WO2023138233A1 - 模型传输方法、装置、电子设备及可读存储介质 - Google Patents

模型传输方法、装置、电子设备及可读存储介质 Download PDF

Info

Publication number
WO2023138233A1
WO2023138233A1 PCT/CN2022/136415 CN2022136415W WO2023138233A1 WO 2023138233 A1 WO2023138233 A1 WO 2023138233A1 CN 2022136415 W CN2022136415 W CN 2022136415W WO 2023138233 A1 WO2023138233 A1 WO 2023138233A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
transmission
slices
slice
path
Prior art date
Application number
PCT/CN2022/136415
Other languages
English (en)
French (fr)
Inventor
王碧舳
刘慧焘
许晓东
董辰
韩书君
Original Assignee
北京邮电大学
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 北京邮电大学 filed Critical 北京邮电大学
Publication of WO2023138233A1 publication Critical patent/WO2023138233A1/zh

Links

Images

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • H04L41/145—Network analysis or design involving simulating, designing, planning or modelling of a network
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/14—Network analysis or design
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
    • Y02D30/00—Reducing energy consumption in communication networks
    • Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks

Definitions

  • the present disclosure relates to the technical field of communication, and in particular to a model transmission method, device, electronic equipment and readable storage medium.
  • network nodes tend to be intelligent.
  • the intelligentization of network nodes has led to rapid expansion of information space, and even dimensional disasters, which has exacerbated the difficulty of representing information carrying space, making it difficult to match traditional network service capabilities with high-dimensional information space.
  • the amount of data transmitted through communication is too large, and the information business service system cannot continue to meet people's needs for complex, diverse, and intelligent information transmission.
  • Using artificial intelligence models to encode, disseminate, and decode business information can significantly reduce the amount of data transmission in communication services and greatly improve the efficiency of information transmission.
  • These models are relatively stable, and have reusability and dissemination. The dissemination and reuse of models will help to enhance network intelligence while reducing overhead and resource waste, forming an intelligent network with extremely intelligent nodes and a minimal network.
  • the core of Intent-Driven Network lies in the transmission model.
  • the network has a storage function, and the model is stored in the network, which may be stored on the end user side or in the cloud.
  • Each node can absorb the models of other nodes on the network to realize self-evolution. Therefore, the efficiency of the propagation model directly determines the efficiency of communication.
  • transmission through the same path will lead to a large delay in model transmission, resulting in a decrease in communication efficiency. Therefore, an efficient model transmission method and device are urgently needed.
  • the present disclosure provides a method, device, device and storage medium for transmitting models in an Intent-Driven Network.
  • a model transmission method including:
  • n is a positive integer greater than 1;
  • the model to be transmitted in the sending end node is divided into n groups of model slices in proportion, and a path selection strategy corresponding to the n groups of model slices and the n transmission paths is formed;
  • the divided n groups of model slices are respectively transmitted to the receiving end node through the corresponding transmission paths.
  • the dividing the model to be transmitted in the sending end node into n groups of model slices in proportion includes:
  • each set of model slices includes one or more model slices.
  • model transfer method also includes:
  • the current transmission path is used for transmission, and the current model slice is stored in the corresponding routing node during transmission.
  • the transmission capability includes channel capacity of the transmission path.
  • a model transmission device comprising:
  • the path obtaining module is configured to obtain n transmission paths between the sending end node and the receiving end node, where n is a positive integer greater than 1;
  • the model segmentation module is configured to divide the model to be transmitted in the sending end node into n groups of model slices in proportion according to the transmission capabilities of each of the transmission paths, and form a path selection strategy that corresponds one-to-one with the n groups of model slices and the n transmission paths;
  • the model slice transmission module is configured to transmit the segmented n groups of model slices to the receiving end node through corresponding transmission paths.
  • the process of the model splitting module splitting the model to be transmitted in the sending end node into n groups of model slices in proportion includes:
  • each group of the model slices segmented by the model segmentation module includes one or more of the model slices.
  • the path selection module is configured to determine whether the model slice has been stored in any routing node of the current transmission path before transmitting each model slice:
  • the model slice transmission module switches to the next transmission path for transmission, and stores the current model slice in the corresponding routing node during transmission;
  • the model slice transmission module uses the current transmission path for transmission, and stores the current model slice in the corresponding routing node during transmission.
  • the transmission capability includes channel capacity of the transmission path.
  • the present disclosure also provides an electronic device, comprising:
  • the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor, so that the at least one processor can execute the model transmission method described in any one of the above technical solutions.
  • the present disclosure also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to make the computer execute the model transmission method according to any one of the above-mentioned embodiments.
  • the present disclosure also provides a computer program product, including a computer program, when the computer program is executed by a processor, the model transmission method according to any one of the above-mentioned embodiments is implemented.
  • the present disclosure can significantly reduce the transmission delay and improve the efficiency of model transmission by dividing the model and transmitting it through different transmission paths.
  • Fig. 1 is a flow chart of the steps of the model transmission method in the embodiment of the present disclosure
  • FIG. 2 is a schematic diagram of the first model slice transmission in an embodiment of the present disclosure
  • FIG. 3 is a schematic diagram of a second model slice transmission in an embodiment of the present disclosure.
  • FIG. 4 is a schematic diagram of a second model slice transmission in an embodiment of the present disclosure.
  • FIG. 5 is a schematic diagram of a second model slice transmission in an embodiment of the present disclosure.
  • Fig. 6 is a functional block diagram of the first model transmission device in an embodiment of the present disclosure.
  • FIG. 7 is a functional block diagram of a second model transmission device in an embodiment of the present disclosure.
  • FIG. 8 is a flow chart of model slice propagation in the model transfer method in an embodiment of the present disclosure.
  • the present disclosure discloses that business information is propagated mainly through an artificial intelligence model in an intelligent network.
  • the first business information to be propagated is compressed into the second business information related to the artificial intelligence model, which greatly reduces the data traffic in the network, and the compression efficiency far exceeds the traditional compression algorithm.
  • the sending-end device extracts the first service information by using a pre-configured first model to obtain the second service information to be transmitted; the sending-end device transmits the second service information to the receiving-end device.
  • the receiving end device receives the second service information, and uses the pre-configured second model to restore the second service information to obtain the third service information; the third service information restored by the second model may have a slight difference in quality compared with the original first service information, but the two are consistent in content, and the user experience is almost the same.
  • the transmitting end device transmits the second service information to the receiving end device, it further includes: an updating module judging whether the receiving end device needs to update the second model, and transmitting a pre-configured third model to the receiving end device when it is judged that an update is required, and the receiving end device uses the third model to update the second model. Processing business information through pre-trained artificial intelligence models can significantly reduce the amount of data transmission in communication services and greatly improve the efficiency of information transmission.
  • Model propagation and reuse will help enhance network intelligence while reducing overhead and resource waste.
  • the model can be divided into several model slices according to different segmentation rules, the above model slices can also be transmitted between different network nodes, and the model slices can be assembled into a model.
  • Model slices can be distributed and stored on multiple network nodes. When a network node finds that it lacks or needs to update a certain model or a certain model slice, it can make a request to the surrounding nodes that may have the slice.
  • Both the transmission of the business information and the transmission of the model take place in the communication network, and the communication transmission is performed based on a network protocol.
  • the network nodes passed on the path for transmitting the service information and the model include an intelligent simplified router.
  • the functions of the I-Driven Router include but are not limited to business information transmission, model transmission, absorbing model self-update, security protection and other functions.
  • the transmission function of the Intelligent-Driven Router involves the transmission of business information or models from the source node to the sink node, and there are multiple paths between the source node and the sink node.
  • the model transmission function of the Smart-Driven Router can transmit model slices. By rationally arranging model slices to take multiple paths, multiple transmission model slices can be used to improve the model transmission rate.
  • the present disclosure also discloses a model transmission method, as shown in FIG. 1 , including:
  • Step S101 obtaining n transmission paths between the sending node and the receiving node, where n is a positive integer greater than 1;
  • Step S102 according to the transmission capacity of each transmission path, the model to be transmitted in the sending end node is divided into n groups of model slices in proportion, and a path selection strategy corresponding to n groups of model slices and n transmission paths is formed;
  • step S103 the divided n groups of model slices are respectively transmitted to the receiving end node through corresponding transmission paths.
  • the core of the disclosed model transmission method is to divide the model to be transmitted into multiple groups of model slices and transmit them simultaneously through different transmission paths. Compared with transmitting the entire model through a single transmission path, rationally arranging multiple transmission paths to transmit model slices at the same time can effectively reduce transmission delay and improve communication efficiency.
  • node A is a sending end node
  • node B is a receiving end node
  • the deep learning model is divided into n groups according to the amount of data, and the n groups of slices are transmitted on multiple paths to ensure that multiple model slices reach the destination at the same time.
  • the advantage of splitting the model in this way is that it can increase the transmission rate of the model in the network and improve the efficiency of network communication. For example, as shown in Figure 2, there are two transmission paths between nodes A and B, which are transmission paths L1 and L2 respectively.
  • the model to be transmitted is divided into two model slices, where slice 1 and slice 2 are transmitted through paths L1 and L2 respectively; at the same time, the size of slice 1 and slice 2 can be divided according to the transmission capabilities of the transmission paths L1 and L2.
  • the transmission capability can include the channel capacity of the transmission path, that is, the maximum transmission rate of the channel. For example, assuming that the channel capacities of paths L1 and L2 are the same or have little difference, the model can be equally divided and transmitted, and the transmission delay can be reduced to half of the original; assuming that the channel capacity of the transmission path L1 is twice that of the transmission path L2, then we can divide the model according to the ratio of 2:1. After segmentation, the size of slice 1 is 1/3 of the model, and the size of slice 2 is 2/3 of the model.
  • Slice 1 is transmitted through the transmission path L2 with a smaller channel capacity
  • slice 2 is transmitted through the transmission path L1 with a larger channel capacity. Transmission, by segmenting the model and reasonably allocating the transmission paths of the model slices, the efficiency of model transmission can be maximized.
  • Each group of model slices can be one model slice or multiple model slices.
  • the model can also be divided into 10 model slices. Assuming that there are 2 transmission paths, slices 1-5 can be transmitted through one transmission path, and slices 6-10 can be transmitted through another transmission path.
  • dividing the model to be transmitted in the sending end node into n groups of model slices in proportion includes:
  • the total transmission delay is obtained as the objective function by modeling the segmentation delay t1 of the model to be transmitted into model slices and the transmission delay t2 of each model slice on the corresponding transmission path;
  • the model segmentation ratio and path selection strategy are obtained by solving the minimized objective function, and the model to be transmitted is divided into n groups of model slices according to the model segmentation ratio.
  • the objective function min ⁇ t 1 +t 2 ⁇ can be obtained by modeling. Assume that there is a matrix N, and the data in it is to record the path number (1, 2, ..., n) taken by each slice (1, 2, ..., n). They are in a one-to-one correspondence.
  • K is the segmentation ratio, such as [0.12, 0.32, 0.24, 0.32].
  • the segmentation delay t 1 and the transmission delay t can be calculated according to K and N 2 . Specifically, it can be modeled through deep learning. The input parameters of the deep neural network are set to the segmentation ratio K and matrix N, and the output parameters are t 1 + t 2 .
  • the best K and N are found to minimize the value of the objective function min ⁇ t 1 +t 2 ⁇ , that is, the total transmission delay of the model is minimized, the transmission capacity of each transmission path is maximized, and the transmission efficiency of the model is improved by using limited communication resources.
  • the model transmission method also includes:
  • some routing nodes of the transmission path may have stored part of the model slices of the model, while other routing nodes of the transmission path have not stored the model slices.
  • some routing nodes of the transmission path may have stored part of the model slices of the model, while other routing nodes of the transmission path have not stored the model slices.
  • the model to be transmitted is divided into 10 model slices, and the sending node A has slices 1-10, while the receiving node B only has slices 1-7, so the sending node A is required to transmit its required slices 8, 9, and 10.
  • the transmission paths L1 and L2 between nodes A and B can be transmitted.
  • routing node C in the first transmission path L1 has backed up slices 1-5
  • routing node D has backed up slices 1-5 and 8, and has no backup slices 9 and 10
  • routing node E on the second transmission path L2 has backed up slices 1-4
  • routing node F has backed up slices 1-3, 6-9, and has no backup slice 8; therefore, slices 9 and 10 required by receiving node B can be transmitted through the first transmission path L1, and slice 8 Transmission via the second transmission path L2.
  • the three model slices are divided into two groups of model slices, and the transmission delay can be effectively reduced by transmitting the model slices through two transmission paths at the same time; in addition, when transmitting the model slices, the transmission path that does not store the model slices should be selected as much as possible, so that the model slices can be shared with the routing nodes that do not back up the model slices, and useful model slices can be saved in the routing nodes, which improves the sharing of the network.
  • slices 9 and 10 can be backed up on the routing nodes (C, D) of the first transmission path L1 at the same time.
  • node C has stored slices 1-5
  • node D has stored slices 1-5 and slice 8. Therefore, slices 9 and 10 can be backed up on nodes C and D.
  • the slice 8 can be backed up on the node E that does not have the slice 8, and the node F that has backed up the slice 8 does not need to back up again.
  • Figure 5 shows that after model slices 8, 9, and 10 arrive at node B, node B now has all the slices 1-10 of the model. This transmission method can not only reduce the transmission delay of the model, but also improve the sharing of the model.
  • slice 8 can be directly transmitted to node B by routing node D that has backed up slice 8
  • slice 9 can be directly transmitted to node B by routing node F that has backed up slice 9, and only slice 10 is transmitted from node A to node B, which can further improve the transmission efficiency of model slices and reduce the bandwidth occupied by transmission model slices.
  • this transmission method requires that the model has formed a certain degree of sharing in the network. Through this transmission method, the sharing of model slices can be maximized.
  • any one of the model transfer methods in the above embodiments can be applied to the field of smart medical care.
  • the medical industry will incorporate more artificial intelligence technologies.
  • medical services will become truly intelligent.
  • the Zhijian network can take advantage of its own propagation model.
  • a service node has already trained the model, when another service node has a service demand, the node does not need to waste a lot of time training the model. It can directly request a node that has trained a similar model from the network, and directly transmit the model through model slices. This not only leaves useful slices in the network nodes, but also improves the model transmission rate and saves precious time for diagnosis and treatment.
  • the present disclosure also provides a model transmission device, as shown in Figure 6, comprising:
  • the path obtaining module 601 is configured to obtain n transmission paths between the sending end node and the receiving end node, where n is a positive integer greater than 1;
  • the model segmentation module 602 is configured to divide the model to be transmitted in the sending end node into n groups of model slices in proportion according to the transmission capacity of each transmission path, and form a path selection strategy for one-to-one correspondence between the n groups of model slices and the n transmission paths;
  • the model slice transmission module 603 is configured to transmit the divided n groups of model slices to the receiving end node through corresponding transmission paths.
  • the core of the disclosed model transmission method is to divide the model to be transmitted into multiple groups of model slices and transmit them simultaneously through different transmission paths.
  • rationally arranging multiple transmission paths to transmit model slices at the same time can effectively reduce transmission delay and improve communication efficiency.
  • the path acquisition module 601 is used to obtain n transmission paths with model slice transmission capabilities between nodes A and B.
  • node A needs to send a model to node B, it does not transmit from one path, but uses the model segmentation module 602 to divide the model into n groups according to the amount of data.
  • the model slice transmission module 603 allocates n groups of model slices for transmission on multiple transmission paths, ensuring that multiple model slices arrive at the receiving end node at the same time, so the advantages of model segmentation and transmission It can improve the transmission rate of the model in the network and improve the efficiency of network communication.
  • the model to be transmitted is divided into two model slices, where slice 1 and slice 2 are transmitted through paths L1 and L2 respectively; at the same time, the size of slice 1 and slice 2 can be divided according to the transmission capabilities of the transmission paths L1 and L2.
  • the transmission capability can include the channel capacity of the transmission path, that is, the maximum transmission rate of the channel.
  • the model can be equally divided and transmitted, and the transmission delay can be reduced to half of the original; assuming that the channel capacity of the transmission path L1 is twice that of the transmission path L2, then we can divide the model according to the ratio of 2:1. After segmentation, the size of slice 1 is 1/3 of the model, and the size of slice 2 is 2/3 of the model. Slice 1 is transmitted through the transmission path L2 with a smaller channel capacity, and slice 2 is transmitted through the transmission path L1 with a larger channel capacity. Transmission, by segmenting the model and reasonably allocating the transmission paths of the model slices, the efficiency of model transmission can be maximized.
  • Each group of model slices can be one model slice or multiple model slices.
  • the model can also be divided into 10 model slices. Assuming that there are 2 transmission paths, slices 1-5 can be transmitted through one transmission path, and slices 6-10 can be transmitted through another transmission path.
  • the process of the model segmentation module 602 dividing the model to be transmitted in the sending end node into n groups of model slices in proportion includes:
  • the total transmission delay is obtained as the objective function by modeling the segmentation delay of the model to be transmitted into model slices and the transmission delay of each model slice on the corresponding transmission path;
  • the model segmentation ratio and path selection strategy are obtained by solving the minimized objective function, and the model to be transmitted is divided into n groups of model slices according to the model segmentation ratio.
  • the model segmentation module 602 in this embodiment can obtain the objective function min ⁇ t 1 +t 2 ⁇ through modeling. Assume that there is a matrix N, and the data in it is to record the path number (1, 2, ..., n) of each slice (1, 2, ..., n). 1 and transmission delay t 2 . Specifically, it can be modeled through deep learning. The input parameters of the deep neural network are set to the segmentation ratio K and matrix N, and the output parameters are t 1 + t 2 .
  • the best K and N are found to minimize the value of the objective function min ⁇ t 1 +t 2 ⁇ , that is, the total transmission delay of the model is minimized, the transmission capacity of each transmission path is maximized, and the transmission efficiency of the model is improved by using limited communication resources.
  • the model transmission device also includes:
  • the path selection module 604 is configured to determine whether the model slice has been stored in any routing node of the current transmission path before transmitting each model slice:
  • the model slice transmission module 603 switches to the next transmission path for transmission, and stores the current model slice in the corresponding routing node during the transmission process. It should be noted that in this step, some model slices have changed the transmission path, which is not completely consistent with the initially determined path selection strategy, which may lead to a mismatch in the transmission capacity of the transmission path. Therefore, each routing node needs to be synchronized to ensure that the model slices can reach the receiving end node at the same time;
  • the model slice transmission module 603 uses the current transmission path for transmission, and stores the current model slice in the corresponding routing node during transmission.
  • some routing nodes of the transmission path may have stored part of the model slices of the model, while other routing nodes of the transmission path have not stored the model slices.
  • some routing nodes of the transmission path may have stored part of the model slices of the model, while other routing nodes of the transmission path have not stored the model slices.
  • the sending node A has slices 1-10
  • the receiving node B only has slices 1-7, so the sending node A is required to transmit the required slices 8, 9, and 10.
  • the transmission paths L1 and L2 between nodes A and B can be transmitted.
  • routing node C in the first transmission path L1 has backed up slices 1-5
  • routing node D has backed up slices 1-5 and 8, and has no backup slices 9 and 10
  • routing node E on the second transmission path L2 has backed up slices 1-4
  • routing node F has backed up slices 1-3, 6-9, and has no backup slice 8; therefore, slices 9 and 10 required by receiving node B can be transmitted through the first transmission path L1, and slice 8 Transmission via the second transmission path L2.
  • the three model slices are divided into two groups of model slices, and the transmission delay can be effectively reduced by transmitting the model slices through two transmission paths at the same time; in addition, the path selection module 604 selects the transmission path that does not store the model slices for transmission as much as possible, so that the model slices can be shared with the routing nodes that do not back up the model slices, and useful model slices are saved in the routing nodes, which improves the sharing of the network.
  • slices 9 and 10 can be backed up on the routing nodes (C, D) of the first transmission path L1 at the same time.
  • node C has stored slices 1-5
  • node D has stored slices 1-5 and slice 8. Therefore, slices 9 and 10 can be backed up on nodes C and D.
  • the slice 8 can be backed up on the node E that does not have the slice 8, and the node F that has backed up the slice 8 does not need to back up again.
  • Figure 5 shows that after model slices 8, 9, and 10 arrive at node B, node B now has all the slices 1-10 of the model. This transmission method can not only reduce the transmission delay of the model, but also improve the sharing of the model.
  • slice 8 can be directly transmitted to node B by routing node D that has backed up slice 8
  • slice 9 can be directly transmitted to node B by routing node F that has backed up slice 9, and only slice 10 is transmitted from node A to node B, which can further improve the transmission efficiency of model slices and reduce the bandwidth occupied by transmission model slices.
  • this transmission method requires that the model has formed a certain degree of sharing in the network. Through this transmission method, the sharing of model slices can be maximized.
  • any one of the model transmission devices in the above embodiments can be applied in the field of smart medical care.
  • the medical industry will incorporate more artificial intelligence technologies.
  • medical services will become truly intelligent.
  • the Zhijian network can take advantage of its own propagation model.
  • a service node has already trained the model, when another service node has a service demand, the node does not need to waste a lot of time training the model. It can directly request a node that has trained a similar model from the network, and directly transmit the model through model slices. This not only leaves useful slices in the network nodes, but also improves the model transmission rate and saves precious time for diagnosis and treatment.
  • the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.
  • electronic device is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers.
  • Electronic devices may also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices.
  • the components shown herein, their connections and relationships, and their functions, are by way of example only, and are not intended to limit implementations of the disclosure described and/or claimed herein.
  • the device includes a computing unit that can perform various appropriate actions and processes according to a computer program stored in a read only memory (ROM) or loaded from a storage unit into a random access memory (RAM). In RAM, various programs and data necessary for device operation are also stored.
  • the computing unit, ROM, and RAM are connected to each other through a bus. Input/output (I/O) interfaces are also connected to the bus.
  • I/O interface Multiple components in the device are connected to the I/O interface, including: input units, such as keyboards, mice, etc.; output units, such as various types of displays, speakers, etc.; storage units, such as magnetic disks, optical discs, etc.; and communication units, such as network cards, modems, wireless communication transceivers, etc.
  • the communication unit allows the device to exchange information/data with other devices over a computer network such as the Internet and/or various telecommunication networks.
  • Computing units may be various general and/or special purpose processing components having processing and computing capabilities. Some examples of computing units include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processors (DSPs), and any suitable processors, controllers, microcontrollers, etc.
  • the calculation unit executes various methods and processes described above, such as the model transmission method in the above embodiments.
  • the model transfer method may be implemented as a computer software program tangibly embodied on a machine-readable medium, such as a storage unit.
  • part or all of the computer program may be loaded and/or installed on the device via a ROM and/or a communication unit.
  • One or more steps of the model transfer method described above may be performed when the computer program is loaded into RAM and executed by the computing unit.
  • the computing unit may be configured to execute the model transfer method by any other suitable means (eg, by means of firmware).
  • Various implementations of the systems and techniques described above herein may be implemented in digital electronic circuitry, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), load programmable logic devices (CPLDs), computer hardware, firmware, software, and/or combinations thereof.
  • FPGAs field programmable gate arrays
  • ASICs application specific integrated circuits
  • ASSPs application specific standard products
  • SOCs systems on chips
  • CPLDs load programmable logic devices
  • computer hardware firmware, software, and/or combinations thereof.
  • Program codes for implementing the model transfer method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to processors or controllers of general-purpose computers, special purpose computers, or other programmable data processing devices, so that the program codes cause the functions/operations specified in the flowcharts and/or block diagrams to be implemented when executed by the processors or controllers.
  • the program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM, Random Access Memory), read-only memory (ROM, Read-Only Memory), erasable programmable read-only memory (EPROM, Electrical Programmable Read Only Memory or flash memory), optical fiber, compact disc read-only memory (CD-ROM, Compact Disc Read-Only Memory) ), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
  • RAM Random Access Memory
  • ROM read-only memory
  • EPROM Electrical Programmable Read Only Memory or flash memory
  • CD-ROM compact disc read-only memory
  • CD-ROM Compact Disc Read-Only Memory
  • magnetic storage devices or any suitable combination of the foregoing.
  • the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (Cathode Ray Tube, cathode ray tube) or LCD (Liquid Crystal Display, liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer.
  • a display device e.g., a CRT (Cathode Ray Tube, cathode ray tube) or LCD (Liquid Crystal Display, liquid crystal display) monitor
  • a keyboard and pointing device e.g., a mouse or trackball
  • Other kinds of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form, including acoustic input, voice input, or tactile input.
  • the systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or any combination of such back-end, middleware, or front-end components.
  • the components of the system can be interconnected by any form or medium of digital data communication, eg, a communication network. Examples of communication networks include: Local Area Network (LAN, Local Area Network), Wide Area Network (WAN, Wide Area Network) and the Internet.
  • a computer system may include clients and servers.
  • Clients and servers are generally remote from each other and typically interact through a communication network.
  • the relationship of client and server arises by computer programs running on the respective computers and having a client-server relationship to each other.
  • the server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.
  • steps may be reordered, added or deleted using the various forms of flow shown above.
  • each step described in the present disclosure may be executed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in the present disclosure can be achieved, no limitation is imposed herein.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Signal Processing (AREA)
  • Data Exchanges In Wide-Area Networks (AREA)

Abstract

提供一种模型传输方法、装置、电子设备及可读存储介质。具体实现方案为：获取发送端节点与接收端节点之间的n个传输路径(S101)，其中，n为大于1的正整数；根据各个传输路径的传输能力，按照比例将发送端节点中待传输的模型切分成n组模型切片，并形成将n组模型切片与n个传输路径一一对应的路径选择策略(S102)；将切分好的n组模型切片分别通过对应的传输路径向接收端节点传输(S103)。

Description

模型传输方法、装置、电子设备及可读存储介质 技术领域
本公开涉及通信技术领域，尤其涉及模型传输方法、装置、电子设备及可读存储介质。
背景技术
在未来的万物智联网络中，网络节点趋向于智能化，网络节点智能化导致了信息空间快速扩张、甚至维度灾难，加剧了表征信息承载空间的难度，导致传统的网络服务能力与高维信息空间难以匹配，通信传输的数据量过大，信息业务服务系统无法持续满足人们复杂、多样和智能化信息传输的需求。而通过人工智能模型来编码、传播、解码业务信息，可显著降低通信业务中的数据传输量，极大地提升了信息传输效率。这些模型相对稳定，并具有复用性、传播性。模型的传播和复用将有助于增强网络智能，同时降低开销和资源浪费，形成节点极智、网络极简的智简网络。
智简网络的核心在于传输模型，网络具有存储功能，模型存于网络，可能存储在终端用户侧，也可能存储在云端。各个节点可以吸收网路上其它节点的模型实现自我进化，因此，传播模型的效率直接决定了通信的效率。但由于某些模型较大，通过同一条路径传输就会导致模型传输的时延较大，导致通信效率降低，因此，现亟需一种高效的模型传输方法和装置。
发明内容
本公开提供了一种用于在智简网络中传输模型的方法、装置、设备以及存储介质。
根据本公开的一方面，提供了一种模型传输方法，包括：
获取发送端节点与接收端节点之间的n个传输路径，其中，n为大于1的正整数；
根据各个所述传输路径的传输能力，按照比例将所述发送端节点中待传输的模型切分成n组模型切片，并形成将n组所述模型切片与n个所述 传输路径一一对应的路径选择策略；
将切分好的n组所述模型切片分别通过对应的所述传输路径向所述接收端节点传输。
可选的，所述按照比例将所述发送端节点中待传输的模型切分成n组模型切片包括：
通过对所述待传输的模型切分成所述模型切片的切分时延和各个所述模型切片在对应的所述传输路径上的传输时延进行建模得到总传输时延作为目标函数；
通过求解最小化所述目标函数得到模型切分比例以及所述路径选择策略，并按所述模型切分比例将所述待传输的模型切分成n组所述模型切片。
可选的，每组所述模型切片包括一个或多个所述模型切片。
可选的，所述模型传输方法还包括：
在传输每一个所述模型切片之前，判断所述模型切片是否已经存储于当前的所述传输路径的任意一个路由节点中：
若是，则切换至下一条所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中；
若否，则使用当前的所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中。
可选的，所述传输能力包括所述传输路径的信道容量。
根据本公开的另一方面，提供了一种模型传输装置，包括：
路径获取模块，被配置为获取发送端节点与接收端节点之间的n个传输路径，其中，n为大于1的正整数；
模型切分模块，被配置为根据各个所述传输路径的传输能力，按照比例将所述发送端节点中待传输的模型切分成n组模型切片，并形成将n组所述模型切片与n个所述传输路径一一对应的路径选择策略；
模型切片传输模块，被配置为将切分好的n组所述模型切片分别通过对应的所述传输路径向所述接收端节点传输。
可选的，所述模型切分模块按照比例将所述发送端节点中待传输的模型切分成n组模型切片的过程包括：
通过对所述待传输的模型切分成所述模型切片的切分时延和各个所述模型切片在对应的所述传输路径上的传输时延进行建模得到总传输时延作为目标函数；
通过求解最小化所述目标函数得到模型切分比例以及所述路径选择策略，并按所述模型切分比例将所述待传输的模型切分成n组所述模型切片。
可选的，所述模型切分模块切分出的每组所述模型切片包括一个或多个所述模型切片。
可选的，还包括：
路径选择模块，被配置为在传输每一个所述模型切片之前，判断所述模型切片是否已经存储于当前的所述传输路径的任意一个路由节点中：
响应于所述模型切片已经存储于对应的所述传输路径中，所述模型切片传输模块切换至下一条所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中；
响应于所述模型切片未存储于对应的所述传输路径中，所述模型切片传输模块使用当前的所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中。
可选的，所述传输能力包括所述传输路径的信道容量。
本公开还提供了一种电子设备，包括：
至少一个处理器；以及
与所述至少一个处理器通信连接的存储器；其中，
所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行上述技术方案中任一项所述的模型传输方法。
本公开还提供了一种存储有计算机指令的非瞬时计算机可读存储介质，其中，所述计算机指令用于使所述计算机执行根据上述实施例中任一项所述的模型传输方法。
本公开还提供了一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现根据上述实施例中任一项所述的模型传输方法。
本公开通过上述技术方案中的模型传输方法、装置、电子设备以及存储介质，通过将模型切分后通过不同的传输路径进行传输，可显著地降低传输时延，提升模型传输的效率。
应当理解，本部分所描述的内容并非旨在标识本公开的实施例的关键或重要特征，也不用于限制本公开的范围。本公开的其它特征将通过以下的说明书而变得容易理解。
附图说明
附图用于更好地理解本方案，不构成对本公开的限定。其中：
图1是本公开实施例中的模型传输方法的步骤流程图；
图2是本公开实施例中的第一种模型切片传输示意图；
图3是本公开实施例中的第二种模型切片传输示意图；
图4是本公开实施例中的第二种模型切片传输示意图；
图5是本公开实施例中的第二种模型切片传输示意图；
图6是本公开实施例中的第一种模型传输装置的原理框图；
图7是本公开实施例中的第二种模型传输装置的原理框图；
图8是本公开实施例中的模型传输方法的模型切片传播流程图。
具体实施方式
以下结合附图对本公开的示范性实施例做出说明，其中包括本公开实施例的各种细节以助于理解，应当将它们认为仅仅是示范性的。因此，本领域普通技术人员应当认识到，可以对这里描述的实施例做出各种改变和修改，而不会背离本公开的范围和精神。同样，为了清楚和简明，以下的描述中省略了对公知功能和结构的描述。
本公开公开了智简网络中主要通过人工智能模型传播业务信息，通过使用人工智能模型将待传播的第一业务信息压缩为与所述人工智能模型相关的第二业务信息，极大地降低了网络中的数据通信量，压缩效率远超 传统的压缩算法。其中，发送端设备利用预先配置的第一模型对所述第一业务信息进行提取并得到待传输的第二业务信息；所述发送端设备向接收端设备传输所述第二业务信息。接收端设备接收所述第二业务信息，并利用预先配置的第二模型对所述第二业务信息进行恢复处理得到第三业务信息；经第二模型恢复的第三业务信息比起原先的第一业务信息会有些许质量上的差异，但两者在内容上是一致的，给用户的体验几乎是无差异的。在所述发送端设备向接收端设备传输所述第二业务信息之前，还包括：更新模块判断所述接收端设备是否需要对所述第二模型进行更新，并在判断需要更新时向所述接收端设备传输预先配置的第三模型，所述接收端设备利用所述第三模型对所述第二模型进行更新。通过预先训练好的人工智能模型对业务信息进行处理，可显著降低通信业务中的数据传输量，极大地提升了信息传输效率。这些模型相对稳定，并具有复用性、传播性。模型的传播和复用将有助于增强网络智能，同时降低开销和资源浪费。所述模型能够根据不同切分规则切分为若干个模型切片，上述模型切片也可以在不同的网络节点之间传输，模型切片可以组装成模型。模型切片可以分散存储在多个网络节点上。当网络节点请发现自己缺少或需要更新某模型或某模型切片时，可以通过请求的方式，向周围可能具有该切片的节点请求。
传输所述业务信息、传输所述模型均发生在通信网络中，基于网络协议进行通信传输。传输所述业务信息、传输所述模型的路径上经过的网络节点包括智简路由器。智简路由器的功能包括但不限于业务信息传输、模型传输，吸收模型自我更新，安全保护等功能。智简路由器的传输功能，涉及将业务信息或模型从信源节点传输到信宿节点，信源节点和信宿节点之间存在多个路径。智简路由器的模型传输功能，可以对模型切片进行传输，通过合理安排模型切片走多个路径，多路传输模型切片，提高模型传输速率。
本公开还公开了一种模型传输方法，如图1所示，包括：
步骤S101，获取发送端节点与接收端节点之间的n个传输路径，其中，n为大于1的正整数；
步骤S102，根据各个传输路径的传输能力，按照比例将发送端节点 中待传输的模型切分成n组模型切片，并形成将n组模型切片与n个传输路径一一对应的路径选择策略；
步骤S103，将切分好的n组模型切片分别通过对应的传输路径向接收端节点传输。
具体地，本公开的模型传输方法的核心在于将待传输的模型切分成多组模型切片后经过不同的传输路径同时进行传输，相比单个传输路径传输整个模型，合理地安排多个传输路径同时传输模型切片可以有效地减少传输时延，提高通信效率。
示例性地，假设节点A为发送端结点，节点B为接收端结点，节点A、B之间有n条具备模型切片传输能力的传输路径。当节点A需要向节点B发送模型时，并非是从一条路径传输，而是如图8所示将深度学习模型按照数据量切分成n组，将n组切片在多条路径上传输，保障多个模型切片同时到达信宿，这样将模型切分传输的优点是可以提高网络中模型的传输速率，提高了网络通信的效率。例如图2所示，节点A、B之间具有2条传输路径，分别是传输路径L1、L2，将待传输的模型分为2个模型切片，其中，切片1和切片2分别通过路径L1、L2进行传输；同时，切片1和切片2的大小可以根据传输路径L1、L2的传输能力来切分，传输能力可以包括传输路径的信道容量，即信道的最大传输速率。例如，假设路径L1、L2的信道容量大小相同或者差异不大时，可以将模型进行均分后传输，传输时延可以降低到原来的一半；假设传输路径L1的信道容量是传输路径L2的2倍，那么我们可以将模型按照2:1的比例来切分，切分后切片1的大小是模型的1/3，切片2的大小是模型的2/3，将切片1通过信道容量较小的传输路径L2传输，切片2则通过信道容量较大的传输路径L1传输，通过对模型进行切分，并合理地分配模型切片的传输路径，可以使得模型传输的效率最大化。
需要说明的是，上述实施例仅为一种可选的实施方式，本公开的核心在于通过对模型进行切分并分开传输来降低传输时延，模型的切分方法不限于上述比例。每一组模型切片可以是一个模型切片，也可以是多个模型切片。例如，模型也可以切分为10个模型切片，假设现有2条传输路径，可以将切片1-5通过一条传输路径传输，切片6-10通过另一条传输路径传 输。
作为一种可选的实施方式，按照比例将发送端节点中待传输的模型切分成n组模型切片包括：
通过对待传输的模型切分成模型切片的切分时延t 1和各个模型切片在对应的传输路径上的传输时延t 2进行建模得到总传输时延作为目标函数；
通过求解最小化目标函数得到模型切分比例以及路径选择策略，并按模型切分比例将待传输的模型切分成n组模型切片。
具体地，本实施例中可以通过建模得到目标函数min{t 1+t 2}，假设有一个矩阵N，里面的数据是记录每个切片(1,2,…,n)走的路径号(1,2,…,n),它们是一一对应的关系，K为切分比例，比如[0.12,0.32,0.24,0.32]，根据建模的模型可以根据K和N计算出切分时延t 1和传输时延t 2。具体可以通过深度学习建模，深度神经网络的输入参数设置为切分比例K和矩阵N,输出参数为t 1+t 2，通过多次训练找出最佳的K和N，以使目标函数min{t 1+t 2}的值最小，即模型的总传输时延最小，最大化的利用各个传输路径的传输能力，利用有限的通信资源提升模型的传输效率。
作为一种可选的实施方式，模型传输方法还包括：
在传输每一个模型切片之前，判断模型切片是否已经存储于当前的传输路径的任意一个路由节点中：
若是，则切换至下一条传输路径进行传输，并在传输过程中将当前的模型切片存储于对应的路由节点中，需要说明的是，由于该步骤中部分模型切片改变了传输路径，与初步确定的路径选择策略不完全一致，可能导致传输路径的传输能力不匹配，因此需要各个路由节点同步，保证模型切片能够同时到达接收端节点；
若否，则使用当前的传输路径进行传输，并在传输过程中将当前的模型切片存储于对应的路由节点中。
具体地，在对待传输的模型进行切分后，可能某些传输路径的路由节点已经存储了该模型的部分模型切片，而其它的传输路径的路由节点还未存储该模型切片。例如，如图3所示，假设待传输的模型被切分为10个模型切片，发送端节点A中具有切片1-10，而接端节点B仅有切片1-7， 因此要求发送端节点A传输其所需的切片8、切片9、切片10。这时节点A和B之间的传输路径L1和L2可以传输，可以图3中看到第一传输路径L1中的路由节点C已经备份了切片1-5，路由节点D已经备份了切片1-5、切片8，没有备份切片9和10；第二传输路径L2的路由节点E备份了切片1-4，路由节点F备份了切片1-3、6-9，没有备份切片8；于是可以将接收端节点B所需的切片9、切片10通过第一传输路径L1传输，切片8通过第二传输路径L2传输。本实施例中3个模型切片分为2组模型切片，通过2条传输路径同时传输模型切片，可以有效地降低传输时延；此外，在传输模型切片时，尽可能选择没有存储该模型切片的传输路径，这样就能够与没有备份该模型切片的路由节点共享模型切片，在路由节点保存有用的模型切片，提高了网络的共享性。
进一步地，如图4所示，在利用第一传输路径L1传输切片9和10时，同时可以将切片9和切片10备份在第一传输路径L1的路由节点(C、D)上，例如，节点C已存储切片1-5，节点D已存储切片1-5、切片8，因此在节点C和D上可以备份切片9和10，这样，节点C或D需要切片9和10时就不需要再次通过节点A来传输，有利于节省通信资源。同样地，在利用第二传输路径L2传输切片8时，可以将切片8备份在没有切片8的节点E上，而已备份切片8的节点F则无需再次备份。图5所示为模型切片8、9、10到达节点B之后，节点B此时具有了模型所有的切片1-10。通过该传输方法不仅可以降低模型的传输时延，还可以提高模型的共享性。
作为另外一种可选的实施方式，切片8可以直接由已经备份了切片8的路由节点D传输给节点B，切片9可以直接由已备份了切片9的路由节点F传输给节点B，仅切片10由节点A传输给节点B，这样可以进一步地提升模型切片的传输效率，减少传输模型切片所占用的带宽。但这种传输方式需要模型在网络中已经形成了一定的共享性，通过这种传输方式可以最大化利用模型切片的共享性。
示例性地，上述实施例中的任意一个模型传输方法可以应用于智慧医疗领域。随着科技的进步，医疗行业将融入更多人工智能技术。随着智慧医疗的建设，医疗服务将走向真正意义上的智能化。利用广泛的医疗数据， 训练有效的深度学习模型，根据患者的数据，可以有效分析帮助诊断患者状况，辅助进行治疗，从而使患者用较短的治疗时间和基本的治疗费用，就可以享受到优质的医疗服务。在智慧医疗场景下，智简网络可以利用自身传播模型的优势，在一个服务节点已经训练过模型的情况下，当另一个服务节点有服务需求时，该节点不需要浪费大量时间训练模型，可以直接从网络中请求已训练好类似模型的节点，直接把该模型通过模型切片方式传输过来，这样既在网络节点中留下了有用切片，又提高了模型传输速率，节约了宝贵的诊疗时间。
本公开还提供了一种模型传输装置，如图6所示，包括：
路径获取模块601，被配置为获取发送端节点与接收端节点之间的n个传输路径，其中，n为大于1的正整数；
模型切分模块602，被配置为根据各个传输路径的传输能力，按照比例将发送端节点中待传输的模型切分成n组模型切片，并形成将n组模型切片与n个传输路径一一对应的路径选择策略；
模型切片传输模块603，被配置为将切分好的n组模型切片分别通过对应的传输路径向接收端节点传输。
具体地，本公开的模型传输方法的核心在于将待传输的模型切分成多组模型切片后经过不同的传输路径同时进行传输，相比单个传输路径传输整个模型，合理地安排多个传输路径同时传输模型切片可以有效地减少传输时延，提高通信效率。例如，假设节点A为发送端结点，节点B为接收端结点，首先通过路径获取模块601获取节点A、B之间存在的n条具备模型切片传输能力的传输路径，当节点A需要向节点B发送模型时，并非是从一条路径传输，而是通过模型切分模块602将模型按照数据量切分成n组，模型切片传输模块603将n组模型切片分配在多条传输路径上传输，保障多个模型切片同时到达接收端节点，这样将模型切分传输的优点是可以提高网络中模型的传输速率，提高了网络通信的效率。例如图2所示，节点A、B之间具有2条传输路径，分别是传输路径L1、L2，将待传输的模型分为2个模型切片，其中，切片1和切片2分别通过路径L1、L2进行传输；同时，切片1和切片2的大小可以根据传输路径L1、L2的 传输能力来切分，传输能力可以包括传输路径的信道容量，即信道的最大传输速率。例如，假设路径L1、L2的信道容量大小相同或者差异不大时，可以将模型进行均分后传输，传输时延可以降低到原来的一半；假设传输路径L1的信道容量是传输路径L2的2倍，那么我们可以将模型按照2:1的比例来切分，切分后切片1的大小是模型的1/3，切片2的大小是模型的2/3，将切片1通过信道容量较小的传输路径L2传输，切片2则通过信道容量较大的传输路径L1传输，通过对模型进行切分，并合理地分配模型切片的传输路径，可以使得模型传输的效率最大化。
需要说明的是，上述实施例仅为一种可选的实施方式，本公开的核心在于通过对模型进行切分并分开传输来降低传输时延，模型的切分方法不限于上述比例。每一组模型切片可以是一个模型切片，也可以是多个模型切片。例如，模型也可以切分为10个模型切片，假设现有2条传输路径，可以将切片1-5通过一条传输路径传输，切片6-10通过另一条传输路径传输。
作为一种可选的实施方式，模型切分模块602按照比例将发送端节点中待传输的模型切分成n组模型切片的过程包括：
通过对待传输的模型切分成模型切片的切分时延和各个模型切片在对应的传输路径上的传输时延进行建模得到总传输时延作为目标函数；
通过求解最小化目标函数得到模型切分比例以及路径选择策略，并按模型切分比例将待传输的模型切分成n组模型切片。
具体地，本实施例中的模型切分模块602可以通过建模得到目标函数min{t 1+t 2}，假设有一个矩阵N，里面的数据是记录每个切片(1,2,…,n)走的路径号(1,2,…,n),它们是一一对应的关系，K为切分比例，比如[0.12,0.32,0.24,0.32]，根据建模的模型可以根据K和N计算出切分时延t 1和传输时延t 2。具体可以通过深度学习建模，深度神经网络的输入参数设置为切分比例K和矩阵N,输出参数为t 1+t 2，通过多次训练找出最佳的K和N，以使目标函数min{t 1+t 2}的值最小，即模型的总传输时延最小，最大化的利用各个传输路径的传输能力，利用有限的通信资源提升模型的传输效率。
作为一种可选的实施方式，如图7所示，模型传输装置还包括：
路径选择模块604，被配置为在传输每一个模型切片之前，判断模型切片是否已经存储于当前的传输路径的任意一个路由节点中：
响应于模型切片已经存储于对应的传输路径中，模型切片传输模块603切换至下一条传输路径进行传输，并在传输过程中将当前的模型切片存储于对应的路由节点中，需要说明的是，由于该步骤中部分模型切片改变了传输路径，与初步确定的路径选择策略不完全一致，可能导致传输路径的传输能力不匹配，因此需要各个路由节点同步，保证模型切片能够同时到达接收端节点；
响应于模型切片未存储于对应的传输路径中，模型切片传输模块603使用当前的传输路径进行传输，并在传输过程中将当前的模型切片存储于对应的路由节点中。
具体地，在对待传输的模型进行切分后，可能某些传输路径的路由节点已经存储了该模型的部分模型切片，而其它的传输路径的路由节点还未存储该模型切片。例如，如图3所示，假设待传输的模型被切分为10个模型切片，发送端节点A中具有切片1-10，而接端节点B仅有切片1-7，因此要求发送端节点A传输其所需的切片8、切片9、切片10。这时节点A和B之间的传输路径L1和L2可以传输，可以图3中看到第一传输路径L1中的路由节点C已经备份了切片1-5，路由节点D已经备份了切片1-5、切片8，没有备份切片9和10；第二传输路径L2的路由节点E备份了切片1-4，路由节点F备份了切片1-3、6-9，没有备份切片8；于是可以将接收端节点B所需的切片9、切片10通过第一传输路径L1传输，切片8通过第二传输路径L2传输。本实施例中3个模型切片分为2组模型切片，通过2条传输路径同时传输模型切片，可以有效地降低传输时延；此外，路径选择模块604尽可能选择没有存储该模型切片的传输路径进行传输，这样就能够与没有备份该模型切片的路由节点共享模型切片，在路由节点保存有用的模型切片，提高了网络的共享性。
进一步地，如图4所示，在利用第一传输路径L1传输切片9和10时，同时可以将切片9和切片10备份在第一传输路径L1的路由节点(C、D)上，例如，节点C已存储切片1-5，节点D已存储切片1-5、切片8，因此在节点C和D上可以备份切片9和10，这样，节点C或D需要切片9和 10时就不需要再次通过节点A来传输，有利于节省通信资源。同样地，在利用第二传输路径L2传输切片8时，可以将切片8备份在没有切片8的节点E上，而已备份切片8的节点F则无需再次备份。图5所示为模型切片8、9、10到达节点B之后，节点B此时具有了模型所有的切片1-10。通过该传输方法不仅可以降低模型的传输时延，还可以提高模型的共享性。
作为另外一种可选的实施方式，切片8可以直接由已经备份了切片8的路由节点D传输给节点B，切片9可以直接由已备份了切片9的路由节点F传输给节点B，仅切片10由节点A传输给节点B，这样可以进一步地提升模型切片的传输效率，减少传输模型切片所占用的带宽。但这种传输方式需要模型在网络中已经形成了一定的共享性，通过这种传输方式可以最大化利用模型切片的共享性。
示例性地，上述实施例中的任意一个模型传输装置可以应用于智慧医疗领域。随着科技的进步，医疗行业将融入更多人工智能技术。随着智慧医疗的建设，医疗服务将走向真正意义上的智能化。利用广泛的医疗数据，训练有效的深度学习模型，根据患者的数据，可以有效分析帮助诊断患者状况，辅助进行治疗，从而使患者用较短的治疗时间和基本的治疗费用，就可以享受到优质的医疗服务。在智慧医疗场景下，智简网络可以利用自身传播模型的优势，在一个服务节点已经训练过模型的情况下，当另一个服务节点有服务需求时，该节点不需要浪费大量时间训练模型，可以直接从网络中请求已训练好类似模型的节点，直接把该模型通过模型切片方式传输过来，这样既在网络节点中留下了有用切片，又提高了模型传输速率，节约了宝贵的诊疗时间。
根据本公开的实施例，本公开还提供了一种电子设备、一种可读存储介质和一种计算机程序产品。
具体地，电子设备旨在表示各种形式的数字计算机，诸如，膝上型计算机、台式计算机、工作台、个人数字助理、服务器、刀片式服务器、大型计算机、和其它适合的计算机。电子设备还可以表示各种形式的移动装置，诸如，个人数字处理、蜂窝电话、智能电话、可穿戴设备和其它类似 的计算装置。本文所示的部件、它们的连接和关系、以及它们的功能仅仅作为示例，并且不意在限制本文中描述的和/或者要求的本公开的实现。
设备包括计算单元，其可以根据存储在只读存储器(ROM)中的计算机程序或者从存储单元加载到随机访问存储器(RAM)中的计算机程序，来执行各种适当的动作和处理。在RAM中，还可存储设备操作所需的各种程序和数据。计算单元、ROM以及RAM通过总线彼此相连。输入/输出(I/O)接口也连接至总线。
设备中的多个部件连接至I/O接口，包括：输入单元，例如键盘、鼠标等；输出单元，例如各种类型的显示器、扬声器等；存储单元，例如磁盘、光盘等；以及通信单元，例如网卡、调制解调器、无线通信收发机等。通信单元允许设备通过诸如因特网的计算机网络和/或各种电信网络与其他设备交换信息/数据。
计算单元可以是各种具有处理和计算能力的通用和/或专用处理组件。计算单元的一些示例包括但不限于中央处理单元(CPU)、图形处理单元(GPU)、各种专用的人工智能(AI)计算芯片、各种运行机器学习模型算法的计算单元、数字信号处理器(DSP)、以及任何适当的处理器、控制器、微控制器等。计算单元执行上文所描述的各个方法和处理，例如上述实施例中的模型传输方法。例如，在一些实施例中，模型传输方法可被实现为计算机软件程序，其被有形地包含于机器可读介质，例如存储单元。在一些实施例中，计算机程序的部分或者全部可以经由ROM和/或通信单元而被载入和/或安装到设备上。当计算机程序加载到RAM并由计算单元执行时，可以执行上文描述的模型传输方法的一个或多个步骤。备选地，在其他实施例中，计算单元可以通过其他任何适当的方式(例如，借助于固件)而被配置为执行模型传输方法。
本文中以上描述的系统和技术的各种实施方式可以在数字电子电路系统、集成电路系统、现场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、芯片上系统的系统(SOC)、负载可编程逻辑设备(CPLD)、计算机硬件、固件、软件、和/或它们的组合中实现。这些各种实施方式可以包括：实施在一个或者多个计算机程序中，该一个或者多个计算机程序可在包括至少一个可编程处理器的可编程系统上执行和/ 或解释，该可编程处理器可以是专用或者通用可编程处理器，可以从存储系统、至少一个输入装置、和至少一个输出装置接收数据和指令，并且将数据和指令传输至该存储系统、该至少一个输入装置、和该至少一个输出装置。
用于实施本公开的模型传输方法的程序代码可以采用一个或多个编程语言的任何组合来编写。这些程序代码可以提供给通用计算机、专用计算机或其他可编程数据处理装置的处理器或控制器，使得程序代码当由处理器或控制器执行时使流程图和/或框图中所规定的功能/操作被实施。程序代码可以完全在机器上执行、部分地在机器上执行，作为独立软件包部分地在机器上执行且部分地在远程机器上执行或完全在远程机器或服务器上执行。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM，RandomAccessMemory)、只读存储器(ROM，Read-OnlyMemory)、可擦除可编程只读存储器(EPROM，Electrical Programmable Read Only Memory或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM，Compact Disc Read-Only Memory)、光学储存设备、磁储存设备、或上述内容的任何合适组合。
为了提供与用户的交互，可以在计算机上实施此处描述的系统和技术，该计算机具有：用于向用户显示信息的显示装置(例如，CRT(Cathode Ray Tube，阴极射线管)或者LCD(Liquid Crystal Display，液晶显示器)监视器)；以及键盘和指向装置(例如，鼠标或者轨迹球)，用户可以通过该键盘和该指向装置来将输入提供给计算机。其它种类的装置还可以用于提供与用户的交互；例如，提供给用户的反馈可以是任何形式的传感反馈(例如，视觉反馈、听觉反馈、或者触觉反馈)；并且可以用任何形式 (包括声输入、语音输入或者、触觉输入)来接收来自用户的输入。
可以将此处描述的系统和技术实施在包括后台部件的计算系统(例如，作为数据服务器)、或者包括中间件部件的计算系统(例如，应用服务器)、或者包括前端部件的计算系统(例如，具有图形用户界面或者网络浏览器的用户计算机，用户可以通过该图形用户界面或者该网络浏览器来与此处描述的系统和技术的实施方式交互)、或者包括这种后台部件、中间件部件、或者前端部件的任何组合的计算系统中。可以通过任何形式或者介质的数字数据通信(例如，通信网络)来将系统的部件相互连接。通信网络的示例包括：局域网(LAN，Local Area Network)、广域网(WAN，Wide Area Network)和互联网。
计算机系统可以包括客户端和服务器。客户端和服务器一般远离彼此并且通常通过通信网络进行交互。通过在相应的计算机上运行并且彼此具有客户端-服务器关系的计算机程序来产生客户端和服务器的关系。服务器可以是云服务器，也可以为分布式系统的服务器，或者是结合了区块链的服务器。
应该理解，可以使用上面所示的各种形式的流程，重新排序、增加或删除步骤。例如，本公开中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行，只要能够实现本公开公开的技术方案所期望的结果，本文在此不进行限制。
上述具体实施方式，并不构成对本公开保护范围的限制。本领域技术人员应该明白的是，根据设计要求和其他因素，可以进行各种修改、组合、子组合和替代。任何在本公开的精神和原则之内所作的修改、等同替换和改进等，均应包含在本公开保护范围之内。

Claims (13)

  1. 一种模型传输方法，其特征在于，包括：
    获取发送端节点与接收端节点之间的n个传输路径(S101)，其中，n为大于1的正整数；
    根据各个所述传输路径的传输能力，按照比例将所述发送端节点中待传输的模型切分成n组模型切片，并形成将n组所述模型切片与n个所述传输路径一一对应的路径选择策略(S102)；
    将切分好的n组所述模型切片分别通过对应的所述传输路径向所述接收端节点传输(S103)。
  2. 根据权利要求1所述的模型传输方法，其特征在于，所述按照比例将所述发送端节点中待传输的模型切分成n组模型切片包括：
    通过对所述待传输的模型切分成所述模型切片的切分时延和各个所述模型切片在对应的所述传输路径上的传输时延进行建模得到总传输时延作为目标函数；
    通过求解最小化所述目标函数得到模型切分比例以及所述路径选择策略，并按所述模型切分比例将所述待传输的模型切分成n组所述模型切片。
  3. 根据权利要求1或2所述的模型传输方法，其特征在于，每组所述模型切片包括一个或多个所述模型切片。
  4. 根据权利要求3所述的模型传输方法，其特征在于，所述模型传输方法还包括：
    在传输每一个所述模型切片之前，判断所述模型切片是否已经存储于当前的所述传输路径的任意一个路由节点中：
    若是，则切换至下一条所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中；
    若否，则使用当前的所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中。
  5. 根据权利要求1-4中任一所述的模型传输方法，其特征在于，所述传输能力包括所述传输路径的信道容量。
  6. 一种模型传输装置，其特征在于，包括：
    路径获取模块(601)，被配置为获取发送端节点与接收端节点之间的n个传输路径，其中，n为大于1的正整数；
    模型切分模块(602)，被配置为根据各个所述传输路径的传输能力，按照比例将所述发送端节点中待传输的模型切分成n组模型切片，并形成将n组所述模型切片与n个所述传输路径一一对应的路径选择策略；
    模型切片传输模块(603)，被配置为将切分好的n组所述模型切片分别通过对应的所述传输路径向所述接收端节点传输。
  7. 根据权利要求6所述的模型传输装置，其特征在于，所述模型切分模块按照比例将所述发送端节点中待传输的模型切分成n组模型切片的过程包括：
    通过对所述待传输的模型切分成所述模型切片的切分时延和各个所述模型切片在对应的所述传输路径上的传输时延进行建模得到总传输时延作为目标函数；
    通过求解最小化所述目标函数得到模型切分比例以及所述路径选择策略，并按所述模型切分比例将所述待传输的模型切分成n组所述模型切片。
  8. 根据权利要求6或7所述的模型传输装置，其特征在于，所述模型切分模块切分出的每组所述模型切片包括一个或多个所述模型切片。
  9. 根据权利要求8所述的模型传输装置，其特征在于，还包括：
    路径选择模块(604)，被配置为在传输每一个所述模型切片之前，判断所述模型切片是否已经存储于当前的所述传输路径的任意一个路由节点中：
    响应于所述模型切片已经存储于对应的所述传输路径中，所述模型切片传输模块切换至下一条所述传输路径进行传输，并在 传输过程中将当前的所述模型切片存储于对应的所述路由节点中；
    响应于所述模型切片未存储于对应的所述传输路径中，所述模型切片传输模块使用当前的所述传输路径进行传输，并在传输过程中将当前的所述模型切片存储于对应的所述路由节点中。
  10. 根据权利要求6-9中任一所述的模型传输装置，其特征在于，所述传输能力包括所述传输路径的信道容量。
  11. 一种电子设备，包括：
    至少一个处理器；以及
    与所述至少一个处理器通信连接的存储器；其中，
    所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行权利要求1-5中任一项所述的模型传输方法。
  12. 一种存储有计算机指令的非瞬时计算机可读存储介质，其中，所述计算机指令用于使所述计算机执行根据权利要求1-5中任一项所述的模型传输方法。
  13. 一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现根据权利要求1-5中任一项所述的模型传输方法。
PCT/CN2022/136415 2022-01-20 2022-12-04 模型传输方法、装置、电子设备及可读存储介质 WO2023138233A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210065202.9A CN116527510A (zh) 2022-01-20 2022-01-20 模型传输方法、装置、电子设备及可读存储介质
CN202210065202.9 2022-01-20

Publications (1)

Publication Number Publication Date
WO2023138233A1 true WO2023138233A1 (zh) 2023-07-27

Family

ID=87347757

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/136415 WO2023138233A1 (zh) 2022-01-20 2022-12-04 模型传输方法、装置、电子设备及可读存储介质

Country Status (2)

Country Link
CN (1) CN116527510A (zh)
WO (1) WO2023138233A1 (zh)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105847139A (zh) * 2016-03-25 2016-08-10 乐视控股（北京）有限公司 数据传输方法、装置及系统
CN110022261A (zh) * 2019-05-20 2019-07-16 北京邮电大学 基于sctp-cmt传输协议的多路径传输方法及设备
CN110769465A (zh) * 2019-10-24 2020-02-07 Oppo（重庆）智能科技有限公司 数据传输方法及装置、电子设备及可读存储介质
CN111263402A (zh) * 2020-01-17 2020-06-09 广东中兴新支点技术有限公司 数据传输方法、装置及电子设备
US20200366612A1 (en) * 2019-05-14 2020-11-19 Vmware, Inc. Congestion avoidance in a slice-based network

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN105847139A (zh) * 2016-03-25 2016-08-10 乐视控股（北京）有限公司 数据传输方法、装置及系统
US20200366612A1 (en) * 2019-05-14 2020-11-19 Vmware, Inc. Congestion avoidance in a slice-based network
CN110022261A (zh) * 2019-05-20 2019-07-16 北京邮电大学 基于sctp-cmt传输协议的多路径传输方法及设备
CN110769465A (zh) * 2019-10-24 2020-02-07 Oppo（重庆）智能科技有限公司 数据传输方法及装置、电子设备及可读存储介质
CN111263402A (zh) * 2020-01-17 2020-06-09 广东中兴新支点技术有限公司 数据传输方法、装置及电子设备

Also Published As

Publication number Publication date
CN116527510A (zh) 2023-08-01

Similar Documents

Publication Publication Date Title
US20220036230A1 (en) Quantum entangled state processing method, device, and storage medium
CN111475274B (zh) 云协同多任务调度方法及装置
WO2023221360A1 (zh) 深度学习模型的训练方法、装置、系统、设备和介质
JP2023025146A (ja) エンドツーエンドの自己適応に基づく分散型トレーニング方法、装置、及び機器
CN112769897A (zh) 边缘计算消息的同步方法、装置、电子设备及存储介质
JP7405933B2 (ja) 量子チャンネルクラシック容量の推定方法及び装置、電子機器と媒体
CN113645146B (zh) 基于新流密度的软件定义网络控制器负载均衡方法及系统
WO2023138233A1 (zh) 模型传输方法、装置、电子设备及可读存储介质
Kiamari et al. Coding for edge-facilitated wireless distributed computing with heterogeneous users
CN111680791B (zh) 适用于异构环境中的通信方法、装置、系统
Ju et al. eDeepSave: Saving DNN inference using early exit during handovers in mobile edge environment
CN115480882A (zh) 一种分布式边缘云资源调度方法及系统
WO2023179800A1 (zh) 通信接收方法及其装置
Li et al. Federated clouds for efficient multitasking in distributed artificial intelligence applications
CN106294445B (zh) 基于跨机房Hadoop集群的数据存储的方法及装置
Cheng et al. Efficient deep learning approach for computational offloading in mobile edge computing networks
WO2023138234A1 (zh) 模型管理方法、装置、组网架构、电子设备及存储介质
Alam et al. Derandomized distributed multi-resource allocation with little communication overhead
WO2023138231A1 (zh) 一种网络模型的残差传播方法和残差传播装置
WO2023179801A1 (zh) 数据处理方法、装置、通信系统、电子设备及存储介质
WO2023138232A1 (zh) 模型训练方法、装置、电子设备及存储介质
CN105704004B (zh) 业务数据处理方法和装置
CN113630330B (zh) 软件定义网络多控制器负载均衡方法及系统
CN117978851B (zh) 会话连接方法、交互方法、装置、设备及介质
Sinthiya et al. Low-cost Task Offloading Scheme for Mobile Edge Cloud and Internet Cloud Using Genetic Algorithm

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: 22921657

Country of ref document: EP

Kind code of ref document: A1