KR102641628B1 - Method and system for providing service using segmented deep learning model - Google Patents

Abstract

A service provision method and system using a segmented deep learning model are disclosed. A service provision method according to an embodiment is a head model for an input image from a first module including the head model among a head model and a body model generated by dividing a pre-generated deep learning model. Receiving a service request including a first output value, inputting the first output value included in the received service request into the body model included in a second module to obtain a second output value, and obtaining the second output value It may include providing a service to the first module based on .

Description

Service provision method and system using segmented deep learning model {METHOD AND SYSTEM FOR PROVIDING SERVICE USING SEGMENTED DEEP LEARNING MODEL}

The explanation below relates to a service provision method and system using a segmented deep learning model.

SaaS (Software as a Service) is software that allows only the services that users need among the various functions of the software to be available through a network. This SaaS-type deep learning model service can provide various functions using deep learning models, such as face recognition functions and image classification functions, through the network. Users can call and use the desired function through the service's deep learning model API (Application Programming Interface).

However, SaaS-type deep learning model services have the problem that image transmission through a public network is essential. Regulations on the transmission of sensitive information, such as a person's face, through external networks are being strengthened, and there are many cases where some users want to use the service but do not want images to be collected from the client terminal to the server. .

[Prior document number]

Korean Patent Publication No. 10-2021-0082630

The previously created deep learning model is divided into a head model and a body model, and the data primarily processed through the head model is transferred from the client terminal containing the head model to the server containing the body model through a public network. A method of providing a service that can provide a SaaS-type deep learning model service without transmitting the image directly through a public network by transmitting it to and having the server secondarily process the data that was primarily processed through the body model; and Provides a system.

It provides technology to properly divide a previously created deep learning model into a head model and a body model.

Provides a service that covers data protection issues between the customer's system and the service provider's modules by using a first module containing a head model and a second module containing a body model within one system, such as the customer's legacy system. Provides methods and systems.

In the service provision method of a computer system including at least one processor, the head model among a head model and a body model generated by dividing a previously generated deep learning model by the at least one processor Receiving a service request including a first output value of the head model for an input image from a first module including; obtaining a second output value by inputting, by the at least one processor, the first output value included in the received service request into the body model included in a second module; and providing a service to the client terminal based on the second output value by the at least one processor.

According to one side, the first module and the second module may be included in the computer system.

According to another aspect, the second module is included in the computer system, and the first module is included in a separate computer device different from the computer system, and the second module is included in the computer device through network communication between the computer system and the separate computer device. It may be characterized as being linked to 2 modules.

According to another aspect, the deep learning model is selected based on at least one of the number of cumulative parameters processed by each of the nodes included in the deep learning model and the amount of data that must be transferred from the head model to the body model. It may be divided into the head model and the body model based on the node.

According to another aspect, the service providing method includes dividing the deep learning model into the head model and the body model by the at least one processor; and distributing the head model by the at least one processor.

According to another aspect, the dividing step includes selecting a candidate group including candidate nodes selected from among the nodes based on the weight value of each node included in the deep learning model; Based on at least one of the number of cumulative parameters processed by each of the nodes included in the deep learning model and the amount of data that must be transferred from the head model to the body model, a node that is a basis for division among the nodes is selected. selection step; and dividing the deep learning model into the head model and the body model based on the selected node.

According to another aspect, the step of selecting the candidate group includes setting the same first weight value for each of the nodes included in the deep learning model when the node is a parent node, and setting the same first weight value when the node is a child node. setting a second weight value; And a step of selecting a set of nodes for which a first weight value is set among nodes excluding the highest parent node and a leaf node as the candidate group, wherein the second weight value is set to the first weight value by the parent node. It may be characterized as including a value divided by the number of child nodes it contains.

According to another aspect, the service providing method further includes connecting an interface between the head model included in the first module and the body model included in the second module by the at least one processor, , the service request may be generated based on the connected interface.

According to another aspect, the cost added to the user of the first module in relation to the provision of the service may be affected by the capacity of the head model.

According to another aspect, the deep learning model may be divided into the head model and the body model in consideration of the capacity of the head model.

A computer program stored on a computer-readable recording medium is provided in conjunction with a computer system to execute the method on the computer system.

Provided is a computer-readable recording medium on which a computer program for executing the above method on a computer device is recorded.

At least one processor implemented to execute computer-readable instructions, wherein the head model and the body model are generated by dividing a previously generated deep learning model by the at least one processor. The body receives a service request including a first output value of the head model for an input image from a first module including a head model, and the second module includes the first output value included in the received service request. A computer device is provided, characterized in that it obtains a second output value by inputting it into a model, and provides a service to the first module based on the second output value.

The previously created deep learning model is divided into a head model and a body model, and the data primarily processed through the head model is transferred from the client terminal containing the head model to the server containing the body model through a public network. By transmitting it to and having the server secondarily process the data that was primarily processed through the body model, it is possible to provide a SaaS-type deep learning model service without transmitting the image directly through a public network.

It is possible to provide technology for appropriately dividing a previously created deep learning model into a head model and a body model.

Within one system, such as a customer's legacy system, a first module including a head model and a second module including a body model can be used to cover data protection issues between the customer's system and the service provider's modules.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention.
Figure 2 is a block diagram showing an example of a computer device according to an embodiment of the present invention.
Figure 3 is a diagram showing an example of a general view of a service provision system according to an embodiment of the present invention.
Figure 4 is a flowchart showing an example of a service provision method according to an embodiment of the present invention.
Figure 5 is a flowchart showing an example of a deep learning model segmentation method according to an embodiment of the present invention.
Figure 6 is a diagram showing an example of dividing a deep learning model in one embodiment of the present invention.
Figure 7 is a flowchart showing an example of a deep learning model segmentation method according to an embodiment of the present invention.
Figure 8 is a diagram illustrating an example of a process for dividing a deep learning model in one embodiment of the present invention.

Hereinafter, embodiments will be described in detail with reference to the accompanying drawings.

The service provision system according to embodiments of the present invention may be implemented by at least one computer device, and the service provision method according to embodiments of the present invention is performed through at least one computer device included in the service provision system. It can be. The computer program according to an embodiment of the present invention may be installed and driven in the computer device, and the computer device may perform the service providing method according to the embodiment of the present invention under the control of the driven computer program. The above-described computer program can be combined with a computer device and stored in a computer-readable recording medium to enable the computer to execute the service provision method.

1 is a diagram illustrating an example of a network environment according to an embodiment of the present invention. The network environment in FIG. 1 shows an example including a plurality of electronic devices 110, 120, 130, and 140, a plurality of servers 150 and 160, and a network 170. Figure 1 is an example for explaining the invention, and the number of electronic devices or servers is not limited as in Figure 1. In addition, the network environment in FIG. 1 only explains one example of environments applicable to the present embodiments, and the environment applicable to the present embodiments is not limited to the network environment in FIG. 1.

The plurality of electronic devices 110, 120, 130, and 140 may be fixed terminals or mobile terminals implemented as computer devices. Examples of the plurality of electronic devices 110, 120, 130, and 140 include smart phones, mobile phones, navigation devices, computers, laptops, digital broadcasting terminals, PDAs (Personal Digital Assistants), and PMPs (Portable Multimedia Players). ), tablet PC, etc. For example, in FIG. 1, the shape of a smartphone is shown as an example of the electronic device 110. However, in embodiments of the present invention, the electronic device 110 actually communicates with other devices through the network 170 using a wireless or wired communication method. It may refer to one of various physical computer devices capable of communicating with electronic devices 120, 130, 140 and/or servers 150, 160.

The communication method is not limited, and may include not only a communication method utilizing a communication network that the network 170 may include (for example, a mobile communication network, wired Internet, wireless Internet, and a broadcast network), but also short-range wireless communication between devices. For example, the network 170 may include a personal area network (PAN), a local area network (LAN), a campus area network (CAN), a metropolitan area network (MAN), a wide area network (WAN), and a broadband network (BBN). , may include one or more arbitrary networks such as the Internet. Additionally, the network 170 may include any one or more of network topologies including a bus network, star network, ring network, mesh network, star-bus network, tree or hierarchical network, etc. Not limited.

Each of the servers 150 and 160 is a computer device or a plurality of computers that communicate with a plurality of electronic devices 110, 120, 130, 140 and a network 170 to provide commands, codes, files, content, services, etc. It can be implemented with devices. For example, the server 150 provides services (e.g., cloud service, game service, content provision service, group call service) to a plurality of electronic devices 110, 120, 130, and 140 connected through the network 170. (or voice conference service), messaging service, mail service, social network service, map service, translation service, financial service, payment service, search service, etc.).

Figure 2 is a block diagram showing an example of a computer device according to an embodiment of the present invention. Each of the plurality of electronic devices 110, 120, 130, and 140 described above or each of the servers 150 and 160 may be implemented by the computer device 200 shown in FIG. 2.

As shown in FIG. 2, this computer device 200 may include a memory 210, a processor 220, a communication interface 230, and an input/output interface 240. The memory 210 is a computer-readable recording medium and may include a non-permanent mass storage device such as random access memory (RAM), read only memory (ROM), and a disk drive. Here, non-perishable large-capacity recording devices such as ROM and disk drives may be included in the computer device 200 as a separate permanent storage device that is distinct from the memory 210. Additionally, an operating system and at least one program code may be stored in the memory 210. These software components may be loaded into the memory 210 from a computer-readable recording medium separate from the memory 210. Such separate computer-readable recording media may include computer-readable recording media such as floppy drives, disks, tapes, DVD/CD-ROM drives, and memory cards. In another embodiment, software components may be loaded into the memory 210 through the communication interface 230 rather than a computer-readable recording medium. For example, software components may be loaded into memory 210 of computer device 200 based on computer programs installed by files received over network 170.

The processor 220 may be configured to process instructions of a computer program by performing basic arithmetic, logic, and input/output operations. Commands may be provided to the processor 220 by the memory 210 or the communication interface 230. For example, processor 220 may be configured to execute received instructions according to program code stored in a recording device such as memory 210.

The communication interface 230 may provide a function for the computer device 200 to communicate with other devices (eg, the storage devices described above) through the network 170. For example, a request, command, data, file, etc. generated by the processor 220 of the computer device 200 according to a program code stored in a recording device such as memory 210 is transmitted to the network ( 170) and can be transmitted to other devices. Conversely, signals, commands, data, files, etc. from other devices may be received by the computer device 200 through the communication interface 230 of the computer device 200 via the network 170. Signals, commands, data, etc. received through the communication interface 230 may be transmitted to the processor 220 or memory 210, and files, etc. may be stored in a storage medium (as described above) that the computer device 200 may further include. It can be stored as a permanent storage device).

The input/output interface 240 may be a means for interfacing with the input/output device 250. For example, input devices may include devices such as a microphone, keyboard, or mouse, and output devices may include devices such as displays and speakers. As another example, the input/output interface 240 may be a means for interfacing with a device that integrates input and output functions, such as a touch screen. The input/output device 250 may be configured as a single device with the computer device 200.

Additionally, in other embodiments, computer device 200 may include fewer or more components than those of FIG. 2 . However, there is no need to clearly show most prior art components. For example, the computer device 200 may be implemented to include at least some of the input/output devices 250 described above, or may further include other components such as a transceiver, a database, etc.

Figure 3 is a diagram showing an example of a general view of a service provision system according to an embodiment of the present invention. The service providing system 300 according to the embodiment of FIG. 3 may include a client terminal 310 and a server 320. Here, the client terminal 310 and the server 320 may each be implemented by at least one computer device 200. The client terminal 310 may also be expressed as an edge terminal.

First, the previously created deep learning model can be divided into a head model (311) and a body model (321). Here, the previously created deep learning model may mean a deep learning model that has completed learning. The type or learning of the deep learning model may vary depending on the service the service provider wishes to provide through the server 320. You will be able to easily understand the deep learning model itself and how to learn the deep learning model through already well-known technologies. The head model 311 may constitute the head portion of the deep learning model, and the body model 321 may constitute the body portion of the deep learning model.

Segmentation of this deep learning model may be performed by the server 320, but may also be performed by a computer device separate from the server 320. The divided head model 311 may be distributed to the client terminal 310, and the body model 321 may be placed on the server 320. Distribution of the head model 311 may also be performed by the server 320, but may also be performed by a computer device separate from the server 320. For example, when the head model 311 and the body model 321 are created by dividing the deep learning model in the server 320, the server 320 distributes the head model 311 to the client terminal 310. You can. If the head model 311 and the body model 321 are created by dividing the deep learning model in another computer device, the other computer device uses the head model 311 as the client terminal 310 and the body model ( 321) can be distributed to each server 320. These other computer devices may be implemented in the form of other nodes forming a cloud network together with the server 320, but are not limited thereto. The method of dividing the deep learning model into the head model 311 and the body model 321 will be described in more detail later.

Meanwhile, the service providing system 300 according to embodiments of the present invention can provide a service (or function) for processing images using a deep learning model in the form of SaaS (Software as a Service). . The embodiment of Figure 3 shows an example of providing a service for recognizing faces in images. To this end, the client terminal 310 may further include a face detector 312 that detects the face area 340 in the original image 330.

At this time, the client terminal 310 does not directly transmit the image of the detected face area 340 to the server 320, but an intermediate calculation value processed through the head model 311 included in the client terminal 310 ( An intermediate operation value can be created and transmitted to the server 320. At this time, the intermediate calculation value may be the first output value of the head model 311. The client terminal 310 may transmit a service request including this intermediate calculation value to the server 320 through a network (eg, network 170). At this time, the server 320 receives the service request, inputs the intermediate operation value included in the received service request into the body model 321 included in the server 320, and enters the final output value, which is the second output value of the body model 321. The output value can be obtained. Thereafter, the server 320 may provide a service corresponding to the service request to the client terminal 310 based on the final output value. In the embodiment of FIG. 3, a face recognition service according to the detected face area 340 may be provided to the client terminal 310.

At this time, the head model 311 and the body model 321 are divided from a previously created deep learning model. In this case, the final output value as the second output value obtained by inputting the first output value (intermediate operation value) obtained by inputting the image of the face area 340 into the head model 311 and inputting the image of the face area 340 into the body model 321 is the face area 340. It may be the same as the third output value obtained by inputting the image of to the deep learning model. On the other hand, the intermediate calculation value transmitted from the client terminal 310 to the server 320 is a mixed data value 350, and it is very difficult to obtain an image of the face area 340 from the intermediate calculation value. Accordingly, the service providing system 300 can provide an image processing service using a deep learning model in SaaS form without transmitting the image through a public network.

As such, according to embodiments of the present invention, in providing a SaaS-type deep learning model service, users can receive the service without exposing sensitive data to the service provider, thereby solving the problem of sharing sensitive information. .

Additionally, the amount of resource usage between the client terminal 310 and the server 320 can be adjusted through embodiments of the present invention. For example, by adjusting the capacity of the head model 311 and body model 321 created by dividing the deep learning model, the amount of resource usage required by each of the client terminal 310 and the server 320 can be adjusted. . Accordingly, the service provider can adjust the cost added to the user of the client terminal 310 in relation to the provision of the service based on the capacity of the head model 311. In other words, the entity dividing the deep learning model can divide the deep learning model into the head model 311 and the body model 321 by considering the capacity of the head model 311, and the service provider can divide the deep learning model into the head model 311 and the body model 321 by taking this capacity into consideration. Costs added to the user of the client terminal 310 in relation to the provision of the service can be adjusted.

Meanwhile, in the embodiment of FIG. 3, the head model 311 is installed on the client terminal 310 and the body model 321 is installed on the server 320, using network communication between the client terminal 310 and the server 320. An example in which the head model 311 and the body model 321 are linked has been described. However, depending on the embodiment, the head model 311 and the body model 321 may be installed in one physical electronic device. For example, within one system, such as a customer's legacy system, a first module including a head model 311 and a second module including a body model 321 may exist. In this case, it is possible to cover data protection issues between the customer's system module and the service provider's module.

Figure 4 is a flowchart showing an example of a service provision method according to an embodiment of the present invention. The service provision method according to this embodiment may be performed by the computer device 200 implementing the server 320 described above. At this time, the processor 220 of the computer device 200 may be implemented to execute control instructions according to the code of an operating system included in the memory 210 or the code of at least one computer program. Here, the processor 220 causes the computer device 200 to perform steps 410 to 440 included in the method of FIG. 4 according to control instructions provided by code stored in the computer device 200. can be controlled.

In step 410, the computer device 200 may connect an interface between the head model included in the client terminal and the body model included in the computer device 200. Here, the client terminal may correspond to the client terminal 310 described above with reference to FIG. 3. As described above, a head model and a body model can be created as the previously created deep learning model is divided, the head model can be distributed to a client terminal, and the body model is a computer device ( 200). At this time, the computer device 200 may connect the interface between the head model and the body model. The connection of the interface actually means a connection so that when the client terminal transmits the first output value of the head model, the computer device 200 uses the transmitted first output value as an input of the body model corresponding to the head model. You can.

In step 420, the computer device 200 may receive a service request including the first output value of the head model for the input image from the client terminal through the network. At this time, the client terminal may generate a service request based on the connected interface in step 410. More specifically, the client terminal may generate a service request through an API call of the deep learning model service, and this API call may be based on the interface connected in step 410. Meanwhile, the input image may refer to an input image from the perspective of a deep learning model. Taking the embodiment of FIG. 3 as an example, the image of the face area 340 detected in the original image 330 and input as the head model may be the input image described in step 420. Considering the case where the original image 330 is input as a head model for the image classification service, the original image 330 may be used as the input image as is.

In step 430, the computer device 200 may obtain a second output value by inputting the first output value included in the received service request into the body model. At this time, as already described, the second output value may be the same as the output value when input to the deep learning model before segmenting the input image. In other words, the computer device 200 implementing the server 320 can obtain the output value of the deep learning model for the input image without directly receiving the input image from the client terminal.

In step 440, the computer device 200 may provide a service to the client terminal based on the second output value. Examples of services include, but are not limited to, facial recognition services and image classification services, as described above. For example, it will be easy to understand that the service provision method according to embodiments of the present invention can be applied to all types of services provided by processing images through a deep learning model.

Figure 5 is a flowchart showing an example of a deep learning model segmentation method according to an embodiment of the present invention. As already described, segmentation of the deep learning model may be performed by the server 320 or may be performed by a separate computer device. The embodiment of FIG. 5 describes an example performed by the computer device 200 implementing the server 320. Steps 510 and 520 included in the deep learning model segmentation method of FIG. 5 may be performed before step 410 of FIG. 4.

In step 510, the computer device 200 may divide the deep learning model into a head model and a body model. To this end, step 510 may include steps 511 and 512.

In step 511, the computer device 200 selects one of the nodes based on at least one of the number of cumulative parameters processed by each of the nodes included in the deep learning model and the amount of data that must be transferred from the head model to the body model. You can select a node that becomes the basis for division. The operations of the layers of the deep learning model are already described, and the weights are included in the matrix. At this time, there are sections where there is a break between matrices and operations between the matrices, and nodes in these sections may be selected according to the number of accumulated parameters and/or the amount of data.

For example, at each node of a deep learning model, the amount of calculation is determined based on the number of accumulated parameters, and when a graph file is uploaded to the inference engine, the deep learning model processes calculations sequentially one by one and outputs the results. Since using, how long the calculation time will take and how many resources will be used can be determined depending on the number of accumulated parameters, and the computer device 200 can use the head model and body model based on the calculation time, resource usage, etc. An appropriate section for dividing can be determined. For example, the computer device 200 may receive in advance the calculation time and/or resource usage defined for segmentation of the deep learning model. In this case, the computer device 200 may select a node with a cumulative number of parameters that satisfies the pre-entered calculation time and/or resource usage as a reference node for dividing the deep learning model into a head model and a body model. there is. In addition to the number of these accumulated parameters, the amount of data to be sent from the head to the body described above can be used to determine an appropriate section for dividing the head model and the body model. For example, the computer device 200 uses a section of a specific node that transmits a specific amount of data (for example, the amount of data included in a preset range) through the network as a section for dividing the deep learning model. You can decide. Depending on the embodiment, the complexity of the layer may be used to split the deep learning model into a head model and a body model. For example, the computer device 200 may calculate the complexity of a layer for each node based on the number of layers included in each node, and based on nodes with low complexity of these layers, the deep learning model may be used to create a head model and a body. It can be divided into models.

In step 512, the computer device 200 may divide the deep learning model into a head model and a body model based on the selected node. At this time, the selected node may be included in either the head model or the body model.

At step 520, computer device 200 may distribute the head model. In one example, computer device 200 may distribute the head model to client devices, but embodiments in which multiple client devices are present may also be considered. In this case, the computer device 200 may distribute the head model to each of a plurality of client devices. Meanwhile, when another computer device divides the deep learning model into a head model and a body model, the other computer device can distribute the head model to the client device and the body model to the server 320, respectively.

In this way, the deep learning model is based on the head model and It can be divided into body models. At this time, the capacity of the head model may affect the cost added to the user of the client terminal in relation to the provision of the service. For example, as the capacity of the head model increases, resource usage at the client terminal may increase. Conversely, the resource usage of the server 320 may decrease. Therefore, the service provider can adjust the cost added to the user of the client terminal in relation to the provision of services according to the capacity of the head model, and when necessary, consider the capacity of the head model and use the deep learning model as the head model and body model. It can be divided into As a more specific example, a number of candidate nodes that serve as the basis for division are selected based on at least one of the number of cumulative parameters processed by each of the nodes included in the deep learning model and the amount of data that must be transferred from the head model to the body model. Afterwards, the deep learning model can be divided based on the node that is finally selected according to the capacity of the head model.

Figure 6 is a diagram showing an example of dividing a deep learning model in one embodiment of the present invention. Figure 6 generally shows a deep learning model and shows an example of dividing this deep learning model into a head model 610 and a body model 620. At this time, the embodiment of FIG. 6 shows an example in which the “Add” node is selected as the basis for division. In the embodiment of Figure 6, the output of the head model 610 is the output of the "Add" node, and deriving an input image to the deep learning model through the output value becomes a very difficult problem. Therefore, even if the output value of the head model 610 is exposed through a public network, sensitive information such as the input image can be prevented from being exposed.

Meanwhile, through the above embodiments, it can be understood that the output value of the head model may be more than one depending on which node the deep learning model is divided based on. Additionally, it will be easy to understand that connecting the interface between the head model and the body model includes connecting the output of the head model and the input of the body model.

Figure 7 is a flowchart showing an example of a deep learning model segmentation method according to an embodiment of the present invention. The deep learning model segmentation method according to this embodiment can be performed by the computer device 200. At this time, the computer device 200 of this embodiment may be the same physical electronic device as the computer device 200 that implements the server 320 described above, but according to the embodiment, the computer device 200 that implements the server 320 ) may be a different physical electronic device. For example, the deep learning model may be divided into a head model and a body model by the computer device 200 according to this embodiment, with the head model being divided into the client terminal 310 and the body model being divided into the server 320. It can be distributed. If the computer device 200 according to this embodiment is the same physical electronic device as the computer device 200 implementing the server 320, the head model may be distributed to the client terminal 310, and the body model may be distributed to the client terminal 310. It may be installed on the computer device 200.

At this time, the processor 220 of the computer device 200 according to this embodiment may be implemented to execute control instructions according to the code of an operating system included in the memory 210 or the code of at least one computer program. Here, the processor 220 causes the computer device 200 to perform steps 710 to 730 included in the method of FIG. 7 according to control instructions provided by code stored in the computer device 200. can be controlled.

In step 710, the computer device 200 may select nodes to be included in the candidate group for the split point among the nodes of the previously created deep learning model. At this time, the nodes of the deep learning model can be divided into parent nodes or child nodes. In this case, the weight value of the parent node starts with 1, and the weight value of the child node can be determined by dividing the weight of all parent nodes by the number of children of the parent node. For example, when one parent node has two child nodes, when the weight value of the parent node is 1, the weight value of each of the two child nodes may be 0.5. The weight value of each node may be determined by the computer device 200. Meanwhile, since the nodes in the candidate group are used as dividing points for dividing the deep learning model, the highest parent node and leaf node that cannot be used as a standard for dividing the deep learning model may be excluded from the selection of the candidate group. In other words, with a weight value of 1, nodes other than the top parent node and leaf node can be selected as a candidate group.

In step 720, the computer device 200 may select a node as a split point among the nodes included in the candidate group based on at least one of the number and shape value of the cumulative parameters of each of the nodes included in the candidate group. there is. The number of accumulated parameters may refer to the number of accumulated parameters associated with the computational amount of each node, and the type value may be related to the data transmission amount according to node division. For example, the amount of data to be transmitted from the head node to the body node may be determined according to the type value. If the output size of the node shown in Figure 6 {1 × 32 × 32 × 32}, the shape value can be 32,768 (1 × 32 × 32 × 32).

In step 730, the computer device 200 may divide the deep learning model into a head model and a body model based on the node as the selected division point. As already explained, the head model and body model can be installed in different physical electronic devices and used to provide services through deep learning models so that sensitive information in the input image is not exposed.

Figure 8 is a diagram illustrating an example of a process for dividing a deep learning model in one embodiment of the present invention. Figure 8 briefly represents the nodes included in the deep learning model as ovals. At this time, the numbers in the oval are the weights of the corresponding nodes. As already explained, the weight of the parent node can be 1, and the weight of the child node can be the weight of the parent node divided by the number of children of the parent node of the child node. . At this time, the nodes to be included in the candidate group may have a weight value of 1 and may be nodes excluding the highest parent node and leaf node. The example in Figure 8 shows candidate node 1 and candidate node 2 that satisfy the corresponding conditions. At this time, the node that serves as the division point of the deep learning model among candidate node 1 and candidate node 2 may be determined by at least one of the number and shape value of the accumulated parameters. As already explained, the accumulated parameters are related to the amount of computation of the node, and the type value may be related to the amount of data transmission according to node division. Therefore, depending on which candidate node the deep learning model is divided into, the amount of computation of the head node or the amount of data transmission from the head node to the body node can be determined.

As such, according to embodiments of the present invention, the previously generated deep learning model is divided into a head model and a body model and is primarily processed through the head model in a client terminal including the head model. By transmitting data to a server containing a body model through a public network and having the server secondarily process the data that was primarily processed through the body model, SaaS-type deep image processing is possible without directly transmitting the image through a public network. A learning model service can be provided. Additionally, it is possible to provide technology for appropriately dividing a previously created deep learning model into a head model and a body model. In addition, data protection issues between the customer's system and the service provider's module can be covered by using a first module including a head model and a second module including a body model within one system, such as the customer's legacy system. .

The device described above may be implemented with hardware components or a combination of hardware components and software components. For example, the devices and components described in the embodiments include a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, a microcomputer, a field programmable gate array (FPGA), and a programmable logic unit (PLU). It may be implemented using one or more general-purpose or special-purpose computers, such as a logic unit, microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. Additionally, a processing device may access, store, manipulate, process, and generate data in response to the execution of software. For ease of understanding, a single processing device may be described as being used; however, those skilled in the art will understand that a processing device includes multiple processing elements and/or multiple types of processing elements. It can be seen that it may include. For example, a processing device may include a plurality of processors or one processor and one controller. Additionally, other processing configurations, such as parallel processors, are possible.

Software may include a computer program, code, instructions, or a combination of one or more of these, which may configure a processing unit to operate as desired, or may be processed independently or collectively. You can command the device. The software and/or data may be embodied in any type of machine, component, physical device, computer storage medium or device for the purpose of being interpreted by or providing instructions or data to the processing device. there is. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. At this time, the medium may continuously store a computer-executable program, or temporarily store it for execution or download. In addition, the medium may be a variety of recording or storage means in the form of a single or several pieces of hardware combined. It is not limited to a medium directly connected to a computer system and may be distributed over a network. Examples of media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical recording media such as CD-ROMs and DVDs, magneto-optical media such as floptical disks, And there may be something configured to store program instructions, including ROM, RAM, flash memory, etc. Additionally, examples of other media include recording or storage media managed by app stores that distribute applications, sites or servers that supply or distribute various other software, etc.

As described above, although the embodiments have been described with limited examples and drawings, various modifications and variations can be made by those skilled in the art from the above description. For example, the described techniques are performed in a different order than the described method, and/or components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or other components are used. Alternatively, appropriate results may be achieved even if substituted or substituted by an equivalent.

Therefore, other implementations, other embodiments, and equivalents of the claims also fall within the scope of the claims described below.

Claims (12)

In a service provision method of a computer system including at least one processor,
An image using the deep learning model from the first module including the head model among the head model and body model generated by the at least one processor through division of the learned deep learning model. Receiving a request for processing service, wherein the request includes a first output value of the head model for an input image;
obtaining a second output value by inputting, by the at least one processor, the first output value included in the received image processing service request into the body model included in a second module; and
Providing the image processing service to the first module based on the second output value, by the at least one processor
Including,
The head model and the body model are generated by dividing the deep learning model based on a first node selected from among the nodes included in the deep learning model,
The first node is selected again from candidate nodes selected from among the nodes based on the weight values of each of the nodes,
The weight value of each of the nodes is determined depending on whether each of the nodes is a parent node and the number of child nodes included in the parent node of each of the nodes.
A service provision method characterized by . According to paragraph 1,
A service provision method, characterized in that the first module and the second module are included in the computer system. According to paragraph 1,
the second module is included in the computer system,
A method of providing a service, wherein the first module is included in a separate computer device that is different from the computer system and is linked to the second module through network communication between the computer system and the separate computer device. According to paragraph 1,
The first node is selected from among the candidate nodes based on at least one of the number of cumulative parameters processed by each of the nodes and the amount of data to be transferred from the head model to the body model. How we provide services. According to paragraph 1,
Splitting the deep learning model into the head model and the body model by the at least one processor; and
distributing, by the at least one processor, the head model
A service provision method further comprising: According to clause 5,
The dividing step is,
selecting a candidate group including the candidate nodes;
selecting the first node as a basis for division in the candidate group based on at least one of the number of cumulative parameters processed by each of the nodes and the amount of data to be transferred from the head model to the body model; and
Splitting the deep learning model into the head model and the body model based on the selected first node.
A service provision method comprising: According to clause 6,
The step of selecting the candidate group is,
For each of the nodes included in the deep learning model, setting the same first weight value when the node is a parent node and setting a second weight value when the node is a child node; and
Selecting a set of nodes for which a first weight value is set among nodes excluding the highest parent node and leaf node as the candidate group.
Including,
The second weight value includes the first weight value divided by the number of child nodes included in the parent node.
A service provision method characterized by . According to paragraph 1,
Connecting an interface between the head model included in the first module and the body model included in the second module, by the at least one processor.
It further includes,
The request for the image processing service occurs through the connected interface.
A service provision method characterized by . According to paragraph 1,
A service provision method, wherein the cost added to the user of the first module in relation to the provision of the image processing service is affected by the capacity of the head model. According to clause 9,
A service provision method characterized in that the deep learning model is divided into the head model and the body model in consideration of the capacity of the head model. A computer program stored in a computer-readable recording medium in combination with a computer system to execute the method of any one of claims 1 to 10 on the computer system. At least one processor implemented to execute computer readable instructions
Including,
By the at least one processor,
Receiving a request for an image processing service using the deep learning model from the first module including the head model among the head model and body model created through division of the deep learning model for which learning has been completed, The request includes a first output value of the head model for the input image,
Obtaining a second output value by inputting the first output value included in the received image processing service request into the body model included in a second module,
Providing the image processing service to the first module based on the second output value,
The head model and the body model are generated by dividing the deep learning model based on a first node selected from among the nodes included in the deep learning model,
The first node is selected again from candidate nodes selected from among the nodes based on the weight values of each of the nodes,
The weight value of each of the nodes is determined depending on whether each of the nodes is a parent node and the number of child nodes included in the parent node of each of the nodes.
A computer device characterized by a.

Images