A neural network model training method, device, chip and equipment

Abstract

The application discloses a neural network model training method and device, a chip and equipment. The method comprises the following steps: obtaining user behavior sample data collected by a sensor; the user behavior sample data at least comprises first sample data of a previous moment and second sample data of a next moment; taking the first sample data as input of a first neural network model to obtain a predicted label of the first sample data; determining a real label corresponding to the first sample data based on the predicted label of the first sample data and the second sample data; and training the first neural network model based on the first sample data and the real label to obtain a trained first neural network model. In this way, the first neural network model of the end-side device is fine-tuned by using the user behavior sample data, so that the trained first neural network model can take into account the use habits of the user and improve the adaptability of the model. Meanwhile, the model is trained on the end-side device, which is beneficial to protecting the privacy of the user.

Description

Technical Field

[0001] This application relates to artificial intelligence technology, and more particularly to a method, apparatus, chip, and device for training a neural network model. Background Technology

[0002] The existing model training model is "cloud-side training, edge-side inference." This means that data collected on the edge is transmitted to the cloud, the cloud uses this data to train the model, the trained model is then transmitted back to the edge, and the edge uses the trained model for direct inference. To protect user privacy, the data transmitted from the edge to the cloud does not include user-related data. Therefore, the model trained on the cloud cannot accommodate the usage habits of each user, resulting in poor adaptability. Summary of the Invention

[0003] This application aims to provide a method, apparatus, chip, and device for training neural network models.

[0004] The technical solution of this application is implemented as follows:

[0005] Firstly, a neural network model training method is provided, applied to an edge device, the method comprising:

[0006] Acquire user behavior sample data collected by sensors; wherein, the user behavior sample data includes at least the first sample data from the previous moment and the second sample data from the next moment;

[0007] The first sample data is used as the input to the first neural network model to obtain the predicted label corresponding to the first sample data;

[0008] Based on the predicted label of the first sample data and the second sample data, determine the true label corresponding to the first sample data;

[0009] The first neural network model is trained based on the first sample data and the corresponding real labels to obtain the trained first neural network model.

[0010] Secondly, a neural network model training device is provided for use in an edge device, the device comprising:

[0011] An acquisition unit is used to acquire user behavior sample data collected by sensors; wherein the user behavior sample data includes at least a first sample data from the previous moment and a second sample data from the next moment.

[0012] The processing unit is configured to take the first sample data as input to the first neural network model to obtain the predicted label corresponding to the first sample data, and to determine the true label corresponding to the first sample data based on the predicted label of the first sample data and the second sample data.

[0013] The training unit is used to train the first neural network model based on the first sample data and the corresponding real labels to obtain the trained first neural network model.

[0014] Thirdly, a modem chip is provided, comprising: a processor and a memory configured to store a computer program capable of running on the processor, wherein the processor is configured to perform the steps of the method of the first aspect when running the computer program.

[0015] Fourthly, a computer-readable storage medium is provided having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the method of the first aspect.

[0016] This application discloses a neural network model training method, apparatus, chip, and device. It utilizes user behavior sample data to fine-tune a first neural network model on a device-side interface, enabling the trained model to adapt to user habits and improve its adaptability. Furthermore, this application performs model training on a device-side interface, which helps protect user privacy. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the first process of the neural network model training method in the embodiments of this application;

[0018] Figure 2 This is a schematic diagram of tag allocation in an embodiment of this application;

[0019] Figure 3 This is a schematic diagram of the neural network model training framework in an embodiment of this application;

[0020] Figure 4 This is a schematic diagram of the second process of the neural network model training method in the embodiments of this application;

[0021] Figure 5 This is a schematic diagram of training data as an example of an embodiment of this application;

[0022] Figure 6 This is a schematic diagram of the structure of the neural network model training device in the embodiments of this application;

[0023] Figure 7 This is a schematic diagram of the modem chip structure in an embodiment of this application. Detailed Implementation

[0024] In order to gain a more detailed understanding of the features and technical content of the embodiments of this application, the implementation of the embodiments of this application will be described in detail below with reference to the accompanying drawings. The accompanying drawings are for reference and illustration only and are not intended to limit the embodiments of this application.

[0025] This application provides a method for training a neural network model. Figure 1 This is a schematic diagram of the first process of the neural network model training method in this application embodiment, applied to an edge device.

[0026] like Figure 1 As shown, the specific training method for this neural network model may include:

[0027] Step 101: Acquire user behavior sample data collected by the sensor; wherein the user behavior sample data includes at least the first sample data of the previous moment and the second sample data of the next moment.

[0028] When a user performs a corresponding operation on the device, the device's sensors collect sample data of the user's behavior under that operation. The first sample data and the second sample data can be data collected by the same sensor at different times, or data collected by different sensors at different times. Here, the first sample data from the previous moment is used to predict the user's predicted operation on the device at the next moment, while the second sample data from the next moment is the user's actual operation on the device at the next moment.

[0029] For example, in a scenario where a user unlocks the camera with a gesture and presses the shutter button to take a picture, the camera sensor on the device collects the user's gesture data, activates the camera based on the gesture data, and assumes the user wants to take a picture. Then, the pressure sensor collects the pressure data corresponding to the shutter button, and activates the shutter function based on the pressure data, indicating that the user indeed wants to take a picture. Here, the user's gesture data can be understood as the first sample data, and the pressure data as the second sample data.

[0030] Step 102: Use the first sample data as input to the first neural network model to obtain the predicted label corresponding to the first sample data.

[0031] In some embodiments, the first neural network model is distributed from the cloud-side device to the edge device, and the file types supported by the model training framework of the edge device include the file type of the first neural network model.

[0032] In other embodiments, the method further includes: acquiring a second neural network model sent by a cloud-side device; and when it is determined that the file type of the second neural network model does not belong to the file type supported by the model training framework of the edge device, converting the file type of the second neural network model to the file type of the first neural network model supported by the model training framework of the edge device using a conversion tool.

[0033] In other words, the cloud-side device sends the second neural network model to the edge device. When the file type of the second neural network model is not a file type supported by the model training framework of the edge device, it is necessary to convert the file type of the second neural network model to a file type supported by the model training framework of the edge device. Here, the file types supported by the model training framework of the edge device include multiple types, and conversion to one of them is sufficient. The converted second neural network model is the first neural network model.

[0034] Here, the model training framework of the edge device supports functions such as gradient calculation, parameter optimization, parameter quantization / quantization-aware training.

[0035] Step 103: Based on the predicted label of the first sample data and the second sample data, determine the true label corresponding to the first sample data.

[0036] It should be noted that, based on the predicted label of the first sample data of the user at the previous moment, the predicted operation of the user on the terminal device at the next moment is predicted, while the second sample data is used to reflect the actual operation of the user at the next moment. The predicted operation may be the same as the actual operation or it may be different. Here, based on the predicted label of the first sample data and the second sample data, the actual label corresponding to the first sample data is determined. In this way, the effectiveness of training the first neural network model can be guaranteed when executing step 104.

[0037] In some embodiments, step 103 specifically includes: if the predicted label is a positive sample label, and the second sample data indicates that the predicted label is a correct positive sample label, then the true label is determined to be a positive sample label; if the predicted label is a positive sample label, and the second sample data indicates that the predicted label is an incorrect positive sample label, then the true label is determined to be a negative sample label.

[0038] The predicted label is a negative sample label, the second sample data indicates that the predicted label is a correct negative sample label, and the true label is determined to be a negative sample label;

[0039] The predicted label is a negative sample label, the second sample data indicates that the predicted label is an incorrect negative sample label, and the true label is determined to be a positive sample label.

[0040] Positive samples are the samples to be tested, and negative samples are the samples other than the samples to be tested.

[0041] In other words, if the predicted label of the first sample data at the previous time step is a positive sample label, and the predicted label of the second sample data at the next time step is a correct positive sample label, then the predicted label is correct and is the true label. If the predicted label of the first sample data at the previous time step is a positive sample label, and the predicted label of the second sample data at the next time step is an incorrect positive sample label, then the predicted label is incorrect and is corrected to a negative sample label; the corrected predicted label is the true label. If the predicted label of the first sample data at the previous time step is a negative sample label, and the predicted label of the second sample data at the next time step is a correct negative sample label, then the predicted label is correct and is the true label. If the predicted label of the first sample data at the previous time step is a negative sample label, and the predicted label of the second sample data at the next time step is an incorrect negative sample label, then the predicted label is incorrect and is corrected to a positive sample label; the corrected predicted label is the true label.

[0042] Here, this application embodiment provides a label allocation diagram. Figure 2 This is a schematic diagram of tag allocation in an embodiment of this application, such as... Figure 2 As shown, tag assignment specifically includes:

[0043] Step 201: Determine whether the preset first step is detected based on the first sample data of the previous moment; if detected, determine that the predicted label of the user's first sample data of the previous moment is a positive sample label, and proceed to step 202; if not detected, determine that the predicted label of the user's first sample data of the previous moment is a negative sample label, and proceed to step 205.

[0044] Step 202: Execute step 203 if the second sample data at the next moment detects the preset second step; if detected, execute step 204 if not detected.

[0045] Step 203: Indicate that the predicted label of the first sample data is the correct positive label.

[0046] Step 204: Indicate that the predicted label of the first sample data is a false positive label.

[0047] Step 205: Execute step 206 if the preset first step is detected based on the second sample data at the next moment; if detected, execute step 207.

[0048] Step 206: Indicate the predicted label of the first sample data as a false negative label.

[0049] Step 207: Indicate that the predicted label of the first sample data is the correct negative sample label.

[0050] For example, the edge device collects the user's first gesture data (which can be understood as the first sample data) through a camera sensor. If the camera is detected to be activated, the user's first gesture data is determined to be a positive sample. Then, pressure data (which can be understood as the second sample data) is collected through a pressure sensor. If the camera function is detected to be activated, the user's first gesture data is determined to be a correct positive sample; if the camera function is not detected to be activated, the user's first gesture data is determined to be an incorrect positive sample. If the camera is not detected to be activated, the user's first gesture data is determined to be a negative sample. Then, the device continues to collect the user's second gesture data (which can be understood as the second sample data) through the camera sensor. If the camera is detected to be activated, the user's first gesture data is determined to be an incorrect negative sample; if the camera is not detected to be activated, the user's first gesture data is determined to be a correct negative sample.

[0051] Step 104: Train the first neural network model based on the first sample data and the corresponding real labels to obtain the trained first neural network model.

[0052] The model training process is the same as the existing model training process, and will not be elaborated here.

[0053] After step 104 is executed, the trained first neural network model replaces the original first neural network model. Before entering the next model training, the trained first neural network model is used to label the first sample data.

[0054] In some embodiments, the method further includes: acquiring original sample data and corresponding sample labels; wherein the original sample data is the data used when training the second neural network model on the cloud-side device; step 104 specifically includes: training the first neural network model based on the first sample data and corresponding real labels, and the original sample data and corresponding sample labels, to obtain the trained first neural network model.

[0055] Here, the first sample data and its corresponding real labels, as well as the original sample data and its corresponding sample labels used on the cloud-side device, are used as training data to train the first neural network model. This avoids the risk of model collapse caused by using only the user's first sample data for model training. At the same time, using the user's first sample data enables the trained first neural network model to adapt to the user.

[0056] In some embodiments, the data type of the original sample data and the user behavior sample data is half-precision floating-point number (fp16). Reducing the data precision shortens the training time and thus improves the training speed.

[0057] It should be noted that the training method for neural network models is executed by the application processor (AP) of the edge device. Specifically, the AP can be a graphics processing unit (GPU) or an embedded neural network processing unit (NPU), both of which have floating-point computing capabilities.

[0058] Here, the execution entity for steps 101 to 104 can be the processor of the end-side device.

[0059] By employing the above technical solution, the first neural network model on the edge device is fine-tuned using user behavior sample data, enabling the trained first neural network model to cater to user habits and improve the model's adaptability. Furthermore, since the edge device no longer needs to upload data, wait for cloud-side device scheduling tasks for model training, or download updated models, training time is shortened, and bandwidth consumption for data transmission can be eliminated.

[0060] Based on the above embodiments, this application provides a neural network model training framework. Figure 3 This is a schematic diagram of the neural network model training framework in the embodiments of this application, as shown below. Figure 3 As shown, the neural network model training framework includes cloud-side device training 30, edge-side device inference 31, and edge-side device training 32.

[0061] Cloud-side device training 30: The second neural network model is pre-trained using the original sample data and corresponding sample labels.

[0062] Edge device inference 31: Using the user's first sample data from the previous moment as input to the first neural network model, the predicted label corresponding to the first sample data is obtained. Based on the predicted label of the first sample data and the second sample data from the next moment, the true label corresponding to the first sample data is determined.

[0063] Here, if the file types supported by the model training framework of the edge device include the file type of the first neural network model, then the second neural network model is the first neural network model. If the file types supported by the model training framework of the edge device do not include the file type of the first neural network model, a conversion tool is used to convert the file type of the second neural network model to the file type of the first neural network model supported by the model training framework of the edge device; that is, the second neural network model is the converted first neural network model.

[0064] Edge device training 32: Use the first sample data and its corresponding real label, as well as the original sample data and its corresponding sample label used on the cloud device, as training data to train the first neural network model and obtain the trained first neural network model.

[0065] After that, the trained first neural network model replaced the original first neural network model. Before entering the next model training, the trained first neural network model was used to label the first sample data.

[0066] To better illustrate the purpose of this application, further examples are provided based on the above embodiments. Figure 4 This is a schematic diagram of the second process of the neural network model training method in this application embodiment, applied to an edge device, such as... Figure 4 As shown, the specific training method for this neural network model includes:

[0067] Step 401: Obtain the original sample data and corresponding sample labels, and obtain the user behavior sample data collected by the sensor; wherein, the original sample data is the data used when training the second neural network model on the cloud-side device, and the user behavior sample data includes at least the first sample data of the previous time step and the second sample data of the next time step.

[0068] Step 402: Obtain the second neural network model sent by the cloud-side device.

[0069] Step 403: When it is determined that the file type of the second neural network model is not a file type supported by the model training framework of the edge device, the file type of the second neural network model is converted to the file type of the first neural network model supported by the model training framework of the edge device using a conversion tool.

[0070] Step 404: When it is determined that the file type of the second neural network model belongs to the file type supported by the model training framework of the edge device, the second neural network model is used as the first neural network model.

[0071] Step 405: Use the first sample data as input to the first neural network model to obtain the predicted label corresponding to the first sample data.

[0072] Step 406: Based on the predicted label of the first sample data and the second sample data, determine the true label corresponding to the first sample data.

[0073] In some embodiments, step 406 specifically includes: the predicted label is a positive sample label, the second sample data indicates that the predicted label is a correct positive sample label, and the true label is determined to be a positive sample label;

[0074] The predicted label is a positive sample label, the second sample data indicates that the predicted label is an incorrect positive sample label, and the true label is determined to be a negative sample label;

[0075] The predicted label is a negative sample label, the second sample data indicates that the predicted label is a correct negative sample label, and the true label is determined to be a negative sample label;

[0076] The predicted label is a negative sample label, the second sample data indicates that the predicted label is an incorrect negative sample label, and the true label is determined to be a positive sample label.

[0077] Step 407: When the preset training conditions are met, the first neural network model is trained based on the first sample data and the corresponding real labels, and the original sample data and the corresponding sample labels, to obtain the trained first neural network model.

[0078] Here, the first sample data and its corresponding real labels, as well as the original sample data and its corresponding sample labels used on the cloud-side device, are used as training data to train the first neural network model, avoiding single-source training.

[0079] Using only the user's first sample data for model training can cause the model to fail to converge (risk of model collapse), while using the user's first sample data can make the trained first neural network model adapt to the user.

[0080] In some embodiments, the preset training conditions include: the amount of the first sample data is greater than or equal to the first preset data amount; and / or, the charging state of the end device is trickle charging or fully charged.

[0081] 5. In the first possible implementation, when the amount of data of the first sample data collected by the sensor...

[0082] The model training condition is met only when the amount of data is greater than or equal to the first preset amount (e.g., 10), which makes the trained first neural network model more effective.

[0083] In the second possible implementation, the model training conditions are met only when the device on the edge is in trickle charging or fully charged state. This reduces user perception and power consumption. Trickle charging is used to compensate for capacity loss caused by self-discharge after the battery is fully charged.

[0084] In the third possible implementation, the model training conditions are met only when the amount of the first sample data collected by the sensor is greater than or equal to the first preset amount of data, and the charging status of the end device is in trickle charging or full charge. This makes the trained first neural network model more effective, while also reducing user perception and power consumption.

[0085] 5. In some embodiments, acquiring user behavior sample data collected by sensors includes:

[0086] When the amount of the first sample data collected is greater than the second preset amount of data, the second preset amount of first sample data before the current time is obtained; wherein, the second preset amount of data is the maximum amount of first sample data required for training the neural network model.

[0087] It's important to note that during model training, a minimum training data size (i.e., the first preset data size) is typically set. Training begins when this minimum size is met. For example, if the initial sample data contains 100 data points, and the first sample data contains at least 10 data points (the first preset data size), training starts. An upper limit, or maximum data size, is also usually set. For instance, if the maximum data size is set to 1000 data points, the initial sample data contains 100 data points, and the first sample data contains more than 900 data points (the second preset data size), then the oldest collected data needs to be replaced with new data. This means acquiring the 900 most recently collected first sample data points. Using the most recently collected first sample data for model training ensures the effectiveness of the trained neural network model.

[0088] Figure 5 This is a schematic diagram of training data as an example of an embodiment of this application, such as... Figure 5 As shown, the original sample data volume is set to 100 records. The volume of the first sample data is increased from 0 to 10, so that the training data volume equals the sum of the original sample data volume and the first sample data volume, which is 110 records, and model training can begin. Then, 890 new first sample data are collected, bringing the training data volume to 1000 records, reaching the upper limit of the training data. Subsequent collection of new first sample data replaces the earliest collected data with the new data.

[0089] To implement the method of the embodiments of this application, based on the same inventive concept, a neural network model training device is also provided in the embodiments of this application. Figure 6 This is a schematic diagram of the structure of the neural network model training device in the embodiments of this application, applied to an edge device, such as... Figure 6 As shown, the neural network model training device 60 includes:

[0090] The acquisition unit 601 is used to acquire user behavior sample data collected by the sensor; wherein, the user behavior sample data includes at least the first sample data of the previous moment and the second sample data of the next moment;

[0091] The processing unit 602 is used to take the first sample data as input to the first neural network model to obtain the predicted label corresponding to the first sample data, and to determine the real label corresponding to the first sample data based on the predicted label of the first sample data and the second sample data.

[0092] The training unit 603 is used to train the first neural network model based on the first sample data and the corresponding real labels to obtain the trained first neural network model.

[0093] By employing the above technical solution, the first neural network model of the edge device is fine-tuned using user behavior sample data, enabling the trained first neural network model to take into account user habits and improve the model's adaptability. Furthermore, this application performs model training on the edge device, which helps protect user privacy.

[0094] In some embodiments, the training unit 603 is further configured to train the first neural network model based on the first sample data and the corresponding real labels when a preset training condition is met.

[0095] In some embodiments, the preset training conditions include: the amount of the first sample data is greater than or equal to the first preset data amount; and / or, the charging state of the end device is trickle charging or fully charged.

[0096] In some embodiments, the acquisition unit 601 is further configured to acquire original sample data and corresponding sample labels; wherein the original sample data is the data used when training the second neural network model on the cloud-side device; the training unit 603 is further configured to train the first neural network model based on the first sample data and corresponding real labels, and the original sample data and corresponding sample labels, to obtain the trained first neural network model.

[0097] In some embodiments, the acquisition unit 601 is further configured to acquire the second preset amount of first sample data before the current time when the amount of the first sample data collected is greater than the second preset amount of data; wherein the second preset amount of data is the maximum amount of first sample data required for training the neural network model.

[0098] In some embodiments, the processing unit 602 is specifically used to determine that the true label is a positive sample label when the predicted label is a positive sample label, the second sample data indicates that the predicted label is a correct positive sample label, and the true label is a positive sample label.

[0099] The predicted label is a positive sample label, the second sample data indicates that the predicted label is an incorrect positive sample label, and the true label is determined to be a negative sample label;

[0100] The predicted label is a negative sample label, the second sample data indicates that the predicted label is a correct negative sample label, and the true label is determined to be a negative sample label;

[0101] The predicted label is a negative sample label, the second sample data indicates that the predicted label is an incorrect negative sample label, and the true label is determined to be a positive sample label.

[0102] In some embodiments, the acquisition unit 601 is further configured to acquire the second neural network model sent by the cloud-side device; the processing unit 602 is further configured to, when determining that the file type of the second neural network model does not belong to the file type supported by the model training framework of the edge device, convert the file type of the second neural network model to the file type of the first neural network model supported by the model training framework of the edge device through a conversion tool.

[0103] This application also provides another modem chip. Figure 7 This is a schematic diagram of the modem chip structure in an embodiment of this application, as shown below. Figure 7 As shown, the chip 70 includes: a processor 701 and a memory 702 configured to store computer programs capable of running on the processor;

[0104] The processor 701 is configured to execute the method steps in the foregoing embodiments when running a computer program.

[0105] Of course, in practical applications, such as Figure 7 As shown, the various components in the chip 70 are coupled together via a bus system 703. It can be understood that the bus system 703 is used to implement communication between these components. In addition to a data bus, the bus system 703 also includes a power bus, a control bus, and a status signal bus. However, for clarity, in... Figure 7The general designated all buses as Bus System 703.

[0106] In practical applications, the aforementioned processor can be at least one of the following: Application-Specific Integrated Circuit (ASIC), Digital Signal Processing Device (DSPD), Programmable Logic Device (PLD), Field-Programmable Gate Array (FPGA), controller, microcontroller, and microprocessor. It is understood that, for different devices, the electronic devices used to implement the functions of the aforementioned processor can also be other types, and the embodiments of this application do not specifically limit this.

[0107] The aforementioned memory can be volatile memory, such as random-access memory (RAM); or non-volatile memory, such as read-only memory (ROM), flash memory, hard disk drive (HDD), or solid-state drive (SSD); or a combination of the above types of memory, and provides instructions and data to the processor.

[0108] In an exemplary embodiment, this application also provides a computer-readable storage medium for storing a computer program.

[0109] Optionally, the computer-readable storage medium can be applied to any of the methods in the embodiments of this application, and the computer program causes the computer to execute the corresponding processes implemented by the processor in the various methods of the embodiments of this application. For the sake of brevity, it will not be described in detail here.

[0110] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and 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. In addition, the coupling, direct coupling, or communication connection between the various components shown or discussed can be through some interfaces, and the indirect coupling or communication connection between devices or units can be electrical, mechanical, or other forms.

[0111] The units described above as separate components may or may not be physically separate. 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 the units may be selected to achieve the purpose of this embodiment according to actual needs.

[0112] Furthermore, in the various embodiments of the present invention, all functional units can be integrated into one processing module, or each unit can be a separate unit, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in a combination of hardware and software functional units. Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0113] The methods disclosed in the several method embodiments provided in this application can be arbitrarily combined without conflict to obtain new method embodiments.

[0114] The features disclosed in the several product embodiments provided in this application can be arbitrarily combined without conflict to obtain new product embodiments.

[0115] The features disclosed in the several method or device embodiments provided in this application can be arbitrarily combined without conflict to obtain new method or device embodiments.

[0116] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A neural network model training method, applied to an edge device, characterized in that, The method includes: Acquire user behavior sample data collected by sensors; wherein, the user behavior sample data includes at least the first sample data from the previous moment and the second sample data from the next moment; The first sample data is used as the input to the first neural network model to obtain the predicted label corresponding to the first sample data; Based on the predicted label of the first sample data and the second sample data, determine the true label corresponding to the first sample data; The first neural network model is trained based on the first sample data and the corresponding real labels to obtain the trained first neural network model. The method further includes: acquiring original sample data and corresponding sample labels; wherein, the original sample data is the data used when training the second neural network model on the cloud-side device; the step of training the first neural network model based on the first sample data and corresponding real labels includes: training the first neural network model based on the first sample data and corresponding real labels, and the original sample data and corresponding sample labels, to obtain the trained first neural network model. The first neural network model is obtained based on the second neural network model.

2. The method according to claim 1, characterized in that, The step of training the first neural network model based on the first sample data and the corresponding real labels includes: When the preset training conditions are met, the first neural network model is trained based on the first sample data and the corresponding real labels.

3. The method according to claim 2, characterized in that, The preset training conditions include: the amount of the first sample data is greater than or equal to the first preset data amount; and / or, the charging state of the end device is trickle charging or full charge.

4. The method according to claim 1, characterized in that, The acquisition of user behavior sample data collected by sensors includes: When the amount of the first sample data collected is greater than the second preset amount of data, the second preset amount of first sample data before the current time is obtained; wherein, the second preset amount of data is the maximum amount of first sample data required for training the neural network model.

5. The method according to claim 1, characterized in that, The step of determining the true label corresponding to the first sample data based on the predicted label of the first sample data and the second sample data includes: The predicted label is a positive sample label, the second sample data indicates that the predicted label is a correct positive sample label, and the true label is determined to be a positive sample label; The predicted label is a positive sample label, the second sample data indicates that the predicted label is an incorrect positive sample label, and the true label is determined to be a negative sample label; The predicted label is a negative sample label, the second sample data indicates that the predicted label is a correct negative sample label, and the true label is determined to be a negative sample label; The predicted label is a negative sample label, the second sample data indicates that the predicted label is an incorrect negative sample label, and the true label is determined to be a positive sample label.

6. The method according to claim 1, characterized in that, The method further includes: Obtain the second neural network model sent by the cloud-side device; When it is determined that the file type of the second neural network model does not belong to the file type supported by the model training framework of the edge device, the file type of the second neural network model is converted to the file type of the first neural network model supported by the model training framework of the edge device using a conversion tool.

7. A neural network model training device, applied to an edge device, characterized in that, The apparatus is used to perform the method according to any one of claims 1 to 6.

8. A modem chip, comprising a processor configured to perform the method of any one of claims 1 to 6.

9. An end-side device comprising the modem chip as claimed in claim 8.

Images