Model training method, resource management method for model training and related device

Abstract

The invention provides a model training method, a resource management method for model training and a related device. The method comprises the following steps: a first terminal device receives a first sub-model in a global model sent by a network device, wherein the first sub-model is determined according to a first segmentation layer; the first terminal device carries out training on the first sub-model; wherein the global model further comprises a second sub-model, and training of the first sub-model and training of the second sub-model are jointly used for determining a first local model; the first segmentation layer is determined according to the capability of the first terminal device, and / or the training duration of the first sub-model is used for determining whether the training round of the first local model participating in model aggregation is the current training round.

Description

Technical Field

[0001] The present application relates to the field of wireless communication technology, and more specifically, to a model training method, a resource management method for model training, and related devices. Background Art

[0002] With the development of communication technology, the integration of edge computing and artificial intelligence helps to utilize edge resources in distributed intelligent services. In edge networks, artificial intelligence models can be trained through various learning methods to obtain intelligent models with excellent performance.

[0003] However, related training methods may waste the powerful computing power of base stations or cause large latency overhead. Therefore, how to efficiently train artificial intelligence models is an urgent problem that needs to be solved. Summary of the Invention

[0004] The present application provides a model training method, a resource management method for model training, and related devices. The following introduces various aspects of the embodiments of the present application.

[0005] In a first aspect, a model training method is provided, including: a first terminal device receives a first sub-model in a global model, and the first sub-model is determined according to a first segmentation layer; the first terminal device trains the first sub-model; wherein the global model also includes a second sub-model, and the training of the first sub-model and the second sub-model is jointly used to determine a first local model; the first segmentation layer is determined according to the capability of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0006] According to a second aspect, a model training method is provided, comprising: a network device sends a first sub-model in a global model to a first terminal device, wherein the first sub-model is determined according to a first segmentation layer; the network device trains a second sub-model in the global model; wherein the training of the first sub-model and the second sub-model is jointly used to determine a first local model; the first segmentation layer is determined according to the capability of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0007] In a third aspect, a resource management method for model training is provided, comprising: decoupling the model training process into a single training round; determining a first constraint for the single training round; wherein the first constraint is used to determine a first management scheme, the first management scheme including multiple splitting layers corresponding to multiple terminal devices that perform the model training, the multiple terminal devices including a first terminal device, the multiple splitting layers including a first splitting layer corresponding to the first terminal device, and the first splitting layer being related to the capabilities of the first terminal device.

[0008] In a fourth aspect, a model training device is provided, which is a first terminal device, and the first terminal device includes: a receiving unit for receiving a first sub-model in a global model, and the first sub-model is determined according to a first segmentation layer; a processing unit for training the first sub-model; wherein the global model also includes a second sub-model, and the training of the first sub-model and the second sub-model is jointly used to determine a first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0009] In a fifth aspect, a model training device is provided, which is a network device, and the network device includes: a sending unit for sending a first sub-model in a global model to a first terminal device, wherein the first sub-model is determined according to a first segmentation layer; a processing unit for training a second sub-model in the global model; wherein the training of the first sub-model and the second sub-model is jointly used to determine a first local model; the first segmentation layer is determined according to the capability of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0010] In a sixth aspect, a resource management device for model training is provided, the resource management device comprising: a first processing unit for decoupling the model training process into a single training round; a second processing unit for determining a first constraint for the single training round; wherein the first constraint is used to determine a first management scheme, the first management scheme including multiple splitting layers corresponding to multiple terminal devices that perform the model training, the multiple terminal devices including a first terminal device, the multiple splitting layers including a first splitting layer corresponding to the first terminal device, and the first splitting layer being related to the capabilities of the first terminal device.

[0011] In a seventh aspect, a communication device is provided, comprising a memory and a processor, wherein the memory is used to store a program, and the processor is used to call the program in the memory to execute the method described in any one of the first to third aspects.

[0012] In an eighth aspect, a device is provided, comprising a processor for calling a program from a memory to execute the method as described in any one of the first to third aspects.

[0013] In a ninth aspect, a chip is provided, comprising a processor for calling a program from a memory so that a device equipped with the chip executes the method described in any one of the first to third aspects.

[0014] In a tenth aspect, a computer-readable storage medium is provided, on which a program is stored, wherein the program enables a computer to execute the method as described in any one of the first to third aspects.

[0015] In an eleventh aspect, a computer program product is provided, comprising a program, wherein the program enables a computer to execute the method according to any one of the first to third aspects.

[0016] In a twelfth aspect, a computer program is provided, wherein the computer program enables a computer to execute the method as described in any one of the first to third aspects.

[0017] In the embodiment of the present application, the first terminal device trains the first sub-model in the global model, and the training result and the training result of the second sub-model in the global model jointly determine the first local model corresponding to the first terminal device. The first segmentation layer of the first sub-model is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the first local model participates in the model aggregation of the current training round. In the current training round of the global model, the way in which the segmentation layer is associated with the capabilities of the terminal device and the training results corresponding to the terminal device may not participate in the model aggregation takes into account the differences in model training of different terminal devices, thereby improving training efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a wireless communication system applied in the embodiments of the present application.

[0019] Figure 2 It is a flowchart of a model training method provided in an embodiment of the present application.

[0020] Figure 3 yes Figure 2 A schematic diagram of a possible implementation of the method shown.

[0021] Figure 4 yes Figure 2A flowchart of a possible implementation of the method is shown.

[0022] Figure 5 This is a flowchart of a resource management method for model training provided in an embodiment of the present application.

[0023] Figure 6 yes Figure 5 A flowchart of a possible implementation of the method is shown.

[0024] Figure 7 It is a structural diagram of a model training device provided in an embodiment of the present application.

[0025] Figure 8 It is a structural diagram of another model training device provided in an embodiment of the present application.

[0026] Figure 9 This is a structural diagram of a resource management device for model training provided in an embodiment of the present application.

[0027] Figure 10 This is a schematic block diagram of a communication device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0028] The following will describe the technical solutions in the embodiments of this application in conjunction with the drawings in the embodiments of this application. Obviously, the embodiments described are part of the embodiments of this application, not all of the embodiments. With respect to the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0029] The embodiments of the present application can be applied to various communication systems. For example, the embodiments of the present application can be applied to global system of mobile communication (GSM) system, code division multiple access (CDMA) system, wideband code division multiple access (WCDMA) system, general packet radio service (GPRS), long term evolution (LTE) system, advanced long term evolution (LTE-A) system, new radio (NR) system, NR system evolution system, LTE-based access to unlicensed spectrum (LTE-U) system on unlicensed spectrum, NR-based access to unlicensed spectrum (NR-U) system on unlicensed spectrum, NTN system, universal mobile telecommunication system (UMTS), wireless local area networks (WLAN), wireless fidelity (WiFi), and fifth generation communication (5G) system. The embodiments of the present application may also be applied to other communication systems, such as future communication systems, such as sixth-generation (6G) mobile communication systems or satellite communication systems.

[0030] Traditional communication systems support a limited number of connections and are easy to implement. However, with the development of communication technology, communication systems can not only support traditional cellular communications, but also support one or more other types of communications. For example, a communication system can support one or more of the following communications: device to device (D2D) communication, machine to machine (M2M) communication, machine type communication (MTC), enhanced machine type communication (eMTC), vehicle to vehicle (V2V) communication, and vehicle to everything (V2X) communication, etc. The embodiments of the present application can also be applied to communication systems that support the above-mentioned communication methods.

[0031] The communication system in the embodiment of the present application can be applied to a carrier aggregation (CA) scenario, a dual connectivity (DC) scenario, and a standalone (SA) networking scenario.

[0032] The communication system in the embodiments of the present application can be applied to unlicensed spectrum. The unlicensed spectrum can also be considered a shared spectrum. Alternatively, the communication system in the embodiments of the present application can also be applied to licensed spectrum. The licensed spectrum can also be considered a dedicated spectrum.

[0033] The embodiments of the present application can be applied to NTN systems. As examples, the NTN systems may include 4G-based NTN systems, NR-based NTN systems, Internet of Things (IoT)-based NTN systems, and narrowband Internet of Things (NB-IoT)-based NTN systems.

[0034] A communication system may include one or more terminal devices. The terminal devices mentioned in the embodiments of the present application may also be referred to as user equipment (UE), access terminal, subscriber unit, subscriber station, mobile station, mobile station (MS), mobile terminal (MT), remote station, remote terminal, mobile device, user terminal, terminal, wireless communication device, user agent or user device, etc.

[0035] In some embodiments, the terminal device may be a station (ST) in a WLAN. In some embodiments, the terminal device may be a cellular phone, a cordless phone, a session initiation protocol (SIP) phone, a wireless local loop (WLL) station, a personal digital assistant (PDA) device, a handheld device with wireless communication capabilities, a computing device or other processing device connected to a wireless modem, an in-vehicle device, a wearable device, a terminal device in a next-generation communication system (e.g., a NR system), or a terminal device in a future-evolved public land mobile network (PLMN) network.

[0036] In some embodiments, a terminal device may be a device that provides voice and / or data connectivity to a user. For example, the terminal device may be a handheld device, an in-vehicle device, etc. with wireless connection capabilities. As some specific examples, the terminal device may be a mobile phone, a tablet computer, a laptop computer, a PDA, a mobile internet device (MID), a wearable device, a virtual reality (VR) device, an augmented reality (AR) device, a wireless terminal in industrial control, a wireless terminal in self-driving, a wireless terminal in remote medical surgery, a wireless terminal in a smart grid, a wireless terminal in transportation safety, a wireless terminal in a smart city, a wireless terminal in a smart home, etc.

[0037] In some embodiments, the terminal device can be deployed on land. For example, the terminal device can be deployed indoors or outdoors. In some embodiments, the terminal device can be deployed on the water, such as on a ship. In some embodiments, the terminal device can be deployed in the air, such as on an airplane, a balloon, or a satellite.

[0038] In addition to the terminal device, the communication system may also include one or more network devices. The network device in the embodiment of the present application may be a device for communicating with the terminal device, and the network device may also be referred to as an access network device or a radio access network device. The network device may be, for example, a base station. The network device in the embodiment of the present application may refer to a radio access network (RAN) node (or device) that connects the terminal device to a wireless network. A base station can broadly cover various names as follows, or be replaced with the following names, such as: NodeB, evolved NodeB (eNB), next generation NodeB (gNB), relay station, transmission point (TRP), transmission point (TP), master station (MeNB), secondary station (SeNB), multi-standard radio (MSR) node, home base station, network controller, access node, wireless node, access point (AP), transmission node, transceiver node, baseband unit (BBU), remote radio unit (RRU), active antenna unit (AAU), remote radio head (RRH), central unit (CU), distributed unit (DU), positioning node, etc. A base station can be a macro base station, a micro base station, a relay node, a donor node or the like, or a combination thereof. A base station can also refer to a communication module, a modem or a chip used to be set in the aforementioned device or apparatus. The base station can also be a mobile switching center and a device that performs base station functions in D2D, V2X, and M2M communications, a network-side device in a 6G network, or a device that performs base station functions in future communication systems. The base station can support networks with the same or different access technologies. The embodiments of this application do not limit the specific technology and specific device form used by the network equipment.

[0039] Base stations can be fixed or mobile. For example, a helicopter or drone can be configured to act as a mobile base station, and one or more cells can move based on the location of the mobile base station. In other examples, a helicopter or drone can be configured to act as a device that communicates with another base station.

[0040] In some deployments, the network device in the embodiments of the present application may refer to a CU or a DU, or the network device may include a CU and a DU. The gNB may also include an AAU.

[0041] By way of example and not limitation, in embodiments of the present application, a network device may be mobile, for example, a mobile device. In some embodiments of the present application, the network device may be a satellite or balloon station. In some embodiments of the present application, the network device may also be a base station located on land, water, or the like.

[0042] In an embodiment of the present application, the network device can provide services for a cell, and the terminal device communicates with the network device through the transmission resources used by the cell (for example, frequency domain resources, or spectrum resources). The cell can be a cell corresponding to the network device (for example, a base station). The cell can belong to a macro base station or a base station corresponding to a small cell. The small cells here may include: metro cells, micro cells, pico cells, femto cells, etc. These small cells have the characteristics of small coverage and low transmission power, and are suitable for providing high-speed data transmission services.

[0043] For example, Figure 1 This is a schematic diagram of the architecture of a communication system provided in an embodiment of the present application. Figure 1 As shown, the communication system 100 may include a network device 110, which may be a device that communicates with a terminal device 120 (or a communication terminal or terminal). The network device 110 may provide communication coverage for a specific geographic area and may communicate with terminal devices within the coverage area.

[0044] Figure 1 A network device and two terminal devices are shown as an example. In some embodiments of the present application, the communication system 100 may include multiple network devices and the coverage area of ​​each network device may include other numbers of terminal devices, which is not limited in the embodiments of the present application.

[0045] In the embodiments of this application, Figure 1 The communication system shown also includes other network entities such as a mobility management entity (MME) and an access and mobility management function (AMF), which are not limited in the embodiments of the present application.

[0046] It should be understood that the device with communication function in the network / system in the embodiment of the present application can be called a communication device. Figure 1Taking the communication system 100 shown as an example, the communication equipment may include a network device 110 and a terminal device 120 with communication functions. The network device 110 and the terminal device 120 may be the specific devices described above and will not be repeated here; the communication equipment may also include other devices in the communication system 100, such as a network controller, a mobile management entity and other network entities, which is not limited in the embodiments of the present application.

[0047] To facilitate the detailed explanation of the innovative features of the technical solution, we first introduce some relevant technical knowledge involved in the embodiments of this application. The following related technologies can be combined with the technical solutions of the embodiments of this application as optional solutions, and they all fall within the scope of protection of the embodiments of this application. The embodiments of this application include at least part of the following contents.

[0048] With the continuous development of communication technology, intelligent services have increasingly higher performance requirements for artificial intelligence (AI) models. For example, driven by the vision of pan-intelligence in 6G networks, the integration of edge computing and artificial intelligence has developed rapidly by making full use of edge resources for distributed intelligent services. Among them, distributed intelligent services include autonomous driving, smart transportation, target detection, etc. Edge resources used to provide distributed intelligent services include edge computing, communication, storage, etc.

[0049] Edge computing enables edge devices in a network to train artificial intelligence models, resulting in high-performance intelligent models. However, efficiently achieving high-performance intelligent models is a pressing challenge, especially for resource-constrained wireless networks. These constraints can include limited computing power, storage capacity, or wireless resources on edge devices.

[0050] Currently, the main methods for training AI models in edge networks include federated learning, split learning, and federated split learning. The following describes each of these methods, using the example of training deep neural networks in AI models on base stations and end devices.

[0051] Federated learning, as a distributed intelligence paradigm, can train AI models to enable intelligent services. Typically, a federated learning paradigm in an edge network consists of a base station and multiple end devices. The federated learning process is divided into multiple training rounds. In a given training round, the federated learning process includes the following steps S11-S13.

[0052] S11: The base station broadcasts the global model to all terminal devices. The global model is a deep neural network.

[0053] S12: All terminal devices use local data sets to perform global model training in parallel to obtain local models, and transmit their local models to the base station through wireless channels respectively.

[0054] S13: The base station aggregates the local models transmitted by all terminal devices to obtain a new global model.

[0055] The base station and the terminal device repeat the above process of S11-S13 until the global model meets the preset convergence conditions, at which point federated learning is completed.

[0056] As shown in processes S11-13, in federated learning, terminal devices need to train a complete AI model. However, terminal devices with limited resources (such as computing and storage) cannot afford to train large AI models. Furthermore, the base station is only responsible for aggregating the local models uploaded by the terminal devices, and the base station's powerful computing power is not fully utilized, thus wasting its powerful computing power.

[0057] Split learning, another distributed intelligence paradigm, is primarily targeted at training AI models in resource-constrained wireless networks to enable intelligent services. The split learning paradigm in edge networks also consists of a base station and multiple end devices. In addition, the base station is equipped with a powerful edge server. Typically, the split learning process is also divided into multiple training rounds. During a training round, the end device, base station, and edge server execute processes S21-S27.

[0058] S21: The edge server divides the global model (deep neural network) into a device-side global model and a server-side global model. The device-side global model is the global model located on the terminal device. The cut point used to divide the global model is called a cut layer.

[0059] S22: The base station transmits the device-side global model to a terminal device. The server-side global model is placed in the edge server to assist the terminal device in training the global model.

[0060] S23: The terminal device receiving the global model performs forward propagation (FP) of the device-side global model using the local data set to obtain the output activation of the neural network. The output activation is, for example, an output feature vector (smashed data, SD). The terminal device can transmit the output activation of the neural network and the corresponding data label to the base station via a wireless channel.

[0061] S24: The edge server continues to execute the forward propagation and backward propagation (BP) of the server-side global model to obtain the output activation gradient of the neural network.

[0062] S25: The base station transmits the output activation gradient of the neural network to the terminal device, and the edge server performs an update of the server-side global model to obtain a local server-side model.

[0063] S26: The terminal device performs back propagation of the device-side global model and updates the device-side global model to obtain a local device-side model.

[0064] S27: The terminal device transmits the local device-side model to the base station through the wireless channel. The base station acts as a relay to send the local device-side model to the next terminal device. When all terminal devices complete one training, a training round is completed.

[0065] The model training device repeats the above process of S21-S27 until the global model meets the preset convergence conditions, at which point the segmentation learning is completed.

[0066] As can be seen from processes S21-27, in segmented learning, the base station and terminal devices perform collaborative training of the global model in a serial manner. However, this approach means that only one terminal device interacts with the base station at any given moment, leaving the remaining terminal devices idle. This results in a large amount of unused distributed idle resources and incurs significant latency overhead.

[0067] Federated segmentation learning, as a hybrid distributed intelligence paradigm, enables efficient training of AI models in resource-constrained wireless networks, enabling intelligent services. The federated segmentation learning paradigm for edge networks also consists of a base station and multiple end devices. To implement segmentation learning, the base station is equipped with a powerful edge server. The federated segmentation learning process is typically divided into multiple training rounds. During a training round, the end device, base station, and edge server execute processes S31-S37.

[0068] S31: The edge server divides the global model (deep neural network) into a device-side global model and a server-side global model.

[0069] S32: The base station broadcasts the device-side global model to all terminal devices. At the same time, the server-side global model is replicated as many times as the number of terminal devices to assist all terminal devices in training the global model.

[0070] S33: All terminal devices use their local datasets to perform forward propagation of their device-side global models in parallel, obtaining the output activations of the neural network. All terminal devices transmit the output activations of the neural network and the corresponding data labels to the base station via wireless channels.

[0071] S34: The edge server continues to execute the forward propagation and back propagation of the server-side global model for each terminal device to obtain the output activation gradient of the neural network.

[0072] S35: The base station transmits the output activation gradient of the neural network to each terminal device respectively, and the edge server executes the server-side global model update for each terminal device to obtain a local server-side model.

[0073] S36: All terminal devices perform back propagation of the device-side global model in parallel and perform updates on the device-side global model to obtain local device-side models. All terminal devices transmit their local device-side models to the base station via the wireless channel.

[0074] S37: The edge server combines the local device-side model and the local server-side model into a local model for each terminal device, and performs aggregation of the local models to obtain a global model.

[0075] The model training device repeats the above S31-S37 process until the global model meets the preset convergence conditions, at which point the federated segmentation learning is completed.

[0076] As can be seen from processes S31-37, in federated segmented learning, all terminal devices perform the same amount of device-side model training. Due to the differences in terminal device resources, terminal devices with powerful computing capabilities are not fully utilized in AI model training. In addition, the edge server needs to wait for all terminal devices to complete local training before performing global model aggregation in each training round. In other words, the latency of each round depends on the slowest terminal device, which results in a large latency overhead.

[0077] In addition, in wireless communications, the time-varying allocation of wireless channels and wireless resources will also affect the training quality of the global model. The allocation of wireless resources, for example, is the allocation of resources such as bandwidth and computing frequency. Due to the rapid changes in wireless channels, wireless resources may not be reasonably allocated, so it is impossible to ensure that the artificial intelligence model can perform efficient training in the long term. In particular, for wireless networks with limited resources and heterogeneous device capabilities, the quality of model training is more easily affected. For example, within each training round, the allocation of wireless channels and wireless resources can remain unchanged for a short time, but it is difficult to remain unchanged within multiple rounds, making it impossible to ensure long-term efficient training.

[0078] In response to the problems of the above-mentioned multiple model training methods, the embodiment of the present application proposes a method for training a model. This method takes into account the differences in terminal device resources and makes full use of the powerful computing power of the base station to undertake the model training task, thereby improving the training efficiency on the basis of ensuring the global model training performance. For ease of understanding, the following is combined with Figure 2 The method of training the model is described in detail.

[0079] Figure 2 This description is based on the interaction between a first terminal device and a network device. The first terminal device is one of the aforementioned terminal devices that possesses certain computing and communication capabilities. In some embodiments, the first terminal device can train a model based on local data samples. This model can be an artificial intelligence / machine learning model. In some embodiments, the first terminal device can send data samples and / or computing models to the network device. In some embodiments, the first terminal device can receive the computing model sent by the network device.

[0080] As an example, the network device may send the artificial intelligence model to the first terminal device via broadcasting.

[0081] In some embodiments, the first terminal device may be any terminal device in the edge network. The first terminal device may store a variety of data samples. Exemplarily, the data samples stored by the first terminal device may include local data samples used to train the model.

[0082] The first terminal device is any terminal device among the multiple terminal devices participating in the model training. For example, the multiple terminal devices can jointly train the artificial intelligence model with the network device. For example, the multiple terminal devices can provide data samples for the model training.

[0083] The network device is any of the aforementioned communication devices that provide services to multiple terminal devices. In some embodiments, the network device is a communication device with powerful computing capabilities. For example, the network device can be the base station that broadcasts the global model to multiple terminal devices based on federated learning, or the base station that assists multiple terminal devices in training the global model based on segmented learning, without limitation.

[0084] In some embodiments, the network device is equipped with an edge server with strong computing power to participate in model training. As an example, the edge server can be used to determine at least one of the following information under the first constraint: the communication bandwidth of each terminal device performing model training; the computing frequency allocation method of the edge server; the selection of multiple terminal devices for determining the first aggregation period; and the multiple segmentation layers including the first segmentation layer. Figure 5 The first constraint condition and various types of information are described.

[0085] The network device can communicate with multiple terminal devices including the first terminal device. In some embodiments, the network device can receive data samples or intermediate data of model training sent by multiple terminal devices. In some embodiments, the network device can send a global model of a certain training round, segmented sub-models, and / or intermediate data of model training to multiple terminal devices.

[0086] See also Figure 2 In step S210, the first terminal device receives the first sub-model in the global model. In some embodiments, the first sub-model received by the first terminal device may be from a network device, such as Figure 2 In some embodiments, the first sub-model received by the first terminal device may come from a third-party device that assists the network device in model training.

[0087] As an example, the third-party device is a server deployed outside the base station and can communicate with the first terminal device.

[0088] The global model can be any artificial intelligence / machine learning model that supports intelligent services, and this embodiment of the application does not limit this. Types of global models include but are not limited to: convolutional neural network models, recurrent neural network models, long short-term memory networks, etc.

[0089] In some embodiments, the global model may be determined by a network device. For example, when performing federated segmentation learning on the global model, the network device needs to aggregate the distributed learning results of multiple terminal devices to determine the global model of the current training round.

[0090] The global model can be the model trained in the previous training round. The training method of the global model generally includes multiple training rounds. A training round can refer to a training process or a learning process, also known as a learning round. For example, in the federated learning described above, one training round is used to complete the process of S11-S13. For another example, in the segmented learning described above, one training round is used to complete the process of S21-S27. For another example, in the federated segmented learning described above, one training round is used to complete the process of S31-S37.

[0091] As an example, the training process of the global model can be decoupled according to Lyapunov optimization theory to optimize the training effect. Any training round in the global model training can be any single training round obtained based on this decoupling.

[0092] In some embodiments, the global model can be segmented and trained based on segmented learning or federated segmented learning. The global model can be segmented to form at least one sub-model. In other words, the global model can be composed of multiple segmented sub-models.

[0093] For example, after a global model is split, multiple sub-models can be trained on different devices to improve training efficiency. When a sub-model is trained on a terminal device, it is called a device-side global model. When a sub-model is trained on a server, it is called a server-side global model.

[0094] It should be understood that the training of the global model needs to be performed based on the local data of the terminal device. Therefore, the device-side global model generally includes an input terminal for data samples. The terminal device can input local data through the input terminal of the device-side global model and perform forward propagation. In addition, the terminal device can also use backpropagation on the device-side global model to obtain a trained local model, also known as a local model.

[0095] In some embodiments, for segmented learning or federated segmented learning, the training rounds of the global model can be set based on the training period of the terminal device or the server. For example, a training round mainly includes the communication duration and training duration on the terminal device side.

[0096] As an example, for the terminal side, the training rounds can be done without considering the model aggregation time on the server side. Because the computing power of the server side is relatively strong, the model aggregation time can be relatively negligible.

[0097] In some embodiments, a training round can be divided into multiple training cycles and communication cycles. That is, the training duration and communication duration need to be considered within a training round. For a terminal device, a communication cycle can refer to an upload cycle and / or a download cycle.

[0098] As an example, in federated segmentation learning, the base station and the terminal device collaborate to train for T training rounds, where T is a positive integer. A training round may include a device-side global model download cycle, H terminal device and base station collaborative training cycles, and a device-side local model upload cycle, where H is a positive integer. The device-side global model download cycle, the terminal device and base station collaborative training cycle, and the device-side local model upload cycle are in a serial relationship. Taking the training of neural networks as an example, each terminal device and base station collaborative training cycle includes a terminal device forward propagation small cycle, a neural network output activation upload small cycle, an edge server-side forward and backward propagation small cycle, a neural network output activation gradient download small cycle, and a terminal device backpropagation small cycle. Each small cycle is in a serial relationship, which will be combined later. Figure 4 Each of the above cycles and small cycles is exemplified.

[0099] Optionally, the length of each cycle in each training round depends on the resources of the terminal device, the wireless channel status, the communication and computing capabilities of the base station, the adaptive partitioning scheme of the global model, and the semi-asynchronous aggregation strategy of the local model.

[0100] In some embodiments, during the model training process, the global model in step S210 may be the global model applied to any training round. That is, the global model may be the model trained in any training round. In some embodiments, the global model may be the model determined in the previous training round before any training round. That is, the global model may be the model determined after the completion of any training round other than the training round in which the model converged.

[0101] The training of the global model is performed jointly by the network device and multiple terminal devices including the first terminal device. In an embodiment of the present application, the network device can send the sub-models after segmenting the global model to multiple terminal devices by broadcasting, so that all terminal devices participating in the model training receive the sub-models trained in the current training round.

[0102] The sub-model received by the first terminal device is the first sub-model in the global model. The first sub-model is a sub-model obtained by segmenting the global model for the first terminal device. The first sub-model can also be called the device-side global model corresponding to the first terminal device.

[0103] The first sub-model is determined based on the first segmentation layer. That is, for the first terminal device, the global model is segmented at the first segmentation layer. The segmented portion used for training the first terminal device is the first sub-model.

[0104] As an example, when the first terminal device is the nth terminal device among N terminal devices that perform model training, the first segmentation layer can be l n Wherein, N is an integer greater than or equal to 1, that is, a positive integer; 1≤n≤N.

[0105] In some embodiments, at least two of all participating terminal devices correspond to different segmentation layers. In other words, at least two of the multiple participating terminal devices receive different sub-models. This means that different terminal devices require different sub-models for training, helping to address the varying training times caused by differences in computing / storage capabilities across multiple terminal devices, thereby reducing training latency caused by heterogeneity among the different terminal devices.

[0106] In some embodiments, the first terminal device is one of multiple terminal devices performing model training. The multiple sub-models received by the multiple terminal devices are determined based on the global model and multiple different segmentation layers. Therefore, the multiple sub-models include the first sub-model, and the multiple segmentation layers include the first segmentation layer. As an example, the multiple segmentation layers correspond to the multiple terminal devices in a one-to-one relationship. As an example, the multiple segmentation layers correspond to the multiple terminal devices in a one-to-many relationship.

[0107] As an example, different sub-models may refer to different sub-model sizes. For example, the number of computational layers of the two sub-models corresponding to the two terminal devices may be different. Another example is that the bit sizes of the sub-models corresponding to the two terminal devices after quantization are different.

[0108] The first sub-model may be determined based on one or more factors, which may include one or more of the following: capability of the first terminal device, number of data samples of the first terminal device, and wireless resource status between the first terminal device and the network device.

[0109] In some embodiments, the first segmentation layer can be determined based on the capabilities of the first terminal device, thereby dividing the global model for the multiple terminal devices based on the capability differences of the multiple terminal devices. Exemplarily, the first segmentation layer can be determined based on the capability level corresponding to the first terminal device. For example, the second terminal device performing model training corresponds to the third sub-model, and when the capability level of the first terminal device is higher than that of the second terminal device, the first sub-model is larger than the third sub-model.

[0110] As an example, all terminal devices performing model training can be divided into multiple terminal device sets based on capability levels. The terminal devices in each terminal device set correspond to the same capability level. For example, a first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

[0111] In some embodiments, the first segmentation layer can be determined based on the number of data samples of the first terminal device, thereby reducing the difference in training time caused by the difference in the number of data samples. The number of data samples is also the number of data set samples. Exemplarily, the first segmentation layer can be determined based on the range of the number of data set samples corresponding to the first terminal device. For example, the number of data set samples of the first terminal device is within a first range, and the number of data set samples of the second terminal device performing model training is within a second range. When the lower limit value of the first range is greater than or equal to the upper limit value of the second range, the first sub-model is smaller than the third sub-model corresponding to the second terminal device.

[0112] In some embodiments, the first segmentation layer can be based on the radio resource status between the first terminal device and the network device. The radio resource status can include one or more parameters such as communication bandwidth, channel environment, and channel quality. For example, the edge server can adaptively determine the first segmentation layer for the first terminal device based on the radio resource status of the first terminal device, thereby performing global model segmentation.

[0113] In some embodiments, the first segmentation layer may be determined by comprehensively considering the aforementioned multiple information to balance the influence of multiple factors. It should be understood that the process of determining the segmentation layer based on one or more of the aforementioned factors by the network device may be dynamically adjusted, also known as an adaptive process.

[0114] After the global model is segmented at the first segmentation layer, the portion used for training on the network device side is the second sub-model. The network side can train the second sub-model via the server, so the second sub-model can also be called the server-side global model corresponding to the first terminal device. The first sub-model and the second sub-model can constitute the global model. In other words, the global model includes the first sub-model and the second sub-model.

[0115] In some embodiments, the first terminal device executing step S210 may indicate the start of the current training round. That is, after each training round starts, the network device sends the first sub-model in the global model to the first terminal device.

[0116] In step S220, the first terminal device trains the first sub-model. Correspondingly, the second sub-model in the global model corresponding to the first sub-model is trained on the network device. As an example, the first terminal device and the network device train two sub-models in the global model separately and exchange intermediate parameters, thereby achieving collaborative training of the global model.

[0117] As an example, the first terminal device can use the stochastic gradient descent method to collaborate with the base station to train the global model.

[0118] In some embodiments, the first terminal device may train the first sub-model based on some or all of the data samples in the data set. The data set may be a data set consisting of local data or other data, which is not limited here.

[0119] The data set samples of the first terminal device may include but are not limited to pictures, audio, signals, etc. The embodiments of the present application are not limited thereto.

[0120] In some embodiments, the network device trains the second sub-model based on the intermediate data uploaded by the first terminal device, and sends the training parameters of the second sub-model to the first terminal device so that the first terminal device can update the first sub-model.

[0121] As an example, the intermediate data uploaded by the first terminal device may be a neural network output activation, such as the SD mentioned above.

[0122] As an example, the intermediate data uploaded by the first terminal device also includes a sample tag or a device identifier to facilitate identification by the network device.

[0123] As an example, the training parameter sent by the network device to the first terminal device may be a gradient of activation output by a neural network.

[0124] As an example, step S220 may include a process of interaction between the first terminal device and the network device. Exemplarily, the first terminal device performs forward propagation on the first sub-model, obtains output activation at or before the split layer, and uploads it. The network device performs forward propagation and backward propagation on the second sub-model based on the output activation to obtain the gradient of the output activation. The network device sends the gradient of the output activation to the first terminal device and updates the second sub-model to the server-side local model. The first terminal device performs backward propagation on the first sub-model based on the gradient of the output activation, thereby updating the first sub-model to the device-side local model.

[0125] In some embodiments, the training of the first sub-model and the second sub-model is jointly used to determine a first local model. The first local model is a local model corresponding to the first terminal device obtained after the first terminal device and the network device collaboratively train the global model. The first local model can be used by the network device to perform model aggregation to obtain an updated global model.

[0126] As an example, when the first terminal device is the nth terminal device, the first local model of the tth training round (1≤t≤T) is:

[0127]

[0128] Among them, ⊕ is the concatenation symbol.

[0129] In some embodiments, the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round, so as to avoid the entire training duration being affected by the longer training duration of a certain terminal device. In other words, the terminal device performing model training can determine whether to use the current training results for model aggregation based on actual conditions. Alternatively, the network device can determine whether the terminal devices participating in model aggregation include the first terminal device based on actual conditions.

[0130] The training duration of the first sub-model can be replaced by the training duration of the global model corresponding to the first terminal device, that is, the training duration of the global model jointly trained by the first terminal device and the network device. The training duration of the first sub-model may change dynamically with changes in state parameters.

[0131] In some embodiments, the training duration of the first sub-model is determined based on one or more of the following information: the position of the first segmentation layer in the global model; the capabilities / number of data samples of the first terminal device; the status of wireless resources between the first terminal device and the network device; and the computing frequency allocated by the network device to the first terminal device.

[0132] In some embodiments, the current training round may be a single training round that is currently being trained. As can be seen from the foregoing, if the training round in which the first local model participates in model aggregation is the current training round, the first terminal device is the terminal device participating in the model aggregation this time. If the training round in which the first local model participates in model aggregation is not the current training round, the first terminal device continues to train the first sub-model in one or more training rounds after the current training round. It can be seen that the first terminal device does not need to terminate the current training. Only when the training of the first sub-model is completed and the status of the first local model meets the requirements will it participate in model aggregation.

[0133] As an example, the state of the first local model in the current training round can be represented by a parameter to indicate whether it can participate in model aggregation. For example, when the state of the first local model is 1, the first local model participates in model aggregation in the current training round; when the state of the first local model is 0, the first local model does not participate in model aggregation in the current training round. The reverse is also true.

[0134] In some embodiments, the relationship between the training duration of the first sub-model and the first aggregation period is used to determine whether the parameters of the first partial model are aggregated with the model of the current training round. The first aggregation period is a parameter preset in the current training round for determining the model aggregation time.

[0135] As an example, when the training duration of the first sub-model is longer than the duration of the first aggregation period, the first local model does not participate in model aggregation in the current training round; when the training duration of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in model aggregation in the current training round. In other words, after the training duration of the first sub-model exceeds the set first aggregation period, the first local model determined by the first terminal device in the current training round will undergo model aggregation in subsequent training rounds.

[0136] In some embodiments, the length of each aggregation period in each training round depends on the terminal devices participating in the local model aggregation. As an example, the first aggregation period can be determined based on the states of N local models of N terminal devices performing model training and a first threshold. The first threshold can be the minimum number of local models participating in the model aggregation, such as N min .

[0137] For example, the judgment rule for local model aggregation can be determined as follows:

[0138]

[0139] Among them, m n,t is the state of the local model of the nth terminal device. n,t =1 indicates that the nth terminal device has obtained the local model and can participate in the local model aggregation, and vice versa. min Aggregate the number of terminal devices participating in the local model.

[0140] In some embodiments, after the network device aggregates the local models corresponding to some or all terminal devices in the current training round, it can obtain a global model for the next training round. The global model for the next training round can be determined based on multiple weighting coefficients corresponding to the multiple terminal devices participating in the model aggregation. That is, the network device performs model aggregation based on weighting. The weighting coefficients can also be called weight coefficients.

[0141] In some embodiments, the multiple weighting coefficients are determined based on the aggregation interval for multiple terminal devices to participate in model aggregation and / or the bias parameter for controlling model aggregation. The aggregation interval for participating in model aggregation can indicate whether the terminal device participates in model aggregation in every training round. The bias parameter for controlling model aggregation can indicate the direction or target of model aggregation.

[0142] For example, when the aggregation interval is 1, the terminal device participates in model aggregation in every training round. When the aggregation interval is greater than 1, the terminal device participates in model aggregation only after multiple training rounds.

[0143] As an example, when the bias parameter is greater than 1, model aggregation can reduce the risk of global model training being biased towards end devices with powerful computing capabilities. When the bias coefficient is less than 1, model aggregation accelerates global model convergence at the expense of global model training being biased towards end devices with powerful computing capabilities. When the bias coefficient is equal to 1, the weighting coefficient is only related to the dataset samples of the end device and is not affected by the heterogeneity of the end device.

[0144] For example, when the current training round is the t-th training round in T training rounds, the global model w of the t+1-th training round is t+1 for:

[0145]

[0146] Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices at the tth training round, ρ n,t represents the weighted coefficient of the nth terminal device in the tth training round, represents the local model of the nth terminal device at the tth training round.

[0147] When the nth terminal device is the terminal device set S participating in the model aggregation t One of them, weighting coefficient ρ n,t for:

[0148]

[0149] Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t Denotes the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t represents the aggregation interval of the kth terminal device.

[0150] Depend on Figure 2 As can be seen, the embodiments of the present application propose an adaptive federated segmentation learning paradigm. This method fully considers the differences in terminal device capabilities, the number of data samples, and wireless resources. For example, this method can enable terminal devices with weak computing power to perform fewer device-side global model training tasks, while terminal devices with strong computing power can take on more device-side global model training tasks.

[0151] Furthermore, in this method, except for the device-side global model, the remaining server-side global models are trained with the assistance of network devices. The collaborative training of all terminal devices and network devices is carried out in parallel, thereby improving the efficiency of global model training. It can be seen that the model training method in the embodiment of the present application can not only fully utilize the powerful computing power of network devices such as base stations, but also overcome the limited resources of terminal devices in wireless networks, while improving the training efficiency while ensuring the performance of global model training.

[0152] For ease of understanding, the following Figure 3 and Figure 4 The adaptive federated segmentation learning method proposed in the embodiment of the present application is schematically illustrated. Figure 3 A structural diagram of an adaptive federated segmentation learning paradigm provided in an embodiment of the present application. Figure 4 A flowchart of an adaptive federated segmentation learning paradigm based on semi-asynchronous local model aggregation is provided in an embodiment of the present application.

[0153] See also Figure 3 The federated segmentation learning system may include one network device (base station 310) and N terminal devices. The base station 310 is equipped with an edge server 320 with powerful computing capabilities. The N terminal devices are terminal devices 301, ..., terminal devices 30n to terminal devices 30N. Figure 3 The three terminal devices shown correspond to different capabilities and data sample quantities. The data set D1 of the terminal device 301 is smaller than the data set D n , dataset D n Data set D smaller than 30N of terminal equipment N N terminal devices and the base station may perform local collaborative training, and the collaborative training may be performed for T training rounds.

[0154] like Figure 3 As shown, in collaborative training, the terminal device will execute the download of the device-side global model (first sub-model), upload of the neural network output activation, download of the gradient of the neural network output activation, and upload of the device-side local model.

[0155] Continue to see Figure 3 , when entering the tth training round, the base station 310 or the edge server 320 may have a global model 330, that is, w t During the global model download cycle on the device side, the edge server will adaptively divide the global model w for the nth terminal device according to the wireless resource status and terminal device capabilities of the terminal device. t For the device-side global model (First sub-model 333) and the server-side global model (Second sub-model 334). For example, the first sub-model 331 of the terminal device 301 is The second sub-model 332 is The first sub-model 335 of the terminal device 30N is The second sub-model 336 is

[0156] Furthermore, the base station transmits the device-side global model to the nth terminal device Server-side global model It is placed at the base station.

[0157] In the τ(1≤τ≤H)th terminal device and base station collaborative training cycle, the nth terminal device uses the data set D nSample execution device-side global model The forward propagation of the neural network obtains the output activation and transmits it and the corresponding sample label to the base station. Then, the edge server at the base station executes the server-side global model The forward propagation and back propagation of the neural network get the gradient of the output activation.

[0158] The base station transmits the gradient of the output activation of the neural network to the nth terminal device and updates the server-side global model Get the server-side local model The nth terminal device executes the device-side global model Update the device-side global model after backpropagation Get the local model on the device The above process is repeated H times to obtain the local model of the device side Server-side local models

[0159] During the device-side local model upload cycle, the nth terminal device uploads the device-side local model Transmitted to the base station, the edge server will With its server-side local model Combine into local models Then, the edge server models the local models of the N terminal devices according to the weight coefficients (weighted coefficients) of the N common terminal devices to obtain the global model w for the next training round. t+1 .

[0160] Among them, the weight coefficient of the nth terminal device can be expressed as ρ n,t When the local models of all N terminal devices participate in model aggregation, ρ n,t To satisfy When only the local models of some terminal devices participate in model aggregation, ρ n,t Only the case of terminal devices participating in model aggregation is considered.

[0161] See also Figure 4 In step S410, when entering the preset training round, the edge server adaptively divides the global model for each terminal device into a device-side global model and a server-side global model (second sub-model), and transmits the device-side global model to each terminal device. The device-side global model is the first sub-model, and the server-side global model is the second sub-model.

[0162] The training process of step S410 occurs within the device-side global model download cycle. For example, in the tth training round, when the terminal device and the base station adopt the frequency-division multiple access (FDMA) transmission mode, the device-side global model download cycle of the nth terminal device is It can be determined as:

[0163]

[0164] in, For the nth terminal device, split layer l n The global model size of the device side is divided, is the downlink (DL) transmission rate of the nth terminal device. Optionally, can be expressed as:

[0165]

[0166] Among them, B n,t is the communication bandwidth allocated to the nth terminal device, is the transmission power of the base station to the nth terminal device, h n,t is the attenuation coefficient of the wireless channel between the nth terminal device and the base station, and N0 is the additive white Gaussian noise power.

[0167] In step S420, the terminal device uses the local data set to collaboratively train the global model of the previous round with the base station. This training process occurs within H terminal device and base station collaborative training cycles. As described above, each terminal device and base station collaborative training cycle can include a terminal device forward propagation mini-cycle, a neural network output activation upload mini-cycle, an edge server-side forward and backward propagation mini-cycle, a neural network output activation gradient download mini-cycle, and a terminal device backward propagation mini-cycle. The following is an exemplary description of each of the multiple mini-cycles.

[0168] The nth terminal device uses the stochastic gradient descent method to train the global model in collaboration with the base station. The terminal device forward propagates a small period for:

[0169]

[0170] Where b is the number of samples that the nth terminal device performs in a terminal device and base station collaborative training cycle, The number of floating point operations (FLOPs) that need to be considered when training a data sample based on the device-side global model and performing forward propagation for the nth terminal device, f nis the central processing unit (CPU) frequency of the nth terminal device, q n is the number of FLOPs of the n-th terminal device in each cycle.

[0171] After the nth terminal device completes training, the neural network output activates the upload small cycle It can be determined as:

[0172]

[0173] Among them, for the nth terminal device, the segmentation layer is l n The divided device-side global model, is the size of the neural network output activation for a single sample, is the uplink (UL) transmission rate of the nth terminal device. Optionally, can be expressed as:

[0174]

[0175] in, is the transmission power of the nth terminal device to the base station.

[0176] After the base station receives the output and activates it, the edge server side forward and reverse propagation small cycle It can be determined as:

[0177]

[0178] Among them, when the edge server performs its server-side global model training for the nth terminal device, The number of floating-point computing units required to perform forward and backward propagation for a data sample, is the CPU frequency allocated by the edge server to the nth terminal device, q s is the number of FLOPs per cycle of the edge server.

[0179] After the edge server determines the gradient of the output activation, the gradient of the neural network output activation is downloaded in a small cycle It can be determined as:

[0180]

[0181] For the nth terminal device, the segmentation layer is l n The divided device-side global model, The gradient size of the neural network output activation corresponding to a single sample.

[0182] After receiving the gradient of the output activation, the terminal device backpropagates a small cycle It can be determined as:

[0183]

[0184] Among them, when the nth terminal device trains a data sample on its device-side global model, The number of FLOPs required to perform backpropagation for the end device.

[0185] It's important to note that the terminal device's forward propagation mini-cycle, the neural network output activation upload mini-cycle, the edge server's forward and backward propagation mini-cycle, and the neural network output activation gradient download mini-cycle are serially related to the terminal device's backward propagation mini-cycle. The collaborative training cycle of multiple terminal devices and base stations includes H such serial processes.

[0186] After the nth terminal device completes H terminal device and base station collaborative training cycles, the nth terminal device and the edge server can respectively obtain a device-side local model and a server-side local model, which can be used to determine the first local model.

[0187] As an example, the local model on the device It can be determined as:

[0188]

[0189] in, is the initial device-side global model (first sub-model) of the nth terminal device, η t is the learning rate, It is the device-side gradient formed by the n-th terminal device during the τ-th collaborative training cycle between the terminal device and the base station.

[0190] As an example, a server-side local model It can be determined as:

[0191]

[0192] in, is the initial server-side global model (second sub-model) given by the edge server to the nth terminal device, It is the server-side gradient formed by the edge server for the nth terminal device during the τth terminal device and base station collaborative training cycle.

[0193] In step S430, the terminal device transmits the device-side local model to the base station. The device-side local model upload process here occurs within the device-side local model upload cycle. As an example, the device-side local model upload cycle of the nth terminal device is It can be determined as:

[0194]

[0195] In step S440, it is determined whether the local models can be aggregated. If yes, step S450 is executed, otherwise step S420 is executed. Optionally, the base station can use specific rules to determine whether local model aggregation can be performed.

[0196] In step S450, the base station obtains a global model by weighted aggregation of local models. For example, after obtaining the device-side local model of the nth terminal device, the base station combines it with the corresponding server-side local model to obtain a local model (the first local model), and then performs weighted aggregation of the local models to obtain a global model.

[0197] In step S460, it is determined whether convergence has been achieved. If so, step S470 is executed, otherwise step S410 is executed. The base station may use specific rules to determine whether the global model has reached convergence. In some embodiments, the global model w t+1 The convergence judgment rule can be determined as:

[0198] |F(w t+1 )―F(w t )|≤ε;

[0199] Among them, F(w) is the loss function calculated based on the global model w, which can be used to measure the training effect of the global model, and ε is the preset convergence accuracy.

[0200] It should be understood that the above embodiments are merely illustrative of how to determine whether a global model has converged. In addition to the convergence determination rules mentioned above, any rule that can determine whether a global model has reached objective convergence can be applied to the embodiments of this application. The embodiments of this application are not limiting.

[0201] In step S470 , the adaptive federated segmentation learning based on semi-asynchronous local model aggregation is ended.

[0202] After determining that the global model meets convergence requirements, the base station broadcasts the device-side global model and a training termination instruction to all end devices. The end devices and edge servers stop training the device-side global model and the server-side global model, respectively. The base station and end devices can then release the communication and computing resources used for collaborative training.

[0203] Depend on Figure 3 and Figure 4It can be seen that the embodiment of the present application proposes a semi-asynchronous local model aggregation strategy for the differences in device resources and the heterogeneity of wireless channels. This strategy can reduce the waiting delay of each training round and further improve the efficiency of global model training. On the one hand, all terminal devices can perform local collaborative training in parallel with the base station to obtain a local model. On the other hand, once the preset aggregation period arrives, the edge server on the base station side does not need to wait for all terminal devices to complete the local collaborative training, and directly performs aggregation on the local model of the terminal device that has completed the local collaborative training. Terminal devices that have not completed the collaborative training can continue to train in the next training round without being terminated or interrupted. By implementing a semi-asynchronous local model aggregation strategy, the training delay overhead is significantly reduced.

[0204] The adaptive federated segmentation learning paradigm based on semi-asynchronous local model aggregation provided by the embodiment of the present application can fully utilize the powerful computing power of network equipment to assist resource-constrained terminal devices in completing the training tasks of artificial intelligence models, and maximize the use of the computing resources of the terminal devices. Therefore, the training efficiency of the global model is accelerated on the basis of ensuring the performance of the global model. In addition, the use of a semi-asynchronous local model aggregation strategy can also reduce the latency overhead caused by the global model training process and improve the aggregation efficiency.

[0205] The present application also provides an artificial intelligence model learning system. The learning system includes a network device and multiple terminal devices, wherein any of the multiple terminal devices executes the method described above for terminal devices to execute, and the network device executes the method described above for network devices to execute.

[0206] Combined with the above Figures 2 to 4 This paper introduces a method for model training based on terminal device capabilities or training duration. As previously mentioned, time-varying wireless channels and the allocation of wireless resources can also affect the quality and efficiency of global model training. For example, due to rapidly changing wireless channels, the base station cannot make the optimal global resource allocation decision for executing AI model training tasks.

[0207] Based on this, embodiments of the present application also propose a resource management method for model training. This resource management can include the management of wireless resources and the management of terminal device training resources. For example, in a time-varying wireless channel, this resource management method can ensure efficient global model training through an online management solution for wireless resources.

[0208] For ease of understanding, the following Figure 5 The resource management method proposed in the embodiment of the present application is described in detail. Figure 5The resource management method shown can be executed by a network device such as a base station, or by an edge server deployed outside the base station. Figure 2 The model training method shown in Figure 2 is as follows. For simplicity, Figure 2 Terms that have been explained in will not be repeated here.

[0209] See also Figure 5 In step S510, the model training process is decoupled into a single training round. In other words, the model training task is decoupled into a single training round task to improve training efficiency.

[0210] In some embodiments, decoupling of model training can be performed based on Lyapunov optimization theory. Lyapunov optimization theory is widely used in random networks. This theory analyzes, controls, and optimizes networks with random events, time variations, and uncertainty, demonstrating that time-averaged constrained optimization can be achieved in general random networks.

[0211] As an example, a simple Lyapunov drift plus penalty framework can be used to optimize the average latency of wireless network throughput, power, distortion, etc. to approximately make optimal decisions.

[0212] Alternatively, Lyapunov optimization can decouple multi-level stochastic optimization problems into sub-problems that are deterministic at each level, while providing theoretical guarantees for the long-term stability of the network. As an example, based on Lyapunov optimization theory, a base station can decouple the training of an AI model into a single training round.

[0213] In some embodiments, the decoupling of model training can be determined according to Lyapunov optimization theory and the second constraint condition, that is, according to Lyapunov optimization theory and the second constraint condition, the process of model training is decoupled into a single training round.

[0214] As an example, the second constraint may be related to system energy consumption. For example, the base station may utilize Lyapunov optimization theory to decouple the global model training task into a single training round of global model training tasks under the constraint of system energy consumption.

[0215] In step S520, a first constraint is determined for a single training round.

[0216] In some embodiments, the first constraint condition may be related to at least one of the following information: the duration of model training; the energy consumption of multiple terminal devices and network devices. The duration of model training is used to limit the application duration of the constraint condition. The energy consumption of multiple terminal devices and network devices is used to constrain the energy consumption of the training device for model training during the model training process. Exemplarily, the first constraint condition may be a long-term low energy consumption condition. The online management solution determined based on the first constraint condition is used to perform long-term, low-energy, adaptive, and stable segmentation collaborative training of the global model.

[0217] The first constraint condition is used to determine a first management scheme. The first management scheme can perform resource optimization on a wireless network that performs a global model training task of a single training round, thereby ensuring optimal execution of the global model training task of a single training round.

[0218] In some embodiments, the first management solution may include a solution for multiple terminal devices to perform model training, and also include an online management solution for wireless resources. For example, the base station may make an optimal online resource management solution based on the first constraint condition.

[0219] In some embodiments, the first management scheme may also be determined based on the wireless network status of a single training round. Optionally, the wireless network status of a single training round may include information such as channel status, terminal device status, and edge server status.

[0220] The first management scheme includes multiple segmentation layers corresponding to multiple terminal devices that perform model training, the multiple terminal devices including a first terminal device, and the multiple segmentation layers including a first segmentation layer corresponding to the first terminal device.

[0221] In some embodiments, the first management scheme may further include one or more of the following information: the communication bandwidth of each terminal device performing model training; the method for allocating computing frequency of the edge server; and the selection of multiple terminal devices used to determine the first aggregation period. For example, the online resource management scheme may include the allocation of communication bandwidth and computing frequency of the edge server, the selection of terminal devices participating in local model aggregation, and the selection of segmentation layers for the global model.

[0222] As an example, the allocation of communication bandwidth and edge server computing frequency can be formulated as a convex optimization problem with a single training round, focused on maintaining long-term low energy consumption in the wireless network. Furthermore, the base station or server processing unit can employ the Lagrangian duality method to iteratively solve the problem, yielding an asymptotically optimal allocation of communication bandwidth and edge server computing frequency.

[0223] Optionally, the communication bandwidth and edge server computing frequency allocation scheme may also be obtained by other methods, which is not limited in the embodiments of the present application.

[0224] As an example, the selection of terminal devices participating in local model aggregation can be formulated as an integer linear programming problem with a single training round, based on maintaining long-term low energy consumption in the wireless network. Furthermore, the processing unit of the base station or server can employ a branch-and-bound algorithm to iteratively solve the problem, resulting in an asymptotically optimal solution for selecting terminal devices participating in local model aggregation.

[0225] Optionally, the selection of terminal devices participating in local model aggregation can also be obtained by other methods, which is not limited in the embodiments of the present application.

[0226] As an example, selecting the global model's split layer can be formulated as an integer optimization problem with a single training round, focused on maintaining long-term low energy consumption in the wireless network. Furthermore, the base station or server's processing unit can employ an exhaustive search algorithm to iteratively solve the problem and find the optimal global model split layer selection solution.

[0227] Optionally, the segmentation layer selection scheme of the global model can also be obtained by other methods, which is not limited in the embodiment of the present application.

[0228] In some embodiments, the first management solution can be distributed to all terminal devices performing model training, so that multiple edge devices can collaborate and complete the training task of the global model in the current training round. For example, after determining the multiple segmentation layers corresponding to the multiple terminal devices, the network device can send the corresponding multiple device-side global models to the multiple terminal devices.

[0229] As an example, in a single training round, the terminal device and the base station perform one device-side global model download cycle, H terminal device and base station collaborative training cycles, and one device-side local model upload cycle. As mentioned above, each terminal device and base station collaborative training cycle includes a terminal device forward propagation mini-cycle, a neural network output activation upload mini-cycle, an edge server-side forward and backward propagation mini-cycle, a neural network output activation gradient download mini-cycle, and a terminal device backpropagation mini-cycle.

[0230] For ease of understanding, the following takes the resource management method for model training executed by the base station as an example, combined with Figure 6 The process of the wireless resource online management solution for time-varying wireless channels and ensuring efficient training of the global model proposed in the embodiment of the present application is schematically explained.

[0231] See also Figure 6 In step S610, the base station decouples the training task of the artificial intelligence model into a single training round task.

[0232] In step S620, the base station develops an optimal online resource management plan, specifically including steps S621 through S623. In step S621, terminal devices participating in local model aggregation are selected. In step S622, communication bandwidth and frequency are allocated to the edge server. In step S623, the global model segmentation layer is selected.

[0233] In step S630, the terminal device and the base station perform a single training round task of the global model.

[0234] In step S640, it is determined whether the global model has reached convergence. If so, step S650 is executed, otherwise step S610 is executed. Optionally, the base station may use specific rules to determine whether the global model has reached convergence. t+1 The convergence judgment rules are as described above.

[0235] In step S650, the online wireless resource management process ends. After determining that the global model meets the convergence conditions, the base station broadcasts a termination instruction to all terminal devices. The terminal devices and edge servers stop training the device-side global model and the server-side global model, respectively. The base station and terminal devices then release the wireless resources used for collaborative training.

[0236] Depend on Figure 5-Figure 6 It can be seen that the online wireless resource management solution provided by the embodiments of the present application can ensure efficient global model training in an environment with rapidly changing wireless channels. This resource management method not only achieves optimal utilization of computing and communication resources, but also accelerates the training efficiency of the global model. At the same time, maintaining long-term low energy consumption of the wireless network as a constraint meets the demand for low energy efficiency of the network in executing artificial intelligence model tasks.

[0237] Combined with the above Figures 1 to 6 , describes the method embodiment of the present application in detail. Figures 7 to 10 , the device embodiment of the present application is described in detail. It should be understood that the description of the device embodiment corresponds to the description of the method embodiment, so that parts not described in detail can be referred to the previous method embodiment.

[0238] Figure 7 7 is a schematic block diagram of a model training device according to an embodiment of the present application. The model training device 700 may be a first terminal device for model training. The first terminal device may be any of the terminal devices described above. Figure 7 The model training device 700 shown includes a receiving unit 710 and a processing unit 720 .

[0239] The receiving unit 710 may be configured to receive a first sub-model in the global model, where the first sub-model is determined according to the first segmentation layer.

[0240] The processing unit 720 can be used to train the first sub-model; wherein the global model also includes a second sub-model, and the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0241] Optionally, the first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

[0242] Optionally, the first segmentation layer is determined according to the capability level corresponding to the first terminal device, and the second terminal device that performs model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

[0243] Optionally, the first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

[0244] Optionally, the first segmentation layer is further determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and the network device.

[0245] Optionally, the training duration of the first sub-model is determined based on one or more of the following information: the position of the first segmentation layer in the global model; the capability / number of data samples of the first terminal device; the status of wireless resources between the first terminal device and the network device; and the computing frequency allocated by the network device to the first terminal device.

[0246] Optionally, when the training round in which the first local model participates in model aggregation is not the current training round, the processing unit is further configured to train the first sub-model in one or more training rounds after the current training round.

[0247] Optionally, the current training round corresponds to the first aggregation period. When the training time of the first sub-model is longer than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training time of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

[0248] Optionally, the first aggregation period is determined based on the states of N local models of N terminal devices that perform model training and a first threshold, where N is a positive integer.

[0249] Optionally, when the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

[0250] Optionally, the model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

[0251] Optionally, the multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameter for controlling the model aggregation.

[0252] Optionally, the current training round is the t-th training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1-th training round is t+1 for:

[0253]

[0254] Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices at the tth training round, ρ n,t represents the weighted coefficient of the nth terminal device in the tth training round, represents the local model of the nth terminal device at the tth training round.

[0255] Optionally, the nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for:

[0256]

[0257] Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t Denotes the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t represents the aggregation interval of the kth terminal device.

[0258] Optionally, the network device sending the first sub-model includes an edge server, which is used to determine at least one of the following information under the first constraint condition: the communication bandwidth of each terminal device performing model training; the computing frequency allocation method of the edge server; the selection of multiple terminal devices for determining the first aggregation period; and multiple segmentation layers including the first segmentation layer.

[0259] Figure 8 800 is a schematic block diagram of a model training device provided in an embodiment of the present application. The model training device 800 can be any of the network devices for model training described above. Figure 8 The model training device 800 shown includes a sending unit 810 and a processing unit 820.

[0260] The sending unit 810 can be used to send the first sub-model in the global model to the first terminal device, where the first sub-model is determined according to the first segmentation layer.

[0261] Processing unit 820 can be used to train the second sub-model in the global model; wherein, the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

[0262] Optionally, the first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

[0263] Optionally, the first segmentation layer is determined according to the capability level corresponding to the first terminal device, and the second terminal device that performs model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

[0264] Optionally, the first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

[0265] Optionally, the first segmentation layer is further determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and the network device.

[0266] Optionally, the training duration of the first sub-model is determined based on one or more of the following information: the position of the first segmentation layer in the global model; the capability / number of data samples of the first terminal device; the status of wireless resources between the first terminal device and the network device; and the computing frequency allocated by the network device to the first terminal device.

[0267] Optionally, when the training round in which the first local model participates in model aggregation is not the current training round, the processing unit is further configured to train the second sub-model in one or more training rounds after the current training round.

[0268] Optionally, the current training round corresponds to the first aggregation period. When the training time of the first sub-model is longer than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training time of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

[0269] Optionally, the first aggregation period is determined based on the states of N local models of N terminal devices that perform model training and a first threshold, where N is a positive integer.

[0270] Optionally, when the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

[0271] Optionally, the model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

[0272] Optionally, the multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameter for controlling the model aggregation.

[0273] Optionally, the current training round is the t-th training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1-th training round is t+1 for:

[0274]

[0275] Among them, 1≤n≤N, w n,t represents the local model state of the nth terminal device among N terminal devices at the tth training round, ρ n,t represents the weighted coefficient of the nth terminal device in the tth training round, represents the local model of the nth terminal device at the tth training round.

[0276] Optionally, the nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for:

[0277]

[0278] Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t Denotes the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t represents the aggregation interval of the kth terminal device.

[0279] Optionally, the network device includes an edge server, which is used to determine at least one of the following information under the first constraint condition: the communication bandwidth of each terminal device performing model training; the computing frequency allocation method of the edge server; the selection of multiple terminal devices for determining the first aggregation period; and multiple segmentation layers including the first segmentation layer.

[0280] Figure 9 Schematic block diagram of a resource management device for model training provided in an embodiment of the present application. The resource management device 900 can be any network device or edge server for model training described above. Figure 9 The resource management device 900 shown includes a first processing unit 910 and a second processing unit 920 .

[0281] The first processing unit 910 can be used to decouple the model training process into a single training round.

[0282] The second processing unit 920 can be used to determine a first constraint condition for a single training round; wherein the first constraint condition is used to determine a first management scheme, the first management scheme includes multiple segmentation layers corresponding to multiple terminal devices that perform model training, the multiple terminal devices include a first terminal device, the multiple segmentation layers include a first segmentation layer corresponding to the first terminal device, and the first segmentation layer is related to the capabilities of the first terminal device.

[0283] Optionally, the first processing unit 910 is configured to decouple the model training process into a single training round according to Lyapunov optimization theory and a second constraint.

[0284] Optionally, the first management scheme further includes one or more of the following information: the communication bandwidth of each terminal device performing model training; the computing frequency allocation method of the edge server; and the selection of multiple terminal devices for determining the first aggregation period.

[0285] Optionally, the first constraint is related to at least one of the following information: the duration of model training; and energy consumption of multiple terminal devices and network devices.

[0286] Figure 10Shown is a schematic structural diagram of a communication device according to an embodiment of the present application. Figure 10 The dotted line in the figure indicates that the unit or module is optional. The apparatus 1000 can be used to implement the method described in the above method embodiment. The apparatus 1000 can be a chip, a terminal device or a network device.

[0287] The device 1000 may include one or more processors 1010. The processor 1010 may support the device 1000 to implement the method described in the method embodiment above. The processor 1010 may be a general-purpose processor or a special-purpose processor. For example, the processor may be a CPU. Alternatively, the processor may be another general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may be any conventional processor, etc.

[0288] The apparatus 1000 may further include one or more memories 1020. The memories 1020 store programs that can be executed by the processor 1010, causing the processor 1010 to perform the methods described in the above method embodiments. The memories 1020 may be independent of the processor 1010 or integrated into the processor 1010.

[0289] The apparatus 1000 may further include a transceiver 1030. The processor 1010 may communicate with other devices or chips via the transceiver 1030. For example, the processor 1010 may transmit and receive data with other devices or chips via the transceiver 1030.

[0290] The present application also provides a computer-readable storage medium for storing a program. The computer-readable storage medium can be applied to a terminal device or network device provided in the present application, and the program enables a computer to execute the method performed by the terminal device or network device in each embodiment of the present application.

[0291] The computer-readable storage medium may be any available medium that can be read by a computer or a data storage device such as a server or data center that includes one or more available media. The available medium may be a magnetic medium, an optical medium, or a semiconductor medium. Examples of computer storage media include, but are not limited to, phase-change random access memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc-read only memory (CD-ROM), solid state disk (SSD), digital video disc (DVD) or other optical storage, magnetic cassette, tape / disk storage or other magnetic storage device, or any other non-transmission medium.

[0292] The present application also provides a computer program product. The computer program product includes a program. The computer program product can be applied to a terminal device or network device provided in the present application, and the program causes a computer to execute the method performed by the terminal device or network device in each embodiment of the present application.

[0293] In the above embodiments, it can be implemented in whole or in part by software, hardware, firmware or any combination thereof. When implemented using software, it can 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. When the computer program instructions are loaded and executed on a computer, the process or function described in the embodiment of the present application is generated in whole or in part. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium. For example, the computer instructions can be transmitted from one website, computer, server or data center to another website, computer, server or data center via wired (e.g., coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0294] The present application also provides a computer program that can be applied to a terminal device or network device provided in the present application, and enables a computer to execute the method performed by the terminal or network device in each embodiment of the present application.

[0295] The terms "system" and "network" are used interchangeably in this application. Furthermore, the terms used in this application are intended only to explain the specific embodiments of this application and are not intended to limit this application. The terms "first," "second," "third," and "fourth," etc. in the specification and claims of this application and the accompanying drawings are used to distinguish between different objects, rather than to describe a specific order.

[0296] It should be noted that the terms "comprises," "includes," and "has," and any variations thereof, are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus comprising a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, article, or apparatus. In the absence of further qualification, an element defined by the phrase "comprising a..." does not preclude the presence of other identical elements in the process, method, article, or apparatus comprising the element.

[0297] In the embodiments of this application, the term "indication" may refer to a direct indication, an indirect indication, or an indication of an association. For example, "A indicates B" may refer to a direct indication of B, e.g., B can obtain information through A; it may refer to an indirect indication of B, e.g., A indicates C, e.g., B can obtain information through C; or it may refer to an association between A and B.

[0298] In the embodiments of the present application, the term "corresponding" may indicate a direct or indirect correspondence between the two, or an association relationship between the two, or a relationship between indication and indication, configuration and configuration, etc.

[0299] In the embodiments of the present application, "pre-definition" or "pre-configuration" may be implemented by pre-storing corresponding codes, tables, or other methods that can be used to indicate relevant information in a device (e.g., including a terminal device and a network device). The present application does not limit the specific implementation method. For example, pre-definition may refer to a definition in a protocol.

[0300] In the embodiments of the present application, the "protocol" may refer to a standard protocol in the communication field, for example, it may include an LTE protocol, a NR protocol, and related protocols used in future communication systems, and this application does not limit this.

[0301] In the embodiments of the present application, determining B based on A does not mean determining B only based on A. B can also be determined based on A and / or other information.

[0302] In the embodiments of this application, the term "and / or" is simply a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0303] In the embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0304] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0305] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0306] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0307] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above method embodiments can be implemented by means of software plus the necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk), and includes a number of instructions for enabling a service classification device (which can be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods described in each embodiment of the present application. It should be noted that the serial numbers of the various implementation methods are only for description and do not represent the pros and cons of the embodiments.

[0308] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of this application. Therefore, the scope of protection of this application should be based on the scope of protection of the claims.

Claims

1. A model training method, characterized in that: include: The first terminal device receives a first sub-model in the global model, where the first sub-model is determined according to the first segmentation layer; The first terminal device trains the first sub-model; Among them, the global model also includes a second sub-model, and the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

2. The model training method according to claim 1, characterized in that: The first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

3. The model training method according to claim 1 or 2, characterized in that: The first segmentation layer is determined according to the capability level corresponding to the first terminal device. The second terminal device that executes the model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

4. The model training method according to any one of claims 1 to 3, characterized in that: The first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

5. The model training method according to any one of claims 1 to 4, characterized in that: The first segmentation layer is also determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and a network device.

6. The model training method according to any one of claims 1 to 5, characterized in that: The training duration of the first sub-model is determined according to one or more of the following information: The position of the first segmentation layer in the global model; Capabilities / number of data samples of the first terminal device; The state of the wireless resources between the first terminal device and the network device; The network device allocates a computing frequency to the first terminal device.

7. The model training method according to any one of claims 1 to 6, characterized in that: When the training round in which the first local model participates in model aggregation is not the current training round, the model training method further includes: In one or more training rounds after the current training round, the first terminal device trains the first sub-model.

8. The model training method according to any one of claims 1 to 7, characterized in that: The current training round corresponds to the first aggregation period. When the training duration of the first sub-model is greater than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training duration of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

9. The model training method according to claim 8, characterized in that: The first aggregation period is determined based on the states of N local models of N terminal devices that perform the model training and a first threshold, where N is a positive integer.

10. The model training method according to any one of claims 1 to 9, characterized in that: When the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

11. The model training method according to any one of claims 1 to 10, characterized in that: The model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

12. The model training method according to claim 11, characterized in that: The multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameters for controlling the model aggregation.

13. The model training method according to claim 11 or 12, characterized in that: The current training round is the tth training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1th training round t+1 for: Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices in the tth training round, ρ n,t represents the weight coefficient of the nth terminal device in the tth training round, Represents the local model of the nth terminal device in the tth training round.

14. The model training method according to claim 13, characterized in that: The nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for: Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t represents the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t Represents the aggregation interval of the k-th terminal device.

15. The model training method according to any one of claims 1 to 14, characterized in that: The network device sending the first sub-model includes an edge server, and the edge server is used to determine at least one of the following information under the first constraint condition: The communication bandwidth of each terminal device performing the model training; The computing frequency allocation method of the edge server; selection of a plurality of terminal devices for determining a first aggregation period; A plurality of segmentation layers including the first segmentation layer.

16. A model training method, characterized in that: include: The network device sends a first sub-model in the global model to the first terminal device, where the first sub-model is determined according to the first segmentation layer; The network device trains a second sub-model in the global model; Among them, the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

17. The model training method according to claim 16, characterized in that: The first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

18. The model training method according to claim 16 or 17, characterized in that: The first segmentation layer is determined according to the capability level corresponding to the first terminal device. The second terminal device that executes the model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

19. The model training method according to any one of claims 16 to 18, characterized in that: The first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

20. The model training method according to any one of claims 16-19, characterized in that: The first segmentation layer is also determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and the network device.

21. The model training method according to any one of claims 16-20, characterized in that: The training duration of the first sub-model is determined according to one or more of the following information: The position of the first segmentation layer in the global model; Capabilities / number of data samples of the first terminal device; A wireless resource status between the first terminal device and the network device; The network device allocates a computing frequency to the first terminal device.

22. The model training method according to any one of claims 16-21, characterized in that: When the training round in which the first local model participates in model aggregation is not the current training round, the model training method further includes: In one or more training rounds after the current training round, the network device trains the second sub-model.

23. The model training method according to any one of claims 16 to 22, characterized in that: The current training round corresponds to the first aggregation period. When the training duration of the first sub-model is greater than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training duration of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

24. The model training method according to claim 23, characterized in that: The first aggregation period is determined based on the states of N local models of N terminal devices that perform the model training and a first threshold, where N is a positive integer.

25. The model training method according to any one of claims 16-24, characterized in that: When the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

26. The model training method according to any one of claims 16-25, characterized in that: The model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

27. The model training method according to claim 26, characterized in that: The multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameters for controlling the model aggregation.

28. The model training method according to claim 26 or 27, characterized in that: The current training round is the tth training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1th training round t+1 for: Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices in the tth training round, ρ n,t represents the weight coefficient of the nth terminal device in the tth training round, Represents the local model of the nth terminal device in the tth training round.

29. The model training method according to claim 28, characterized in that: The nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for: Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and a n,t represents the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t Represents the aggregation interval of the k-th terminal device.

30. The model training method according to any one of claims 16-29, characterized in that: The network device includes an edge server, and the edge server is used to determine at least one of the following information under a first constraint: The communication bandwidth of each terminal device performing the model training; The computing frequency allocation method of the edge server; selection of a plurality of terminal devices for determining a first aggregation period; A plurality of segmentation layers including the first segmentation layer.

31. A resource management method for model training, characterized in that: include: Decoupling the process of training the model into individual training rounds; Determining a first constraint for the single training round; Among them, the first constraint condition is used to determine a first management scheme, the first management scheme includes multiple segmentation layers corresponding to multiple terminal devices that execute the model training, the multiple terminal devices include a first terminal device, the multiple segmentation layers include a first segmentation layer corresponding to the first terminal device, and the first segmentation layer is related to the capabilities of the first terminal device.

32. The resource management method according to claim 31, characterized in that: The process of decoupling the model training into a single training round includes: According to Lyapunov optimization theory and the second constraint, the model training process is decoupled into a single training round.

33. The resource management method according to claim 31 or 32, characterized in that: The first management plan also includes one or more of the following information: The communication bandwidth of each terminal device performing the model training; How the computing frequency of edge servers is allocated; A selection of a plurality of terminal devices is used to determine a first aggregation period.

34. The resource management method according to any one of claims 31 to 33, characterized in that: The first constraint condition is related to at least one of the following information: The duration of training of the model; The energy consumption of the multiple terminal devices and the network devices.

35. A model training device, characterized in that: The model training device is a first terminal device, and the first terminal device includes: A receiving unit, configured to receive a first sub-model in the global model, wherein the first sub-model is determined according to the first segmentation layer; A processing unit, configured to train the first sub-model; Among them, the global model also includes a second sub-model, and the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

36. The model training device according to claim 35, characterized in that: The first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

37. The model training device according to claim 35 or 36, characterized in that: The first segmentation layer is determined according to the capability level corresponding to the first terminal device. The second terminal device that executes the model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

38. The model training device according to any one of claims 35 to 37, characterized in that: The first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

39. The model training device according to any one of claims 35 to 38, characterized in that: The first segmentation layer is also determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and a network device.

40. The model training device according to any one of claims 35 to 39, characterized in that: The training duration of the first sub-model is determined according to one or more of the following information: The position of the first segmentation layer in the global model; Capabilities / number of data samples of the first terminal device; The state of the wireless resources between the first terminal device and the network device; The network device allocates a computing frequency to the first terminal device.

41. The model training device according to any one of claims 35 to 40, characterized in that: When the training round in which the first local model participates in model aggregation is not the current training round, the processing unit is further used to train the first sub-model in one or more training rounds after the current training round.

42. The model training device according to any one of claims 35 to 41, characterized in that: The current training round corresponds to the first aggregation period. When the training duration of the first sub-model is greater than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training duration of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

43. The model training device according to claim 42, characterized in that: The first aggregation period is determined based on the states of N local models of N terminal devices that perform the model training and a first threshold, where N is a positive integer.

44. The model training device according to any one of claims 35 to 43, characterized in that: When the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

45. The model training device according to any one of claims 35 to 44, characterized in that: The model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

46. ​​The model training device according to claim 45, characterized in that: The multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameters for controlling the model aggregation.

47. The model training device according to claim 45 or 46, characterized in that: The current training round is the tth training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1th training round t+1 for: Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices in the tth training round, ρ n,t represents the weight coefficient of the nth terminal device in the tth training round, Represents the local model of the nth terminal device in the tth training round.

48. The model training device according to claim 47, characterized in that: The nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for: Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t represents the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t Represents the aggregation interval of the k-th terminal device.

49. The model training device according to any one of claims 35 to 48, characterized in that: The network device sending the first sub-model includes an edge server, and the edge server is used to determine at least one of the following information under the first constraint condition: The communication bandwidth of each terminal device performing the model training; The computing frequency allocation method of the edge server; selection of a plurality of terminal devices for determining a first aggregation period; A plurality of segmentation layers including the first segmentation layer.

50. A model training device, characterized in that: The model training device is a network device, and the network device includes: A sending unit, configured to send a first sub-model in the global model to a first terminal device, where the first sub-model is determined according to the first segmentation layer; A processing unit, configured to train a second sub-model in the global model; Among them, the training of the first sub-model and the second sub-model are jointly used to determine the first local model; the first segmentation layer is determined according to the capabilities of the first terminal device, and / or the training duration of the first sub-model is used to determine whether the training round in which the first local model participates in model aggregation is the current training round.

51. The model training device according to claim 50, characterized in that: The first terminal device is one of a plurality of terminal devices, and the plurality of sub-models received by the plurality of terminal devices are respectively determined according to the global model and a plurality of different segmentation layers, the plurality of sub-models include the first sub-model, and the plurality of segmentation layers include the first segmentation layer.

52. The model training device according to claim 50 or 51, characterized in that: The first segmentation layer is determined according to the capability level corresponding to the first terminal device. The second terminal device that executes the model training corresponds to the third sub-model. When the capability level of the first terminal device is higher than the capability level of the second terminal device, the first sub-model is larger than the third sub-model.

53. The model training device according to any one of claims 50-52, characterized in that: The first terminal device belongs to a first terminal device set, and all terminal devices in the first terminal device set correspond to the same capability level.

54. The model training device according to any one of claims 50-53, characterized in that: The first segmentation layer is also determined according to the number of data samples of the first terminal device and / or the state of wireless resources between the first terminal device and the network device.

55. The model training device according to any one of claims 50-54, characterized in that: The training duration of the first sub-model is determined according to one or more of the following information: The position of the first segmentation layer in the global model; Capabilities / number of data samples of the first terminal device; A wireless resource status between the first terminal device and the network device; The network device allocates a computing frequency to the first terminal device.

56. The model training device according to any one of claims 50-55, characterized in that: When the training round in which the first local model participates in model aggregation is not the current training round, the processing unit is further used to train the second sub-model in one or more training rounds after the current training round.

57. The model training device according to any one of claims 50-56, characterized in that: The current training round corresponds to the first aggregation period. When the training duration of the first sub-model is greater than the duration of the first aggregation period, the first local model does not participate in the model aggregation of the current training round; when the training duration of the first sub-model is less than or equal to the duration of the first aggregation period, the first local model participates in the model aggregation of the current training round.

58. The model training device according to claim 57, characterized in that: The first aggregation period is determined based on the states of N local models of N terminal devices that perform the model training and a first threshold, where N is a positive integer.

59. The model training device according to any one of claims 50-58, characterized in that: When the state of the first local model in the current training round is 1, the first local model participates in the model aggregation of the current training round; when the state of the first local model in the current training round is 0, the first local model does not participate in the model aggregation of the current training round.

60. The model training device according to any one of claims 50-59, characterized in that: The model aggregation of the current training round is used to determine the global model of the next training round, and the global model of the next training round is determined according to multiple weighting coefficients corresponding to multiple terminal devices participating in the model aggregation.

61. The model training device according to claim 60, characterized in that: The multiple weighting coefficients are respectively determined according to the aggregation interval of the multiple terminal devices participating in the model aggregation and / or the bias parameters for controlling the model aggregation.

62. The model training device according to claim 60 or 61, characterized in that: The current training round is the tth training round among T training rounds, where T is a positive integer, 1≤t≤T, and the global model w of the t+1th training round t+1 for: Among them, 1≤n≤N, m n,t represents the local model state of the nth terminal device among N terminal devices in the tth training round, ρ n,t represents the weight coefficient of the nth terminal device in the tth training round, Represents the local model of the nth terminal device in the tth training round.

63. The model training device according to claim 62, characterized in that: The nth terminal device is a set of terminal devices S participating in model aggregation t One of the weighted coefficients ρ of the nth terminal device in the tth training round n,t for: Among them, D n represents the number of data samples of the nth terminal device, γ represents the bias parameter for controlling model aggregation, and α n,t represents the aggregation interval of the nth terminal device, D k Indicates S t The number of data samples of the kth terminal device, α k,t Represents the aggregation interval of the k-th terminal device.

64. The model training device according to any one of claims 50 to 63, characterized in that: The network device includes an edge server, and the edge server is used to determine at least one of the following information under a first constraint: The communication bandwidth of each terminal device performing the model training; The computing frequency allocation method of the edge server; selection of a plurality of terminal devices for determining a first aggregation period; A plurality of segmentation layers including the first segmentation layer.

65. A resource management device for model training, characterized in that: The resource management device comprises: A first processing unit, configured to decouple the model training process into a single training round; A second processing unit, configured to determine a first constraint condition for the single training round; Among them, the first constraint condition is used to determine a first management scheme, the first management scheme includes multiple segmentation layers corresponding to multiple terminal devices that execute the model training, the multiple terminal devices include a first terminal device, the multiple segmentation layers include a first segmentation layer corresponding to the first terminal device, and the first segmentation layer is related to the capabilities of the first terminal device.

66. The resource management device according to claim 65, characterized in that: The first processing unit is also used to decouple the model training process into a single training round according to Lyapunov optimization theory and a second constraint condition.

67. The resource management device according to claim 65 or 66, characterized in that: The first management plan also includes one or more of the following information: The communication bandwidth of each terminal device performing the model training; How the computing frequency of edge servers is allocated; A selection of a plurality of terminal devices is used to determine a first aggregation period.

68. The resource management device according to any one of claims 65-67, characterized in that: The first constraint condition is related to at least one of the following information: The duration of training of the model; The energy consumption of the multiple terminal devices and the network devices.

69. A communication device, characterized in that: It comprises a memory and a processor, wherein the memory is used to store a program, and the processor is used to call the program in the memory to execute the method as described in any one of claims 1-15 or 16-30 or 31-34.

70. A device, characterized in that Comprising a processor, configured to call a program from a memory to execute a method as described in any one of claims 1-15 or 16-30 or 31-34.

71. A chip, characterized in that: It comprises a processor, which is used to call a program from a memory, so that a device equipped with the chip executes the method as described in any one of claims 1-15 or 16-30 or 31-34.

72. A computer-readable storage medium, characterized in that A program is stored thereon, the program causing a computer to execute the method as claimed in any one of claims 1-15, 16-30 or 31-34.

73. A computer program product, characterized in that It comprises a program which causes a computer to execute the method as claimed in any one of claims 1-15 or 16-30 or 31-34.

74. A computer program, characterized in that The computer program causes a computer to execute the method according to any one of claims 1-15 or 16-30 or 31-34.

Images