System and method for determining memory capacity when learning cell images

Abstract

This system (200) for assessing memory capacity during learning of a cell image (80) comprises: a learning processing unit (10) provided with a first processor (10a) for training a learning model (21), and a memory (10b); a selection unit (45) for selecting between a training mode in which the learning model is trained and a verification mode in which verification as to whether the memory capacity is running short is carried out; an assessment unit (12d) for assessing, in the verification mode, whether the memory capacity has run short; and a display unit (121) for issuing a notification based on a first assessment result (32).

Description

[Technical field]

[0001] The present invention relates to a memory capacity determination system and a memory capacity determination method for learning cell images. [Background technology]

[0002] Conventionally, a technology for generating a learning model for analyzing cell images has been disclosed. Such a technology for generating a learning model for analyzing cell images is disclosed, for example, in JP 2021-64115 A.

[0003] Patent Publication No. 2021-64115 discloses a configuration for generating a trained model by performing machine learning using training data in which a cell image is used as an input image and a stained image obtained by staining the cytoskeleton is used as a correct image. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-64115 Summary of the Invention [Problem to be solved by the invention]

[0005] Here, although not disclosed in JP 2021-64115 A, machine learning is performed using a processor such as a GPU (Graphics Processing Unit) and a memory that serves as a working area during learning. If the machine learning ends normally, a trained model is generated. On the other hand, if the machine learning ends abnormally, the data used for learning and the learning conditions are changed, and machine learning is performed again. One of the causes of abnormal termination of machine learning is considered to be insufficient memory capacity during learning. However, in the configuration for performing machine learning as disclosed in JP 2021-64115 A, the user cannot know that the machine learning has not been completed due to insufficient memory capacity. Therefore, a technology that allows the user to know that the machine learning has not been completed due to insufficient memory capacity is desired.

[0006] The present invention has been made to solve the above-mentioned problems, and one object of the present invention is to provide a memory capacity determination system and a memory capacity determination method for learning cell images, which allow a user to understand that machine learning has not been completed due to insufficient memory capacity. [Means for solving the problem]

[0007] In order to achieve the above-mentioned object, a memory capacity determination system during learning of a cell image in a first aspect of the present invention comprises a processor that uses cell images to learn a learning model under a first learning condition a predetermined number of times, a learning processing unit that has a memory to be used as a working area for the learning process of the learning model, a selection unit that selects between a learning mode in which the learning model is learned by performing learning processing by the learning processing unit, and a verification mode in which it is verified whether or not there is insufficient memory capacity when performing the learning process by the learning processing unit, a judgment unit that judges whether or not there is insufficient memory capacity during the learning process of the learning model in the verification mode, and an alarm unit that issues an alarm based on the judgment result by the judgment unit.

[0008] A method for determining memory capacity during learning of a cell image in a second aspect of the present invention includes the steps of: learning a learning model under a first learning condition using cell images a predetermined number of times; selecting a learning mode in which the learning model is learned by performing a learning process of the learning model; and a verification mode in which it is verified whether the memory capacity used for the working area when performing the learning process of the learning model is insufficient; determining whether the memory capacity has become insufficient in the learning process of the learning model; and notifying the result of the determination of whether the memory capacity has become insufficient. Effect of the Invention

[0009] In the memory capacity determination system for learning a cell image in the first aspect and the memory capacity determination method for learning a cell image in the second aspect, in a verification mode in which verification is performed to verify whether or not the memory capacity used for the working area is insufficient when performing a learning process of a learning model, a determination is made as to whether or not the memory capacity is insufficient. Here, the memory capacity secured as the working area is equal in both the learning mode and the verification mode. Therefore, by performing verification in the verification mode, it is possible to determine whether or not the memory capacity is insufficient in the learning mode. In addition, the determination result as to whether or not the memory capacity is insufficient is notified. Therefore, by performing verification in the verification mode before performing learning of a learning model in the learning mode, a user can know whether or not the memory capacity is insufficient before performing a learning process in the learning mode. As a result, the user can know that machine learning was not completed due to a lack of memory capacity. In addition, since it is possible to select the learning mode and the verification mode, when it is not necessary to perform verification in the verification mode, the learning model can be trained in the learning mode without performing a learning process in the verification mode. As a result, it is possible to improve the user's convenience. [Brief description of the drawings]

[0010] [Figure 1]1 is a block diagram showing a memory capacity determination system including an image processing device according to an embodiment of the present invention. [Diagram 2] FIG. 2 is a schematic diagram for explaining a working area and a memory capacity during learning. [Diagram 3] 4 is a functional block diagram for explaining functions of a second processor of the image processing device. FIG. [Figure 4] 13A and 13B are diagrams showing an example of a screen displayed when memory capacity is not insufficient when performing a learning process in a verification mode in the image processing device according to the present embodiment. [Diagram 5] FIG. 13 is a diagram showing an example of a screen displayed when memory capacity is not insufficient during learning processing in a verification mode. [Figure 6] 11A and 11B are diagrams illustrating an example of a screen displayed when memory capacity is insufficient when performing a learning process in a verification mode in the image processing device according to the present embodiment. [Figure 7] 13A and 13B are diagrams illustrating an example of a screen displayed when memory capacity is insufficient during learning processing in a verification mode. [Figure 8] 10 is a flowchart illustrating a memory capacity verification process in the image processing apparatus according to the present embodiment. [Figure 9] 11 is a flowchart for explaining a process of optimizing a first learning condition while changing a batch size in the image processing device according to the present embodiment. [Figure 10] FIG. 13 is a block diagram showing a memory capacity determination system including an image processing device according to a modified example. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0011] DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, preferred embodiments of the present invention will be described with reference to the accompanying drawings.

[0012] The configuration of a memory capacity determination system 200 for studying a cell image 80 including an image processing device 100 according to this embodiment and a memory capacity determination method for studying a cell image 80 will be described with reference to FIGS. 1 to 7.

[0013] Image Processing System The memory capacity determination system 200 during learning of a cell image 80 shown in Figure 1 is a memory capacity determination system during learning that is capable of determining whether the capacity of the memory 10b is insufficient when learning a learning model 21 for performing analytical processing on a cell image 80 captured by a user performing cell culture, etc.

[0014] Overview of the memory capacity determination system The memory capacity determination system 200 for learning a cell image 80 includes an image processing device 100, a computer 120, and an imaging device .

[0015] FIG. 1 shows an example of a memory capacity determination system 200 constructed in a client-server model. The computer 120 functions as a client terminal in the memory capacity determination system 200. The image processing device 100 functions as a server in the memory capacity determination system 200. The image processing device 100, the computer 120, and the imaging device 130 are connected to each other via a network 140 so as to be able to communicate with each other. The image processing device 100 performs various information processing in response to a request (processing request) from the computer 120 operated by a user. The image processing device 100 performs learning of a learning model 21 for analyzing a cell image 80 in response to the request. In this embodiment, the image processing device 100 trains the learning model 21 to classify cells captured in the cell image 80. For example, the image processing device 100 trains the learning model 21 to classify whether a cell captured in the cell image 80 is a normal cell or not, and whether a cell captured in the cell image 80 is aged or not.

[0016] Furthermore, in response to a request, the image processing device 100 determines whether or not the memory 10b will be insufficient when learning the learning model 21 for analyzing the cell image 80. Furthermore, the image processing device 100 transmits notification content 30 according to the determination result of whether or not the memory 10b will be insufficient to the computer 120. Acceptance of operations on the image processing device 100 and display of the analysis results and images after analysis performed by the image processing device 100 are performed on a GUI (Graphical User Interface) displayed on a display unit 121 of the computer 120.

[0017] The network 140 connects the image processing device 100, the computer 120, and the imaging device 130 so that they can communicate with each other. The network 140 may be, for example, a local area network (LAN) constructed within a facility. The network 140 may be, for example, the Internet. When the network 140 is the Internet, the memory capacity determination system 200 during learning of the cell image 80 may be a system constructed in the form of cloud computing.

[0018] The computer 120 is a so-called personal computer, and includes a processor and a storage unit. The display unit 121 and the input unit 122 are connected to the computer 120. The display unit 121 is, for example, a liquid crystal display device. The display unit 121 may be an electroluminescence display device, a projector, or a head-mounted display. The input unit 122 is, for example, an input device including a mouse and a keyboard. The input unit 122 may be a touch panel. One or more computers 120 are provided in the memory capacity determination system 200 during learning of the cell image 80. In this embodiment, the display unit 121 is configured to perform a notification based on the first determination result 32 by the determination unit 12d (see FIG. 3) described later. In addition, the selection unit 45 described later is displayed on the display unit 121. The display unit 121 is an example of the "notification unit" in the claims.

[0019] The imaging device 130 captures an image of a cell to generate a cell image 80. The imaging device 130 can transmit the generated cell image 80 to the computer 120 and / or the image processing device 100 via the network 140. The imaging device 130 captures a microscopic image of the cell. The imaging device 130 performs imaging using an imaging method such as bright-field observation, dark-field observation, phase contrast observation, or differential interference observation. Depending on the imaging method, one or more types of imaging devices 130 are used. The memory capacity determination system 200 during learning of the cell image 80 may be provided with one or more imaging devices 130.

[0020] The image processing device 100 includes a learning processing unit 10, an estimation unit 11, a second processor 12, and a storage unit 13.

[0021] The learning processing unit 10 includes a first processor 10a and a memory 10b.

[0022] The first processor 10a is configured to perform learning of the learning model 21 under the first learning condition 22 using the cell image 80 a preset number of times. The first processor 10a includes, for example, a GPU or an FPGA (Field-Programmable Gate Array) configured for image processing. In this embodiment, an example in which the first processor 10a is a GPU will be described.

[0023] The memory 10b is used as a working area for the learning process of the learning model 21. The memory 10b includes, for example, a RAM (Random Access Memory).

[0024] The estimation unit 11 is configured to perform estimation processing of the cell image 80 using the trained model 25. The estimation unit 11 includes, for example, a GPU or an FPGA configured for image processing.

[0025] The second processor 12 includes a CPU (Central Processing Unit), an FPGA, an ASIC (Application Specific Integrated Circuit), etc. The second processor 12 executes a predetermined program 20, whereby the image processing device 100 performs arithmetic processing.

[0026] The storage unit 13 includes a non-volatile storage device. The non-volatile storage device is, for example, a hard disk drive, a solid state drive, or the like. The storage unit 13 stores various programs 20 executed by the second processor 12. The storage unit 13 also stores an epoch number 23, which is the number of times the learning processing unit 10 repeats the learning process in the learning mode. The storage unit 13 also stores a second learning condition 24, which is a learning condition when the learning process in the verification mode was previously performed, and a second judgment result 33 under the second learning condition 24. The storage unit 13 also stores notification content 30. The storage unit 13 also stores a cell image 80.

[0027] The selection unit 45 is configured to select a learning mode in which the learning model 21 is learned by the learning processing unit 10, and a verification mode in which verification is performed as to whether or not the capacity of the memory 10b is insufficient when the learning processing unit 10 performs the learning processing. The selection unit 45 is a push button on a GUI displayed on the display unit 121. Details of the selection unit 45 will be described later.

[0028] The image processing device 100 is configured to perform learning of the learning model 21 using the cell image 80 stored in the storage unit 13 in response to a request from the computer 120. Specifically, the learning processing unit 10 included in the image processing device 100 performs learning of the learning model 21.

[0029] The image processing device 100 performs a learning process in a learning mode in response to a request from a user, thereby learning the learning model 21. Specifically, in the learning mode, the learning processing unit 10 repeatedly performs the learning process under the first learning condition 22 a preset number of times (the number of epochs 23). The learning process in the learning mode is performed by the learning processing unit 10.

[0030] 2, when the learning processing unit 10 performs the learning process of the learning model 21, the memory 10b is used as a working area 50. The memory 10b includes an available capacity 51 and an unavailable capacity 52 that is used by other programs and cannot be used for the learning process. If the capacity secured as the working area 50 when the learning processing unit 10 performs the learning of the learning model 21 is larger than the available capacity 51, a capacity shortage (out of memory) of the memory 10b occurs, and the learning of the learning model 21 ends abnormally.

[0031] Therefore, in this embodiment, the image processing device 100 performs a learning process in a verification mode to determine whether or not the capacity of the memory 10b is insufficient when learning the learning model 21. In the verification mode, the learning process is performed a number of times less than the number of times of the learning process in the learning mode. Specifically, in the verification mode, one learning process is performed. In addition, the image processing device 100 acquires the notification content 30 (see FIG. 1) based on the determination result of whether or not the capacity of the memory 10b is insufficient. In addition, the image processing device 100 transmits the acquired notification content 30 to the computer 120 (see FIG. 1). The computer 120 that receives the notification content 30 displays the notification content 30 on the display unit 121 (see FIG. 1). Note that the learning process in the learning model 21 and the learning process in the verification mode are not performed in parallel. That is, the image processing device 100 executes either the learning process in the learning mode or the learning process in the verification mode.

[0032] <Determination process by the second processor> 3, the second processor 12 includes a learning condition registration unit 12a, a learning condition change unit 12b, an image size increase unit 12c, a judgment unit 12d, and a notification content acquisition unit 12e as functional blocks. In other words, the second processor 12 executes the program 20 stored in the storage unit 13 to function as the learning condition registration unit 12a, the learning condition change unit 12b, the image size increase unit 12c, the judgment unit 12d, and the notification content acquisition unit 12e.

[0033] The learning condition registration unit 12a is configured to register the first learning condition 22. Note that registering the first learning condition 22 means that the learning condition registration unit 12a acquires the first learning condition 22 input by the operator via the input unit 122 and stores the acquired first learning condition 22 in the memory unit 13. In addition, the learning condition registration unit 12a outputs the first learning condition 22 to the learning processing unit 10. In addition, the learning condition registration unit 12a outputs the first learning condition 22 to the determination unit 12d.

[0034] The first learning condition 22 includes a type 22a of the learning model 21 (see FIG. 1), a batch size 22c (see FIG. 1) of the cell images 80 used for learning the learning model 21, a size 22b (see FIG. 1) of the cell images 80, and a number 22d (see FIG. 1) of the discrimination classes to be learned by the learning model 21. The batch size 22c means the number of cell images 80 used for learning the learning model 21. The type 22a of the learning model 21 means the classification of the algorithm of the learning model 21. The number 22d of discrimination classes is the number of classes to be classified in the cell images 80. The batch size 22c of the cell images 80 is an example of the "number of cell images" in the claims.

[0035] The learning condition change unit 12b is configured to change the content of the first learning condition 22. In this embodiment, the learning condition change unit 12b is configured to change the batch size 22c of the cell image 80 among the first learning conditions 22. The learning condition change unit 12b outputs the changed first learning condition 222 to the learning processing unit 10. The learning condition change unit 12b also stores the changed first learning condition 222 in the storage unit 13. The learning condition change unit 12b also outputs the changed first learning condition 222 to the judgment unit 12d. The change of the content of the first learning condition 22 by the learning condition change unit 12b is executed when optimizing the first learning condition 22, which will be described later. That is, when performing the first learning process in the verification mode, the change of the content of the first learning condition 22 by the learning condition change unit 12b is not performed.

[0036] In this embodiment, when performing the learning process in the verification mode, in order to increase the load on the memory 10b and to increase the capacity secured as the working area 50 (see FIG. 2), the image size increasing unit 12c is configured to increase the size 22b of the cell image 80. For example, the image size increasing unit 12c is configured to increase the size of the cell image 80 by 1.1 to 1.2 times. The image size increasing unit 12c outputs the cell image 80a whose size has been increased to the learning processing unit 10. Note that the process of increasing the size of the cell image 80 by the image size increasing unit 12c is performed only in the verification mode. That is, the process of increasing the size of the cell image 80 by the image size increasing unit 12c is not performed in the learning mode.

[0037] The learning processing unit 10 performs a learning process in the verification mode using the first learning condition 22 input from the learning condition registration unit 12a, the cell image 80a whose size has been increased input from the image size increase unit 12c, and the learning model 21. That is, the learning processing unit 10 is configured to perform a learning process in the verification mode using the cell image 80a whose size has been increased by the image size increase unit 12c. In this embodiment, the learning processing unit 10 is configured to perform a learning process in the verification mode fewer times than the number of learning processes in the learning mode. Specifically, the learning processing unit 10 is configured to perform a single learning process in the verification mode.

[0038] Furthermore, the learning processing unit 10 outputs information 31 on the end of the learning process when the learning process in the verification mode is performed to the determination unit 12d. The information 31 on the end of the learning process is status information when the learning process in the verification mode is completed. Specifically, when the learning process in the verification mode is normally completed, the learning processing unit 10 outputs status information on the normal end of the learning process in the verification mode to the determination unit 12d as information 31 on the end of the learning process. When the learning process in the verification mode is abnormally completed, the learning processing unit 10 outputs status information on the abnormal end of the learning process in the verification mode to the determination unit 12d as information 31 on the end of the learning process. When the learning process in the verification mode is abnormally completed, the learning processing unit 10 outputs log information on the abnormal end to the determination unit 12d.

[0039] The determination unit 12d is configured to determine whether or not the capacity of the memory 10b is insufficient during the learning process of the learning model 21 in the verification mode. Specifically, the determination unit 12d determines whether or not the capacity of the memory 10b is insufficient based on the information 31 on the completion of the learning process input from the learning processing unit 10. That is, when information on the status that the learning process in the verification mode has been normally completed is input as the information 31 on the completion of the learning process, the determination unit 12d determines that the capacity of the memory 10b is not insufficient.

[0040] When status information indicating that the learning process in the verification mode has ended abnormally is input as the information 31 indicating the end of the learning process, the determination unit 12d acquires whether or not the capacity of the memory 10b is insufficient based on the log information of the abnormal end. Specifically, the determination unit 12d determines whether or not the capacity of the memory 10b is insufficient depending on whether or not the log information of the abnormal end includes information indicating that the capacity of the memory 10b is insufficient. The determination unit 12d outputs a first determination result 32, which is a determination result of whether or not the capacity of the memory 10b is insufficient, to the notification content acquisition unit 12e. Note that the first determination result 32 includes status information indicating that the learning process in the verification mode has ended normally without an out-of-memory condition and status information indicating that the learning process in the verification mode has ended abnormally with an out-of-memory condition. The determination unit 12d determines whether or not the capacity of the memory 10b is insufficient only in the verification mode, and not in the learning mode.

[0041] The notification content acquisition unit 12e is configured to acquire the notification content 30 stored in the storage unit 13 based on the first judgment result 32. The notification content 30 includes a message 30a (see FIG. 5) indicating that the out-of-memory verification is successful, and a message 30b (see FIG. 7) indicating that the out-of-memory verification is unsuccessful. When out-of-memory has not occurred, the notification content acquisition unit 12e acquires the message 30a indicating that the out-of-memory verification is successful as the notification content 30. When out-of-memory has occurred, the notification content acquisition unit 12e acquires the message 30b indicating that the out-of-memory verification is unsuccessful as the notification content 30. The notification content acquisition unit 12e outputs the notification content 30 to the display unit 121. In addition, in the learning mode, the notification content acquisition unit 12e acquires a message indicating whether the learning of the learning model 21 has ended normally or abnormally as the notification content 30. That is, in the learning mode, the notification content acquisition unit 12e does not acquire the notification content 30 regarding whether or not an out-of-memory state has occurred.

[0042] The display unit 121 notifies whether or not the capacity of the memory 10b is insufficient when performing the learning process in the verification mode, based on the input notification content 30. Specifically, the display unit 121 is configured to display the notification content 30 input from the notification content acquisition unit 12e.

[0043] Here, if the learning process in the verification mode has been performed in the past under the second learning condition 24 that is equal to the first learning condition 22, it is considered that the determination result of whether or not the capacity of the memory 10b is insufficient in the learning process in the verification mode will be the same result. Therefore, in this embodiment, the determination unit 12d is configured to determine whether or not the second learning condition 24 that is equal to the first learning condition 22 is stored in the storage unit 13. Furthermore, the learning processing unit 10 is configured not to execute the learning process of the first learning condition 22 in the verification mode when the second learning condition 24 that is equal to the first learning condition 22 is stored in the storage unit 13. Note that the second learning condition 24 is the first learning condition 22 when the verification was performed in the past. That is, the second learning condition 24 includes the type of the learning model 21, the type of the cell image 80, the batch size, and the number of the discrimination classes. Moreover, the first learning condition 22 and the second learning condition 24 being equal means that the type of the learning model 21, the type of the cell image 80, the batch size, and the number of the discrimination classes are all equal.

[0044] When the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13, the notification content acquisition unit 12e acquires a second judgment result 33 based on the second learning condition 24. Furthermore, the notification content acquisition unit 12e acquires notification content 30 based on the second judgment result 33, and outputs the acquired notification content 30 to the display unit 121. Note that the second judgment result 33 includes the same content as the first judgment result 32. That is, when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13, the notification content acquisition unit 12e acquires the notification content 30 based on the second judgment result 33 by a configuration similar to the configuration for acquiring the notification content 30 based on the first judgment result 32.

[0045] The display unit 121 is configured to notify a second judgment result 33 based on the second learning condition 24 when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13. Specifically, the display unit 121 displays the notification content 30 acquired by the notification content acquisition unit 12e based on the second judgment result 33 and input to the display unit 121.

[0046] In addition, in this embodiment, when the learning processing unit 10 performs the learning processing in the verification mode, the estimation unit 11 (see FIG. 1) is configured to execute, in parallel with the learning processing by the learning processing unit 10, an estimation process of the cell image 80 by the already learned trained model 25 (see FIG. 1) using the memory 10b. In this embodiment, the estimation process is also performed in the working area 50 used for the learning process in the verification mode. That is, the capacity of the memory 10b secured as the working area 50 includes the area of ​​the memory 10b for performing the learning process in the verification mode and the area of ​​the memory 10b for performing the estimation process. In this embodiment, the memory 10b with the larger capacity secured is selected as the trained model 25 to be used for the estimation process. The estimation process is, for example, a process of classifying the cells shown in the cell image 80 and estimating which classification each pixel of the cell image 80 belongs to.

[0047] <First, optimization of learning conditions> Here, the capacity of the working area 50 (see FIG. 2) used when performing the learning process in the verification mode varies depending on the first learning condition 22. Therefore, depending on the first learning condition 22, there are cases where not all of the available capacity 51 of the memory 10b is reserved as the working area 50. In other words, even if the first learning condition 22 is changed to one that increases the capacity of the memory 10b reserved as the working area 50 and the learning process in the verification mode is performed, there are cases where there is still room in the available capacity 51 of the memory 10b and an out-of-memory situation does not occur.

[0048] Furthermore, if the memory 10b becomes out of memory when the learning process is performed in the verification mode, the user changes the first learning conditions 22 to adjust the first learning conditions 22 so that the first learning conditions 22 do not become out of memory. However, if the user manually changes the first learning conditions 22 and repeatedly performs the learning process in the verification mode to obtain optimal first learning conditions 22, the burden on the user increases.

[0049] Therefore, in this embodiment, the learning processing unit 10 is configured to repeatedly execute the learning process in the verification mode while changing the first learning condition 22 by the learning condition changing unit 12b. That is, in this embodiment, the first learning condition 22 can be optimized by the learning processing unit 10 and the learning condition changing unit 12b. In this embodiment, the learning processing unit 10 is configured to repeatedly execute the learning process in the verification mode while changing the batch size 22c by the learning condition changing unit 12b, for example. Note that the optimization of the first learning condition 22 is configured to be performed by an operational input by the user.

[0050] If out-of-memory does not occur during the learning process in the verification mode, the learning processing unit 10 is configured to repeatedly execute the learning process in the verification mode while increasing the batch size 22c. Also, if out-of-memory occurs during the learning process in the verification mode, the learning processing unit 10 is configured to repeatedly execute the learning process in the verification mode while decreasing the batch size 22c.

[0051] Moreover, the judgment unit 12d is configured to acquire the optimal first learning condition 22 under which the capacity of the memory 10b does not run short, based on the first judgment result 32. The judgment unit 12d is configured to acquire, as the optimal first learning condition 22, the maximum batch size 22c of the cell images 80 under which the capacity of the memory 10b does not run short.

[0052] <Screen display of learning process in verification mode> Next, a configuration for performing a learning process in the verification mode by the image processing device 100 of this embodiment will be described with reference to FIGS.

[0053] First, with reference to FIG. 4 and FIG. 5, a screen example when the learning process in the verification mode is successful will be described.

[0054] 4 shows an example of a learning condition setting screen 111a displayed on the display unit 121 (see FIG. 1). The learning condition setting screen 111a displays a learning name input field 40, a model selection field 41, a dataset selection field 42, an epoch number input field 43, a batch size input field 44, a verify button 45a, a start learning button 45b, and a cancel button 46.

[0055] The learning name input field 40 is an input field for inputting a learning name. The model selection field 41 is a selection field for selecting the learning model 21 (see FIG. 1) to be learned. The dataset selection field 42 is a selection field for selecting a dataset to be used for learning the learning model 21. The epoch number input field 43 is an input field for inputting the epoch number 23. The batch size input field 44 is an input field for inputting the batch size 22c. The verification button 45a is a push button on the GUI for starting the learning process in the verification mode. The learning start button 45b is a push button on the GUI for starting the learning process in the learning mode. The cancel button 46 is a push button on the GUI for canceling the learning process.

[0056] A learning model 21 is selected in the model selection field 41, whereby a type 22a (see FIG. 1) of the learning model 21 in the first learning condition 22 (see FIG. 1) is set. A dataset is selected in the dataset selection field 42, whereby a size 22b (see FIG. 1) of the cell image 80 (see FIG. 1) and a number 22d (see FIG. 1) of discrimination classes in the first learning condition 22 are set. A batch size value input in the batch size input field 44 sets a batch size 22c (see FIG. 1) in the first learning condition 22. The number of epochs 23 (see FIG. 1) input in the number of epochs input field 43 is set as the number of epochs during learning in the learning mode. That is, the number of epochs 23 input in the number of epochs input field 43 is not used in the learning process in the verification mode.

[0057] When the verification button 45a is pressed, a learning process in the verification mode is executed according to the first learning condition 22 that has been set.

[0058] In the learning process in the verification mode, when the learning process is normally completed without occurrence of out-of-memory, the display unit 121 displays the verification result notification screen 111b shown in Fig. 5. Specifically, the display unit 121 displays the notification content 30 transmitted from the image processing device 100 as the verification result notification screen 111b. A message 30a indicating that the out-of-memory verification was successful is displayed on the verification result notification screen 111b. Note that, when displaying the verification result notification screen 111b, for example, the display unit 121 displays it as a pop-up screen on the learning condition setting screen 111a shown in Fig. 4.

[0059] Next, with reference to FIG. 6 and FIG. 7, a description will be given of an example of a screen that is displayed when the learning process in the verification mode fails.

[0060] A learning condition setting screen 111c shown in Fig. 6 is the same as the learning condition setting screen 111a shown in Fig. 4, except that the batch size 22c input in the batch size input field 44 is different. In the example shown in Fig. 6, "11" is input as an example of the batch size 22c when an out-of-memory condition occurs in the learning process in the verification mode.

[0061] When the verification button 45a is pressed on the learning condition setting screen 111c, the display unit 121 displays a verification result notification screen 111d as shown in FIG. 7. Specifically, the display unit 121 displays the notification content 30 transmitted from the image processing device 100 as the verification result notification screen 111d. The verification result notification screen 111d displays a message 30b indicating that the out-of-memory verification has failed. The message 30b indicating that the out-of-memory verification has failed includes information 30c urging the user to reduce the batch size 22c and information 30d urging the user to reduce the size 22b of the cell image 80. That is, in this embodiment, the display unit 121 is configured to notify the user of information 30c urging the user to reduce the batch size 22c of the cell image 80 under the first learning condition 22 when the capacity of the memory 10b is insufficient. Moreover, the display unit 121 is configured to further notify information 30d that prompts the user to reduce the size 22b of the cell image 80 under the first learning condition 22 when the capacity of the memory 10b is insufficient.

[0062] Next, a memory capacity verification process in the memory capacity determination method during learning of a cell image 80 according to this embodiment will be described with reference to FIG.

[0063] In step 101, the learning condition registration unit 12a (see FIG. 3) registers the first learning condition 22. Specifically, the learning condition registration unit 12a stores the first learning condition 22 set on the learning condition setting screen 111a (see FIG. 4) in the memory unit 13 (see FIG. 1).

[0064] In step 102, a learning mode is selected. Specifically, the second processor 12 (see FIG. 1) selects between a learning mode in which the learning model 21 (see FIG. 1) is learned by performing a learning process on the learning model 21 based on an input from the selection unit 45 (see FIG. 4) and a verification mode in which verification is performed as to whether or not the capacity of the memory 10b (see FIG. 1) used as a working area when performing the learning process on the learning model 21 is insufficient. More specifically, the second processor 12 selects the verification mode when the verification button 45a (see FIG. 4) is pressed. Also, the second processor 12 selects the learning mode when the learning start button 45b (see FIG. 4) is pressed.

[0065] In step 103, the second processor 12 (see FIG. 1) determines whether or not the verification mode has been selected. If the verification mode has not been selected, the process proceeds to step 104. If the verification mode has been selected, the process proceeds to step 105.

[0066] In step 104, the learning processing unit 10 (see FIG. 1) performs learning processing in the learning mode. Specifically, the learning processing unit 10 performs learning of the learning model 21 under the first learning condition 22 using the cell image 80 a preset number of times (the number of epochs 23). Thereafter, the processing ends. Note that the preset number of times (the number of epochs 23) is a value input in the number of epochs input field 43 on the learning condition setting screen 111a (see FIG. 4).

[0067] When the process proceeds from step 103 to step 105, in step 105, the judgment unit 12d (see FIG. 3) judges whether or not verification has been performed under the same learning conditions in the past. Specifically, the judgment unit 12d judges whether or not the second learning conditions 24 (see FIG. 1) equal to the first learning conditions 22 (see FIG. 1) are stored in the storage unit 13 (see FIG. 1). If the second learning conditions 24 equal to the first learning conditions 22 are stored in the storage unit 13, the process proceeds to step 106. If the second learning conditions 24 equal to the first learning conditions 22 are not stored in the storage unit 13, the process proceeds to step 107.

[0068] In step 106, the display unit 121 displays the second judgment result 33 (see FIG. 1) based on the second learning condition 24. Then, the process ends.

[0069] When the process proceeds from step 105 to step 107, in step 107, the learning processing unit 10 executes a learning process in the verification mode under the first learning condition 22. In this embodiment, in parallel with the learning process in the verification mode in step 107, an estimation process of the cell image 80 is executed by the estimation unit 11. In this embodiment, the learning process in the verification mode is executed once, regardless of the value input in the epoch number input field 43 on the learning condition setting screen 111a (see FIG. 4).

[0070] In step 108, the judgment unit 12d judges whether or not the capacity of the memory 10b (see FIG. 1) is insufficient during the learning process of the learning model 21 in the verification mode. Specifically, the judgment unit 12d judges whether or not the capacity of the memory 10b is insufficient based on the learning process end information 31 (see FIG. 3). If the capacity of the memory 10b is insufficient, the process proceeds to step 109. If the capacity of the memory 10b is not insufficient, the process proceeds to step 110.

[0071] In step 109, the display unit 121 notifies the first judgment result 32 as to whether or not the capacity of the memory 10b is insufficient. In step 109, the display unit 121 notifies that the capacity of the memory 10b is insufficient. Specifically, the display unit 121 notifies that the capacity of the memory 10b is insufficient by displaying a verification result notification screen 111d (see FIG. 7). Thereafter, the process ends.

[0072] When the process proceeds from step 108 to step 110, in step 110, the display unit 121 notifies that the capacity of the memory 10b is sufficient. Specifically, the display unit 121 notifies that the capacity of the memory 10b is sufficient by displaying a verification result notification screen 111b (see FIG. 5). Then, the process ends.

[0073] Next, a process in which the image processing device 100 according to the present embodiment optimizes the first learning condition 22 (see FIG. 1) will be described with reference to Fig. 9. Note that the optimization process of the first learning condition 22 shown in Fig. 9 is performed after the learning process in the verification mode is executed once.

[0074] In step 201, the judgment unit 12d (see FIG. 3) judges whether or not out-of-memory of the memory 10b has occurred in the learning process in the verification mode. If out-of-memory has occurred, the process proceeds to step 202. If out-of-memory has not occurred, the process proceeds to step 206.

[0075] In step 202, the learning condition change unit 12b (see FIG. 3) decreases the batch size 22c by one.

[0076] In step 203, the learning processing unit 10 executes the learning process in the verification mode under the first learning condition 222 (see FIG. 3) after the batch size 22c has been reduced by one.

[0077] In step 204, the determination unit 12d determines whether or not out-of-memory of the memory 10b has occurred in the learning process in the verification mode under the first learning condition 222 after the batch size 22c has been reduced by one in step 203. If out-of-memory of the memory 10b has occurred, the process proceeds to step 202. If out-of-memory of the memory 10b has not occurred, the process proceeds to step 205.

[0078] In step 205, the learning condition registration unit 12a (see FIG. 3) stores the first learning condition 222 in the storage unit 13 (see FIG. 1). Then, the process ends. That is, the process of steps 202 to 205 is a process of acquiring an optimal batch size 22c by repeatedly performing the learning process in the verification mode while reducing the batch size 22c.

[0079] Furthermore, when the process proceeds from step 201 to step 206, in step 206, the learning condition change unit 12b (see FIG. 3) increases the batch size 22c by one.

[0080] In step 207, the learning processing unit 10 executes the learning process in the verification mode under the first learning condition 222 (see FIG. 3) after the batch size 22c has been increased by one.

[0081] In step 208, the determination unit 12d determines whether or not out-of-memory of the memory 10b has occurred in the learning process in the verification mode under the first learning condition 222 after the batch size 22c has been increased by one in step 207. If out-of-memory of the memory 10b has occurred, the process proceeds to step 209. If out-of-memory of the memory 10b has not occurred, the process proceeds to step 206.

[0082] In step 209, the learning condition change unit 12b (see FIG. 3) decreases the batch size 22c by one.

[0083] In step 210, the learning condition registration unit 12a (see FIG. 3) stores the first learning condition 222 obtained by decreasing the batch size 22c by one in step 209 in the storage unit 13 (see FIG. 1). Then, the process ends. That is, the process of steps 206 to 210 is a process of acquiring an optimal batch size 22c by repeatedly executing the learning process in the verification mode while increasing the batch size 22c.

[0084] (Effects of this embodiment) In this embodiment, the following effects can be obtained.

[0085] In this embodiment, as described above, the memory capacity determination system 200 when learning a cell image 80 includes a processor (first processor 10a) that uses the cell image 80 to learn the learning model 21 under the first learning conditions 22 a preset number of times, a learning processing unit 10 having a memory 10b used as a working area for the learning process of the learning model 21, a selection unit 45 that selects between a learning mode in which the learning model 21 is learned by performing a learning process by the learning processing unit 10 and a verification mode in which it is verified whether or not the capacity of the memory 10b is insufficient when performing the learning process by the learning processing unit 10, a judgment unit 12d that judges whether or not the capacity of the memory 10b is insufficient during the learning process of the learning model 21 in the verification mode, and an alarm unit (display unit 121) that issues an alarm based on the judgment result (first judgment result 32) by the judgment unit 12d.

[0086] Here, the capacity of the memory 10b secured as the working area 50 is equal in both the learning mode and the verification mode. Therefore, by performing verification in the verification mode in which verification is performed to verify whether or not the capacity of the memory 10b used as the working area when performing the learning process of the learning model 21 is insufficient, it is possible to determine whether or not the capacity of the memory 10b is insufficient due to the learning mode. In addition, the determination result (first determination result 32) as to whether or not the capacity of the memory 10b is insufficient is notified. Therefore, by performing verification in the verification mode before performing learning of the learning model 21 in the learning mode, the user can know whether or not the capacity of the memory 10b is insufficient before performing the learning process in the learning mode. As a result, the user can know that the machine learning was not completed due to the insufficient memory capacity. In addition, since it is possible to select the learning mode and the verification mode, when it is not necessary to perform verification in the verification mode, the learning model 21 can be trained in the learning mode without performing the learning process in the verification mode. As a result, it is possible to improve the convenience of the user.

[0087] Furthermore, in this embodiment, as described above, the method for determining memory capacity when learning a cell image 80 includes the steps of: learning the learning model 21 using the cell image 80 a predetermined number of times under the first learning conditions 22; selecting a learning mode in which the learning model 21 is learned by performing a learning process on the learning model 21; and a verification mode in which verification is performed as to whether or not the capacity of the memory 10b used as a working area when performing the learning process on the learning model 21 is insufficient; determining whether or not the capacity of the memory 10b has become insufficient in the learning process of the learning model 21 in the verification mode; and notifying the determination result of whether or not the capacity of the memory 10b has become insufficient (first determination result 32).

[0088] This makes it possible to provide a memory capacity determination system 200 for determining memory capacity during learning of a cell image 80, which allows the user to understand that machine learning has not been completed due to insufficient memory capacity, similar to the memory capacity determination system 200 for determining memory capacity during learning of a cell image 80.

[0089] In addition, in the above embodiment, the following additional effects can be obtained by configuring as follows.

[0090] That is, in this embodiment, as described above, the learning processing unit 10 is configured to execute the learning process in the verification mode a number of times less than the number of times of the learning process in the learning mode. This makes it possible to shorten the time required for the learning process in the verification mode compared to the time required for the learning process in the learning mode. As a result, it is possible to shorten the time required for determining whether the capacity of the memory 10b is insufficient.

[0091] In the present embodiment, as described above, the learning processing unit 10 is configured to execute one learning process in the verification mode. This can further reduce the time required for the learning process in the verification mode. As a result, the time required for determining whether the capacity of the memory 10b is insufficient can be further reduced.

[0092] In addition, in this embodiment, as described above, the image size increasing unit 12c that increases the size 22b of the cell image 80 is further provided, and the learning processing unit 10 is configured to perform the learning processing in the verification mode using the cell image 80a whose size has been increased by the image size increasing unit 12c. This makes it possible to increase the capacity of the memory 10b secured as the working area 50 when performing the learning processing in the verification mode. Therefore, since it is possible to perform the learning processing in the verification mode with an extra capacity of the memory 10b secured as the working area 50, it is possible to make the determination condition for whether or not the capacity of the memory 10b is insufficient stricter. As a result, it is possible to perform the learning processing of the learning model 21 in the verification mode under stricter conditions, and it is possible to easily prevent the capacity of the memory 10b from being insufficient in the learning mode.

[0093] In the present embodiment, as described above, the vehicle further includes a storage unit 13 that stores the second learning condition 24, which is a learning condition when the learning process in the verification mode was previously performed, and a judgment result (second judgment result 33) under the second learning condition 24, and the judgment unit 12d is configured to judge whether or not the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13, the learning processing unit 10 is configured to not execute the learning process in the verification mode when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13, and the notification unit (display unit 121) is configured to notify the judgment result (second judgment result 33) under the second learning condition 24 when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13. As a result, when verification has already been performed under the second learning condition 24 equal to the first learning condition 22, notification is performed under the second judgment result 33 without performing the learning process in the verification mode. Therefore, it is possible to notify the user whether or not the capacity of the memory 10b is insufficient without performing the learning process in the verification mode. As a result, it is possible to prevent the learning process in the verification mode from being executed redundantly due to the first learning condition 22.

[0094] In addition, in this embodiment, as described above, the notification unit (display unit 121) is configured to notify the user of information 30c that prompts the user to reduce the number of images (batch size 22c) of the cell images 80 in the first learning condition 22 when the capacity of the memory 10b is insufficient. This allows the user to understand that the number of images (batch size 22c) of the cell images 80 in the first learning condition 22 should be reduced to avoid the capacity of the memory 10b being insufficient. As a result, it becomes possible to understand the items of the first learning condition 22 that are used to avoid the capacity of the memory 10b being insufficient, and therefore even if the user has a low level of proficiency, the first learning condition 22 can be easily adjusted.

[0095] Furthermore, in this embodiment, as described above, the notification unit (display unit 121) is configured to further notify the user of information 30d that prompts the user to reduce the size 22b of the cell image 80 in the first learning condition 22 when the capacity of the memory 10b is insufficient. This allows the user to understand that the size 22b of the cell image 80 should be reduced in addition to reducing the number of images (batch size 22c) of the cell image 80. As a result, even if the user has a low level of proficiency, the first learning condition 22 can be more easily adjusted to avoid the capacity of the memory 10b being insufficient.

[0096] In addition, in this embodiment, as described above, when the learning processing unit 10 performs the learning processing in the verification mode, the estimation unit 11 is further provided, which performs the estimation processing of the cell image 80 by the already trained trained model 25 using the memory 10b in parallel with the learning processing by the learning processing unit 10. As a result, the estimation processing of the cell image 80 by the trained model 25 is performed in parallel with the learning processing of the learning model 21, so that the learning processing in the verification mode can be performed with an increased load on the memory 10b compared to the case where only the learning processing of the learning model 21 is performed. As a result, even if the estimation processing by the trained model 25 is performed in parallel while the learning processing of the learning model 21 is being performed in the learning mode, it is possible to prevent the capacity of the memory 10b from running short in the learning mode.

[0097] In addition, as described above, the present embodiment further includes a learning condition change unit 12b that changes the contents of the first learning conditions 22, and the learning processing unit 10 is configured to repeatedly execute the learning process in the verification mode while changing the first learning conditions 22 by the learning condition change unit 12b, and the judgment unit 12d is configured to acquire the optimal first learning conditions 22 that do not cause a shortage of capacity in the memory 10b based on the judgment result (first judgment result 32). As a result, the optimal first learning conditions 22 that do not cause a shortage of capacity in the memory 10b are acquired, and therefore the learning process of the learning model 21 in the learning mode can be performed with the optimal first learning conditions 22 without depending on the proficiency of the user.

[0098] In the present embodiment, as described above, the learning condition change unit 12b is configured to change the number of cell images 80 (batch size 22c) among the first learning conditions 22, and the determination unit 12d is configured to acquire the maximum number of cell images 80 (batch size 22c) that does not cause a capacity shortage in the memory 10b as the optimal first learning condition 22. This allows the learning process of the learning model 21 to be performed using the maximum number of cell images 80 (batch size 22c) that does not cause a capacity shortage in the memory 10b. As a result, the learning accuracy of the learning model 21 can be improved while preventing the capacity of the memory 10b from running short.

[0099] In addition, as described above, the present embodiment further includes a learning condition registration unit 12a that registers the first learning condition 22 including the type 22a of the learning model 21, the number of images (batch size 22c) of the cell images 80 used for learning the learning model 21, the size 22b of the cell images 80, and the number 22d of the discrimination classes to be trained by the learning model 21. This makes it possible to grasp whether or not the capacity of the memory 10b is insufficient when training the learning model 21 based on the registered type 22a of the learning model 21, the number of images (batch size 22c) of the cell images 80 used for training the learning model 21, the size 22b of the cell images 80, and the number 22d of the discrimination classes to be trained by the learning model 21.

[0100] [Variations] The embodiments disclosed herein should be considered to be illustrative and not restrictive in all respects. The scope of the present invention is indicated by the claims, not by the description of the embodiments above, and further includes all modifications (variations) within the meaning and scope of the claims.

[0101] For example, in the above embodiment, the image processing device 100 functions as a server of the memory capacity determination system 200 during learning of the cell image 80 constructed in a client-server model, but the present invention is not limited to this. In the present invention, for example, as shown in FIG. 10, the image processing device 100 may be configured by an independent computer. In the example of FIG. 10, the image processing device 100 is configured by a computer 300 including a second processor 310 and a storage unit 320. A display unit 330 and an input unit 340 are connected to the computer 300. The computer 300 is connected to the imaging device 130 so as to be able to communicate with the image capturing device 130. The second processor 310 of the computer 300 includes, as functional blocks, the learning condition registration unit 12a, the learning condition change unit 12b, the image size increase unit 12c, the determination unit 12d, and the notification content acquisition unit 12e shown in the above embodiment (see FIG. 3).

[0102] In the above embodiment and the modified example shown in FIG. 10, an example was shown in which a single second processor 12 (310) executes each process as the learning condition registration unit 12a, the learning condition change unit 12b, the image size increase unit 12c, the judgment unit 12d, and the notification content acquisition unit 12e, but the present invention is not limited to this. Each process of determining whether or not the capacity of the memory 10b is insufficient during learning of the cell image 80 may be shared and executed by multiple processors. Each process may be executed by a separate processor. The multiple processors may be provided in separate computers. In other words, the image processing device 100 may be configured by multiple computers.

[0103] In the above embodiment, the learning processing unit 10 executes the learning process less times in the verification mode than in the learning mode, but the present invention is not limited to this. For example, the learning processing unit 10 may be configured to execute the learning process the same number of times as the learning process in the learning mode or more times in the verification mode. However, when the learning processing unit 10 executes the learning process the same number of times as the learning process in the learning mode or more times in the verification mode, the time required to verify whether the capacity of the memory 10b is insufficient increases. Therefore, it is preferable that the learning processing unit 10 executes the learning process less times in the verification mode than in the learning mode.

[0104] In the above embodiment, the learning processing unit 10 executes the learning process in the verification mode using the cell image 80a whose size has been increased by the image size increaser 12c. However, the present invention is not limited to this. For example, the learning processing unit 10 may execute the learning process in the verification mode using the cell image 80, instead of the cell image 80a whose size has been increased. In this case, the second processor 12 may not include the image size increaser 12c.

[0105] In the above embodiment, an example of a configuration in which the learning processing unit 10 does not execute the learning process in the verification mode by the first learning condition 22 when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13 has been shown, but the present invention is not limited to this. For example, the learning processing unit 10 may be configured to execute the learning process in the verification mode by the first learning condition 22 even when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13. However, when the learning processing unit 10 executes the learning process in the verification mode by the first learning condition 22 even when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13, the learning processing unit 10 will execute a learning process in the verification mode that is essentially unnecessary. Therefore, it is preferable that the learning processing unit 10 is configured not to execute the learning process in the verification mode by the first learning condition 22 when the second learning condition 24 equal to the first learning condition 22 is stored in the storage unit 13.

[0106] In the above embodiment, the memory capacity determination system 200 includes an estimation unit 11 that uses the memory 10b to execute an estimation process of the cell image 80 based on the trained model 25 in parallel with the learning process by the learning processing unit 10 when the learning processing unit 10 performs the learning process in the verification mode. However, the present invention is not limited to this. The memory capacity determination system 200 does not need to include the estimation unit 11.

[0107] In the above embodiment, the memory capacity determination system 200 includes the learning condition change unit 12b, and the learning processing unit 10 repeatedly executes the learning process in the verification mode while changing the first learning condition 22 by the learning condition change unit 12b. However, the present invention is not limited to this. For example, the learning processing unit 10 may be configured not to repeatedly execute the learning process in the verification mode while changing the first learning condition 22 by the learning condition change unit 12b. In this case, the memory capacity determination system 200 does not need to include the learning condition change unit 12b.

[0108] In the above embodiment, the determination unit 12d has been configured to obtain the maximum batch size 22c of cell images 80 that does not cause a shortage of capacity of the memory 10b as the optimal first learning condition 22, but the present invention is not limited to this. For example, the determination unit 12d may be configured to obtain the maximum size of cell images 80 that does not cause a shortage of capacity of the memory 10b as the optimal first learning condition 22.

[0109] In the above embodiment, the first processor 10a and the estimation unit 11 are different processors, but the present invention is not limited to this. For example, the first processor 10a and the estimation unit 11 may be configured by the same processor.

[0110] In the above embodiment, the image processing device 100 is configured to learn the learning model 21 and analyze the cell image 80, but the present invention is not limited to this. For example, the image processing device 100 may be configured as a learning device that only learns the learning model 21.

[0111] In the above embodiment, the image processing device 100 has been described as having the learning model 21 learn to classify cells depicted in the cell image 80, but the present invention is not limited to this. For example, the image processing device 100 may be configured to have the learning model learn to improve the resolution of cells depicted in the cell image 80. The content that the image processing device 100 has the learning model learn can be arbitrarily set by the user.

[0112] [Aspects] It will be appreciated by those skilled in the art that the exemplary embodiments described above are examples of the following aspects.

[0113] (Item 1) A learning processing unit including a processor that performs learning of a learning model under a first learning condition using cell images a preset number of times, and a memory that is used as a working area for the learning process of the learning model; a selection unit that selects between a learning mode in which the learning model is learned by performing a learning process by the learning processing unit and a verification mode in which verification is performed as to whether or not the capacity of the memory is insufficient when the learning process is performed by the learning processing unit; a determination unit that determines whether or not the capacity of the memory is insufficient in the learning process of the learning model in the verification mode; A memory capacity determination system for use in learning a cell image, comprising: a notification unit that issues a notification based on a determination result by the determination unit.

[0114] (Item 2) The memory capacity determination system during learning of cell images described in item 1, wherein the learning processing unit is configured to execute the learning process a number of times in the verification mode that is less than the number of learning processes in the learning mode.

[0115] (Item 3) 3. The memory capacity determination system during learning of a cell image according to item 2, wherein the learning processing unit is configured to execute a single learning process in the verification mode.

[0116] (Item 4) An image size increasing unit that increases the size of the cell image, The memory capacity determination system during learning of a cell image described in any one of items 1 to 3, wherein the learning processing unit is configured to perform learning processing in the verification mode using the cell image whose size has been increased by the image size increase unit.

[0117] (Item 5) A storage unit is further provided that stores a second learning condition, which is a learning condition when the learning process in the verification mode was previously performed, and the determination result under the second learning condition, the determination unit is configured to determine whether the second learning condition that is equal to the first learning condition is stored in the storage unit, the learning processing unit is configured not to execute the learning process in the verification mode when the second learning condition equal to the first learning condition is stored in the storage unit, The memory capacity determination system during learning of a cell image described in any one of items 1 to 4, wherein the notification unit is configured to notify the determination result based on the second learning condition when the second learning condition equal to the first learning condition is stored in the memory unit.

[0118] (Item 6) The memory capacity determination system during learning of cell images described in any one of items 1 to 5, wherein the notification unit is configured to notify information encouraging the user to reduce the batch size of the cell images under the first learning condition when the memory capacity is insufficient.

[0119] (Item 7) The memory capacity determination system during learning of cell images described in item 6, wherein the notification unit is configured to further notify information encouraging the user to reduce the size of the cell image under the first learning condition when the memory capacity is insufficient.

[0120] (Item 8) The memory capacity determination system during learning of a cell image described in any one of items 1 to 7, further comprising an estimation unit that, when performing learning processing in the verification mode by the learning processing unit, uses the memory to perform an estimation process of the cell image using an already learned model in parallel with the learning processing by the learning processing unit.

[0121] (Item 9) A learning condition change unit that changes the content of the first learning condition, the learning processing unit is configured to repeatedly execute a learning process in the verification mode while changing the first learning condition by the learning condition changing unit, The memory capacity determination system for learning cell images described in any one of items 1 to 8, wherein the determination unit is configured to obtain the optimal first learning condition under which the memory capacity does not run short based on the determination result.

[0122] (Item 10) the learning condition change unit is configured to change a batch size of the cell images among the first learning conditions, The memory capacity determination system for learning cell images described in item 9, wherein the determination unit is configured to obtain the maximum batch size of the cell images that does not cause a shortage of memory capacity as the optimal first learning condition.

[0123] (Item 11) The memory capacity determination system for use in learning cell images described in any one of items 1 to 10, further comprising a learning condition registration unit that registers the first learning conditions including the type of learning model, the batch size of cell images used to learn the learning model, the size of the cell images, and the number of discrimination classes to be trained in the learning model.

[0124] (Item 12) learning the learning model under the first learning condition a preset number of times using the cell image; A step of selecting a learning mode in which the learning model is learned by performing a learning process of the learning model, and a verification mode in which verification is performed as to whether or not the capacity of the memory used for the working area when performing the learning process of the learning model is insufficient; determining whether or not the capacity of the memory is insufficient in a learning process of the learning model in the verification mode; and notifying the result of the determination as to whether or not the memory capacity is insufficient. [Explanation of symbols]

[0125] 10 Learning processing unit 10a First processor 10b Memory 12a Learning condition registration section 12b Learning condition change section 12c Image size increase section 12d Judgment section 13 Storage section 21 Learning Model 22 First Learning Condition 22a Types of learning models 22b Cell image size 22c Batch size (number of cell images) 22d Number of discrimination classes 23 epochs (pre-set number of times) 24 Second Learning Condition 25 Pre-trained models 32 First judgment result (judgment result) 33 Second judgment result (judgment result) 45 Selection section 45a Validation button 45b Start learning button 50 working areas 80 Cell Images 80a Cell image with increased size 121, 330 Display unit (notification unit) 200 Memory Capacity Judgment System for Cell Images during Learning

Claims

1. A learning processing unit including a processor that performs learning of a learning model under a first learning condition using cell images a preset number of times, and a memory that is used as a working area for the learning process of the learning model; a selection unit that selects between a learning mode in which the learning model is learned by performing a learning process by the learning processing unit and a verification mode in which verification is performed as to whether or not the capacity of the memory is insufficient when the learning process is performed by the learning processing unit; a determination unit that determines whether or not the capacity of the memory is insufficient in the learning process of the learning model in the verification mode; A memory capacity determination system for use in learning a cell image, comprising: a notification unit that issues a notification based on a determination result by the determination unit.

2. The memory capacity determination system for learning cell images according to claim 1 , wherein the learning processing unit is configured to execute the learning process a number of times in the verification mode that is less than the number of times the learning process is executed in the learning mode.

3. The system for determining memory capacity during learning of a cell image according to claim 2 , wherein the learning processing unit is configured to execute a single learning process in the verification mode.

4. An image size increasing unit that increases the size of the cell image, The memory capacity determination system for learning cell images as described in claim 1, wherein the learning processing unit is configured to perform learning processing in the verification mode using the cell image whose size has been increased by the image size increasing unit.

5. a storage unit configured to store a second learning condition, which is a learning condition when a learning process in the verification mode was previously performed, and the determination result under the second learning condition, the determination unit is configured to determine whether the second learning condition that is equal to the first learning condition is stored in the storage unit, the learning processing unit is configured not to execute the learning process in the verification mode when the second learning condition equal to the first learning condition is stored in the storage unit, The memory capacity determination system for determining a memory capacity during learning of a cell image as described in claim 1, wherein the notification unit is configured to notify the determination result based on the second learning condition when the second learning condition equal to the first learning condition is stored in the memory unit.

6. The memory capacity determination system for learning cell images as described in claim 1, wherein the notification unit is configured to notify information encouraging the reduction of the number of cell images under the first learning condition when the memory capacity is insufficient.

7. The memory capacity determination system for learning cell images as described in claim 6, wherein the notification unit is configured to further notify information encouraging the user to reduce the size of the cell image under the first learning condition when the memory capacity becomes insufficient.

8. The memory capacity determination system for learning cell images as described in claim 1, further comprising an estimation unit that, when performing learning processing in the verification mode by the learning processing unit, uses the memory to perform an estimation process of the cell image using an already learned model in parallel with the learning processing by the learning processing unit.

9. A learning condition change unit that changes the content of the first learning condition, the learning processing unit is configured to repeatedly execute a learning process in the verification mode while changing the first learning condition by the learning condition changing unit, The memory capacity determination system for learning cell images as described in claim 1 , wherein the determination unit is configured to obtain the optimal first learning conditions that do not cause a shortage of memory capacity based on the determination result.

10. the learning condition change unit is configured to change the number of the cell images in the first learning condition, The memory capacity determination system for learning cell images as described in claim 9, wherein the determination unit is configured to obtain the maximum number of cell images that does not cause a shortage of memory capacity as the optimal first learning condition.

11. The memory capacity determination system for learning cell images as described in claim 1, further comprising a learning condition registration unit that registers the first learning conditions including the type of learning model, the number of cell images used to learn the learning model, the size of the cell images, and the number of discrimination classes to be learned by the learning model.

12. learning the learning model under the first learning condition a preset number of times using the cell image; A step of selecting a learning mode in which the learning model is learned by performing a learning process of the learning model, and a verification mode in which verification is performed as to whether or not the capacity of the memory used for the working area when performing the learning process of the learning model is insufficient; determining whether or not the capacity of the memory is insufficient in a learning process of the learning model in the verification mode; and notifying the result of the determination as to whether or not the memory capacity is insufficient.

Images