TWI765250B - Method for selecting deep learning algorithm and device for selecting deep learning algorithm - Google Patents

Abstract

The invention provides a method for selecting a deep learning algorithm, includes the following steps: acquiring a type of a problem to be processed; selecting a corresponding data set according to the type of the problem, and dividing the selected data set into training data and test data; calculating the similarity of the training data; adjusting a batch size of the training data according to the similarity; selecting a plurality of deep learning algorithms according to the type of the problem, and training the selected deep learning algorithm by using the training data to acquire the corresponding algorithm model; verifying the algorithm model by using the test data, and selecting the deep learning algorithm with the best verification result. The method reduces the difficulty of selection of the deep learning algorithm and is highly versatile. The invention also provides a device for selecting a deep learning algorithm.

Description

Selection method of deep learning algorithm and selection device of deep learning algorithm

The invention belongs to the field of deep learning algorithms, and particularly relates to a method for selecting a deep learning algorithm and a device for selecting a deep learning algorithm.

At present, Convolutional Neural Network (CNN) is widely used as a tool for computer vision. A suitable CNN algorithm model can be designed for any database, and the CNN algorithm model can be used to train the samples in the database, so as to obtain the relationship between the samples in the database and the labels of the samples. However, finding a suitable algorithm from complex CNN algorithms usually requires professionals with professional backgrounds to obtain through a large number of analysis experiments, which is difficult for ordinary people.

In view of the above situation, it is necessary to provide a method for selecting a deep learning algorithm and a device for selecting a deep learning algorithm, so as to reduce the difficulty of selecting a deep learning algorithm for solving a specific problem.

A first aspect of the present invention provides a method for selecting a deep learning algorithm, and the method for selecting a deep learning algorithm includes: acquiring a problem type to be processed; Select a corresponding data set according to the problem type, and divide the selected data set into training data and test data; calculate the similarity of the training data; adjust the batch size of the training data according to the similarity; Select multiple deep learning algorithms for the problem type, and use the training data to train the selected deep learning algorithm to obtain a corresponding algorithm model; use the test data to verify the algorithm model, and select to verify The deep learning algorithm with the best results.

Further, the method for calculating the similarity of the training data includes the steps of randomly selecting a plurality of sample groups in the training data, each of the sample groups including two samples, calculating the similarity of each sample group respectively, and then The average of the similarities of the plural sample groups is calculated, and the average is set as the similarity of the training data.

Further, the relationship between the similarity of the training data and the batch size satisfies: the batch size is inversely proportional to the similarity.

Further, the problem types include image classification, object segmentation and object identification, and the deep learning algorithm is a convolutional neural network roadmap algorithm, including ResNet, AlexNet, VGG and Inception.

Further, the similarity of the training data of the convolutional neural network road algorithm is calculated by using the structural similarity.

A second aspect of the present invention provides a device for selecting a deep learning algorithm. The device for selecting a deep learning algorithm includes: a processor; and a memory, where a plurality of program modules are stored in the memory, and the plurality of program modules are composed of The processor operates and performs the following steps: Obtain the type of the problem to be processed; select a corresponding data set according to the problem type, and divide the selected data set into training data and test data; calculate the similarity of the training data; adjust the training according to the similarity batch size of data; select a variety of deep learning algorithms according to the problem type, and use the training data to train the selected deep learning algorithm to obtain a corresponding algorithm model; use the test data to verify the The algorithm model is selected, and the deep learning algorithm with the best verification result is selected.

Further, the method for calculating the similarity of the training data includes the steps of randomly selecting a plurality of sample groups in the training data, each of the sample groups including two samples, calculating the similarity of each sample group respectively, and then The average of the similarities of the plural sample groups is calculated, and the average is set as the similarity of the training data.

Further, the relationship between the similarity of the training data and the batch size satisfies: the batch size is inversely proportional to the similarity.

Further, the problem types include image classification, object segmentation and object identification, and the deep learning algorithm is a convolutional neural network roadmap algorithm, including ResNet, AlexNet, VGG and Inception.

Further, the similarity of the training data of the convolutional neural network road algorithm is calculated by using the structural similarity.

The invention utilizes the similarity of the training data to flexibly adjust the batch size of the training data, improves the quality of the training data, and makes the algorithm model trained by the training data more accurate, and then uses the test data to verify the algorithm model, and selects the one that solves the specific problem. The optimal algorithm, the present invention reduces the difficulty of selecting a deep learning algorithm and has universality.

100: Selection Device for Deep Learning Algorithms

10: Memory

20: Processor

30: Display unit

40: Input unit

6: Selection system for deep learning algorithms

61: Information acquisition module

62: Select Mods

63: Similarity calculation module

64: Adjustment Mods

65: Training Module

66: Verification Module

FIG. 1 is a schematic diagram of a hardware architecture of a device for selecting a deep learning algorithm in an embodiment of the present invention.

FIG. 2 is a schematic diagram of a module of a deep learning algorithm selection system of the deep learning algorithm selection device shown in FIG. 1 .

FIG. 3 is a schematic flowchart of a method for selecting a deep learning algorithm according to an embodiment of the present invention.

In order to more clearly understand the above objects, features and advantages of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features of the embodiments can be combined with each other unless there is conflict.

In the following description, many specific details are set forth to facilitate a full understanding of the present invention, and the described embodiments are only a part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs. The terms used herein in the description of the present invention are for the purpose of describing specific embodiments only, and are not intended to limit the present invention.

As used herein, the term "and/or" includes any and all combinations of one or more of the associated listed items.

Referring to FIG. 1 , the present invention provides an apparatus 100 for selecting a deep learning algorithm, which is used to reduce the difficulty of selecting a deep learning algorithm for solving a specific problem.

Among them, deep learning (DL) is one of the technologies and research fields of machine learning. By establishing artificial neural networks (ANNs) with a hierarchical structure, the convolution of artificial intelligence is realized in computing systems. Convolutional Neural Networks (CNN) is one of the representative algorithms of deep learning.

The following only takes the convolutional neural network as an example to illustrate, but this method is not limited to the convolutional neural network.

The deep learning algorithm selection device 100 includes a memory 10 , a processor 20 , a display unit 30 and an input unit 40 , and the display unit 30 and the memory 10 are respectively electrically connected to the processor 20 .

The memory 10 is used to store various data in the deep learning algorithm selection device 100 , such as various deep learning algorithms and the like. In this embodiment, the memory 10 may include, but is not limited to, a read-only memory (Read-Only Memory, ROM), a random access memory (Random Access Memory, RAM), and a programmable read-only memory (Programmable Read-Only Memory, PROM) , Erasable Programmable Read-Only Memory (EPROM), One-time Programmable Read-Only Memory (OTPROM), Electronic Erasable Rewritable Read-Only Memory (Electrically-Erasable Programmable Read-Only Memory, EEPROM), Compact Disc Read-Only Memory (CD-ROM), or other optical disk storage, magnetic disk storage, tape storage, or a computer capable of carrying or storing data any other medium that is readable.

The processor 20 may be a central processing unit (CPU, Central Processing Unit), a microprocessor, a digital word processing chip or any processor chip capable of performing data processing functions.

The display unit 30 is used to display the output results of training data or test data, and the like. For example, to display the test result of the deep learning algorithm model, in this embodiment, the display unit 30 may be, but not limited to, a display device such as a touch display screen, a liquid crystal display screen, or the like.

The input unit 40 is used for the user to input various information and control commands, such as confirming the type of problem to be solved by the algorithm. In this embodiment, the input unit 40 may be, but not limited to, a remote control, a mouse, a voice input device, a touch screen, and the like.

Referring to FIG. 2 , the deep learning algorithm selection device 100 also runs a deep learning algorithm selection system 6 . FIG. 2 is a system for selecting a deep learning algorithm in an embodiment of the present invention 6 Schematic diagram of the functional module. In this embodiment, the selection system 6 of the deep learning algorithm includes one or more computer instructions in the form of a program, and the one or more computer instructions in the form of a program are stored in the memory 10 and executed by the processor 20 . The deep learning algorithm selection system 6 includes an information acquisition module 61 , a selection module 62 , a similarity calculation module 63 , an adjustment module 64 , a training module 65 and a verification module 66 .

The information acquisition module 61 is used to acquire the types of problems to be solved by the deep learning algorithm, and the types of problems include image classification, object segmentation and object identification, but are not limited thereto.

The selection module 62 is used for selecting a corresponding data set according to the problem type, dividing the selected data set into training data and test data, and also for selecting all corresponding deep learning algorithms according to the problem type.

The similarity calculation module 63 is used to calculate the similarity of the training data.

In one embodiment, the similarity of training data is calculated using structural similarity.

The structural similarity index (SSIM) is an index to measure the similarity of two images, and it measures the image similarity from three aspects: brightness, contrast, and structure. The range of similarity is 0 to 1. The greater the value of the similarity between the two images, the greater the similarity between the two images. When the two images are exactly the same, the value of the similarity SSIM is equal to 1.

In one embodiment, there are 100 images in the training data, 20 groups of sample images are randomly selected, two for each group, a total of 40 images, and the similarity of the 20 groups of sample images is calculated by using the structural similarity, and then 20 The average value of the similarity of the group sample images, and the average value is set as the similarity of the training data.

It can be understood that, in actual operation, the number of samples is flexibly selected according to the number of training materials in the actual scene, the similarity of each sample group is calculated separately, and then the average value of the similarity of the selected sample groups is calculated.

It can be understood that in other embodiments, other methods for calculating training data are also included.

The adjustment module 64 is used to adjust the batch size of the corresponding training data according to the similarity.

The training module 65 is used to train the deep learning algorithm by using the training data to obtain the algorithm model.

The verification module 66 uses the test data to verify the algorithm model, so as to obtain the deep learning algorithm with the best verification result.

Please refer to FIG. 3 , the present invention also provides a method for selecting a deep learning algorithm. FIG. 3 is a schematic flowchart of the method for selecting a deep learning algorithm. The method for selecting a deep learning algorithm includes the following steps:

In step 301, the type of the problem to be processed is acquired.

Specifically, the information acquisition module 61 acquires the types of questions to be processed, and the types of questions include but are not limited to image classification, object segmentation, and object identification.

Step 302, select a corresponding data set according to the question type, and divide the selected data set into training data and test data.

Specifically, the selection module 62 selects the corresponding data set according to the question type, and divides the selected data set into training data and test data.

For example, Common Object in Context (COCO) is an image dataset maintained by Microsoft, which can classify images, including: more than 300,000 images, more than 2 million instances, and more than 80 types of objects.

The training data is used to train the deep learning algorithm, and the algorithm model is obtained, and the test data is used to verify the accuracy of the algorithm model.

The ratio of training data to test data is generally 4:1. It is understandable that the ratio between training data and test data can be specifically adjusted according to the actual scene, for example, but not limited to, the ratio is 3:1 or 5:1.

Step 303: Calculate the similarity of the training data.

Specifically, the similarity calculation module 63 calculates the similarity of the training data.

In one embodiment, the similarity calculation module 63 uses the structural similarity to calculate the similarity of the training data.

The range of similarity is 0 to 1. The greater the value of the similarity between the two images, the greater the similarity between the two images. When the two images are exactly the same, the value of the similarity is equal to 1.

Specifically, multiple sample groups in the training data are randomly selected, each sample group includes two images, the similarity of each sample group is calculated separately, and then the average value of the similarity of each sample group is calculated.

Step 304: Adjust the batch size of the training data according to the similarity.

Specifically, the adjustment module 64 adjusts the batch size of the corresponding training data according to the similarity, so that the weight of the deep learning algorithm can be accurately trained by using the adjusted training data.

In one embodiment, the similarity is the value of SSIM, and the corresponding relationship between the similarity and the batch size is shown in Table 1. The batch size is the amount of training data in each input to the CNN algorithm.

That is, the batch size is inversely proportional to the similarity, the larger the similarity, the smaller the batch size, and the smaller the similarity, the larger the batch size.

It can be understood that in other embodiments, the batch size is a range, as shown in Table 2.

You can flexibly choose to adjust the batch size according to the actual scene and the size of the similarity.

Step 305 , select multiple deep learning algorithms according to the problem type, and train the selected deep learning algorithm by using the training data to obtain a corresponding algorithm model.

Specifically, the selection module 62 selects a complex deep learning algorithm according to the problem type, and the training module 65 uses the training data to train the selected deep learning algorithm, and obtains corresponding algorithm models respectively.

In one embodiment, the CNN algorithm used to deal with the image classification problem includes ResNet, AlexNet, VGG, Inception, and the like.

Step 306 , verify all the algorithm models by using the test data, and select the deep learning algorithm with the best verification result.

Specifically, the verification module 66 uses the test data to verify all the algorithm models, and selects the deep learning algorithm with the best verification result.

It can be understood that after step 304, there are also steps, Calculate the similarity of the test data.

Specifically, the similarity calculation module 63 calculates the similarity of the test data; and adjusts the batch size of the test data according to the similarity.

Specifically, the adjustment module 64 adjusts the batch size of the corresponding training data according to the similarity.

The method for selecting a deep learning algorithm provided by the present invention utilizes the similarity of the training data to flexibly adjust the batch size of the training data, trains the selected deep learning algorithm based on the training data, and accurately obtains the corresponding algorithm model, and uses the training data to train the selected deep learning algorithm. The algorithm model is verified by the test data, and the selection method of the optimal deep learning algorithm is obtained.

The present invention can flexibly adjust the batch size of the training data by utilizing the similarity of the training data, thereby improving the quality of the training data, making the algorithm model obtained from the training data more accurate, and then using the verification test data to verify the algorithm model, and selecting a solution to solve the problem. The optimal algorithm for a specific problem, the above method reduces the difficulty of deep learning algorithm selection, and has strong versatility.

It will be apparent to those skilled in the art that the present invention is not limited to the details of the above-described exemplary embodiments, but that the present invention may be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments are to be regarded in all respects as illustrative and not restrictive, the scope of the present invention being defined by the appended claims rather than the foregoing description, and are therefore intended to fall within All changes within the meaning and scope of equivalents of the claims are encompassed within the present invention. should not Treat any reference sign in a claim as the claim to which the restriction relates. Furthermore, it is clear that the word "comprising" does not exclude other means or steps, and the singular does not exclude the plural.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention in any form. In addition, those skilled in the art can also make other changes within the spirit of the present invention. Of course, these changes made according to the spirit of the present invention should all be included within the scope of the claimed protection of the present invention.

Claims (8)

A method for selecting a deep learning algorithm, which is improved in that the method for selecting a deep learning algorithm includes: acquiring a problem type to be processed; selecting a corresponding data set according to the problem type, and classifying the selected data set into training data and test data; calculate the similarity of the training data; adjust the batch size of the training data according to the similarity, and the relationship between the similarity of the training data and the batch size satisfies: The batch size is inversely proportional to the similarity, the larger the similarity, the smaller the batch size, the smaller the similarity, the larger the batch size; choose a variety of depths according to the problem type Learning an algorithm, and using the training data to train the selected deep learning algorithm to obtain a corresponding algorithm model; using the test data to verify the algorithm model, and selecting the depth with the best verification result Learn algorithms. The method for selecting a deep learning algorithm according to claim 1, wherein the method for calculating the similarity of the training data includes the following steps: randomly selecting a plurality of sample groups in the training data, and each sample group includes For two samples, the similarity of each of the sample groups is calculated respectively, and then the average of the similarities of the plural sample groups is calculated, and the average is set as the similarity of the training data. The method for selecting a deep learning algorithm according to claim 1, wherein the problem types include image classification, object segmentation and object identification, and the deep learning algorithm is a convolutional neural network road algorithm, including ResNet, AlexNet, VGG and Inception. The method for selecting a deep learning algorithm according to claim 3, wherein the similarity of the training data of the convolutional neural network road algorithm is calculated by using structural similarity. A device for selecting a deep learning algorithm, which is improved in that the device for selecting a deep learning algorithm comprises: a processor; and a memory, wherein a plurality of program modules are stored in the memory, and the plurality of program modules are The processor runs and executes the following steps: acquiring the problem type to be processed; selecting a corresponding data set according to the problem type, and dividing the selected data set into training data and test data; calculating the similarity of the training data; The similarity adjusts the batch size of the training data, and the relationship between the similarity of the training data and the batch size satisfies: the batch size is inversely proportional to the similarity, and the similarity The larger the value is, the smaller the batch size is, the smaller the similarity is, the larger the batch size is; select multiple deep learning algorithms according to the problem type, and use the training data to train the selected A deep learning algorithm is used to obtain a corresponding algorithm model; the algorithm model is verified using the test data, and the deep learning algorithm with the best verification result is selected. The device for selecting a deep learning algorithm according to claim 5, wherein the method for calculating the similarity of the training data includes the following steps: randomly selecting a plurality of sample groups in the training data, each of the sample groups includes For two samples, the similarity of each of the sample groups is calculated respectively, and then the average of the similarities of the plural sample groups is calculated, and the average is set as the similarity of the training data. The device for selecting a deep learning algorithm according to claim 5, wherein the problem types include image classification, object segmentation and object identification, and the deep learning algorithm is a convolutional neural network road algorithm, including ResNet, AlexNet, VGG and Inception. The device for selecting a deep learning algorithm according to claim 7, wherein the similarity of the training data of the convolutional neural network road algorithm is calculated using structural similarity.

Images