是由发送端设备发送的N个第一信号在信道中叠加得到的。
通过上述模型传输的方式，一个或者多个发送端设备可根据接收端设备发送的任务需求信息来确定多个神经网络模型，并通过联合传输的方式将组合后的神经网络模型发送至接收端设备，相较于传统的相互独立的模型传输方式，能够显著减少传输过程中的资源开销与时延，并有效降低接收端设备对目标神经网络模型预留的存储空间。
图18为本申请实施例提供的另一例模型传输方法的流程示意图。该方法包括步骤S1810至步骤S1860。下面对步骤S1810至S1860进行详细的说明。与图14所示的模型传输的方法不同的是，该图18中所示的方法的步骤S1810至S1830和步骤S1850可由为接收端设备提供服务的服务器(如，云端设备)执行。应理解，步骤S1810至S1860与上述图14中的步骤S1410至S1460相似，此处为了避免赘述，省略其详细说明。
根据上述模型传输的方式，可通过引入云端设备来根据接收端设备发送的任务需求信 息确定多个神经网络模型，且该云端设备通过联合传输的方式将组合后的神经网络模型发送至接收端设备，相较于传统的相互独立的模型传输方式，能够显著减少传输过程中的资源开销与时延，有利于控制AI应用的多样性和有效性，并有效降低接收端设备对目标神经网络模型预留的存储空间。
图19是本申请实施例提供的模型传输的装置的示意性框图。图19所示的模型传输的装置包括获取单元1910、处理单元1920和发送单元1930。
获取单元1910、处理单元1920和发送单元1930可以用于执行本申请实施例的模型传输的方法，例如，可以用于执行上述图12中的方法1200。
获取单元1910，用于接收N个第一模型的信息和融合辅助信息，该融合辅助信息包括目标模型的外部特征值，该目标模型对应第二任务。
处理单元1920，根据该N个第一模型的信息和该融合辅助信息确定N个第一信号，并发送该N个第一信号，其中，该N个第一模型对应N个第一任务，N为大于或等于2的整数，该第二任务与该第一任务不同。
发送单元1930，用于向接收端设备发送该N个第一信号。
应理解，在本申请实施例中，该获取单元1910还可以是接收单元、或者处理单元、或者是一个具备接收和处理功能的单元。
在一种可能的实现方式中，当该获取单元1910为接收单元时，发送端设备可接收云端设备或者接收端设备发送的第一模型的信息；当该获取单元1910为处理单元时，发送端设备可用于在本地自行确定该第一模型的信息；当获取单元1910为具备接收和处理功能的单元时，发送端设备即可接收云端设备或者接收端设备发送的第一模型的信息，也可在本地自行确定该第一模型的信息。
具体地，处理单元1920通过第一模块确定该N个第一信号。
可选地，作为一个实施例，该第一模块是基于发送N个第一信号的信道的参数确定的。
上述信道的参数包括以下中的一个或多个：信道状态信息CSI、信噪比SNR、链路质量。其中，信道状态信息可包括：信道质量指示CQI、秩指示RI、预编码矩阵PMI。
示例性地，发送端设备可根据信道参数中的信道状态信息和信噪比来获得神经网络模型的配置信息，并进一步确定与上述N个第一模型一一对应的N个第一模块。
具体地，本申请实施例提供的模型传输的装置还包括第一模块，该第一模块用于根据所述N个第一模型和该融合辅助信息确定N个第一信号。
可选地，作为一个实施例，该N个第一模型中的第i个第一模型对应的第一模块是根据该N个第一模型中除第i个第一模型以外的至少一个第一模型的信息确定的，i∈[1，N]。
可选地，作为一个实施例，该N个第一模块是根据训练数据训练得到的，该训练数据包括M个第一训练信号和目标模型，该M个第一训练信号与M个训练模型一一对应，目标模型能够满足该M个训练模型对应的任务，M为大于或等于2的整数。
可选地，作为一个实施例，该处理单元1920还用于获取M个原始模型，并将每个第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将该M个第二训练信号在同一信道中叠加以获得第三训练信号，并对第三训练信号进行解析以获得中间模型，并根据中间模型与目标模型的偏差，对M个原始模型的参数进行调节处理，以获得模型 库，从模型库中获取第一模块，M个原始模型与M个第一训练信号一一对应，M为大于或等于2的整数，该模型库包括经过调节处理后的原始模型，并且经过调节处理后的中间模型与目标模型的偏差在预设范围内。
可选地，作为一个实施例，发送单元1930还用于向接收端设备发送N个第一模型的融合辅助信息。
需要说明的是，上述装置以功能单元的形式体现。这里的术语“单元”可以通过软件和/或硬件形式实现，对此不作具体限定。
例如，“单元”可以是实现上述功能的软件程序、硬件电路或二者结合。所述硬件电路可能包括应用特有集成电路(application specific integrated circuit，ASIC)、电子电路、用于执行一个或多个软件或固件程序的处理器(例如共享处理器、专有处理器或组处理器等)和存储器、合并逻辑电路和/或其它支持所描述的功能的合适组件。
因此，在本申请的实施例中描述的各示例的单元，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
图20是本申请实施例提供的神经网络模型的训练装置的硬件结构示意图。图20所示的神经网络模型的训练装置2000(该装置2000具体可以是一种计算机设备)包括存储器2010、处理器2020、通信接口2030以及总线2040。其中，存储器2010、处理器2020、通信接口2030通过总线2040实现彼此之间的通信连接。
存储器2010可以是只读存储器(read only memory，ROM)，静态存储设备，动态存储设备或者随机存取存储器(random access memory，RAM)。存储器2010可以存储程序，当存储器2010中存储的程序被处理器2020执行时，处理器2020用于执行本申请实施例的神经网络模型的训练方法的各个步骤。具体地，处理器2020可以执行上文中图12所示的方法中1200。
处理器2020可以采用通用的中央处理器(central processing unit，CPU)，微处理器，应用专用集成电路(application specific integrated circuit，ASIC)，图形处理器(graphics processing unit，GPU)或者一个或多个集成电路，用于执行相关程序，以实现本申请方法实施例的神经网络模型的训练方法。
处理器2020还可以是一种集成电路芯片，具有信号的处理能力，例如，可以是图9所示的芯片。在实现过程中，本申请的神经网络模型的训练方法的各个步骤可以通过处理器2020中的硬件的集成逻辑电路或者软件形式的指令完成。
上述处理器2020还可以是通用处理器、数字信号处理器(digital signal processing，DSP)、专用集成电路(ASIC)、现成可编程门阵列(field programmable gate array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。可以实现或者执行本申请实施例中的公开的各方法、步骤及逻辑框图。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。结合本申请实施例所涉及的方法的步骤可以直接体现为硬件译码处理器执行完成，或者用译码处理器中的硬件及软件模块组合执行完成。软件模块可以位于随机存储器，闪存、只读存储器，可编程只读存储器或者电可擦写可编程存储器、寄存器等本领域成熟的存储介质中。该存储介质位于存储器2010，处理器2020 读取存储器2010中的信息，结合其硬件完成图19所示的训练装置中包括的单元所需执行的功能，或者，执行本申请方法实施例的图12所示的神经网络模型的传输方法。
通信接口2030使用例如但不限于收发器一类的收发装置，来实现装置2000与其他设备或通信网络之间的通信。例如，可以通过通信接口2030获取训练数据。
总线2040可包括在装置2000各个部件(例如，存储器2010、处理器2020、通信接口2030)之间传送信息的通路。
应注意，尽管上述装置2000仅示出了存储器、处理器、通信接口，但是在具体实现过程中，本领域的技术人员应当理解，装置2000还可以包括实现正常运行所必须的其他器件。同时，根据具体需要，本领域的技术人员应当理解，装置2000还可包括实现其他附加功能的硬件器件。此外，本领域的技术人员应当理解，装置2000也可仅仅包括实现本申请实施例所必须的器件，而不必包括图20中所示的全部器件。
应理解，本申请实施例中的处理器可以为中央处理单元(central processing unit，CPU)，该处理器还可以是其他通用处理器、数字信号处理器(digital signal processor，DSP)、专用集成电路(application specific integrated circuit，ASIC)、现成可编程门阵列(field programmable gate array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件等。通用处理器可以是微处理器或者该处理器也可以是任何常规的处理器等。
还应理解，本申请实施例中的存储器可以是易失性存储器或非易失性存储器，或可包括易失性和非易失性存储器两者。其中，非易失性存储器可以是只读存储器(read-only memory，ROM)、可编程只读存储器(programmable ROM，PROM)、可擦除可编程只读存储器(erasable PROM，EPROM)、电可擦除可编程只读存储器(electrically EPROM，EEPROM)或闪存。易失性存储器可以是随机存取存储器(random access memory，RAM)，其用作外部高速缓存。通过示例性但不是限制性说明，许多形式的随机存取存储器(random access memory，RAM)可用，例如静态随机存取存储器(static RAM，SRAM)、动态随机存取存储器(DRAM)、同步动态随机存取存储器(synchronous DRAM，SDRAM)、双倍数据速率同步动态随机存取存储器(double data rate SDRAM，DDR SDRAM)、增强型同步动态随机存取存储器(enhanced SDRAM，ESDRAM)、同步连接动态随机存取存储器(synchlink DRAM，SLDRAM)和直接内存总线随机存取存储器(direct rambus RAM，DRRAM)。
上述实施例，可以全部或部分地通过软件、硬件、固件或其他任意组合来实现。当使用软件实现时，上述实施例可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令或计算机程序。在计算机上加载或执行所述计算机指令或计算机程序时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以为通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集合的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质。半导体介质 可以是固态硬盘。
应理解，本文中术语“和/或”，仅仅是一种描述关联对象的关联关系，表示可以存在三种关系，例如，A和/或B，可以表示：单独存在A，同时存在A和B，单独存在B这三种情况，其中A,B可以是单数或者复数。另外，本文中字符“/”，一般表示前后关联对象是一种“或”的关系，但也可能表示的是一种“和/或”的关系，具体可参考前后文进行理解。
本申请中，“至少一个”是指一个或者多个，“多个”是指两个或两个以上。“以下至少一项(个)”或其类似表达，是指的这些项中的任意组合，包括单项(个)或复数项(个)的任意组合。例如，a,b,或c中的至少一项(个)，可以表示：a,b,c,a-b,a-c,b-c,或a-b-c，其中a,b,c可以是单个，也可以是多个。
应理解，在本申请的各种实施例中，上述各过程的序号的大小并不意味着执行顺序的先后，各过程的执行顺序应以其功能和内在逻辑确定，而不应对本申请实施例的实施过程构成任何限定。
本领域普通技术人员可以意识到，结合本文中所公开的实施例描述的各示例的单元及算法步骤，能够以电子硬件、或者计算机软件和电子硬件的结合来实现。这些功能究竟以硬件还是软件方式来执行，取决于技术方案的特定应用和设计约束条件。专业技术人员可以对每个特定的应用来使用不同方法来实现所描述的功能，但是这种实现不应认为超出本申请的范围。
所属领域的技术人员可以清楚地了解到，为描述的方便和简洁，上述描述的系统、装置和单元的具体工作过程，可以参考前述方法实施例中的对应过程，在此不再赘述。
在本申请所提供的几个实施例中，应该理解到，所揭露的系统、装置和方法，可以通过其它的方式实现。例如，以上所描述的装置实施例仅仅是示意性的，例如，所述单元的划分，仅仅为一种逻辑功能划分，实际实现时可以有另外的划分方式，例如多个单元或组件可以结合或者可以集成到另一个系统，或一些特征可以忽略，或不执行。另一点，所显示或讨论的相互之间的耦合或直接耦合或通信连接可以是通过一些接口，装置或单元的间接耦合或通信连接，可以是电性，机械或其它的形式。
所述作为分离部件说明的单元可以是或者也可以不是物理上分开的，作为单元显示的部件可以是或者也可以不是物理单元，即可以位于一个地方，或者也可以分布到多个网络单元上。可以根据实际的需要选择其中的部分或者全部单元来实现本实施例方案的目的。
另外，在本申请各个实施例中的各功能单元可以集成在一个处理单元中，也可以是各个单元单独物理存在，也可以两个或两个以上单元集成在一个单元中。
所述功能如果以软件功能单元的形式实现并作为独立的产品销售或使用时，可以存储在一个计算机可读取存储介质中。基于这样的理解，本申请的技术方案本质上或者说对现有技术做出贡献的部分或者该技术方案的部分可以以软件产品的形式体现出来，该计算机软件产品存储在一个存储介质中，包括若干指令用以使得一台计算机设备(可以是个人计算机，服务器，或者网络设备等)执行本申请各个实施例所述方法的全部或部分步骤。而前述的存储介质包括：U盘、移动硬盘、只读存储器(read-only memory，ROM)、随机存取存储器(random access memory，RAM)、磁碟或者光盘等各种可以存储程序代码的介质。
以上所述，仅为本申请的具体实施方式，但本申请的保护范围并不局限于此，任何熟悉本技术领域的技术人员在本申请揭露的技术范围内，可轻易想到变化或替换，都应涵盖在本申请的保护范围之内。因此，本申请的保护范围应以所述权利要求的保护范围为准。

Claims (40)

  1. 一种模型传输的方法，其特征在于，所述方法包括：
    获取N个第一模型的信息，所述N个第一模型对应N个第一任务，其中，N为大于或等于2的整数；
    获取融合辅助信息，所述融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同；
    根据所述N个第一模型的信息和所述融合辅助信息确定N个第一信号；
    发送所述N个第一信号。
  2. 根据权利要求1所述的方法，其特征在于，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，所述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，所述模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
  3. 根据权利要求1或2所述的方法，所述N个第一信号是第一模块对所述N个第一模型的信息和所述融合辅助信息处理后得到的。
  4. 根据权利要求3所述的方法，其特征在于，所述第一模块是根据训练数据训练得到的，所述训练数据包括M个第一训练信号和目标模型，所述M个第一训练信号与M个训练模型一一对应，所述目标模型能够满足所述M个训练模型对应的任务，M为大于或等于2的整数。
  5. 根据权利要求3所述的方法，其特征在于，所述第一模块是根据发送所述N个第一信号的信道的参数确定的。
  6. 根据权利要求3所述的方法，其特征在于，所述N个第一模型中的第i个模型对应的第一模块是根据所述N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
  7. 根据权利要求4所述的方法，其特征在于，所述方法还包括：
    获取M个原始模型，所述M个原始模型与所述M个第一训练信号一一对应，M为大于或等于2的整数；
    将每个所述第一训练信号输入所对应的原始模型，以获得M个第二训练信号；
    将所述M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，所述第三训练信号与中间模型对应；
    根据所述中间模型与所述目标模型的偏差，对所述M个原始模型的参数进行调节处理，以获得模型库，所述模型库包括经过所述调节处理后的原始模型，并且经过所述调节处理后的中间模型与所述目标模型的偏差在预设范围内；
    从所述模型库中获取所述第一模块。
  8. 根据权利要求1至7中任一项所述的方法，其特征在于，所述方法还包括：
    发送所述融合辅助信息。
  9. 根据权利要求1至8中任一项所述的方法，其特征在于，所述融合辅助信息来自所述N个第一信号的接收端装置，或者
    所述融合辅助信息来自为所述N个第一信号的接收端装置提供服务的服务器。
  10. 一种模型传输的系统，其特征在于，所述模型传输系统包括：
    发送端装置，用于获取N个第一模型的信息和融合辅助信息，根据所述N个第一模型的信息和所述融合辅助信息确定N个第一信号，并发送所述N个第一信号，其中，所述N个第一模型对应N个第一任务，N为大于或等于2的整数，所述融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同；
    接收端装置，用于接收第二信号，并对接收到的所述第二信号进行解析，以获取目标模型，其中，所述第二信号是由所述N个第一信号得到的。
  11. 根据权利要求10所述的系统，其特征在于，所述第二信号是由所述N个第一信号叠加后得到的。
  12. 根据权利要求10或11所述的系统，其特征在于，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，所述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，所述模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
  13. 根据权利要求10或11所述的系统，其特征在于，所述N个第一信号是第一模块对所述N个第一模型的信息和所述融合辅助信息处理后得到的。
  14. 根据权利要求13所述的系统，其特征在于，所述第一模块是根据训练数据训练得到的，所述训练数据包括M个第一训练信号和目标模型，所述M个第一训练信号与M个训练模型一一对应，所述目标模型能够满足所述M个训练模型对应的任务，M为大于或等于2的整数。
  15. 根据权利要求13所述的系统，其特征在于，所述第一模块是基于发送所述N个第一信号的信道的参数确定的。
  16. 根据权利要求13所述的系统，其特征在于，所述N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
  17. 根据权利要求14所述的系统，其特征在于，所述发送端装置还用于获取M个原始模型，并将每个所述第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将所述M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，所述第三训练信号与中间模型对应，并根据所述中间模型与所述目标模型的偏差，对所述M个原始模型的参数进行调节处理，以获得模型库，从所述模型库中获取所述第一模块，所述M个原始模型与所述M个第一训练信号一一对应，M为大于或等于2的整数，所述模型库包括经过所述调节处理后的原始模型，并且经过所述调节处理后的中间模型与所述目标模型的偏差在预设范围内。
  18. 根据权利要求10至17中任一项所述的系统，其特征在于，所述发送端装置还用于向所述接收端装置发送所述融合辅助信息。
  19. 根据权利要求10至18中任一项所述的系统，其特征在于，所述融合辅助信息来自所述接收端装置，或者
    所述融合辅助信息来自为所述接收端装置提供服务的服务器。
  20. 一种模型传输的系统，其特征在于，所述模型传输系统包括：
    第一发送端装置，用于获取N个第一模型中的S个第一模型的信息和第一融合辅助信息，根据所述N个第一模型中的S个第一模型的信息和所述第一融合辅助信息确定S个第一信号，并发送所述S个第一信号，其中，所述N个第一模型中的S个模型对应S个第一任务，S为大于或等于1的整数，所述第一融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同；
    第二发送端装置，用于获取N个第一模型中的K个第一模型的信息和第二融合辅助信息，根据所述N个第一模型中的K个第一模型的信息和所述第二融合辅助信息确定K个第一信号，并发送所述K个第一信号，其中，所述N个第一模型中的K个模型对应K个第一任务，K为大于或等于1的整数，所述第二融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务，所述第二任务与所述N个第一任务不同，所述S个第一模型与所述K个第一模型的交集为空集；
    接收端装置，用于接收第二信号，并对所述接收到的所述第二信号进行解析，以获取目标模型，其中，所述第二信号是由所述S个第一信号和所述K个第一信号得到的。
  21. 根据权利要求20所述的系统，其特征在于，所述第二信号是由所述S个第一信号和所述K个第一信号叠加后得到的。
  22. 根据权利要求20或21所述的系统，其特征在于，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，所述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，所述模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
  23. 根据权利要求20或21所述的系统，其特征在于，所述S个第一信号是第一发送端装置通过对应的第一模块对所述S个第一模型中的S个第一模型的信息和所述第一融合辅助信息处理后得到的，所述K个第一信号是第二发送端装置通过对应的第一模块对所述K个第一模型中的K个第一模型的信息和所述第二融合辅助信息处理后得到的。
  24. 根据权利要求23所述的系统，其特征在于，所述第一发送端装置对应的第一模块或所述第二发送端装置对应的第一模块，是根据训练数据训练得到的，所述训练数据包括M个第一训练信号和目标模型，所述M个第一训练信号与M个训练模型一一对应，所述目标模型能够满足所述M个训练模型对应的任务，M为大于或等于2的整数。
  25. 根据权利要求23所述的系统，其特征在于，所述第一发送端装置对应的第一模块或所述第二发送端装置对应的第一模块是基于发送对应第一信号的信道的参数确定的。
  26. 根据权利要求23所述的系统，其特征在于，所述S个第一模型中的第i个模型对应的第一模块是根据S个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的；所述K个第一模型中的第j个模型对应的第一模块是根据K个第一模型中除第一模型中的第j个模型以外的至少一个第一模型的信息确定的。
  27. 根据权利要求24所述的系统，其特征在于，所述发送端装置还用于获取M个原始模型，并将每个所述第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将所述M个第二训练信号在同一信道中叠加以获得第三训练信号，其中，所述第三训练信号与中间模型对应，并根据所述中间模型与所述目标模型的偏差，对所述M个原始模型的参数进行调节处理，以获得模型库，从所述模型库中获取所述第一模块，所述M 个原始模型与所述M个第一训练信号一一对应，M为大于或等于2的整数，所述模型库包括经过所述调节处理后的原始模型，并且经过所述调节处理后的中间模型与所述目标模型的偏差在预设范围内。
  28. 一种模型传输的装置，其特征在于，包括：
    获取单元，用于获取融合辅助信息和N个第一模型的信息，所述融合辅助信息包括目标模型的外部特征值，所述目标模型对应第二任务；
    处理单元，用于根据所述N个第一模型的信息和所述融合辅助信息确定N个第一信号，并发送所述N个第一信号，其中，所述N个第一模型对应N个第一任务，N为大于等于2的整数，所述第二任务与所述第一任务不同；
    发送单元，用于向接收端设备发送所述N个第一信号。
  29. 根据权利要求28所述的装置，其特征在于，第一模型的信息或目标模型的信息包括模型的外部特征值和/或模型的参数，其中，所述模型的外部特征值包括以下信息中的一个或者多个：神经网络计算图信息、优化器信息、超参数信息，所述模型的参数包括以下参数中的一个或者多个：权重矩阵、权重向量、偏置矩阵、偏置向量。
  30. 根据权利要求28或29所述的装置，其特征在于，所述处理单元通过第一模块确定所述N个第一信号。
  31. 根据权利要求30所述的装置，其特征在于，所述第一模块是根据训练数据训练得到的，所述训练数据包括M个第一训练信号和目标模型，所述M个第一训练信号与M个训练模型一一对应，所述目标模型能够满足所述M个训练模型对应的任务，M为大于或等于2的整数。
  32. 根据权利要求30所述的装置，其特征在于，所述第一模块是基于发送所述N个第一信号的信道的参数确定的。
  33. 根据权利要求30所述的装置，其特征在于，所述N个第一模型中的第i个模型对应的第一模块是根据N个第一模型中除第一模型中的第i个模型以外的至少一个第一模型的信息确定的。
  34. 根据权利要求31所述的装置，其特征在于，所述处理单元还用于获取M个原始模型，并将每个所述第一训练信号输入所对应的原始模型，以获得M个第二训练信号，并将所述M个第二训练信号在同一信道中叠加以获得第三训练信号，并对所述第三训练信号进行解析以获得中间模型，并根据所述中间模型与所述目标模型的偏差，对所述M个原始模型的参数进行调节处理，以获得模型库，从所述模型库中获取所述第一模块，所述M个原始模型与所述M个第一训练信号一一对应，M为大于或等于2的整数，所述模型库包括经过所述调节处理后的原始模型，并且经过所述调节处理后的中间模型与所述目标模型的偏差在预设范围内。
  35. 根据权利要求28至34中任一项所述的装置，其特征在于，所述发送单元还用于发送所述融合辅助信息。
  36. 根据权利要求28至35中任一项所述的装置，其特征在于，所述融合辅助信息来自所述接收端装置，或者
    所述融合辅助信息来自为所述接收端装置提供服务的服务器。
  37. 一种模型传输的装置，其特征在于，包括处理器和存储器，所述存储器用于存储 程序指令，所述处理器用于调用所述程序指令来执行权利要求1至9中任一项所述的方法。
  38. 一种计算机可读存储介质，其特征在于，所述计算机可读存储用于设备执行的程序代码，该程序代码用于实现如权利要求1至9中任一项所述的方法。
  39. 一种芯片，其特征在于，所述芯片包括处理器与数据接口，所述处理器通过所述数据接口读取存储器上存储的指令，以执行如权利要求1至9中任一项所述的方法。
  40. 一种计算机程序产品，其特征在于，所述计算机程序产品包括用于设备执行的程序代码，该程序代码用于实现如权利要求1至9中任一项所述的方法。
PCT/CN2023/089764 2022-05-31 2023-04-21 一种模型传输的方法及装置 WO2023231635A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210609813.5 2022-05-31
CN202210609813.5A CN117220734A (zh) 2022-05-31 2022-05-31 一种模型传输的方法及装置

Publications (1)

Publication Number Publication Date
WO2023231635A1 true WO2023231635A1 (zh) 2023-12-07

Family

ID=89026883

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/089764 WO2023231635A1 (zh) 2022-05-31 2023-04-21 一种模型传输的方法及装置

Country Status (2)

Country Link
CN (1) CN117220734A (zh)
WO (1) WO2023231635A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117650834B (zh) * 2024-01-29 2024-04-12 南京理工大学 基于知识蒸馏的天地一体化网络时空流量预测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190050714A1 (en) * 2017-08-09 2019-02-14 Ants Technology (Hk) Limited Modular distributed artificial neural networks
CN110766151A (zh) * 2019-10-29 2020-02-07 中国人民解放军国防科技大学 一种基于场景的开放式神经网络模型管理系统
WO2022041015A1 (zh) * 2020-08-26 2022-03-03 华为技术有限公司 神经网络模型优化方法及装置
US20220086057A1 (en) * 2020-09-11 2022-03-17 Qualcomm Incorporated Transmission of known data for cooperative training of artificial neural networks
US20220108687A1 (en) * 2020-10-07 2022-04-07 General Dynamics Mission Systems, Inc. Signal processor employing neural network trained using evolutionary feature selection

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190050714A1 (en) * 2017-08-09 2019-02-14 Ants Technology (Hk) Limited Modular distributed artificial neural networks
CN110766151A (zh) * 2019-10-29 2020-02-07 中国人民解放军国防科技大学 一种基于场景的开放式神经网络模型管理系统
WO2022041015A1 (zh) * 2020-08-26 2022-03-03 华为技术有限公司 神经网络模型优化方法及装置
US20220086057A1 (en) * 2020-09-11 2022-03-17 Qualcomm Incorporated Transmission of known data for cooperative training of artificial neural networks
US20220108687A1 (en) * 2020-10-07 2022-04-07 General Dynamics Mission Systems, Inc. Signal processor employing neural network trained using evolutionary feature selection

Also Published As

Publication number Publication date
CN117220734A (zh) 2023-12-12

Similar Documents

Publication Publication Date Title
Yang et al. Edge intelligence for autonomous driving in 6G wireless system: Design challenges and solutions
WO2020221200A1 (zh) 神经网络的构建方法、图像处理方法及装置
WO2022105714A1 (zh) 数据处理方法、机器学习的训练方法及相关装置、设备
WO2022042713A1 (zh) 一种用于计算设备的深度学习训练方法和装置
Zhao et al. Orbital collaborative learning in 6G space-air-ground integrated networks
CN113326930B (zh) 数据处理方法、神经网络的训练方法及相关装置、设备
CN108710906B (zh) 基于轻量级网络LightPointNet的实时点云模型分类方法
CN111797983A (zh) 一种神经网络构建方法以及装置
WO2023279975A1 (zh) 一种模型处理方法、联邦学习方法及相关设备
WO2023231635A1 (zh) 一种模型传输的方法及装置
CN110794965B (zh) 一种基于深度强化学习的虚拟现实语言任务卸载方法
WO2022012668A1 (zh) 一种训练集处理方法和装置
WO2023125628A1 (zh) 神经网络模型优化方法、装置及计算设备
CN113469367A (zh) 一种联邦学习方法、装置及系统
CN112115830B (zh) 一种基于比特域特征提取的目标分布式融合识别方法
Zhang et al. Deep learning techniques for advancing 6G communications in the physical layer
CN111931901A (zh) 一种神经网络构建方法以及装置
CN115018039A (zh) 一种神经网络蒸馏方法、目标检测方法以及装置
CN113516227A (zh) 一种基于联邦学习的神经网络训练方法及设备
CN113536970A (zh) 一种视频分类模型的训练方法及相关装置
CN116229226A (zh) 一种适用于光电吊舱的双通道图像融合目标检测方法
US20230422054A1 (en) Method for performing federated learning in wireless communication system, and apparatus therefor
Cui et al. Sensing-assisted communication beamforming based on multi-modal feature extraction for high-reliable IoV
WO2024109910A1 (zh) 一种生成模型训练方法、数据转换方法以及装置
Li et al. Design of Fruit-Carrying Monitoring System for Monorail Transporter in Mountain Orchard

Legal Events

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

Ref document number: 23814821

Country of ref document: EP

Kind code of ref document: A1