JP7037180B2 - Learning data discriminator and learning data discriminator - Google Patents

Description

The present invention relates to a learning data discriminating device and a learning data discriminating program, and more particularly to a technique for discriminating which learning algorithm is suitable for a data set prepared for learning.

In recent years, expectations for machine learning, which is a field of artificial intelligence (AI), are increasing. Machine learning refers to the process of creating a model from a large number of data, and is roughly divided into supervised learning and unsupervised learning depending on the nature of the data used for learning. Semi-supervised learning and reinforcement learning are also available as intermediate learning methods between supervised learning and unsupervised learning.

Machine learning is rapidly being applied to various products and services, and appropriate algorithms are applied according to the application. For example, Patent Document 1 discloses that a learning algorithm group in which a plurality of types of learning algorithms are prepared is provided, and an algorithm used for actual learning is selected according to a problem to be solved.

Whichever training algorithm is used, it is necessary to collect the data used for training to build the model. And just as there is an appropriate learning algorithm depending on the application, there is appropriate data to build the model according to the learning algorithm. In the past, developers had to manually determine which algorithm the collected dataset could be used to train. Therefore, there is a problem that the data set cannot be used properly for learning unless the developer is familiar with machine learning of various algorithms.

Japanese Unexamined Patent Publication No. 5-298277

The present invention has been made to solve such a problem, and it is appropriate to determine which learning algorithm the data set is suitable for, even if the developer is not proficient in machine learning of various algorithms. The purpose is to be able to do it.

In order to solve the above-mentioned problems, in the present invention, the existence or nonexistence of a folder is determined as a storage format of the data set provided for learning, and a learning algorithm suitable for the data set is presented according to the result of the determination. I try to do it.

According to the present invention configured as described above, a learning algorithm suitable for the data set is presented depending on whether or not the data set provided for training is stored in the folder. It eliminates the need for developers to manually determine which algorithms the dataset can be used to train. As a result, even a developer who is not proficient in machine learning of various algorithms can appropriately determine which learning algorithm the data set is suitable for.

It is a figure which shows the configuration example of the communication system which applied the learning data discrimination apparatus by this embodiment. It is a block diagram which shows the functional structure example of the learning data discrimination apparatus by this embodiment.

Hereinafter, an embodiment of the present invention will be described with reference to the drawings. FIG. 1 is a diagram showing a configuration example of a communication system to which the learning data discrimination device according to the present embodiment is applied. As shown in FIG. 1, the communication system of the present embodiment includes a user terminal 10 and a server device 20, and both are configured to be connectable via a communication network 30 such as the Internet or a mobile phone network.

The server device 20 corresponds to the learning data discriminating device of the present embodiment, and discriminates which learning algorithm the data set transmitted from the user terminal 10 is suitable for. For example, the user operates the user terminal 10 to collectively store a data set composed of a plurality of data in a compressed file, and upload the compressed file to the server device 20. The server device 20 decompresses the compressed file transmitted from the user terminal 10 and determines which learning algorithm the provided data set is suitable for.

FIG. 2 is a block diagram showing a functional configuration example of the learning data discrimination device according to the present embodiment mounted on the server device 20. As shown in FIG. 2, the learning data discriminating device of the present embodiment includes a data set acquisition unit 21, a storage format determination unit 22, a data size determination unit 23, and an algorithm presentation unit 24 as its functional configuration. Further, the learning data discriminating device of the present embodiment includes a corresponding table storage unit 25 as a storage medium.

Each of the above functional blocks 21 to 24 can be configured by any of hardware, DSP (Digital Signal Processor), and software. For example, when configured by software, each of the above functional blocks 21 to 24 is actually configured to include a computer CPU, RAM, ROM, etc., and is a program stored in a recording medium such as RAM, ROM, hard disk, or semiconductor memory. Is realized by the operation of.

The data set acquisition unit 21 acquires the data set provided for learning from the user terminal 10. As described above, when the data set is sent as a compressed file, the data set acquisition unit 21 decompresses the compressed file and takes out the data set. In this embodiment, the data set acquired by the data set acquisition unit 21 is assumed to be a set of a plurality of image data.

The storage format determination unit 22 determines the presence or absence of a folder as the storage format of the data set acquired by the data set acquisition unit 21. That is, when the user uploads the data set from the user terminal 10 to the server device 20, when a plurality of data are transmitted without being put in the folder, the storage format determination unit 22 is the data acquired by the data set acquisition unit 21. Judge that the save format of the set is "no folder". On the other hand, when the user uploads the data set from the user terminal 10 to the server device 20, when a plurality of data are put in one or a plurality of folders and transmitted, the storage format determination unit 22 is performed by the data set acquisition unit 21. It is determined that the storage format of the acquired data set is "with folder".

When the storage format determination unit 22 determines that the storage format of the data set is "there is a folder", the data size determination unit 23 has an aspect ratio of a plurality of image data constituting the data set of 2: 1 or 1. : Determine whether or not it is unified with 2. The aspect ratio of the image data here means the ratio between the vertical size and the horizontal size of the rectangular image.

The algorithm presentation unit 24 determines and presents a learning algorithm suitable for the data set according to the result of the determination by the storage format determination unit 22 and the data size determination unit 23. At the time of this determination, the algorithm presentation unit 24 refers to the corresponding table storage unit 25. The correspondence table storage unit 25 stores table information showing the correspondence relationship between each determination element consisting of the presence / absence of a folder and the aspect ratio of image data and a learning algorithm suitable for the content of each determination element. There is.

The algorithm presentation unit 24 transmits, for example, a suitable learning algorithm determined by using the table information to the user terminal 10 and presents it to the user. As the presentation method, for example, a method of displaying the name of the learning algorithm as text information on the screen of the user terminal 10 can be considered. As another example of the presentation method, the following may be used. That is, the algorithm presentation unit 24 stores information indicating the determined suitable learning algorithm in the server device 20, and when a request for providing information is made through access to the server device 20 from the user terminal 10, a predetermined presentation screen is displayed. Information on the learning algorithm is provided to the user terminal 10 through.

The specific processing contents by the algorithm presentation unit 24 will be described below. When the storage format determination unit 22 determines that the storage format of the data set is "no folder", the algorithm presentation unit 24 presents at least unsupervised learning as a learning algorithm suitable for the data set.

On the other hand, when the storage format determination unit 22 determines that the storage format of the data set is "with a folder", the algorithm presentation unit 24 presents at least supervised learning as a learning algorithm suitable for the data set. The algorithm presentation unit 24 may present at least one of image classification and GAN as a specific example of supervised learning.

When a user puts a dataset in a folder, the folder is always given a folder name. Usually, the folder name is a name that represents a concept common to the data sets contained therein. In this case, the folder name may be used as teacher data for image classification. Therefore, when the storage format determination unit 22 determines that the storage format of the data set is "with a folder", the algorithm presentation unit 24 presents image classification as one of the learning algorithms suitable for the data set with a folder. do.

Also, when a user puts a dataset in a folder, the image data put in the same folder may all indicate images of the same concept. In this case, there is a possibility that the image data in the same folder can be used as the correct answer data of GAN. For example, when the data set is divided into a plurality of folders, it is possible to specify one of the folders and perform GAN learning. Therefore, when the storage format determination unit 22 determines that the storage format of the data set is "with a folder", the algorithm presentation unit 24 presents GAN as one of the learning algorithms suitable for the data set with a folder. ..

Here, as a specific example of supervised learning in which a data set with folders is suitable, an example of presenting at least one of image classification and GAN has been described, but other algorithms for supervised learning are presented. You may do so. For example, when the data set acquired by the data set acquisition unit 21 is stored in a plurality of folders, CycleGAN or DiscoGAN, which are variations of GAN, may be presented.

In this case, it is possible to select any two folders from the plurality of folders for learning. That is, learning CycleGAN such as selecting any two folders from a plurality of folders and performing style conversion between the image data stored in one folder and the image data stored in the other folder. May be possible. In addition, DiscoGAN learning such as selecting any two folders from a plurality of folders and grasping the relationship (attribute) between the image data stored in one folder and the image data stored in the other folder. May be able to do.

Further, in the algorithm presentation unit 24, the storage format determination unit 22 determines that the data set storage format is “with a folder”, and the data size determination unit 23 determines that the aspect ratio of the plurality of image data is 2: 1 or 1: 1. If it is determined that the data set is unified in 2, pix2pix may be further presented as a learning algorithm suitable for the data set.

The learning algorithm of pix2pix tries to express the image transformation hidden between two images by DNN (deep neural network). When the aspect ratio of a plurality of image data stored in a folder is unified to 2: 1 or 1: 2, two images having an aspect ratio of 1: 1 are placed one above the other in one image data. Alternatively, it can be estimated that they may have been recorded side by side. Therefore, in this case, there is a possibility that the learning of the pix2pix can be performed using the two images recorded side by side, and the algorithm presenting unit 24 uses the pix2pix as one of the learning algorithms suitable for the data set. To present.

In contrast to the above datasets with folders, in the case of datasets without folders, they are not necessarily image data with a common concept. Nor does the folder name that should be the teacher data exist. Therefore, when the storage format determination unit 22 determines that the storage format of the data set is "no folder", the algorithm presentation unit 24 basically performs unsupervised learning as a learning algorithm suitable for the data set. Present.

However, the algorithm presentation unit 24 may further present pix2pix as a learning algorithm suitable for a data set without folders. This is because there is a possibility that two images are recorded side by side in one image data. Even for the data set without folders, the data size determination unit 23 determines whether or not the aspect ratio of the image data is unified at 2: 1 or 1: 2, and it is determined that they are unified. In some cases, pix2pix may be presented as a learning algorithm suitable for the dataset.

As described in detail above, in the present embodiment, the presence / absence of a folder is determined as the storage format of the data set provided for learning, and a learning algorithm suitable for the data set is presented according to the result of the determination. I am doing it. Further, in the present embodiment, in addition to the data set storage format, the data size of a plurality of image data is also determined, and the data set is set according to whether or not the aspect ratio is unified at 2: 1 or 1: 2. I try to present a suitable learning algorithm.

According to the present embodiment configured in this way, whether or not the data set provided for learning is stored in the folder, and the aspect ratio of the plurality of image data are unified to 2: 1 or 1: 2. Since the learning algorithm suitable for the data set is presented depending on whether or not the data set is present, it is necessary for the user to manually determine which algorithm the data set collected by the user can be used for learning. It disappears. As a result, even a developer who is not proficient in machine learning of various algorithms can appropriately determine which learning algorithm the data set is suitable for.

In the above embodiment, in the system configuration in which the user terminal 10 and the server device 20 are connected by the communication network 30, the learning data discrimination device of the present embodiment is mounted on the server device 20, and the server device 20 is the user terminal 10. Although an example of acquiring a data set from is described, the present invention is not limited thereto. For example, the learning data discrimination device of the present embodiment may be mounted on the user terminal 10 so that the user terminal 10 acquires a data set from an external server device, a removable storage medium, or the like.

Further, the data size determination unit 23 shown in the above embodiment is not an essential configuration in the present invention, and may be omitted.

Further, in the above embodiment, an example in which image data is used as an example of a data set has been described, but the present invention is not limited thereto. For example, the present embodiment can be applied even when voice data, text data, or data in other formats are used.

In addition, the above embodiments are merely examples of the embodiment of the present invention, and the technical scope of the present invention should not be construed in a limited manner. That is, the present invention can be implemented in various forms without departing from its gist or its main features.

10 User terminal 20 Server device (learning data discrimination device)
21 Data set acquisition unit 22 Storage format determination unit 23 Data size determination unit 24 Algorithm presentation unit 25 Corresponding table storage unit

Claims (7)

A dataset acquisition unit that acquires the dataset provided for training,
As the storage format of the data set acquired by the above data set acquisition unit, a storage format determination unit that determines the existence of a folder and a storage format determination unit.
A learning data discrimination device including an algorithm presentation unit that presents a learning algorithm suitable for the data set according to the result of determination by the storage format determination unit. The above data set is a set of multiple image data.
When the storage format determination unit determines that the storage format of the data set is no folder, the algorithm presentation unit presents at least unsupervised learning as a learning algorithm suitable for the data set, and the storage format determination unit. The data discriminating device for learning according to claim 1, wherein when it is determined that the storage format of the data set has a folder, at least supervised learning is presented as a learning algorithm suitable for the data set. 2. The algorithm presenting unit further presents pix2pix as a learning algorithm suitable for the data set when the storage format determination unit determines that the storage format of the data set is no folder. The learning data discriminating device described in 1. When the storage format determination unit determines that the storage format of the data set has a folder, the algorithm presentation unit presents at least one of image classification and GAN as a learning algorithm suitable for the data set. The learning data discriminating device according to claim 2 or 3. When the storage format determination unit determines that the data set has a folder and the data set is divided into a plurality of folders, the algorithm presentation unit determines that the data set has a folder. The learning data discriminating device according to any one of claims 2 to 4, wherein at least one of CycleGAN and DiscoGAN is presented as a suitable learning algorithm. Further, a data size determination unit for determining whether or not the aspect ratio of a plurality of image data constituting the above data set is unified at 2: 1 or 1: 2 is provided.
When the data size determination unit determines that the aspect ratio of the plurality of image data is unified at 2: 1 or 1: 2, the algorithm presentation unit can be used as a learning algorithm suitable for the data set. The data discriminating device for learning according to any one of claims 2 to 5, further comprising presenting pix2pix. A dataset acquisition method, which acquires the dataset provided for training.
As the storage format of the data set acquired by the data set acquisition means, the storage format determination means for determining the existence of a folder and the learning algorithm for which the data set is suitable according to the result of the determination by the storage format determination means. A data discriminating program for learning to make a computer function as a means of presenting an algorithm.

Images