system

Abstract

Reduce the development costs of learning models. [Solution] System (1) is a system for controlling machine learning of a model on a computing infrastructure comprising multiple nodes configured to communicate via a network (NW). The system comprises acquisition means (111) for acquiring learning data information relating to learning data and time information indicating the learning deadline, and selection means (112) for selecting one or more nodes from the multiple nodes to be used for machine learning of the model based on node information relating to the multiple nodes, the learning data information, and the time information. The node information includes cost information relating to the usage costs. The selection means selects one or more nodes such that machine learning is completed within the learning deadline indicated by the time information and the cost required for machine learning is minimized.

Description

Technical Field

[0001] The present invention relates to a system, and more specifically, to the technical field of a system for selecting a data center for implementing machine learning.

Background Art

[0002] Services using a trained model (i.e., AI (Artificial Intelligence)) generated by machine learning have been proposed. For example, in Patent Document 1, the types and installation environments of signs or markings around a vehicle, the driving situation of the vehicle, the position of the vehicle, and the line-of-sight direction of the driver of the vehicle are input into the trained model, and a system for providing safe driving support information based on the output of the trained model is described.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] In the technology described in Patent Document 1, a trained model is used in an in-vehicle device, but a trained model may also be used in a data center having higher processing power than the in-vehicle device. For example, in the service for providing safe driving support information described in Patent Document 1, it is required to process relatively little data from the vehicle in real time. On the other hand, in machine learning for generating a learning model, it is required to process a large amount of data. That is, the performance required for the data center used to provide the above service is different from the performance required for the data center for implementing machine learning. By the way, the cost required for developing a learning model often varies depending on the data center used for developing the learning model (in other words, for implementing machine learning). If no measures are taken, there is a technical problem that the development cost of the learning model increases.

[0005] This invention has been made in view of the above-mentioned problems, and aims to provide a system that allows for the selection of a data center in a way that reduces the development costs of a learning model. [Means for solving the problem]

[0006] A system according to one aspect of the present invention is a system for controlling machine learning of a model on a computing infrastructure comprising a plurality of nodes configured to communicate via a network, comprising: acquisition means for acquiring learning data information relating to learning data and time information indicating a learning deadline; node information relating to the plurality of nodes; and selection means for selecting one or more nodes from the plurality of nodes to be used for machine learning of the model, based on the learning data information and the time information, wherein the node information includes cost information relating to usage costs, and the selection means selects one or more nodes such that machine learning is completed within the learning deadline indicated by the time information and the cost required for machine learning is small. [Brief explanation of the drawing]

[0007] [Figure 1] This is a conceptual diagram illustrating the concept of the system according to the embodiment. [Figure 2] This is a block diagram showing the configuration of the information processing device according to the embodiment. [Figure 3] This figure shows an example of computing infrastructure information. [Figure 4] This figure shows an example of job information. [Figure 5] This figure shows an example of an image used to input job information. [Figure 6] This figure shows an example of an image illustrating the results of machine learning. [Modes for carrying out the invention]

[0008] Embodiments of the system will be described with reference to Figures 1 to 6.

[0009] (system) A system according to an embodiment will be described with reference to Figure 1. In Figure 1, System 1 comprises data centers DC1, DC2, and DC3, and clouds CL1 and CL2, which are connected to each other via a network NW. The number of data centers in System 1 may be two or less, or four or more. The number of clouds in System 1 may be one, or three or more.

[0010] Furthermore, the locations of data centers DC1, DC2, and DC3 are arbitrary. For example, data center DC1 may be in Japan, data center DC2 may be in the United States, and data center DC3 may be in Europe. For example, data center DC1 may be in Aichi Prefecture, data center DC2 may be in Kyushu, and data center DC3 may be in Hokkaido.

[0011] Furthermore, at least one of the data centers DC1, DC2, and DC3 may be a containerized data center. Used batteries from BEVs (Battery Electric Vehicles) may be used as at least part of the power supply for the containerized data center.

[0012] At least one of data centers DC1, DC2, and DC3 may be a data center owned by the company (i.e., an on-premises data center). Data centers DC1, DC2, and DC3 may include data centers provided by other providers (i.e., hosted data centers). At least one of clouds CL1 and CL2 may be a public cloud where the environment built by the cloud provider is shared with other users. Clouds CL1 and CL2 may include a hosted private cloud where a specific user has exclusive use of the cloud environment provided by the cloud provider. Note that a hosted data center and a hosted private cloud may be the same concept.

[0013] Furthermore, data centers DC1, DC2, and DC3, as well as clouds CL1 and CL2, may be referred to as "nodes." The network NW may also be referred to as a "link." Therefore, System 1 can be described as a computing infrastructure comprising multiple nodes configured to communicate via the network NW.

[0014] System 1 includes a database DB. The database DB contains training data used for machine learning. The training data included in the database DB may be training data from a commercially available training dataset. The training data included in the database DB may be training data based on data collected from multiple vehicles (e.g., connected cars).

[0015] In System 1, machine learning using at least a portion of the training data contained in the database DB may be performed in at least a portion of the data centers DC1, DC2, and DC3, as well as in at least a portion of the clouds CL1 and CL2.

[0016] (Configuration of information processing device) System 1 includes an information processing device 100. The information processing device 100 will be described with reference to Figure 2. In Figure 2, the information processing device 100 includes an arithmetic unit 110, a storage device 120, a communication device 130, an input device 140, and an output device 150. The arithmetic unit 110, the storage device 120, the communication device 130, the input device 140, and the output device 150 may be connected via a data bus 160.

[0017] Furthermore, the information processing device 100 does not necessarily have to include at least one of the input device 140 and the output device 150. In this case, at least one of the input device 140 and the output device 150 may be connected to the information processing device 100 via an input / output port (not shown) of the information processing device 100 (i.e., at least one of the input device 140 and the output device 150 may be externally attached to the information processing device 100).

[0018] The computing device 110 may have one or more processors. The processor may be, for example, at least one of a CPU (Central Processing Unit) and a GPU (Graphics Processing Unit).

[0019] The storage device 120 may have one or more memories. The memory may be, for example, at least one of a RAM (Random Access Memory), a ROM (Read Only Memory), a hard disk drive, a magneto-optical disk drive, and an SSD (Solid State Drive).

[0020] The communication device 130 may be capable of communicating with a device external to the information processing device 100. Note that the communication device 130 may perform wired communication or wireless communication.

[0021] The input device 140 is a device that can receive input of information to the information processing device 100 from the outside. The input device 140 may include an operating device (for example, a keyboard, a mouse, a touch panel, etc.) that can be operated by a user of the information processing device 100. The input device 140 may include a recording medium reader that can read information recorded on a recording medium detachable from the information processing device 100, such as a USB (Universal Serial Bus) memory. Note that when information is input to the information processing device 100 via the communication device 130 (in other words, when the information processing device 100 acquires information via the communication device 130), the communication device 130 may function as an input device.

[0022] The output device 150 is a device capable of outputting information to the outside of the information processing device 100. The output device 150 may output visual information such as characters and images, auditory information such as sounds, or tactile information such as vibrations. The output device 150 may include, for example, at least one of a display, speaker, printer, and vibration motor. The output device 150 may also be capable of outputting information to a recording medium that can be attached to or detached from the information processing device 100, such as a USB memory stick. When the information processing device 100 outputs information via the communication device 130, the communication device 130 may function as an output device.

[0023] The storage device 120 is capable of storing desired data. The storage device 120 may store a computer program that the arithmetic unit 110 will execute. The storage device 120 may also temporarily store data that the arithmetic unit 110 will use temporarily when the arithmetic unit 110 is executing a computer program.

[0024] Furthermore, the computer program may be recorded on a recording medium that is readable by a computer and is not temporary. In this case, the information processing device 100 may read the computer program from the recording medium using a recording medium reading device (not shown). As a result, the computer program may be stored in the storage device 120. Furthermore, at least one of the following may be used as the recording medium: an optical disc, a magnetic medium, a magneto-optical disc, a semiconductor memory, and any other medium capable of storing a program.

[0025] Furthermore, the computer program may be obtained from an external device (not shown) of the information processing device 100 via the communication device 130. In other words, the information processing device 100 may download the computer program via the communication device 130. As a result, the computer program may be stored in the storage device 120.

[0026] The arithmetic unit 110 may, together with the storage device 120 in which the computer program is stored, execute the processing that the information processing device 100 should perform. In other words, the arithmetic unit 110 may, together with the storage device 120 and the computer program stored in the storage device 120, execute the processing that the information processing device 100 should perform. For example, by the arithmetic unit 110 executing the computer program, a logical functional block for executing the processing that the information processing device 100 should perform may be realized within the arithmetic unit 110.

[0027] For example, the arithmetic unit 110 may have an acquisition unit 111, a selection unit 112, and a determination unit 113 as the above-mentioned functional blocks. Furthermore, the arithmetic unit 110 may also have an acquisition unit 111, a selection unit 112, and a determination unit 113 as physically implemented processing circuits. At least one of the acquisition unit 111, the selection unit 112, and the determination unit 113 may be implemented in a form that combines a logical functional block and a physical processing circuit (i.e., hardware). Details of the acquisition unit 111, the selection unit 112, and the determination unit 113 will be described later.

[0028] The storage device 120 stores computing resource information 121 and job information 122. Computing resource information 121 is information about computing resources available for machine learning. For example, computing resource information 121 may be information about data centers DC1, DC2, and DC3, and clouds CL1 and CL2, respectively. For example, as shown in Figure 3, computing resource information 121 may be information indicating the computing performance, availability, and failure information of each data center. For example, a data center may be represented by information for identifying the data center. For example, computing performance may be represented in FLOPS (Floating-Point Operations Per Second). For example, availability may be represented by the number of available cores.

[0029] Furthermore, the term "data center" in Figure 3 is not limited to data centers DC1, DC2, and DC3, but also includes clouds CL1 and CL2. As mentioned above, data centers DC1, DC2, and DC3, as well as clouds CL1 and CL2, may be referred to as "nodes." For this reason, the computing resource information 121 may also be referred to as node information. In addition to "data center," "computational performance," "availability," and "usage fee," the computing resource information 121 may include other items.

[0030] Job information 122 is information about a job related to machine learning. For example, as shown in Figure 4, job information 122 may be information indicating the learning deadline, dataset, and data volume for each job. For example, the learning deadline may be a date indicating the learning deadline, or it may be the period from the present until the learning deadline. For example, the dataset may be represented by information to identify the dataset used for machine learning. For example, the data volume may be information indicating the amount of data in the dataset. In addition to "job," "learning deadline," "dataset," and "data volume," job information 122 may include other items.

[0031] When a user of the information processing device 100 registers a job, the image 20 shown in Figure 5 may be displayed on the display of the output device 150 as an example. For example, the user may enter the necessary information into at least one of the multiple input fields included in the image 20 via the input device 140. When the user presses the "OK" button included in the image 20 via the input device 140, the information entered by the user is registered in the job information 122.

[0032] For example, information entered in the input field for "Job Name" in Image 20 may be stored in the "Job" field of Job Information 122. For example, information entered in the input field for "Dataset" in Image 20 may be stored in the "Dataset" field of Job Information 122. For example, information entered in the input field for "Learning Deadline" in Image 20 may be stored in the "Learning Deadline" field of Job Information 122. For example, the information processing device 100 may determine the amount of data in the dataset based on the information entered in the input field for "Dataset" in Image 20. The information processing device 100 may store the determined amount of data in the "Amount of Data" field of Job Information 122.

[0033] (Operation of information processing device) Next, the operation of the information processing device 100 will be described. Here, we will describe the process by which the information processing device 100 selects a data center to perform machine learning corresponding to the job included in the job information 122. Hereafter, "data center" is a concept that includes not only data centers DC1, DC2, and DC3, but also clouds CL1 and CL2.

[0034] The acquisition unit 111 of the computing unit 110 acquires the amount of data and the learning deadline for a job included in the job information 122. The selection unit 112 of the computing unit 110 may calculate the computing performance required to complete the machine learning corresponding to the job within the learning deadline based on the acquired amount of data and the acquired learning deadline.

[0035] The selection unit 112 may extract one or more data centers capable of meeting the calculated computing performance based on the computing resource information 121 (in other words, capable of completing machine learning for a single job within the learning deadline). The selection unit 112 selects a data center from the extracted one or more data centers to perform machine learning for a single job in order to minimize the cost required for machine learning. The selection unit 112 may also select one data center to perform machine learning. The selection unit 112 may also select multiple data centers to perform machine learning. If the selection unit 112 selects multiple data centers, the determination unit 113 of the computing device 110 determines the training data to be input to each of the multiple data centers based on the dataset for a single job.

[0036] The selection unit 112, via the communication device 130, causes the selected data center to perform machine learning corresponding to a single job. For example, the selection unit 112 may register a single job in the queue related to the selected data center. At this time, the information processing device 100 may send a dataset related to the single job from the database DB to the selected data center based on the job information 122. If the selection unit 112 selects multiple data centers, the information processing device 100 may send learning data related to the single job from the database DB to each of the multiple data centers based on the decision result by the decision unit 113.

[0037] Here, data center usage fees vary from data center to data center. On-premises data is relatively inexpensive, while public cloud data is relatively expensive. Hosted data is often more expensive than on-premises data but cheaper than public cloud data.

[0038] For example, if the extracted data servers include both on-premises and public cloud types, the selection unit 112 may select the on-premises type to minimize the cost of machine learning. For example, if the extracted data servers include both hosted and public cloud types, the selection unit 112 may select the hosted type to minimize the cost of machine learning. Furthermore, if the selection unit 112 selects multiple data centers to perform machine learning for a single job, the selection unit 112 may prioritize selecting the on-premises type to minimize the cost of machine learning.

[0039] When machine learning corresponding to a job is completed, the information processing device 100 may obtain result information indicating the result of the job from the data center. The information processing device 100 may store this result information in the storage device 120. The user of the information processing device 100 may have the information processing device 100 display the result information via the input device 140. In this case, the information processing device 100 may display the image 30 shown in Figure 6 on a display as an example of an output device 150.

[0040] (Examples of application) The trained model generated by machine learning using System 1 described above may be applied, for example, to the advanced driver assistance functions (Advanced Drive / Advanced Drive Assistance System) of a vehicle.

[0041] For example, a base model related to advanced driver assistance functions may be generated as a trained model by machine learning using a commercially available training dataset contained in the database DB and System 1. Furthermore, the above base model may be fine-tuned by machine learning using training data based on data collected from multiple vehicles driving in a specific region, contained in the database DB, and System 1. As a result, a training model related to advanced driver assistance functions optimized for a specific region may be generated. LoRA (Low-Rank Adaptation) may be used for fine-tuning.

[0042] (Technical effects) To provide a safer and more comfortable driving environment for vehicles, the use of AI is being considered. For example, a trained model (i.e., AI) related to advanced driver assistance functions can be executed on an in-vehicle system to provide assistance with the operation of peripheral devices such as air conditioners and audio systems, and to support safer driving. In addition to or instead of the in-vehicle system, the trained model related to advanced driver assistance functions can be executed on a server on the network, and more comprehensive services can be provided to the vehicle user via the communication device installed in the vehicle. A server providing such services needs to respond to user requests in real time. However, the data input to such a server is relatively small.

[0043] For example, to develop AI related to advanced driver assistance functions, the server that performs machine learning (corresponding to the servers included in the data center mentioned above) needs to process a large amount of data. However, if the predetermined development schedule is followed, real-time response is not required. In other words, if the predetermined development schedule is followed, the processing time does not need to be short. Thus, the performance requirements for the server that runs the trained model and the server that performs machine learning are different.

[0044] As mentioned above, usage fees often vary depending on the data center. In System 1 according to this embodiment, the information processing device 100 selects a data center such that machine learning is completed within the learning period and the cost of machine learning is minimized. In other words, in System 1, a data center with relatively low costs is preferentially selected as the data center for performing machine learning while adhering to a predetermined development schedule. Therefore, according to System 1 according to this embodiment, a data center can be selected in a way that suppresses the development costs of the learning model.

[0045] Furthermore, System 1 may include an on-premise data center that reliably meets the company's computing needs, and at least one of a hosted private cloud and / or public cloud that meets the remaining portion of the company's computing needs. The information processing device 100 may select a data center such that machine learning is completed within the learning period and the costs required for machine learning are minimized. With this configuration, the company can meet its computing needs while suppressing the development costs of the learning model.

[0046] (First variation) The computing resource information 121 may further include environmental impact information indicating the environmental burden related to the data center. For example, the environmental impact information may include indicators that show the environmental burden. For example, the environmental impact information may include information indicating the type of energy used by the data center. The type of energy may include, for example, green energy, renewable energy, fossil fuels, etc.

[0047] For example, the selection unit 112 of the information processing device 100 may select a data center for machine learning based on the computing resource information 121 such that the cost of machine learning is reduced and the environmental impact is minimized. With this configuration, it is possible to suppress both the development costs of the learning model and the environmental impact.

[0048] (Second variation) The computing resource information 121 may further include power information regarding the power situation in the region, including the data center. Information indicating the power situation may include, for example, the amount of power generated by solar power generation, the amount of power generated by wind power generation, the amount of energy stored in storage batteries, whether or not output curtailment is being implemented, etc.

[0049] For example, the selection unit 112 of the information processing device 100 may select a data center for machine learning based on the computing resource information 121 in such a way that the cost required for machine learning is reduced. For example, the selection unit 112 may prioritize selecting data centers in areas with relatively large power supply surplus capacity from among data centers with relatively low usage fees.

[0050] For example, the amount of electricity generated by solar and wind power is susceptible to weather conditions. If the amount of electricity generated by at least one of the solar and wind power sources exceeds the amount used, at least one of them will be temporarily shut down. In other words, there may be cases where solar and wind power cannot be fully utilized. Data centers consume a relatively large amount of electricity. By configuring the system as described above, it is possible to suppress the development costs of learning models while also preventing the temporary shutdown of at least one of the solar and wind power sources.

[0051] (Third variation) The computing resource information 121 may further include environmental load information indicating the environmental burden on the data center and power information regarding the power situation in the region including the data center. The selection unit 112 of the information processing device 100 may select a data center for machine learning based on the computing resource information 121 such that the cost required for machine learning is reduced and the environmental burden is reduced. For example, the selection unit 112 may prioritize selecting a data center in a region with relatively large power supply surplus among data centers that have relatively low usage fees and relatively small environmental burdens.

[0052] For example, the selection unit 112 may calculate a monetary score related to usage fees, an environmental score related to environmental impact, and a power score related to the power supply status. Here, the monetary score may be smaller the lower the usage fees. The environmental score may be smaller the lower the environmental impact. The power score may be smaller the larger the power supply surplus.

[0053] For example, the selection unit 112 may calculate the score for a data center as “w1 × (monetary score) + w2 × (environmental score) + w3 × (power score)”. Here, “w1”, “w2”, and “w3” are weights. Weight w1 is greater than weights w2 and w3. The relative magnitudes of weights w2 and w3 may be determined according to the user's policy. The selection unit 112 may also select the data center with the smallest score as the data center to perform machine learning on.

[0054] With this configuration, for example, it is possible to reduce the development costs of the learning model, reduce the environmental impact, and furthermore, prevent the temporary shutdown of at least one of the solar and wind power generation systems.

[0055] Aspects of the invention derived from the embodiments and modifications described above are described below.

[0056] A system according to one aspect of the invention is a system for controlling machine learning of a model on a computing infrastructure comprising a plurality of nodes configured to communicate via a network, comprising: acquisition means for acquiring learning data information relating to learning data and time information indicating a learning deadline; node information relating to the plurality of nodes; and selection means for selecting one or more nodes from the plurality of nodes to be used for machine learning of the model, based on the learning data information and the time information, wherein the node information includes cost information relating to usage costs, and the selection means selects one or more nodes such that machine learning is completed within the learning deadline indicated by the time information and the cost required for machine learning is small.

[0057] In the above-described embodiment, "data centers DC1, DC2, and DC3, and clouds CL1 and CL2" correspond to an example of a "node," "acquisition unit 111" corresponds to an example of an "acquisition means," "selection unit 112" corresponds to an example of a "selection means," and "computational resource information 121" corresponds to an example of "node information."

[0058] In one example of the system, the plurality of nodes may include at least two of the following: on-premises, private cloud, and public cloud.

[0059] In other examples of the system, the node information may include performance information relating to the computing performance of each of the multiple nodes, and the selection means may select one or more nodes such that the end of the machine learning process approaches the learning period indicated by the time information, and the cost required for machine learning is reduced.

[0060] In other examples of the system, when two or more nodes are selected as the one or more nodes, the system may include a determination means that determines the learning data to be input to each of the two or more nodes based on the learning data information. In the above embodiment, the "determination unit 113" corresponds to an example of the "determination means".

[0061] In other examples of the system, the node information may include environmental load information relating to the environmental load of each of the multiple nodes, and the selection means may select one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information, the cost required for machine learning is small, and the environmental load is small.

[0062] In other examples of the system, the node information may include power information relating to the local power situation, and the selection means may select one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information and the cost required for machine learning is small.

[0063] In other examples of the system, the node information may include environmental load information relating to the environmental load of each of the multiple nodes and power information relating to the power situation in the region, and the selection means may select one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information, the cost required for machine learning is small, and the environmental load is small.

[0064] In other examples of the system, an output means may be provided to output a report on machine learning when the machine learning of the model is completed, and the report may include information on costs. In the above embodiment, "output device 150" corresponds to an example of "output means".

[0065] In other examples of the system, output means may be provided to output an input screen that includes an input field for entering the learning deadline.

[0066] A control method according to one aspect of the present invention is a control method for controlling machine learning of a model on a computing infrastructure comprising a plurality of nodes configured to communicate via a network, comprising: an acquisition step of acquiring learning data information relating to learning data and time information indicating a learning deadline; and a selection step of selecting one or more nodes from the plurality of nodes to be used for machine learning of the model, based on node information relating to the plurality of nodes, the learning data information, and the time information, wherein the node information includes cost information relating to usage costs, and in the selection step, the one or more nodes are selected such that machine learning is completed within the learning deadline indicated by the time information and the cost required for machine learning is minimized.

[0067] The present invention is not limited to the embodiments described above, and can be modified as appropriate without contradicting the gist or idea of ​​the invention as can be read from the claims and specification as a whole. Systems involving such modifications are also included within the technical scope of the present invention. [Explanation of Symbols]

[0068] 1...System, 100...Information Processing Device, 111...Acquisition Unit, 112...Selection Unit, 113...Decision Unit, DB...Database, DC1, DC2, DC3...Data Center, CL1, CL2...Cloud, NW...Network

Claims

1. A system for controlling machine learning models in a computing infrastructure comprising multiple nodes configured to communicate via a network, A means for acquiring learning data information related to the learning data and time information indicating the learning deadline, A selection means for selecting one or more nodes from the plurality of nodes to be used for machine learning of the model, based on node information relating to the plurality of nodes, training data information, and time information. Equipped with, The node information includes cost information regarding usage fees, The selection means selects one or more nodes such that machine learning is completed within the learning period indicated by the time information and the cost required for machine learning is minimized. system.

2. The aforementioned plurality of nodes include at least two of the following: on-premises, private cloud, and public cloud. The system according to claim 1.

3. The node information includes performance information regarding the computing performance of each of the multiple nodes, The selection means selects one or more nodes such that the end of the machine learning process approaches the learning period indicated by the time information, and the cost required for machine learning is reduced. The system according to claim 1.

4. When two or more nodes are selected as the one or more nodes, the system includes a determination means for determining the learning data to be input to each of the two or more nodes based on the learning data information. The system according to claim 1.

5. The node information includes environmental load information relating to the environmental load of each of the multiple nodes, The selection means selects one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information, the cost required for machine learning is low, and the environmental impact is low. The system according to claim 1.

6. The node information includes power information regarding the local power situation, The selection means selects one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information and the cost required for machine learning is minimized. The system according to claim 1.

7. The node information includes environmental load information relating to the environmental load of each of the multiple nodes, and power information relating to the power situation in the region. The selection means selects one or more nodes based on the node information such that machine learning is completed within the learning period indicated by the time information, the cost required for machine learning is low, and the environmental impact is low. The system according to claim 1.

8. The system includes an output means for outputting a report on machine learning when the machine learning of the aforementioned model is completed. The aforementioned report includes information on costs. The system according to claim 1.

9. The system includes an output means that outputs an input screen which includes an input field for entering the learning deadline. The system according to claim 1.

Images