TWI786430B - Device and method for optimizing model conversion of deep learning model, and storage medium - Google Patents

Abstract

A method for optimizing model conversion of deep learning model includes: performing a model conversion on a first deep learning model to obtain a second deep learning model; obtaining weight arrangements of the first deep learning model and the second deep learning model according to deep learning frameworks of the first deep learning model and the second deep learning model; performing model quantifications on the first deep learning model and the second deep learning model; analyzing a weight similarity between the first deep learning model and the second deep learning model based on the weight arrangements and the model quantifications of the first deep learning model and the second deep learning model, and establishing a weight analysis report based on analyzed results of the weight similarity; testing the first learning model and the second deep learning model according to a preset test set to establish a model performance estimation report; obtaining one or more optimization suggestions for the second deep learning model based on the weight analysis report and the model performance estimation report; and using an optimized second deep learning model to process objective data.

Description

Model conversion optimization device, method and storage medium for deep learning model

The present invention relates to the technical field of model conversion optimization, in particular to a model conversion optimization device and method for a deep learning model, and a computer-readable storage medium.

Existing deep learning frameworks include TensorFlow, Caffe, NCNN, Pytorch and other frameworks. In the process from algorithm development to algorithm deployment, different deep learning frameworks are often used. For example, the Tensorflow framework is used for algorithm development and debugging, but in the terminal deployment stage, the NCNN framework may be used, which poses the problem of model conversion. In order to ensure that the model can still maintain good prediction accuracy after the conversion, it is necessary to analyze the model before and after the conversion, and the existing model analysis is generally carried out by personnel with rich deep learning knowledge, relying too much on the professionalism of analysts In addition, there are problems of high time consumption and low efficiency, which lead to poor performance of the model finally deployed on the terminal.

In view of this, it is necessary to provide a model conversion optimization device, method, and computer-readable storage medium for a deep learning model, which can automatically perform in-depth analysis before and after model conversion, and avoid poor performance of the converted model to the greatest extent. situation.

An embodiment of the present invention provides a model conversion optimization method for a deep learning model, The method includes: performing model conversion on the first deep learning model to obtain a second deep learning model, wherein the first deep learning model and the second deep learning model have different deep learning frameworks; according to the first The deep learning framework of the deep learning model obtains the weight arrangement of the first deep learning model and obtains the weight arrangement of the second deep learning model according to the deep learning framework of the second deep learning model; The deep learning model and the second deep learning model carry out model quantification; based on the weight arrangement and model quantification results of the first deep learning model, and the weight arrangement and model quantification results of the second deep learning model, it is obtained The weight similarity between the first deep learning model and the second deep learning model is used to construct a weight analysis report; the first deep learning model and the second deep learning model are respectively analyzed according to a preset test set Performing a test to construct a model performance estimation report; obtaining an optimization suggestion for the second deep learning model according to the weight analysis report and the model performance estimation report, so as to perform model optimization on the second deep learning model; And using the optimized second deep learning model to process the data to be processed.

One embodiment of the present invention provides a model conversion optimization device for a deep learning model, the device includes a processor and a memory, the memory stores a number of computer programs, and the processor is used to execute the computer programs stored in the memory When implementing the following steps: carry out model conversion to the first deep learning model to obtain a second deep learning model, wherein the first deep learning model and the second deep learning model have different deep learning frameworks; according to the first The deep learning framework of the deep learning model obtains the weight arrangement of the first deep learning model and obtains the weight arrangement of the second deep learning model according to the deep learning framework of the second deep learning model; The deep learning model and the second deep learning model carry out model quantification; based on the weight arrangement and model quantification results of the first deep learning model, and the weight arrangement and model quantification results of the second deep learning model, it is obtained The weight similarity between the first deep learning model and the second deep learning model is used to construct a weight analysis report; the first deep learning model and the second deep learning model are respectively analyzed according to a preset test set Tests are performed to build a model performance estimation report; based on the The weight analysis report and the model performance estimation report obtain optimization suggestions for the second deep learning model, so as to optimize the second deep learning model; and use the optimized second deep learning model to process data to process.

An embodiment of the present invention provides a computer-readable storage medium, the computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be executed by one or more processors to implement the following steps: A deep learning model performs model conversion to obtain a second deep learning model, wherein the first deep learning model and the second deep learning model have different deep learning frameworks; according to the deep learning framework of the first deep learning model Obtain the weight arrangement of the first deep learning model and obtain the weight arrangement of the second deep learning model according to the deep learning framework of the second deep learning model; for the first deep learning model and the second deep learning model Two deep learning models perform model quantification; based on the weight arrangement and model quantification results of the first deep learning model, and the weight arrangement and model quantification results of the second deep learning model, the first deep learning model is obtained Weight similarity with the second deep learning model to construct a weight analysis report; respectively test the first deep learning model and the second deep learning model according to a preset test set to construct model performance an estimation report; obtain an optimization suggestion for the second deep learning model according to the weight analysis report and the model performance estimation report, so as to perform model optimization on the second deep learning model; and use the optimized second The deep learning model processes the data to be processed.

Compared with the conventional technology, the above-mentioned deep learning model conversion optimization device, method and computer-readable storage medium can realize automatic in-depth analysis before and after the conversion of the deep learning model, and the analysis efficiency is high, and can be based on the analysis results Provide model optimization suggestions to avoid poor performance of the converted model to the greatest extent.

10: Memory

20: Processor

30:Model conversion optimization program

101: Conversion Module

102: Determine the module

103: Quantization module

104:First Building Mod

105:Second Building Module

106: Optimize the module

107: Processing module

100:Model conversion optimization device

S300, S302, S304, S306, S308, S310, S312: steps

FIG. 1 is a functional module diagram of a model conversion optimization device according to an embodiment of the present invention.

FIG. 2 is a functional module diagram of a model conversion optimization program according to an embodiment of the present invention.

Fig. 3 is a flow chart of a model conversion optimization method according to an embodiment of the present invention.

Please refer to FIG. 1 , which is a schematic diagram of a preferred embodiment of the model conversion optimization device of the present invention.

The model conversion optimization device 100 can provide model optimization suggestions for the model conversion of deep learning models between different frameworks, and can verify the prediction effect of the models before conversion and after conversion. The model conversion optimization apparatus 100 may include a memory 10 , a processor 20 and a model conversion optimization program 30 stored in the memory 10 and operable on the processor 20 . When the processor 20 executes the model conversion optimization program 30, the steps in the embodiment of the model conversion optimization method are implemented, such as steps S300-S312 shown in FIG. 3 . Alternatively, when the processor 20 executes the model conversion optimization program 30, the functions of the modules in FIG. 2 are implemented, such as the modules 101-107.

The model conversion optimization program 30 can be divided into one or more modules, and the one or more modules are stored in the memory 10 and executed by the processor 20 to implement the present invention. The one or more modules may be a series of computer program instruction segments capable of accomplishing specific functions, and the instruction segments are used to describe the execution process of the model conversion optimization program 30 in the model conversion optimization device 100 . For example, the model conversion optimization program 30 can be divided into conversion module 101, determination module 102, quantization module 103, first construction module 104, second construction module 105, optimization module 106 and processing Module 107. For the specific functions of each module, see the functions of each module in Figure 2 below.

Those skilled in the art can understand that the schematic diagram is only an example of the model conversion optimization device 100, and does not constitute a limitation to the model conversion optimization device 100, and may include more or less components than those shown in the illustration, or combine certain components , or different components, for example, the model conversion optimization apparatus 100 may also include input and output devices, communication modules, bus bars, and the like.

The processor 20 can be a central processing unit (Central Processing Unit, CPU), also It can be other general-purpose processors, digital signal processors (Digital Signal Processor, DSP), application specific integrated circuits (Application Specific Integrated Circuit, ASIC), off-the-shelf programmable gate arrays (Field-Programmable Gate Array, FPGA) or other possible Program logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or the processor 20 can also be any conventional processor, etc. The processor 20 can use various interfaces and buses to connect various parts of the model conversion optimization device 100 .

The memory 10 can be used to store the model conversion optimization program 30 and/or modules, and the processor 20 runs or executes the computer programs and/or modules stored in the memory 10, and calls the data stored in the memory 10, Various functions of the model conversion optimization device 100 are realized. Memory 10 can include high-speed random access memory, and can also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart memory card (Smart Media Card, SMC), secure digital (Secure Digital (SD) card, flash memory card (Flash Card), at least one disk memory device, flash memory device, or other volatile solid state memory device.

FIG. 2 is a functional module diagram of a preferred embodiment of the model conversion optimization program of the present invention.

Referring to Fig. 2, the model conversion optimization program 30 may include a conversion module 101, a determination module 102, a quantization module 103, a first construction module 104, a second construction module 105, an optimization module 106 and a processing module 107. In one embodiment, the above-mentioned modules may be programmable software instructions stored in the memory 10 and called and executed by the processor 20 . It can be understood that, in other embodiments, the above-mentioned modules can also be program instructions or firmware solidified in the processor 20 .

The conversion module 101 is used to perform model conversion on the first deep learning model to obtain a second deep learning model.

In one embodiment, the first deep learning model and the second deep learning model preferably have different deep learning frameworks. For example, in the model development stage, use the Tensorflow framework for development and testing to obtain the first deep learning model, and in the terminal deployment stage, you need to use The model of the NCNN framework is deployed to the terminal, so the first deep learning model with the Tensorflow framework needs to be converted into a second deep learning model with the NCNN framework, and then the second deep learning model is deployed to the terminal. The terminal may be an electronic device such as a mobile phone or a tablet computer. The conversion module 101 can perform model conversion on the first deep learning model based on a preset deep learning model conversion tool to obtain a second deep learning model. The preset deep learning model conversion tool may be an existing model converter, which is not limited here.

It can be understood that the model functions of the first deep learning model and the second deep learning model are the same, and the model functions of the first deep learning model and the second deep learning model can be determined according to actual development requirements. For example, the first deep learning model and the second deep learning model can be used for image recognition, voice recognition, image processing, data collection, or natural language processing.

It can be understood that both the first deep learning model and the second deep learning model can use samples in a training sample library for model training, and the samples in the training sample library are divided into a training set and a verification set. The training set is used for model training, and the verification set is used to test the performance of the model.

The determination module 102 is used to obtain the weight arrangement of the first deep learning model according to the deep learning framework of the first deep learning model and obtain the second deep learning model according to the deep learning framework of the second deep learning model. How the weights of the model are arranged.

In an embodiment, when the model transformation of the first deep learning model is completed to obtain the second deep learning model, the determining module 102 can obtain the second deep learning model according to the deep learning framework of the second deep learning model. 2. The weight arrangement of the deep learning model. The determination module 102 can also obtain the weight arrangement of the first deep learning model according to the deep learning framework of the first deep learning model. The weight arrangement of the first deep learning model may represent the arrangement of weight parameters in the first deep learning model, and the weight arrangement of the second deep learning model may be To characterize the arrangement of weight parameters in the second deep learning model.

The quantization module 103 is used for performing model quantization on the first deep learning model and the second deep learning model.

In one embodiment, the quantization module 103 may perform a quantization operation of a specified quantization type on the first deep learning model and the second deep learning model. For example, the quantization type may be to quantize only the model weight, or to quantize both the model weight and the model startup output.

In one embodiment, the quantization module 103 can obtain the weight distribution information of the first deep learning model based on statistics of the weight arrangement of the first deep learning model, and then according to the weight distribution information of the first deep learning model Model quantization is performed on the first deep learning model. The weight distribution information of the first deep learning model may represent the distribution of weight parameters in the first deep learning model. The quantization module 103 can also obtain the weight distribution information of the second deep learning model based on statistics of the weight arrangement of the second deep learning model, and then calculate the weight distribution information of the second deep learning model according to the weight distribution information of the second deep learning model. Model quantization for deep learning models. The weight distribution information of the second deep learning model may represent the distribution of weight parameters in the second deep learning model.

In one embodiment, the quantization module 103 can perform model quantization on the first deep learning model according to the weight distribution information of the first deep learning model. The weight distribution information of the deep learning model determines the number of bits (bits) and precision (such as int or float) for model quantization of the first deep learning model, and then quantifies the first deep learning model based on the determined number of bits and precision. Model quantization for deep learning models. The quantization module 103 can perform model quantization on the second deep learning model according to the weight distribution information of the second deep learning model. Determine the number of bits and precision for model quantization of the second deep learning model, and then perform model quantization on the second deep learning model based on the determined number of bits and precision.

The first construction module 104 is configured to obtain the first deep learning model based on the weight arrangement and model quantization results of the first deep learning model, and the weight arrangement and model quantization results of the second deep learning model. and the weight similarity between the second deep learning model to construct a weight analysis report.

In one embodiment, when the quantization module 103 completes the model quantization of the first deep learning model and the second deep learning model, the first construction module 104 may base on the weight of the first deep learning model Arrangement and model quantification results, and the weight arrangement and model quantification results of the second deep learning model are analyzed to obtain the weight similarity between the first deep learning model and the second deep learning model, which can then be based on The obtained weight similarity is analyzed to construct the weight analysis report.

In an embodiment, the weight analysis report may further include weight distribution information of the first deep learning model and weight distribution information of the second deep learning model.

The second construction module 105 respectively tests the first deep learning model and the second deep learning model according to a preset test set to build a model performance estimation report.

In one embodiment, the model performance estimation report includes the test results of the model prediction performance of the first deep learning model and the test results of the model prediction performance of the second deep learning model. The data of the preset test set may include a training sample library for model training of the first deep learning model and the second deep learning model, and the samples of the training sample library are divided into a training set and a verification set. set.

For example, the function of the first deep learning model and the second deep learning model is image recognition, and the training sample library includes 1000 image samples, wherein the training set includes 800 image samples, and the verification set includes 200 sample image. In the model training process of the first deep learning model, the image samples of the training set are used to train the first deep learning model, and the image samples of the verification set are used to verify the trained first deep learning model. predictive effect. In the second deep learning model During the training process of the type model, the image samples of the training set are also used to train the second deep learning model, and the image samples of the verification set are used to verify the prediction effect of the trained second deep learning model. The preset test set includes all image samples in the training sample library, that is, the second construction module 105 uses 1000 image samples to test the first deep learning model and uses 1000 image samples to test the first deep learning model. Two deep learning models are tested, so as to build the model performance estimation report according to the test results.

In one embodiment, the data of the preset test set may also include live sample sets not included in the training sample library, that is, sample data that the model has not been exposed to during the training process. When the second deep learning model is deployed on the terminal, the on-site sample set may refer to sample data acquired by the terminal in an actual test usage scenario. For example, the second deep learning model is a face recognition model used for face recognition, and the on-site sample set may be face images collected by the terminal during use.

In one embodiment, the data of the preset test set may also include an augmented sample set generated by augmenting the verification set. For example, operations such as flipping and mirroring may be performed on the samples in the verification set to obtain the augmented sample set.

For example, the verification set can be 200 image samples, and 2000 image samples are obtained by performing operations such as translation, flipping, rotation, contrast adjustment, Gaussian noise, and color transformation on multiple image samples in the verification set. The 2000 image samples constitute the augmented sample set.

In one embodiment, the second construction module 105 respectively tests the first deep learning model and the second deep learning model according to a preset test set, so as to construct a model performance estimation report. Specific ways may include: The second construction module 105 uses the training sample library, the on-site sample set and the expanded sample set to test the first deep learning model and the second deep learning model, and then according to the training samples The first test result obtained from the library, the second test result obtained from the on-site sample set, and the third test result obtained from the expanded sample set construct the model performance estimation report.

The optimization module 106 is used to obtain optimization suggestions for the second deep learning model according to the weight analysis report and the model performance estimation report, so as to optimize the second deep learning model.

In one embodiment, when the weight analysis report and the model performance estimation report are obtained, the optimization module 106 can obtain the second deep learning model according to the weight analysis report and the model performance estimation report The optimization suggestion is used to perform model optimization on the second deep learning model, thereby improving the model prediction accuracy of the second deep learning model deployed on the terminal, and saving computing resources occupied by the model.

In one embodiment, when the optimization module 106 obtains an optimization suggestion for the second deep learning model according to the weight analysis report and the model performance estimation report, the model developer or end user can control the second deep learning model. Two deep learning models execute the optimization suggestion.

The processing module 107 is used to use the optimized second deep learning model to process the data to be processed.

In an implementation manner, the data to be processed may refer to input data obtained by a terminal deployed with the second deep learning model in an actual usage scenario and input to the second deep learning model. After the second deep learning model executes the optimization suggestion, the processing module 107 can use the optimized second deep learning model to process the data to be processed to obtain a processing result corresponding to the data to be processed.

Fig. 3 is a flowchart of a model conversion optimization method for a deep learning model in an embodiment of the present invention. According to different requirements, the order of the steps in the flowchart can be changed, and some steps can be omitted.

Step S300, performing model conversion on the first deep learning model to obtain a second deep learning model.

In one embodiment, the first deep learning model and the second deep learning model Model optimization has different deep learning frameworks. For example, in the model development stage, the Tensorflow framework is used for development and testing to obtain the first deep learning model, and in the terminal deployment stage, the model of the NCNN framework needs to be deployed to the terminal, so the first deep learning model with the Tensorflow framework needs to be The learning model is converted into a second deep learning model with an NCNN framework, and then the second deep learning model is deployed to the terminal. The terminal may be an electronic device such as a mobile phone or a tablet computer. The first deep learning model may be converted based on a preset deep learning model conversion tool to obtain a second deep learning model. The preset deep learning model conversion tool may be an existing model converter, which is not limited here.

It can be understood that the model functions of the first deep learning model and the second deep learning model are the same, and the model functions of the first deep learning model and the second deep learning model can be determined according to actual development requirements. For example, the first deep learning model and the second deep learning model can be used for image recognition, voice recognition, image processing, data collection, or natural language processing.

Step S302, obtaining the weight arrangement of the first deep learning model according to the deep learning framework of the first deep learning model and obtaining the weights of the second deep learning model according to the deep learning framework of the second deep learning model Arrangement.

In one embodiment, when the model conversion of the first deep learning model is completed to obtain the second deep learning model, the second deep learning model can be obtained according to the deep learning framework of the second deep learning model weighted arrangement. The weight arrangement manner of the first deep learning model may also be obtained according to the deep learning framework of the first deep learning model. The weight arrangement of the first deep learning model can represent the arrangement of weight parameters in the first deep learning model, and the weight arrangement of the second deep learning model can represent the arrangement of weight parameters in the second deep learning model. The arrangement of weight parameters.

Step S304, performing an operation on the first deep learning model and the second deep learning model Row model quantization.

In an embodiment, a quantization operation of a specified quantization type may be performed on the first deep learning model and the second deep learning model. For example, the quantization type may be to quantize only the model weight, or to quantize both the model weight and the model startup output.

In one embodiment, the weight distribution information of the first deep learning model can be statistically obtained based on the weight arrangement of the first deep learning model, and then the first deep learning model can be calculated according to the weight distribution information of the first deep learning model. A deep learning model for model quantization. The weight distribution information of the first deep learning model may represent the distribution of weight parameters in the first deep learning model. The weight distribution information of the second deep learning model can also be obtained statistically based on the weight arrangement of the second deep learning model, and then the second deep learning model can be calculated according to the weight distribution information of the second deep learning model. Model quantization. The weight distribution information of the second deep learning model may represent the distribution of weight parameters in the second deep learning model.

In an embodiment, the step of performing model quantization on the first deep learning model according to the weight distribution information of the first deep learning model may include: determining the weight distribution information of the first deep learning model to the The number of bits (bits) and precision (such as int or float) for model quantization of the first deep learning model, and then model quantization of the first deep learning model based on the determined number of bits and precision. The step of performing model quantification on the second deep learning model according to the weight distribution information of the second deep learning model may include: determining to quantify the second deep learning model according to the weight distribution information of the second deep learning model The number of bits and precision of model quantization, and then perform model quantization on the second deep learning model based on the determined number of bits and precision.

Step S306, based on the weight arrangement and model quantization results of the first deep learning model, and the weight arrangement and model quantification results of the second deep learning model, the first deep learning model and the second deep learning model are obtained. Weight similarity between deep learning models to build weight analysis reports.

In one embodiment, when the model quantization of the first deep learning model and the second deep learning model is completed, it may be based on the weight arrangement and model quantization results of the first deep learning model, and the The weight arrangement mode of the second deep learning model and the model quantification result are analyzed to obtain the weight similarity between the first deep learning model and the second deep learning model, and then the weight similarity obtained based on the analysis can be used to construct the Weight analysis report.

In an embodiment, the weight analysis report may further include weight distribution information of the first deep learning model and weight distribution information of the second deep learning model.

Step S308, respectively test the first deep learning model and the second deep learning model according to a preset test set, so as to build a model performance estimation report.

In one embodiment, the model performance estimation report includes the test results of the model prediction performance of the first deep learning model and the test results of the model prediction performance of the second deep learning model. The data of the preset test set may include a training sample library for model training of the first deep learning model and the second deep learning model, and the samples of the training sample library are divided into a training set and a verification set. set.

For example, the function of the first deep learning model and the second deep learning model is image recognition, and the training sample library includes 1000 image samples, wherein the training set includes 800 image samples, and the verification set includes 200 sample image. In the model training process of the first deep learning model, the image samples of the training set are used to train the first deep learning model, and the image samples of the verification set are used to verify the trained first deep learning model. predictive effect. In the model training process of the second deep learning model, the image samples of the training set are also used to train the second deep learning model, and the image samples of the verification set are used to verify the trained second deep learning model prediction effect. The preset test set includes all image samples in the training sample library, that is, using 1000 image samples to test the first deep learning model and using 1000 image samples to test the second deep learning model , so as to construct the model performance estimation report according to the test results.

In one embodiment, the data of the preset test set may also include live sample sets not included in the training sample library, that is, sample data that the model has not been exposed to during the training process. When the second deep learning model is deployed on the terminal, the on-site sample set may refer to sample data acquired by the terminal in an actual test usage scenario. For example, the second deep learning model is a face recognition model used for face recognition, and the on-site sample set may be face images collected by the terminal during use.

In one embodiment, the data of the preset test set may also include an augmented sample set generated by augmenting the verification set. For example, operations such as flipping and mirroring may be performed on the samples in the verification set to obtain the augmented sample set.

For example, the verification set can be 200 image samples, and 2000 image samples are obtained by performing operations such as translation, flipping, rotation, contrast adjustment, Gaussian noise, and color transformation on multiple image samples in the verification set. The 2000 image samples constitute the augmented sample set.

In one embodiment, the step of testing the first deep learning model and the second deep learning model according to a preset test set to construct a model performance estimation report may include: using the training sample library, The on-site sample set and the expanded sample set test the first deep learning model and the second deep learning model, and then according to the first test result obtained from the training sample library, the on-site sample set The second test result obtained and the third test result obtained from the amplified sample set construct the model performance estimation report.

Step S310, obtaining an optimization suggestion for the second deep learning model according to the weight analysis report and the model performance estimation report, so as to optimize the second deep learning model.

In one embodiment, when the weight analysis report and the model performance estimation report are obtained, an optimization suggestion for the second deep learning model can be obtained according to the weight analysis report and the model performance estimation report, To perform model optimization on the second deep learning model, and then can The model prediction accuracy of the second deep learning model deployed on the terminal is improved, and computing resources occupied by the model are saved.

In one embodiment, when the optimization suggestion of the second deep learning model is obtained according to the weight analysis report and the model performance estimation report, the model developer or end user can control the second deep learning model Execute the optimization suggestion.

Step S312, using the optimized second deep learning model to process the data to be processed.

In an implementation manner, the data to be processed may refer to input data obtained by a terminal deployed with the second deep learning model in an actual usage scenario and input to the second deep learning model. After the second deep learning model executes the optimization suggestion, the optimized second deep learning model may be used to process the data to be processed, so as to obtain a processing result corresponding to the data to be processed.

The model conversion optimization device, method and computer-readable storage medium for the above-mentioned deep learning model can realize automatic in-depth analysis before and after the conversion of the deep learning model, and the analysis efficiency is high, and model optimization suggestions can be given based on the analysis results to realize Minimize situations where the converted model performs poorly.

In summary, the present invention meets the requirements of an invention patent, and a patent application is filed according to law. However, what is described above is only a preferred embodiment of the present invention, and the scope of the present invention is not limited to the above-mentioned embodiments. For example, all equivalent modifications or changes made by those who are familiar with the technology of this case according to the spirit of the present invention are acceptable. Should be covered in the scope of the following patent applications.

S300, S302, S304, S306, S308, S310, S312: steps

Claims (10)

A model conversion optimization method for a deep learning model, the method comprising: performing model conversion on a first deep learning model to obtain a second deep learning model, wherein the first deep learning model is different from the second deep learning model The deep learning framework of the first deep learning model; the weight arrangement of the first deep learning model is obtained according to the deep learning framework of the first deep learning model and the second deep learning model is obtained according to the deep learning framework of the second deep learning model The weight arrangement of the first deep learning model and the second deep learning model are quantified; based on the weight arrangement and model quantization results of the first deep learning model, and the second depth The weight arrangement of the learning model and the model quantification results are analyzed to obtain the weight similarity between the first deep learning model and the second deep learning model, so as to construct a weight analysis report; Test the first deep learning model and the second deep learning model to build a model performance estimation report; obtain optimization suggestions for the second deep learning model according to the weight analysis report and the model performance estimation report , to perform model optimization on the second deep learning model; and use the optimized second deep learning model to process the data to be processed. The model conversion optimization method for a deep learning model as described in Claim 1, wherein the step of quantifying the first deep learning model and the second deep learning model includes: based on the first deep learning model Obtain the weight distribution information of the first deep learning model according to the weight arrangement method, and perform model quantization on the first deep learning model according to the weight distribution information of the first deep learning model; and based on the weight distribution information of the first deep learning model; Obtaining the weight distribution information of the second deep learning model according to the weight arrangement method of the second deep learning model, and according to the weight distribution information of the second deep learning model The second deep learning model performs model quantization. The model conversion optimization method for a deep learning model as described in Claim 2, wherein the weight analysis report further includes the weight distribution information of the first deep learning model and the weight distribution of the second deep learning model Information. The model conversion optimization method for a deep learning model as described in Claim 2, wherein the step of performing model quantization on the first deep learning model according to the weight distribution information of the first deep learning model includes: according to the weight distribution information of the first deep learning model The weight distribution information of the first deep learning model determines the number of bits and precision for model quantization of the first deep learning model, so as to quantify the first deep learning model based on the determined number of bits and the precision. Model quantization for deep learning models. The model conversion optimization method for a deep learning model as described in Claim 2, wherein the step of performing model quantization on the second deep learning model according to the weight distribution information of the second deep learning model includes: according to the weight distribution information of the second deep learning model The weight distribution information of the second deep learning model determines the number of bits and precision for model quantization of the second deep learning model, so as to quantify the second deep learning model based on the determined number of bits and the precision. Model quantization for deep learning models. The model conversion optimization method of the deep learning model as described in claim 1, wherein the data of the preset test set includes training samples for model training of the first deep learning model and the second deep learning model library, the samples of the training sample library are divided into training set and verification set. The model conversion optimization method of the deep learning model as described in claim item 6, wherein the data of the preset test set also includes an on-site sample set not included in the training sample library and/or expanding the verification set The augmented sample set generated by the increase. The model conversion optimization method of the deep learning model as described in claim item 7, wherein the The step of respectively testing the first deep learning model and the second deep learning model according to the preset test set to construct the model performance estimation report includes: using the training sample library, the field sample set and the expanded sample set to test the first deep learning model and the second deep learning model; and the first test result obtained according to the training sample library and the second The test result and the third test result obtained from the amplified sample set construct the model performance estimation report. A model conversion and optimization device for a deep learning model, the device includes a processor and a memory, and a plurality of computer programs are stored in the memory, and the processor is used to execute the computer programs stored in the memory to achieve claim 1 The steps of the model conversion optimization method of the deep learning model described in any one of to 8. A computer-readable storage medium, the computer-readable storage medium stores a plurality of instructions, and the plurality of instructions can be executed by one or more processors, so as to implement any one of the requirements 1 to 8 The steps of the model conversion optimization method for the deep learning model described above.

Images