CN110874550A - Data processing method, apparatus, device and system - Google Patents

Abstract

The application discloses a data processing method, which comprises the following steps: the edge device obtains an initial neural network model that includes at least one neuron. The edge device inputs the data to be processed into the trained neural network model to obtain result data; the result data is obtained by processing data to be processed by using neurons in the trained neural network model, the trained neural network model is obtained by training a pruning neural network model by using N groups of training samples, and the pruning neural network model is obtained by pruning at least one neuron in the initial neural network model according to respective activation information of at least one neuron. Therefore, the accuracy of data processing is improved, the calculation amount is reduced, and the waste of calculation resources is avoided.

Description

Data processing method, apparatus, device and system

technical field

The present invention relates to the field of computer technology, and in particular, to a data processing method, apparatus, device and system.

Background technique

With the development of deep learning technology, especially the popularization of convolutional neural network, it is widely used in fields such as image processing and face recognition. At present, considering the versatility of the model, the design cost and other reasons, a general neural network model is usually designed in the same application scenario for data processing.

In the video surveillance scenario, the actual surveillance environment is not considered, such as the height and angle of the camera installation. If the same neural network model is used to process images collected by different cameras during data processing, it will affect the accuracy of image processing. In addition, considering the versatility of the neural network model, it is designed with many neurons and many network layers, so that there will be a waste of computing resources in some specific scenarios.

SUMMARY OF THE INVENTION

The present application discloses a data processing method, device, device and system, which can propose a neural network model suitable for the edge device itself or the current scene, so as to perform corresponding data processing, thereby improving the accuracy of data processing and avoiding the need for Waste of computing resources.

In a first aspect, the present application discloses a data processing method, comprising: obtaining a trained neural network model by an edge device, and inputting data to be processed into the trained neural network model to obtain result data of the data to be processed. The trained neural network model is obtained by using N groups of training samples to train the pruned neural network model. The pruned neural network model is obtained after pruning the initial neural network model according to the activation information of each neuron in the initial neural network model.

Considering the large amount of computation for model training and the limited computing resources of edge devices, neural network model training is usually performed on the data center side. In other words, the data center may perform pruning processing on the neurons in the initial neural network model according to the respective activation information of each neuron in the initial neural network model to obtain a pruned neural network model. Further, the data center uses N groups of training samples to train the pruned neural network model to obtain the trained neural network model. Furthermore, the data center sends the trained neural network model to the edge device, so that the edge device can perform data processing based on the trained neural network model.

The activation information of the neuron refers to the relevant information generated by the neuron when the neuron in the initial neural network model is used for data processing. The activation information includes, but is not limited to, activation value, activation times, and average activation value, which will be described in detail below.

By implementing the above process, a neural network model suitable for the edge device itself (or the scene where the edge device is currently located) can be proposed, and then the neural network model is used for data processing, which can improve the accuracy of data processing. In addition, compared with the general neural network model in the prior art, the model scale can be reduced, the data processing rate can be improved, and the waste of computing resources can be avoided.

In a possible implementation, each set of training samples includes input data and output data. The output data is obtained by processing the input data by using the initial neural network model. Specifically, the edge device obtains the initial neural network model from the data center. Then, collect multiple sets of input data of the own device or the current scene of the own device, and then use the initial neural network model to process the multiple sets of input data to obtain corresponding output data. Further, each set of input data and output data is regarded as a set of training samples, so that multiple sets of training samples can be obtained. Send multiple sets of training samples to the data center, so that the data center can use the multiple sets of training samples to retrain the initial neural network model to obtain a trained neural network suitable for the edge device (or the current scene of the edge device). network model.

Optionally, when using the initial neural network model for data processing, the activation information of each neuron in the initial neural network model can also be recorded each time.

By implementing the above steps, the edge device can obtain a customized neural network model after training, the neural network model after training can be adapted to the training samples of itself or the scene in which it is located, and more accurate results can be obtained. Moreover, the training samples do not need manual annotation, which can reduce user operations and improve the accuracy of neural network model training.

In a possible implementation manner, the activation information includes at least one of the following: activation value, activation times, and average activation value. Among them, the activation value is the output value of the neuron each time the neuron in the initial neural network model is used for data processing. The number of activations is the number of times that the activation value of the neuron is less than or equal to the preset threshold (the fourth threshold) during data processing using the neurons in the initial neural network model for many times. The average activation value is the average value of the activation value of the neuron in the data processing using the neurons in the initial neural network model for many times. Correspondingly, the process that the data center performs pruning processing on the initial neural network model specifically includes at least one of the following: when the activation information includes an activation value, the data center is based on the activation of at least one neuron in the initial neural network model. value, delete (prune) neurons whose activation value is less than or equal to the first threshold in the initial neural network model. When the activation information includes activation times, the data center deletes neurons whose activation times are less than or equal to the second threshold in the initial neural network model according to the activation times of at least one neuron in the initial neural network model. When the activation information includes an average activation value, the data center deletes neurons whose average activation value is less than or equal to the third threshold in the initial neural network model according to the respective average activation value of at least one neuron in the initial neural network model.

By implementing the above process, the data center can prune and delete neurons in the model itself according to the activation information of each neuron in the initial neural network model to obtain a pruned neural network model. It is convenient for subsequent training of the pruned neural network model to obtain a trained neural network model adapted to the edge device or the scene in which it is located. In this way, the model scale can be reduced, the waste of computing resources can be reduced, and the efficiency of data processing can be improved.

In a possible implementation manner, the trained neural network model may be obtained by updating the parameters in the pruned neural network model by using a loss function in the data center. The loss function is used to indicate the error loss between the training data and the prediction data, where the training data is the data output before inputting the input data into the fully connected layer obtained in the initial neural network model. The prediction data is the data output before inputting the input data into the fully connected layer obtained in the pruned neural network model. Specifically, the data center can use N groups of training samples to train the pruned neural network model for multiple times. In each training process, the loss function is used to modify the parameters of the model, and the one with the smallest loss function value is selected as the neural network after training. parameters of the network model to obtain the trained neural network model.

By implementing the above process, the data center can use the loss function to train the trained neural network model with higher accuracy, which is convenient for the subsequent edge device side to directly use the trained neural network model for data processing, thereby improving the accuracy of data processing. Spend.

In a second aspect, the present application provides a data processing apparatus, the apparatus comprising functional modules or units for executing the method described in the first aspect or any possible implementation manner of the first aspect.

In a third aspect, the present application provides an edge device (such as a smart camera, a roadside monitoring device, etc.), including a processor, a memory, a communication interface and a bus; the processor, the communication interface, and the memory communicate with each other through the bus; the communication interface, for receiving and sending data; a memory for storing instructions; and a processor for invoking the instructions in the memory to execute the method described in the first aspect or any possible implementation manner of the first aspect.

In a fourth aspect, the present application provides a data processing system, including a data center and an edge device; wherein, the data center is used to store an initial neural network model, and use N groups of training samples to train the initial neural network model, to obtain the trained neural network model. The edge device includes a processor, a memory, a communication interface and a bus; the processor, the communication interface, and the memory communicate with each other through the bus; the communication interface is used to receive and send data; the memory is used to store instructions; to execute the method described in the first aspect or any possible implementation manner of the first aspect.

In a fifth aspect, the present application provides a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, when the computer-readable storage medium runs on a computer, the computer executes the method described in the first aspect.

In a sixth aspect, the present application provides a computer program product comprising instructions which, when run on a computer, cause the computer to perform the methods described in the above aspects.

On the basis of the implementation manners provided by the above aspects, the present application may further combine to provide more implementation manners.

Description of drawings

FIG. 1 is a schematic structural diagram of a YOLO model provided by an embodiment of the present invention.

FIG. 2 is a schematic diagram of a network framework of a data processing system provided by an embodiment of the present invention.

FIG. 3 is a schematic flowchart of a data processing method provided by an embodiment of the present invention.

4A-4B are schematic structural diagrams of two network layers provided by an embodiment of the present invention.

FIG. 5 is a schematic structural diagram of a data processing apparatus provided by an embodiment of the present invention.

FIG. 6 is a schematic structural diagram of an edge device provided by an embodiment of the present invention.

Detailed ways

The technical solutions in the embodiments of the present invention will be described in detail below with reference to the accompanying drawings of the present invention.

First, some technical terms involved in the present invention are introduced.

Edge devices refer to devices installed on the edge network side. For example, in a video surveillance scenario, the edge device may specifically be a surveillance device or a smart camera installed on the road.

A neural network model refers to a complex network system formed by interconnecting a large number of simple processing units (called neurons). The neural network model is described based on the mathematical model of neurons, and has the capabilities of large-scale parallelism, distributed storage and processing, self-adaptation and self-learning. See below for a description of neurons. In practical applications, the neural network model is composed of at least one of the following network layers: convolutional layer, activation layer, pooling layer, and fully connected layer, etc. The number of each network layer deployed is not limited in the present invention, which can be for one or more. Each network layer consists of one or more neurons, and the parameters (weights) of the neurons can also be referred to as model parameters of the neural network model. Neurons in each network layer can be connected to each other, or neurons in different network layers can be connected to each other, as shown in FIG. 4A below, neurons in two adjacent network layers can be connected to each other.

Neurons, also known as nodes. Each neuron represents a specific output function, also known as the excitation function. When using the neural network for data processing, the essence is to use the neurons in the neural network model to perform data processing, that is, to use the excitation function of the neuron to perform data processing. Understandably, due to different neural network models, the excitation function of the same neuron in different neural network models may also be different, which is not limited in the present invention.

Convolutional layer refers to the network layer used for feature extraction. In the convolution layer, a convolution operation can be performed on the input data to extract the deep-level feature data of the input data. Taking the input data as an image, the image is input into the convolution layer, and the convolution layer is used to perform the convolution operation on the image to obtain the hidden and deep feature image of the image.

Pooling layer refers to the network layer used for data compression. The activation layer refers to the activation of the input data, and its essence is to perform the operation of a specific function to enhance the nonlinear expression ability of the data. The fully connected layer refers to the network layer that acts as a classifier in the neural network model. Each neuron (node) in the fully connected layer is connected with the neurons in the previous network layer to connect the previous network layer. The feature data output by the neurons is combined to obtain the final output result of the neural network model.

In the present invention, neural network models include but are not limited to convolutional neural network models, recurrent neural network models, deep neural network models, feedforward neural network models, deep belief network models, generative adversarial network models, and other neural network models. For ease of description, the following embodiments of the present invention take the YOLO model in the convolutional neural network model as an example for further description. Among them, the YOLO model is a network model used for target detection in the convolutional neural network model, for example, to detect target objects included in the image, such as vehicles, dogs, etc. The relevant embodiments involved in the YOLO model are briefly introduced below.

FIG. 1 is a schematic diagram of a logical structure of a neural network model provided by an embodiment of the present invention, and the neural network model may also be called a YOLO model. As shown in the figure, the YOLO model includes 9 network layers, of which the first 7 network layers (shown as layers 1-7) are convolutional layers, and the last 2 network layers (shown as layers 8 and 9) Both are fully connected layers. The number of neurons included in each network layer may be different.

Activation information refers to relevant information when a neuron is activated, such as the activation value of the neuron, the number of activations, and the average activation value. The activation of neurons here refers to the use of neurons in the neural network model to perform data operations (data calculations), and it is also considered that the neurons are activated for use.

The activation value refers to the output value of the neuron when it is activated. In other words, it refers to the output value of the neuron obtained when the neuron is used to perform data operations on the input data.

The number of activations refers to the number of times that the output value of the neuron is greater than the preset threshold when the neuron is activated multiple times. In other words, it refers to the number of times that the activation value (output value) of each neuron is greater than the preset threshold when the neuron is used for data calculation multiple times. The preset threshold is a user or system self-defined setting. Specifically, the preset threshold value may be obtained by the user or the system according to a series of experimental data statistics, or an empirical value set according to actual experience, etc., which is not limited in the present invention.

The average activation value refers to the average value of the output value of the neuron when the neuron is activated multiple times. In other words, it refers to the average value of the activation values of the neuron obtained for multiple times when the neuron is used for data calculation.

Exemplarily, in an application scenario of image recognition, the edge device is a smart camera. The edge device can periodically collect multiple images in the current scene, and input each image into the neural network model for processing to obtain corresponding result data. The resulting data is used to indicate the category to which the image belongs, eg forest, beach, starry sky. Wherein, every time an image is processed by using the neural network model, the essence is to use the neurons in the neural network model to process the image. Correspondingly, in multiple image processing, the edge device will use the neurons in the neural network model to process the multiple images multiple times. In each image processing process, the edge device will record the activation information of the neuron, such as activation value, activation times, and average activation value, etc., so as to facilitate the subsequent use of the activation information of the neuron to update the neural network model, which will be described in detail below. described.

Loss function, also known as cost function. Refers to the optimization function of the neural network model, which is used to measure the degree of error of the prediction. The essence of the training process of the neural network model is the process of finding the minimum value of the loss function, that is, the process in which the difference (also called the loss value) between the predicted result and the actual result is the smallest, or the predicted result is the closest to the actual result. Understandably, in the actual neural network model training process, the data center obtains the model parameters with the smallest value of the loss function by training the model parameters in the neural network model multiple times, as the model parameters in the trained neural network model. , so as to obtain the trained neural network model.

For example, taking the neural network model as an image recognition model as an example, assuming that the training samples for training the image recognition model are multiple images, each image consists of one or more pixels, and each pixel corresponds to its own pixel. value (that is, the real data entered). Each image includes a target object, and the pixel value of the image area where the target object is located is different from the pixel values of other areas in the image. In other words, the target object can be distinguished or identified by the pixel value of the image area where the target object is located.

Correspondingly, the data center inputs multiple images into the image recognition model, and can perform inference calculation on the pixels in each image to obtain the prediction data of the pixels. Further, the data center will also use a preset loss function to compare the predicted data of a pixel in multiple images with the real data of the pixel to correct the model parameters in the image recognition model. Repeat the above operation, the data center can find the one with the smallest difference between the predicted data and the real data of the pixels in the multiple images, and use the model parameters of the image recognition model trained (corrected) this time as the image after training. The model parameters of the recognition model are obtained, thereby also obtaining the trained image recognition model.

FIG. 2 is a schematic diagram of a network framework of a data processing system according to an embodiment of the present invention. As shown, the data processing system 100 includes a data center 102 and edge devices 104 . Wherein, the data center 102, also known as the cloud, includes multiple servers, which are deployed with a trained neural network model (specifically, the initial neural network model or the trained neural network model below), to For download and use by edge devices 104.

In practical applications, considering that the data processing process will lead to high energy consumption requirements and low computing power of edge devices, the training of the neural network model is usually completed in the data center 102 .

The training process of the neural network model is briefly described below. Specifically, the data center obtains an initial neural network model (a neural network model to be trained or to be optimized) and a training sample set. Further, the initial neural network model is trained multiple times using the training sample set. In the multiple training process, the one with the smallest loss function value is selected, and the model parameters of the neural network model generated by this training are obtained as the model parameters of the trained neural network model, so as to obtain the trained neural network model. The training of the neural network model will be described in detail below in the present invention.

Wherein, the initial neural network model can be determined according to the actual application scenario. Exemplarily, taking an image recognition scene as an example, the initial neural network model may be a convolutional neural network model, such as the YOLO model mentioned above. Taking the speech recognition scene as an example, the initial neural network model may be a deep neural network model, which may include one or more network layers, and neurons in two adjacent network layers may be connected to each other, which is not limited in the present invention. .

The training sample set includes N groups of training samples, and each group of training samples is used to train the initial neural network model to obtain a trained neural network model. Each set of training samples includes input data and output data, and the input data and output data correspond one-to-one. N is a positive integer. In different application scenarios, the training samples (ie input data and output data) may be different.

For example, in an image recognition scenario, suppose an edge device wants to use a neural network model to identify images that include the sun at different times. Correspondingly, when selecting a training sample set, the input data should include sun images at different time periods in a day, and the output data should include the position of the sun in each sun image, or other time periods used to identify the sun in the sun image. In this way, the neural network model obtained by training can accurately identify the time period classification of the sun in the image.

Correspondingly, after obtaining the above-mentioned training sample set (that is, multiple sets of training samples including input data and output data), the data center can use multiple sun images in the training sample set and the real time period of the sun in each sun image, The initial neural network model is trained to obtain the trained neural network model. Specifically, the data center can input each sun image into the initial neural network model, and use the neural network model to process the pixels in each sun image to identify the sun according to the position of the sun in the sun image or other identification methods. The characteristic information of the time period is used to predict and obtain the predicted time period of the sun in the sun image. Further, the data center uses a preset loss function to calculate the difference between the actual time period in which the sun is located in the sun image and the predicted time period. By repeating the above operation process many times, the data center can determine the model parameters of the neural network model with the smallest difference value, and use them as model parameters of the trained neural network model, thereby obtaining the trained neural network model.

In the present invention, the data center uses training samples including input data and output data to train the neural network model. Specifically, the initial neural network model is trained in an unsupervised and intelligent training manner without human participation. Compared with the traditional technology, the use of artificially labeled input data to realize the training of neural network models can avoid errors caused by human participation, thereby improving the accuracy of neural network model training without human participation, and can also improve the neural network. Ease of model training.

For example, in an image recognition application scenario, in the process of training a neural network model with traditional techniques, an image with an annotation is usually selected as a training sample. vehicles, etc. It is convenient to directly use the labeled images to train the initial neural network model.

However, in the present invention, the data center chooses to use unlabeled images as training samples for training the initial neural network model. Specifically, the training sample includes input data (here, an image) and output data (here, an object included in the image). The output data is the image as the input data of the initial neural network model, and the pixel points in the image are calculated by using the initial neural network model, so as to calculate the pixel points in the image according to the feature information of the object included in the image (for example, the contour feature of the object). , object identification, etc.) to predict the objects included in the image. It can be seen that the present invention uses unmarked training samples to train the neural network model, which can improve the convenience of training the neural network model; at the same time, avoid errors caused by human marking, and can improve the accuracy of the neural network model.

The edge device 104 can obtain the initial neural network model or the trained neural network model from the data center 102, so as to use the obtained neural network model to perform corresponding data processing. In different application scenarios, the actual data processing using the acquired neural network model is also different.

Exemplarily, in an image recognition scenario, the input data is an image, and the image includes feature information of a vehicle to be recognized, such as an identifier of the vehicle, a contour of the vehicle, and other features used to identify the vehicle. Correspondingly, the edge device can input the image into the trained neural network model, and obtain the vehicle included in the image according to the feature information of the vehicle in the image. For another example, in a speech recognition scenario, the input data is the speech to be recognized. Correspondingly, the edge device can input the speech to be recognized into the trained neural network model, and recognize and process the characteristic information such as frequency or wavelength included in the speech to be recognized, so as to obtain the text information corresponding to the speech to be recognized, etc. Not limited.

Optionally, the above training sample set (N groups of training samples) may be collected by the edge device 104 and then sent to the data center 102, so that the data center 102 can use the training sample set to The neural network model is trained. Specifically, the edge device 104 may first obtain the initial neural network model from the data center 102 . Then, the edge device can acquire N sets of input data through its own device or other edge devices, and input each set of input data into the initial neural network model to obtain corresponding output data. Based on this principle, the edge device can obtain a training sample set including N groups of training samples, and each group of training samples includes input data and output data, which will not be repeated here.

Understandably, in different application scenarios, the types of input data or output data may be different, which may include but are not limited to image data, voice data, text data, etc. Repeat. Specifically, when the edge device has the data collection capability, the edge device can collect corresponding input data through its own device. Taking an image processing scenario as an example, the edge device has a camera, and the edge device can use the camera to collect one or more images of the current scene to use it as input data. On the contrary, when the edge device does not have the data collection capability, the edge device can collect the corresponding input data through other edge devices with the data collection capability. Further, the edge device obtains the corresponding training samples according to the obtained input data, for details, please refer to the relevant descriptions in the foregoing embodiments, which will not be repeated here.

In the present invention, the number of edge devices 104 is not limited, and it may be one or more, and the figure shows n as an example, where n is a positive integer. The edge device may specifically be a device such as a smart camera, a smart camera, a smart phone, a tablet computer, a handheld computer, a notebook, and the like, which is not limited in the present invention.

The following example illustrates a specific implementation process for the edge device to obtain the training sample set. Taking the image classification application scenario as an example, the edge device is a monitoring device. Correspondingly, the edge device can collect multiple images in the current scene as input data for the neural network model. Further, the edge device can obtain the initial neural network model from the data center, and use the initial neural network model to process multiple input images as input data to obtain the classification to which each input image belongs. For example, the input image is People images, star images, seascape images, beach images, and more. Correspondingly, the edge device may take each input image and the category to which the input image belongs as a set of training samples, so as to obtain a training sample set including multiple sets of training samples. Optionally, the edge device can send the training sample set to the data center, so that the data center can use the training sample set to retrain the initial neural network model, which will not be repeated here.

Next, relevant embodiments of the data processing method involved in the present invention are introduced. Please refer to FIG. 3 , which is a schematic flowchart of a data processing method provided by an embodiment of the present invention. The method shown in Figure 3 includes the following implementation steps:

Step S301, the edge device obtains an initial neural network model from the data center.

In the present invention, the initial neural network model may also be referred to as a full model, which may be a general model obtained by training a data center using a training sample set in advance, and the general model is suitable for all or part of application scenarios. For example, the YOLO model commonly used in the field of object detection, etc.

Step S302, the edge device obtains N groups of input data, respectively inputs the N groups of input data into the initial neural network model, and obtains the output data of the N groups of input data and the respective activation information of each neuron in the initial neural network model.

When the edge device uses the initial neural network model for data processing (inference calculation), it can collect the corresponding input data. In different application scenarios, the input data may be different, for example, it may be image data, text data, voice data, etc. For details, please refer to the relevant descriptions in the foregoing embodiments, which will not be repeated here. Further, the edge device can input the input data into the initial neural network model for calculation, and obtain corresponding output data. Based on this principle, the edge device uses the initial neural network model to perform N inference calculations. That is to say, the edge device can collect N groups of input data, and use the initial neural network model N times to process the N groups of input data to obtain output data corresponding to each of the N groups of input data.

Optionally, the edge device uses a set of input data and output data obtained each time as a set of training samples. When using N sets of input data to calculate the initial neural network model, the edge device can obtain N sets of training samples, each set of training samples includes input data and output data, and the output data is the output data using the initial neural network model. Input data to process the obtained data. Among them, each time the edge device obtains an output data, the initial neural network model needs to be used.

Optionally, each time the edge device uses the initial neural network model to process a set of input data, it essentially uses the neurons in the initial neural network model to perform computation on the input data. During the calculation process, the edge device may also record the activation information generated when each neuron in the initial neural network model performs calculation (is activated). In other words, each time the edge device performs data computation using each neuron in the initial neural network model, it can record the individual activation information of each neuron. The activation information includes, but is not limited to, any one or a combination of the following: activation value, activation times, and average activation value. For the related introduction of the activation information, reference may be made to the foregoing embodiments, and details are not repeated here.

Specifically, each time the edge device uses the neurons in the initial neural network model to calculate the input data, it can record the respective activation values of each neuron in the initial neural network model. When the edge device uses the neurons in the initial neural network model for data processing multiple times (for example, M times), the edge device can count and record the respective activation times and the average activation value of each neuron in the initial neural network model. and other information. Wherein, the number of activations may be the number of times that the M activation values of the neuron are greater than or equal to the preset threshold when the neurons in the initial neural network model are used for data processing, as the activation times of the neuron. M is a positive integer less than or equal to N. The average activation value may be the average value of M activation values of the neuron when the neuron in the initial neural network model is used for data processing M times, as the average activation value of the neuron.

Step S303: The edge device sends the N groups of training samples and the activation information of each neuron in the initial neural network model to the data center.

Accordingly, the data center receives N sets of training samples and the respective activation information of each neuron. The training sample includes input data and output data. The output data is the data obtained by taking the input data as the input of the initial neural network model, and calculating the data. For the input data and output data, please refer to the relevant descriptions in the foregoing embodiments, which are not repeated here. Repeat.

The edge device sends the N groups of training samples and the respective activation information of each neuron to the data center, so that the data center can use these information to retrain the initial neural network model to obtain a model suitable for the edge device (or the edge device deployment). scene) neural network model. In this way, it is convenient to train their own neural network models for different edge devices, so as to meet the actual needs of different edge devices and improve the practicability of model processing. In other words, the present invention can retrain a personalized neural network model according to edge devices or edge device deployment scenarios to meet the real-time requirements of edge devices.

Step S304, the data center performs pruning processing on the neurons in the initial neural network model according to the respective activation information of each neuron to obtain a pruned neural network model.

In the present invention, the data center deletes neurons satisfying any one or more of the following conditions according to the activation information of each neuron in the initial neural network model, so as to obtain a corresponding pruned neural network model. Specifically, the conditions include:

1) The activation value of the neuron is less than or equal to the first threshold;

2) The number of activations of neurons is less than or equal to the second threshold;

3) The average activation value of the neuron is less than or equal to the third threshold.

In this way, the neurons that are not applicable to the edge device or the deployment scenario of the edge device are easily trimmed, so as to reduce the scale of the model and the amount of computation, thereby saving computing resources.

The first threshold, the second threshold and the third threshold may be specifically set by the user or the system, and they may be the same or different, which is not limited in the present invention. Exemplarily, if the system wants to obtain a pruned neural network model with higher computational accuracy, the above three thresholds can be set to be relatively large, such as 5 and so on. Conversely, if the system wants to obtain a pruned neural network model with lower computational accuracy, the above three thresholds can be set smaller, such as 0.01 and so on. Optionally, the above three thresholds may specifically be obtained by the system according to a series of statistical data, or experience values set by the user according to actual experience, etc., which are not limited in the present invention.

For example, FIG. 4A shows a schematic structural diagram of two network layers in the YOLO model. As shown in FIG. 4A , the Nth network layer (referred to as the Nth layer, specifically, can be any one of layers 1 to 8 in the YOLO model in FIG. 1 ) includes 6 neurons, which are O _n1 , O _n2 , ... _On6 . Layer N+1 includes 4 neurons, which are O _(n+1)1 , O _(n+1)2 , ... O _(n+1)4 . The neurons in the two adjacent layers are fully connected, that is, each neuron in the Nth layer is connected to all the neurons in the N+1th layer. During the pruning process, it is assumed that the activation values of the 6 neurons in the Nth layer are 0, 1, 0, 0, 1, and 1 respectively; the activation values of the 4 neurons in the N+1th layer are 1, 0, 1, 1 respectively. Correspondingly, the data center adopts the pruning and deletion of neurons whose activation value is less than or equal to 0, and the pruned neural network model as shown in Fig. 4B can be obtained. That is, the data center deletes the neurons _On1 , _On3 and _On4 in the Nth layer, and deletes the neurons O _(n+1)2 in the N+1th layer, so as to obtain as shown in FIG. 4B including A pruned neural network model with two network layers.

Step S305, the data center trains the pruned neural network model according to the N groups of training samples, and obtains the trained neural network model.

In the present invention, in the process of training the pruned neural network model, the data center will use the loss function to update the model parameters in the pruned neural network model to obtain the trained neural network model. The loss function is used to indicate the error loss between the training data and the prediction data, the training data is the data obtained by inputting the input data into the first neural network model, and the prediction data is inputting the input data into the second neural network model data obtained in .

The first neural network model is a network model other than the preset classification algorithm in the initial neural network model. The second neural network model is a network model other than the preset classification algorithm in the pruning neural network model. The preset classification algorithm is an algorithm or rule used in the model to calculate the output result, such as the image classification rule softmax and so on.

In practical applications, the preset classification algorithm is usually designed in the fully connected layer of the model. When the preset classification algorithm is designed in the fully connected layer, the training data may specifically be the data output before the fully connected layer obtained by inputting the input data into the initial neural network model. The above-mentioned prediction data may specifically be the data output before the fully connected layer obtained by inputting the input data into the pruning neural network model.

For example, referring to the YOLO model shown in FIG. 1, the initial neural network model is the first YOLO model, the pruning neural network model is the second YOLO model, and the first YOLO model and the second YOLO model both include 7 Convolutional layer and 2 fully connected layers, but the neurons included in each network layer can be different. Correspondingly, the training data in this example may be the data output from the last convolutional layer (ie, the seventh convolutional layer, or a network layer before the fully-connected layer) obtained by inputting the input data into the first YOLO model. The prediction data here may be the data output from the seventh convolutional layer (ie, the last convolutional layer) obtained by inputting the input data into the second YOLO model.

Understandably, in order to ensure the accuracy of model training, the data center can use N groups of training samples to train the pruned neural network model multiple times to obtain a trained neural network model with higher accuracy. During the model training process, the data center will use the preset loss function to modify the model parameters in the model to obtain the trained neural network model. Among them, the model training process is essentially a process in which the data center continuously calculates the value of the loss function, and selects the model parameter corresponding to the minimum value of the loss function as the model parameter of the neural network model after training, so as to obtain the post-training model parameters. neural network model. It is understandable that in different neural network models, the specific expression of the loss function may be different, and the present invention will be described below with an example.

Step S306, the data center updates the initial neural network model to the trained neural network model.

Step S307, the edge device obtains the trained neural network model from the data center.

Step S308: The edge device acquires the data to be processed, inputs the data to be processed into the trained neural network model, and obtains result data corresponding to the data to be processed.

The data center can save the trained neural network model. Optionally, the data center can replace the initial neural network model with a trained neural network model, that is, update the initial neural network model to a trained neural network model for the edge device to download and use.

Optionally, the data center can train each edge device to obtain a trained neural network model suitable for its own device or its own scene according to the training principle of the above-mentioned neural network model. After the data center trains the respective trained neural network models suitable for the at least one edge device for the at least one edge device, the data center may use the trained neural network model and the edge device (specifically, the identifier of the edge device). ) is associated and saved to identify which edge device the trained neural network model is for personalized training, or which edge device is suitable for application.

Correspondingly, the edge device can obtain the trained neural network model corresponding to the edge device from the data center according to actual needs. Specifically, the edge device may send an acquisition request to the data center, where the acquisition request carries the identifier of the edge device (eg, device name, device ID number, etc.), and the acquisition request is used to request to acquire the trained neural network adapted by the edge device network model. After receiving the acquisition request, the data center queries the trained neural network model corresponding to the identifier of the edge device according to the identifier of the edge device in the request, and sends it to the edge device.

Correspondingly, the edge device side can receive the trained neural network model sent by the data center, so as to facilitate subsequent data processing based on the trained neural network model. For example, an edge device can obtain data to be processed through its own device or other devices, and input the data to be processed into the trained neural network model for processing to obtain corresponding result data, which is used to indicate that the data to be processed corresponds to the result of. Wherein, in different application scenarios, the data to be processed and the result data are different.

Exemplarily, in an image classification scenario, the data to be processed is an image to be classified, and the image is composed of at least one pixel. Correspondingly, the edge device can input the image to be classified into the neural network model after training, and use each neuron in the neural network model after training to calculate each pixel in the image to be classified to obtain the image to be classified. The result data corresponding to the classified image. The result data is used to indicate the category to which the to-be-categorized image belongs, for example, it can be a person image, a starry sky image, a beach image, a forest image, and the like.

For another example, in a speech recognition scenario, the data to be processed may be speech to be recognized. Correspondingly, the edge device can input the speech to be recognized into the trained neural network model, and use each neuron in the trained neural network model to calculate the speech to be recognized to obtain result data corresponding to the speech to be recognized. The result data here may be text information corresponding to the speech to be recognized. In other words, speech-to-text translation can be achieved using a trained neural network model.

In order to facilitate the understanding of the content of the solution of the present invention, the YOLO model for vehicle detection is described in detail below.

First, the data center can collect surveillance images of different traffic intersections, and use these surveillance images to train to obtain the initial YOLO model. The edge device can obtain the initial YOLO model from the data center according to actual needs, so as to facilitate subsequent retraining of the model based on its own or its own scene. Further, the edge device can collect vehicle images within a preset period of time (for example, one month, etc.) in the current scene, and input the vehicle image into the initial YOLO model for processing to identify the characteristic information of the vehicle in the vehicle image. (e.g. vehicle identification, license plate number, vehicle outline, etc.). Further, the edge device obtains the target vehicle included in the vehicle image according to the feature information of the vehicle in the vehicle image. At the same time, the activation value of each neuron in the initial YOLO model can also be recorded. The edge device also sends the vehicle image, the target vehicle included in the vehicle image, and the activation value of each neuron in the initial YOLO model to the data center.

Correspondingly, the data center deletes (prunes) neurons with an activation value of 0 in the initial YOLO model according to the activation value of each neuron to obtain a corresponding pruned neural network model. Further, the data center uses the vehicle image and the target vehicle included in the vehicle image as training samples, and uses multiple sets of training samples to retrain the pruning neural network model. Specifically, the data center can use the loss function of the following formula (1) to adjust the model parameters of the pruned neural network model to obtain the trained YOLO model.

Among them, loss is the loss function. i is the pixel point i included in the vehicle image. t is the total number of pixels constituting the vehicle image, in other words, the vehicle image includes t pixels. P _i is the data output after processing the pixel i by using the pruning neural network model that removes the fully connected layer (specifically, the classification rule softmax deployed in the fully connected layer). F _i is the data output after processing pixel i by using the initial YOLO model that removes the fully connected layer (specifically, the classification rule softmax deployed in the fully connected layer).

The data center trains and adjusts the model parameters in the pruned neural network model according to the loss function shown in the above formula (1). Optionally, in order to ensure the accuracy of the model, a set of model parameters with the smallest loss function value is found during the training process as the model parameters of the YOLO model after training, so as to obtain the YOLO model after training. The present invention is not limited or detailed here. Optionally, the data center can associate and save the identity of the edge device and the trained YOLO model, so that the edge device side can subsequently obtain the trained YOLO model from the data center according to actual needs.

Correspondingly, the edge device can obtain the trained YOLO model from the data center, which is convenient for subsequent vehicle detection using the trained YOLO model. Exemplarily, it is assumed that the edge device is a monitoring device deployed on a road, and the edge device can acquire an image to be processed, and the image to be processed includes the vehicle to be detected. Correspondingly, the edge device takes the image to be processed as the input of the YOLO model after training, and uses the neurons of each network layer in the YOLO model after training to calculate each pixel point of the image to be processed, and identifies the image to be processed. The characteristic information of the vehicle to be detected (such as vehicle identification, license plate number, vehicle outline, etc.) is obtained, and then the vehicle to be detected is known, so that vehicle detection can be realized conveniently and efficiently.

By implementing the embodiments of the present invention, different trained neural network models can be trained for different edge devices or deployment scenarios of edge devices to form customized neural network models, and edge devices can perform data analysis based on the customized neural network models. And processing, can improve the accuracy and processing efficiency. Moreover, in the embodiment of the present invention, it is not necessary to manually mark the collected data, and the neural network model of the data center analyzes and processes the sample data sent by each edge device to obtain processed data, which further improves the processing efficiency of each edge device. . Since each edge device can process data according to the customized neural network model, the processing results can be obtained in a short time, which reduces the time-consuming of data processing.

The related embodiments of the data processing method provided by the embodiments of the present invention are described in detail above with reference to FIG. 1 to FIG. 3 . The data processing apparatus, device, and system provided by the embodiments of the present invention will be described below with reference to FIG. 5 to FIG. 6 .

Please refer to FIG. 5 , which is a schematic structural diagram of a data processing apparatus provided by an embodiment of the present invention. The data processing apparatus 500 shown in FIG. 5, applied to the edge device side, may include an acquisition module 501 and a processing module 502; wherein,

The obtaining module 501 is configured to obtain an initial neural network model, where the initial neural network model includes at least one neuron, and the neuron is used to process the input data of the initial neural network model to obtain: activation information of the neuron;

The processing module 502 is used to input the data to be processed into the trained neural network model to obtain result data;

Wherein, the result data is obtained by using the neurons in the trained neural network model to process the data to be processed, and the trained neural network model is obtained by using N groups of training samples to prune neural networks. obtained by training a network model, the pruned neural network model is obtained after pruning at least one neuron in the initial neural network model according to the activation information of the at least one neuron, and the The activation information is the respective information of the at least one neuron when the data processing is performed by using the at least one neuron in the initial neural network model, and N is a positive integer.

In a possible implementation manner, the training sample includes input data and output data, wherein the output data is obtained by computing the input data using the initial neural network model.

In a possible implementation manner, the activation information includes at least one of the following: activation value, activation times, and average activation value; The pruning process of at least one neuron in the network model includes: when the activation information includes an activation value, according to the respective activation value of the at least one neuron, the activation value in the initial neural network model is less than or equal to The neuron corresponding to the first threshold is deleted; when the activation information includes the activation times, according to the respective activation times of the at least one neuron, the activation times in the initial neural network model are less than or equal to the second threshold. The corresponding neuron is deleted; when the activation information includes an average activation value, according to the respective average activation value of the at least one neuron, the average activation value in the initial neural network model is less than or equal to the third. The neuron corresponding to the threshold is deleted; wherein, the activation value is the output value of the neuron each time the neuron in the initial neural network model is used for data processing, and the activation times is M times of use The number of times that the activation value of the neuron in the initial neural network model is greater than or equal to the fourth threshold in data processing, and the average activation value is M times using the neural network in the initial neural network model. The average value of the activation value of the neuron in the data processing, M is a positive integer less than or equal to N.

In a possible implementation manner, the trained neural network model is also obtained by using a loss function to update parameters in the pruned neural network model; wherein the loss function is used to indicate the training data and The error loss between prediction data, the training data is the data output before the full connection layer obtained by inputting the input data in the training sample into the initial neural network model, and the prediction data is the The input data in the training sample is input into the pruning neural network model, and the data obtained before the fully connected layer is obtained.

In a possible implementation manner, the obtaining module 501 is specifically configured to obtain an initial neural network model from a data center.

It should be understood that the apparatus 500 in this embodiment of the present invention may be implemented by an application-specific integrated circuit (ASIC), or a programmable logic device (PLD), and the PLD may be a complex program logic device ( complex programmable logical device (CPLD), field-programmable gate array (FPGA), generic array logic (GAL) or any combination thereof. When the data processing method shown in FIG. 3 can also be implemented by software, the apparatus and its respective modules can also be software modules.

The data processing apparatus 500 provided in this embodiment of the present invention can be correspondingly used to execute the method provided by the foregoing embodiment of the present invention, and the functions and/or other operations performed by each module in the apparatus 500 are respectively for executing the flow of the corresponding method in FIG. 3 above. The steps, for the sake of brevity, are not repeated here.

By implementing the embodiments of the present invention, different trained neural network models can be designed for different edge devices or deployment scenarios of edge devices. It is convenient for subsequent data processing using a neural network model suitable for the edge device or the deployment scenario of the edge device, which improves the accuracy of data processing.

FIG. 6 is a schematic structural diagram of an edge device provided by an embodiment of the present invention. The edge device 600 shown in FIG. 6 may include one or more processors 601, a communication interface 602, and a memory 603. The processor 601, the communication interface 602, and the memory 603 may be connected by a bus, or by other means such as wireless transmission. achieve communication. The embodiment of the present invention takes the connection through the bus 604 as an example, wherein the memory 603 is used for storing instructions, and the processor 601 is used for executing the instructions stored in the memory 503 . The memory 603 stores program codes, and the processor 601 can call the program codes stored in the memory 603 to perform the following operations:

obtaining an initial neural network model, the initial neural network model includes at least one neuron, and the neuron is used to process the input data of the initial neural network model to obtain activation information of the neuron;

Input the data to be processed into the trained neural network model to obtain the result data;

Wherein, the result data is obtained by using the neurons in the trained neural network model to process the data to be processed, and the trained neural network model is obtained by using N groups of training samples to prune neural networks. obtained by training a network model, the pruned neural network model is obtained after pruning at least one neuron in the initial neural network model according to the activation information of the at least one neuron, and the The activation information is the respective information of the at least one neuron when the data processing is performed by using the at least one neuron in the initial neural network model, and N is a positive integer.

Optionally, in this embodiment of the present invention, the processor 601 may call program codes stored in the memory 603 to execute all or part of the steps described in the method embodiment described in FIG. 3 above, and/or other content described in the text Wait, I won't go into details here.

It should be understood that the processor 601 may be constituted by one or more general-purpose processors, such as a central processing unit (Central Processing Unit, CPU). The processor 601 can be used to run the programs of the following functional modules in the related program codes. The functional module may specifically include, but is not limited to, functional modules such as the acquisition module and/or the processing module described above. That is, the processor 601 executing the program code may function as any one or more of the above-mentioned functional modules. For details about each functional module mentioned here, reference may be made to the relevant descriptions in the foregoing embodiments, which will not be repeated here.

Communication interface 602 may be a wired interface (eg, an Ethernet interface) or a wireless interface (eg, a cellular network interface or using a wireless local area network interface) for communicating with other modules/devices. For example, the communication interface 602 in this embodiment of the present application may be specifically configured to receive an initial neural network model or a trained neural network model sent by the data center.

The memory 603 may include a volatile memory (Volatile Memory), such as a random access memory (Random Access Memory, RAM); the memory may also include a non-volatile memory (Non-Volatile Memory), such as a read-only memory (Read-Only Memory) ROM), flash memory (Flash Memory), hard disk (Hard Disk Drive, HDD) or solid-state drive (Solid-State Drive, SSD); the memory 603 may also include a combination of the above-mentioned types of memory. The memory 603 may be used to store a set of program codes, so that the processor 601 can call the program codes stored in the memory 603 to implement the functions of the above-mentioned functional modules involved in the embodiments of the present invention.

It should be understood that the edge device 600 according to the embodiment of the present invention may correspond to the data processing apparatus 500 shown in FIG. 5 in the embodiment of the present invention, and may correspond to the edge device in executing the method shown in FIG. 3 according to the embodiment of the present invention The device side is the operation steps of the execution main body, and the above steps and other operations and/or functions of each module in the edge device are to implement the corresponding processes of each method in FIG. 3 , and are not repeated here for brevity.

It should be noted that FIG. 6 is only a possible implementation manner of the embodiment of the present invention. In practical applications, the edge device may further include more or less components, which is not limited here. For content not shown or described in the embodiments of the present invention, reference may be made to the relevant descriptions in the embodiments described in the foregoing FIG. 1 to FIG. 3 , and details are not repeated here.

By implementing the embodiments of the present invention, different trained neural network models can be designed for different edge devices or deployment scenarios of edge devices. It is convenient for subsequent data processing using a neural network model suitable for the edge device or the deployment scenario of the edge device, which improves the accuracy of data processing.

An embodiment of the present invention further provides a data processing system, where the data processing system includes the data center 102 and the edge device 104 as shown in FIG. 2 above. The initial neural network model or the trained neural network model is deployed in the data center 102 . The edge device includes a processor, a memory, a communication interface and a bus; the processor, the communication interface, and the memory communicate with each other through the bus; the communication interface is used to receive and send data; the memory is used to store instructions; The instructions in the memory execute all or part of the implementation steps described in the method embodiment shown in FIG. 3 , and details are not repeated here.

The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination. When implemented in software, the above-described embodiments may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer may be a general purpose computer, special purpose computer, computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer readable storage medium to another computer readable storage medium, for example, the computer instructions may be downloaded from a website site, computer, server or data center Transmission to another website site, computer, server, or data center is by wire (eg, coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server, a data center, or the like that contains one or more sets of available media. The usable media may be magnetic media (eg, floppy disks, hard disks, magnetic tapes), optical media (eg, DVDs), or semiconductor media. The semiconductor medium may be a solid state drive (SSD).

The above descriptions are merely specific embodiments of the present invention. Those skilled in the art can think of changes or substitutions according to the specific embodiments provided by the present invention, which should be included within the protection scope of the present invention.

Claims (12)

1. a data processing method, is characterized in that, described method comprises: The edge device acquires an initial neural network model, the initial neural network model includes at least one neuron, and the neuron is used to process the input data of the initial neural network model to obtain the activation of the neuron information; The edge device inputs the data to be processed into the trained neural network model to obtain result data; Wherein, the result data is obtained by using the neurons in the trained neural network model to process the data to be processed, and the trained neural network model is obtained by using N groups of training samples to prune neural networks. obtained by training a network model, the pruned neural network model is obtained after pruning at least one neuron in the initial neural network model according to the activation information of the at least one neuron, and the The activation information is the respective information of the at least one neuron when the data processing is performed by using the at least one neuron in the initial neural network model, and N is a positive integer. 2 . The method according to claim 1 , wherein the training sample includes input data and output data, wherein the output data is obtained by calculating the input data using the initial neural network model. 3 . . 3. The method according to claim 1 or 2, wherein the activation information comprises at least one of the following: activation value, activation times and average activation value; The performing pruning processing on at least one neuron in the initial neural network model according to the respective activation information of the at least one neuron includes: When the activation information includes an activation value, according to the respective activation value of the at least one neuron, delete the neuron whose activation value is less than or equal to the first threshold in the initial neural network model; When the activation information includes the activation times, according to the respective activation times of the at least one neuron, delete the neurons whose activation times are less than or equal to the second threshold in the initial neural network model; When the activation information includes an average activation value, according to the respective average activation value of the at least one neuron, delete the neuron corresponding to the average activation value in the initial neural network model that is less than or equal to the third threshold; Wherein, the activation value is the output value of the neuron every time the neuron in the initial neural network model is used for data processing, and the activation times is M times using the neuron in the initial neural network model The number of times that the activation value of the neuron is greater than or equal to the fourth threshold in the data processing of the neuron, and the average activation value is M times using the neuron in the initial neural network model to perform the neuron in the data processing. The average value of the activation values, M is a positive integer less than or equal to N. 4. The method according to any one of claims 1-3, wherein the trained neural network model is obtained by training the pruned neural network model using N groups of training samples, comprising: The trained neural network model is obtained by using a loss function to update the parameters in the pruned neural network model based on N groups of training samples; Wherein, the loss function is used to indicate the error loss between the training data and the prediction data, and the training data is inputting the input data in the training sample into the initial neural network model, before the fully connected layer obtained The output data, the prediction data is the data output before the fully connected layer obtained by inputting the input data in the training sample into the pruning neural network model. 5. The method according to claims 1-4, wherein the acquisition of the initial neural network model by the edge device comprises: The edge device obtains the initial neural network model from the data center. 6. An edge device, characterized in that the edge device comprises an acquisition module and a processing module; wherein, The obtaining module is configured to obtain an initial neural network model, where the initial neural network model includes at least one neuron, and the neuron is used to process the input data of the initial neural network model to obtain the activation information of the neuron; The processing module is used to input the data to be processed into the trained neural network model to obtain the result data; Wherein, the result data is obtained by using the neurons in the trained neural network model to process the data to be processed, and the trained neural network model is obtained by using N groups of training samples to prune neural networks. obtained by training a network model, the pruned neural network model is obtained after pruning at least one neuron in the initial neural network model according to the activation information of the at least one neuron, and the The activation information is the respective information of the at least one neuron when the data processing is performed by using the at least one neuron in the initial neural network model, and N is a positive integer. 7 . The edge device according to claim 6 , wherein the training sample includes input data and output data, wherein the output data is obtained by calculating the input data using the initial neural network model. 8 . of. 8. The edge device according to claim 6 or 7, wherein the activation information comprises at least one of the following: activation value, activation times and average activation value; The performing pruning processing on at least one neuron in the initial neural network model according to the respective activation information of the at least one neuron includes: When the activation information includes an activation value, according to the respective activation value of the at least one neuron, delete the neuron whose activation value is less than or equal to the first threshold in the initial neural network model; When the activation information includes the activation times, according to the respective activation times of the at least one neuron, delete the neurons whose activation times are less than or equal to the second threshold in the initial neural network model; When the activation information includes an average activation value, according to the respective average activation value of the at least one neuron, delete the neuron corresponding to the average activation value in the initial neural network model that is less than or equal to the third threshold; Wherein, the activation value is the output value of the neuron every time the neuron in the initial neural network model is used for data processing, and the activation times is M times using the neuron in the initial neural network model The number of times that the activation value of the neuron is greater than or equal to the fourth threshold in the data processing of the neuron, and the average activation value is M times using the neuron in the initial neural network model to perform the neuron in the data processing. The average value of the activation values, M is a positive integer less than or equal to N. 9. The edge device according to any one of claims 6-8, wherein the trained neural network model is also obtained by using a loss function to update parameters in the pruned neural network model ; Wherein, the loss function is used to indicate the error loss between the training data and the prediction data, and the training data is inputting the input data in the training sample into the initial neural network model, before the fully connected layer obtained The output data, the prediction data is the data output before the fully connected layer obtained by inputting the input data in the training sample into the pruning neural network model. 10. The edge device according to any one of claims 6-9, wherein, The obtaining module is specifically used to obtain the initial neural network model from the data center. 11. An edge device, comprising a memory and a processor coupled with the memory; the memory is used for storing instructions, and the processor is used for executing the instructions; wherein, the processor executes the The operation steps of the method according to any one of the preceding claims 1-5 are performed when instructed. 12. A data processing system, comprising a data center and an edge device, wherein the data center is used to store the initial neural network model and train the initial neural network model to Obtain the trained neural network model; the edge device is the edge device according to any one of claims 6-9 above; or, the edge device according to claim 11 above.

Images