KR102549144B1 - Method to reconfigure a machine learning cluster without interruption - Google Patents

Abstract

In the present specification, in a method in which a first node joins a cluster for distributed learning, a learning model for the distributed learning is received from a second node included in the cluster, and a background thread participate in the cluster for the distributed learning, the background thread generates a tensor related to the gradient value for the distributed learning, and through the tensor, an Allreduce operation for the distributed learning and apply the first gradient values obtained through the all-reduce operation to the learning model.

Description

How to reconstruct a machine learning cluster without disruption {METHOD TO RECONFIGURE A MACHINE LEARNING CLUSTER WITHOUT INTERRUPTION}

The present specification relates to a method and apparatus for reconfiguring a machine learning cluster without interruption for efficiency of cloud resources in a distributed learning environment.

For synchronous distributed learning in machine learning, all computing nodes can learn with the same learning model, and at the end of each iteration, the model can be updated by collecting gradient values and obtaining an average value.

Due to the nature of this learning progress, existing distributed deep learning frameworks do not provide a function to add computing nodes without stopping because it is difficult to synchronize the learning model without stopping learning.

An object of the present specification is to provide a method for adding a computing node without interrupting the learning process during distributed learning in a distributed learning environment.

In addition, an object of the present specification is to provide a method for synchronizing a learning model of a newly added computing node with a model currently in progress in a distributed learning environment,

The technical problems to be achieved by this specification are not limited to the above-mentioned technical problems, and other technical problems not mentioned are clear to those skilled in the art from the detailed description of the invention below. will be understandable.

One aspect of the present specification is a method for a first node to join a cluster for distributed learning, comprising: receiving a learning model for the distributed learning from a second node included in the cluster; Involving a background thread in the cluster for distributed learning, wherein the background thread creates a tensor related to a gradient value for distributed learning; performing an allreduce operation for the distributed learning through the tensor; and applying the obtained first gradient values to the learning model through the all-reduce operation. can include

In addition, the second node may stop the distributed learning in order to deliver the learning model to the first node when the first node requests to join the cluster.

In addition, the tensor may have the same type as the gradient value for the distributed learning and be initialized to 0.

In addition, the step of applying the first gradient values obtained through the all-reduce operation to the learning model may include storing the obtained first gradient values in a memory; and sequentially applying the first gradient values stored in the memory to the learning model through a main thread; may further include.

In addition, receiving a second gradient value related to the next repeated distributed learning from the node of the cluster; applying the second gradient value to the learning model; and joining the cluster; may further include.

Another aspect of the present specification is an apparatus for implementing a first node joining a cluster for distributed learning, comprising: a transceiver for transmitting and receiving a signal; Memory; and an AI processor for functionally controlling the transceiver and the memory. Including, the AI processor receives the learning model for the distributed learning from the second node included in the cluster through the transceiver, participates a background thread in the cluster for the distributed learning, , The background thread generates a tensor related to the gradient value for the distributed learning, performs an Allreduce operation for the distributed learning through the tensor, and performs the Allreduce operation Through, it is possible to apply the obtained first gradient values to the learning model.

According to the embodiments of the present specification, in a distributed learning environment, it is possible to provide a method for adding or removing computing nodes without interrupting the learning process during distributed learning.

Also, an object of the present specification is to provide a method of synchronizing a learning model of a newly added computing node with a model currently being trained in a distributed learning environment.

Effects obtainable in the present specification are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below. .

1 is a block diagram for explaining an electronic device related to the present specification.
2 is a block diagram of an AI device according to an embodiment of the present specification.
3 is an example of a distributed learning model to which the present specification can be applied.
4 is an example of a cluster reconstruction method to which the present specification can be applied.
5 is an embodiment to which the present specification may be applied.
6 is an example of a device in general to which the present disclosure may be applied.
The accompanying drawings, which are included as part of the detailed description to aid understanding of the present specification, provide examples of the present specification and describe technical features of the present specification together with the detailed description.

Hereinafter, the embodiments disclosed in this specification will be described in detail with reference to the accompanying drawings, but the same or similar components are given the same reference numerals regardless of reference numerals, and redundant description thereof will be omitted. The suffixes "module" and "unit" for components used in the following description are given or used together in consideration of ease of writing the specification, and do not have meanings or roles that are distinct from each other by themselves. In addition, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of a related known technology may obscure the gist of the embodiment disclosed in this specification, the detailed description thereof will be omitted. In addition, the accompanying drawings are only for easy understanding of the embodiments disclosed in this specification, the technical idea disclosed in this specification is not limited by the accompanying drawings, and all changes included in the spirit and technical scope of this specification , it should be understood to include equivalents or substitutes.

Terms including ordinal numbers, such as first and second, may be used to describe various components, but the components are not limited by the terms. These terms are only used for the purpose of distinguishing one component from another.

It is understood that when an element is referred to as being "connected" or "connected" to another element, it may be directly connected or connected to the other element, but other elements may exist in the middle. It should be. On the other hand, when an element is referred to as “directly connected” or “directly connected” to another element, it should be understood that no other element exists in the middle.

Singular expressions include plural expressions unless the context clearly dictates otherwise.

In this application, terms such as "comprise" or "have" are intended to designate that there is a feature, number, step, operation, component, part, or combination thereof described in the specification, but one or more other features It should be understood that the presence or addition of numbers, steps, operations, components, parts, or combinations thereof is not precluded.

1 is a block diagram for explaining an electronic device related to the present specification.

Referring to FIG. 1 , the electronic device 100 includes a wireless communication unit 110, an input unit 120, a sensing unit 140, an output unit 150, an interface unit 160, a memory 170, and a control unit 180. ) and a power supply unit 190. The components shown in FIG. 1 are not essential to implement an electronic device, so an electronic device described in this specification may have more or fewer components than those listed above.

More specifically, among the components, the wireless communication unit 110 is between the electronic device 100 and the wireless communication system, between the electronic device 100 and other electronic devices 100, or between the electronic device 100 and an external server. It may include one or more modules enabling wireless communication between Also, the wireless communication unit 110 may include one or more modules that connect the electronic device 100 to one or more networks.

The wireless communication unit 110 may include at least one of a broadcast reception module 111, a mobile communication module 112, a wireless Internet module 113, a short-distance communication module 114, and a location information module 115. .

The input unit 120 includes a camera 121 or video input unit for inputting a video signal, a microphone 122 for inputting an audio signal, or a user input unit 123 for receiving information from a user, for example , a touch key, a push key (mechanical key, etc.). Voice data or image data collected by the input unit 120 may be analyzed and processed as a user's control command.

The sensing unit 140 may include one or more sensors for sensing at least one of information within the electronic device, environmental information surrounding the electronic device, and user information. For example, the sensing unit 140 may include a proximity sensor 141, an illumination sensor 142, a touch sensor, an acceleration sensor, a magnetic sensor, and gravity. Sensor (G-sensor), gyroscope sensor (gyroscope sensor), motion sensor (motion sensor), RGB sensor, infrared sensor (IR sensor), finger scan sensor, ultrasonic sensor , an optical sensor (eg, a camera (see 121)), a microphone (see 122), a battery gauge, an environmental sensor (eg, a barometer, a hygrometer, a thermometer, a radiation detection sensor, It may include at least one of a heat detection sensor, a gas detection sensor, etc.), a chemical sensor (eg, an electronic nose, a healthcare sensor, a biometric sensor, etc.). Meanwhile, the electronic device disclosed in this specification may combine and utilize information sensed by at least two or more of these sensors.

The output unit 150 is for generating an output related to sight, hearing, or touch, and includes at least one of a display unit 151, a sound output unit 152, a haptic module 153, and an optical output unit 154. can do. The display unit 151 may implement a touch screen by forming a mutual layer structure or integrally with the touch sensor. Such a touch screen may function as a user input unit 123 providing an input interface between the electronic device 100 and the user and provide an output interface between the electronic device 100 and the user.

The interface unit 160 serves as a passage for various types of external devices connected to the electronic device 100 . The interface unit 160 connects a device equipped with a wired/wireless headset port, an external charger port, a wired/wireless data port, a memory card port, and an identification module. It may include at least one of a port, an audio input/output (I/O) port, a video input/output (I/O) port, and an earphone port. In response to the external device being connected to the interface unit 160, the electronic device 100 may perform appropriate control related to the connected external device.

Also, the memory 170 stores data supporting various functions of the electronic device 100 . The memory 170 may store a plurality of application programs (application programs or applications) running in the electronic device 100 , data for operating the electronic device 100 , and commands. At least some of these application programs may be downloaded from an external server through wireless communication. In addition, at least some of these application programs may exist on the electronic device 100 from the time of shipment for basic functions of the electronic device 100 (eg, incoming and outgoing calls, outgoing functions, message receiving, and outgoing functions). Meanwhile, the application program may be stored in the memory 170, installed on the electronic device 100, and driven by the control unit 180 to perform an operation (or function) of the electronic device.

The controller 180 controls general operations of the electronic device 100 in addition to operations related to the application program. The control unit 180 may provide or process appropriate information or functions to the user by processing signals, data, information, etc. input or output through the components described above or by running an application program stored in the memory 170.

In addition, the controller 180 may control at least some of the components discussed in conjunction with FIG. 1 in order to drive an application program stored in the memory 170 . Furthermore, the controller 180 may combine and operate at least two or more of the components included in the electronic device 100 to drive the application program.

The power supply unit 190 receives external power and internal power under the control of the controller 180 and supplies power to each component included in the electronic device 100 . The power supply unit 190 includes a battery, and the battery may be a built-in battery or a replaceable battery.

At least some of the components may operate in cooperation with each other in order to implement an operation, control, or control method of an electronic device according to various embodiments described below. Also, the operation, control, or control method of the electronic device may be implemented on the electronic device by driving at least one application program stored in the memory 170 .

Hereinafter, prior to examining various embodiments implemented through the electronic device 100 discussed above, the components listed above will be examined in more detail.

First, looking at the wireless communication unit 110, the broadcast reception module 111 of the wireless communication unit 110 receives a broadcast signal and/or broadcast-related information from an external broadcast management server through a broadcast channel. The broadcast channel may include a satellite channel and a terrestrial channel. Two or more broadcast receiving modules may be provided in the electronic device 100 to receive simultaneous broadcasting of at least two broadcast channels or to switch broadcast channels.

The mobile communication module 112 complies with technical standards or communication methods for mobile communication (eg, Global System for Mobile communication (GSM), Code Division Multi Access (CDMA), Code Division Multi Access 2000 (CDMA2000), EV) -DO(Enhanced Voice-Data Optimized or Enhanced Voice-Data Only), WCDMA(Wideband CDMA), HSDPA(High Speed Downlink Packet Access), HSUPA(High Speed Uplink Packet Access), LTE(Long Term Evolution), LTE-A (Long Term Evolution-Advanced), etc.) to transmit and receive radio signals with at least one of a base station, an external terminal, and a server on a mobile communication network.

The wireless signal may include a voice call signal, a video call signal, or various types of data according to text/multimedia message transmission/reception.

The wireless Internet module 113 refers to a module for wireless Internet access, and may be built into or external to the electronic device 100 . The wireless Internet module 113 is configured to transmit and receive radio signals in a communication network according to wireless Internet technologies.

Wireless Internet technologies include, for example, WLAN (Wireless LAN), Wi-Fi (Wireless-Fidelity), Wi-Fi (Wireless Fidelity) Direct, DLNA (Digital Living Network Alliance), WiBro (Wireless Broadband), WiMAX (World Interoperability for Microwave Access), HSDPA (High Speed Downlink Packet Access), HSUPA (High Speed Uplink Packet Access), LTE (Long Term Evolution), LTE-A (Long Term Evolution-Advanced), etc., and the wireless Internet module ( 113) transmits and receives data according to at least one wireless Internet technology within a range including Internet technologies not listed above.

From the viewpoint that wireless Internet access by WiBro, HSDPA, HSUPA, GSM, CDMA, WCDMA, LTE, LTE-A, etc. is made through a mobile communication network, the wireless Internet module (113) performing wireless Internet access through the mobile communication network ) may be understood as a kind of the mobile communication module 112.

The short range communication module 114 is for short range communication, and includes Bluetooth??, Radio Frequency Identification (RFID), Infrared Data Association (IrDA), Ultra Wideband (UWB), ZigBee, Near field communication (NFC), wireless-fidelity (Wi-Fi), Wi-Fi Direct, wireless universal serial bus (USB), and magnetic secure transmission (MST) technologies may be used to support short-range communication. . The short-range communication module 114 may be used between the electronic device 100 and a wireless communication system, between the electronic device 100 and other electronic devices 100, or between the electronic device 100 through wireless area networks. ) and a network where another electronic device 100 (or an external server) is located. The local area wireless communication network may be a local area wireless personal area network (Wireless Personal Area Networks).

Here, the other electronic device 100 is a wearable device capable of exchanging (or interlocking) data with the electronic device 100 according to the present specification (for example, a smart watch), smart glasses (smart glass) or HMD (head mounted display). The short-range communication module 114 may detect (or recognize) a wearable device capable of communicating with the electronic device 100 around the electronic device 100 . Furthermore, if the detected wearable device is a device authorized to communicate with the electronic device 100 according to the present specification, the control unit 180 transmits at least a portion of data processed by the electronic device 100 to the short-range communication module ( 114) can be transmitted to the wearable device. Accordingly, the user of the wearable device may use data processed by the electronic device 100 through the wearable device. For example, according to this, when a call is received in the electronic device 100, the user makes a phone call through the wearable device, or when a message is received in the electronic device 100, the user receives the received message through the wearable device. It is possible to check the message.

The location information module 115 is a module for obtaining the location (or current location) of the electronic device, and representative examples thereof include a Global Positioning System (GPS) module or a Wireless Fidelity (WiFi) module. For example, when an electronic device utilizes a GPS module, the location of the electronic device may be obtained using a signal transmitted from a GPS satellite. As another example, when an electronic device utilizes a Wi-Fi module, the location of the electronic device may be obtained based on information about the Wi-Fi module and a wireless access point (AP) that transmits or receives a wireless signal. If necessary, the location information module 115 may perform any function among other modules of the wireless communication unit 110 in order to obtain data on the location of the electronic device in substitution or addition. The location information module 115 is a module used to obtain the location (or current location) of the electronic device, and is not limited to a module that directly calculates or acquires the location of the electronic device.

Next, the input unit 120 is for inputting image information (or signals), audio information (or signals), data, or information input from a user. For inputting image information, the electronic device 100 has one Alternatively, a plurality of cameras 121 may be provided. The camera 121 processes an image frame such as a still image or a moving image obtained by an image sensor in a video call mode or a photographing mode. The processed image frame may be displayed on the display unit 151 or stored in the memory 170 . Meanwhile, the plurality of cameras 121 provided in the electronic device 100 may be arranged to form a matrix structure, and through the cameras 121 forming the matrix structure, various angles or focal points may be provided to the electronic device 100. A plurality of image information having may be input. In addition, the plurality of cameras 121 may be arranged in a stereo structure to obtain left and right images for realizing a stereoscopic image.

The microphone 122 processes external sound signals into electrical voice data. The processed voice data may be utilized in various ways according to the function (or application program being executed) being performed in the electronic device 100 . Meanwhile, various noise cancellation algorithms for removing noise generated in the process of receiving an external sound signal may be implemented in the microphone 122 .

The user input unit 123 is for receiving information from a user, and when information is input through the user input unit 123, the control unit 180 can control the operation of the electronic device 100 to correspond to the input information. . The user input unit 123 is a mechanical input means (or a mechanical key, for example, a button located on the front, rear or side of the electronic device 100, a dome switch, a jog wheel, jog switch, etc.) and a touch input means. As an example, the touch input means consists of a virtual key, soft key, or visual key displayed on a touch screen through software processing, or a part other than the touch screen. It can be made of a touch key (touch key) disposed on. On the other hand, the virtual key or visual key can be displayed on the touch screen while having various forms, for example, graphic (graphic), text (text), icon (icon), video (video) or these can be made of a combination of

Meanwhile, the sensing unit 140 senses at least one of information in the electronic device, surrounding environment information surrounding the electronic device, and user information, and generates a sensing signal corresponding thereto. Based on these sensing signals, the controller 180 may control driving or operation of the electronic device 100 or perform data processing, functions, or operations related to an application program installed in the electronic device 100 . Representative sensors among various sensors that may be included in the sensing unit 140 will be described in more detail.

First, the proximity sensor 141 refers to a sensor that detects the presence or absence of an object approaching a predetermined detection surface or an object existing nearby without mechanical contact by using the force of an electromagnetic field or infrared rays. The proximity sensor 141 may be disposed in an inner area of the electronic device covered by the touch screen as described above or in the vicinity of the touch screen.

Examples of the proximity sensor 141 include a transmission type photoelectric sensor, a direct reflection type photoelectric sensor, a mirror reflection type photoelectric sensor, a high frequency oscillation type proximity sensor, a capacitance type proximity sensor, a magnetic type proximity sensor, an infrared proximity sensor, and the like. If the touch screen is a capacitive type, the proximity sensor 141 may be configured to detect the proximity of a conductive object as a change in electric field according to the proximity of the object. In this case, the touch screen (or touch sensor) itself may be classified as a proximity sensor.

On the other hand, for convenience of description, an action of approaching a touch screen without contacting an object to recognize that the object is located on the touch screen is referred to as a "proximity touch", and the touch An act of actually touching an object on a screen is called a "contact touch". The location at which an object is proximity-touched on the touch screen means a location at which the object corresponds vertically to the touch screen when the object is proximity-touched. The proximity sensor 141 may detect a proximity touch and a proximity touch pattern (eg, proximity touch distance, proximity touch direction, proximity touch speed, proximity touch time, proximity touch position, proximity touch movement state, etc.) there is. Meanwhile, as described above, the controller 180 processes data (or information) corresponding to the proximity touch operation and proximity touch pattern detected through the proximity sensor 141, and furthermore, visual information corresponding to the processed data. It can be output on the touch screen. Furthermore, the controller 180 may control the electronic device 100 to process different operations or data (or information) depending on whether a touch to the same point on the touch screen is a proximity touch or a contact touch. .

The touch sensor detects a touch (or touch input) applied to the touch screen (or the display unit 151) using at least one of various touch methods such as a resistive film method, a capacitive method, an infrared method, an ultrasonic method, and a magnetic field method. detect

As an example, the touch sensor may be configured to convert a change in pressure applied to a specific part of the touch screen or capacitance generated in a specific part into an electrical input signal. The touch sensor may be configured to detect a location, area, pressure upon touch, capacitance upon touch, and the like, at which a touch object that applies a touch to the touch screen is touched on the touch sensor. Here, the touch object is an object that applies a touch to the touch sensor, and may be, for example, a finger, a touch pen, a stylus pen, or a pointer.

As such, when there is a touch input to the touch sensor, the corresponding signal(s) is sent to the touch controller. The touch controller processes the signal(s) and then transmits corresponding data to the controller 180. Thus, the controller 180 can know which area of the display unit 151 has been touched. Here, the touch controller may be a component separate from the controller 180 or may be the controller 180 itself.

Meanwhile, the controller 180 may perform different controls or the same control according to the type of the touch object that touches the touch screen (or a touch key provided other than the touch screen). Whether to perform different controls or the same control according to the type of the touch object may be determined according to an operating state of the electronic device 100 or an application program being executed.

On the other hand, the touch sensor and proximity sensor described above are independently or combined, short (or tap) touch, long touch, multi-touch, drag touch on the touch screen ), flick touch, pinch-in touch, pinch-out touch, swipe touch, hovering touch, etc. Touch can be sensed.

The ultrasonic sensor may recognize location information of a sensing object by using ultrasonic waves. Meanwhile, the controller 180 may calculate the location of the wave generating source through information detected by the optical sensor and the plurality of ultrasonic sensors. The position of the wave generating source can be calculated using the property that light is much faster than ultrasonic waves, that is, the time for light to reach the optical sensor is much faster than the time for ultrasonic waves to reach the ultrasonic sensor. More specifically, the location of the wave generating source may be calculated by using light as a reference signal and a time difference between the arrival time of ultrasonic waves.

Meanwhile, looking at the configuration of the input unit 120, the camera 121 includes at least one of a camera sensor (eg, CCD, CMOS, etc.), a photo sensor (or image sensor), and a laser sensor.

The camera 121 and the laser sensor may be combined with each other to detect a touch of a sensing target for a 3D stereoscopic image. A photo sensor may be stacked on a display device, and the photo sensor is configured to scan a motion of a sensing target close to the touch screen. More specifically, the photo sensor scans the content placed on the photo sensor by mounting photo diodes and transistors (TRs) in rows/columns and using electrical signals that change according to the amount of light applied to the photo diodes. That is, the photo sensor calculates the coordinates of the sensing object according to the change in light, and through this, the location information of the sensing object can be obtained.

The display unit 151 displays (outputs) information processed by the electronic device 100 . For example, the display unit 151 may display execution screen information of an application program driven in the electronic device 100 or UI (User Interface) and GUI (Graphic User Interface) information according to such execution screen information. .

Also, the display unit 151 may be configured as a 3D display unit that displays a 3D image.

A 3D display method such as a stereoscopic method (glasses method), an auto stereoscopic method (non-glasses method), or a projection method (holographic method) may be applied to the 3D display unit.

The audio output unit 152 may output audio data received from the wireless communication unit 110 or stored in the memory 170 in a call signal reception, a call mode or a recording mode, a voice recognition mode, a broadcast reception mode, and the like. The sound output unit 152 also outputs sound signals related to functions performed by the electronic device 100 (eg, call signal reception sound, message reception sound, etc.). The sound output unit 152 may include a receiver, a speaker, a buzzer, and the like.

The haptic module 153 generates various tactile effects that a user can feel. A representative example of the tactile effect generated by the haptic module 153 may be vibration. The strength and pattern of the vibration generated by the haptic module 153 may be controlled by a user's selection or a setting of a controller. For example, the haptic module 153 may synthesize and output different vibrations or sequentially output them.

In addition to vibration, the haptic module 153 responds to stimuli such as an array of pins that move vertically with respect to the contact skin surface, air blowing force or suction force through a nozzle or suction port, rubbing against the skin surface, electrode contact, and electrostatic force. It is possible to generate various tactile effects, such as the effect of heat absorption and the effect of reproducing the feeling of coolness and warmth using an element capable of absorbing heat or generating heat.

The haptic module 153 not only transmits a tactile effect through direct contact, but also can be implemented so that a user can feel a tactile effect through a muscle sense such as a finger or an arm. Two or more haptic modules 153 may be provided according to the configuration of the electronic device 100 .

The light output unit 154 outputs a signal for notifying occurrence of an event using light from a light source of the electronic device 100 . Examples of events occurring in the electronic device 100 may include message reception, call signal reception, missed calls, alarms, schedule notifications, e-mail reception, and information reception through applications.

A signal output from the light output unit 154 is implemented as the electronic device emits light of a single color or multiple colors to the front or back side. The signal output may be terminated when the electronic device detects the user's confirmation of the event.

The interface unit 160 serves as a passage for all external devices connected to the electronic device 100 . The interface unit 160 receives data from an external device or receives power and transmits it to each component inside the electronic device 100, or transmits data inside the electronic device 100 to an external device. For example, a wired/wireless headset port, an external charger port, a wired/wireless data port, a memory card port, and a port for connecting a device having an identification module. The interface unit 160 may include an audio I/O (Input/Output) port, a video I/O (Input/Output) port, an earphone port, and the like.

On the other hand, the identification module is a chip that stores various types of information for authenticating the right to use the electronic device 100, and includes a user identification module (UIM), a subscriber identity module (SIM), and universal user authentication. module (universal subscriber identity module; USIM) and the like. A device equipped with an identification module (hereinafter referred to as 'identification device') may be manufactured in the form of a smart card. Accordingly, the identification device may be connected to the terminal 100 through the interface unit 160 .

In addition, the interface unit 160 becomes a passage through which power from the cradle is supplied to the electronic device 100 when the electronic device 100 is connected to an external cradle, or a user inputs power from the cradle. It may be a passage through which various command signals are transmitted to the electronic device 100 . Various command signals or the power input from the cradle may operate as a signal for recognizing that the electronic device 100 is correctly mounted on the cradle.

The memory 170 may store programs for operation of the controller 180 and may temporarily store input/output data (eg, phonebook, message, still image, video, etc.). The memory 170 may store data related to vibration and sound of various patterns output when a touch is input on the touch screen.

The memory 170 may be a flash memory type, a hard disk type, a solid state disk type, a silicon disk drive type, or a multimedia card micro type. ), card-type memory (eg SD or XD memory, etc.), RAM (random access memory; RAM), SRAM (static random access memory), ROM (read-only memory; ROM), EEPROM (electrically erasable programmable read -only memory), a programmable read-only memory (PROM), a magnetic memory, a magnetic disk, and an optical disk. The electronic device 100 may operate in relation to a web storage that performs a storage function of the memory 170 on the Internet.

Meanwhile, as described above, the controller 180 controls operations related to application programs and general operations of the electronic device 100 in general. For example, if the state of the electronic device satisfies a set condition, the controller 180 may execute or release a locked state that restricts a user's control command input to applications.

In addition, the controller 180 may perform control and processing related to voice calls, data communications, video calls, etc., or perform pattern recognition processing capable of recognizing handwriting input or drawing input performed on the touch screen as characters and images, respectively. can Furthermore, the control unit 180 may control any one or a combination of the components described above in order to implement various embodiments described below on the electronic device 100 according to the present specification.

The power supply unit 190 receives external power and internal power under the control of the controller 180 and supplies power necessary for the operation of each component. The power supply unit 190 includes a battery, and the battery may be a built-in battery capable of being charged, or may be detachably coupled to the terminal body for charging.

In addition, the power supply unit 190 may have a connection port, and the connection port may be configured as an example of an interface 160 electrically connected to an external charger that supplies power to charge the battery.

As another example, the power supply unit 190 may charge the battery in a wireless manner without using the connection port. In this case, the power supply unit 190 uses at least one of an inductive coupling method based on magnetic induction from an external wireless power transmitter or a magnetic resonance coupling method based on electromagnetic resonance. power can be delivered. In this specification, the electronic device 100 may include a terminal and a server.

2 is a block diagram of an AI device according to an embodiment of the present specification.

The AI device 20 may include an electronic device including an AI module capable of performing AI processing or a server including the AI module. In addition, the AI device 20 may be included in at least a portion of the electronic device 100 shown in FIG. 1 to perform at least a portion of AI processing together.

The AI device 20 may include an AI processor 21, a memory 25 and/or a communication unit 27.

The AI device 20 is a computing device capable of learning a neural network, and may be implemented in various electronic devices such as a server, a desktop PC, a notebook PC, and a tablet PC.

The AI processor 21 may learn a neural network using a program stored in the memory 25 . In particular, the AI processor 21 may learn a neural network for recognizing vehicle-related data. Here, the neural network for recognizing vehicle-related data may be designed to simulate the structure of a human brain on a computer, and may include a plurality of network nodes having weights that simulate neurons of the human neural network. A plurality of network modes may transmit and receive data according to a connection relationship, respectively, so as to simulate synaptic activity of neurons that transmit and receive signals through synapses. Here, the neural network may include a deep learning model developed from a neural network model. In the deep learning model, a plurality of network nodes may exchange data according to a convolution connection relationship while being located in different layers. Examples of neural network models are deep neural networks (DNN), convolutional deep neural networks (CNN), recurrent Boltzmann machines (RNNs), restricted Boltzmann machines (RBMs), deep trust It includes various deep learning techniques such as deep belief networks (DBN) and deep Q-network, and can be applied to fields such as computer vision, voice recognition, natural language processing, and voice/signal processing.

Meanwhile, the processor performing the functions described above may be a general-purpose processor (eg, CPU), or may be an AI-only processor (eg, GPU) for artificial intelligence learning.

The memory 25 may store various programs and data necessary for the operation of the AI device 20 . The memory 25 may be implemented as a non-volatile memory, a volatile memory, a flash-memory, a hard disk drive (HDD), or a solid state drive (SDD). The memory 25 is accessed by the AI processor 21, and reading/writing/modifying/deleting/updating of data by the AI processor 21 can be performed. In addition, the memory 25 may store a neural network model (eg, the deep learning model 26) generated through a learning algorithm for data classification/recognition according to an embodiment of the present specification.

Meanwhile, the AI processor 21 may include a data learning unit 22 that learns a neural network for data classification/recognition. The data learning unit 22 may learn criteria regarding which training data to use to determine data classification/recognition and how to classify and recognize data using the training data. The data learning unit 22 may acquire learning data to be used for learning and learn the deep learning model by applying the obtained learning data to the deep learning model.

The data learning unit 22 may be manufactured in the form of at least one hardware chip and mounted on the AI device 20 . For example, the data learning unit 22 may be manufactured in the form of a dedicated hardware chip for artificial intelligence (AI), or manufactured as a part of a general-purpose processor (CPU) or a graphics-only processor (GPU) for the AI device 20. may be mounted. Also, the data learning unit 22 may be implemented as a software module. When implemented as a software module (or a program module including instructions), the software module may be stored in a computer-readable, non-transitory computer readable recording medium (non-transitory computer readable media). In this case, at least one software module may be provided by an Operating System (OS) or an application.

The data learning unit 22 may include a training data acquisition unit 23 and a model learning unit 24 .

The training data acquisition unit 23 may acquire training data required for a neural network model for classifying and recognizing data. For example, the learning data acquisition unit 23 may acquire vehicle data and/or sample data to be input to a neural network model as learning data.

The model learning unit 24 may learn to have a criterion for determining how to classify predetermined data by using the obtained training data. In this case, the model learning unit 24 may learn the neural network model through supervised learning using at least some of the learning data as a criterion. Alternatively, the model learning unit 24 may learn the neural network model through unsupervised learning in which a decision criterion is discovered by self-learning using learning data without guidance. In addition, the model learning unit 24 may learn the neural network model through reinforcement learning using feedback about whether the result of the situation judgment according to learning is correct. In addition, the model learning unit 24 may train the neural network model using a learning algorithm including error back-propagation or gradient decent.

When the neural network model is learned, the model learning unit 24 may store the learned neural network model in memory. The model learning unit 24 may store the learned neural network model in a memory of a server connected to the AI device 20 through a wired or wireless network.

The data learning unit 22 further includes a training data pre-processing unit (not shown) and a learning data selection unit (not shown) to improve the analysis result of the recognition model or save resources or time required for generating the recognition model. You may.

The learning data pre-processing unit may pre-process the acquired data so that the acquired data can be used for learning for situation determination. For example, the learning data pre-processing unit may process the acquired data into a preset format so that the model learning unit 24 can use the acquired learning data for learning for image recognition.

In addition, the learning data selector may select data necessary for learning from among the learning data acquired by the learning data acquisition unit 23 or the learning data preprocessed by the preprocessor. The selected learning data may be provided to the model learning unit 24 . For example, the learning data selection unit may select only data about an object included in the specific region as training data by detecting a specific region among images acquired through a vehicle camera.

In addition, the data learning unit 22 may further include a model evaluation unit (not shown) to improve the analysis result of the neural network model.

The model evaluation unit inputs evaluation data to the neural network model, and when an analysis result output from the evaluation data does not satisfy a predetermined criterion, it may cause the model learning unit 22 to learn again. In this case, the evaluation data may be predefined data for evaluating the recognition model. For example, the model evaluator may evaluate that the predetermined criterion is not satisfied when the number or ratio of the evaluation data for which the analysis result is inaccurate among the analysis results of the learned recognition model for the evaluation data exceeds a preset threshold. there is.

The communication unit 27 may transmit the AI processing result by the AI processor 21 to an external electronic device.

Here, the external electronic device may be defined as a terminal or a client. Also, the AI device 20 may be implemented through a server or network.

On the other hand, the AI device 20 shown in FIG. 2 has been functionally divided into an AI processor 21, a memory 25, a communication unit 27, etc., but the above-mentioned components are integrated into one module and the AI module Note that it can also be called

3 is an example of a distributed learning model to which the present specification can be applied.

Referring to FIG. 3, nodes for distributed learning form a ring structure and are connected through a framework such as Tensorflow or Pytorch to perform distributed learning. In order to collect gradient values of nodes connected in such a ring structure, a Ring Allreduce method may be used. Ring Allreduce is an operation that reduces the target array of all processors to a single array and returns the resulting array to all processes.

For example, the collected gradient value can be divided by the number of nodes. A set of gradient values divided in this way may be referred to as a chunk. Each node can pass one chunk to the next node in parallel, and the delivered chunk can be merged with each node's chunk.

That is, each node can add a local chunk to the received chunk and forward it to the next node. All chunks are passed through the ring, and chunks can be merged at each node. Through this, when a chunk visits all nodes (for example, when the process is performed as many as the number of nodes), each node may hold one chunk in which the results of all nodes are combined.

A method of adding or removing computing nodes in this distributed learning model may be as follows.

1. Checkpoint method: Basically, when the node configuration is changed, learning is stopped and learning is resumed with the changed configuration. At this time, the computing node that did not participate in learning before does not have an up-to-date learning model, so the learning model should be saved (checkpoint) in the form of a file when learning was stopped and passed to the new computing node. .

2. Broadcast method: When a demand for changing node configuration occurs during learning, one of the existing computing nodes delivers information including the changed configuration and learning model to all nodes using broadcast communication. After completing this synchronization process, learning resumes.

In method 1, learning is suspended from the point of saving the learning model until learning is resumed with a new configuration, and in method 2, learning is suspended until the synchronization process of transmitting the learning model through broadcast communication is finished. Therefore, it is not possible to fully utilize the changing computing resources in the long-term learning process. That is, the efficiency of providing cloud resources is greatly reduced. In particular, since the section where the largest processing time occurs is the part where the learning model is transferred to a new computing node, a new technique to minimize this time is needed.

4 is an example of a cluster reconstruction method to which the present specification can be applied.

Referring to FIG. 4 , the cluster reconfiguration method of the present specification may be performed through a background thread and a main thread of a node to be added.

1. When a new node is added while learning of existing clusters is in progress, learning model synchronization process

For example, when adding a new node to existing clusters is requested, one of the existing computing nodes may complete learning in progress and transfer the model learned so far to the additionally requested node. At this time, since the learned model should not be changed during the transmission process, the node transmitting the learned model may be temporarily excluded from the learning process.

The node to be added participates in the communication pool to which the existing computing nodes are connected to the background thread in order to receive the gradient values calculated later while receiving the learning model. In addition, the background thread of the node that delivers the learned model can also participate in the communication pool to which existing computing nodes are connected, and the operation of the next node to be added can be performed in the same way. This is because, after the node to be added receives the learning model, the learning model changes as the learning of the existing computing nodes progresses while the learning preparation process is performed and the learning model is loaded.

The background thread of the node to be added prepares a tensor of the same type as the gradient value exchanged by existing computing nodes, and initializes the value to 0. If the existing computing nodes perform Allreduce operation to obtain the average value of the gradient, the background thread can participate in the Allreduce operation through the previously prepared tensor. That is, since the Allreduce operation is derived as a result of the sum of tensors transmitted by each node, the node to be added can receive the same gradient value as the existing computing nodes without affecting the learning operation process.

The node to be added can sequentially store the gradient values obtained through the above process in memory, and the Main Thread can secure the latest learning model by sequentially applying the gradient values stored in memory to the learning model.

2. Cluster participation process of nodes that have secured the latest learning model

The node to be added that has secured the latest learning model can apply the last received gradient value to the latest learning model at the same time that the existing nodes finish learning the next iteration. Upon application, the new node participates in learning along with the existing computing nodes. Then, the new node can perform an Allreduce operation with the gradient value calculated through learning without preparing a tensor to receive the gradient value.

Through this, in the present specification, the configuration of the learning cluster can be freely changed during the learning process, so that it can respond to changes in computing resources used in the learning process that is performed for a long time, and the efficiency of using cloud resources can be greatly increased.

5 is an embodiment to which the present specification may be applied.

Referring to FIG. 5 , a node may be implemented physically or virtualized by the electronic device 100 . This electronic device 100 may include a GPU (Graphic Processing Unit).

The first node to be added receives a learning model for distributed learning from the second node included in the existing cluster (S510). For example, the second node may stop the distributed learning in order to transfer the learning model to the first node when the first node requests to join the cluster. This can be managed by a system node that manages the distributed learning.

The first node makes the background thread participate in the cluster for distributed learning (S520). More specifically, the background thread can create a tensor of the same shape as the gradient value used in distributed learning. For example, this tensor has the same data type as the gradient value for distributed learning, and each data field can be initialized to 0.

The first node performs an Allreduce operation for distributed learning through tensors (S530).

The first node applies the obtained gradient value to the learning model through an allreduce operation (S540). For example, the first node may store gradient values obtained in an all-reduce operation in a memory through a tensor, and sequentially apply them to a learning model in a main thread. Through this, the first node can acquire the latest learning model without stopping the learning process of the existing cluster.

Thereafter, the first node receives the gradient value obtained through the distributed learning through the next iteration, and applies it to the learning model obtained immediately before. After that, the first node can join the cluster and perform distributed learning like nodes included in the existing cluster without generating a separate tensor.

Device General to which this specification may apply

Referring to FIG. 6 , a server (X200) according to the proposed embodiment may include a communication module (X210), a processor (X220), and a memory (X230). The communication module X210 is also referred to as a radio frequency (RF) unit. The communication module X210 may be configured to transmit various signals, data, and information to an external device and to receive various signals, data, and information to an external device. The server X200 may be connected to an external device by wire and/or wirelessly. The communication module X210 may be implemented by being separated into a transmission unit and a reception unit. The processor X220 may control overall operations of the server X200, and may be configured to perform a function of calculating and processing information to be transmitted and received by the server X200 to and from an external device. Also, the processor X220 may be configured to perform a server operation proposed in this specification. The processor X220 may control the communication module X110 to transmit data or messages to a UE, another vehicle, or another server according to the proposal of this specification. The memory X230 may store information processed for a predetermined period of time, and may be replaced with components such as a buffer.

In addition, the terminal device (X100) and the server (X200) as described above may include a GPU, and for their specific configuration, the matters described in the various embodiments of the present specification described above are applied independently or two or more embodiments are simultaneously applied. It can be implemented to be applied, and descriptions of overlapping contents are omitted for clarity.

The above specification can be implemented as computer readable code on a medium on which a program is recorded. The computer-readable medium includes all types of recording devices in which data that can be read by a computer system is stored. Examples of computer-readable media include Hard Disk Drive (HDD), Solid State Disk (SSD), Silicon Disk Drive (SDD), ROM, RAM, CD-ROM, magnetic tape, floppy disk, optical data storage device, etc. , and also includes those implemented in the form of a carrier wave (eg, transmission over the Internet). Accordingly, the above detailed description should not be construed as limiting in all respects and should be considered illustrative. The scope of this specification should be determined by reasonable interpretation of the appended claims, and all changes within the equivalent scope of this specification are included in the scope of this specification.

In addition, although services and embodiments have been described above, this is only an example and does not limit the present specification, and those skilled in the art to which this specification belongs will not deviate from the essential characteristics of the present service and embodiments. It will be seen that various modifications and applications not exemplified above are possible. For example, each component specifically shown in the embodiments can be modified and implemented. And differences related to these modifications and applications should be construed as being included in the scope of the present specification as defined in the appended claims.

Claims (10)

A method in which a first node performed by a computing device joins a cluster for distributed learning,
receiving a learning model for the distributed learning from a second node included in the cluster;
Involving a background thread in the cluster for distributed learning, wherein the background thread creates a tensor related to a gradient value for distributed learning;
performing an allreduce operation for the distributed learning through the tensor; and
applying first gradient values obtained through the all-reduce operation to the learning model;
Including, method.
According to claim 1,
The second node is
When the first node requests to join the cluster, stopping the distributed learning to deliver the learning model to the first node.
According to claim 1,
The tensor has the same type as the gradient value for the distributed learning and is initialized to 0.
According to claim 1,
The step of applying the first gradient values obtained through the all-reduce operation to the learning model
storing the obtained first gradient values in a memory; and
sequentially applying the first gradient values stored in the memory to the learning model through a main thread;
Further comprising a method.
According to claim 1,
Receiving, from a node of the cluster, a second gradient value related to the next repeated distributed learning;
applying the second gradient value to the learning model; and
joining the cluster;
Further comprising a method.
An apparatus for implementing a first node joining a cluster for distributed learning,
a transceiver for transmitting and receiving signals;
Memory; and
an AI processor for functionally controlling the transceiver and the memory;
Including,
The AI processor
Through the transceiver, the learning model for the distributed learning is received from the second node included in the cluster, a background thread participates in the cluster for the distributed learning, and the background thread is configured to participate in the distributed learning cluster. Generating a tensor related to a gradient value for , performing an Allreduce operation for the distributed learning through the tensor, and obtaining a first gradient through the Allreduce operation Apparatus for applying values to the learning model.
According to claim 6,
The second node is
When the first node requests to join the cluster, stopping the distributed learning to deliver the learning model to the first node.
According to claim 6,
The tensor has the same type as the gradient value for the distributed learning and is initialized to 0.
According to claim 6,
The AI processor
To apply the first gradient values obtained through the all-reduce operation to the learning model,
Apparatus for storing the obtained first gradient values in the memory and sequentially applying the first gradient values stored in the memory to the learning model through a main thread.
According to claim 6,
The AI processor
Receives, from the node of the cluster, a second gradient value related to the next repeated distributed learning, applies the second gradient value to the learning model, and joins the cluster.

Images