WO2021016932A1 - Data processing method and apparatus, and computer-readable storage medium - Google Patents

Abstract

Disclosed are a data processing method and apparatus, and a computer-readable storage medium. The method comprises: determining an error degree of an initial neural network on the basis of target training data and a real value corresponding to the target training data; then if the error degree does not fall within a preset range, adjusting parameters in the initial neural network on the basis of the error degree; next, on the basis of a specified data conversion algorithm, carrying out data conversion on the adjusted parameters; after the data conversion is completed, continuing to train the initial neural network on the basis of the target training data and the real value corresponding to the target training data until a loss degree falls within a preset range; and finally, determining the initial neural network of which the loss degree falls within the preset range as a target neural network. Thus, in a subsequent application process, the data conversion processing does not need to be carried out on the parameters, such that the processing effect of the neural network on data to be processed can be guaranteed.

Description

Data processing method, device and computer readable storage medium Technical field

This specification belongs to the field of artificial intelligence, and particularly relates to a data processing method, device and computer-readable storage medium.

Background technique

At present, in order to improve the computational efficiency of neural networks, when the neural network is used to process the data to be processed, the data involved in the processing process, for example, the data to be processed and the parameters in the neural network, are fixed-pointed. Further, in order to achieve fixed-point processing, the expression range of the data is generally set in advance, and if the data exceeds this range, the data is truncated.

In the prior art, in order to further optimize the processing process of the neural network, data conversion algorithms are often introduced. Specifically, when the neural network is used to process the data to be processed, the specified data conversion algorithm is used to perform the processing of the data and parameters to be processed. Data conversion to optimize the calculation operations involved in the processing.

However, the operation of data conversion processing on parameters based on the specified data conversion algorithm may cause the parameters to exceed the data expression range, which in turn will cause the parameters to be truncated, which reduces the processing accuracy of the neural network and reduces the use of the neural network to process the data. Treatment effect.

Summary of the invention

This specification provides a data processing method, device, and computer-readable storage medium to solve the problem that the processing effect of the neural network on the data to be processed is reduced due to the introduction of a specified data conversion algorithm.

In order to solve the above technical problems, this manual is implemented as follows:

In the first aspect, an embodiment of this specification provides a data processing method, which includes:

Determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data;

If the degree of error is not within the preset range, adjust the parameters in the initial neural network based on the degree of error;

Perform data conversion processing on the adjusted parameters based on a designated data conversion algorithm;

After completing the data conversion process, continue training the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the preset range;

The initial neural network whose loss degree is within the preset range is determined as the target neural network.

In the second aspect, the embodiments of this specification provide a data processing method, which includes:

Perform fixed-point processing on the data to be processed to obtain fixed-point processing data;

Perform data conversion processing on the fixed-point processing data based on a designated data conversion algorithm to obtain converted data;

Based on the pre-training parameters in the target neural network, the converted data is processed to obtain a processing result; wherein, the target neural network is generated using the method in the first aspect.

In the third aspect, an embodiment of this specification provides a data processing device, which includes:

The first determining module is configured to determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data;

An adjustment module, configured to adjust parameters in the initial neural network based on the degree of error if the degree of error is not within a preset range;

The conversion module is configured to perform data conversion processing on the adjusted parameters based on a specified data conversion algorithm;

The training module is used to continue training the initial neural network based on the target training data and the true value corresponding to the target training data after the data conversion processing is completed, until the loss degree is within the expected Set within

The second determining module is used to determine the initial neural network whose loss degree is within the preset range as the target neural network.

In a fourth aspect, an embodiment of this specification provides a data processing device, which includes:

The first processing module is used to perform fixed-point processing on the data to be processed to obtain fixed-point processing data;

The conversion module is configured to perform data conversion processing on the fixed-point processing data based on a specified data conversion algorithm to obtain converted data;

The second processing module is configured to process the converted data based on the pre-training parameters in the target neural network to obtain a processing result; wherein the target neural network is generated by using the device in the third aspect.

In a fifth aspect, the embodiments of the present specification provide a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the first aspect or the second aspect is implemented. The steps of the data processing method.

In a sixth aspect, the embodiments of this specification provide a movable platform with limited computing power, and the movable platform is used to implement the steps of the data processing method in the first aspect or the second aspect.

In a seventh aspect, an embodiment of this specification provides an image acquisition device including a processor, and the image acquisition device is configured to implement the steps of the data processing method in the first aspect or the second aspect.

In an eighth aspect, the embodiments of this specification provide an unmanned aerial vehicle that is used to implement the steps of the data processing method in the first aspect or the second aspect.

In a ninth aspect, the embodiments of the present specification provide a handheld stabilized PTZ which is used to implement the steps of the data processing method in the first aspect or the second aspect.

In the embodiment of this specification, the error degree of the initial neural network can be determined based on the target training data and the true value corresponding to the target training data. Then, if the error degree is not within the preset range, the error degree of the initial neural network is determined based on the error degree. The parameters are adjusted, and then, based on the specified data conversion algorithm, the adjusted parameters are processed for data conversion. After the data conversion process is completed, based on the target training data and the true value corresponding to the target training data, continue to train the initial neural network. Until the loss degree is within the preset range, finally, the initial neural network with the loss degree within the preset range is determined as the target neural network. In the embodiment of this specification, the data conversion process of the parameters is simulated in advance during the generation process of the target neural network, so that the finally generated parameters in the target neural network can achieve the effect after the data conversion process, so that the subsequent In the application process, only the data to be processed need to be processed for data conversion, without the need for data conversion processing for parameters, which can avoid the problem of reduced accuracy due to truncated parameters, and realize the introduction of specified data conversion algorithms while ensuring The processing effect of the data to be processed by the neural network.

Description of the drawings

Figure 1 is a data processing method provided by an embodiment of this specification;

Figure 2 is another data processing method provided by an embodiment of this specification;

Figure 3 is a flow chart of the steps of a data processing method provided by an embodiment of this specification;

FIG. 4 is a flowchart of the steps of another data processing method provided by an embodiment of this specification;

Figure 5 is a block diagram of a data processing device provided by an embodiment of this specification;

Figure 6 is a block diagram of a data processing device provided by an embodiment of this specification;

FIG. 7 is a schematic diagram of the hardware structure of a terminal for implementing each embodiment of this specification;

FIG. 8 is a block diagram of a computing processing device provided by an embodiment of this specification;

Fig. 9 is a block diagram of a portable or fixed storage unit provided by an embodiment of the specification.

Detailed ways

The technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this specification.

Fig. 1 is a data processing method provided by an embodiment of this specification. As shown in Fig. 1, the method may include:

Step 101: Determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data.

In the embodiments of this specification, the initial neural network can be built based on the data processing that needs to be performed. The data processing that needs to be performed is different, and the structure of the initial neural network can be different. Accordingly, the target training data and the target training data correspond to The true value of can be different. For example, the initial neural network is a neural network used to enhance the sharpness of the image. The target training data can be a sample image, and the true value corresponding to the target training data can be the same as the sample image. The image content and definition is greater than the real image of the sample image. Taking the initial neural network as a neural network used to classify images as an example, the target training data may be a sample image, and the true value corresponding to the target training data may be the true category of the sample image. Further, the initial neural network can include multiple levels, different parameters can be defined in different levels, the parameters in each level can be initial parameters set randomly, and each level can perform different operations. For example, The initial neural network may include an input layer, a feature extraction layer, a fully connected layer, an output layer, and so on. Further, the initial neural network can be a convolutional neural network, where the convolutional neural network is a deep feedforward artificial neural network, mainly used in the field of image processing, and the artificial neurons of the convolutional neural network can respond to a part of the coverage The surrounding unit inside contains a large number of convolution operations when processing the image.

Further, when determining the initial degree of error based on the target training data and the true value corresponding to the target training data, the target training data can be first input into the initial neural network, and accordingly, each level in the initial neural network will respond to the target in turn The training data is processed, and finally a predicted value is output. For example, taking the level as the convolutional layer as an example, the level can be the convolution processing on the processing results of the target training data output by the previous level. Taking the level as the output layer as an example, the level can be the target training The final processing result of the data, that is, the predicted value, is output.

Then, based on the preset loss function, the predicted value, and the true value corresponding to the target training data, the error degree of the initial neural network can be calculated. Specifically, the predicted value and the true value corresponding to the target training data can be input into the predicted value. Set the loss function, and then do a gradient operation on the loss function, calculate the gradient value of the loss function, and get the error degree of the initial neural network. The gradient value can be calculated by using the back propagation algorithm, and the error degree can be Indicates the degree of deviation between the predicted value of the initial neural network survival and the true value corresponding to the target sample data.

Step 102: If the degree of error is not within a preset range, adjust the parameters in the initial neural network based on the degree of error.

In the embodiments of this specification, the preset range can be set according to actual application scenarios and actual requirements, and the embodiments of this specification do not limit it. Further, if the degree of error is within the preset range, it can be considered that the degree of deviation between the predicted value and the true value corresponding to the target sample data is sufficiently small. At this time, it can be considered that the processing capacity of the initial neural network can meet the demand. The neural network can process the data correctly. On the contrary, if the degree of error is not within the preset range, it can be considered that the deviation between the predicted value and the true value corresponding to the target sample data is large. At this time, it can be considered as the initial The processing capacity of the neural network cannot meet the demand, and the initial neural network cannot process the data correctly. Therefore, the parameters in the initial neural network can be adjusted based on the degree of error. Further, the adjustment can be made based on the stochastic gradient descent method. Specifically, the product of the gradient value and the preset step length can be calculated first, and then the difference between the parameter and the product can be calculated, and then the adjusted parameter can be obtained. In this way, the method of adjusting the parameters based on the stochastic gradient descent method can reduce the error degree of the neural network toward the direction of maximum convergence, thereby increasing the convergence speed of the neural network model, and thereby improving the efficiency of generating the neural network.

Of course, other parameter adjustment methods can also be used to adjust the parameters, for example, directly calculating the difference between the parameter and the preset fixed value as the adjusted parameter, which is not limited in the embodiment of this specification.

Step 103: Perform data conversion processing on the adjusted parameters based on the designated data conversion algorithm.

In the embodiment of this specification, the specified data conversion algorithm may be an algorithm that needs to be introduced in the application process, and the specified data conversion algorithm may be an algorithm that can improve the calculation effect. For example, the specified data conversion algorithm may be based on the remainder theorem Data conversion algorithm, or, the specified data conversion algorithm can also be a data conversion algorithm based on Lagrangian interpolation theorem, or, the specified data conversion algorithm can also be a data conversion algorithm based on the remainder theorem and based on Lagrangian interpolation Theorem's data conversion algorithm.

Further, since the parameters in the target neural network are often adjusted continuously during the process of generating the target neural network, in this step, the parameters can be pre-transformed during the generation process, so that even if the parameters are truncated The problem will not affect the accuracy of the final neural network model. At the same time, through the data conversion processing of the parameters in the process of generating the target neural network, the parameters in the final generated target neural network can be achieved The effect after the data conversion processing. In this way, in the subsequent application process, there is no need to perform data conversion processing on the parameters, thereby avoiding the problem of reduced accuracy due to the truncation of the parameters during the application process.

Step 104: After completing the data conversion process, continue training the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the preset range Inside.

In the embodiment of this specification, since the parameters in the initial neural network have been updated in the foregoing steps, the training of the initial neural network can be continued based on the target training data and the true values corresponding to the target training data. The training process can start from determining the degree of error of the initial neural network, and re-execute the above processing process. In this way, in the process of repeated training, the parameters in the initial neural network are continuously adjusted to improve the processing capacity of the initial neural network until the loss of the initial neural network falls within the preset range during a certain round of training. , You can stop training.

Step 105: Determine the initial neural network whose loss degree is within the preset range as the target neural network.

In the embodiments of this specification, continuous training is used to reduce the error degree of the initial neural network. Further, if the loss degree is within a preset range, it can be considered that the processing capacity of the initial neural network can meet the requirements, and the initial neural network can be correct. The data is processed accordingly, so the initial neural network can be determined as the target neural network. Furthermore, the generated target neural network can be deployed on a common platform for application. For example, the platform can be a simplified instruction set computer microprocessor (Advanced RISC Machines, ARM), digital signal processor (Digital Signal Processing, DSP) ), Field-Programmable Gate Array (FPGA), Graphics Processing Unit (GPU), etc.

In summary, the data processing method provided by the embodiments of this specification can determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data. Then, if the error degree is not within the preset range, it is based on The degree of error adjusts the parameters in the initial neural network, and then, based on the specified data conversion algorithm, performs data conversion processing on the adjusted parameters. After completing the data conversion processing, based on the target training data and the true value of the target training data, Continue to train the initial neural network until the loss degree is within the preset range, and finally, the initial neural network with the loss degree within the preset range is determined as the target neural network. In the embodiment of this specification, the data conversion processing of the parameters is simulated in advance during the generation of the target neural network, so that the parameters in the final generated target neural network can achieve the effect after the data conversion processing, so that the subsequent In the application process, only the data to be processed need to be processed for data conversion, without the need for data conversion processing for parameters, which can avoid the problem of reduced accuracy due to truncated parameters, and realize the introduction of specified data conversion algorithms while ensuring The processing effect of the data to be processed by the neural network.

Fig. 2 is another data processing method provided by an embodiment of this specification. As shown in Fig. 2, the method may include:

Step 201: Perform fixed-point processing on the initial training data to obtain fixed-point training data.

In this step, the fixed-point processing refers to the conversion of floating-point numbers to fixed-point representations according to the fixed-point rules. In this way, by performing fixed-point processing on the initial training data, the floating-point numbers in the initial training data can be converted into fixed-point numbers. Points, among them, floating-point number refers to the number of which the position of the decimal point will float and its integer and decimal places will change. Fixed-point number refers to the number determined by the integer and decimal places.

Specifically, during the conversion, each floating-point number in the initial training data can be estimated first, and then based on the data expression range of the preset fixed-point number to perform calibration to determine the calibration value, where the calibration value represents It is the number of digits after the decimal point. The larger the calibration value, the greater the accuracy of the data but the greater the range of expression required, and the higher the hardware requirements of the terminal. On the contrary, the smaller the calibration value, the greater the accuracy of the data. However, the required expression range will be smaller, and the hardware requirements of the terminal will be lower. Correspondingly, when determining the calibration value, it is possible to select a larger one while ensuring that the preset data expression range is not exceeded. The value is used as the calibration value to ensure data accuracy. Finally, the decimal place of the floating-point number is truncated based on the scaled value. For example, suppose the floating-point number is 3.1415926 and the scaled value is 4, then the floating-point number can be converted to 3.1415. Of course, it can also be based on the truncated decimal place. The value of is rounded to the last decimal place after truncation. For example, the floating-point number can be converted to 3.1416.

Step 202: Perform inverse fixed-point processing on the fixed-point training data to obtain the target training data.

In this step, the inverse fixed-point processing refers to the conversion of a fixed-point number into a floating-point number according to the inverse fixed-point rule. Specifically, when the conversion is performed, the fixed-point number can be converted into a fixed-point number according to the number of decimal places. Number, the decimal places of the fixed-point number are complemented, so that the number of decimal places is the same as the number before the conversion, and then the conversion is realized. The complemented value can be generated by a random generation algorithm. In this step, the floating-point number after the inverse fixed-point processing is used as the target training data, so that in the subsequent process of generating the target neural network, floating-point numbers can be used for training, and the expression range of floating-point numbers is relatively large. Therefore, to a certain extent It can reduce the probability of data overflow in each calculation operation during training. At the same time, affected by the accuracy, there will be a loss of accuracy in the floating-point number after the completion. In actual application scenarios, in the subsequent application stage, when the trained target neural network is used to process the real data to be processed, it is often The data to be processed will be fixed-pointed first, which will bring a certain loss of data accuracy. Therefore, in the embodiment of this specification, the fixed-point processing and inverse fixed-point processing can be simulated in advance in the training phase. The resulting data loss can further improve the processing accuracy when using the trained target neural network to process the data to be processed in the subsequent application stage.

Step 203: Determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data.

Specifically, for the specific implementation of this step, refer to the foregoing step 102, which is not described in detail in the embodiment of this specification.

Step 204: If the degree of error is not within the preset range, adjust the parameters in the initial neural network based on the degree of error.

Specifically, for the specific implementation of this step, refer to the foregoing step 102, which is not described in detail in the embodiment of this specification.

Step 205: Perform data conversion processing on the adjusted parameters based on the designated data conversion algorithm.

In this step, taking the specified data conversion algorithm as a data conversion algorithm based on the remainder theorem as an example, the data conversion algorithm based on the remainder theorem can perform linear decomposition and linear combination for each item involved in the calculation, using non-time-consuming arithmetic operations, That is, the addition operation replaces the time-consuming operation, that is, the multiplication operation, in order to reduce the algorithm time.

Specifically, this step can be implemented through the following substeps (1) to (2):

Sub-step (1): Perform convolution processing on the adjusted weight matrix based on the transformation matrix defined in the designated data transformation algorithm and the transpose matrix of the transformation matrix.

In this step, assuming that the weight matrix after convolution processing is C, then C can be expressed as:

C=[GgG ^T ]⊙[B ^T dB];

Among them, ⊙ means that the array elements are multiplied in order, g means the preset convolution kernel, G means the change matrix used to transform the convolution kernel, G ^T means the transpose matrix of G, and d means convolution in the weight matrix Each element of the calculation, as an example, assume that the weight matrix is

Then d can be [d1, d2, d3, d4], B represents the change matrix used to transform the weight matrix, B ^T represents the transpose matrix of B, where G, B and their respective transpose matrices are Pre-set matrices, the specific content of these matrices may be preset based on the size of the preset convolution kernel and the size of the weight matrix. Compared with the calculation method that directly performs the convolution processing, the calculation method represented by the above formula can increase the addition calculation and reduce the multiplication calculation in the calculation operation. Because the addition calculation takes less time and the multiplication calculation takes more time, Therefore, the processing efficiency can be improved to a certain extent.

Further, after the convolution processing is performed, the weight matrix after the convolution processing can also be fixed-point processing, and then after the fixed-point processing is completed, the weight matrix after the convolution processing is inverse fixed-point processing. Specifically, the implementation of the fixed-point processing can refer to the description in the above step 201, and the implementation of the inverse fixed-point processing can refer to the description in the above step 202, which is not repeated in the embodiment of the present specification. Since the expression range of floating-point numbers is relatively large, in the embodiment of this specification, after the parameters are fixed-point, and then inversely fixed-point into floating-point numbers, the probability of data overflow in each calculation operation during training can be reduced to a certain extent. At the same time, affected by the accuracy, the floating-point number obtained after the inverse fixed-pointization will have a precision loss. In actual application scenarios, in the subsequent application stage, the trained target neural network is used to process the real data to be processed. , The parameters may also be fixed-point processing, which will bring a certain loss of data accuracy. Therefore, in the embodiment of this specification, the fixed-point processing and inverse fixed-point processing of the parameters during the training phase can be simulated in advance The data loss caused by fixed-point processing can further improve the processing accuracy when using the trained target neural network to process the data to be processed in the subsequent application stage.

In the embodiments of this specification, the parameters are

Sub-step (2): based on the inverse transformation matrix defined in the specified data conversion algorithm and the transposed matrix of the inverse transformation matrix, transform the weight matrix after the convolution processing.

Specifically, assuming that the weight matrix after completing the data conversion processing is Y, then Y can be expressed as: Y=A ^T CA

Among them, A represents the inverse change matrix used to transform the weight matrix C after the convolution processing, ^AT represents the transposed matrix of A, and both A and ^AT are preset matrices. The specific content of these matrices can be It is predetermined based on the size of C, and the size of C may be predetermined based on the size of the preset convolution kernel and the size of the weight matrix.

Step 206: After completing the data conversion process, continue training the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the preset range Inside.

Specifically, for the specific implementation of this step, refer to the foregoing step 102, which is not described in detail in the embodiment of this specification.

Step 207: Determine the initial neural network whose loss degree is within the preset range as the target neural network.

Specifically, for the specific implementation of this step, refer to the foregoing step 102, which is not described in detail in the embodiment of this specification.

In summary, the data processing method provided by the embodiment of this specification performs fixed-point processing and inverse fixed-point processing on the initial training data to generate target training data. In this way, the data loss caused by fixed-point processing can be simulated in advance. Furthermore, in the subsequent application stage, the processing accuracy when using the trained target neural network to process the data to be processed can be improved. Then, based on the target training data and the true value corresponding to the target training data, the error level of the initial neural network is determined. Then, if the error level is not within the preset range, the parameters in the initial neural network are adjusted based on the error level, and then , Based on the specified data conversion algorithm, perform data conversion processing on the adjusted parameters. After completing the data conversion processing, continue training the initial neural network based on the target training data and the true value of the target training data until the loss Set the range, and finally, determine the initial neural network whose loss degree is within the preset range as the target neural network. In the embodiment of this specification, the data conversion processing of the parameters is simulated in advance during the generation of the target neural network, so that the parameters in the final generated target neural network can achieve the effect after the data conversion processing, so that the subsequent In the application process, only the data to be processed need to be processed for data conversion, without the need for data conversion processing for parameters, which can avoid the problem of reduced accuracy due to truncated parameters, and realize the introduction of specified data conversion algorithms while ensuring The processing effect of the data to be processed by the neural network.

Fig. 3 is a flow chart of the steps of a data processing method provided by an embodiment of this specification. As shown in Fig. 3, the method may include:

Step 301: Perform fixed-point processing on the data to be processed to obtain fixed-point processing data.

As an example, the target neural network is a convolutional neural network used to process images, and the data to be processed is an image to be processed. In this step, in the image matrix corresponding to the image to be processed, the value of each element The value is converted to a fixed-point number, and then the fixed-point processing data is obtained. Wherein, the element in the image matrix is the pixel in the image to be processed, and the value of the element is the pixel value of the pixel. Specifically, the implementation of converting a floating-point number to a fixed-point number can refer to the content in the foregoing step 201, which is not limited in the embodiment of the present specification.

Since the integer and decimal places of fixed-point numbers are determined, and the integer and decimal places of floating-point numbers will fluctuate, the increase in the number of digits in the calculation results based on floating-point numbers is often greater than that based on fixed-point numbers. The number of digits of the calculation result, in this way, the terminal needs to provide a register with more digits to support the calculation operation. In the embodiment of this specification, the fixed-point processing data is obtained through the fixed-point processing of the data to be processed, so that the subsequent generation process can be calculated with fixed-point data. In this way, there is no need to provide a register with a large number of bits. The cost of the terminal can be saved. At the same time, because the fixed-point number has a fixed number of digits, the calculation amount is often smaller than that of a floating-point number with a floating number. Therefore, in this step, the fixed-point This method can be applied to more terminals with limited computing power.

Step 302: Perform data conversion processing on the fixed-point processing data based on a designated data conversion algorithm to obtain converted data.

In the embodiment of this specification, the specified data conversion algorithm may be an algorithm that can improve the calculation effect. For example, the specified data conversion algorithm may be a data conversion algorithm based on the remainder theorem, or the specified data conversion algorithm may also be based on pull The data conversion algorithm of Grange's Interpolation Theorem, these two algorithms can perform linear decomposition and linear combination for each item involved in the calculation, and use non-time-consuming calculation operations, namely addition operations, instead of time-consuming calculation operations, namely multiplication operations, and then It can reduce algorithm time and improve processing efficiency. It should be noted that when the convolution kernel of the data conversion algorithm based on the Lagrangian interpolation theorem is large, the number of additions will increase at a speed far exceeding the size of the convolution kernel, which will eventually lead to an increase in the time consumed by the addition operation It even exceeds the time consumed by the multiplication operation saved. Therefore, when selecting the specified data conversion algorithm, you can choose based on the size of the convolution kernel used to ensure that the specified data conversion algorithm introduced can improve processing The effect of efficiency. It should be noted that after the data conversion processing of the fixed-point processing data, data overflow may be caused. Therefore, the embodiment of this specification can provide more storage bits for the converted data to store, so as to avoid data overflow and ensure subsequent The processing process can proceed normally.

Step 303: Based on the pre-training parameters in the target neural network, process the converted data to obtain a processing result.

In this step, the target neural network may be generated by the method shown in the foregoing neural network generation method embodiment. Specifically, the converted data can be input into the target neural network, and then based on the pre-training parameters in the target neural network and the converted data, the dot product operation can be performed. Specifically, the various levels of the target neural network can be based on the The parameters in the level are multiplied by the converted data, and then the result is output to the next level to continue processing based on the next level. Finally, the result of the last layer of the target neural network can be used as the point multiplication result Further, data conversion and inverse processing can be performed on the result of the dot multiplication operation to obtain the processing result. For example, if the specified data conversion algorithm is a data conversion algorithm based on the remainder theorem, then the result of the dot multiplication operation is subjected to data conversion inversion processing The process can be based on the inverse transformation matrix and the transpose matrix of the inverse transformation matrix defined in the data transformation algorithm based on the remainder theorem, and the specific transformation method can refer to the relevant description in the foregoing step 205. The embodiment of this specification is here Do not repeat it.

Further, when the prior art introduces the specified data conversion algorithm in the application stage, in order to avoid truncating the parameters beyond the data expression range after the specified data conversion algorithm is processed on the parameters, it is often necessary to provide more storage locations for the parameters for storage. The terminal to which the target neural network needs to be applied is configured with a register with a large number of storage bits, which in turn leads to a large application cost. However, in the embodiment of this specification, since there is no need to perform specified data conversion processing on the parameters in the application stage, accordingly, there is no need to provide more storage locations for the parameters, thereby reducing the application cost.

In summary, the data processing method provided in the embodiments of this specification can perform fixed-point processing on the data to be processed to obtain fixed-point processed data, and then perform data conversion processing on the fixed-point processed data based on the specified data conversion algorithm to obtain the converted data Then, input the converted data into the target neural network, and process the converted data based on the pre-training parameters in the target neural network to obtain the processing result. Among them, in the process of generating the target neural network, the data conversion processing of the parameters is simulated in advance, so that the finally generated parameters in the target neural network can achieve the effect after the data conversion processing. Therefore, in the embodiment of this specification only It is necessary to perform data conversion processing on the data to be processed, without the need to perform data conversion processing on the parameters, thereby avoiding the problem of reduced accuracy due to truncation of the parameters, thereby achieving the introduction of the specified data conversion algorithm while ensuring the neural network to process the data Treatment effect.

Fig. 4 is a step flow chart of another data processing method provided by an embodiment of this specification. As shown in Fig. 4, the method may include:

Step 401: Determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data.

Specifically, the implementation of this step can refer to the above step 101, which is not limited in the embodiment of this specification.

Step 402: If the degree of error is not within a preset range, adjust the parameters in the initial neural network based on the degree of error.

Specifically, the implementation of this step can refer to the foregoing step 102, which is not limited in the embodiment of this specification.

Step 403: Perform data conversion processing on the adjusted parameters based on the designated data conversion algorithm.

Specifically, the implementation of this step can refer to the above step 103, which is not limited in the embodiment of this specification.

Step 404: After completing the data conversion process, continue training the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the preset range Inside.

Specifically, the implementation of this step can refer to the above step 104, which is not limited in the embodiment of this specification.

Step 405: Determine the initial neural network whose loss degree is within the preset range as the target neural network.

Specifically, the implementation of this step can refer to the above step 105, which is not limited in the embodiment of the specification.

Step 406: Perform fixed-point processing on the data to be processed to obtain fixed-point processing data.

Specifically, the implementation of this step can refer to the above step 201, which is not limited in the embodiment of this specification.

Step 407: Perform data conversion processing on the fixed-point processing data based on the designated data conversion algorithm to obtain converted data.

Specifically, the implementation of this step can refer to the above step 202, which is not limited in the embodiment of this specification.

Step 408: Based on the pre-training parameters in the target neural network, process the converted data to obtain a processing result.

Specifically, the implementation of this step can refer to the above step 203, which is not limited in the embodiment of this specification.

In summary, the data processing method provided by the embodiments of this specification can determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data. Then, if the error degree is not within the preset range, it is based on The degree of error adjusts the parameters in the initial neural network, and then, based on the specified data conversion algorithm, performs data conversion processing on the adjusted parameters. After completing the data conversion processing, based on the target training data and the true value of the target training data, Continue to train the initial neural network until the degree of loss is within the preset range. Determine the initial neural network with the degree of loss within the preset range as the target neural network. Then, you can perform fixed-point processing on the data to be processed to obtain fixed-point processing Data, and then based on the specified data conversion algorithm, the fixed-point processing data is converted into data to obtain the converted data. Then, the converted data is input into the target neural network, and based on the pre-training parameters in the target neural network, the converted data The data is processed to obtain the processing result. Among them, in the process of generating the target neural network, the data conversion processing of the parameters is simulated in advance, so that the finally generated parameters in the target neural network can achieve the effect after the data conversion processing. Therefore, in the embodiment of this specification only It is necessary to perform data conversion processing on the data to be processed, without the need to perform data conversion processing on the parameters, thereby avoiding the problem of reduced accuracy due to truncation of the parameters, thereby achieving the introduction of the specified data conversion algorithm while ensuring the neural network to process the data Treatment effect.

FIG. 5 is a block diagram of a data processing device provided by an embodiment of this specification. As shown in FIG. 5, the data processing device 50 may include:

The first determining module 501 is configured to determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data.

The adjustment module 502 is configured to adjust the parameters in the initial neural network based on the degree of error if the degree of error is not within the preset range.

The conversion module 503 is configured to perform data conversion processing on the adjusted parameters based on a specified data conversion algorithm.

The training module 504 is configured to, after completing the data conversion process, continue to train the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the Within the preset range.

The second determining module 505 is configured to determine the initial neural network whose loss degree is within the preset range as the target neural network.

In summary, the data processing device provided by the embodiment of this specification can determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data. Then, if the error degree is not within the preset range, it is based on The degree of error adjusts the parameters in the initial neural network, and then, based on the specified data conversion algorithm, performs data conversion processing on the adjusted parameters. After completing the data conversion processing, based on the target training data and the true value of the target training data, Continue to train the initial neural network until the loss degree is within the preset range. Finally, the initial neural network with the loss degree within the preset range is determined as the target neural network. In the embodiment of this specification, the data conversion processing of the parameters is simulated in advance during the generation of the target neural network, so that the parameters in the final generated target neural network can achieve the effect after the data conversion processing, so that the subsequent In the application process, only the data to be processed need to be processed for data conversion, without the need for data conversion processing for parameters, which can avoid the problem of reduced accuracy due to truncated parameters, and realize the introduction of specified data conversion algorithms while ensuring The effect of the neural network on the data to be processed.

Optionally, the device 50 further includes:

The first processing module is used to perform fixed-point processing on initial training data to obtain fixed-point training data;

The second processing module is configured to perform inverse fixed-point processing on the fixed-point training data to obtain the target training data.

Optionally, the designated data conversion algorithm is a data conversion algorithm based on the remainder theorem;

Optionally, the specified data conversion algorithm is a data conversion algorithm based on the Lagrangian interpolation theorem.

Optionally, the specified data conversion algorithm is a data conversion algorithm based on the remainder theorem and a data conversion algorithm based on the Lagrange interpolation theorem.

Optionally, the parameter is a weight matrix;

The conversion module 503 is used to:

Performing convolution processing on the adjusted weight matrix based on the transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the transformation matrix;

The weight matrix after the convolution processing is transformed based on the inverse transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the inverse transformation matrix.

Optionally, the conversion module 503 is further used for:

Performing fixed-point processing on the weight matrix after convolution processing;

After the fixed-point processing is completed, the inverse fixed-point processing is performed on the weight matrix after the convolution processing.

Optionally, the target neural network is a convolutional neural network, and the target training data is an image.

In summary, the data processing device provided by the embodiment of this specification performs fixed-point processing and inverse fixed-point processing on initial training data to generate target training data. In this way, the data loss caused by fixed-point processing can be simulated in advance. Furthermore, in the subsequent application stage, the processing accuracy when using the trained target neural network to process the data to be processed can be improved. Then, based on the target training data and the true value corresponding to the target training data, the error level of the initial neural network is determined. Then, if the error level is not within the preset range, the parameters in the initial neural network are adjusted based on the error level, and then , Based on the specified data conversion algorithm, perform data conversion processing on the adjusted parameters. After completing the data conversion processing, continue training the initial neural network based on the target training data and the true value of the target training data until the loss Set the range, and finally, determine the initial neural network whose loss degree is within the preset range as the target neural network. In the embodiment of this specification, the data conversion processing of the parameters is simulated in advance during the generation of the target neural network, so that the parameters in the final generated target neural network can achieve the effect after the data conversion processing, so that the subsequent In the application process, only the data to be processed need to be processed for data conversion, without the need for data conversion processing for parameters, which can avoid the problem of reduced accuracy due to truncated parameters, and realize the introduction of specified data conversion algorithms while ensuring The effect of the neural network on the data to be processed.

FIG. 6 is a block diagram of a data processing device provided by an embodiment of this specification. As shown in FIG. 6, the data processing device 60 may include:

The first processing module 601 is configured to perform fixed-point processing on the data to be processed to obtain fixed-point processing data;

The conversion module 602 is configured to perform data conversion processing on the fixed-point processing data based on a specified data conversion algorithm to obtain converted data;

The second processing module 603 is configured to process the converted data based on the pre-training parameters in the target neural network to obtain the processing result; wherein, the target neural network is generated by the data processing device in the foregoing embodiment of.

Optionally, the second processing module 603 is configured to:

Performing a dot multiplication operation based on the pre-training parameters and the converted data;

Perform data conversion inverse processing on the result of the dot multiplication operation to obtain the processing result.

Optionally, the data to be processed is an image.

In summary, the data processing device provided by the embodiments of this specification can perform fixed-point processing on the data to be processed to obtain fixed-point processed data, and then perform data conversion processing on the fixed-point processed data based on the specified data conversion algorithm to obtain the converted data Then, input the converted data into the target neural network, and process the converted data based on the pre-training parameters in the target neural network to obtain the processing result. Among them, in the process of generating the target neural network, the data conversion processing of the parameters is simulated in advance, so that the finally generated parameters in the target neural network can achieve the effect after the data conversion processing. Therefore, in the embodiment of this specification only It is necessary to perform data conversion processing on the data to be processed, without the need to perform data conversion processing on the parameters, thereby avoiding the problem of reduced accuracy due to truncation of the parameters, thereby achieving the introduction of the specified data conversion algorithm while ensuring the neural network to process the data Treatment effect.

FIG. 7 is a schematic diagram of the hardware structure of a terminal for implementing the various embodiments of this specification. The terminal 700 includes, but is not limited to: a radio frequency unit 701, a network module 702, an audio output unit 703, an input unit 704, a sensor 705, a display unit 706, User input unit 707, interface unit 708, memory 709, processor 710, power supply 711 and other components. Those skilled in the art can understand that the terminal structure shown in FIG. 12 does not constitute a limitation on the terminal, and the terminal may include more or fewer components than shown in the figure, or combine some components, or arrange different components. In the embodiments of this specification, terminals include, but are not limited to, mobile phones, tablet computers, notebook computers, palmtop computers, vehicle-mounted terminals, wearable devices, and pedometers.

It should be understood that, in the embodiment of this specification, the radio frequency unit 701 can be used for receiving and sending signals in the process of sending and receiving information or talking. Specifically, after receiving downlink data from the base station, it is processed by the processor 710; Uplink data is sent to the base station. Generally, the radio frequency unit 701 includes but is not limited to an antenna, at least one amplifier, a transceiver, a coupler, a low noise amplifier, a duplexer, and the like. In addition, the radio frequency unit 701 can also communicate with the network and other devices through a wireless communication system.

The terminal provides users with wireless broadband Internet access through the network module 702, such as helping users to send and receive emails, browse web pages, and access streaming media.

The audio output unit 703 may convert the audio data received by the radio frequency unit 701 or the network module 702 or stored in the memory 709 into audio signals and output them as sounds. Moreover, the audio output unit 703 may also provide audio output related to a specific function performed by the terminal 700 (for example, call signal reception sound, message reception sound, etc.). The audio output unit 703 includes a speaker, a buzzer, a receiver, and the like.

The input unit 704 is used to receive audio or video signals. The input unit 704 may include a graphics processing unit (GPU) 7041 and a microphone 7042, and the graphics processor 7041 is used to capture still pictures or video images obtained by an image capture terminal (such as a camera) in a video capture mode or an image capture mode. Data is processed. The processed image frame may be displayed on the display unit 706. The image frame processed by the graphics processor 7041 may be stored in the memory 709 (or other storage medium) or sent via the radio frequency unit 701 or the network module 702. The microphone 7042 can receive sound, and can process such sound into audio data. The processed audio data can be converted into a format that can be sent to the mobile communication base station via the radio frequency unit 701 for output in the case of a telephone call mode.

The terminal 700 further includes at least one sensor 705, such as a light sensor, a motion sensor, and other sensors. Specifically, the light sensor includes an ambient light sensor and a proximity sensor. The ambient light sensor can adjust the brightness of the display panel 7061 according to the brightness of the ambient light. The proximity sensor can close the display panel 7061 and/or when the terminal 700 is moved to the ear. Or backlight. As a kind of motion sensor, the accelerometer sensor can detect the magnitude of acceleration in various directions (usually three-axis), and can detect the magnitude and direction of gravity when stationary, and can be used to identify terminal posture (such as horizontal and vertical screen switching, related games, Magnetometer attitude calibration), vibration recognition related functions (such as pedometer, tap), etc.; sensor 705 can also include fingerprint sensors, pressure sensors, iris sensors, molecular sensors, gyroscopes, barometers, hygrometers, thermometers, infrared Sensors, etc., will not be repeated here.

The display unit 706 is used to display information input by the user or information provided to the user. The display unit 706 may include a display panel 7061, and the display panel 7061 may be configured in the form of a liquid crystal display (LCD), an organic light-emitting diode (OLED), etc.

The user input unit 707 may be used to receive inputted numeric or character information, and generate key signal input related to user settings and function control of the terminal. Specifically, the user input unit 707 includes a touch panel 7071 and other input devices 7072. The touch panel 7071, also called a touch screen, can collect user touch operations on or near it (for example, the user uses any suitable objects or accessories such as fingers, stylus, etc.) on the touch panel 7071 or near the touch panel 7071. operating). The touch panel 7071 may include two parts, a touch detection terminal and a touch controller. Among them, the touch detection terminal detects the user's touch position, detects the signal brought by the touch operation, and transmits the signal to the touch controller; the touch controller receives the touch information from the touch detection terminal, converts it into contact coordinates, and then sends it To the processor 710, the command sent by the processor 710 is received and executed. In addition, the touch panel 7071 can be implemented in multiple types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the touch panel 7071, the user input unit 707 may also include other input devices 7072. Specifically, other input devices 7072 may include, but are not limited to, a physical keyboard, function keys (such as volume control buttons, switch buttons, etc.), trackball, mouse, and joystick, which will not be repeated here.

Further, the touch panel 7071 can be overlaid on the display panel 7061. When the touch panel 7071 detects a touch operation on or near it, it transmits it to the processor 710 to determine the type of the touch event. The type of event provides corresponding visual output on the display panel 7061. Although the touch panel 7071 and the display panel 7061 are used as two independent components to realize the input and output functions of the terminal, in some embodiments, the touch panel 7071 and the display panel 7061 can be integrated to realize the input and output of the terminal. The output function is not limited here.

The interface unit 708 is an interface for connecting an external terminal and the terminal 700. For example, the external terminal may include a wired or wireless headset port, an external power source (or battery charger) port, a wired or wireless data port, a memory card port, a port for connecting a terminal with an identification module, audio input/output (I/O) port, video I/O port, headphone port, etc. The interface unit 708 may be used to receive input (for example, data information, power, etc.) from an external terminal and transmit the received input to one or more elements in the terminal 700 or may be used to communicate between the terminal 700 and the external terminal. Transfer data between.

The memory 709 can be used to store software programs and various data. The memory 709 may mainly include a program storage area and a data storage area. The program storage area may store an operating system, an application program required by at least one function (such as a sound playback function, an image playback function, etc.), etc.; Data (such as audio data, phone book, etc.) created by the use of mobile phones. In addition, the memory 709 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.

The processor 710 is the control center of the terminal. It uses various interfaces and lines to connect various parts of the entire terminal. It executes by running or executing software programs and/or modules stored in the memory 709, and calling data stored in the memory 709. Various functions of the terminal and processing data, so as to monitor the terminal as a whole. The processor 710 may include one or more processing units; preferably, the processor 710 may integrate an application processor and a modem processor, where the application processor mainly processes the operating system, user interface, application programs, etc., and the modem The processor mainly deals with wireless communication. It can be understood that the foregoing modem processor may not be integrated into the processor 710.

The terminal 700 may also include a power source 711 (such as a battery) for supplying power to various components. Preferably, the power source 711 may be logically connected to the processor 710 through a power management system, so as to manage charging, discharging, and power consumption management through the power management system. Features.

In addition, the terminal 700 includes some functional modules not shown, which will not be repeated here.

Optionally, the embodiment of this specification also provides a terminal, including a processor 710, a memory 709, and a computer program stored on the memory 709 and running on the processor 710. When the computer program is executed by the processor 710, Each process of the foregoing data processing method embodiment is implemented, and the same technical effect can be achieved. In order to avoid repetition, details are not repeated here.

The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in One place, or it can be distributed to multiple network units. Some or all of the modules may be selected according to actual needs to achieve the objectives of the solutions of the embodiments. Those of ordinary skill in the art can understand and implement it without creative work.

Each component embodiment of this specification can be implemented by hardware, or by software modules running on one or more processors, or by a combination of them. Those skilled in the art should understand that a microprocessor or a digital signal processor may be used in practice to implement some or all of the functions of some or all of the components in the computing processing device according to the embodiments of this specification. This specification can also be implemented as a device or device program (for example, a computer program and a computer program product) for executing part or all of the methods described herein. Such a program for realizing this specification may be stored on a computer-readable medium, or may have the form of one or more signals. Such signals can be downloaded from Internet websites, or provided on carrier signals, or provided in any other form.

Correspondingly, the embodiments of the present specification also provide a computer-readable storage medium on which a computer program is stored, and the computer program is executed by a processor to implement the steps of the data processing method, or data Processing method steps.

Furthermore, the embodiment of the present specification also provides a movable platform with limited computing power, and the movable platform is used to implement the steps of the above-mentioned data processing method. Further, the embodiment of the present specification also provides an image acquisition device, the image acquisition device includes a processor, and the image acquisition device is configured to implement the steps of the above data processing method. Limited computing power refers to the limited bandwidth and/or computing resources allocated to data processing on the mobile platform due to multiple concurrent tasks or hardware limitations. For example, the mobile platform has a processor processing capacity of 800M, and the computing process requires more than 1G computing resources. Based on the method provided by the embodiment of this specification, the calculation amount of the system can be greatly reduced, thereby making it possible to use the limited computing power platform. Further, the embodiment of the present specification also provides an unmanned aerial vehicle, which is used to implement the steps of the aforementioned data processing method. The UAV includes a power unit, a flight control unit, and an image acquisition unit. The embodiments of the present specification can process the image data acquired by the image processing unit, and simplify the calculation complexity and calculation amount of the image processing unit through the data processing method, thereby reducing the calculation load of the drone processor. The data processing result can be sent to the flight control unit to control the operation of the power unit.

Furthermore, the embodiment of this specification also provides a handheld stabilized PTZ, which is used to implement the steps of the above data processing method. The handheld stabilized PTZ includes a control unit, a control motor, and a PTZ shaft arm. The handheld stabilized PTZ may also have a camera device or an external image acquisition device through a communication interface. The embodiment of this specification can process the image data acquired by the image processing unit, and simplify the calculation complexity and calculation amount of the image processing unit through the data processing method, thereby reducing the calculation load of the handheld stable pan/tilt. The data processing result can be sent to the control part to control the operation of the motor.

Further, FIG. 8 is a block diagram of a computing processing device provided by an embodiment of this specification. As shown in FIG. 8, FIG. 8 shows a computing processing device that can implement the method according to this specification. The computing processing device traditionally includes a processor 810 and a computer program product in the form of a memory 820 or a computer readable medium. The memory 820 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read Only Memory), EPROM, hard disk, or ROM. The memory 820 has a storage space 830 for executing program codes of any method steps in the above methods. For example, the storage space 830 for program codes may include various program codes for implementing various steps in the above method. These program codes can be read out from or written into one or more computer program products. These computer program products include program code carriers such as hard disks, compact disks (CDs), memory cards or floppy disks. Such a computer program product is usually a portable or fixed storage unit as described with reference to FIG. 9. The storage unit may have storage segments, storage spaces, etc. arranged similarly to the memory 820 in the computing processing device of FIG. 8. The program code can be compressed in an appropriate form, for example. Generally, the storage unit includes computer-readable codes, that is, codes that can be read by, for example, a processor such as 810. These codes, when run by a computing processing device, cause the computing processing device to perform each of the methods described above. step.

The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

The “one embodiment”, “an embodiment” or “one or more embodiments” referred to herein means that a specific feature, structure, or characteristic described in combination with the embodiment is included in at least one embodiment in this specification. In addition, please note that the word examples "in one embodiment" herein do not necessarily all refer to the same embodiment.

In the instructions provided here, a lot of specific details are explained. However, it can be understood that the embodiments of the present specification can be practiced without these specific details. In some instances, well-known methods, structures and technologies are not shown in detail, so as not to obscure the understanding of this specification.

In the claims, any reference signs placed between parentheses should not be constructed as a limitation to the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "a" or "an" preceding an element does not exclude the presence of multiple such elements. This description can be realized by means of hardware including several different elements and by means of a suitably programmed computer. In the unit claims enumerating several devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. do not indicate any order. These words can be interpreted as names.

Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this specification, but not to limit them; although this specification has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: The technical solutions recorded in the foregoing embodiments are modified, or some of the technical features thereof are equivalently replaced; these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this specification.

Claims (27)

1. A data processing method, characterized in that the method includes:

   Determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data;

   If the degree of error is not within the preset range, adjust the parameters in the initial neural network based on the degree of error;

   Perform data conversion processing on the adjusted parameters based on a designated data conversion algorithm;

   After completing the data conversion process, continue training the initial neural network based on the target training data and the true value corresponding to the target training data until the loss degree is within the preset range;

   The initial neural network whose loss degree is within the preset range is determined as the target neural network.
2. The method according to claim 1, characterized in that, before determining the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data, the method further comprises:

   Perform fixed-point processing on the initial training data to obtain fixed-point training data;

   Perform inverse fixed-point processing on the fixed-point training data to obtain the target training data.
3. The method according to claim 1, wherein the designated data conversion algorithm is a data conversion algorithm based on the remainder theorem.
4. The method according to claim 1, wherein the specified data conversion algorithm is a data conversion algorithm based on Lagrangian interpolation theorem.
5. The method according to claim 1, wherein the designated data conversion algorithm is a data conversion algorithm based on the remainder theorem and a data conversion algorithm based on the Lagrangian interpolation theorem.
6. The method according to claim 3, wherein the parameter is a weight matrix;

   The performing data conversion processing on the adjusted parameters based on a specified data conversion algorithm includes:

   Performing convolution processing on the adjusted weight matrix based on the transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the transformation matrix;

   The weight matrix after the convolution processing is transformed based on the inverse transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the inverse transformation matrix.
7. 7. The method according to claim 6, wherein after performing convolution processing on the adjusted weight matrix based on the transformation matrix defined in the specified data transformation algorithm and the transpose matrix of the transformation matrix, Before the conversion of the weight matrix after convolution processing based on the inverse transformation matrix defined in the specified data conversion algorithm and the transpose matrix of the inverse transformation matrix, the method further includes:

   Performing fixed-point processing on the weight matrix after convolution processing;

   After the fixed-point processing is completed, the inverse fixed-point processing is performed on the weight matrix after the convolution processing.
8. The method according to any one of claims 1 to 7, wherein the target neural network is a convolutional neural network, and the target training data is an image.
9. A data processing method, characterized in that the method includes:

   Perform fixed-point processing on the data to be processed to obtain fixed-point processing data;

   Perform data conversion processing on the fixed-point processing data based on a designated data conversion algorithm to obtain converted data;

   Based on the pre-training parameters in the target neural network, the converted data is processed to obtain a processing result; wherein the target neural network is generated by the method according to any one of claims 1 to 8.
10. The method according to claim 9, wherein the processing the converted data based on the pre-training parameters in the target neural network to obtain the processing result comprises:

    Performing a dot multiplication operation based on the pre-training parameters and the converted data;

    Perform data conversion inverse processing on the result of the dot multiplication operation to obtain the processing result.
11. The method according to claim 9 or 10, wherein the data to be processed is an image.
12. A data processing device, characterized in that the device includes:

    The first determining module is configured to determine the error degree of the initial neural network based on the target training data and the true value corresponding to the target training data;

    An adjustment module, configured to adjust parameters in the initial neural network based on the degree of error if the degree of error is not within a preset range;

    The conversion module is configured to perform data conversion processing on the adjusted parameters based on a specified data conversion algorithm;

    The training module is used to continue training the initial neural network based on the target training data and the true value corresponding to the target training data after the data conversion processing is completed, until the loss degree is within the expected Set within

    The second determining module is used to determine the initial neural network whose loss degree is within the preset range as the target neural network.
13. The device according to claim 12, wherein the device further comprises:

    The first processing module is used to perform fixed-point processing on initial training data to obtain fixed-point training data;

    The second processing module is configured to perform inverse fixed-point processing on the fixed-point training data to obtain the target training data.
14. The device according to claim 12, wherein the designated data conversion algorithm is a data conversion algorithm based on the remainder theorem.
15. The device according to claim 12, wherein the specified data conversion algorithm is a data conversion algorithm based on Lagrangian interpolation theorem.
16. The device according to claim 12, wherein the specified data conversion algorithm is a data conversion algorithm based on the remainder theorem and a data conversion algorithm based on the Lagrangian interpolation theorem.
17. The device according to claim 14, wherein the parameter is a weight matrix;

    The conversion module is used for:

    Performing convolution processing on the adjusted weight matrix based on the transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the transformation matrix;

    The weight matrix after the convolution processing is transformed based on the inverse transformation matrix defined in the specified data transformation algorithm and the transposed matrix of the inverse transformation matrix.
18. The device according to claim 17, wherein the conversion module is further configured to:

    Performing fixed-point processing on the weight matrix after convolution processing;

    After completing the fixed-point processing, perform inverse fixed-point processing on the weight matrix after the convolution processing.
19. The device according to any one of claims 12 to 18, wherein the target neural network is a convolutional neural network, and the target training data is an image.
20. A data processing device, characterized in that the device includes:

    The first processing module is used to perform fixed-point processing on the data to be processed to obtain fixed-point processing data;

    The conversion module is configured to perform data conversion processing on the fixed-point processing data based on a specified data conversion algorithm to obtain converted data;

    The second processing module is used to process the converted data based on the pre-training parameters in the target neural network to obtain the processing result; wherein, the target neural network uses any one of claims 12 to 19 Generated by the described device.
21. The device according to claim 20, wherein the second processing module is configured to:

    Performing a dot multiplication operation based on the pre-training parameters and the converted data;

    Perform data conversion inverse processing on the result of the dot multiplication operation to obtain the processing result.
22. The device according to claim 20 or 21, wherein the data to be processed is an image.
23. A computer-readable storage medium, wherein a computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the steps of the data processing method according to any one of claims 1 to 11 are realized .
24. A movable platform with limited computing power, characterized in that the movable platform is used to implement the steps of the data processing method according to any one of claims 1 to 11.
25. An image acquisition device comprising a processor, wherein the image acquisition device is used to implement the steps of the data processing method according to any one of claims 1 to 11.
26. An unmanned aerial vehicle, characterized in that the unmanned aerial vehicle is used to implement the steps of the data processing method according to any one of claims 1 to 11.
27. A handheld stabilized PTZ, characterized in that the handheld stabilized PTZ is used to implement the steps of the data processing method according to any one of claims 1 to 11.

Images