WO2022064656A1 - 処理システム、処理方法及び処理プログラム - Google Patents
処理システム、処理方法及び処理プログラム Download PDFInfo
- Publication number
- WO2022064656A1 WO2022064656A1 PCT/JP2020/036394 JP2020036394W WO2022064656A1 WO 2022064656 A1 WO2022064656 A1 WO 2022064656A1 JP 2020036394 W JP2020036394 W JP 2020036394W WO 2022064656 A1 WO2022064656 A1 WO 2022064656A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- inference
- processing
- model
- server device
- data
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N5/00—Computing arrangements using knowledge-based models
- G06N5/04—Inference or reasoning models
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0495—Quantised networks; Sparse networks; Compressed networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/0464—Convolutional networks [CNN, ConvNet]
Definitions
- the present invention relates to a processing system, a processing method and a processing program.
- resources such as the calculation amount and memory of the device used in the edge device are referred to as a device other than the edge device (hereinafter, for convenience, cloud) which is physically and logically located farther from the user than the edge device. ), It is poor. For this reason, if a process with a large computational load is performed by an edge device, it may take a large amount of time to complete the process, or it may take a long time to complete other processes that do not have a large computational load. There is. For example, there is a case where a process wait occurs for other processes while a process having a large amount of calculation is being performed by the edge device.
- Non-Patent Document 1 proposes application of so-called adaptive learning to the edge cloud. That is, in the method described in Non-Patent Document 1, a trained model trained using general-purpose training data in the cloud is expanded on an edge device, and learning is performed in the cloud using the data acquired by the edge device. By re-learning the model that has been used, the operation that takes advantage of the cloud and the edge device is realized.
- the above edge cloud network is expected to be applied to applications such as automatic analysis of surveillance camera images, automatic driving, and smart speakers.
- real-time performance is important as well as accuracy, but communication costs and delays between the edge and cloud have been problems.
- the present invention has been made in view of the above, and an object of the present invention is to provide a processing system, a processing method, and a processing program capable of reducing the amount of data transfer from the edge device to the server device and reducing the delay. do.
- the processing system is a processing system performed by using an edge device and a server device, and the edge device uses the first model.
- the inference result by the first inference unit is output, and when the reliability is equal to or less than the threshold value, the inference unit has a determination unit that outputs the feature amount of the data to be processed to the server device. It has a second inference unit that executes inference processing on the processing target data based on the feature amount of the processing target data output from the edge device by using the second model having higher inference accuracy than the first model. It is characterized by that.
- the present invention it is possible to reduce the amount of data transfer from the edge device to the server device and reduce the delay.
- FIG. 1 is a diagram illustrating an outline of a processing method of the processing system according to the embodiment.
- FIG. 2-1 is a diagram illustrating an example of DNN1 and DNN2.
- FIG. 2-2 is a diagram illustrating an example of DNN1 and DNN2.
- FIG. 3 is a diagram schematically showing an example of the configuration of the processing system according to the embodiment.
- FIG. 4 is a diagram illustrating an example of selecting a DNN base model.
- FIG. 5 is a diagram showing an outline of the structure of YOLO v3.
- FIG. 6-1 is a diagram illustrating an example of the structure of DNN.
- FIG. 6-2 is a diagram illustrating an example of the structure of DNN.
- FIG. 7 is a sequence diagram showing a processing flow of the processing system according to the embodiment.
- FIG. 2-1 is a diagram illustrating an example of DNN1 and DNN2.
- FIG. 2-2 is a diagram illustrating an example of DNN1 and DNN2.
- FIG. 8 is a diagram showing the processing time of each DNN selected as an example in the processing system.
- FIG. 9 is a distribution diagram of the entropy of the test data obtained based on the inference result of DNN.
- FIG. 10 is a diagram showing the relationship between the offload rate and the overall accuracy.
- FIG. 11 is a diagram showing an example of a computer in which an edge device and a server device are realized by executing a program.
- FIG. 1 is a diagram illustrating an outline of a processing method of the processing system according to the embodiment.
- the high-precision model and the lightweight model constitute a model cascade.
- an edge device using a high-speed and low-precision lightweight model for example, DNN1 (first model)
- a low-speed and high-precision high-precision model for example, DNN2 (second model)
- a server device is a device that is physically and logically located far from the user.
- the edge device is an IoT device and various terminal devices that are physically and logically close to the user, and has less resources than a server device.
- DNN1 and DNN2 are models that output inference results based on the input processing target data.
- the DNN1 takes an image as an input and infers the probability of each class of the object appearing in the image based on the feature amount extracted from the image.
- the DNN2 infers the probability of each class of the object to be inferred in the image to be inferred by using the feature amount extracted by the DNN1 (intended to output a predetermined intermediate layer, hereinafter referred to as a feature map) as an input.
- both DNN1 and DNN2 make inferences for the same image.
- the edge device it is determined whether to adopt the inference result of DNN1 or DNN2 based on the comparison result between the reliability (described later) and the threshold value.
- the reliability is a value for determining whether the processing target data should be processed in the edge device or the server device.
- the reliability may be obtained, for example, based on the entropy of the inference result by DNN1.
- the reliability may be set to a value that becomes smaller as the entropy of the inference result of DNN1 becomes higher. Specific examples will be described later.
- the reliability of the inference of the DNN1 classification for the object shown in the input image is acquired. Then, in the processing system, when the acquired reliability exceeds a predetermined threshold value, the inference result of DNN1 is adopted. That is, the inference result of the lightweight model is output as the final estimation result of the model cascade. On the other hand, in the processing system, when the reliability is equal to or less than a predetermined threshold value, the inference result obtained by inputting the feature map extracted by DNN1 into DNN2 is output as the final inference result.
- the processing system according to the embodiment selects the edge device or the server device based on the comparison result between the reliability and the threshold value as to which of the edge device and the server device should process the data to be processed. Then, the processing target data is processed. Therefore, the processing system according to the embodiment can control whether the processing is executed in the edge device or the cloud.
- DNN1 and DNN2 are diagrams illustrating an example of DNN1 and DNN2.
- the DNN has an input layer for inputting data, a plurality of intermediate layers for variously converting data input from the input layer, and an output layer for outputting so-called inferred results such as probability and likelihood.
- the output value output from each layer may be irreversible if the input data needs to maintain anonymity.
- the processing system uses DNN2 on the cloud side and DNN1 on the edge side.
- a feature extraction layer Bf2 that extracts the feature amount of the input image and outputs it as a feature map, and a layer that performs processing using the extracted feature amount (for example, reflected in the input image).
- a high-precision model with a layer that detects an object and infers the probability of each class of the object that appears in the input image (hereinafter collectively referred to as the detection layer Bd2) and a detection layer Bd2 (second execution unit) is adopted. (See Fig. 2-1).
- the training data is used in advance to train DNN2'before placement.
- the trained feature extraction layer Bf2 of DNN2 ′ is used as it is on the edge side.
- a DNN that is placed in front of the detection layer Bd1 (see arrow Y1 in FIG. 2-1) and that combines the feature extraction layer Bf2 (first extraction unit) and the detection layer Bd1 (first execution unit) is edged. Let it be DNN1 on the side (see Fig. 2-2).
- the detection layer Bd1 infers the probability of each class of the object reflected in the input image by using the feature map (feature amount of a certain layer) extracted by the feature extraction layer Bf2 of DNN1.
- the DNN1 arranged on the edge side deletes the feature amount extraction layer Bf1 (see FIG. 2-1) from the original DNN1 and arranges the feature amount extraction layer Bf2 of the trained DNN2'instead (see FIG. 2-2). ). That is, DNN1 composed of a combination of the feature amount extraction layer Bf2 of DNN2'and the detection layer Bd1 of DNN1'is arranged on the edge side. Of the DNN1 shown in FIG. 2-2, the feature amount extraction layer Bf2 may be fixed and learning may be performed again. Further, the DNN2 arranged on the server side has a configuration in which the feature amount extraction layer Bf2 is deleted from the initial DNN2'(see FIG. 2-1) (see FIG. 2-2).
- the parameters of the feature extraction layer Bf2 are fixed to the parameters after the initial training of DNN2', and the detection layer Bd1 in the subsequent stage is trained using the learning data.
- both the feature extraction layer Bf2 and the detection layer Bd1 are trained using the learning data.
- learning may be performed in cooperation with DNN2 and DNN1.
- DNN2 may be retrained using the training data used by DNN1.
- training may be performed using common learning data between DNN1 and DNN2.
- the feature extraction layer Bf2 in the initial DNN2' is arranged as the feature extraction layer Bf2 of the DNN1 on the edge side. Therefore, the detection layer Bd2 of the DNN2 on the cloud side can execute the inference process using the feature map output by the feature extraction layer Bf2 of the DNN1 on the edge side (see the arrow Y2 in FIG. 2-2). .. Therefore, in the processing system according to the present embodiment, it can be said that the same feature map is shared between the edge and the cloud, and inference processing can be performed on each of the edge side and the cloud side.
- the execution of the feature extraction process can be omitted, so that the calculation time can be shortened and the delay can be reduced.
- the data output from the edge side to the cloud side is not an image to be processed but a feature map extracted from the image, it is possible to reduce the amount of data transfer from the edge side to the cloud side. Since it is sufficient that the same feature map can be shared between the edge and the cloud, the minimum condition that the feature extraction layer Bf2 of DNN1 actually arranged on the edge side and the feature extraction layer Bf2 of the initial DNN2'must be satisfied is The sizes of the layers to be connected are the same. This is because the detection layer of DNN1 is relearned again.
- the parameter values may differ between the feature extraction layer Bf2 of DNN1 and the feature extraction layer Bf2 of DNN2'.
- FIG. 3 is a diagram schematically showing an example of the configuration of the processing system according to the embodiment.
- the processing system 100 includes a server device 20 and an edge device 30. Further, the server device 20 and the edge device 30 are connected via the network N.
- the network N is, for example, the Internet.
- the server device 20 is a server provided in a cloud environment.
- the edge device 30 is, for example, an IoT device and various terminal devices.
- a predetermined program is read into a computer or the like including a ROM (Read Only Memory), a RAM (Random Access Memory), a CPU (Central Processing Unit), etc., and the CPU loads the predetermined program. It is realized by executing it.
- so-called accelerators represented by GPUs, VPUs (Vision Processing Units), FPGAs (Field Programmable Gate Arrays), ASICs (Application Specific Integrated Circuits), and dedicated AI (Artificial Intelligence) chips are also used.
- the server device 20 and the edge device 30 each have a NIC (Network Interface Card) or the like, and may communicate with other devices via a telecommunication line such as a LAN (Local Area Network) or the Internet. It is possible.
- the server device 20 has an inference unit 21 (second inference unit) that makes inferences using the trained high-precision model DNN2.
- DNN2 contains information such as model parameters.
- DNN2 has a detection layer Bd2 as described above.
- the inference unit 21 uses DNN2 to execute inference processing on the input image based on the feature map of the input image output from the edge device 30.
- the inference unit 21 returns the quantized feature map output from the edge device 30 to the FP32 and uses it as an input to the detection layer Bd2 of the DNN2.
- the inference unit 21 acquires an inference result (for example, a probability for each class of an object shown in an image) as an output of DNN2.
- the inference unit 21 receives the inference data, that is, the output value of Bf2 of the feature extraction layer of DNN1, and outputs the inference result.
- the feature map is a feature quantity of data whose label is unknown.
- the inference data is an image.
- the server device 20 and the edge device 30 form a model cascade. Therefore, the inference unit 21 does not always make inferences.
- the inference unit 21 accepts the input of the quantized feature map and performs inference by the detection layer Bd2 of the DNN2.
- the edge device 30 has an inference unit 31 (first inference unit) having a trained lightweight model DNN1, a determination unit 32, and a quantization unit 33.
- the inference unit 31 makes inferences using the trained lightweight model DNN1.
- DNN1 contains information such as model parameters.
- the DNN1 has a feature extraction layer Bf2 and a detection layer Bd1 of the trained DNN2'(see FIG. 2-1) as described above.
- the inference unit 31 uses DNN1 to extract the feature amount of the input image as a feature map, and executes inference processing on the input image based on the extracted feature map.
- the inference unit 31 inputs an image to be processed into DNN1 and acquires an inference result.
- the inference unit 31 extracts the feature amount of the processing target data using DNN1, and executes the inference processing for the processing target data based on the extracted feature amount.
- the inference unit 31 accepts the input of the image to be processed, processes the image to be processed, and outputs the inference result (for example, the probability for each class of the object to be reflected in the image).
- the determination unit 32 determines whether to adopt the inference result of the edge device 30 or the server device 20 by comparing the reliability with a predetermined threshold value.
- the determination unit 32 When the reliability exceeds a predetermined threshold value, the determination unit 32 outputs the inference result inferred by the inference unit 31. When the reliability is equal to or less than a predetermined threshold value, the determination unit 32 outputs a feature map, which is an output of the feature extraction layer Bf2 connected to DNN1 on the edge side, to the server device 20, and performs inference processing to the server device 20. It is determined that the arranged DNN2 is executed.
- the reliability may be set to a value that becomes smaller as the entropy of the inference result of DNN1 becomes larger.
- the entropy of the inference result of DNN1 may be obtained by the equation (1).
- C is a set of all labels c output by DNN1, and y is the probability of each label.
- the quantized unit 33 quantizes the feature map extracted by the feature extraction layer Bf2 of the DNN 1 and quantizes the quantized feature map into the server device 20. Output to.
- FIG. 4 is a diagram illustrating an example of selecting a base model of DNN1 and DNN2.
- darknet19 which is a back-end model of YOLOv2 which is relatively lightweight and high speed
- YOLOv3 Darknet53
- the selected NN is an example, and if it is a high-precision NN, a high-speed NN and a NN whose accuracy is lower than that of a high-precision NN, and a NN that can share the feature quantity extraction layer, how? You may make a choice.
- the edge device 30 and the server device 20 may be configured to have different depths in the same NN.
- FIG. 5 is a diagram showing an outline of the structure of YOLO v3. It is a figure explaining the selection example of the base model of DNN1 and DNN2.
- YOLOv3 has a convolutional layer (feature extraction layer) Bf-1 having a Residual block Bf-1 for feature extraction, and a detection unit Bd2-1 which is a network (FPN) for object detection.
- FPN network
- YOLOv3 maintains high accuracy. Therefore, it is desirable to avoid damaging the configuration in Residual block Bf-1. Subsequently, in YOLOv3, it is desirable to have a configuration in which the feature map is received in the previous stage of the detection unit Bd2-1 so that the detection unit Bd2-1 can execute the detection using the feature map as it is. The deeper the layer (the larger the number), the heavier the calculation of the model and the larger the number of parameters to be used. Therefore, it is desirable that the DNN1 of the edge device 30 is a model having a shallow layer in order to be lightweight and high speed.
- FIG. 6-1 and 6-2 are diagrams illustrating an example of the structure of DNN1 and DNN2.
- the pre-stage of YOLOv2 in which the trained feature extraction layer Bf2-1 holding the Residual block Bf-1 is selected as it is as the base model of the edge device 30 in order to avoid damaging the configuration in the Residual block Bf-1 of YOLOv3. (Arrow Y11 in FIG. 6-1), and used as DNN1 of the edge device 30. Therefore, in the edge device 30, the feature map is extracted from the image in the feature extraction layer Bf2-1 using the YOLOv2 reconstructed in this way, and the input image is used in the detection layer Bd1-1 using the extracted feature map. Infer the probability of each class of objects reflected in.
- the edge device 30 outputs Result1 assuming that Result1 is reliable (see arrow Y11 in FIG. 6-2). ) End the process.
- the feature extraction layer Bf2-1 in the edge device 30 has a structure common to the feature extraction layer Bf2-1 of YOLOv3 which is the base model of the server device 20, the feature extraction layer Bf2-1 The feature map output from the server device 20 can also be shared by the detection unit Bd2-1 of the server device 20.
- YOLOv3 having a configuration in which the feature extraction layer Bf2-1 is deleted from YOLOv3 is applied as DNN2.
- the edge device 30 when Result1 is equal to or less than the threshold value, it is considered that Result1 is unreliable, and the feature map of the feature extraction layer Bf2-1 is output to the edge device 30 after being quantized (arrow Y12 in FIG. 6-2). reference). Then, in the server device 20, this feature map is returned to FP32, the feature map is input to the detection unit Bd2-1 of the reconstructed YOLOv3, and the inference result (Result2) of the detection unit Bd2-1 is output (Result2). See arrow Y13 in FIG. 6-2). Therefore, the calculation range in the server device 20 is the calculation range in the detection unit Bd2-1. In other words, in the server device 20, the calculation in the feature extraction layer Bf2-1 can be omitted.
- FIG. 7 is a sequence diagram showing a processing flow of the processing system according to the embodiment.
- the edge device 30 when the inference unit 31 accepts the input of the image (step S1), the input image is input to the DNN1.
- the feature extraction layer Bf2 extracts the feature amount of the input image as a feature map (step S2) and outputs it to the determination unit 32 (step S3).
- the detection layer Bd1 executes inference processing for the input image based on the feature map, for example, detection processing (step S4), and outputs the inference process to the determination unit 32 (step S5).
- the determination unit 32 calculates the reliability based on the inference result by the inference unit 31 (step S6), compares the calculated reliability with a predetermined threshold value, and determines whether or not the reliability is equal to or less than the predetermined threshold value. (Step S7).
- step S7 When the reliability is not equal to or less than the threshold value (step S7: No), that is, when the reliability exceeds the threshold value, the determination unit 32 outputs the inference result inferred by DNN1 of the inference unit 31 (step S8).
- step S7 when the reliability is equal to or less than the threshold value (step S7: Yes), the determination unit 32 outputs the feature map to the quantization unit 33 (step S9), and the quantization unit 33 quantizes the feature map (step). S10), transmission to the server device 20 (step S11).
- the inference unit 21 returns the quantized feature map output from the edge device 30 to the FP32 and uses it as an input to the detection layer Bd2 of the DNN2.
- the detection layer Bd2 executes inference processing for the input image, for example, detection processing, based on the feature map output from the edge device 30 (step S12).
- the server device 20 transmits the inference result of the DNN 2 to the edge device 30 (step S13), and outputs the inference result from the edge device 30 (step S14). In the actual embodiment, it is assumed that the inference result is returned to the user, and the final inference result is output from the edge device 30.
- the inference result of DNN2 may be output from the server device 20, or the server device 20 may be configured to hold the inference result as it is.
- the edge device 30 may transmit the inference result to the server device 20 when the inference result is used on the server device 20 side.
- FIG. 8 is a diagram showing the processing time of each DNN selected as an example in the processing system 100.
- the edge device DNN1 is a model in which the trained feature extraction layer Bf2-1 holding the Residual block Bf-1 is placed in front of YOLOv2 and trained.
- DNN2 of the server device 20 is YOLOv3 which has been learned.
- the processing time is also shown for YOLO v2, which is the base model of the edge device 30.
- the edge device DNN1 has a feature extraction layer Bf2-1 minutes deeper than YOLOv2, so that the processing time is longer than YOLOv2, but the inference accuracy is improved compared to YOLOv2.
- the processing time per image of the edge device DNN1 is about twice as fast as that of the server device 20 DNN2 YOLOv3. In this way, even when the feature extraction layer Bf2-1 of YOLOv3 is arranged in place of the feature extraction layer of YOLOv2, the DNN1 of the edge device can maintain the speeding up of processing while improving the inference accuracy. It is possible.
- FIG. 9 is a distribution diagram of the entropy of the test data obtained based on the inference result of DNN1.
- Offload rate 0 is a state in which all data is processed by the edge device 30 and the original accuracy (acc_origin) is low when not quantized
- Offload rate 1 is a state in which all data is a server. It is processed by the apparatus 20 and has a high original accuracy (acc_origin) when it is not quantized.
- Int4 quantization, int6 quantization, and int8 quantization are performed on the feature map by the feature extraction layer Bf2 of the edge device 30.
- int4 quantization when int6 quantization, and int8 quantization, when int8 quantization is performed on the feature map by the feature extraction layer Bf2 of the edge device 30, the accuracy is improved.
- the amount of transferred data could be reduced by 75% compared to the case without quantization with almost no loss. Therefore, it is desirable to apply int8 quantization to the quantization of the feature map.
- a cost term may be provided so that the feature map output by the feature extraction layer Bf2 becomes more sparse.
- int6 and / or int4 show an offload rate and / or overall accuracy close to int8 in FIG. 10, it shows an offload rate and / or overall accuracy close to int8 among int6 or int4.
- the smaller of the quantization bit rates may be selected.
- the quantized feature map may be compressed.
- an image coding method such as HEVC or VVC may be adopted.
- a general-purpose compression method such as ZIP may be adopted. It should be noted that the same property as the above-mentioned natural image is intended to be a general property incorporated in image coding such that the correlation between adjacent pixels is often high.
- dequantization may be performed when quantization is performed, and decoding may be performed when compression is performed. Quantization and / or compression may be performed by an edge device or by another device physically or logically located closer to the edge device (rather than the server device). The dequantization and / or decoding may be performed by the server device or by another device physically or logically located closer to the server device (rather than the edge device).
- the offload rate exceeds 0.4 (entropy threshold value is 0.5)
- the threshold value is set to 0.5, it is considered that the offload rate and the accuracy are balanced. In this way, by setting the threshold value according to the balance between the offload rate and the accuracy, it is possible to adjust the offload rate and the overall accuracy according to each use case.
- the threshold value for deciding whether to infer on the edge device side or the server device side As an example, it is conceivable to use the accuracy as a whole as a reference. If the inference on the edge device side or the inference on the server device side does not change the accuracy so much, the edge device side can process it, and if the accuracy drops when the edge device side processes it, the server device side can process it.
- a threshold that can be processed is desirable. For example, in the case of FIG. 10, it can be seen that when the quantization bit rate is 8 (acc_int8), the threshold value of the offload rate should be 0.4. Then, the threshold value of the entropy of the inference result on the edge side may be determined so that the offload rate becomes 0.4. In the case of FIG. 10, the entropy threshold value may be about 0.5.
- the feature extraction layer Bf2 in the DNN2'before being arranged in the server device 20 is arranged as the feature extraction layer Bf2 of the DNN1 of the edge device 30, so that the feature map output by the edge device 30 is obtained.
- the execution of the feature extraction process can be omitted, so that the calculation time of the entire system can be shortened and the delay can be reduced.
- the output data from the edge device 30 to the server device 20 is not an image to be processed but a feature map extracted from the image, it is possible to reduce the amount of data transfer from the edge device 30 to the server device 20.
- the edge device transmits a feature map that is neither a processed image or an encoded image nor a commonly used feature quantity such as a frequency signal to the server device, it is a secret to a third party. It is also possible to improve the sex. In order to ensure confidentiality, it is sufficient to impose a constraint that the relationship between the feature map and the input data (target image) is irreversible when learning.
- DNN1 in which the feature extraction layer Bf2 in DNN2'before being arranged in the server device 20 is arranged in front of the detection layer Bd1 in the edge device 30.
- this DNN1 we actually performed a classification task and evaluated that it is effective as a lightweight model in terms of both accuracy and speed.
- the reliability used in determining which inference result of the edge device 30 or the server device 20 is to be adopted is formulated as the entropy of the inference result of DNN1 and the threshold value is set. This made it possible to adjust the offload rate and overall accuracy according to the use case during actual operation.
- an image has been described as an example of the data to be processed, but the detection result may be detected by various sensors, not limited to the image.
- the case where YOLOv2 and YOLOv3 are applied as the base model of DNN1 and DNN2 has been described as an example, but the base model of DNN1 and DNN2 may be appropriately set according to the task.
- edge devices 30 or server devices 20 there may be a plurality of edge devices 30 or server devices 20, and there may be a plurality of edge devices 30 and server devices 20.
- each component of each of the illustrated devices is functional and conceptual, and does not necessarily have to be physically configured as shown in the figure. That is, the specific form of distribution / integration of each device is not limited to the one shown in the figure, and all or part of them may be functionally or physically distributed / physically in arbitrary units according to various loads and usage conditions. Can be integrated and configured. Further, each processing function performed by each device may be realized by a CPU and a program analyzed and executed by the CPU, or may be realized as hardware by wired logic.
- FIG. 11 is a diagram showing an example of a computer in which the edge device 30 and the server device 20 are realized by executing a program.
- the computer 1000 has, for example, a memory 1010 and a CPU 1020. Further, the accelerator described above may be provided to assist the calculation.
- the computer 1000 also has a hard disk drive interface 1030, a disk drive interface 1040, a serial port interface 1050, a video adapter 1060, and a network interface 1070. Each of these parts is connected by a bus 1080.
- the memory 1010 includes a ROM (Read Only Memory) 1011 and a RAM 1012.
- the ROM 1011 stores, for example, a boot program such as a BIOS (Basic Input Output System).
- BIOS Basic Input Output System
- the hard disk drive interface 1030 is connected to the hard disk drive 1090.
- the disk drive interface 1040 is connected to the disk drive 1100.
- a removable storage medium such as a magnetic disk or an optical disk is inserted into the disk drive 1100.
- the serial port interface 1050 is connected to, for example, a mouse 1110 and a keyboard 1120.
- the video adapter 1060 is connected to, for example, the display 1130.
- the hard disk drive 1090 stores, for example, an OS (Operating System) 1091, an application program 1092, a program module 1093, and program data 1094. That is, the program that defines each process of the edge device 30 and the server device 20 is implemented as a program module 1093 in which a code that can be executed by a computer is described.
- the program module 1093 is stored in, for example, the hard disk drive 1090.
- the program module 1093 for executing the same processing as the functional configuration in the edge device 30 and the server device 20 is stored in the hard disk drive 1090.
- the hard disk drive 1090 may be replaced by an SSD (Solid State Drive).
- the setting data used in the processing of the above-described embodiment is stored as program data 1094 in, for example, a memory 1010 or a hard disk drive 1090. Then, the CPU 1020 reads the program module 1093 and the program data 1094 stored in the memory 1010 and the hard disk drive 1090 into the RAM 1012 and executes them as needed.
- the program module 1093 and the program data 1094 are not limited to those stored in the hard disk drive 1090, but may be stored in, for example, a removable storage medium and read by the CPU 1020 via the disk drive 1100 or the like. Alternatively, the program module 1093 and the program data 1094 may be stored in another computer connected via a network (LAN (Local Area Network), WAN (Wide Area Network), etc.). Then, the program module 1093 and the program data 1094 may be read from another computer by the CPU 1020 via the network interface 1070.
- LAN Local Area Network
- WAN Wide Area Network
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Data Mining & Analysis (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computational Linguistics (AREA)
- Artificial Intelligence (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Image Analysis (AREA)
- Apparatus For Radiation Diagnosis (AREA)
Abstract
Description
[実施の形態の概要]
本発明の実施の形態について説明する。本発明の実施の形態では、学習済みの高精度モデル及び軽量モデルを使って推論処理を行う処理システムについて説明する。なお、実施の形態の処理システムでは、推論処理において用いるモデルとして、DNN(Deep Neural Network)を用いた場合を例に説明する。実施の形態の処理システムでは、どのようなニューラルネットワークを用いてもよいし、学習済みモデルに代えて低演算量の信号処理と高演算量の信号処理を用いてもよい。
次に、DNN1、DNN2について説明する。図2-1及び図2-2は、DNN1及びDNN2の一例を説明する図である。DNNは、データが入る入力層、入力層から入力されたデータを様々に変換する複数の中間層、確率や尤度など、いわゆる推論した結果を出力する出力層を有する。各層から出力される出力値は、入力されるデータが匿名性を保つ必要がある場合は非可逆としてもよい。
次に、処理システムの構成について説明する。図3は、実施の形態に係る処理システムの構成の一例を模式的に示す図である。
図4は、DNN1、DNN2のベースモデルの選定例について説明する図である。本実施の形態では、物体のクラスごとの確率を推論するモデルとして、図4に示すように、比較的軽量かつ高速であるYOLOv2のバックエンドモデルであるdarknet19(以下、YOLOv2と記載する。)を、エッジ装置30のベースモデルとして選定し、比較的高精度であるYOLOv3のバックエンドモデルであるdarknet53(以下、YOLOv3と記載する。)を、サーバ装置20のベースモデルとして選定した。選定したNNは例であり、高精度なNNと、高速であるNNであって高精度なNNよりも精度は落ちるNNとであって、特徴量抽出層を共有できるNNであればどのように選択をしてもよい。簡単な例では、同一のNNで、エッジ装置30とサーバ装置20とで深さが異なるように構成してもよい。
図6-1及び図6-2は、DNN1,DNN2の構造の一例を説明する図である。YOLOv3のResidual block Bf-1内の構成を損なうことを避けるため、Residual block Bf-1を保持する学習済みの特徴抽出層Bf2-1を、そのまま、エッジ装置30のベースモデルとして選定したYOLOv2の前段に配置し(図6-1の矢印Y11)、エッジ装置30のDNN1とした。したがって、エッジ装置30では、このように再構築したYOLOv2を用いて、特徴抽出層Bf2-1において画像から特徴マップを抽出し、抽出した特徴マップを用いて、検知層Bd1-1において、入力画像に写る物体のクラスごとの確率を推論する。
図7は、実施の形態に係る処理システムの処理の流れを示すシーケンス図である。図7に示すように、まず、エッジ装置30において、推論部31は、画像の入力を受け付けると(ステップS1)、入力画像をDNN1に入力する。DNN1では、特徴抽出層Bf2が、入力画像の特徴量を特徴マップとして抽出し(ステップS2)、判定部32に出力する(ステップS3)。DNN1では、検知層Bd1が、特徴マップを基に入力画像に対する推論処理、例えば、検知処理を実行し(ステップS4)、判定部32に出力する(ステップS5)。
エッジ装置30のDNN1の処理時間と、サーバ装置のDNN2の処理時間とを評価した。タスクは、分類であり、テストデータは、ImageNet(各クラス300枚、計3000画像)であり、HWは、Platform: NVIDIA GeForceRTX2070+AMD 3600である。図8は、処理システム100において一例として選定した各DNNの処理時間を示す図である。
エッジ装置30のDNN1を用いて、テストデータのエントロピーの分布を可視化した。タスク、テストデータ、HWは、図8と同じ条件である。図9は、DNN1の推論結果を基に求めたテストデータのエントロピーの分布図である。
そこで、エントロピーの閾値(オフロード率)の変動に伴う推論結果の全体精度の変動を求める評価実験3を行った。なお、閾値は、オフロード率に連動し、オフロード率を下げる場合には閾値を上げる。この評価実験では、int4、int6、int8で、エッジ装置30の特徴抽出層Bf2によって抽出された特徴マップに対する量子化を行った。なお、タスク、テストデータ、HWは、図8と同じ条件である。図10は、オフロード率と全体精度との関係を示す図である。図10において、「Offload rate 0」は、全てのデータがエッジ装置30により処理され、量子化しない場合のオリジナル精度(acc_origin)が低い状態であり、「Offload rate 1」は、すべてのデータがサーバ装置20により処理され、量子化しない場合のオリジナル精度(acc_origin)が高い状態である。エッジ装置30の特徴抽出層Bf2による特徴マップに対しては、int4量子化、int6量子化、int8量子化を行っている。
このように、実施の形態では、サーバ装置20に配置する前のDNN2´における特徴抽出層Bf2を、エッジ装置30のDNN1の特徴抽出層Bf2として配置することで、エッジ装置30が出力した特徴マップを、サーバ装置20においても共有することができる。すなわち、サーバ装置20のDNN2の検知層Bd2は、エッジ装置30のDNN1の特徴抽出層Bf2が出力した特徴マップを用いて推論処理を実行することが可能になる。
図示した各装置の各構成要素は機能概念的なものであり、必ずしも物理的に図示の如く構成されていることを要しない。すなわち、各装置の分散・統合の具体的形態は図示のものに限られず、その全部又は一部を、各種の負荷や使用状況等に応じて、任意の単位で機能的又は物理的に分散・統合して構成することができる。さらに、各装置にて行なわれる各処理機能は、その全部又は任意の一部が、CPU及び当該CPUにて解析実行されるプログラムにて実現され、あるいは、ワイヤードロジックによるハードウェアとして実現され得る。
図11は、プログラムが実行されることにより、エッジ装置30及びサーバ装置20が実現されるコンピュータの一例を示す図である。コンピュータ1000は、例えば、メモリ1010、CPU1020を有する。また、演算を補助するために前述したアクセラレータを備えてもよい。また、コンピュータ1000は、ハードディスクドライブインタフェース1030、ディスクドライブインタフェース1040、シリアルポートインタフェース1050、ビデオアダプタ1060、ネットワークインタフェース1070を有する。これらの各部は、バス1080によって接続される。
21,31 推論部
30 エッジ装置
32 判定部
33 量子化部
100 処理システム
Claims (6)
- エッジ装置とサーバ装置とを用いて行われる処理システムであって、
前記エッジ装置は、
第1のモデルを用いて、処理対象データの特徴量を抽出し、抽出した特徴量を基に前記処理対象データに対する推論処理を実行する第1の推論部と、
前記第1の推論部による推論結果に対する信頼度が閾値を超えた場合には前記第1の推論部による推論結果を出力し、前記信頼度が閾値以下である場合には前記処理対象データの特徴量を前記サーバ装置に出力する判定部と、
を有し、
前記サーバ装置は、
前記第1のモデルよりも推論精度が高い第2のモデルを用いて、前記エッジ装置から出力された前記処理対象データの特徴量を基に前記処理対象データに対する推論処理を実行する第2の推論部と、
を有することを特徴とする処理システム。 - 前記第1のモデルは、
前記処理対象データの特徴量を抽出する第1の抽出部と、
前記処理対象データの特徴量を基に第1の推論処理を実行する第1の実行部と、
を有し、
前記第2のモデルは、
前記処理対象データの特徴量を基に第2の推論処理を実行する第2の実行部
を有し、
前記第2の実行部は、前記第1の抽出部によって抽出された前記処理対象データの特徴量を基に、前記第2の推論処理を実行することを特徴とする請求項1に記載の処理システム。 - 前記判定部は、量子化した前記処理対象データの特徴量を前記サーバ装置に出力することを特徴とする請求項1または2に記載の処理システム。
- 前記信頼度は、前記第1の推論部による推論結果のエントロピーに基づくことを特徴とする請求項1~3のいずれか一つに記載の処理システム。
- エッジ装置とサーバ装置とを用いて行われる処理システムが実行する処理方法であって、
前記エッジ装置が、第1のモデルを用いて、処理対象データの特徴量を抽出し、抽出した特徴量を基に前記処理対象データに対する推論処理を実行する第1の推論工程と、
前記エッジ装置が、前記第1の推論工程における推論結果に対する信頼度が閾値を超えた場合には前記第1の推論工程における推論結果を出力し、前記信頼度が閾値以下である場合には前記処理対象データの特徴量を前記サーバ装置に出力する判定工程と、
前記サーバ装置が、前記第1のモデルよりも推論精度が高い第2のモデルを用いて、前記エッジ装置から出力された前記処理対象データの特徴量を基に前記処理対象データに対する推論処理を実行する第2の推論工程と、
を含んだことを特徴とする処理方法。 - 方法をコンピュータに実行させる処理プログラムであって、
エッジ装置としてのコンピュータに、
第1のモデルを用いて、処理対象データの特徴量を抽出し、抽出した特徴量を基に前記処理対象データに対する推論処理を実行する第1の推論ステップと、
前記第1の推論ステップにおける推論結果に対する信頼度が閾値を超えた場合には前記第1の推論ステップにおける推論結果を出力し、前記信頼度が閾値以下である場合には前記処理対象データの特徴量を出力する判定ステップと、
を実行させ、
サーバ装置としてのコンピュータに、
前記第1のモデルよりも推論制度が高い第2のモデルを用いて、前記エッジ装置から出力された前記処理対象データの特徴量を基に前記処理対象データに対する推論処理を実行する第2の推論ステップ、
を実行させることを特徴とする処理プログラム。
Priority Applications (5)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/JP2020/036394 WO2022064656A1 (ja) | 2020-09-25 | 2020-09-25 | 処理システム、処理方法及び処理プログラム |
| EP20955250.4A EP4220498B8 (en) | 2020-09-25 | 2020-09-25 | Processing system, processing method, and processing program |
| CN202080105384.7A CN116194933B (zh) | 2020-09-25 | 2020-09-25 | 处理系统、处理方法以及记录介质 |
| JP2022551064A JP7452679B2 (ja) | 2020-09-25 | 2020-09-25 | 処理システム、処理方法及び処理プログラム |
| US18/027,604 US20250292120A1 (en) | 2020-09-25 | 2020-09-25 | Processing system, processing method, and processing program |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| PCT/JP2020/036394 WO2022064656A1 (ja) | 2020-09-25 | 2020-09-25 | 処理システム、処理方法及び処理プログラム |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2022064656A1 true WO2022064656A1 (ja) | 2022-03-31 |
Family
ID=80846432
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/JP2020/036394 Ceased WO2022064656A1 (ja) | 2020-09-25 | 2020-09-25 | 処理システム、処理方法及び処理プログラム |
Country Status (5)
| Country | Link |
|---|---|
| US (1) | US20250292120A1 (ja) |
| EP (1) | EP4220498B8 (ja) |
| JP (1) | JP7452679B2 (ja) |
| CN (1) | CN116194933B (ja) |
| WO (1) | WO2022064656A1 (ja) |
Cited By (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115984166A (zh) * | 2022-11-04 | 2023-04-18 | 慧视云创(北京)科技有限公司 | 一种边云协同的图像变化检测方法及系统 |
| WO2024029347A1 (ja) * | 2022-08-04 | 2024-02-08 | ソニーセミコンダクタソリューションズ株式会社 | 情報処理装置、情報処理方法、及び情報処理システム |
| WO2024057578A1 (ja) * | 2022-09-12 | 2024-03-21 | 日本電信電話株式会社 | 抽出システム、抽出方法および抽出プログラム |
| WO2024079902A1 (ja) * | 2022-10-14 | 2024-04-18 | 日本電気株式会社 | 処理制御システム、処理制御装置、および処理制御方法 |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20230111874A1 (en) * | 2021-10-12 | 2023-04-13 | Oracle International Corporation | Device emulations in a notebook session |
| WO2025182439A1 (ja) * | 2024-03-01 | 2025-09-04 | ソニーセミコンダクタソリューションズ株式会社 | 情報処理装置、情報処理方法およびプログラム |
| US20250378690A1 (en) * | 2024-04-25 | 2025-12-11 | Moodmetrics Ai | Method and system for extracting inherent user feature using artificial intelligence |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2008309945A (ja) * | 2007-06-13 | 2008-12-25 | Kddi Corp | パターンマッチング方法および装置ならびにその特徴量正規化方法および装置 |
| JP2017126158A (ja) * | 2016-01-13 | 2017-07-20 | 日本電信電話株式会社 | 2値分類学習装置、2値分類装置、方法、及びプログラム |
| JP2018045369A (ja) * | 2016-09-13 | 2018-03-22 | 株式会社東芝 | 認識装置、認識システム、認識方法およびプログラム |
| WO2019220620A1 (ja) * | 2018-05-18 | 2019-11-21 | 日本電気株式会社 | 異常検出装置、異常検出方法及びプログラム |
| JP2020065225A (ja) * | 2018-10-19 | 2020-04-23 | 株式会社日立製作所 | 計算機、センシングシステム、及びデータ通信方法 |
Family Cites Families (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP6171660B2 (ja) * | 2013-07-22 | 2017-08-02 | 富士ゼロックス株式会社 | 情報処理装置、情報処理システム及びプログラム |
| JP6632288B2 (ja) * | 2014-12-12 | 2020-01-22 | キヤノン株式会社 | 情報処理装置、情報処理方法、プログラム |
| US11250336B2 (en) * | 2017-12-28 | 2022-02-15 | Intel Corporation | Distributed and contextualized artificial intelligence inference service |
| KR102075791B1 (ko) * | 2019-04-10 | 2020-03-02 | 주식회사 와이드티엔에스 | 에지 컴퓨팅에서 IoT 디바이스 연동을 이용한 패스트 데이터 처리 시스템 |
| CN110166973A (zh) * | 2019-05-29 | 2019-08-23 | 北京工业大学 | 一种基于nb-iot技术和边缘计算的超高层施工过程结构健康智能监测系统 |
| US20210174163A1 (en) * | 2019-12-10 | 2021-06-10 | International Business Machines Corporation | Edge inference for artifical intelligence (ai) models |
| CN111445026B (zh) * | 2020-03-16 | 2023-08-22 | 东南大学 | 面向边缘智能应用的深度神经网络多路径推理加速方法 |
-
2020
- 2020-09-25 US US18/027,604 patent/US20250292120A1/en active Pending
- 2020-09-25 JP JP2022551064A patent/JP7452679B2/ja active Active
- 2020-09-25 EP EP20955250.4A patent/EP4220498B8/en active Active
- 2020-09-25 WO PCT/JP2020/036394 patent/WO2022064656A1/ja not_active Ceased
- 2020-09-25 CN CN202080105384.7A patent/CN116194933B/zh active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2008309945A (ja) * | 2007-06-13 | 2008-12-25 | Kddi Corp | パターンマッチング方法および装置ならびにその特徴量正規化方法および装置 |
| JP2017126158A (ja) * | 2016-01-13 | 2017-07-20 | 日本電信電話株式会社 | 2値分類学習装置、2値分類装置、方法、及びプログラム |
| JP2018045369A (ja) * | 2016-09-13 | 2018-03-22 | 株式会社東芝 | 認識装置、認識システム、認識方法およびプログラム |
| WO2019220620A1 (ja) * | 2018-05-18 | 2019-11-21 | 日本電気株式会社 | 異常検出装置、異常検出方法及びプログラム |
| JP2020065225A (ja) * | 2018-10-19 | 2020-04-23 | 株式会社日立製作所 | 計算機、センシングシステム、及びデータ通信方法 |
Non-Patent Citations (4)
| Title |
|---|
| MORITA, KAZUNARI : "Use of Homomorphic Encryption for Secret Person Re-identification", IEICE TECHNICAL REPORT, vol. 117, no. 443, 12 February 2018 (2018-02-12), JP , pages 53 - 58, XP009535719, ISSN: 0913-5685 * |
| OKAYAMA YOSHIAKI: "Overview of Edge with the Cloud in Support of IoT", NIKKEI SYSTEMS, no. 302, 26 May 2018 (2018-05-26), JP , pages 50 - 53, XP009543447, ISSN: 1881-1620 * |
| OKOSHI ET AL.: "Proposal and Evaluation of DNN Model Operation Method with Cloud/Edge Collaboration", PROCEEDINGS OF THE 80TH NATIONAL CONVENTION, no. 1, 13 March 2018 (2018-03-13), pages 3 - 4 |
| See also references of EP4220498A4 |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2024029347A1 (ja) * | 2022-08-04 | 2024-02-08 | ソニーセミコンダクタソリューションズ株式会社 | 情報処理装置、情報処理方法、及び情報処理システム |
| WO2024057578A1 (ja) * | 2022-09-12 | 2024-03-21 | 日本電信電話株式会社 | 抽出システム、抽出方法および抽出プログラム |
| WO2024057374A1 (ja) * | 2022-09-12 | 2024-03-21 | 日本電信電話株式会社 | 抽出システム、抽出方法および抽出プログラム |
| WO2024079902A1 (ja) * | 2022-10-14 | 2024-04-18 | 日本電気株式会社 | 処理制御システム、処理制御装置、および処理制御方法 |
| CN115984166A (zh) * | 2022-11-04 | 2023-04-18 | 慧视云创(北京)科技有限公司 | 一种边云协同的图像变化检测方法及系统 |
Also Published As
| Publication number | Publication date |
|---|---|
| EP4220498A1 (en) | 2023-08-02 |
| EP4220498B1 (en) | 2025-09-03 |
| CN116194933A (zh) | 2023-05-30 |
| EP4220498B8 (en) | 2025-10-15 |
| JP7452679B2 (ja) | 2024-03-19 |
| CN116194933B (zh) | 2026-01-06 |
| JPWO2022064656A1 (ja) | 2022-03-31 |
| EP4220498A4 (en) | 2024-06-19 |
| US20250292120A1 (en) | 2025-09-18 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2022064656A1 (ja) | 処理システム、処理方法及び処理プログラム | |
| CN111488985B (zh) | 深度神经网络模型压缩训练方法、装置、设备、介质 | |
| US20190156213A1 (en) | Gradient compressing apparatus, gradient compressing method, and non-transitory computer readable medium | |
| CN112001914A (zh) | 深度图像补全的方法和装置 | |
| US10929676B2 (en) | Video recognition using multiple modalities | |
| CN113505848B (zh) | 模型训练方法和装置 | |
| JP2023512570A (ja) | 画像処理方法および関連装置 | |
| WO2022246986A1 (zh) | 数据处理方法、装置、设备及计算机可读存储介质 | |
| WO2023231794A1 (zh) | 一种神经网络参数量化方法和装置 | |
| CN118503685B (zh) | 基于设备属性和被动流量特征的设备指纹提取方法和系统 | |
| CN115019057A (zh) | 图像特征提取模型确定方法及装置、图像识别方法及装置 | |
| CN113343979B (zh) | 用于训练模型的方法、装置、设备、介质和程序产品 | |
| CN115147895A (zh) | 人脸鉴伪方法、装置及计算机程序产品 | |
| WO2025148920A1 (zh) | 三维图像生成方法、装置、电子设备及介质 | |
| JP7552903B2 (ja) | 処理方法、処理システム及び処理プログラム | |
| JP7546254B2 (ja) | 処理システム、処理方法及び処理プログラム | |
| CN113139463A (zh) | 用于训练模型的方法、装置、设备、介质和程序产品 | |
| WO2021232708A1 (zh) | 一种图像处理方法及电子设备 | |
| CN114756425B (zh) | 智能监控方法、装置、电子设备及计算机可读存储介质 | |
| CN114065913B (zh) | 模型量化方法、装置及终端设备 | |
| CN112966600B (zh) | 用于拥挤人群计数的自适应多尺度上下文聚合方法 | |
| US20240338939A1 (en) | Processing method, and processing system | |
| CN116704445A (zh) | 一种堆填土图像的识别方法、装置、电子设备及存储介质 | |
| CN118710692B (zh) | 图像生成方法、模型训练方法、装置及电子设备 | |
| CN114463584B (zh) | 图像处理、模型训练方法、装置、设备、存储介质及程序 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 20955250 Country of ref document: EP Kind code of ref document: A1 |
|
| ENP | Entry into the national phase |
Ref document number: 2022551064 Country of ref document: JP Kind code of ref document: A |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| ENP | Entry into the national phase |
Ref document number: 2020955250 Country of ref document: EP Effective date: 20230425 |
|
| WWG | Wipo information: grant in national office |
Ref document number: 2020955250 Country of ref document: EP |
|
| WWP | Wipo information: published in national office |
Ref document number: 18027604 Country of ref document: US |
