WO2021138783A1 - Data processing method and apparatus, and computer readable storage medium - Google Patents
Data processing method and apparatus, and computer readable storage medium Download PDFInfo
- Publication number
- WO2021138783A1 WO2021138783A1 PCT/CN2020/070549 CN2020070549W WO2021138783A1 WO 2021138783 A1 WO2021138783 A1 WO 2021138783A1 CN 2020070549 W CN2020070549 W CN 2020070549W WO 2021138783 A1 WO2021138783 A1 WO 2021138783A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- data set
- training
- training model
- model
- data processing
- Prior art date
Links
- 238000003672 processing method Methods 0.000 title claims abstract description 15
- 238000012549 training Methods 0.000 claims abstract description 237
- 238000000034 method Methods 0.000 claims abstract description 48
- 238000012545 processing Methods 0.000 claims description 136
- 238000002372 labelling Methods 0.000 claims description 39
- 238000011156 evaluation Methods 0.000 claims description 32
- 238000012360 testing method Methods 0.000 claims description 22
- 238000012790 confirmation Methods 0.000 claims description 12
- 238000004590 computer program Methods 0.000 claims description 7
- 238000013473 artificial intelligence Methods 0.000 description 16
- 238000010586 diagram Methods 0.000 description 15
- 230000008569 process Effects 0.000 description 8
- 238000013480 data collection Methods 0.000 description 7
- 230000011218 segmentation Effects 0.000 description 6
- 241000282472 Canis lupus familiaris Species 0.000 description 5
- 238000004891 communication Methods 0.000 description 5
- 241000282326 Felis catus Species 0.000 description 4
- 238000013527 convolutional neural network Methods 0.000 description 4
- 238000001514 detection method Methods 0.000 description 4
- 230000009471 action Effects 0.000 description 3
- 230000006870 function Effects 0.000 description 3
- 230000003993 interaction Effects 0.000 description 3
- 238000005516 engineering process Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 238000013528 artificial neural network Methods 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000010801 machine learning Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 238000005457 optimization Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000008439 repair process Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
- 230000011664 signaling Effects 0.000 description 1
- 230000000007 visual effect Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR 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; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Definitions
- This application relates to the field of artificial intelligence technology, in particular to a data processing method and device, and a computer-readable storage medium.
- Artificial Intelligence is a new technological science that studies and develops theories, methods, technologies and application systems for simulating, extending and expanding human intelligence.
- the research fields of artificial intelligence include robotics, language recognition, image recognition, natural speech processing, and expert systems.
- the realization of artificial intelligence mostly relies on training models, which are models generated through independent learning on a large amount of training data.
- the type of data may include, but is not limited to, images, videos, audios, objects, texts, and so on.
- the embodiments of the present application disclose a data processing method and device, and a computer-readable storage medium, which can efficiently and conveniently generate and deploy a training model.
- an embodiment of the present application provides a data processing method, which includes:
- an embodiment of the present application provides a data processing device, the data processing device includes: an input unit, a processing unit, and an output unit;
- the input unit is used to obtain a first data set for the target item
- the processing unit is configured to generate a first labeled data set corresponding to the first data set according to the identification purpose of the target item; perform model training on the first labeled data set to generate a first training model;
- the output unit is configured to output the first training model when the processing unit determines that the accuracy information of the first training model satisfies a deployment condition.
- an embodiment of the present application provides a data processing device, including a processor and a memory, the memory is used to store computer instructions, and when the processor executes the computer instructions, the data processing device Perform the method described in the first aspect above.
- an embodiment of the present application provides a computer-readable storage medium that stores a computer program that can implement the method described in the first aspect when the computer program is executed by a processor.
- data collection can be achieved by acquiring the first data set for the target project; data labeling can be realized by generating the first labeled data set; model generation can be realized by performing model training on the first labeled data set ; Model deployment can be realized by judging whether the training model meets the deployment conditions; thus, the training model can be generated and deployed efficiently, conveniently and flexibly.
- Managing the series of processes of data collection, data labeling, model generation and model deployment in the form of projects can improve the convenience and feasibility of training model applications.
- FIG. 1 is a schematic diagram of a network architecture provided by an embodiment of the present application.
- FIG. 2 is a schematic flowchart of a data processing method provided by an embodiment of the present application.
- FIG. 3 is a schematic diagram of a scene of a data processing method provided by an embodiment of the present application.
- Fig. 4 is a schematic diagram of an interface for creating a project provided by an embodiment of the present application.
- FIG. 5 is a schematic diagram of an interface of a certain plan provided by an embodiment of the present application.
- FIG. 6 is a schematic structural diagram of a data processing device provided by an embodiment of the present application.
- Fig. 7 is a schematic structural diagram of a terminal device disclosed in an embodiment of the present application.
- the training model refers to the model obtained by adaptively adjusting the selected algorithm (or called autonomous learning) through a large amount of training data.
- the training model can be applied to machine learning, language recognition, image recognition and other fields.
- the training model is applied to the field of image recognition, which can realize the recognition of cats or dogs in the image.
- a data set refers to an unlabeled data set, which can include one or more unlabeled data.
- Unlabeled data means that there is no trace of labeling on the data. For example, there is no mark on the image.
- Annotated data set refers to an annotated data set, which can include one or more annotated data.
- Labeled data that is, there are traces of labeling on the data. For example, there are marking traces on the image.
- the data is introduced by taking an image as an example.
- FIG. 1 is a schematic diagram of a network architecture provided by an embodiment of this application.
- the network architecture scene may include a user, a data processing device 101, and an application device 102.
- the data processing device 101 may be a terminal device, or a device matched with the terminal device, such as a processor.
- Terminal devices may include, but are not limited to, personal computers (PC), notebook computers, tablet computers, smart phones (such as Android phones, etc.), mobile Internet devices (Mobile Internet Devices, MID), and so on.
- the data processing device 101 has the ability to generate a training model.
- the application device 102 may include, but is not limited to, robots, aircraft, automobiles, smart home appliances, wearable devices, virtual reality (VR) devices, surveillance camera devices, smart phones, tablet computers, MIDs, PCs and other devices.
- the application device 102 has the ability to deploy a training model.
- the data processing device 101 shown in FIG. 1 takes a PC as an example, and the application equipment 102 takes a smart robot, a car, and an aircraft as examples.
- the shape and quantity of each device are for example, and do not constitute a limitation to the embodiment of the present application.
- the data processing device 101 can create a project according to the project creation instruction input by the user.
- the project can be an artificial intelligence (AI) project.
- AI artificial intelligence
- the AI project can be used to achieve a certain purpose, for example, to achieve image matching. Recognition of cats or dogs, etc.
- the data processing device 101 may generate a labeled data set corresponding to the data set according to the identification purpose of the item, perform model training on the labeled data set, and generate a training model.
- the data processing device 101 can output the training model, and can output the training model to the application device 102 to deploy the training model on the application device 102, so that the application device 102 can achieve the identification purpose .
- the data processing device 101 may also deploy the training model by itself, that is, deploy the training model on the data processing device 101, so that the data processing device 101 can achieve the identification purpose.
- the data processing device 101 may also output the training model to a third-party platform, so as to deploy the training model on the third-party platform.
- the third-party platform can be an image recognition platform or an AI visual recognition platform, etc.
- the data processing device 101 receives the project creation instruction to create the AI item 1, it creates the AI item 1, and the AI item 1 is used to recognize the puppy A.
- the data processing device 101 receives an image set for AI item 1 (for example, including 20 images), it generates an annotated image set corresponding to the image set according to the identified puppy A, and an image of puppy A is attached to the annotated image set. Is marked (for example, puppy A in a certain image is circled by a dotted line).
- the data processing device 101 performs model training on the labeled image set, and generates a training model. If the training model meets the deployment conditions, the data processing device 101 can output the training model to the surveillance camera equipment, so that the surveillance camera equipment can realize the Recognition of dog A.
- the embodiments of the present application can be applied to research and development scenarios, test scenarios, and usage scenarios, so that the embodiments of the present application have a wide range of applications. Even users who do not know the training model can also use the embodiments of the present application to control the application device 102 or the data processing device 101 to realize the identification of the designated item.
- FIG. 2 is a schematic flowchart of a data processing method provided by an embodiment of this application. As shown in Figure 2, the data processing method includes but is not limited to the following steps:
- Step 201 The data processing device obtains the first data set of the target item.
- the target item can be any AI item, for example, an AI item for recognizing human faces, an AI item for recognizing specific creatures, or an AI item for users to recognize parts defects. Different items can have different identification purposes.
- the user can input a project creation instruction for the user display interface of the data processing device, and the data processing device can create a project according to the project creation instruction.
- the first data set of the target item can be understood as a data set used to achieve the identification purpose of the target item.
- the specific number is not limited in the embodiment of the present application, for example, 20 images.
- the first data set may be an initial data set, that is, a data set initially obtained, or a data set after subsequent adjustments.
- the data processing device may receive a data upload instruction input by the user for the target project, and the data upload instruction may carry the first data set for uploading the first data set to the data processing Device, so that the data processing device obtains the first data set.
- the data processing device may obtain the first data set for the target item through the data collection device.
- the data collection device can be a camera of a terminal device or a surveillance camera device. For example, after taking a series of photos of puppy A, the camera transmits these photos of puppy A to the data processing device.
- one project can correspond to multiple plans, one plan can correspond to one data set, and one training model can be obtained by executing one plan.
- data set 1 is a data set of plan 1 for the target project, and execution plan 1 can get training model 1
- data set 2 is a data set of plan 2 for the target project, and execution plan 2 can get training model 2.
- Step 202 The data processing device generates a first annotation data set corresponding to the first data set according to the identification purpose of the target item.
- the identification purpose of the target item may include item type and label type.
- Item types can include, but are not limited to, target detection, object segmentation, etc.
- Target detection is used to detect targets, such as detecting pets or human faces
- object segmentation is used to segment or separate objects, such as combining people and scenes in an image. Perform segmentation, etc.
- Marking types can include, but are not limited to, internal chamfering, foreign matter, plane repair, etc. The marking types can be added or deleted or modified by the user, or they can be provided by the system.
- the first labeled data set refers to a data set in which the first data in the first data set is labeled.
- the number of the first annotated data set can be the same as the number of the first data set.
- the first data set includes 100 photos, and the first annotated data set also includes 100 photos.
- the 100 photos include photos of puppy A. Is marked.
- the data processing device may generate the first annotation data set by calling the annotation model. Specifically, according to the identification purpose of the target item, the data processing device calls the labeling model to perform labeling processing on the first data in the first data set to generate the first labeling data set.
- the labeling model is a training model that has been trained, meets the requirements, and can be used directly to implement labeling of data.
- the recognition purpose of the target item is to detect puppy A
- the data processing device calls the labeling model to label each image in 20 images (that is, the first data set) according to the purpose of detecting puppy A, and the image is labeled Puppy A in, thus get 20 annotated images (that is, the first annotated data set).
- Generating the first annotation data set by calling the annotation model can save workload and improve annotation efficiency.
- the data processing device may generate the first annotation data set through an annotation instruction. Specifically, the data processing device generates the first annotation data set according to the identification purpose of the target item and according to the annotation instruction for the first data in the first data set. That is, the user can input a labeling instruction for each first data in the first data set.
- the labeling instruction is related to the recognition purpose. If the recognition purpose is target detection, the labeling instruction can be to circle the object to be recognized, etc.; if the recognition purpose is an object For segmentation, the labeling instruction may be a dividing line for dividing the object. After the user confirms the labeling instruction of a certain data, the data processing device may save the labelled data, and then generate the first labeling data set.
- the data processing device For example, if the purpose of identifying the target item is to detect puppy A, then the data processing device according to the purpose of detecting puppy A, according to the user's annotation instructions for each image in the 20 images (ie the first data set), save the annotations Of 20 images (that is, the first annotated data set).
- the accuracy of the first annotation data set is higher by the way of labeling instructions than by calling the labeling model.
- the data processing device may also generate the first annotation data set by calling the annotation model and combining the annotation instructions. For example, according to the identification purpose of the target item, the data processing device first calls the labeling model to label the first data in the first data set, and then adjusts according to the labeling instructions input by the user, and finally generates the first labeling data set. This combination method can further improve the accuracy of the first annotation data set.
- the user can check the first labeled data set to ensure the accuracy of the first labeled data set.
- the user can input at least one confirmation instruction for the first annotated data set. If the data processing device receives at least one confirmation instruction, it can be determined that the first annotated data set meets the training condition, and step 203 can be executed.
- each of the at least one confirmation instruction can be from a user with different authority, for example, two confirmation instructions, one from the user operating the terminal device, one from the system inspector or administrator; or the first one from the user with the first authority
- the second time comes from a user with the second authority, the level of the second authority is higher than the first authority.
- the data processing device may count the number of marked data, the number of invalid data, and the number of marked objects in the first marked data set.
- the number of labeled data refers to how many data are labeled, for example, how many images out of 100 images are labeled.
- the amount of invalid data refers to the amount of data that is not related to the purpose of recognition. For example, the purpose of recognition is to detect puppy A, but an image is an image of a building, and the image has nothing to do with puppy A. Then the image It can be considered invalid data.
- the number of labeled objects refers to how many objects are labeled in the data. For example, if the recognition purpose is object segmentation, then the number of labeled objects is at least two. The data processing device can also output these statistical results for the user to make corresponding adjustments.
- Step 203 The data processing device performs model training on the first labeled data set to generate a first training model.
- the data processing device calls the target training model to perform the training on the first labeled data set.
- Model training to generate the first training model can be understood as an initial model without any training data input, which is used to train the input training data.
- the target training model can also be described as an artificial intelligence model or a deterministic model.
- the target training model can be a linear model, a convolutional neural network model, or a recurrent neural network model.
- the data processing device inputs the first labeled data set as training data into the target training model to perform model training, thereby generating the first training model.
- the target training model is a convolutional neural network model
- the data processing device inputs 20 images labeled puppy A into the convolutional neural network model.
- the convolutional neural network model can be trained through the model to obtain the first training model. Once the training model meets the deployment conditions, the deployment of the first training model can realize the recognition of puppy A.
- the data processing device may perform model training based on the existing training model to generate the first training model.
- the existing training model is the training model BB
- the data processing device can obtain the first data set based on the accuracy information of the training model BB, generate the first labeled data set, and input the first labeled data set into the training model BB to generate the first data set. Train the model. Since the first data set is acquired based on the training model BB, the accuracy of the first training model is higher than that of the training model BB.
- the data processing device may obtain a data processing server in an idle state, and call the data processing server to perform model training to generate the first training model.
- the data processing server may include, but is not limited to, a graphics processing (Graphics Processing Unit, GPU) server, a text processing server, a central processing (CPU) server, an application-specific integrated circuit (ASIC) server, Tensor Processing Unit (TPU) server, Neural Network Processing Unit (NPU) server, Field-programmable Gate Array (FPGA) server, etc.
- the data processing server may be a GPU server, and the embodiment of the present application takes the GPU server as an example for description.
- the GPU server is a computing service based on graphics applications. It has real-time and high-speed parallel computing and floating-point computing capabilities. It is suitable for application scenarios such as 3D graphics applications, video decoding, deep learning, and scientific computing.
- the GPU server can be located in the Internet cloud or It may be mounted in a data processing device in the form of a GPU processor.
- Step 204 The data processing device outputs the first training model when the accuracy information of the first training model meets the deployment condition.
- the data processing device When the data processing device obtains the first training model, it can judge whether the accuracy information of the first training model meets the deployment condition by referring to the data set. Specifically, the data processing device calls the first training model to test the reference data set, obtains the first test result, and compares the first test result with the reference annotation result corresponding to the reference data set to obtain the accuracy information of the first training model . The data processing device calls the first training model to test the reference data set, that is, the reference data set is input to the first training model for target detection or object segmentation.
- the reference data set can also be described as a test data set, etc.
- the reference data set is a data set carefully selected by the user to meet the identification purpose of the target item, and is an unlabeled data set.
- the first test result (assumed to be R1) refers to the labeling result output by the first training model after the reference data set is input to the first training model.
- the number of data included in the first test result can be the same as the number of data included in the reference data set the same.
- the reference labeling result (assuming R0) can also be described as a test labeling data set, etc.
- the reference labeling result is a labeling data set that the user manually annotates the reference data set and meets the identification purpose of the target item.
- the result of reference labeling has been confirmed many times, and its labeling accuracy is high, and it is used to detect whether the generated training model meets the deployment conditions. It is understandable that the reference data set is the test paper, and the reference marked result is the reference marked answer of the test paper.
- the data processing device compares the first test result (R1) with the reference labeling result (R0) one by one to obtain accuracy information of the first training model.
- the accuracy information may include the overall accuracy and the accuracy of the judgment item.
- the overall accuracy refers to the accuracy of the judgment of the first training model from the overall perspective.
- the accuracy of the evaluation item refers to the accuracy of the first training model from the perspective of a specific index.
- the specific index can be one or more, and a specific index corresponds to the accuracy of a judgment item.
- Specific indicators may include, but are not limited to: recognition type, misrecognition, missed recognition, wrong recognition, multiple recognition, etc.
- the accuracy of the judgment item corresponding to the recognition type represents the recognition accuracy of the recognition type.
- the purpose of the recognition is to identify cats and dogs, then there are two types of recognition, the accuracy of the judgment item corresponding to the recognition of the cat and the accuracy of the judgment item corresponding to the recognition of the dog degree.
- Misrecognition means that compared with R0, there are cases in R1 that should not be marked.
- the purpose of recognition is to identify kitten B. In R0, kitten B is marked, but the image of kitten C in R1 is regarded as If the kitten B is labeled, it is considered that the image in R1 is misrecognized.
- Missing recognition means that compared with R0, there is an unlabeled situation in R1 that should be labeled.
- the purpose of recognition is to identify kitten B. Kitten B is labeled in R0, but there is a certain image in R1 that includes a kitten. B. If the kitten B in the image is not labeled, it is considered that the image in R1 has missed recognition.
- Wrong recognition means that compared with R0, there is a wrong recognition type in R1. For example, the purpose of recognition is to recognize kitten B.
- Kitten B is labeled in R0, but the image of kitten B is labeled as puppy in R1. A, it is considered that the image in R1 is misidentified.
- Multiple recognition means that for a certain image, compared with R0, the recognition results of the same type in R1 are more than the recognition results of that type in R0.
- the purpose of recognition is to recognize puppy A, and the image in R0 A total of 1 puppy A is marked, but 2 puppy A is marked in R1, it is considered that there is multiple recognition in R1.
- the higher the accuracy of the evaluation item corresponding to multiple recognition the lower the probability of multiple recognition.
- the overall accuracy is related to the accuracy of the judgment item.
- the overall accuracy is the weighted average of the accuracy of each evaluation item, or the average of the accuracy of each evaluation item.
- the specific indicators include misrecognition and missed recognition, and the corresponding evaluation item accuracy is 80% and 90%. If the overall accuracy is the average of the accuracy of each evaluation item, then the overall accuracy is 85%.
- the overall accuracy has nothing to do with the accuracy of the judgment item.
- the overall accuracy is used to describe how much data is labeled.
- the reference data set is 100 images, and each of these 100 images includes puppy A. Input these 100 images into the first training model to get R1. There are 87 images of puppy A in R1. In the reference data set R0 corresponding to the reference data set, the puppy A in 100 images is labeled, so the overall accuracy of the first training model is 87%.
- the data processing device may output the accuracy information of the first training model so that the user can learn the accuracy information of the first training model and collect the second data set in a targeted manner.
- the accuracy information of the first training model that satisfies the deployment conditions may include:
- the overall accuracy of the first training model is greater than the first threshold, and the percentage range of the first threshold is (0, 100]. For example, assuming that the first threshold is 85%, the overall accuracy of the first training model is greater than At 85%, the accuracy information of the first training model meets the deployment conditions.
- the accuracy of the evaluation item of the first training model is greater than the second threshold, and the percentage of the second threshold is in the range (0, 100].
- the evaluation item includes evaluation item 1 and evaluation item 2, assuming that the second threshold is 80%, when the accuracy of the first training model is greater than 80%, the accuracy information of the first training model meets the deployment conditions.
- Each evaluation item can correspond to the same second threshold , It can also correspond to different second thresholds. For example, the second threshold corresponding to evaluation item 1 is 80%, and the second threshold corresponding to evaluation item 2 is 85%.
- the overall accuracy of the first training model is greater than the first threshold, and the accuracy of the judgment item of the first training model is greater than the second threshold. For example, assuming that the first threshold is 85%, the evaluation items include evaluation item 1 and evaluation item 2, and the second threshold value corresponding to evaluation item 1 and evaluation item 2 is 80%, then the overall accuracy of the first training model is greater than 85%. And when the accuracy rates of the evaluation items corresponding to the evaluation item 1 and the evaluation item 2 of the first training model are both greater than 80%, the accuracy information of the first training model meets the deployment conditions.
- the specific numerical values of the above-mentioned first threshold and the second threshold may be set by the user or set by default by the system, and the specific numerical values are not limited in the embodiment of the present application.
- the data processing device outputs the first training model when the accuracy information of the first training model meets the deployment condition.
- the first training model is transmitted to the application device, so that the application device can deploy the first training model to achieve the target project.
- the first training model is deployed on the data processing device, so that the data processing device realizes the target project.
- step 201 can implement data collection
- step 202 can implement data labeling
- step 203 can implement model generation
- step 204 can implement model deployment, so that training models can be generated and deployed efficiently, conveniently and flexibly.
- Steps 201 to 204 the series of processes of data collection, data labeling, model generation, and model deployment are managed in the form of projects, which can improve the convenience and feasibility of training model applications.
- the data processing device can obtain the second data set for the target project according to the accuracy information of the first training model, that is, according to the accuracy of the first training model Information, conduct targeted data collection, and the obtained data set is the second data set.
- the accuracy of the evaluation item 1 is that the accuracy of face recognition is 30%
- the accuracy of the evaluation item 2 is that the accuracy of face recognition is 90%
- the accuracy of the evaluation item 1 is not reached.
- the second threshold 80%
- face profile pictures can be collected in a targeted manner to obtain a second data set.
- the data processing device In the case of obtaining the second data set, the data processing device generates a second labeled data set corresponding to the second data set according to the identification purpose of the target item, performs model training on the second labeled data set, and generates a second training model. Second, when the accuracy information of the training model meets the deployment conditions, the second training model is output.
- the data processing device may call the first training model, and perform model training on the second labeled data set to generate the second training model. That is, the second labeled data set is input into the first training model for model training, and the second training model is obtained.
- the data processing device invokes the second training model to test the reference data set, obtains the second test result, compares the second test result with the reference labeling result, and obtains accuracy information of the second training model.
- the data processing device may obtain a third data set for the target item according to the accuracy information of the second training model, and generate a third annotation corresponding to the third data set
- model training is performed on the third labeled data set to generate a third training model
- the third training model is output. It is understandable that when the i-th training model does not meet the deployment conditions, the i+1-th training model is generated according to the i-th training model, and the reference labeling result is used to test whether the i+1-th training model meets the deployment conditions. , And repeat this until a training model that satisfies the deployment conditions is obtained.
- the data processing device can compare the accuracy information of the multiple training models and output a better training model; or the user can compare the accuracy information of the multiple training models choose a better training model to deploy a better training model.
- FIG. 3 is a schematic diagram of a scene of a data processing method provided by an embodiment of this application.
- Figure 3 introduces from the perspective of the interaction between the user and the data processing device, which may include but is not limited to the following steps:
- Step 301 The first user inputs a project creation instruction.
- the first user is a user who operates the data processing device, and the second user is a system inspector or an administrator; or the authority level of the second user is higher than the authority level of the first user.
- the first user can enter the project name, select the project type, upload the template picture, enter the height and width of the picture, and enter or select the label category (or called It is an annotation type), after the first user completes these operations, these operations can be confirmed.
- the first user's confirmation operation of these operations can be understood as inputting an item creation instruction to the data processing device.
- FIG. 4 the schematic diagram of the interface for creating a project shown in FIG. 4 is used as an example, and does not constitute a limitation to the embodiment of the present application.
- Step 302 The data processing device creates a project.
- the data processing device can create the project according to the information carried in the project creation instruction.
- the information carried in the project creation instruction may include the information input or selected by the first user in the interface schematic shown in FIG. 4, for example, including the project name, project type, label category, and so on.
- Step 303 The first user inputs a plan creation instruction.
- the first user can input a plan creation instruction for the project. If the data processing device has not created a plan for the project, the plan creation instruction is used to create the first plan; if the data processing device has created a plan for the project, the plan creation instruction is used to create a new plan.
- the plan creation instruction can carry the plan name, and the first user can name each plan independently. For example, in Figure 5, the plan name is "demo".
- FIG. 5 Refer to the schematic interface diagram of a certain plan shown in FIG. 5, where the first user can click Edit Plan Category under Label Category to edit, modify or delete the label category of the plan. It should be noted that the schematic interface diagram shown in FIG. 5 is used as an example, and does not constitute a limitation to the embodiment of the present application.
- Step 304 The data processing device creates a first plan.
- the first user can execute the plan steps in sequence according to the plan steps in the interface diagram shown in FIG. 5.
- Step 305 The first user inputs a data upload instruction.
- the first user performs the step of uploading pictures, clicks on upload pictures, and selects pictures to be uploaded.
- the pictures to be uploaded can be uploaded in the form of a compressed package.
- the data processing device can receive the compressed package.
- the pictures to be uploaded are pictures related to the purpose of the project.
- Step 306 The data processing device obtains the first data set.
- the data processing device decompresses the compressed package to obtain the first data set.
- the set of pictures uploaded by the first user for the first plan is the first data set
- the set of pictures uploaded for the second plan is the second data set.
- the data processing device may sequentially output each picture in the first data set, and the first user executes the labeling step to manually label each output picture.
- the data processing device can call the labeling model to label each picture.
- manual + labeling model combination to label each picture.
- Step 307 The data processing apparatus generates a first annotation data set corresponding to the first data set.
- the data processing device may generate the first annotation data set.
- Step 308a the first user inputs a confirmation instruction.
- Step 308b the second user inputs a confirmation instruction.
- the first user performs a check step to check whether the images marked in the first annotated data set meet the annotation category, whether the annotation position is correct, and so on. If the first user detects an error, he can modify the error and update the first annotation data set. If the first user checks that all the labels are correct, the first user inputs a confirmation instruction.
- the second user may check the first annotation data set confirmed by the first user again after the first user has checked and confirmed, and input a confirmation instruction after the check is correct.
- the data processing apparatus may perform step 309.
- Step 309 The data processing device performs model training on the first labeled data set to generate a first training model.
- Step 310 When the accuracy information of the first training model meets the deployment condition, the data processing device outputs the first training model.
- step 307 For the specific implementation process of step 307, step 309, and step 310, please refer to the specific description of step 202 to step 204 in the embodiment shown in FIG. 2, which will not be repeated here.
- the first user performs step 303 again to input a plan creation instruction, which is used to create a second plan.
- the first user can upload the second data set based on the accuracy information of the first training model, the data processing device generates the second labeled data set, and after the first user and the second user confirm the second labeled data set, the data processing device will The second data set is input to the first training model for model training to generate a second training model, and when the accuracy information of the second training model meets the deployment conditions, the second training model is output.
- the introduction is introduced from the perspective of the interaction between the user and the data processing device. Even if the first user is not familiar with the training model, the data processing device can be controlled to output the training model that meets the deployment conditions, so that the first user can Autonomously manage the training model to achieve the purpose of the project.
- FIG. 6 is a schematic structural diagram of a data processing device provided by an embodiment of this application.
- the data processing device 60 includes an input unit 601, a processing unit 602, and an output unit 603.
- the input unit 601 is used to obtain the first data set for the target item.
- the processing unit 602 is configured to generate a first labeled data set corresponding to the first data set according to the identification purpose of the target item; perform model training on the first labeled data set to generate a first training model.
- the output unit 603 is configured to output the first training model when the processing unit 602 determines that the accuracy information of the first training model satisfies the deployment condition.
- the processing unit 602 is specifically configured to generate a first labeled data set corresponding to the first data set according to a labeling instruction for the first data in the first data set according to the identification purpose of the target item.
- the processing unit 602 is specifically configured to call the annotation model to perform annotation processing on the first data in the first data set according to the identification purpose of the target item, and generate the first annotation data set corresponding to the first data set.
- the processing unit 602 is further configured to count the number of marked data, the number of invalid data, and the number of marked objects in the first marked data set.
- the processing unit 602 is specifically configured to call the target training model to perform model training on the first labeled data set, and generate the first training model.
- the processing unit 602 is specifically configured to obtain a data processing server in an idle state, call the data processing server to perform model training on the first labeled data set, and generate a first training model.
- the processing unit 602 is further configured to call the first training model to test the reference data set to obtain the test result; compare the test result with the reference annotation result corresponding to the reference data set to obtain the accuracy of the first training model information;
- the output unit 603 is also used to output accuracy information of the first training model.
- the accuracy information includes the overall accuracy and the accuracy of the evaluation item; the accuracy information that satisfies the deployment conditions includes: the overall accuracy is greater than the first threshold; or, the accuracy of the evaluation item is greater than the second threshold; or, the overall accuracy The degree is greater than the first threshold and the accuracy of the judgment item is greater than the second threshold.
- the data processing device 60 further includes a storage unit for storing the first training model and the accuracy information of the first training model.
- the input unit 601 is further configured to obtain information for the target project according to the accuracy information of the first training model when the processing unit 602 determines that the accuracy information of the first training model does not meet the deployment conditions.
- the processing unit 602 is further configured to generate a second labeled data set corresponding to the second data set according to the identification purpose of the target item; perform model training on the second labeled data set to generate a second training model;
- the output unit 603 is further configured to output the second training model when the processing unit 602 determines that the accuracy information of the second training model satisfies the deployment condition.
- the processing unit 602 is specifically configured to call the first training model, perform model training on the second labeled data set, and generate the second training model.
- FIG. 7 is a schematic structural diagram of a terminal device provided by an embodiment of this application.
- the terminal device described in the embodiment of the present application includes: a processor 701, a communication interface 702, and a memory 703.
- the processor 701, the communication interface 702, and the memory 703 may be connected through a bus or in other ways.
- the embodiment of the present application takes the connection through a bus as an example.
- the processor 701 may be a central processing unit (CPU), a network processor (Network Processor, NP), or a combination of a CPU and NP.
- the processor 701 may also be a multi-core CPU or a core used to implement communication identification binding in a multi-core NP.
- the processor 701 may be a hardware chip.
- the hardware chip may be an application-specific integrated circuit (ASIC), a programmable logic device (Programmable Logic Device, PLD), or a combination thereof.
- the PLD may be a Complex Programmable Logic Device (CPLD), a Field-Programmable Gate Array (FPGA), a Generic Array Logic (GAL) or any combination thereof.
- the communication interface 702 can be used for the interaction of sending and receiving information or signaling, as well as the reception and transmission of signals.
- the memory 703 may mainly include a storage program area and a storage data area.
- the storage program area may store an operating system and a stored program required by at least one function (such as a text storage function, a location storage function, etc.); the storage data area may store Data (such as image data, text data) created according to the use of the device, etc., and may include application storage programs, etc.
- the memory 703 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other volatile solid-state storage devices.
- the terminal device shown in FIG. 7 also includes input and output devices, which are used to receive input instructions or selection instructions from the user, and are also used to output accuracy information on the user interface.
- the memory 703 is also used to store program instructions.
- the processor 701 is configured to execute program instructions stored in the memory 703, and when the program instructions are executed, the processor 701 is configured to:
- the processor 701, the communication interface 702, and the memory 703 described in the embodiment of the present application can execute the implementation manner described in the data processing method provided in the embodiment of the present application, and details are not described herein again.
- An embodiment of the present application also provides a computer-readable storage medium, in which a computer program is stored, and the computer program is executed by a processor to implement the method described in the foregoing method embodiment.
- the embodiments of the present application also provide a computer program product containing instructions, which when run on a computer, cause the computer to execute the method described in the above method embodiment.
- the program can be stored in a computer-readable storage medium, and the storage medium can include: Flash disk, read-only memory (Read-Only Memory, ROM), random access device (Random Access Memory, RAM), magnetic disk or optical disk, etc.
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Computing Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- General Health & Medical Sciences (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Medical Informatics (AREA)
- Evolutionary Biology (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Image Analysis (AREA)
Abstract
Description
Claims (16)
- 一种数据处理方法,其特征在于,所述方法包括:A data processing method, characterized in that the method includes:获取针对目标项目的第一数据集;Obtain the first data set for the target project;按照所述目标项目的识别目的生成所述第一数据集对应的第一标注数据集;Generating a first annotation data set corresponding to the first data set according to the identification purpose of the target item;对所述第一标注数据集进行模型训练,生成第一训练模型;Performing model training on the first labeled data set to generate a first training model;在所述第一训练模型的准确度信息满足部署条件的情况下,输出所述第一训练模型。When the accuracy information of the first training model satisfies the deployment condition, output the first training model.
- 根据权利要求1所述的方法,其特征在于,所述按照所述目标项目的识别目的生成所述第一数据集对应的第一标注数据集,包括:The method according to claim 1, wherein the generating a first annotation data set corresponding to the first data set according to the identification purpose of the target item comprises:按照所述目标项目的识别目的,根据针对所述第一数据集中的第一数据的标注指令,生成所述第一数据集对应的第一标注数据集。According to the identification purpose of the target item, a first annotation data set corresponding to the first data set is generated according to an annotation instruction for the first data in the first data set.
- 根据权利要求1所述的方法,其特征在于,所述按照所述目标项目的识别目的生成所述第一数据集对应的第一标注数据集,包括:The method according to claim 1, wherein the generating a first annotation data set corresponding to the first data set according to the identification purpose of the target item comprises:按照所述目标项目的识别目的,调用标注模型对所述第一数据集中的第一数据进行标注处理,生成所述第一数据集对应的第一标注数据集。According to the identification purpose of the target item, a labeling model is invoked to label the first data in the first data set to generate a first labeling data set corresponding to the first data set.
- 根据权利要求1-3任一项所述的方法,其特征在于,所述方法还包括:The method according to any one of claims 1-3, wherein the method further comprises:统计所述第一标注数据集中已标注的数据数量、无效数据的数量和标注的对象数量。Count the number of marked data, the number of invalid data, and the number of marked objects in the first marked data set.
- 根据权利要求1-3任一项所述的方法,其特征在于,所述对所述第一标注数据集进行模型训练,生成第一训练模型,包括:The method according to any one of claims 1 to 3, wherein the performing model training on the first labeled data set to generate a first training model comprises:调用目标训练模型对所述第一标注数据集进行模型训练,生成第一训练模型。The target training model is called to perform model training on the first labeled data set, and a first training model is generated.
- 根据权利要求1-3任一项所述的方法,其特征在于,所述对所述第一标注数据集进行模型训练,生成第一训练模型,包括:The method according to any one of claims 1 to 3, wherein the performing model training on the first labeled data set to generate a first training model comprises:获取处于空闲态的数据处理服务器,调用所述数据处理服务器对所述第一标注数据集进行模型训练,生成第一训练模型。Obtain a data processing server in an idle state, call the data processing server to perform model training on the first labeled data set, and generate a first training model.
- 根据权利要求1所述的方法,其特征在于,所述方法还包括:The method according to claim 1, wherein the method further comprises:调用所述第一训练模型对参考数据集进行测试,得到测试结果;Call the first training model to test the reference data set, and obtain the test result;对比所述测试结果与所述参考数据集对应的参考标注结果,得到并输出所述第一训练模型的准确度信息。The test result is compared with the reference annotation result corresponding to the reference data set to obtain and output the accuracy information of the first training model.
- 根据权利要求7所述的方法,其特征在于,所述准确度信息包括整体准确度和评判项准确度;所述准确度信息满足所述部署条件包括:所述整体准确度大于第一阈值;或,所述评判项准确度大于第二阈值;或,所述整体准确度大于第一阈值且所述评判项准确度大于第二阈值。The method according to claim 7, wherein the accuracy information includes overall accuracy and evaluation item accuracy; the accuracy information satisfying the deployment condition includes: the overall accuracy is greater than a first threshold; Or, the accuracy of the evaluation item is greater than the second threshold; or, the overall accuracy is greater than the first threshold and the accuracy of the evaluation item is greater than the second threshold.
- 根据权利要求7或8所述的方法,其特征在于,所述方法还包括:The method according to claim 7 or 8, wherein the method further comprises:控制所述第一训练模型以及所述第一训练模型的准确度信息的保存。Controlling the storage of the first training model and the accuracy information of the first training model.
- 根据权利要求1或7所述的方法,其特征在于,所述方法还包括:The method according to claim 1 or 7, wherein the method further comprises:在所述第一训练模型的准确度信息不满足所述部署条件的情况下,根据所述第一训练模型的准确度信息,获取针对所述目标项目的第二数据集;In the case that the accuracy information of the first training model does not satisfy the deployment condition, acquiring a second data set for the target item according to the accuracy information of the first training model;按照所述目标项目的识别目的生成所述第二数据集对应的第二标注数据集;Generating a second annotation data set corresponding to the second data set according to the identification purpose of the target item;对所述第二标注数据集进行模型训练,生成第二训练模型;Performing model training on the second labeled data set to generate a second training model;在所述第二训练模型的准确度信息满足所述部署条件的情况下,输出所述第二训练模型。When the accuracy information of the second training model satisfies the deployment condition, output the second training model.
- 根据权利要求10所述的方法,其特征在于,所述对所述第二标注数据集进行模型训练,生成第二训练模型,包括:The method according to claim 10, wherein said performing model training on said second labeled data set to generate a second training model comprises:调用所述第一训练模型,对所述第二标注数据集进行模型训练,生成第二训练模型。Invoking the first training model, performing model training on the second labeled data set, and generating a second training model.
- 根据权利要求1所述的方法,其特征在于,所述方法还包括:The method according to claim 1, wherein the method further comprises:在所述第一标注数据集满足训练条件的情况下,执行对所述第一标注数据集进行模型训练,生成第一训练模型的步骤。In the case that the first labeled data set meets the training condition, the step of performing model training on the first labeled data set to generate a first training model is performed.
- 根据权利要求12所述的方法,其特征在于,所述方法还包括:The method according to claim 12, wherein the method further comprises:在接收到针对所述第一标注数据集的至少一次确认指令的情况下,确定所述第一标注数据集满足所述训练条件。In a case where at least one confirmation instruction for the first labeled data set is received, it is determined that the first labeled data set satisfies the training condition.
- 一种数据处理装置,其特征在于,所述数据处理装置包括用于执行如权利要求1-13任一项所述的各个步骤的单元。A data processing device, characterized in that the data processing device includes a unit for executing each step according to any one of claims 1-13.
- 一种数据处理装置,其特征在于,包括:存储器和处理器,A data processing device, characterized by comprising: a memory and a processor,所述存储器用于存储计算机指令;The memory is used to store computer instructions;当所述处理器执行所述计算机指令时,以使所述数据处理装置执行权利要求1-12任一项所述的方法。When the processor executes the computer instructions, the data processing device executes the method according to any one of claims 1-12.
- 一种计算机可读存储介质,所述计算机可读存储介质中存储有计算机程序,其特征在于:所述计算机程序被处理器执行时实现如权利要求1至12中任一项所述方法的步骤。A computer-readable storage medium in which a computer program is stored, characterized in that: when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 12 are implemented .
Priority Applications (2)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN202080005484.2A CN112805725A (en) | 2020-01-06 | 2020-01-06 | Data processing method and device and computer readable storage medium |
PCT/CN2020/070549 WO2021138783A1 (en) | 2020-01-06 | 2020-01-06 | Data processing method and apparatus, and computer readable storage medium |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
PCT/CN2020/070549 WO2021138783A1 (en) | 2020-01-06 | 2020-01-06 | Data processing method and apparatus, and computer readable storage medium |
Publications (1)
Publication Number | Publication Date |
---|---|
WO2021138783A1 true WO2021138783A1 (en) | 2021-07-15 |
Family
ID=75809347
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
PCT/CN2020/070549 WO2021138783A1 (en) | 2020-01-06 | 2020-01-06 | Data processing method and apparatus, and computer readable storage medium |
Country Status (2)
Country | Link |
---|---|
CN (1) | CN112805725A (en) |
WO (1) | WO2021138783A1 (en) |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN108875963A (en) * | 2018-06-28 | 2018-11-23 | 北京字节跳动网络技术有限公司 | Optimization method, device, terminal device and the storage medium of machine learning model |
CN109034188A (en) * | 2018-06-15 | 2018-12-18 | 北京金山云网络技术有限公司 | Acquisition methods, acquisition device, equipment and the storage medium of machine learning model |
CN109242013A (en) * | 2018-08-28 | 2019-01-18 | 北京九狐时代智能科技有限公司 | A kind of data mask method, device, electronic equipment and storage medium |
CN109583325A (en) * | 2018-11-12 | 2019-04-05 | 平安科技(深圳)有限公司 | Face samples pictures mask method, device, computer equipment and storage medium |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US11423324B2 (en) * | 2017-02-23 | 2022-08-23 | International Business Machines Corporation | Training and estimation of selection behavior of target |
-
2020
- 2020-01-06 CN CN202080005484.2A patent/CN112805725A/en active Pending
- 2020-01-06 WO PCT/CN2020/070549 patent/WO2021138783A1/en active Application Filing
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109034188A (en) * | 2018-06-15 | 2018-12-18 | 北京金山云网络技术有限公司 | Acquisition methods, acquisition device, equipment and the storage medium of machine learning model |
CN108875963A (en) * | 2018-06-28 | 2018-11-23 | 北京字节跳动网络技术有限公司 | Optimization method, device, terminal device and the storage medium of machine learning model |
CN109242013A (en) * | 2018-08-28 | 2019-01-18 | 北京九狐时代智能科技有限公司 | A kind of data mask method, device, electronic equipment and storage medium |
CN109583325A (en) * | 2018-11-12 | 2019-04-05 | 平安科技(深圳)有限公司 | Face samples pictures mask method, device, computer equipment and storage medium |
Also Published As
Publication number | Publication date |
---|---|
CN112805725A (en) | 2021-05-14 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN111797893B (en) | Neural network training method, image classification system and related equipment | |
US11030458B2 (en) | Generating synthetic digital assets for a virtual scene including a model of a real-world object | |
CN112183577A (en) | Training method of semi-supervised learning model, image processing method and equipment | |
CN111598164B (en) | Method, device, electronic equipment and storage medium for identifying attribute of target object | |
KR102548732B1 (en) | Apparatus and Method for learning a neural network | |
CN108983979B (en) | Gesture tracking recognition method and device and intelligent equipment | |
US20170185913A1 (en) | System and method for comparing training data with test data | |
CN105051755A (en) | Part and state detection for gesture recognition | |
WO2021217543A1 (en) | Image annotation method, apparatus, device and medium | |
CN109522970B (en) | Image classification method, device and system | |
WO2021184776A1 (en) | Image recognition method and apparatus, computer device and storage medium | |
KR20190081373A (en) | Terminal device and data processing method based on artificial neural network | |
WO2022247112A1 (en) | Task processing method and apparatus, device, storage medium, computer program, and program product | |
WO2020256732A1 (en) | Domain adaptation and fusion using task-irrelevant paired data in sequential form | |
US20220366244A1 (en) | Modeling Human Behavior in Work Environment Using Neural Networks | |
CN111695443A (en) | Intelligent traffic artificial intelligence open platform, method, medium and electronic device | |
CN111124863B (en) | Intelligent device performance testing method and device and intelligent device | |
CN115658523A (en) | Automatic control and test method for human-computer interaction interface and computer equipment | |
US20210312292A1 (en) | Methods and systems for operating applications through user interfaces | |
WO2021138783A1 (en) | Data processing method and apparatus, and computer readable storage medium | |
KR20230068989A (en) | Method and electronic device for performing learning of multi-task model | |
CN114037889A (en) | Image identification method and device, electronic equipment and storage medium | |
CN117593613B (en) | Multitasking learning method and device, storage medium and electronic equipment | |
CN112529038B (en) | Method and device for identifying main board material and storage medium | |
WO2024088031A1 (en) | Data acquisition method and apparatus, and related device |
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: 20911566 Country of ref document: EP Kind code of ref document: A1 |
|
NENP | Non-entry into the national phase |
Ref country code: DE |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20911566 Country of ref document: EP Kind code of ref document: A1 |
|
32PN | Ep: public notification in the ep bulletin as address of the adressee cannot be established |
Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 15/03/2023) |
|
122 | Ep: pct application non-entry in european phase |
Ref document number: 20911566 Country of ref document: EP Kind code of ref document: A1 |