Quantitative model optimization method, device, equipment and computer storage medium
By obtaining the configuration information of the target neural network model, multiple quantization and accuracy loss calculations are automatically performed, which solves the problems of uncontrollable accuracy loss and low efficiency in the existing technology, and realizes efficient and automatic quantitative model optimization.
Patent Information
- Application Number
- CN202111429362.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2021-11-29
- Publication Date
- 2025-04-22
- Estimated Expiration
- 2041-11-29
AI Technical Summary
Existing model quantization techniques have problems with uncontrollable accuracy losses and low efficiency, especially when it is necessary to perform manually during multiple quantizations, resulting in a large number of repetitive operations.
By obtaining the configuration information of the target neural network model, the unrepeated quantized image set is automatically extracted, the target neural network model is quantified multiple times, and the accuracy loss of each quantized model is calculated through the test set to determine the optimal quantized model.
It realizes the quantization model with the lowest accuracy loss automatically, reduces the repetitive operations in the quantization process, and improves the efficiency and accuracy of model quantization.
Smart Images
Figure CN114021719B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of artificial intelligence, and more specifically, to a quantitative model optimization method, apparatus, device and computer storage medium. Background Art
[0002] With the development of deep learning, neural networks are widely used in various fields. In some scenarios, it is necessary to trim the existing neural network model to adjust the model size of the neural network model. At present, model quantization, as an optimization technology that can effectively reduce the model size and accelerate deep learning reasoning, has been widely studied and applied in academia and industry.
[0003] However, although model quantization can reduce the size of the model and improve the inference speed, it has the following disadvantages: since the quantization images used for quantization are randomly extracted during the quantization process, the accuracy loss of the quantized model is uncontrollable, and thus the accuracy loss after quantization cannot be ensured. On the other hand, the existing model quantization efficiency is low. Summary of the invention
[0004] The purpose of the embodiments of the present application is to provide a quantization model optimization method, device, equipment and computer storage medium, so as to automatically obtain a quantization model with the lowest precision loss.
[0005] To this end, the first aspect of the present application discloses a quantitative model optimization method, the method comprising:
[0006] Obtain a target picture set, wherein the target picture set includes m pictures;
[0007] Obtain configuration information for a target neural network model, the configuration information including the number of quantizations for the target neural network model and the number of pictures n required for each quantization of the target neural network model;
[0008] Based on the number of quantizations of the target neural network and the number of pictures n required for quantizing the target neural network model each time, extracting a group of quantized picture sets from the target picture set, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of the quantized picture sets in the a group of quantized picture sets is not repeated;
[0009] quantizing the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a;
[0010] Acquire several test sets and test the b quantization models in sequence based on the test sets, so that each of the quantization models outputs a test result;
[0011] Calculating the accuracy loss of each quantization model based on the test results output by each quantization model;
[0012] The optimal quantization model is determined based on the accuracy loss of each of the quantization models.
[0013] Since the method of the present application obtains the configuration information of the target neural network model, the target neural network model can be repeatedly quantized through the configuration information of the target neural network model, and the multiple quantization results can be compared to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization starts.
[0014] In the first aspect of the present application, as an optional implementation manner, determining the optimal quantization model based on the accuracy loss of each quantization model includes:
[0015] Determine a quantization model with the smallest precision loss value based on the precision loss of each quantization model;
[0016] The quantization model with the smallest precision loss value is used as the optimal quantization model.
[0017] In the first aspect of the present application, as an optional implementation manner, the calculating the accuracy loss of each quantization model based on the test result output by each quantization model includes:
[0018] Comparing the test result with preset data and calculating the difference between the test result and the preset data;
[0019] The accuracy loss of the quantization model is determined according to the difference between the test result and the preset data.
[0020] In the first aspect of the present application, as an optional implementation, n is greater than or equal to 1 and less than or equal to 5.
[0021] In the first aspect of the present application, as an optional implementation, m is greater than or equal to 10.
[0022] The second aspect of the present application discloses a quantitative model optimization device, the device comprising:
[0023] A first acquisition module is used to acquire a target picture set, where the target picture set includes m pictures;
[0024] A second acquisition module is used to acquire configuration information for a target neural network model, wherein the configuration information includes the number of quantizations for the target neural network model and the number of pictures n required for each quantization of the target neural network model;
[0025] A quantization module, configured to extract a group of quantized picture sets from the target picture set based on the number of quantizations of the target neural network and the number of pictures n required for each quantization of the target neural network model, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of the quantized picture sets in the a group of quantized picture sets is not repeated;
[0026] The quantization module is further used to quantize the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a;
[0027] The third acquisition module is further used to acquire several test sets and test the b quantization models in sequence based on the test sets, so that each of the quantization models outputs a test result;
[0028] A calculation module, used for calculating the precision loss of each quantization model based on the test result output by each quantization model;
[0029] A determination module is used to determine the optimal quantization model based on the accuracy loss of each quantization model.
[0030] The device of the embodiment of the present application can obtain the configuration information of the target neural network model, and can repeatedly quantize the target neural network model according to the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual initiation of quantization.
[0031] In the second aspect of the present application, as an optional implementation, the determining module includes:
[0032] A first determination submodule, configured to determine a quantization model with a minimum precision loss value based on the precision loss of each quantization model;
[0033] The second determination submodule is used to take the quantization model with the smallest precision loss value as the optimal quantization model.
[0034] In the second aspect of the present application, as an optional implementation, the calculation module includes:
[0035] A comparison submodule, used for comparing the test result with preset data and calculating the difference between the test result and the preset data;
[0036] The third determination submodule is used to determine the accuracy loss of the quantization model according to the difference between the test result and the preset data.
[0037] The third aspect of the present application discloses a quantitative model optimization device, the quantitative model optimization device comprising:
[0038] Processor; and
[0039] The memory is configured to store machine-readable instructions, which, when executed by the processor, execute the quantitative model optimization method of the first aspect of the present application.
[0040] The quantization model optimization device of the embodiment of the present application can obtain the configuration information of the target neural network model, and can repeatedly quantize the target neural network model according to the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization initiations.
[0041] The fourth aspect of the present application discloses a computer storage medium, wherein the computer storage medium stores a computer program, and the computer program is executed by a processor to implement the quantitative model optimization method of the first aspect of the present application.
[0042] The computer storage medium of the embodiment of the present application can obtain the configuration information of the target neural network model, and can repeatedly quantize the target neural network model according to the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization starts. BRIEF DESCRIPTION OF THE DRAWINGS
[0043] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.
[0044] Figure 1 It is a flowchart of a quantitative model optimization method disclosed in an embodiment of the present application;
[0045] Figure 2 It is a structural schematic diagram of a quantitative model optimization device disclosed in an embodiment of the present application;
[0046] Figure 3 It is a structural diagram of the quantitative model optimization device disclosed in the embodiment of the present application. DETAILED DESCRIPTION
[0047] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.
[0048] Embodiment 1
[0049] See also Figure 1 , Figure 1 Schematic diagram of a quantitative model optimization method disclosed in the embodiment of the present application. Figure 1 As shown, the method of the embodiment of the present application includes the following steps:
[0050] 101. Obtain a target picture set, where the target picture set includes m pictures;
[0051] 102. Obtain configuration information for the target neural network model, where the configuration information includes the number of quantizations for the target neural network model and the number of images n required for each quantization of the target neural network model;
[0052] 103. Based on the number of quantizations of the target neural network and the number of pictures n required for each quantization of the target neural network model, extract a group of quantized picture sets from the target picture set, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of quantized picture sets in the a group of quantized picture sets is not repeated;
[0053] 104. quantize the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a;
[0054] 105. Obtain several test sets and test b quantization models in sequence based on the test sets, so that each quantization model outputs a test result;
[0055] 106. Calculate the accuracy loss of each quantization model based on the test results output by each quantization model;
[0056] 107. Determine the optimal quantization model based on the accuracy loss of each quantization model.
[0057] In an embodiment of the present application, the target image set may be a training set for training an image recognition neural network model, wherein the image recognition neural network model may be a model obtained by training a CNN basic neural network. On the other hand, in an embodiment of the present application, the target neural network model refers to an image recognition neural network model.
[0058] In an embodiment of the present application, the target picture set includes m pictures, where m is greater than or equal to 10, that is, the target picture set includes at least 10 pictures. For example, the target picture set includes 50 face images acquired at different times.
[0059] In an embodiment of the present application, the configuration information of the target neural network model may be pre-stored in a user terminal, or may be obtained through the user terminal through real-time user settings. For example, the user sets the number of quantization times for the image recognition neural network model through the user terminal. Further, the configuration information of the target neural network model also includes the pictures required for each quantization, that is, the number of pictures n required for each quantization of the target neural network model, wherein n is greater than or equal to 1 and less than or equal to 5, that is, each time the target neural network model is quantized, 1-5 pictures are required. Setting the number of pictures n required for each quantization of the target neural network model to 1-5 can improve the quantization effect of the model, wherein the more pictures used for each quantization, the worse the quantization effect. Therefore, in order to ensure the quantization effect, the number of pictures n required for each quantization of the target neural network model needs to be set to 1-5. Further, the user can set n to 1 or n to 2 through the user terminal.
[0060] In an embodiment of the present application, for step 103: based on the number of quantization times of the target neural network and the number of pictures n required for each quantization of the target neural network model, an example of extracting a group of quantized picture sets from the target picture set, assuming that n is 3 and m is 48, then 3 pictures are extracted from m pictures to form 16 groups of picture sets, that is, a=16, wherein each group of the 16 groups of picture sets is different to ensure that each quantization is not repeated.
[0061] In an embodiment of the present application, the test set includes several face images acquired by different cameras at different times. Furthermore, the several face images are used as inputs of b quantization models, so that the quantization models output test results, wherein the test results can be used to identify the category of the object in the test image.
[0062] In an embodiment of the present application, for the step of calculating the precision loss of each quantization model based on the test results output by each quantization model, preferably, a comparison value is obtained, wherein the comparison value is a preset recognition result based on the same test set, and the precision loss of the quantization model can be calculated by comparing the test results output by the quantization model with the comparison value of each quantization model.
[0063] To summarize, since the method of the embodiment of the present application obtains the configuration information of the target neural network model, the target neural network model can be repeatedly quantized through the configuration information of the target neural network model, and the multiple quantization results can be compared to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization initiations.
[0064] In the embodiment of the present application, as an optional implementation, step: determining the optimal quantization model based on the precision loss of each quantization model includes the following sub-steps:
[0065] Determine the quantization model with the minimum precision loss value based on the precision loss of each quantization model;
[0066] The quantization model with the smallest precision loss is taken as the optimal quantization model.
[0067] In this optional implementation, based on the precision loss of each quantization model, the quantization model with the smallest precision loss value can be determined, and then the quantization model with the smallest precision loss value can be used as the optimal quantization model.
[0068] In the embodiment of the present application, as an optional implementation, the step of calculating the accuracy loss of each quantization model based on the test results output by each quantization model includes the following sub-steps:
[0069] Compare the test result with the preset data and calculate the difference between the test result and the preset data;
[0070] The accuracy loss of the quantization model is determined based on the difference between the test result and the preset data.
[0071] In the embodiment of the present application, by comparing the test result with the preset data and calculating the difference between the test result and the preset data, the precision loss of the quantization model can be determined according to the difference between the test result and the preset data.
[0072] Embodiment 2
[0073] See also Figure 2 , Figure 2 Schematic diagram of a quantitative model optimization device disclosed in the embodiment of the present application. Figure 2 As shown, the device of the embodiment of the present application includes the following functional modules:
[0074] A first acquisition module 201 is used to acquire a target picture set, where the target picture set includes m pictures;
[0075] A second acquisition module 202 is used to acquire configuration information for a target neural network model, where the configuration information includes the number of quantizations for the target neural network model and the number of images n required for each quantization of the target neural network model;
[0076] The quantization module 203 is used to extract a group of quantized picture sets from the target picture set based on the number of quantizations of the target neural network and the number of pictures n required for each quantization of the target neural network model, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of quantized picture sets in the a group of quantized picture sets is not repeated;
[0077] The quantization module 203 is further used to quantize the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a;
[0078] The third acquisition module 204 is further used to acquire several test sets and test the b quantization models in sequence based on the test sets, so that each quantization model outputs a test result;
[0079] A calculation module 205, configured to calculate the precision loss of each quantization model based on the test results output by each quantization model;
[0080] The determination module 206 is configured to determine an optimal quantization model based on the precision loss of each quantization model.
[0081] The device of the embodiment of the present application can obtain the configuration information of the target neural network model, and can repeatedly quantize the target neural network model according to the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual initiation of quantization.
[0082] In the embodiment of the present application, as an optional implementation, the determination module 206 includes the following sub-functional modules:
[0083] A first determination submodule is used to determine a quantization model with a minimum precision loss value based on the precision loss of each quantization model;
[0084] The second determination submodule is used to take the quantization model with the smallest precision loss value as the optimal quantization model.
[0085] In the embodiment of the present application, as an optional implementation, the calculation module 205 includes the following sub-functional modules:
[0086] A comparison submodule, used to compare the test result with the preset data and calculate the difference between the test result and the preset data;
[0087] The third determination submodule is used to determine the accuracy loss of the quantization model according to the difference between the test result and the preset data.
[0088] It should be noted that for other descriptions of the embodiments of the present application, please refer to the relevant descriptions of Embodiment 1 of the present application, and the embodiments of the present application will not be elaborated on this.
[0089] Embodiment 3
[0090] See also Figure 3 , Figure 3 Schematic diagram of the structure of the quantitative model optimization device disclosed in the embodiment of the present application. Figure 3 As shown, the embodiment of the present application includes:
[0091] Processor 301; and
[0092] The memory 302 is configured to store machine-readable instructions, which, when executed by the processor 301, execute the quantitative model optimization method of the embodiment of the present application.
[0093] The quantization model optimization device of the embodiment of the present application can execute the quantization model optimization method, and can repeatedly quantize the target neural network model by obtaining the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization starts.
[0094] Embodiment 4
[0095] An embodiment of the present application discloses a computer storage medium, which stores a computer program. The computer program is executed by a processor to execute the quantitative model optimization method of the first embodiment of the present application.
[0096] The computer storage medium of the embodiment of the present application can execute the quantization model optimization method, and can repeatedly quantize the target neural network model by obtaining the configuration information of the target neural network model, and compare the multiple quantization results to obtain the optimal quantization model. Compared with the prior art, the method of the embodiment of the present application can automatically run the model quantization process in a script manner through the configuration information of the target neural network model, thereby reducing a large number of repetitive operations in the quantization process, while the prior art needs to rely on manual execution of the quantization process and requires manual completion of a large number of repetitive operations, such as multiple image screenings and multiple manual quantization starts.
[0097] In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interfaces, and the indirect coupling or communication connection of the devices or units can be electrical, mechanical or other forms.
[0098] In addition, the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.
[0099] Furthermore, the functional modules in the various embodiments of the present application may be integrated together to form an independent part, or each module may exist separately, or two or more modules may be integrated to form an independent part.
[0100] It should be noted that if the function is implemented in the form of a software function module and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application can essentially be embodied in the form of a software product, or the part that contributes to the prior art or the part of the technical solution. The computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM) random access memory (RAM), disk or optical disk, and other media that can store program codes.
[0101] In this document, relational terms such as first and second, etc. are used merely to distinguish one entity or operation from another entity or operation, but do not necessarily require or imply any such actual relationship or order between these entities or operations.
[0102] The above description is only an embodiment of the present application and is not intended to limit the protection scope of the present application. For those skilled in the art, the present application may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.
Claims
1. A quantitative model optimization method, characterized in that: The method comprises: Obtain a target picture set, where the target picture set includes m pictures; Obtain configuration information for a target neural network model, the configuration information including the number of quantizations for the target neural network model and the number of pictures n required for each quantization of the target neural network model; Based on the number of quantizations of the target neural network and the number of pictures n required for quantizing the target neural network model each time, extracting a group of quantized picture sets from the target picture set, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of the quantized picture sets in the a group of quantized picture sets is not repeated; quantizing the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a; Acquire several test sets and test the b quantization models in sequence based on the test sets, so that each of the quantization models outputs a test result; Calculating the accuracy loss of each quantization model based on the test results output by each quantization model; Determine an optimal quantization model based on the accuracy loss of each of the quantization models; And determining the optimal quantization model based on the precision loss of each quantization model includes: Determine a quantization model with the smallest precision loss value based on the precision loss of each quantization model; The quantization model with the smallest precision loss value is used as the optimal quantization model.
2. The method according to claim 1, characterized in that The calculating the accuracy loss of each quantization model based on the test result output by each quantization model comprises: Comparing the test result with preset data and calculating the difference between the test result and the preset data; The accuracy loss of the quantization model is determined according to the difference between the test result and the preset data.
3. The method according to claim 1, characterized in that n is greater than or equal to 1 and less than or equal to 5.
4. The method according to claim 1, characterized in that m is greater than or equal to 10.
5. A quantitative model optimization device, characterized in that: The device comprises: A first acquisition module, used to acquire a target picture set, wherein the target picture set includes m pictures; A second acquisition module is used to acquire configuration information for a target neural network model, wherein the configuration information includes the number of quantizations for the target neural network model and the number of pictures n required for each quantization of the target neural network model; A quantization module, configured to extract a group of quantized picture sets from the target picture set based on the number of quantizations of the target neural network and the number of pictures n required for each quantization of the target neural network model, wherein the value of a is equal to the number of quantizations of the target neural network, and each group of the quantized picture sets in the a group of quantized picture sets is not repeated; The quantization module is further used to quantize the target neural network model according to the a group of quantized picture sets, so as to crop the target neural network model and obtain b quantized models, where b is equal to a; The third acquisition module is further used to acquire several test sets and test the b quantization models in sequence based on the test sets, so that each of the quantization models outputs a test result; A calculation module, used for calculating the precision loss of each quantization model based on the test result output by each quantization model; A determination module, configured to determine an optimal quantization model based on the precision loss of each of the quantization models; And, the determination module includes: A first determination submodule, configured to determine a quantization model with a minimum precision loss value based on the precision loss of each quantization model; The second determination submodule is used to take the quantization model with the smallest precision loss value as the optimal quantization model.
6. The device according to claim 5, characterized in that The computing module comprises: A comparison submodule, used for comparing the test result with preset data and calculating the difference between the test result and the preset data; The third determination submodule is used to determine the accuracy loss of the quantization model according to the difference between the test result and the preset data.
7. A quantitative model optimization device, characterized in that: include: processor; as well as A memory configured to store machine-readable instructions, which, when executed by the processor, execute the quantitative model optimization method as described in any one of claims 1 to 4.
8. A computer storage medium, characterized in that The computer storage medium stores a computer program, and the computer program is executed by a processor according to the quantitative model optimization method as described in any one of claims 1 to 4.
Citation Information
Patent Citations
Quantization model optimization method and device
CN113673532A