JP2020144553A - Storage device and data processing method in the same - Google Patents

Abstract

To reduce consumption of storage capacity of a storage device, reduce the storage capacity, and improve capacity efficiency.SOLUTION: A storage device 104 connected to a server via a network includes: a storage unit for storing original data acquired from an outside; a generation unit for generating processing data by processing original data selected from the original data stored in the storage unit in response to a processing data request from the server; and an evaluation unit for evaluating processing performance when the generation unit generates the processing data prior to the generation of processing data by the generation unit. The generation unit generates processing data based on the evaluated processing performance and sends the generated processing data to the server.SELECTED DRAWING: Figure 7

Description

The present invention relates to a storage device and a data processing method in the storage device, for example, an image system AI (Artificial Intelligence) that uses image data acquired by an IoT (Internet of Things) sensor device such as a surveillance camera and stored in the storage device for machine learning. It is suitable for application to the information processing system of.

Patent Document 1 describes a technique in which data is managed by either a file server at one base or a file server at another base, and data is appropriately referred to based on which file server the data is managed by. Is disclosed. Specifically, the file server stores the data of the file received from the user terminal in the storage device and replicates it to the remote file server, and stubs the data of the file stored in the storage device. When the file related to the access request received from the user terminal is not stubbed, the file server reads the data of the file from the storage device and sends it to the user terminal. On the other hand, when the file related to the access request received from the user terminal is not stubbed, the file server reads the data of the file from the storage device and sends it to the user terminal.

Japanese Patent No. 6231623

With the development of IoT and AI in recent years, there is an information processing system in which a server machine-learns image data acquired by an IoT device such as a camera and stored in a storage device. In such an information processing system, in order to improve the efficiency and accuracy of machine learning, a data augment that adds a processed image (an image that has been processed such as position shift or rotation to the original image) based on the image data. Mentioning is done. In this information processing system, at the time of data augmentation, all the image data to be processed is read from the storage device to the server, the data augmentation is performed by the server, and the data of all the processed images is stored in the storage device. Will be written out.

Therefore, in the above-mentioned conventional technique, in an information processing system in which image data acquired by an IoT device and stored in a storage device is machine-learned by a server, all processed image data is once written to the storage device to store the storage capacity of the storage device. There is a problem that is squeezed.

Furthermore, in an information processing system that performs machine learning, when handling teacher data that exceeds the main storage capacity installed in the system, machine learning data is stored in a secondary storage area consisting of an HDD (Hard Disk Drive) or the like. A system that performs machine learning by partially reading into main memory is used. In such a system, it is known that access to an HDD or the like is made more efficient by bundling a plurality of learning data into large-sized intermediate data. However, since these intermediate data are also once written to the storage device, the problem that the storage capacity of the storage device is compressed becomes greater as the scale of the training data increases in combination with the above-mentioned data augmentation. ..

The present invention has been made in consideration of the above points, and in an information processing system in which data acquired by an IoT device and stored in a storage device is machine-learned by a server, consumption of the storage capacity of the storage device is reduced and storage is performed. We are trying to propose a storage device and a data processing method in the storage device that can reduce the capacity and improve the capacity efficiency.

In order to solve such a problem, in the present invention, the storage device is connected to the server via a network, has a storage unit that stores the original data acquired from the outside, and responds to a request for processed data from the server. A generation unit that processes the original data selected from the original data stored in the storage unit to generate processing data, and the generation unit prior to the generation of the processing data by the generation unit. It has an evaluation unit that evaluates the processing performance when generating processing data, generates the processing data by the generation unit based on the evaluated processing performance, and transmits the generated processing data to the server. It is characterized by.

According to the present invention, it is possible to reduce the equipment cost and the operation cost in the AI information processing system used for machine learning by reducing the storage capacity of the storage device and improving the storage efficiency.

It is a figure which shows an example of the logical configuration of the hardware of the learning system including the storage device of Example 1. FIG. It is a figure which shows an example of the program and data stored in the memory of the storage apparatus of Example 1. FIG. It is a figure which shows an example of the data stored in the storage data area of the storage apparatus of Example 1. FIG. It is a figure which shows an example of the data generation list included in the data generation list set of Example 1. FIG. It is a figure which shows an example of the table which records the preliminary performance evaluation result of Example 1. FIG. It is a figure which shows an example of the program stored in the memory of the learning server apparatus of Example 1. FIG. It is a figure which shows an example of the logical configuration in the data registration phase of the storage apparatus of Example 1. FIG. It is a flowchart which shows an example of the process of the data registration phase in the storage apparatus of Example 1. FIG. It is a figure which shows an example of the logical configuration in the learning phase of the storage apparatus of Example 1. FIG. It is a flowchart which shows an example of the processing of the learning phase in the storage apparatus of Example 1. FIG. It is a flowchart which shows an example of the buffer area release processing in the storage apparatus of Example 1. FIG. It is a figure which shows an example of the logical configuration of the hardware of the learning system including the storage device of Example 2. FIG. It is a figure which shows an example of the program and data stored in the memory of the storage apparatus of Example 2. FIG. It is a figure which shows an example of the data stored in the storage data area of the storage apparatus of Example 2. FIG. It is a figure which shows an example of the extended data generation list included in the data generation list set of Example 2. FIG. It is a figure which shows an example of the table of the preliminary performance evaluation result of Example 2. It is a figure which shows an example of the logical configuration in the data registration phase of the storage apparatus of Example 2. FIG. It is a figure which shows an example of the logical configuration in the learning phase of the storage apparatus of Example 2.

Examples of the present invention will be described in detail below with reference to the drawings. In each of the drawings for explaining the following examples, the configuration requirements having the same or similar functions with the same reference number will be shown, and the description below will be omitted. Further, the examples and modifications can be combined in part or in whole within the scope of the technical idea of the present invention and within the range consistent with them.

In the following description, various information may be described by the expression of "aaa list" or "aaa table", but various information may be expressed by a data structure other than the list and the table. The "aaa list" and the "aaa table" can also be referred to as "aaa information" to show that they do not depend on the data structure. The "aaa list", "aaa table", or "aaa information" is stored in a storage area reserved for a storage resource (for example, memory).

Further, in the following description, the processing flow may be described with "program" as the subject. The program is loaded into a storage resource (for example, memory) by a processor (for example, a CPU (Central Processing Unit)) and executed, so that the program is predetermined while appropriately using at least one of the storage resource and the communication interface. Execute the processing. For example, an element described or illustrated as "xxx program" is a processing unit in which a processor executes a predetermined process by analyzing and executing a program source using a storage resource or a communication interface. Therefore, the subject of processing by the program may be a processor or a device having the processor. Further, the "xxx program" can be paraphrased as a "xxx part".

In addition, a part or all of the processing executed by the processor may be processed by the hardware circuit. The program that defines the processing to be executed by the processor may be, for example, acquired from an external device via a network or acquired via a storage medium and executed by the processor.

Example 1 of the present invention will be described with reference to FIGS. 1 to 10.

<Configuration of learning system including storage device of Example 1>
The logical structure of the hardware of the system which is the premise of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing an example of a logical configuration of hardware of a learning system including the storage device of the first embodiment. The learning system 1S including the storage device 104 of the first embodiment is a system in which a plurality of learning servers 101 and 102 and the storage device 104 are connected by a network 103.

The plurality of learning servers 101 and 102 of the learning system 1S are servers that perform machine learning of teacher data, and are a learning server (sub) 102 that shares learning processing such as product-sum calculation and error backpropagation processing, and learning. It is composed of a learning server (main) 101 that performs initialization and aggregation of learning results in addition to processing. Each of the learning servers 101 and 102 is composed of a CPU 111, a memory 112, and an IF 113. The learning servers 101 and 102 may be equipped with a storage device such as an HDD (Hard Disk Drive). Further, the learning servers 101 and 102 may be equipped with an accelerator for speeding up the learning process. It is also possible to perform the learning process only on the learning server (main) 101 without using the learning server (sub) 102.

The CPU 111 executes various processes including a learning process by using each program, management information, and the like stored in the memory 112.

The memory 112 stores a program executed by the CPU 111, management information used by the program, and the like. The memory 112 may be used for purposes such as storing information other than these, for example, management information of the OS (Operating System). Further, the memory 112 is generally composed of an SDRAM (Synchronous Dynamic Random Access Memory), but may be composed of a storage element other than this.

The IF 113 is an interface for the learning servers 101 and 102 to exchange information with each other and read learning data from the storage device 104 via the network 103. The IF 113 can be selected according to the configuration of the network 103, and can also be divided according to the application. For example, when the learning servers 101 and 102 communicate with each other, Ethernet (registered trademark, the same applies hereinafter) is used, and when the learning data is read from the storage device 104, Fiber Channel is used. It is also possible to obtain redundancy by linking a plurality of IF 113s.

The network 103 includes a concentrator and a cable, and handles communication between the learning servers 101 and 102 or the storage device 104. In this embodiment, the network 103 by Ethernet is taken as an example, but communication standards that are completely or partially different, such as Infiniband (registered trademark) and Fiber Channel, can be used. When different communication standards are mixed, a line concentrator and a cable that match the communication standard can be used together.

The storage device 104 includes an FE (front end) 121, a CPU 122, a BE (back end) 123, a memory 124, and one or more drives 125.

The FE 121 accepts and responds to the communication from the learning servers 101 and 102 to the storage device 104 via the network 103. As a communication standard, Ethernet, Fiber Channel, or the like can be used.

The CPU 122 responds to data access from the learning servers 101, 102 and other servers via the FE 121 by using each program, management information, and the like stored in the memory 124.

The BE 123 mediates access to the drive 125 in which the actual data is stored when the CPU 122 accesses from the learning servers 101, 102 and other servers. As a communication standard, SAS (Serial Attached SCSI), NVMe (Non-Volatile Memory Express), etc. can be used.

The memory 124 stores a program executed by the CPU 122, management information used by the program, and the like. Information other than these, for example, storage firmware management information may be stored in the memory 124. Further, the memory 124 is generally composed of an SDRAM (Synchronous Dynamic Random Access Memory), but may be composed of other storage elements.

One or more drives 125 are storage devices composed of HDDs and SSDs (Solid State Drives). Further, a plurality of drives 125 may be grouped into a unit called a parity group and used as a logical storage area by using a highly reliable technology such as RAID (Redundant Arrays of Independent Disks). The SSD is not limited to the one using the NAND Flash Memory as the storage area, and may be configured by using a storage element such as the Phase Change Memory.

<Programs and data stored in the memory of the storage device of Example 1>
FIG. 2 shows a program and information stored in the memory 124 of the storage device 104. FIG. 2 is a diagram showing an example of a program and data stored in the memory of the storage device of the first embodiment.

The memory 124 includes a data generation management program 201, a buffer size suppression program 202, a pre-performance evaluation program 203, a data generation processing program 204, a parameter reception program 205, an IO observation program 206, and a file emulation program 207. Is stored. Further, the memory 124 has a generated data buffer 211 as a temporary area for temporarily storing data. Although details will be described later, the generated data save area 303 (see FIG. 3) of the storage data area 301 described later also serves as a temporary area for temporarily storing data together with the generated data buffer 211. The memory 124 may also hold other control programs, control data, and the like.

The data generation management program 201 activates the data generation processing program 204, and the augmentation data is obtained with the data generation throughput corresponding to the learning data size required by the learning servers 101 and 102 and the data read throughput of the learning servers 101 and 102. Is dynamically generated and controlled to be stored in the temporary area. The generated augmentation data becomes training data.

The buffer size suppression program 202 releases the temporary area used by the generated augmentation data that has been read from the learning servers 101 and 102, that is, the closed generated augmentation data, and controls the suppression of the expansion of the temporary area. Do.

The pre-performance evaluation program 203 pre-evaluates the unit processing performance (processing performance per process) of data augmentation, that is, the data generation throughput before the actual learning process.

The data generation processing program 204 is started from the data generation management program 201 to perform actual data augmentation.

The parameter receiving program 205 has dynamic parameters that are dynamically determined when the user uses the learning servers 101 and 102, for example, the number of learning servers (subordinate) 102, the number of processing threads in the learning servers 101 and 102, and the memory. The number of teacher data, the number of epochs, etc. read at one time in 112 are received from the learning server (main) 101. Further, after the storage device 104 receives the dynamic parameters, when the response preparation of the augmentation data is completed inside the storage device 104, the learning server (main) 101 can read the augmentation data. Notify that.

The teacher data is a combination of learning data such as an image to be learned and correct answer information for the learning data. The epoch number is a numerical value indicating how many times a certain teacher data group is repeatedly used for learning.

The IO observation program 206 observes events such as opening, reading, and closing of the learning data file by the learning servers 101 and 102, and the data transfer speed of the learning data to the learning servers 101 and 102, and observes the data transfer speed of the learning data to the learning servers 101 and 102, and the data generation management program 201 and the buffer Notify the size suppression program 202.

The file emulation program 207 publishes the created augmentation data stored in the temporary area to the learning servers 101 and 102 as a normal file, or does the augmentation data before generation exist as if it were present? It is virtualized so that the learning servers 101 and 102 are not aware that dynamic data augmentation is performed.

The generated data buffer 211 is a temporary area in which the augmentation data generated by the data generation processing program 204 is stored. The augmentation data stored in the generated data buffer 211 is an entity of the augmentation data that the file emulation program 207 publishes to the learning servers 101 and 102.

<Data stored in the storage data area of the storage device of the first embodiment>
FIG. 3 shows information stored in the storage area of the storage device 104. FIG. 3 is a diagram showing an example of data stored in the storage data area of the storage device of the first embodiment.

The storage data area 301 is a storage area logically configured on the drive 125 included in the storage device 104. The storage data area 301 stores the pre-performance evaluation result 302, the generated data save area 303, the data generation list set 311 and the data augmentation program 312, and the original data 313.

Of these, the pre-performance evaluation result 302 and the generated data save area 303 are data that are stored in the storage data area 301 but are not disclosed to the user. On the other hand, the data generation list set 311 and the data augmentation program 312 and the original data 313 are data such as image data stored in the storage data area 301 by the user. Data other than the above may be stored in the storage data area 301.

The pre-performance evaluation result 302 is an area for storing a table in which the result of the pre-evaluation performed by the pre-performance evaluation program 203 is recorded. Details of the preliminary performance evaluation result 302 will be described later.

The generated data save area 303 contains, among the augmentation data generated by the data generation processing program 204 and stored in the generated data buffer 211, the data that cannot be stored in the generated data buffer 211 due to the capacity limitation. This is an area that is temporarily saved and stored.

The data generation list set 311 describes the relationship between the data augmentation that the user wants to perform on the original data 313 and the data that the learning servers 101 and 102 will use for learning as a result of the data augmentation. It is a set of data generation list 401 (see FIG. 4). The details of the data generation list 401 will be described later.

The data augmentation program 312 is a program or script used for data augmentation and is executed by the CPU 122 of the storage device 104.

The original data 313 is data to be processed (data augmentation) by the user. It is composed of a plurality of files, lists, etc., such as image data for learning and correct answer data.

<Data generation list of Example 1>
The data generation list 401 included in the data generation list set 311 will be described with reference to FIG. FIG. 4 is a diagram showing an example of a data generation list included in the data generation list set of the first embodiment. Each of the plurality of data generation lists 401 included in the data generation list set 311 is identified by being given a unique ID in the data generation list set 311.

The data generation list 401 includes columns for input data 411, data augmentation program 412, options 413, and output data 414. The data generation list 401 is composed of a plurality of lines, and the correspondence between one set of input data 411, the data augmentation program 412, the option 413, and the output data 414 is described in one line. One line corresponds to one output file.

In the input data 411, the file name of the original data 313 to be subjected to data augmentation is described. The file name of the original data 313 to be processed described in the input data 411 also includes the path information of the corresponding original data 313.

In the data augmentation program 412, the program name of the data augmentation to be executed for the original data 313 of the file name described in the input data 411 is described. Option 413 describes options such as arguments to be given when the data augmentation program described in the data augmentation program 412 is executed on the original data 313 with the file name described in the input data 411. Will be done.

The output data 414 executes the data augmentation program described in the data augmentation program 412 with the option described in option 413 given to the original data 313 with the file name described in the input data 411. The output file name of the augmentation data that is the processing result is described. The output file name described in the output data 414 also includes the path information of the corresponding output file. The output file name described in the output data 414 is a file name when the processing result of the data augmentation is published to the learning servers 101 and 102.

In the data augmentation program 412, NONE may be described as a keyword indicating that data augmentation is not performed.

For example, in the first line in FIG. 4, for the input data in which the input data 411 is “imgs01.jpg”, the data augmentation program 412 sets the data augmentation program “img_augmentation” to “-flip”. Indicates to execute with options. The output data 414 is "imgs01-f.jpg".

Further, for example, in the second line in FIG. 4, the data augmentation program 412 is "imgs_aug" for a plurality of input data listed in the list file in which the input data 411 is "imgs02.lst". Indicates that the augmentation program should be executed with the "-flip" and "-pack bin" options. The output data 414 is "imgs02-f.bin". “Imgs02-f.bin” is an intermediate file packed with multiple output data in which multiple input data are data-augmented by the option of “-pack bin”.

<Table for recording the preliminary performance evaluation results of Example 1>
A table for recording the pre-performance evaluation result 302 will be described with reference to FIG. FIG. 5 is a diagram showing an example of a table for recording the preliminary performance evaluation results of Example 1.

The table of pre-performance evaluation 302 includes columns for the data augmentation program 321 and options 322 and unit processing performance 323. The table of the pre-performance evaluation result 302 is composed of a plurality of rows, and the correspondence between one set of data augmentation program 321 and option 322 and the unit processing performance 323 is described in one row.

The combination of the data augmentation program 321 and the option 322 corresponds to the combination of the data augmentation program 412 and the option 413 described in the data generation list 401. That is, the pre-performance evaluation program 203 deduplications from all the combinations of the data augmentation program 412 and the option 413 described in each row of the table of the pre-performance evaluation result 302 are the input data 411 of the corresponding processing target. Execute for one of them and measure the performance per unit thread. Then, the pre-performance evaluation program 203 records the performance per unit thread measured by executing the combination of the data augmentation program 321 and the option 322 in the unit processing performance 323. The unit processing performance 323 is, for example, the amount of processing data (throughput) per unit time in a unit thread.

<Program stored in the memory of the learning server device>
A program stored in the memory 112 of the learning servers 101 and 102 will be described with reference to FIG. FIG. 6 is a diagram showing an example of a program stored in the memory of the learning server device of the first embodiment. The learning program 501 and the parameter transmission library 502 are stored in the memory 112.

The learning program 501 is a program that describes the content of the learning process that the user actually wants to perform. The processing by the learning program 501 is generally divided into an initialization phase and a learning phase.

The learning program 501 determines the parameters used during the learning process in the initialization phase. Specific examples of the parameters include the file name of the data generation list 401 corresponding to the learning data used by the learning program 501, the number of learning servers 101 and 102, the number of threads, the number of teacher data read into the memory 112 at one time, the number of epochs, and the like. Is.

In order to transmit these parameters to the parameter receiving program 205, a parameter transmitting API (Application Programming Interface) is described in the initialization phase portion of the learning program 501.

In general, the learning server (main) 101 and the learning server (slave) 102 are either the learning server (main) 101 or the learning server (slave) 102 based on the identification information of the learning servers 101 and 102. Is determined. The learning program 501 itself is shared between the learning server (main) 101 and the learning server (slave) 102, but the learning server (main) performs processing that needs to be performed only once for the learning servers 101 and 102 as a whole. ) It is possible to perform control such as having only 101 carry out.

The parameter transmission library 502 is an entity of a program called from the parameter transmission API described in the learning program 501. The parameter transmission library 502 has a function of notifying the storage device 104 of the dynamic parameters of the learning program 501 via the parameter receiving program 205, and a learning program 501 from after the dynamic parameter notification until the data access preparation to the storage device 104 is completed. Has a function to wait for the execution of.

<Learning process in the learning system of Example 1>
The procedure for performing the learning process in the learning system 1S of the first embodiment having the above is shown below. The learning process in the learning system 1S has two main phases. It is a data registration phase in which the user's data generation list 401, the data augmentation program 312 and the original data 313 are registered in the storage device 104, and a learning phase in which the teacher data is machine-learned.

<Data registration phase of Example 1>
The operation of the data registration phase of the first embodiment will be described with reference to FIGS. 7 and 8. FIG. 7 is a diagram showing an example of a logical configuration in the data registration phase of the storage device of the first embodiment, and shows the relationship between each program and data of the learning server (main) 101 and the storage device 104 in the data registration phase. FIG. 8 is a flowchart showing an example of processing of the data registration phase in the storage device of the first embodiment, and shows an operation flow of the storage device 104 in the data registration phase.

In the process of the data registration phase of the first embodiment, first, as shown in step S700 of FIG. 7, the learning server (main) 101 generates data in a predetermined data area (folder or the like) of the storage device 104. A certain data generation list 401 included in the list set 311, a data augmentation program 312, and original data 313 are written. The data and programs pointed to by the input data 411 and the data augmentation program 412 described in the data generation list 401 written in the data area of the storage device 104 are the original data 313 and the original data 313 written almost simultaneously with the data generation list 401. Data augmentation program 312.

As shown in FIG. 8, in step S701, the pre-performance evaluation program 203 of the storage device 104 monitors that the data generation list set 311 is updated. When the data generation list 401 included in the data generation list set 311 is newly registered or updated (step S701: YES), the pre-performance evaluation program 203 shifts the process to step S702, and neither new registration nor update is performed. If (step S701: NO), step S701 is repeated.

Subsequently, in step S702, the pre-performance evaluation program 203 confirms whether the data augmentation program 312 and the original data 313 corresponding to the newly registered or updated data generation list 401 have been stored. When the data augmentation program 312 and the original data 313 corresponding to the data generation list 401 are stored in the pre-performance evaluation program 203 (step S702: YES), the process shifts to step S703 and the process is not stored (step S702). : NO), step S702 is repeated, and the completion of storage is waited for.

Subsequently, in step S703, the pre-performance evaluation program 203 includes the data augmentation (data augmentation program and options) described in the data generation list 401 whose new addition or update has been confirmed in the data generation processing program 204. For example, a pre-performance evaluation is performed in which one of the above combinations) is executed for one input data by one thread of the CPU 122.

The data generation processing program 204 creates augmentation data and measures the performance (throughput, etc.) per unit thread at the time of pre-performance evaluation. Subsequently, in step S704, the pre-performance evaluation program 203 records the performance measurement result in the table of the pre-performance evaluation result 302 in the storage data area after the processing by the data generation processing program 204 is completed (see FIG. 5). ..

Subsequently, in step S705, the pre-performance evaluation program 203 causes the data generation processing program 204 to confirm whether or not there is a free area in the generated data buffer 211. When the generated data buffer 211 has a free area (step S705: YES), the pre-performance evaluation program 203 moves the process to step S712, and when there is no free area (step S705: NO), moves the process to step S706. ..

In step S706, the pre-performance evaluation program 203 causes the data generation processing program 204 to confirm whether or not there is a free area in the generated data save area 303. The pre-performance evaluation program 203 shifts the process to step S711 when there is a free area in the generated data save area 303 (step S706: YES), and performs the process in step S707 when there is no free area (step S706: NO). Move.

In step S707, the pre-performance evaluation program 203 causes the data generation processing program 204 to confirm whether a part of the generated data save area 303 can be released. When a part of the generated data save area 303 can be released (step S707: YES), the pre-performance evaluation program 203 shifts the process to step S710, and when it cannot be released (step S707: NO), the pre-performance evaluation program 203 proceeds to step S708. Move the process.

In step S708, the pre-performance evaluation program 203 causes the data generation processing program 204 to discard the augmentation data created during the pre-performance evaluation.

Subsequently, in step S709, the pre-performance evaluation program 203 determines whether or not there is data augmentation in the data generation list 401 for which new addition or update has been confirmed in step S701 for which the performance of step S703 has not been measured. To do. The pre-performance evaluation program 203 shifts the process to step S703 when there is a data augmentation for which performance measurement has not been performed in the data generation list 401 for which new addition or update has been confirmed (step S709: YES). On the other hand, the pre-performance evaluation program 203 shifts the process to step S701 when there is no data augmentation for which performance measurement has not been performed in the data generation list 401 for which new addition or update has been confirmed (step S709: NO). ..

On the other hand, in step S710, since the pre-performance evaluation program 203 has no free area in the generated data save area 303, the data generation processing program 204 has a retention priority among the data stored in the generated data save area 303. Free space is secured by discarding the ones with low. When step S710 is completed, the pre-performance evaluation program 203 shifts the process to step S711.

Further, in step S711, the pre-performance evaluation program 203 saves a part of the augmentation data stored in the generated data buffer 211 to the generated data saving area 303 in the data generation processing program 204. When step S711 is completed, the pre-performance evaluation program 203 shifts the process to step S712.

Further, in step S712, the pre-performance evaluation program 203 causes the data generation processing program 204 to store the augmentation data created at the time of the pre-performance evaluation in step S703 in the generated data buffer 211. When step S712 is completed, the pre-performance evaluation program 203 shifts the process to step S709.

In step S701, the process is transferred from step S709, the pre-performance evaluation program 203 reconfirms the data generation list set 311 and steps again if there is a data augmentation program 312 for which pre-performance evaluation has not been executed. The pre-performance evaluation of S703 is performed, and if it does not exist, the data generation list set 311 is waited to be updated. The above is the operation of the data registration phase.

It is not always the learning server (main) 101 that writes the data generation list set 311 and the data augmentation program 312 and the original data 313 to the predetermined data area (folder or the like) of the storage device 104. Any device can be used as long as it can access the storage device 104.

<Learning phase of Example 1>
The operation of the learning phase of the first embodiment is shown with reference to FIGS. 9, 10, and 11. The learning phase further includes a data generation operation flow and a generation data area release operation flow. FIG. 9 is a diagram showing an example of a logical configuration in the learning phase of the storage device of the first embodiment, and shows the relationship between each program and data of the learning servers 101 and 102 and the storage device 104 in the learning phase. FIG. 10 is a flowchart showing an example of processing in the learning phase in the storage device of the first embodiment, and shows a data generation operation flow of the storage device 104 in the learning phase. FIG. 11 is a flowchart showing an example of the buffer area release process in the storage device of the first embodiment, and shows the release operation flow of the generated data area that is no longer needed.

As shown in FIG. 9, the learning phase starts from the point where the learning server (main) 101 starts the learning program 501. The learning server (main) 101 executes the initialization phase of the learning program 501, and calls the parameter transmission API in the initialization phase. Then, in the learning server (main) 101, the parameter transmission library 502 activated by the parameter transmission API transmits the dynamic parameters of the learning program 501 to the parameter reception program 205 of the storage device 104. As described above, the dynamic parameters include, for example, the number of learning servers (subordinate) 102, the number of processing threads in the learning servers 101 and 102, the number of teacher data read into the memory 112 at one time, the number of epochs, and the like. After that, the parameter transmission library 502 waits for a response from the parameter reception program 205 and causes the learning program 501 to wait.

<Data generation operation flow of Example 1>
As shown in FIG. 10, first, in step S901, the data generation management program 201 determines whether the parameter receiving program 205 has received the dynamic parameters from the parameter transmitting library 502. When the parameter receiving program 205 receives the dynamic parameter from the parameter transmitting library 502 (step S901: YES), the data generation management program 201 shifts the process to step S902. On the other hand, the data generation management program 201 repeats step S901 when the parameter receiving program 205 has not received the dynamic parameters from the parameter transmitting library 502 (step S901: NO).

In step S902, the data generation management program 201 causes the parameter receiving program 205 to hold the dynamic parameters received from the parameter transmission library 502 so that they can be referred to by each program. The parameter receiving program 205 keeps the dynamic parameters available to each program, and notifies the data generation management program 201 and the buffer size suppression program 202 of the start of the learning phase.

Subsequently, in step S903, the data generation management program 201 receives the notification of the start of the learning phase from the parameter receiving program 205, reads the data generation list 401 corresponding to the augmentation data read in the learning phase, and at the time of pre-performance evaluation. It is confirmed whether the generated augmentation data is held in the generated data save area 303. The data generation management program 201 shifts the process to step S904 when the augmentation data generated at the time of pre-performance evaluation is held in the generated data save area 303 (step S903: YES) and is not held (step S903: YES). Step S903: NO), the process is transferred to step S905.

In step S904, the data generation management program 201 reads the augmentation data held in the generated data save area 303 in step S903 from the generated data save area 303 into the generated data buffer 211.

Subsequently, in step S905, the data generation management program 201 causes the data generation processing program 204 to generate the initial data to be used in the first learning, and the file emulation program 207 prepares for a response based on the generated initial data. Let me. The size and number of files of this initial data are determined based on the number of learning servers 101 and 102, the number of threads, and the number of teacher data read into the memory 112 at one time, and the determined size is the required number of data. Is only generated. Further, among the data corresponding to the initial data, the data read into the generated data buffer 211 in step S904 and the data generated at the time of the preliminary performance evaluation and remaining in the generated data buffer 211 are reused. To.

Subsequently, in step S906, when the response preparation of the data generation processing program 204 is completed, the data generation management program 201 notifies the parameter receiving program 205 of the response preparation completion, and causes the parameter transmission library 502 to respond to the response preparation completion. After that, when the parameter transmission library 502 receives the response preparation completion from the parameter reception program 205, the learning program 501 that has been kept on standby is restarted, and data access from the learning servers 101 and 102 to the storage device 104 is started.

Subsequently, in step S907, the data generation management program 201 determines whether data access (file open) from the learning servers 101 and 102 to the storage device 104 has occurred. Here, the IO observation program 206 uses the learning server 101, which uses the data access from the learning servers 101 and 102 to the storage device 104 and the augmentation data stored in the generated data buffer 211 by the file emulation program 207. It monitors the return of data to be accessed to 102. The IO observation program 206 notifies the data generation management program 201 of this data access and data return. At the same time, the IO observation program 206 also monitors the speed at which the learning servers 101 and 102 read the augmentation data and the access interval to the next data.

When the data access (file open) from the learning servers 101 and 102 to the storage device 104 occurs (step S907: YES), the data generation management program 201 shifts the process to step S908. On the other hand, the data generation management program 201 repeats step S907 when data access (file open) from the learning servers 101 and 102 to the storage device 104 has not occurred (step S907: NO).

In step S908, the data generation management program 201 determines whether subsequent data access following the data access in step S907 occurs. Here, the data generation management program 201 does not repeat the data reading for the number of epochs when the succeeding data of the data access notified from the IO observation program 206 in step S907 exists, or there is no succeeding data. If so, it is determined that subsequent data access will occur. The succeeding data is data corresponding to the line next to the line corresponding to the augmentation data currently read in the data generation list 401, and is data that can be determined to be expected to request the subsequent machining data. ..

When the subsequent data access occurs (step S908: YES), the data generation management program 201 shifts the process to step S909, and when the subsequent data access does not occur (step S908: NO), returns the process to step S901. When the data generation management program 201 determines that subsequent data access does not occur, it determines that learning using the learning data being accessed has been completed, and waits for the activation of the next parameter transmission API.

In step S909, the data generation management program 201 determines whether or not to generate the subsequent data depending on whether the generated data buffer 211 or the generated data save area 303 has necessary data corresponding to the subsequent data access. To do. The data generation management program 201 determines that the subsequent data needs to be generated when neither the generated data buffer 211 nor the generated data save area 303 has the necessary data corresponding to the subsequent data access (step S909: YES). , The process is transferred to step S910. On the other hand, the data generation management program 201 determines that the generation of the subsequent data is unnecessary when there is necessary data corresponding to the subsequent data access in any of the generated data buffer 211 and the generated data save area 303 (step S909: NO), the process is transferred to step S907, and the occurrence of data access is awaited.

In step S910, since the data generation management program 201 needs to generate subsequent data, it is based on the dynamic parameters held by the parameter receiving program 205 and the data access status observed by the IO observation program 206. , The learning servers 101 and 102 calculate the throughput for reading the learning data. Based on this calculation result and the unit processing performance (unit thread performance) of the data augmentation program 312 for the succeeding data stored in the pre-performance evaluation result 302, the data generation throughput exceeds the read throughput of the training data. The data generation resource (for example, the number of threads of the CPU 122) is adjusted and determined.

Subsequently, in step S911, the data generation management program 201 tells the data generation processing program 204 that the data generation process of the subsequent data access determined to be required to be generated in step S909 is adjusted by the data generation resource adjusted in step S910. Let it run. The data generation processing program 204 generates augmentation data to be used in subsequent data access and stores it in the generated data buffer 211. When step S911 is completed, the data generation management program 201 shifts the process to step S907 and waits for data access from the learning servers 101 and 102.

<Release operation flow of the generated data area of Example 1>
In steps S1001 and S1002 shown in FIG. 11, the buffer size suppression program 202 executes the same processing as steps S901 and S902 performed by the data generation management program 201 shown in FIG.

Subsequently, in step S1003, the buffer size suppression program 202 receives notification from the parameter receiving program 205 that it receives and holds the parameter transmission API, and the generated data buffer 211 from the learning servers 101 and 102 by the IO observation program 206. It waits for the detection notification of file close (access end) for the generated data, and determines whether the access end has occurred. The buffer size suppression program 202 shifts the process to step S1004 when the file close occurs (step S1003: YES), and repeats step S1003 when the file close does not occur (step S1003: NO).

In step S1004, the buffer size suppression program 202 counts the file close detection notification from the IO observation program 206 for each generated augmentation data, and the number of file closes is the number of simultaneous accesses, that is, the learning servers 101 and 102. Determine if it is equal to the number of. When the number of file closes becomes equal to the number of simultaneous accesses (step S1004: YES), the buffer size suppression program 202 shifts the process to step S1005, and when the number of file closes is less than the number of simultaneous accesses (step S1004: NO), the step The process is returned to S1003.

In step S1005, the buffer size suppression program 202 stores the generated data buffer 211 in which the data generation processing program 204 stores the generated data satisfying the release condition that the number of file closes becomes equal to the number of simultaneous accesses in step S1004. And the buffer area of the generated data save area 303 is released. The generated data in which the number of file closes is equal to the number of the learning servers 101 and 102 can be regarded as the end of access from the learning servers 101 and 102.

Subsequently, in step S1006, the buffer size suppression program 202 determines whether all the buffer areas used in the data generation list set 311 have been released. That is, the buffer size suppression program 202 is included in the data generation list set 311 in the buffer area of the generated data buffer 211 and the generated data save area 303 in which the generated data is stored via the data generation processing program 204. It is determined whether or not the used area of the data generation list 401 in which the augmentation data to be learned is currently described remains.

When all the buffer areas used in the data generation list set 311 are released (step S1006: YES), the buffer size suppression program 202 shifts the process to step S1001 and when all the buffer areas are not released. (Step S1006: NO), the process is transferred to step S1003. The buffer size suppression program 202 releases all the generated data buffers 211 and the generated data save area 303 used in the data generation list 401 included in the data generation list set 311 for the processes of steps S1003 to S1006. Repeat until.

In the above-described first embodiment, the storage device 104 performs data preparation processing (inflating of teacher data by processing image data) of learning data used in the learning system 1S on the storage device 104 side. That is, the storage device 104 performs the data preparation process after receiving the data request from the learning servers 101 and 102.

Here, when the file size of the teacher data generated in the data preparation process becomes large, for example, when packing a large amount of teacher data that cannot be expanded on the memories of the learning servers 101 and 102 and handling it as intermediate data. , The time (latency) from when the storage device 104 receives the data request to when the file is generated and the transmission is started becomes long, which may cause a problem that the learning server side times out.

On the other hand, in the first embodiment, the storage device 104 prepares transmission data in advance before accepting the data request, determines the buffer size and whether or not the buffer area can be released according to the request of the learning server, and buffers the data. Optimal management to minimize the size.

That is, in the first embodiment, the storage device 104 controls data generation according to the speed of data supply to the learning servers 101 and 102, and monitors data access from the plurality of learning servers 101 and 102, which is unnecessary. By releasing the buffer area that has become, the data supply and buffer discard by the storage device are optimally controlled.

In this way, the storage device 104 has two operation flows, a data generation operation flow and a generated data area release operation flow, in which dynamic augmentation data generation before the start of access during learning and after the end of access are performed. Release the storage buffer area for augmentation data in parallel. This makes it possible to prepare training data with the minimum required storage resources and use the augmentation data for training processing without causing a delay when opening a file.

Example 2 of the present invention will be described with reference to FIGS. 12 to 10. The second embodiment is different from the first embodiment in that an accelerator is used in addition to the CPU when generating augmentation data in the storage device.

The logical structure of the hardware of the system which is the premise of the present invention will be described with reference to FIG. FIG. 12 is a diagram showing an example of the logical configuration of the hardware of the learning system including the storage device of the second embodiment. The learning system 2S including the storage device 1104 of the second embodiment is different from the learning system 1S including the storage device 104 of the first embodiment in that the storage device 1104 further has an accelerator 1126. Other than that, the learning system 2S of the second embodiment is the same as the learning system 1S of the first embodiment, and thus the description thereof will be omitted.

The accelerator 1126 executes processing related to data augmentation at high speed instead of or in cooperation with the CPU 122. The accelerator 1126 can be configured by using an ASIC (Application Specific Integrated Circuit), a GPU (Graphics Processing Unit), an FPGA (Field Programmable Gate Array), or the like.

<Program and data stored in the memory of the storage device of Example 2>
FIG. 13 shows programs and information stored in the memory 124 of the storage device 1104. FIG. 13 is a diagram showing an example of a program and data stored in the memory of the storage device of the second embodiment.

The storage device 1104 of the second embodiment is different from the storage device 104 of the first embodiment in that the accelerator-compatible data generation processing program 1204 is stored in the memory 124 instead of the data generation processing program 204. The other programs and data stored in the memory 124 of the storage device 1104 of the second embodiment are the same as the programs and data stored in the memory 124 of the storage device 104 of the first embodiment, and thus the description thereof will be omitted.

The accelerator-compatible data generation processing program 1204 is started from the data generation management program 201 and performs actual data augmentation. The accelerator-corresponding data generation processing program 1204 uses the accelerator 1126 in addition to the CPU 122 to perform processing.

<Data stored in the storage data area of the storage device of the second embodiment>
FIG. 14 shows information stored in the storage area of the storage device 1104. FIG. 14 is a diagram showing an example of data stored in the storage data area of the storage device of the second embodiment.

The storage data area 1301 of the second embodiment is a storage area logically configured on the drive 125 included in the storage device 104. Compared with the storage data area 301 of the first embodiment, the storage data area 1301 of the second embodiment stores the pre-performance evaluation result 1302 instead of the pre-performance evaluation result 302, and further stores the accelerator program 1314. different. The other data stored in the storage data area 1301 of the second embodiment is the same as the data stored in the 301 in the storage data area of the first embodiment, and thus the description thereof will be omitted.

The pre-performance evaluation result 1302 and the generated data save area 303 are data stored in the storage data area 1301 but not disclosed to the user. On the other hand, the data generation list set 311, the data augmentation program 312, the accelerator program 1314, and the original data 313 are data stored in the storage data area 1301 by the user. Data other than the above data may be stored in the storage data area 1301.

The data generation list set 311 includes data augmentation that the user wants to perform on the original data 313, accelerators 1126 to be used, and data that the learning servers 101 and 102 will use for learning as a result of the data augmentation. A set of lists that describe the relationships. In the first embodiment, the data generation list set 311 includes the data generation list 401, but in the second embodiment, the extended data generation list 1401 is included instead of the data generation list 401.

<Extended data generation list of Example 1>
FIG. 15 is a diagram showing an example of a data generation list included in the data generation list set of the second embodiment. Each of the plurality of extended data generation lists 1401 included in the data generation list set 311 is assigned a unique ID in the data generation list set 311 and identified.

The extended data generation list 1401 is different from the data generation list 401 of the first embodiment in that it further includes two columns, an accelerator 1411 and an accelerator program 1412. Since the extended data generation list 1401 is the same as the data generation list 401 of the first embodiment in other respects, the description thereof will be omitted.

The extended data generation list 1401 is composed of a plurality of lines, and the correspondence between a set of input data 411, a data augmentation program 412, an option 413, an accelerator 1411, an accelerator program 1412, and an output data 414 is described in one line. One line corresponds to one output file.

In the extended data generation list 1401, the type of the accelerator 1126 used in the data augmentation for the input data 411 is described in the accelerator 1411, and the program name executed by the accelerator 1126 is described in the accelerator program 1412. Options may be added to the accelerator program.

In addition, NONE may be described as a keyword indicating that the accelerator 1126 is not used in the accelerator 1411. Further, when a plurality of accelerators 1126 are mounted on the storage device 1104, each row of the accelerator 1411 and the accelerator program 1412 is provided for each accelerator 1126, and the accelerator 1126 is used or not used for each accelerator 1126. The name of the program executed by 1126 may be described.

<Table for recording the preliminary performance evaluation results of Example 2>
A table for recording the pre-performance evaluation result 1302 will be described with reference to FIG. FIG. 16 is a diagram showing an example of a table for recording the preliminary performance evaluation results of Example 2.

The table of the pre-performance evaluation result 1302 is different from the table of the pre-performance evaluation result 302 of the first embodiment in that it further includes two columns of the accelerator 1321 and the accelerator program 1322. Since the table of the pre-performance evaluation result 1302 is the same as the table of the pre-performance evaluation result 302 of the first embodiment in other respects, the description thereof will be omitted.

The table of the pre-performance evaluation result 1302 is composed of a plurality of rows, and the correspondence between one set of data augmentation program 321 and option 322, the accelerator 1321, the accelerator program 1322, and the unit processing performance 323 is described in one row.

The combination of the data augmentation program 321 and the option 322, the accelerator 1321 and the accelerator program 1322 is a combination of the data augmentation program 412, the option 413, the accelerator 1411 and the accelerator program 1412 described in the extended data generation list 1401. Corresponds to. That is, the pre-performance evaluation program 203 corresponds to a combination that is deduplicated from all combinations of the data augmentation program 412, the option 413, the accelerator 1411, and the accelerator program 1412 described in each row of the table of the pre-performance evaluation result 1302. It is executed for one of the input data 411 to be processed and the performance per unit thread is measured. Then, the pre-performance evaluation program 203 records the performance per unit thread measured by executing the combination of the data augmentation program 321 and the option 322, the accelerator 1321 and the accelerator program 1322 in the unit processing performance 323.

<Learning process in the learning system of Example 2>
The procedure for performing the learning process in the learning system 2S of the second embodiment having the above is shown below. The learning process in the learning system 2S has a data registration phase and a learning phase, similarly to the learning process in the learning system 1S of the first embodiment. Hereinafter, only the difference from the first embodiment will be described.

<Data registration phase of Example 2>
The operation of the data registration phase of the second embodiment will be described with reference to FIG. FIG. 17 is a diagram showing an example of a logical configuration in the data registration phase of the storage device of the second embodiment, and shows the relationship between each program and data of the learning server (main) 101 and the storage device 1104 in the data registration phase. For the explanation of the processing flow of the data registration phase in the storage device of the second embodiment, FIG. 8 is referred to, and the storage device 104, the data generation processing program 204, and the data generation list 401 of the first embodiment are described in the second embodiment. The description will be replaced with the storage device 1104, the accelerator-compatible data generation processing program 1204, and the extended data generation list 1401.

In the process of the data registration phase of the second embodiment, first, as shown in step S1700 of FIG. 17, the learning server (main) 101 generates data in a predetermined data area (folder or the like) of the storage device 1104. The extended data generation list 1401 included in the list set 311, the data augmentation program 312, the accelerator program 1214, and the original data 313 are written. The data and programs pointed to by the input data 411, the data augmentation program 412, and the accelerator program 1412 described in the extended data generation list 1401 written in the data area of the storage device 1104 are substantially the same as the data generation list 401. The original data 313, the data augmentation program 312, and the accelerator program 1214 are written at the same time.

Further, in the second embodiment, in step S701 of FIG. 8, the pre-performance evaluation program 203 of the storage device 104 monitors that the data generation list set 311 is updated. When the extended data generation list 1401 included in the data generation list set 311 is newly registered or updated (step S701: YES), the pre-performance evaluation program 203 shifts the process to step S702, and the new registration and the update are also performed. If not (step S701: NO), step S701 is repeated.

Further, in the second embodiment, in step S702 of FIG. 8, the pre-performance evaluation program 203 is the data augmentation program 312, the accelerator program 1214, and the original data corresponding to the newly registered or updated extended data generation list 1401. Check if 313 is stored. When the data augmentation program 312 corresponding to the extended data generation list 1401, the accelerator program 1214, and the original data 313 have already been stored (step S702: YES), the pre-performance evaluation program 203 shifts the process to step S703. If it has not been stored (step S702: NO), step S702 is repeated and the storage is completed.

Further, in the second embodiment, in step S703 of FIG. 8, the pre-performance evaluation program 203 is described in the extended data generation list 1401 in which new addition or update is confirmed in the accelerator-compatible data generation processing program 1204. A pre-performance evaluation is performed in which one of the mentions (a combination of a data augmentation program, an option, an accelerator, and an accelerator program) is executed for one input data by, for example, one thread of the CPU 122.

When the extended data generation list 1401 specifies that the accelerator 1126 is used, the accelerator-compatible data generation processing program 1204 also uses the accelerator 1126 to create augmentation data. The accelerator-compatible data generation processing program 1204 creates augmentation data and measures the performance (throughput, etc.) per unit thread at the time of pre-performance evaluation.

Further, in the second embodiment, in step S704 of FIG. 8, the pre-performance evaluation program 203 sets the performance measurement result of the pre-performance evaluation result 1302 in the storage data area after the processing by the accelerator-compatible data generation processing program 1204 is completed. Record on the table (see Figure 16).

Further, in the second embodiment, the same processing as in the first embodiment is performed in steps S705 to S708 and steps S710 to S712 of FIG.

Further, in the second embodiment, in step S709 of FIG. 8, the pre-performance evaluation program 203 includes data in the extended data generation list 1401 for which new addition or update is confirmed in step S701, in which the performance of step S703 is not measured. Determine if there is augmentation. The pre-performance evaluation program 203 shifts the process to step S703 when there is data augmentation for which performance measurement has not been performed in the extended data generation list 1401 for which new addition or update has been confirmed (step S709: YES). On the other hand, the pre-performance evaluation program 203 shifts the process to step S701 when there is no data augmentation for which performance measurement has not been performed in the data generation list 401 for which new addition or update has been confirmed (step S709: NO). .. The above is the operation of the data registration phase of the second embodiment.

It is not always the learning server (mainly) that writes the data generation list set 311, the data augmentation program 312, the accelerator program 1214, and the original data 313 to the predetermined data area (folder, etc.) of the storage device 1104. ) 101, and any device that can access the storage device 1104 may be used.

<Learning phase of Example 2>
FIG. 18 shows the operation of the learning phase of the second embodiment. In the second embodiment as well, the learning phase further includes the data generation operation flow and the release operation flow of the generated data area, as in the first embodiment. FIG. 18 is a diagram showing an example of a logical configuration in the learning phase of the storage device of the second embodiment, and shows the relationship between each program and data of the learning servers 101 and 102 and the storage device 1104 in the learning phase.

For the explanation of the data generation operation flow of the learning phase in the storage device of the second embodiment, FIG. 10 is referred to, and the storage device 104, the data generation processing program 204, and the data generation list 401 of the first embodiment are referred to in the second embodiment. The storage device 1104, the accelerator-compatible data generation processing program 1204, and the extended data generation list 1401 will be described. In the following, only the steps having a difference from the first embodiment will be described.

As shown in FIG. 18, the learning phase of the second embodiment starts from the point where the learning server (main) 101 starts the learning program 501. The learning server (main) 101 executes the initialization phase of the learning program 501, and calls the parameter transmission API in the initialization phase. Then, in the learning server (main) 101, the parameter transmission library 502 activated by the parameter transmission API transmits the dynamic parameters of the learning program 501 to the parameter reception program 205 of the storage device 1104.

<Data generation operation flow of Example 2>
In the second embodiment, in step S903 of FIG. 10, the data generation management program 201 receives the notification of the start of the learning phase from the parameter receiving program 205, and displays the extended data generation list 1401 corresponding to the augmentation data read in the learning phase. It is confirmed whether the augmentation data generated at the time of reading and pre-performance evaluation is held in the generated data save area 303. The data generation management program 201 shifts the process to step S904 when the augmentation data generated at the time of pre-performance evaluation is held in the generated data save area 303 (step S903: YES) and is not held (step S903: YES). Step S903: NO), the process is transferred to step S905.

Further, in the second embodiment, in step S905 of FIG. 10, the data generation management program 201 causes the accelerator-compatible data generation processing program 1204 to generate the initial data to be used at the first learning, and based on the generated initial data. To prepare the file emulation program 207 for response. At this time, when the use of the accelerator 1126 is specified in the extended data generation list 1401, the accelerator 1126 and the accelerator program 1214 are also used. The size and number of files of this initial data are determined based on the number of learning servers 101 and 102, the number of threads, and the number of teacher data read into the memory 112 at one time, and the determined size is the required number of data. Is only generated. Further, among the data corresponding to the initial data, the data read into the generated data buffer 211 in step S904 and the data generated at the time of the preliminary performance evaluation and remaining in the generated data buffer 211 are reused. To.

Further, in the second embodiment, in step S906 of FIG. 10, when the response preparation of the accelerator-compatible data generation processing program 1204 is completed, the data generation management program 201 notifies the parameter receiving program 205 of the response preparation completion, and the response preparation is completed. Is made to respond to the parameter transmission library 502. After that, when the parameter transmission library 502 receives the response preparation completion from the parameter reception program 205, the learning program 501 that has been kept on standby is restarted, and data access from the learning servers 101 and 102 to the storage device 1104 is started.

Further, in the second embodiment, in step S908 of FIG. 10, the data generation management program 201 determines whether a subsequent data access following the data access in step S907 occurs. Here, the data generation management program 201 does not repeat the data reading for the number of epochs when the succeeding data of the data access notified from the IO observation program 206 in step S907 exists, or there is no succeeding data. If so, it is determined that subsequent data access will occur. The subsequent data is the data corresponding to the line following the line corresponding to the augmentation data currently being read in the extended data generation list 1401.

Further, in the second embodiment, in step S911 of FIG. 10, the data generation management program 201 causes the accelerator-compatible data generation processing program 1204 to execute the data generation processing of the subsequent data determined to be necessary to be generated in step S909. When step S911 is completed, the data generation management program 201 shifts the process to step S907, prepares the generated data buffer 211 for the data to be used for the next access, and waits for the data access from the learning servers 101 and 102. ..

<Release operation flow of the generated data area of Example 1>
For the explanation of the release operation flow of the generated data area in the storage device of the second embodiment, FIG. 11 is referred to, and the storage device 104, the data generation processing program 204, and the data generation list 401 of the first embodiment are described in the second embodiment. The description will be replaced with the storage device 1104, the accelerator-compatible data generation processing program 1204, and the extended data generation list 1401. In the following, only the steps having a difference from the first embodiment will be described.

In the second embodiment, in step S1005 of FIG. 11, the buffer size suppression program 202 stores the generated data in which the number of file closes equals the number of simultaneous accesses in step S1004 in the accelerator-compatible data generation processing program 1204. The buffer areas of the completed data buffer 211 and the generated data saving area 303 are released.

Further, in the second embodiment, in step S1006 of FIG. 11, the buffer size suppression program 202 determines whether or not all the buffer areas used in the data generation list set 311 have been released. That is, the buffer size suppression program 202 uses the data generation list set 311 in the buffer area of the generated data buffer 211 in which the generated data is stored and the generated data save area 303 via the accelerator-compatible data generation processing program 1204. Determine if the middle area remains.

Further, in the second embodiment, the buffer size suppression program 202 shifts the processing to step S1001 when all the buffer areas used in the data generation list set 311 are released (step S1006: YES), and all the buffers. If the area is not released (step S1006: NO), the process is transferred to step S1003. The buffer size suppression program 202 releases all the generated data buffers 211 and the generated data save area 303 used in the extended data generation list 1401 included in the data generation list set 311 for the processes of steps S1003 to S1006. Repeat until it is done.

In the above two operation flows, the data generation operation flow and the release operation flow of the generated data area, the dynamic augmentation data generation using the accelerator 1126 before the access start at the time of learning and the augmentation after the access end. It is possible to release the storage buffer area of the station data in parallel. This makes it possible to prepare training data at high speed using an accelerator with the minimum required storage resources, and to use the augmentation data for training processing without causing a delay when opening a file.

The present invention is not limited to the above-described examples, but includes various modifications. For example, the above-described embodiment has been described in detail in order to explain the present invention in an easy-to-understand manner, and is not necessarily limited to those having the described configuration. Further, as long as there is no contradiction, it is possible to replace a part of the configuration of one embodiment with the configuration of another embodiment, or to add the configuration of another embodiment to the configuration of one embodiment. In addition, it is possible to add, delete, replace, integrate, and divide a part of the configuration of each embodiment.

1S, 2S: Learning system, 101: Learning server (main), 102: Learning server (subordinate), 103: Network, 104: Storage device, 112: Memory, 122: CPU, 124: Memory, 125: Drive, 201: Data generation management program, 202: Buffer size suppression program, 203: Pre-performance evaluation program, 204: Data generation processing program, 205: Parameter reception program, 206: IO observation program, 207: File emulation program, 211: Generated data buffer , 301: Storage data area, 302: Preliminary performance evaluation result, 303: Generated data save area, 311: Data generation list set, 312: Data augmentation program, 313: Original data, 321: Data augmentation program, 322: Option, 323: Unit processing performance, 401: Data generation list, 411: Input data, 412: Data augmentation program, 413: Option, 414: Output data, 501: Learning program, 502: Parameter transmission library, 1104 : Storage device, 1126: Accelerator, 1204: Accelerator-compatible data generation processing program,
1214: Accelerator program, 1301: Storage data area, 1302: Pre-performance evaluation result, 1314: Accelerator program, 1321: Accelerator, 1322: Accelerator program, 1401: Extended data generation list, 1411: Accelerator, 1412: Accelerator program

Claims (13)

In a storage device connected to a server via a network
A storage unit that stores the original data acquired from the outside,
A generation unit that generates processing data by processing the original data selected from the original data stored in the storage unit in response to a request for processing data from the server.
Prior to the generation of the processing data by the generation unit, the generation unit has an evaluation unit for evaluating the processing performance when the processing data is generated.
A storage device characterized in that the generation unit generates the processing data based on the evaluated processing performance and transmits the generated processing data to the server. It also has a generation management unit that manages the generation of machining data by the generation unit.
The generation management unit
The data read speed from the storage device to the server is calculated from the parameters related to data read from the storage device to the server and the access status from the server to the storage device notified in advance, and evaluated by the evaluation unit. The generation speed at which the generation unit generates the processing data is adjusted from the generation speed per unit resource of the processed processing data and the calculated data reading speed.
The generator
The storage device according to claim 1, wherein processing data to be transmitted to the server is generated at a generation rate adjusted by the generation management unit. The storage device according to claim 2, wherein the evaluation unit evaluates the generation speed of the processing data per unit resource for each program to perform the processing. When the original data is registered in the storage unit, the evaluation unit causes the generation unit to perform the processing on one original data for each program to perform the processing, so that the unit of the processing data is The storage device according to claim 3, wherein the generation rate per resource is evaluated. It has a storage area for storing the processing data generated by the generation unit, and has a storage area.
The storage device according to claim 1, wherein the processing data stored in the storage area is transmitted to the server in response to a request for processing data from the server. The processing data generated by the generation unit at the time of evaluation of the processing performance by the evaluation unit is stored in the storage area.
When the processing data is requested from the server and the processing data generated by the generation unit at the time of the evaluation is stored in the storage area, the processing data is read out from the storage area. The storage device according to claim 5, wherein the data is transmitted to the server. The generation management unit
After transmitting the machining data to the server in response to the request from the server, it is determined whether or not the request for machining data following the request for the machining data is expected from the server.
The generator
When it is determined by the generation management unit that a request for subsequent processing data is expected, the subsequent processing data to be transmitted to the server in response to the request from the server is generated in advance and stored in the storage area. The storage device according to claim 5, wherein the storage device is stored. Storage area capacity control that releases the storage area in which the processing data is temporarily stored based on the number of servers that simultaneously access the storage device included in the parameters and the number of file closes for the processing data by the server. The storage device according to claim 5, further comprising a unit. The storage device according to claim 1, wherein the original data is image data. The server is a server that performs machine learning of teacher data.
The storage device according to claim 1, wherein the generation unit performs data augmentation on data selected from the original data to generate the processed data to be the teacher data. Has more accelerators,
The storage device according to claim 1, wherein the generation unit uses the accelerator to generate the processing data. The storage device according to claim 11, wherein the evaluation unit evaluates the processing performance for each of the accelerator that performs the processing and the accelerator program executed by the accelerator. In the data processing method in the storage device connected to the server via the network
The storage device
The original data acquired from the outside is stored in the storage unit,
In response to a request for processing data from the server, processing is performed on the original data selected from the original data stored in the storage unit to generate processing data.
Prior to the generation of the processing data, the processing performance at the time of generating the processing data is evaluated.
The processing data is generated based on the evaluated processing performance, and the processing data is generated.
A data processing method in a storage device, which includes each process of transmitting the generated processed data to the server.

Images