CN113762403B - Image processing model quantization method, device, electronic equipment and storage medium - Google Patents
Image processing model quantization method, device, electronic equipment and storage medium Download PDFInfo
- Publication number
- CN113762403B CN113762403B CN202111076409.8A CN202111076409A CN113762403B CN 113762403 B CN113762403 B CN 113762403B CN 202111076409 A CN202111076409 A CN 202111076409A CN 113762403 B CN113762403 B CN 113762403B
- Authority
- CN
- China
- Prior art keywords
- model
- quantization
- quantized
- detection rate
- false detection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Computational Linguistics (AREA)
- Software Systems (AREA)
- Mathematical Physics (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
- Image Analysis (AREA)
Abstract
The embodiment of the application provides an image processing model quantization method, an image processing model quantization device, electronic equipment and a storage medium, wherein the image processing model of a plurality of storage points is obtained to obtain a plurality of models to be quantized; respectively carrying out model quantization on each model to be quantized to obtain each quantized model; obtaining a quantized test picture, and analyzing the quantized test picture by utilizing each quantized model to respectively obtain test results of each quantized model; and determining a target quantization model based on the test result of each quantization model. Model quantization is carried out on the image processing models of a plurality of storage points in one-time training, so that the problem that the accuracy of a quantized model obtained by selecting a single model for quantization due to random distribution of model parameters is low can be solved, and the success rate of model quantization can be increased; and the parameter adjustment of the image processing model is not needed by people, so that the manual workload is reduced, and a good foundation is laid for the quantification of the batch models.
Description
Technical Field
The present application relates to the field of deep learning technologies, and in particular, to an image processing model quantization method, an image processing model quantization device, an electronic device, and a storage medium.
Background
With the development of artificial intelligence technology, deep learning models are increasingly applied to image processing scenes. The quantization of the deep learning model refers to a process of approximating tensor data of continuously valued floating point model weights to a finite discrete value with lower reasoning precision loss, which is a process of approximating 32-bit finite range floating point data with a data type with fewer bits, and the input and output of the deep learning model are still floating point data, thereby achieving the purposes of reducing the size of the deep learning model, reducing the hardware consumption of the deep learning model, accelerating the reasoning speed of the deep learning model and the like.
In the related technology, firstly, training a deep learning model by using a sample picture to obtain a trained deep learning model; and then manually selecting a certain number of quantized reference pictures, and carrying out model quantization on the trained deep learning model by utilizing the quantized reference pictures, thereby obtaining a quantized model. However, by adopting the method, the accuracy of the obtained quantization model is random, and the quantization model with higher accuracy cannot be obtained.
Disclosure of Invention
The embodiment of the application aims to provide an image processing model quantization method, an image processing model quantization device, electronic equipment and a storage medium, so as to obtain a quantization model with high accuracy. The specific technical scheme is as follows:
In a first aspect, an embodiment of the present application provides an image processing model quantization method, including:
acquiring image processing models of a plurality of storage points to obtain a plurality of models to be quantized;
respectively carrying out model quantization on each model to be quantized to obtain each quantized model;
obtaining a quantized test picture, and analyzing the quantized test picture by utilizing each quantized model to obtain a test result of each quantized model;
and determining a target quantization model based on the test result of each quantization model.
In a possible implementation manner, before the step of obtaining the image processing models of the plurality of storage points to obtain a plurality of models to be quantized, the method further includes:
and training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point every time the training times meet the preset storage conditions.
In one possible implementation manner, the performing model quantization on each model to be quantized to obtain each quantized model includes:
acquiring a quantization reference picture and a quantization configuration file, wherein the quantization reference picture is a randomly selected positive sample picture;
And respectively carrying out model quantization on each model to be quantized based on the quantization reference picture and the quantization configuration file to obtain each quantization model.
In one possible implementation, the test results include a detection rate and a false detection rate;
the obtaining the quantized test pictures, analyzing the quantized test pictures by using each quantized model to obtain test results of each quantized model, respectively, including:
obtaining a quantized test picture, wherein the quantized test picture comprises a positive sample picture and a negative sample picture;
analyzing a positive sample picture in the quantized test picture by using each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model;
analyzing a negative sample picture in the quantized test picture by using each quantization model to respectively obtain a second false detection rate of each quantization model;
and respectively obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model.
In one possible implementation manner, the obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model includes:
For each quantization model, determining the number of false detection pictures of the quantization model according to the number of positive sample pictures, the number of negative sample pictures, the first false detection rate and the second false detection rate of the quantization model in the quantization test pictures;
and determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures.
In one possible implementation manner, the determining the target quantization model based on the test result of each quantization model includes:
acquiring a preset detection rate threshold and a preset false detection rate threshold;
filtering out quantization models with the detection rate smaller than a preset detection rate threshold value from the quantization models, and filtering out quantization models with the false detection rate larger than the preset false detection rate threshold value;
and selecting a target quantization model from the filtered quantization models.
In one possible implementation manner, the determining the target quantization model based on the test result of each quantization model includes:
weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model;
Sequencing the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence;
acquiring a preset range threshold value, and determining each weighted value of the selected numerical value in the preset range threshold value in the weighted value sequence to obtain each target weighted value; dividing each target weighted value with continuous sequence in the weighted value sequence into the same weighted value set to obtain each weighted value set;
and selecting a designated target weighted value from the weighted value set with the largest target weighted value number, and taking a quantized model corresponding to the designated target weighted value as a target quantized model.
In a second aspect, an embodiment of the present application provides an image processing model quantization apparatus, including:
the to-be-quantized model acquisition module is used for acquiring image processing models of a plurality of storage points to obtain a plurality of to-be-quantized models;
the model quantization module is used for respectively carrying out model quantization on each model to be quantized to obtain each quantization model;
the model test module is used for acquiring quantized test pictures, analyzing the quantized test pictures by utilizing the quantized models, and respectively obtaining test results of the quantized models;
And the target quantization model determining module is used for determining a target quantization model based on the test result of each quantization model.
In one possible embodiment, the apparatus further comprises:
the model training module is used for training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point when the training times meet the preset storage conditions.
In a possible implementation manner, the model quantization module is specifically configured to: acquiring a quantization reference picture and a quantization configuration file, wherein the quantization reference picture is a randomly selected positive sample picture; and respectively carrying out model quantization on each model to be quantized based on the quantization reference picture and the quantization configuration file to obtain each quantization model.
In one possible implementation, the test results include a detection rate and a false detection rate; the model test module comprises:
the quantitative test picture acquisition sub-module is used for acquiring quantitative test pictures, wherein the quantitative test pictures comprise positive sample pictures and negative sample pictures;
the positive sample picture analysis submodule is used for analyzing the positive sample picture in the quantized test picture by utilizing each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model;
The negative sample picture analysis submodule is used for analyzing the negative sample pictures in the quantized test pictures by utilizing the quantized models to respectively obtain second false detection rates of the quantized models;
and the false detection rate determination submodule is used for respectively obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model.
In a possible implementation manner, the false detection rate determining submodule is specifically configured to determine, for each quantization model, a number of false detection pictures of the quantization model according to a number of positive sample pictures, a number of negative sample pictures, a first false detection rate and a second false detection rate of the quantization model; determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures;
in a possible implementation manner, the target quantization model determining module is specifically configured to: acquiring a preset detection rate threshold and a preset false detection rate threshold; filtering out quantization models with the detection rate smaller than a preset detection rate threshold value from the quantization models, and filtering out quantization models with the false detection rate larger than the preset false detection rate threshold value; selecting a target quantization model from the filtered quantization models;
In a possible implementation manner, the target quantization model determining module is specifically configured to: weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model; sequencing the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence; acquiring a preset range threshold value, and determining each weighted value of the selected numerical value in the preset range threshold value in the weighted value sequence to obtain each target weighted value; dividing each target weighted value with continuous sequence in the weighted value sequence into the same weighted value set to obtain each weighted value set; and selecting a designated target weighted value from the weighted value set with the largest target weighted value number, and taking a quantized model corresponding to the designated target weighted value as a target quantized model.
In a third aspect, an embodiment of the present application provides an electronic device, including a processor and a memory;
the memory is used for storing a computer program;
the processor is used for realizing any one of the image processing model quantization methods when executing the program stored in the memory.
In a fourth aspect, an embodiment of the present application provides a computer readable storage medium having stored therein a computer program which, when executed by a processor, implements the image processing model quantization method according to any one of the present application.
In a fifth aspect, embodiments of the present application provide a computer program product comprising instructions which, when run on a computer, cause the computer to perform the image processing model quantization method according to any of the present application.
The embodiment of the application has the beneficial effects that:
the image processing model quantization method, the image processing model quantization device, the electronic equipment and the storage medium provided by the embodiment of the application acquire the image processing models of a plurality of storage points to acquire a plurality of models to be quantized; respectively carrying out model quantization on each model to be quantized to obtain each quantized model; obtaining a quantized test picture, and analyzing the quantized test picture by utilizing each quantized model to respectively obtain test results of each quantized model; and determining a target quantization model based on the test result of each quantization model. Model quantization is carried out on the image processing models of a plurality of storage points in one-time training, so that the problem that the accuracy of a quantized model obtained by selecting a single model for quantization due to random distribution of model parameters is low can be solved, and the success rate of model quantization can be increased; and the parameter adjustment of the image processing model is not needed by people, so that the manual workload is reduced, and a good foundation is laid for the quantification of the batch models. Of course, it is not necessary for any one product or method of practicing the application to achieve all of the advantages set forth above at the same time.
Drawings
In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings used in the embodiments or the description of the prior art will be briefly described, and it is obvious that the drawings in the following description are only some embodiments of the present application, and other drawings may be obtained according to these drawings for a person having ordinary skill in the art.
FIG. 1 is a first diagram illustrating a quantization method of an image processing model according to an embodiment of the present application;
FIG. 2 is a second schematic diagram of an image processing model quantization method according to an embodiment of the present application;
FIG. 3 is a schematic diagram of a possible implementation of step S102 in an embodiment of the present application;
FIG. 4 is a schematic diagram of a possible implementation of step S103 in an embodiment of the present application;
FIG. 5 is a schematic diagram of a possible implementation of step S104 in an embodiment of the present application;
FIG. 6 is a first schematic diagram of an image processing model quantization apparatus according to an embodiment of the present application;
fig. 7 is a schematic diagram of an electronic device according to an embodiment of the application.
Detailed Description
The following description of the embodiments of the present application will be made clearly and completely with reference to the accompanying drawings, in which it is apparent that the embodiments described are only some embodiments of the present application, but not all embodiments. Based on the embodiments of the present application, all other embodiments obtained by the person skilled in the art based on the present application are included in the scope of protection of the present application.
First, the terms of the present application are explained:
labeling: and marking the target object in the picture in a picture frame mode.
Quantization reference picture: the picture containing the object provides reference and correction data for the quantization tool.
Positive sample picture: and (5) containing a picture of the target object and marking the target object.
Negative sample picture: the picture without the object is used for preventing false recognition under a specific scene.
Quantification of test pictures: a set of pictures comprising positive and negative sample pictures is used to test the performance of the quantization model.
Image processing model: and the file which is output after the deep learning training and can be used for recognition.
Storage point: in the training process of the image processing model, the image processing model is stored once every certain training times, the image processing model stored each time is called a storage point, and the training times of the image processing models of different storage points are different.
In the related art, firstly, a deep learning model to be quantized is selected manually; and then manually selecting a certain number of quantized reference pictures, and carrying out model quantization on the deep learning model to be quantized by using the quantized reference pictures, thereby obtaining a quantized model. The deep learning model to be quantized is selected by artificial subjective selection, and the effect of which model is determined to be better by the artificial selection is selected as the deep learning model to be quantized, or the deep learning model with the best test result is selected as the deep learning model to be quantized by taking the test result of the deep learning model as the reference, or the deep learning model with the largest training frequency is selected as the deep learning model to be quantized.
However, the best deep learning model may have poor test results after quantization. The inventor finds that in the training process of the deep learning model, the distribution of parameters of the deep learning model has certain randomness, so that the accuracy of the quantization model has no positive correlation with the training times of the deep learning model, and the accuracy of the quantization model obtained by carrying out model quantization on the trained deep learning model in the related technology is random and not necessarily optimal, and the quantization model with higher accuracy cannot be obtained.
In view of this, an embodiment of the present application provides an image processing model quantization method, referring to fig. 1, including:
s101, acquiring image processing models of a plurality of storage points to obtain a plurality of models to be quantized.
The image processing model quantization method of the embodiment of the application can be realized by an electronic device, and in one example, the electronic device can be a smart phone, a personal computer, a server or the like.
The image processing model is a model which is obtained based on the training of the deep learning model and is used for image processing or target recognition, for example, the image processing model can be used for recognizing targets such as vehicles, buildings or animals and plants in the image; for example, the image processing model may be used to perform scene classification or style migration on an image. The storage points refer to storage points in the training process of the image processing model, the training times of the image processing models of different storage points are different, and the acquired image processing models of a plurality of storage points are called as a plurality of models to be quantized.
S102, respectively carrying out model quantization on each model to be quantized to obtain each quantized model.
The model to be quantized after model quantization is called as quantization model; the model quantization process can be referred to a model quantization method in the related art, and the embodiment of the application is not particularly limited. In the actual quantization process, the model quantization work can only effectively quantize parameters in a certain range, if the distribution range of the model parameters to be quantized is far beyond the quantization range of the quantization tool, quantization failure can be caused, and it can be understood that the quantization model obtained here is a quantization model with successful quantization.
S103, obtaining quantized test pictures, and analyzing the quantized test pictures by utilizing each quantized model to obtain test results of each quantized model.
And analyzing the quantized test pictures by utilizing the quantized model for each quantized model, so as to obtain a test result of the quantized model. The test result of the quantization model can be parameters such as the accuracy, the error rate, the detection rate or the false detection rate of the quantization model, and the like, and can be specifically set in a self-defined manner according to actual conditions.
S104, determining a target quantization model based on the test result of each quantization model.
And determining a target quantization model based on the test result of each quantization model, and outputting the target quantization model as a quantization model of the image processing model. For example, the test result includes a correct rate, and a quantization model with the highest correct rate may be selected as the target quantization model; for example, the test result includes a detection rate, and a quantization model with the highest detection rate can be selected as a target quantization model; for example, the test result includes a false detection rate, and a quantization model with the lowest false detection rate may be selected as the target quantization model.
The object quantization model inherits the function of an image processing model for image processing or object recognition. For example, the image processing model is used for vehicle identification, and the target quantization model is also used for vehicle identification; for example, the image processing model is used for image scene classification, and then the target quantization model is also used for image scene classification. The image processing models of a plurality of storage points in one training are subjected to model quantization, an optimal target quantization model is determined according to the test result of each quantization model, the accuracy of image processing or target identification of the target quantization model is higher, the problem that the accuracy of the quantization model obtained by selecting a single model for quantization due to random distribution of model parameters is lower can be solved, and the success rate of model quantization can be increased.
In one example, a preset algorithm may also be used to select the best quantization model as the target quantization model. For example, the quantization model with the best comprehensive update can be calculated through algorithms such as difference calculation, curvature analysis, variance analysis or gradient analysis based on the detection rate and false detection rate in the test result of the quantization model.
In one possible implementation manner, the determining the target quantization model based on the test result of each quantization model includes:
step one, weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model.
The specific weighting algorithm may be set in a customized manner according to the actual situation, and in one example, the weighting value=the detection rate-n×the false detection rate.
And step two, sorting the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence.
The sorting can be performed in ascending order or descending order, and all the sorting is within the protection scope of the present application.
Step three, a preset range threshold value is obtained, and each weighted value of the selected numerical value in the preset range threshold value is determined in the weighted value sequence to obtain each target weighted value; dividing each target weighted value which is sequenced continuously in the weighted value sequence into the same weighted value set to obtain each weighted value set.
The preset range threshold value can be set in a self-defined manner according to practical situations, and is specifically related to the calculation mode of the weighted value, for example, the preset range threshold value can be set to be 85-90, 90-94 or 92-95.
And step four, selecting a designated target weighted value from the weighted value set with the largest number of target weighted values, and taking the quantized model corresponding to the designated target weighted value as a target quantized model.
The specified target weight is an extremum in the set of weights with the largest number of target weights, and in one example, the largest target weight may be selected as the specified target weight from the set of weights with the largest number of target weights.
In one example, the preset range threshold is 90-94, and the weighted value sequence is as follows: 70. 93, 75, 82, 93, 92, 95, 89, 93, 92, 94, 99, 87, for example, then the target weighting values are 93 of rank 2, 93 of rank 5, 92 of rank 6, 93 of rank 9, 92 of rank 10, 94 of rank 11, respectively, two sets of weighting values can be obtained, respectively, a set of weighting values a comprising 93 of rank 5, 92 of rank 6, and a set of weighting values B comprising 93 of rank 9, 92 of rank 10, 94 of rank 11. And selecting 94 of the ranking 11 with the largest numerical value from the weighted value set B as a specified target weighted value, wherein the quantization model corresponding to 94 of the ranking 11 is the target quantization model.
In the embodiment of the application, the image processing models of a plurality of storage points in one training are subjected to model quantization, so that the problem of low accuracy of a quantized model obtained by selecting a single model for quantization due to random distribution of model parameters can be solved, and the success rate of model quantization can be increased; and the parameter adjustment of the image processing model is not needed by people, so that the manual workload is reduced, and a good foundation is laid for the quantification of the batch models.
In a possible implementation manner, referring to fig. 2, before the step of obtaining the image processing models of the plurality of storage points to obtain a plurality of models to be quantized, the method further includes:
s201, training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point when the training times meet the preset storage conditions.
The sample pictures comprise positive sample pictures and negative sample pictures, and the training process of the image processing model by using the sample pictures can refer to the training process of the image processing model in the related technology, and the embodiment of the application is not particularly limited.
The preset storage conditions may be set in a user-defined manner according to the actual situation, in one example, each preset training time may be used as a storage point to store an image processing model, for example, the training times N, 2N, 3N, 4N, etc. are used as storage points to respectively store the image processing models of the storage points, where N is the preset time.
In a possible implementation manner, referring to fig. 3, the performing model quantization on each model to be quantized to obtain each quantization model includes:
s1021, obtaining a quantization reference picture and a quantization configuration file.
In one example, the quantized reference picture is a randomly selected positive sample picture. In the embodiment of the application, compared with the manual selection of the quantized reference picture, the method and the device for randomly selecting the quantized reference picture eliminate the problem that subjective judgment of manual selection is not uniform. In one example, the quantization profile may include a model network file, a tag list file, a quantized picture list file, a quantization parameter profile, and the like.
And S1022, respectively carrying out model quantization on each model to be quantized based on the quantization reference picture and the quantization configuration file to obtain each quantization model.
Taking the quantized reference picture as a reference, and respectively carrying out model quantization on each model to be quantized based on the quantization configuration file, thereby obtaining each quantization model.
In one possible implementation, the test results include a detection rate and a false detection rate; referring to fig. 4, the obtaining a quantized test picture, analyzing the quantized test picture by using each quantization model, respectively obtaining a test result of each quantization model, includes:
S1031, obtaining a quantized test picture, wherein the quantized test picture comprises a positive sample picture and a negative sample picture.
S1032, analyzing the positive sample picture in the quantized test picture by utilizing each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model.
S1033, analyzing the negative sample picture in the quantized test picture by utilizing each quantization model to respectively obtain a second false detection rate of each quantization model.
S1034, according to the first false detection rate and the second false detection rate of each quantization model, respectively obtaining the false detection rate of each quantization model.
For any quantization model, determining the false detection rate of the quantization model according to the first false detection rate and the second false detection rate of the quantization model. For example, the first false positive rate and the second false positive rate may be weighted and averaged to obtain the false positive rate. In one example, the obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model includes:
step 1, for each quantization model, determining the number of false detection pictures of the quantization model according to the number of positive sample pictures, the number of negative sample pictures, the first false detection rate and the second false detection rate of the quantization model in the quantization test pictures.
And 2, determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures.
For example, taking the quantization model 1 as an example, the number of positive sample pictures in the quantization test picture is a, the number of negative sample pictures in the quantization test picture is b, the first false detection rate of the quantization model 1 is x, the second false detection rate is y, and then the false detection rate z of the quantization model 1 may be expressed as: z= (ax+by)/(a+b).
In a possible implementation manner, referring to fig. 5, the determining, based on the test result of each quantization model, a target quantization model includes:
s1041, obtaining a preset detection rate threshold and a preset false detection rate threshold.
The preset detection rate threshold value can be set in a self-defined manner according to practical situations, for example, 85%, 90% or 95%, and the preset false detection rate threshold value can be set in a self-defined manner according to practical situations, for example, 10%, 5% or 3%.
S1042, filtering out the quantization models with the detection rate smaller than the preset detection rate threshold value and filtering out the quantization models with the false detection rate larger than the preset false detection rate threshold value.
S1043, selecting a target quantization model from the filtered quantization models.
In one example, the preset false detection rate threshold includes a positive sample false detection rate threshold, a negative sample false detection rate threshold, and a total false detection rate threshold, filtering out quantization models with a detection rate smaller than the preset detection rate threshold from the quantization models, and filtering out quantization models with a false detection rate greater than the preset false detection rate threshold, including: and filtering out the quantization models with the detection rate smaller than a preset detection rate threshold value, the quantization models with the false detection rate larger than a total false detection rate threshold value, the quantization models with the first false detection rate larger than a positive sample false detection rate threshold value and the quantization models with the second false detection rate larger than a negative sample false detection rate threshold value from the quantization models.
One model can be randomly selected from the filtered quantized models to serve as a target quantized model, or the model with the largest detection rate or the smallest false detection rate can be selected from the filtered quantized models to serve as the target quantized model. In one example, a preset algorithm may be utilized to select the best quantization model among the filtered quantization models as the target quantization model. For example, the quantization model with the best comprehensive update can be calculated through algorithms such as difference calculation, curvature analysis, variance analysis or gradient analysis based on the detection rate and false detection rate in the test result of the quantization model.
The embodiment of the application also provides an image processing model quantization device, which comprises:
the quantitative reference picture acquisition module is used for pulling a specified number of quantitative test pictures from the first storage position, wherein the quantitative test pictures are positive sample pictures containing a target object; in one example, a random acquisition mode may be used to acquire a specified number of quantized test pictures without human intervention. Eliminating artificial subjective influence factors.
And the quantized test picture acquisition module is used for pulling a specified number of quantized test pictures from the second storage position, wherein the quantized test pictures comprise positive sample pictures and negative sample pictures, so that the detection rate of the quantized model and the false detection rate of the quantized model can be evaluated during testing.
And the model to be quantized acquisition module is used for pulling the image processing models of all the storage points from the third storage position to serve as the model to be quantized.
And the quantization configuration file generation module is used for generating a quantization configuration file necessary for quantization, wherein the quantization configuration file comprises a model network file, a tag list file, a quantization picture list file, a quantization parameter configuration file and the like.
And the quantization processing module is used for calling the quantization server, and respectively carrying out model quantization on each model to be quantized one by one based on the quantization configuration file and the quantization reference picture to obtain and output quantization models.
And the quantization test module is used for calling the quantization models one by one, analyzing and testing the quantization test pictures on the quantization test server, and respectively storing test results of each quantization model, wherein the test results comprise the detection rate and the false detection rate.
And the quantization completion detection module is used for judging whether the model to be quantized is completely quantized, if not, continuing to operate the quantization processing module and the quantization test module, and if so, starting the quantization result statistics module.
And the quantization result statistics module is used for screening each quantization model by utilizing a preset algorithm based on the test result of each quantization model, selecting and outputting an optimal target quantization model. In one example, the preset algorithm includes, but is not limited to, difference calculation, curvature analysis, variance analysis, gradient analysis, and the like.
The image processing model quantization device provided by the embodiment of the application can be applied to the quantization of Haisi NNIE (Neural Network Inference Engine ), can be suitable for the scenes of batch output quantization models of a large number of model quantization production tasks in engineering application, and can effectively output the quantization models. Compared with the manual selection of the quantized reference pictures, the method and the device have the advantages that the program is selected randomly, and the problems of non-uniform manual intervention and subjective judgment are reduced. In practice, because the parameter distribution of each storage point model has randomness in the model training process, the model quantization can only effectively quantize parameters in a certain range, if the parameter distribution range of the model to be quantized is far beyond the quantization range which can be quantized by a quantization tool, quantization failure can be caused, and if the parameter distribution of the model to be quantized is in a proper range, the model quantization can be successful. According to the method, all the models to be quantized of all the storage points are quantized, and a plurality of quantization models meeting the standard can be obtained from the image processing model generated by one training without adjusting parameters.
The embodiment of the application also provides a quantization device of the image processing model, referring to fig. 6, comprising:
the to-be-quantized model obtaining module 61 is configured to obtain image processing models of a plurality of storage points, so as to obtain a plurality of to-be-quantized models;
the model quantization module 62 is configured to perform model quantization on each of the models to be quantized to obtain each quantized model;
the model test module 63 is configured to obtain quantized test pictures, analyze the quantized test pictures by using each quantization model, and obtain test results of each quantization model respectively;
the target quantization model determining module 64 is configured to determine a target quantization model based on a test result of each of the quantization models.
In one possible embodiment, the apparatus further comprises:
the model training module is used for training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point when the training times meet the preset storage conditions.
In a possible implementation manner, the model quantization module is specifically configured to:
acquiring a quantization reference picture and a quantization configuration file, wherein the quantization reference picture is a randomly selected positive sample picture;
And respectively carrying out model quantization on each model to be quantized based on the quantization reference picture and the quantization configuration file to obtain each quantization model.
In one possible implementation, the test results include a detection rate and a false detection rate; the model test module comprises:
the quantitative test picture acquisition sub-module is used for acquiring quantitative test pictures, wherein the quantitative test pictures comprise positive sample pictures and negative sample pictures;
the positive sample picture analysis submodule is used for analyzing the positive sample picture in the quantized test picture by utilizing each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model;
the negative sample picture analysis submodule is used for analyzing the negative sample pictures in the quantized test pictures by utilizing the quantized models to respectively obtain second false detection rates of the quantized models;
the false detection rate determination submodule is used for respectively obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model;
in a possible implementation manner, the false detection rate determining submodule is specifically configured to determine, for each quantization model, a number of false detection pictures of the quantization model according to a number of positive sample pictures, a number of negative sample pictures, a first false detection rate and a second false detection rate of the quantization model; and determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures.
In a possible implementation manner, the target quantization model determining module is specifically configured to:
acquiring a preset detection rate threshold and a preset false detection rate threshold;
filtering out quantization models with the detection rate smaller than a preset detection rate threshold value from the quantization models, and filtering out quantization models with the false detection rate larger than the preset false detection rate threshold value;
and selecting a target quantization model from the filtered quantization models.
In a possible implementation manner, the target quantization model determining module is specifically configured to:
weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model;
sequencing the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence;
acquiring a preset range threshold value, and determining each weighted value of the selected numerical value in the preset range threshold value in the weighted value sequence to obtain each target weighted value; dividing each target weighted value with continuous sequence in the weighted value sequence into the same weighted value set to obtain each weighted value set;
and selecting a designated target weighted value from the weighted value set with the largest target weighted value number, and taking a quantized model corresponding to the designated target weighted value as a target quantized model.
The embodiment of the application also provides electronic equipment, which comprises: a processor and a memory;
the memory is used for storing a computer program;
the processor is configured to implement any one of the image processing model quantization methods according to the present application when executing the computer program stored in the memory.
Optionally, referring to fig. 7, the electronic device according to the embodiment of the present application further includes a communication interface 72 and a communication bus 74, where the processor 71, the communication interface 72, and the memory 73 perform communication with each other through the communication bus 74.
The communication bus mentioned for the above-mentioned electronic devices may be a PCI (Peripheral Component Interconnect, peripheral component interconnect standard) bus or an EISA (Extended Industry Standard Architecture ) bus, or the like. The communication bus may be classified as an address bus, a data bus, a control bus, or the like. For ease of illustration, the figures are shown with only one bold line, but not with only one bus or one type of bus.
The communication interface is used for communication between the electronic device and other devices.
The Memory may include RAM (Random Access Memory ) or NVM (Non-Volatile Memory), such as at least one disk Memory. Optionally, the memory may also be at least one memory device located remotely from the aforementioned processor.
The processor may be a general-purpose processor, including a CPU (Central Processing Unit ), NP (Network Processor, network processor), etc.; but also DSP (Digital Signal Processing, digital signal processor), ASIC (Application Specific Integrated Circuit ), FPGA (Field-Programmable Gate Array, field programmable gate array) or other programmable logic device, discrete gate or transistor logic device, discrete hardware components.
The embodiment of the application also provides a computer readable storage medium, wherein a computer program is stored in the computer readable storage medium, and the computer program realizes the image processing model quantization method according to any one of the application when being executed by a processor.
In yet another embodiment of the present application, there is also provided a computer program product containing instructions which, when run on a computer, cause the computer to perform the image processing model quantization method of any one of the present application.
In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented in software, may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed on a computer, produces a flow or function in accordance with embodiments of the present application, in whole or in part. The computer may be a general purpose computer, a special purpose computer, a computer network, or other programmable apparatus. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by a wired (e.g., coaxial cable, fiber optic, digital subscriber line), or wireless (e.g., infrared, wireless, microwave, etc.). The computer readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains an integration of one or more available media. The usable medium may be a magnetic medium (e.g., floppy Disk, hard Disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid State Disk (SSD)), etc.
It should be noted that, in this document, the technical features in each alternative may be combined to form a solution, so long as they are not contradictory, and all such solutions are within the scope of the disclosure of the present application. Relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article, or apparatus that comprises the element.
In this specification, each embodiment is described in a related manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for embodiments of the apparatus, electronic device and storage medium, the description is relatively simple as it is substantially similar to the method embodiments, where relevant see the section description of the method embodiments.
The foregoing description is only of the preferred embodiments of the present application and is not intended to limit the scope of the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application are included in the protection scope of the present application.
Claims (10)
1. An image processing model quantization method, comprising:
acquiring image processing models of a plurality of storage points to obtain a plurality of models to be quantized; the storage points refer to storage point positions in the training process of the image processing model;
respectively carrying out model quantization on each model to be quantized to obtain each quantized model;
obtaining a quantized test picture, and analyzing the quantized test picture by utilizing each quantized model to obtain a test result of each quantized model;
determining a target quantization model based on the test result of each quantization model;
before the step of obtaining the image processing models of the plurality of storage points to obtain the plurality of models to be quantized, the method further includes:
and training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point when the training times meet the preset times.
2. The method according to claim 1, wherein the performing model quantization on each model to be quantized to obtain each quantized model includes:
acquiring a quantization reference picture and a quantization configuration file, wherein the quantization reference picture is a randomly selected positive sample picture;
and respectively carrying out model quantization on each model to be quantized based on the quantization reference picture and the quantization configuration file to obtain each quantization model.
3. The method of claim 1, wherein the test results include a detection rate and a false detection rate;
the obtaining the quantized test pictures, analyzing the quantized test pictures by using each quantized model to obtain test results of each quantized model, respectively, including:
obtaining a quantized test picture, wherein the quantized test picture comprises a positive sample picture and a negative sample picture;
analyzing a positive sample picture in the quantized test picture by using each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model;
analyzing a negative sample picture in the quantized test picture by using each quantization model to respectively obtain a second false detection rate of each quantization model;
And respectively obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model.
4. The method of claim 3, wherein the obtaining the false positive rate of each quantization model according to the first false positive rate and the second false positive rate of each quantization model comprises:
for each quantization model, determining the number of false detection pictures of the quantization model according to the number of positive sample pictures, the number of negative sample pictures, the first false detection rate and the second false detection rate of the quantization model in the quantization test pictures;
and determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures.
5. A method according to claim 3, wherein said determining a target quantization model based on test results of each of said quantization models comprises:
acquiring a preset detection rate threshold and a preset false detection rate threshold;
filtering out quantization models with the detection rate smaller than a preset detection rate threshold value from the quantization models, and filtering out quantization models with the false detection rate larger than the preset false detection rate threshold value;
And selecting a target quantization model from the filtered quantization models.
6. A method according to claim 3, wherein said determining a target quantization model based on test results of each of said quantization models comprises:
weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model;
sequencing the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence;
acquiring a preset range threshold value, and determining each weighted value of the selected numerical value in the preset range threshold value in the weighted value sequence to obtain each target weighted value; dividing each target weighted value with continuous sequence in the weighted value sequence into the same weighted value set to obtain each weighted value set;
and selecting a designated target weighted value from the weighted value set with the largest target weighted value number, and taking a quantized model corresponding to the designated target weighted value as a target quantized model.
7. An image processing model quantization apparatus, comprising:
the to-be-quantized model acquisition module is used for acquiring image processing models of a plurality of storage points to obtain a plurality of to-be-quantized models; the storage points refer to storage point positions in the training process of the image processing model;
The model quantization module is used for respectively carrying out model quantization on each model to be quantized to obtain each quantization model;
the model test module is used for acquiring quantized test pictures, analyzing the quantized test pictures by utilizing the quantized models, and respectively obtaining test results of the quantized models;
the target quantization model determining module is used for determining a target quantization model based on the test result of each quantization model;
the model training module is used for training the image processing model by using the sample picture, and storing the current image processing model as the image processing model of the current storage point when the training times meet the preset times.
8. The apparatus of claim 7, wherein the apparatus further comprises:
the model quantization module is specifically configured to: acquiring a quantization reference picture and a quantization configuration file, wherein the quantization reference picture is a randomly selected positive sample picture; based on the quantization reference picture and the quantization configuration file, respectively carrying out model quantization on each model to be quantized to obtain each quantization model;
the test result comprises a detection rate and a false detection rate; the model test module comprises:
The quantitative test picture acquisition sub-module is used for acquiring quantitative test pictures, wherein the quantitative test pictures comprise positive sample pictures and negative sample pictures;
the positive sample picture analysis submodule is used for analyzing the positive sample picture in the quantized test picture by utilizing each quantization model to respectively obtain the detection rate and the first false detection rate of each quantization model;
the negative sample picture analysis submodule is used for analyzing the negative sample pictures in the quantized test pictures by utilizing the quantized models to respectively obtain second false detection rates of the quantized models;
the false detection rate determination submodule is used for respectively obtaining the false detection rate of each quantization model according to the first false detection rate and the second false detection rate of each quantization model;
the false detection rate determining submodule is specifically configured to determine, for each quantization model, the number of false detection pictures of the quantization model according to the number of positive sample pictures, the number of negative sample pictures, the first false detection rate and the second false detection rate of the quantization model in the quantization test pictures; determining the false detection rate of the quantization model according to the number of false detection pictures of the quantization model, the number of positive sample pictures and the number of negative sample pictures in the quantization test pictures;
The target quantization model determining module is specifically configured to: acquiring a preset detection rate threshold and a preset false detection rate threshold; filtering out quantization models with the detection rate smaller than a preset detection rate threshold value from the quantization models, and filtering out quantization models with the false detection rate larger than the preset false detection rate threshold value; selecting a target quantization model from the filtered quantization models;
the target quantization model determining module is specifically configured to: weighting the detection rate and the false detection rate of each quantization model to obtain a weighted value of the quantization model; sequencing the weighted values of the quantization models according to the time sequence of the storage points corresponding to the quantization models to obtain a weighted value sequence; acquiring a preset range threshold value, and determining each weighted value of the selected numerical value in the preset range threshold value in the weighted value sequence to obtain each target weighted value; dividing each target weighted value with continuous sequence in the weighted value sequence into the same weighted value set to obtain each weighted value set; and selecting a designated target weighted value from the weighted value set with the largest target weighted value number, and taking a quantized model corresponding to the designated target weighted value as a target quantized model.
9. An electronic device, comprising a processor and a memory;
the memory is used for storing a computer program;
the processor is configured to implement the method of any one of claims 1-6 when executing a program stored on the memory.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a computer program which, when executed by a processor, implements the method of any of claims 1-6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111076409.8A CN113762403B (en) | 2021-09-14 | 2021-09-14 | Image processing model quantization method, device, electronic equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202111076409.8A CN113762403B (en) | 2021-09-14 | 2021-09-14 | Image processing model quantization method, device, electronic equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN113762403A CN113762403A (en) | 2021-12-07 |
| CN113762403B true CN113762403B (en) | 2023-09-05 |
Family
ID=78795681
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202111076409.8A Active CN113762403B (en) | 2021-09-14 | 2021-09-14 | Image processing model quantization method, device, electronic equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113762403B (en) |
Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102147866A (en) * | 2011-04-20 | 2011-08-10 | 上海交通大学 | Target identification method based on training Adaboost and support vector machine |
| CN110348562A (en) * | 2019-06-19 | 2019-10-18 | 北京迈格威科技有限公司 | The quantization strategy of neural network determines method, image-recognizing method and device |
| CN110929837A (en) * | 2018-09-19 | 2020-03-27 | 北京搜狗科技发展有限公司 | Neural network model compression method and device |
| WO2020165629A1 (en) * | 2019-02-13 | 2020-08-20 | Mipsology SAS | Quality monitoring and hidden quantization in artificial neural network computations |
| CN111598237A (en) * | 2020-05-21 | 2020-08-28 | 上海商汤智能科技有限公司 | Quantization training method, image processing device, and storage medium |
| CN111639745A (en) * | 2020-05-13 | 2020-09-08 | 北京三快在线科技有限公司 | Data processing method and device |
| CN111767833A (en) * | 2020-06-28 | 2020-10-13 | 北京百度网讯科技有限公司 | Model generation method and device, electronic equipment and storage medium |
| CN111797984A (en) * | 2020-06-17 | 2020-10-20 | 宁波物栖科技有限公司 | Quantification and hardware acceleration method and device for multitask neural network |
| CN111860095A (en) * | 2020-03-23 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | State detection model training method and device and state detection method and device |
| CN111860405A (en) * | 2020-07-28 | 2020-10-30 | Oppo广东移动通信有限公司 | Quantification method and device of image recognition model, computer equipment and storage medium |
| CN111967491A (en) * | 2020-06-29 | 2020-11-20 | 北京百度网讯科技有限公司 | Model offline quantization method and device, electronic equipment and storage medium |
| CN112183742A (en) * | 2020-09-03 | 2021-01-05 | 南强智视(厦门)科技有限公司 | Neural network hybrid quantization method based on progressive quantization and Hessian information |
| WO2021077744A1 (en) * | 2019-10-25 | 2021-04-29 | 浪潮电子信息产业股份有限公司 | Image classification method, apparatus and device, and computer readable storage medium |
| EP3857453A1 (en) * | 2019-02-08 | 2021-08-04 | Huawei Technologies Co., Ltd. | Neural network quantization method using multiple refined quantized kernels for constrained hardware deployment |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11734568B2 (en) * | 2018-02-14 | 2023-08-22 | Google Llc | Systems and methods for modification of neural networks based on estimated edge utility |
| GB201805302D0 (en) * | 2018-03-29 | 2018-05-16 | Benevolentai Tech Limited | Ensemble Model Creation And Selection |
| US11645493B2 (en) * | 2018-05-04 | 2023-05-09 | Microsoft Technology Licensing, Llc | Flow for quantized neural networks |
| WO2020056718A1 (en) * | 2018-09-21 | 2020-03-26 | 华为技术有限公司 | Quantization method and apparatus for neural network model in device |
| US20210125066A1 (en) * | 2019-10-28 | 2021-04-29 | Lightmatter, Inc. | Quantized architecture search for machine learning models |
-
2021
- 2021-09-14 CN CN202111076409.8A patent/CN113762403B/en active Active
Patent Citations (14)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102147866A (en) * | 2011-04-20 | 2011-08-10 | 上海交通大学 | Target identification method based on training Adaboost and support vector machine |
| CN110929837A (en) * | 2018-09-19 | 2020-03-27 | 北京搜狗科技发展有限公司 | Neural network model compression method and device |
| EP3857453A1 (en) * | 2019-02-08 | 2021-08-04 | Huawei Technologies Co., Ltd. | Neural network quantization method using multiple refined quantized kernels for constrained hardware deployment |
| WO2020165629A1 (en) * | 2019-02-13 | 2020-08-20 | Mipsology SAS | Quality monitoring and hidden quantization in artificial neural network computations |
| CN110348562A (en) * | 2019-06-19 | 2019-10-18 | 北京迈格威科技有限公司 | The quantization strategy of neural network determines method, image-recognizing method and device |
| WO2021077744A1 (en) * | 2019-10-25 | 2021-04-29 | 浪潮电子信息产业股份有限公司 | Image classification method, apparatus and device, and computer readable storage medium |
| CN111860095A (en) * | 2020-03-23 | 2020-10-30 | 北京嘀嘀无限科技发展有限公司 | State detection model training method and device and state detection method and device |
| CN111639745A (en) * | 2020-05-13 | 2020-09-08 | 北京三快在线科技有限公司 | Data processing method and device |
| CN111598237A (en) * | 2020-05-21 | 2020-08-28 | 上海商汤智能科技有限公司 | Quantization training method, image processing device, and storage medium |
| CN111797984A (en) * | 2020-06-17 | 2020-10-20 | 宁波物栖科技有限公司 | Quantification and hardware acceleration method and device for multitask neural network |
| CN111767833A (en) * | 2020-06-28 | 2020-10-13 | 北京百度网讯科技有限公司 | Model generation method and device, electronic equipment and storage medium |
| CN111967491A (en) * | 2020-06-29 | 2020-11-20 | 北京百度网讯科技有限公司 | Model offline quantization method and device, electronic equipment and storage medium |
| CN111860405A (en) * | 2020-07-28 | 2020-10-30 | Oppo广东移动通信有限公司 | Quantification method and device of image recognition model, computer equipment and storage medium |
| CN112183742A (en) * | 2020-09-03 | 2021-01-05 | 南强智视(厦门)科技有限公司 | Neural network hybrid quantization method based on progressive quantization and Hessian information |
Non-Patent Citations (1)
| Title |
|---|
| 基于权值交互思想的卷积神经网络量化算法;肖国麟等;电子技术应用;第46卷(第10期);第39-41页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN113762403A (en) | 2021-12-07 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108615071B (en) | Model testing method and device | |
| CN108197652B (en) | Method and apparatus for generating information | |
| EP3620982B1 (en) | Sample processing method and device | |
| CN110874604A (en) | Model training method and terminal equipment | |
| CN109816043B (en) | Method and device for determining user identification model, electronic equipment and storage medium | |
| CN110929785A (en) | Data classification method and device, terminal equipment and readable storage medium | |
| CN113902944A (en) | Model training and scene recognition method, device, equipment and medium | |
| CN112085668B (en) | Image tone mapping method based on region self-adaptive self-supervision learning | |
| CN113762403B (en) | Image processing model quantization method, device, electronic equipment and storage medium | |
| CN113283388A (en) | Training method, device and equipment of living human face detection model and storage medium | |
| CN111404835A (en) | Flow control method, device, equipment and storage medium | |
| CN111597383A (en) | Video heat level prediction method and device | |
| CN111104339A (en) | Software interface element detection method and system based on multi-granularity learning, computer equipment and storage medium | |
| CN115546554A (en) | Sensitive image identification method, device, equipment and computer readable storage medium | |
| CN112598020A (en) | Target identification method and system | |
| CN111582446B (en) | System for neural network pruning and neural network pruning processing method | |
| CN112800813B (en) | Target identification method and device | |
| EP1433134A2 (en) | Scalable expandable system and method for optimizing a random system of algorithms for image quality | |
| CN110807480A (en) | Convolution kernel storage method and device in convolution neural network | |
| CN112836819B (en) | Neural network model generation method and device | |
| CN114648646B (en) | Image classification method and device | |
| CN113762382B (en) | Model training and scene recognition method, device, equipment and medium | |
| CN110728615B (en) | Steganalysis method based on sequential hypothesis testing, terminal device and storage medium | |
| CN110991460B (en) | Image recognition processing method, device, equipment and storage medium | |
| CN114692816B (en) | Processing method and equipment of neural network model |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |