WO2023011470A1 - 一种机器学习系统及模型训练方法 - Google Patents

一种机器学习系统及模型训练方法 Download PDF

Info

Publication number
WO2023011470A1
WO2023011470A1 PCT/CN2022/109701 CN2022109701W WO2023011470A1 WO 2023011470 A1 WO2023011470 A1 WO 2023011470A1 CN 2022109701 W CN2022109701 W CN 2022109701W WO 2023011470 A1 WO2023011470 A1 WO 2023011470A1
Authority
WO
WIPO (PCT)
Prior art keywords
module
machine learning
training
trained
target
Prior art date
Application number
PCT/CN2022/109701
Other languages
English (en)
French (fr)
Inventor
程战战
Original Assignee
上海高德威智能交通系统有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 上海高德威智能交通系统有限公司 filed Critical 上海高德威智能交通系统有限公司
Publication of WO2023011470A1 publication Critical patent/WO2023011470A1/zh

Links

Images

Classifications

    • 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
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • 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
    • G06N3/047—Probabilistic or stochastic 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
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Definitions

  • the present application relates to the technical field of machine learning, in particular to a machine learning system and a model training method.
  • machine learning technology is widely used in various technical fields, such as: video surveillance, behavior analysis, image processing and other technical fields.
  • the machine learning model is trained according to the training data of the target scene provided by the user during the development stage of the application software, and integrated into the specific application software for the user to use after training.
  • the machine learning model integrated in the application software adapts to different application scenarios, in related technologies, every time the target scene is changed, it needs to be retrained according to the training data of the target scene provided by the user. For example: for video surveillance software that integrates one or more machine learning models, if it is to be applied to a train station, it needs to be trained with video samples from the train station during the research and development stage; if it is to be applied to an airport, it needs to be trained with Video samples of airports for training. After the training is completed, the application software can be provided to the user.
  • the purpose of the embodiments of the present application is to provide a machine learning system and a model training method, so as to realize adaptive learning of real application scenarios.
  • the specific technical scheme is as follows:
  • a machine learning system includes: a main control module, a raw data acquisition module, a data screening module, a learning module corresponding to at least one learning method, and a sample labeling module;
  • the main control module is used to determine the target module to be called based on the machine learning model to be trained and at least one selected target learning method determined by the user; the target module to be called at least includes the original data acquisition module and each target Each target learning module corresponding to the learning method; based on the preset execution sequence of each module in the system, determine the target calling sequence of each target module; call each target module according to the target calling sequence, and learn the machine learning
  • the model is trained;
  • the current raw data acquisition module is used to obtain the current raw data of the current application scene captured by the data acquisition device;
  • the data screening module is used to obtain the current raw data from the current raw data based on active learning technology when called In the process, the valuable target original data is screened out;
  • the sample labeling module is used to output the data to be labeled to the user for labeling when called, and obtain the current newly added sample; wherein, the data to be labeled is the user from the described The data selected in the current raw data, or the data selected by the user from the target raw data
  • the learning module corresponding to the at least one learning method includes: an incremental learning module and/or a semi-supervised training module; the incremental learning module is used to , based on the training samples in the pre-stored baseline training set and the current newly added samples, the machine learning model to be trained is incrementally trained to obtain the incrementally trained machine learning model; the semi-supervised training module is used for being called When , based on the current original data or the unlabeled data in the target original data, and the training samples in the pre-stored baseline training set, semi-supervised training is performed on the machine learning model to be trained to obtain the machine learning model after semi-supervised training.
  • the target calling sequence includes one of the following sequences: first call the original data acquisition module, then call the data screening module, then call the sample labeling module, and then call the incremental learning module; first call the original The data acquisition module, then calls the sample labeling module, and then calls the incremental learning module; first calls the original data acquisition module, then calls the data screening module, then calls the sample labeling module, and then calls the incremental learning module, and then calls the incremental learning module output
  • the machine learning model after incremental training is used as the updated machine learning model to be trained, and then calls the semi-supervised training module; first calls the original data acquisition module, then calls the sample labeling module, and then calls the incremental learning module, the incremental learning
  • the incrementally trained machine learning model output by the module is used as the updated machine learning model to be trained, and then the semi-supervised training module is called; the original data acquisition module is called first, then the data screening module is called, and then the semi-supervised training module is called.
  • the machine learning system further includes: an offline coding module; the offline coding module is used to train the basic model with the training samples in the baseline training set in advance, and After the model training is completed, the training samples in the baseline training set are input into the basic model, and the basic features are obtained and saved; the basic features are used for incremental training in the incremental learning module or semi-supervised training in the semi-supervised training module A sample representing the base training set during training.
  • an offline coding module is used to train the basic model with the training samples in the baseline training set in advance, and After the model training is completed, the training samples in the baseline training set are input into the basic model, and the basic features are obtained and saved; the basic features are used for incremental training in the incremental learning module or semi-supervised training in the semi-supervised training module A sample representing the base training set during training.
  • the machine learning system further includes: a verification module; the verification module is configured to, when called, based on the pre-stored baseline test set and the current original data or target original data
  • the selected test data is used to test the trained machine learning model output by the incremental learning module and/or the semi-supervised training module, and output the test results to the user for verification.
  • the target calling sequence includes one of the following sequences: first call the original data acquisition module, then call the data screening module, then call the sample labeling module, then call the incremental learning module, and then call the incremental learning module.
  • the incrementally trained machine learning model output by the learning module is used as the updated machine learning model to be trained, and then the verification module is called, and then the semi-supervised training module is called; the original data acquisition module is called first, and then the sample labeling module is called, and then the The incremental learning module uses the incrementally trained machine learning model output by the incremental learning module as the updated machine learning model to be trained, then calls the verification module, and then calls the semi-supervised training module; first calls the original data acquisition module, Then call the data screening module, then call the semi-supervised training module, and then call the verification module.
  • the machine learning system further includes: a publishing module; the publishing module is used to use the trained machine learning model output by the last learning module as the target model when called to publish.
  • a model training method includes the following steps: obtaining the machine learning model to be trained and at least one selected target learning method determined by the user; obtaining the current application captured by the data collection device The current raw data of the scene; based on active learning technology, from the current raw data, screen out valuable target raw data; according to the target learning method, based on the target raw data, the machine learning model to be trained Perform training to obtain the trained machine learning model.
  • the step of screening out valuable target original data from the current original data based on the active learning technology includes: performing the machine learning model to be trained based on a public data set Perform N times of basic training to obtain N initial machine learning models; where N is greater than or equal to 2; for each current original data, input the current original data into N initial machine learning models to obtain N of the current original data initial output results; N current original data with different initial output results are used as valuable target original data to be screened out.
  • the target learning method includes: incremental learning; according to the target learning method, based on the target original data, the machine learning model to be trained is trained to obtain The step of training the machine learning model includes: marking the target original data according to the incremental learning method to obtain the current newly added samples; based on the training samples in the pre-stored baseline training set and the current newly added samples, the The machine learning model to be trained is incrementally trained to obtain the incrementally trained machine learning model.
  • the target learning method includes: semi-supervised learning; according to the target learning method, based on the target original data, the machine learning model to be trained is trained to obtain
  • the step of training the machine learning model includes: performing semi-supervised training on the machine learning model to be trained based on the unmarked data in the target original data and the training samples in the pre-stored baseline training set according to the semi-supervised learning method, Get the machine learning model after semi-supervised training.
  • the target learning method includes: incremental learning and semi-supervised learning; according to the target learning method, based on the target original data, the machine learning model to be trained
  • the step of performing training to obtain the trained machine learning model includes: marking the target original data according to the incremental learning method to obtain the current newly added samples; based on the training samples in the pre-stored baseline training set and the current new Increase the sample, carry out incremental training to the machine learning model to be trained, obtain the machine learning model after the incremental training; use the machine learning model after the incremental training as the updated machine learning model to be trained, according to half Supervised learning method, based on the unlabeled data in the target original data and the training samples in the pre-stored baseline training set, perform semi-supervised training on the updated machine learning model to be trained, and obtain the machine learning model after semi-supervised training .
  • the machine learning model after the incremental training is used as the updated machine learning model to be trained, according to the semi-supervised learning method, based on the target raw data that is not The marked data, and the training samples in the pre-stored baseline training set, perform semi-supervised training on the updated machine learning model to be trained, and before the step of obtaining the machine learning model after semi-supervised training, the method also includes: based on the pre-stored baseline The test set and the test data selected from the current original data or the target original data are used to test the machine learning model after the incremental training, and output the test results to the user for verification; when the verification results meet the requirements, Executing the machine learning model after the incremental training as the updated machine learning model to be trained, according to the semi-supervised learning method, based on the unlabeled data in the target raw data and the pre-stored baseline training set The steps of performing semi-supervised training on the updated machine learning model to be trained to obtain the machine learning model after semi-supervised training.
  • the method further includes: publishing the trained machine learning model as a target model.
  • a model training device which includes: a learning method acquisition module, used to obtain the user-determined machine learning model to be trained and at least one selected target learning method; the current original The data acquisition module is used to obtain the current raw data of the current application scene captured by the data acquisition device; the raw data screening module is used to screen out valuable target raw data from the current raw data based on active learning technology; the model The training module is configured to train the machine learning model to be trained based on the target original data according to the target learning method, so as to obtain the trained machine learning model.
  • an electronic device including a processor, a communication interface, a memory, and a communication bus, wherein, the processor, the communication interface, and the memory complete communication with each other through the communication bus; the memory uses When the computer program is stored; the processor is used to execute any of the steps of the model training method when executing the program stored on the memory.
  • the embodiment of the present application also provides a computer-readable storage medium, where a computer program is stored in the computer-readable storage medium, and when the computer program is executed by a processor, any step of the model training method described above is implemented.
  • the embodiment of the present application also provides a computer program product containing instructions, which, when run on a computer, causes the computer to execute any one of the model training methods described above.
  • the machine learning system includes a main control module, a raw data acquisition module, a data screening module, a sample labeling module, and a learning module corresponding to at least one learning method.
  • the main control module determines the target to be called based on the user's choice module, and determine the target call sequence based on the preset execution order of each module, so as to call the target module according to the target call sequence to train the machine learning model to be trained.
  • the target module needs to include at least the original data acquisition module and each target learning method
  • the corresponding target learning module the raw data acquisition module obtains the current raw data of the current application scene captured by the data acquisition device, the data screening module screens out valuable target raw data from the current raw data, and the sample tagging module outputs the data to be tagged Label the user to obtain new samples.
  • each learning module is called, according to its corresponding learning method, based on the samples in the pre-stored baseline training set and/or new samples, the machine learning model to be trained is trained to obtain A trained machine learning model.
  • the machine learning system provided by the embodiment of the present application can obtain the data of the current scene, and can train the model according to different learning methods selected by the user to obtain the trained machine learning model. In different application scenarios, the developers redesign the model to realize adaptive learning of real application scenarios.
  • FIG. 1 is a schematic structural diagram of a machine learning system provided in an embodiment of the present application
  • Fig. 2a is a second structural schematic diagram of the machine learning system provided by the embodiment of the present application.
  • Fig. 2b is a functional module relationship diagram based on the machine learning system shown in Fig. 2a;
  • Fig. 3a is a schematic flow chart of the module calling sequence in the machine learning system shown in Fig. 2a;
  • Fig. 3b is a second schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 2a;
  • Fig. 3c is a third schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 2a;
  • Fig. 3d is a fourth schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 2a;
  • Fig. 3e is a fifth schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 2a;
  • Fig. 3f is a schematic diagram of a sixth flow chart of the module calling sequence in the machine learning system shown in Fig. 2a;
  • FIG. 4 is a schematic diagram of a third structure of a machine learning system provided in an embodiment of the present application.
  • Fig. 5a is a schematic flow chart of the module calling sequence in the machine learning system shown in Fig. 4;
  • Fig. 5b is a second schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 4;
  • Fig. 5c is a third schematic flow diagram of the module calling sequence in the machine learning system shown in Fig. 4;
  • FIG. 6 is a schematic diagram of a fourth structure of a machine learning system provided in an embodiment of the present application.
  • Fig. 7 is a schematic flow chart of the module calling sequence in the machine learning system shown in Fig. 6;
  • Fig. 8 is a flow chart of the model training method provided by the embodiment of the present application.
  • Fig. 9 is a flow chart of screening out valuable target raw data in the embodiment of the present application.
  • FIG. 10 is a second flow chart of the model training method provided by the embodiment of the present application.
  • FIG. 11 is a third flow chart of the model training method provided by the embodiment of the present application.
  • FIG. 12 is a schematic structural diagram of a model training device provided in an embodiment of the present application.
  • FIG. 13 is a schematic structural diagram of an electronic device provided by an embodiment of the present application.
  • the embodiment of the present application provides a machine learning system and a model training method.
  • the following firstly introduces the machine learning system provided in the embodiment of the present application.
  • the machine learning system provided in the embodiment of the present application is a user-oriented system, and the user can train the model based on the system according to the user's own needs, so as to obtain a machine learning model that can realize the functions required by the user.
  • FIG. 1 is a schematic structural diagram of a machine learning system provided by an embodiment of the present application.
  • the above-mentioned system may include a main control module 100, a raw data acquisition module 110, a data screening module 120, a sample labeling module 130 and at least A learning module 140 corresponding to a learning mode;
  • the above-mentioned main control module 100 is used to determine the target module to be called based on the machine learning model to be trained and at least one selected target learning method determined by the user; the target module to be called at least includes the original data acquisition module and each target learning method Corresponding to each target learning module; based on the preset execution sequence of each module in the system, determine the target call sequence of each target module; call each target module according to the target call sequence, and perform the machine learning model to be trained train;
  • Raw data acquisition module 110 configured to obtain the current raw data of the current application scene captured by the data acquisition device
  • the data screening module 120 is used to screen out valuable target original data from the current original data based on active learning technology when called;
  • the sample labeling module 130 is configured to output the data to be labeled to the user for labeling to obtain the current newly added sample when being called; wherein, the data to be labeled is the data selected by the user from the current original data, or the user selects the data from the data The data selected in the target raw data output by the screening module;
  • Each learning module 140 is used to train the machine learning model to be trained according to its corresponding learning method based on the training samples in the pre-stored baseline training set and/or the current newly added samples when being called to obtain the trained machine learning model.
  • the machine learning system includes a main control module, a raw data acquisition module, a data screening module, a sample labeling module, and a learning module corresponding to at least one learning method.
  • the main control module determines the target to be called based on the user's choice module, and determine the target call sequence based on the preset execution order of each module, so as to call the target module according to the target call sequence to train the machine learning model to be trained.
  • the target module needs to include at least the original data acquisition module and each target learning method
  • the raw data acquisition module obtains the current raw data of the current application scene captured by the data acquisition device
  • the data screening module screens out valuable target raw data from the current raw data
  • the sample tagging module outputs the data to be tagged Label the user to obtain new samples.
  • each learning module is called, according to its corresponding learning method, based on the samples in the pre-stored baseline training set and/or new samples, the machine learning model to be trained is trained to obtain A trained machine learning model.
  • the machine learning system calls the raw data acquisition module 110, the data screening module 120, the sample labeling module 130 and the learning module 140 sequentially through the main control module 100, so that the data of the current scene can be obtained through the raw data acquisition module 110 , and filter the obtained data of the current scene through the data screening module 120, and then use the sample labeling module 130 to label the obtained data of the current scene or the filtered data to obtain incremental samples, and can pass the learning module 140
  • the model is trained according to different learning methods selected by the user to obtain the trained machine learning model. That is to say, when the model is applied to different application scenarios, by Developers redesigned the model to realize adaptive learning of real application scenarios.
  • the above-mentioned user-selectable machine learning model structure can be a machine learning model of any structure, such as AlexNet network, ResNet50 network, etc., and the machine learning model can be used for detection
  • the model may also be a model used for recognition, or a model used for image restoration, etc., which are not specifically limited here.
  • the aforementioned data collection devices may also be different.
  • the above-mentioned data collection device may be a roadside camera device and/or a radio frequency card reader.
  • the aforementioned data acquisition device may be any one or more of infrared sensors, visible light sensors, and smoke sensors.
  • the above-mentioned data screening module 120 screens the above-mentioned raw data based on the active learning technology.
  • the above-mentioned active learning technology can be implemented based on the uncertainty method, or based on the feature distribution method. No specific details are given here. limited.
  • the selection can be stopped.
  • the above-mentioned data screening module obtains valuable data please refer to the description of the method embodiment below, which will not be described in detail here.
  • the above-mentioned learning methods that users can choose may include incremental learning, semi-supervised learning, etc.
  • the above-mentioned learning methods that users can choose may include incremental learning, semi-supervised learning, etc.
  • the above-mentioned learning methods that users can choose may include incremental learning, semi-supervised learning, etc.
  • the above-mentioned incremental learning module 240 is used to perform incremental training on the machine learning model to be trained based on the training samples in the pre-stored baseline training set and the current newly added samples when being called, so as to obtain the incrementally trained machine learning model;
  • the above-mentioned baseline training set is a training data set constructed by developers when training the machine model in the development stage before the above-mentioned machine learning system is opened to users.
  • the original machine model contained in the above-mentioned machine learning system that is, machine learning
  • the machine learning model to be trained at the initial stage of the system can show better performance in the baseline training set.
  • the above performance can be accuracy, running speed, etc.
  • the machine learning model to be trained can be pre-trained on the baseline training set, but limited by the labeling method of the training samples in the baseline training set, the knowledge that the machine learning model to be trained can learn from the baseline training set is limited, and may not be able to Applicable to real application scenarios actually faced by users. Therefore, the machine learning model to be trained needs to be further trained using the incremental learning module 240 .
  • the data labels in the original training set only contain "motor vehicle” and "non-motor vehicle", so the machine learning to be trained
  • the model can only learn how to identify motor vehicles and non-motor vehicles, and if the user needs the machine learning model to be able to recognize pedestrians in addition to recognizing motor vehicles and non-motor vehicles, you need to use the above incremental learning module 240 Train the machine learning model to be trained that is selected by the user above.
  • the above-mentioned sample labeling module 130 outputs the data to be labeled to the user for labeling, and the user can label the above-mentioned data to be labeled as "motor vehicle”, “non-motor vehicle” or "person”, and the above-mentioned data to be labeled after labeling
  • the data is a new sample
  • the incremental learning module 240 can train the above-mentioned machine learning model to be trained based on the above-mentioned new sample or, the new sample and the baseline training set, to obtain a machine learning model after incremental learning training, This enables the model to identify motor vehicles, non-motor vehicles, and people.
  • the above-mentioned incremental learning method may be a distillation method or a dynamic network method, which is not specifically limited in this embodiment of the present application.
  • the above-mentioned semi-supervised training module 250 is used to perform semi-supervised training on the machine learning model to be trained based on the unlabeled data in the current raw data or target raw data and the training samples in the pre-stored baseline training set when being called, to obtain A machine learning model after semi-supervised training.
  • the data in the baseline training set are labeled with "motor vehicles” or "non-motor vehicles".
  • the data (newly added samples) marked by the user is not all the current original data or the target original data, that is, the user often only marks part of the current original data or the target original data. Therefore, the current original data or the target original data still have
  • the above-mentioned semi-supervised training module can perform semi-supervised training on the machine learning model to be trained based on the above-mentioned baseline training set and the unlabeled data in the above-mentioned original data or target original data, and obtain the trained machine learning model. This enables the trained machine learning model to identify vehicles in the current application scenario.
  • the above-mentioned semi-supervised learning can be implemented by using any one of consistency constraints, regular constraints, Mean Teacher, or any combination of multiple methods.
  • the above-mentioned incremental learning module 240 and semi-supervised training module 250 can be trained based on the above-mentioned baseline training set, which ensures the performance of the model on the basic data, that is, ensures that the trained machine learning model obtained after training can be trained on the baseline training set. better model performance.
  • the machine learning system shown in Figure 2a can also include an offline encoding module (not shown in Figure 2a); the offline encoding module is used to train the basic model with the training samples in the baseline training set in advance, and After the basic model training is completed, the training samples in the baseline training set are input into the basic model, and the basic features are obtained and saved; the basic features are used for incremental training in the incremental learning module 240 or semi-supervised training in the semi-supervised training module 250 A sample representing the base training set during training.
  • the model structure of the basic model and the machine learning model to be trained are the same. For example, if it is incremental training based on distillation, the structure of the basic model and the machine learning model to be trained are the same.
  • the structure of the basic model and the machine learning model to be trained can be different. For example, if it is a machine learning model to be trained based on a dynamic network, the model structure of the model will change during the training process. The structure differs from the base model.
  • the above-mentioned offline encoding module trains the basic model based on the above-mentioned baseline training set (basic data).
  • the model parameters are restored/decrypted to obtain the basic features.
  • the above-mentioned raw data acquisition module 110 obtains batch data (that is, a large amount of data, equivalent to the above-mentioned current original data), and the above-mentioned batch data is processed by the data screening module 120 based on active learning technology Carry out data screening to screen out valuable data in the batch data to obtain the target original data (valuable selection), and the user marks the data to be marked in the above target original data to obtain a small amount of labeled data (equivalent to Added samples above) and the rest of the large amount of unlabeled data.
  • batch data that is, a large amount of data, equivalent to the above-mentioned current original data
  • the data screening module 120 based on active learning technology Carry out data screening to screen out valuable data in the batch data to obtain the target original data (valuable selection), and the user marks the data to be marked in the above target original data to obtain a small amount of labeled data (equivalent to Added samples above) and the rest of the large amount of unlabeled data.
  • the above-mentioned incremental learning module 240 can train the above-mentioned machine learning model to be trained based on the above-mentioned newly added samples and the above-mentioned basic features; the above-mentioned semi-supervised training module 250 can train the above-mentioned machine learning model to be trained based on the above-mentioned unlabeled data and the aforementioned basic features The model is trained.
  • the model obtained after the above-mentioned incremental training can be used as the machine learning model to be trained during semi-supervised training, and the model obtained by the above-mentioned semi-supervised training can also be used as the machine learning model to be trained for incremental training.
  • the model obtained after the above-mentioned incremental training can be used as the machine learning model to be trained during semi-supervised training, and the model obtained by the above-mentioned semi-supervised training can also be used as the machine learning model to be trained for incremental training.
  • the model performance of the trained machine learning model output by the machine learning system can be kept not lower than the original model (that is, the machine to be trained learning model) performance.
  • the off-line encoding module for the semi-supervised training module, while adapting to the new scene, it can also maintain the performance on the basic data; for the incremental learning module, it can recognize the new data and also recognize the basic data. ability.
  • the above-mentioned main control module 100 can sequentially call each target module according to the target calling sequence.
  • the above-mentioned calling sequence can be:
  • the above-mentioned raw data acquisition module 110 obtains the raw image data collected by the camera, and the data screening module 120 screens out valuable target raw data from the above-mentioned raw image data, and then the sample tagging module selects the target raw data selected by the user.
  • the data in the target original data is output to the user for labeling, and new samples are obtained.
  • Each data in the new samples is marked as "motor vehicle”, “non-motor vehicle” or “person”, and then the above incremental learning module is based on the above
  • the baseline training set and the new samples are used to train the machine learning model to be trained to obtain a target model (equivalent to the above-mentioned trained machine learning model), which can identify motor vehicles, non-motor vehicles and people in the original data.
  • the calling sequence may be: first call the raw data acquisition module 110 , then call the sample labeling module 130 , and then call the incremental learning module 240 .
  • the above-mentioned data screening module 120 is not invoked in the invocation sequence shown in FIG. 3b, but after the original image data is obtained, the user selects the data to be marked in the original image data, And the above data to be marked is output to the user by the sample labeling module for labeling.
  • the execution content of other modules is the same as that in Fig. 3a, and will not be repeated here.
  • the above-mentioned target calling sequence may be: first call the original data acquisition module 110, then call the data screening module 120, then call the above-mentioned sample labeling module 130, and then The incremental learning module 240 is called, and the incrementally trained machine learning model output by the incremental learning module is used as the updated machine learning model to be trained, and then the semi-supervised training module 250 is called.
  • the corresponding incrementally trained machine learning model is obtained. Since there are still a lot of unlabeled data in the above raw data, therefore, The above incrementally trained machine learning model can be trained based on the above baseline training set and the above unlabeled data, so that the finally obtained trained machine learning model has better performance, such as robustness.
  • the above-mentioned target calling sequence may be: first call the original data acquisition module 110, then call the above-mentioned sample labeling module 130, then call the incremental learning module 240, and The incrementally trained machine learning model output by the incremental learning module is used as the updated machine learning model to be trained, and then the semi-supervised training module 250 is called.
  • the above target calling sequence may be: first call the raw data acquisition module 110 , then call the data screening module 120 , and then call the semi-supervised training module 250 .
  • the data screening module 120 can obtain valuable data after screening the current raw data.
  • the above-mentioned valuable data is unlabeled data.
  • Semi-supervised training is performed on the above-mentioned machine model to be trained with the unlabeled data screened out by the set and the above-mentioned data screening module.
  • the above target invocation sequence may be: first invoking the raw data acquisition module 110 , and then invoking the semi-supervised training module 250 .
  • the user can label the above-mentioned unlabeled data and calculate the output corresponding to the data.
  • the above machine learning system may further include a verification module 460;
  • the above verification module 460 is used to train the incremental learning module 240 and/or the semi-supervised training module 250 based on the pre-stored baseline test set and the test data selected from the current original data or target original data when called. After training, the machine learning model is tested, and the test results are output to the user for verification. The test data should be different from the data used to train the trained machine learning model.
  • the quantity of verification module 460 can be one or more, and a verification module can be set after each learning module 140, for example: a verification module 460 can be set after the incremental learning module 240, or A verification module 460 may be provided after the semi-supervised training module 250 .
  • the verification module 460 disposed behind the learning module 140 is used to verify the performance of the trained machine learning model output by the learning module 140 .
  • the above-mentioned baseline test set may be set by a developer during the development phase, and the above-mentioned test set may be selected by a user.
  • the user's verification criteria may be that the trained machine learning model has good performance in the above-mentioned baseline test set, and has good performance in the above-mentioned original data or the test data in the target original data.
  • the above-mentioned performance is good
  • the criterion for can be that the accuracy rate on each data set is greater than a preset threshold, which is not specifically limited here.
  • the above calling sequence may include the following sequence:
  • Model as the updated machine learning model to be trained, then calls the verification module 460, and then calls the semi-supervised training module 250;
  • the verification module 460 can be used to test the machine learning model after the above-mentioned incremental training based on the above-mentioned baseline test set and the test set selected from the original data. Test, if the above-mentioned model meets the requirements (for example, the accuracy rate is higher than 88%), then call the semi-supervised training module 250 to train the machine learning model after the above-mentioned incremental training; if the above-mentioned model does not meet the requirements, you can reselect The data to be marked is marked, and the machine learning model to be trained is trained based on the above-mentioned newly added samples and the baseline training set.
  • the original data acquisition module 110 is called first, then the sample labeling module 130 is called, and then the incremental learning module 240 is called, and the incrementally trained machine learning model output by the incremental learning module is used as the updated waiting Train the machine learning model, then call the verification module 460, and then call the semi-supervised training module 250;
  • the raw data acquisition module 110 is called first, then the data screening module 120 is called, then the semi-supervised training module 250 is called, and then the verification module 460 is called.
  • the machine learning model after the semi-supervised training can be obtained, and the machine learning model after the semi-supervised training can be called by the verification module 460.
  • the model is verified, and the verification process is similar to the process in Figure 5a above, and will not be repeated here.
  • the above machine learning system may further include a publishing module 670;
  • the publishing module 670 is configured to publish the trained machine learning model output by the last learning module 140 as the target model when called.
  • the finally obtained trained machine learning model can be used as the target model, and the model can realize the functions required by the user, such as vehicle identification.
  • the target model can be released to the corresponding platform, where the corresponding platform refers to the platform suitable for the real application scenario faced by the user, such as the real application scenario is to identify vehicles on the road, then the corresponding platform Vehicle recognition platform, and if the real application scenario is to identify the identity of visitors, then the corresponding platform is a face recognition platform, etc., for the use of relevant personnel.
  • the above target calling sequence may include: first calling the original data acquisition module 110, then calling the data screening module 120, then calling the sample labeling module 130, and then calling the incremental learning module 240, and the incremental learning
  • the learning model replaces the original machine learning model and is published by the publishing module 670 .
  • FIG. 7 The embodiment shown in FIG. 7 is similar to the above-mentioned embodiment in FIG. 5 a , and will not be repeated here.
  • the machine learning system provided by the embodiments of the present application can allow users to flexibly select a learning method, and then perform corresponding model training, which can adapt to different application scenarios and has good robustness.
  • the embodiment of the present application also provides a model training method, as shown in Figure 8, which is a flow chart of the model training method provided in the embodiment of the present application, and the above-mentioned method can specifically be Include the following steps:
  • Step 800 obtaining the machine learning model to be trained and at least one selected target learning method determined by the user;
  • Step 810 obtaining the current raw data of the current application scene captured by the data acquisition device
  • Step 820 screen out valuable target original data from the current original data
  • step 820 may specifically include the following steps:
  • Step 901 Perform N times of basic training on the machine learning model to be trained based on the public data set to obtain N initial machine learning models; where N is greater than or equal to 2;
  • Step 902 for each current original data, respectively input the current original data into N initial machine learning models to obtain N initial output results of the current original data;
  • Step 903 taking N current raw data with different initial output results as valuable target raw data to be screened out.
  • the above-mentioned machine learning model to be trained can be a ResNet50 classification network
  • the above-mentioned public data set can be an ImageNet data set
  • the process of obtaining valuable target original data can be: randomly initialize the above-mentioned ResNet50 classification network , and train the above model based on the above ImageNet dataset; the above process is repeated N times (N ⁇ 2), and N classification models can be obtained.
  • N classification models are used for testing, if the output results are different, it is considered inconsistent, that is, the image data is valuable target original data.
  • the least confidence (Least Confident) method can be used to filter out the target original data from the original data, that is, to select the maximum probability minimum
  • the data is valuable target raw data.
  • the category prediction probability of the first image data is (0.9,0.1)
  • the category prediction result of the second image data is (0.51,0.49) , that is, the probability of the first image being judged as the first class is 0.9
  • the probability of the second image being judged as the first class is 0.51, that is, the second image data is more difficult for the model differentiated, then the second image data is valuable target original data.
  • the entropy method Entropy
  • the expected model change method Exected Model Change
  • etc. may also be used to screen out the target data from the original data, which are not specifically limited here.
  • step 830 according to the target learning method, the machine learning model to be trained is trained based on the target original data to obtain a trained machine learning model.
  • the model training method provided by the embodiment of the present application allows the user to flexibly select a learning method, and then perform corresponding model training, which can adapt to different application scenarios and has good robustness.
  • the above-mentioned target learning method can be an incremental learning method, that is, the target original data can be marked according to the incremental learning method to obtain the current newly added sample; based on the pre-stored baseline
  • the training samples in the training set and the currently added samples are used to perform incremental training on the machine learning model to be trained to obtain a machine learning model after incremental training.
  • the above-mentioned target learning method can be semi-supervised learning, that is, according to the semi-supervised learning method, based on the unlabeled data in the target raw data and the pre-stored baseline training set
  • the training samples are used to conduct semi-supervised training on the machine learning model to be trained to obtain the machine learning model after semi-supervised training.
  • FIG. 10 is a second flow chart of the machine training method provided in the embodiment of the present application.
  • the above step 830 may specifically include the following steps:
  • Step 1010 mark the target original data to obtain the current newly added samples; based on the training samples in the pre-stored baseline training set and the current newly added samples, perform the training on the machine learning model to be trained Incremental training, to obtain the machine learning model after incremental training;
  • Step 1011 using the machine learning model after the incremental training as the updated machine learning model to be trained, according to the semi-supervised learning method, based on the unlabeled data in the target original data and the pre-stored baseline training set
  • the training samples are used to perform semi-supervised training on the updated machine learning model to be trained to obtain the machine learning model after semi-supervised training.
  • step 1011 it may also include:
  • Step 1110 based on the pre-stored baseline test set and the test data selected from the current original data or the target original data, test the machine learning model after the incremental training, and output the test results to the user for verification; if the requirements are met , execute step 1011;
  • step 1011 it may also include:
  • Step 1111 publish the trained machine learning model as the target model.
  • the machine learning model to be trained and at least one target learning method determined by the user are first obtained, and then the current raw data of the current application scene captured by the data acquisition device is obtained, and based on the active learning technology, from The valuable target raw data is screened out from the current raw data, and then the machine learning model to be trained is trained according to the target learning method selected by the user to obtain the trained machine learning model.
  • the model training method provided in the embodiment of the present application can obtain the original data of the current scene, and train the machine learning model to be trained based on the filtered original data of the current scene according to the learning method selected by the user. That is to say, when When the machine learning model is applied to different application scenarios, developers do not need to redesign it, realizing adaptive learning of real scenarios.
  • the embodiment of the present application also provides a model training device, as shown in Figure 12, the above-mentioned device may include:
  • the learning method obtaining module 1200 is used to obtain the machine learning model to be trained and at least one selected target learning method determined by the user;
  • the current raw data obtaining module 1201 configured to obtain the current raw data of the current application scene captured by the data collection device;
  • Raw data screening module 1202 configured to screen valuable target raw data from the current raw data based on active learning technology
  • the model training module 1203 is configured to train the machine learning model to be trained based on the target original data according to the target learning method, to obtain a trained machine learning model.
  • the machine learning model to be trained and at least one target learning method determined by the user are first obtained, and then the current raw data of the current application scene captured by the data acquisition device is obtained, and based on the active learning technology, from The valuable target raw data is screened out from the current raw data, and then the machine learning model to be trained is trained according to the target learning method selected by the user to obtain the trained machine learning model.
  • the model training method provided in the embodiment of the present application can obtain the original data of the current scene, and train the machine learning model to be trained based on the filtered original data of the current scene according to the learning method selected by the user. That is to say, when When the machine learning model is applied to different application scenarios, developers do not need to redesign it, realizing adaptive learning of real scenarios.
  • the embodiment of the present application also provides an electronic device, as shown in FIG. 13 , including a processor 1301, a communication interface 1302, a memory 1303, and a communication bus 1304. complete the mutual communication,
  • the machine learning model to be trained is trained to obtain a trained machine learning model.
  • model training method In the model training method provided in the embodiment of the present application, first obtain the machine learning model to be trained and at least one target learning method determined by the user, and then obtain the current raw data of the current application scene captured by the data acquisition device, and based on the active learning technology, from The valuable target raw data is screened out from the current raw data, and then the machine learning model to be trained is trained according to the target learning method selected by the user to obtain the trained machine learning model.
  • the model training method provided in the embodiment of the present application can obtain the original data of the current scene, and train the machine learning model to be trained based on the filtered original data of the current scene according to the learning method selected by the user. That is to say, when When the machine learning model is applied to different application scenarios, developers do not need to redesign it, realizing adaptive learning of real scenarios.
  • the communication bus mentioned above for the electronic device may be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus or the like.
  • PCI Peripheral Component Interconnect
  • EISA Extended Industry Standard Architecture
  • the communication bus can be divided into an address bus, a data bus, a control bus, and the like. For ease of representation, only one thick line is used in the figure, but it does not mean that there is only one bus or one type of bus.
  • the communication interface is used for communication between the electronic device and other devices.
  • the memory may include a random access memory (Random Access Memory, RAM), and may also include a non-volatile memory (Non-Volatile Memory, NVM), such as at least one disk memory.
  • RAM Random Access Memory
  • NVM non-Volatile Memory
  • the memory may also be at least one storage device located far away from the aforementioned processor.
  • the above-mentioned processor can be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU), a network processor (Network Processor, NP), etc.; it can also be a digital signal processor (Digital Signal Processor, DSP), a dedicated integrated Circuit (Application Specific Integrated Circuit, ASIC), Field-Programmable Gate Array (Field-Programmable Gate Array, FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components.
  • CPU Central Processing Unit
  • NP Network Processor
  • DSP Digital Signal Processor
  • ASIC Application Specific Integrated Circuit
  • FPGA Field-Programmable Gate Array
  • a computer-readable storage medium is also provided.
  • a computer program is stored in the computer-readable storage medium.
  • any of the above-mentioned model training methods is implemented. A step of.
  • a computer program product including instructions is also provided, and when it is run on a computer, it causes the computer to execute any model training method in the above embodiments.
  • all or part of them may be implemented by software, hardware, firmware or any combination thereof.
  • software When implemented using software, it may be implemented in whole or in part in the form of a computer program product.
  • the computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on the computer, the processes or functions according to the embodiments of the present application will be generated in whole or in part.
  • the computer can be a general purpose computer, a special purpose computer, a computer network, or other programmable devices.
  • the computer instructions may be stored in or transmitted from one computer-readable storage medium to another computer-readable storage medium, for example, the computer instructions may be transmitted from a website, computer, server or data center Transmission to another website site, computer, server, or data center by wired (eg, coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (eg, infrared, wireless, microwave, etc.).
  • the computer-readable storage medium may be any available medium that can be accessed by a computer, or a data storage device such as a server or a data center integrated with one or more available media.
  • the available medium may be a magnetic medium (for example, a floppy disk, a hard disk, or a magnetic tape), an optical medium (for example, DVD), or a semiconductor medium (for example, a Solid State Disk (SSD)).
  • SSD Solid State Disk

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Computing Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Molecular Biology (AREA)
  • Probability & Statistics with Applications (AREA)
  • Evolutionary Biology (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Medical Informatics (AREA)
  • Image Analysis (AREA)

Abstract

一种机器学习系统及模型训练方法，该系统包括主控模块（100）、原始数据获取模块（110）、数据筛选模块（120）、样本标注模块（130）和学习模块（140），主控模块（100）按照目标调用顺序调用用户选择的目标模块，对模型进行训练，数据筛选模块（120）从原始数据获取模块（110）获得的当前原始数据中筛选出有价值的目标原始数据，样本标注模块（130）将待标注的数据输出给用户进行标注，学习模块（140）在被调用时，按照其对应的学习方式，对待训练机器学习模型进行训练，得到相应模型。该机器学习系统，可以获取当前场景的数据，并按照用户选择的学习方式对模型进行训练，并不需要在模型应用于不同的应用场景时，由开发人员对模型重新设计，实现了对真实应用场景的自适应学习。

Description

一种机器学习系统及模型训练方法
本申请要求于2021年08月05日提交中国专利局、申请号为202110898727.6发明名称为“一种机器学习系统及模型训练方法”的中国专利申请的优先权，其全部内容通过引用结合在本申请中。
技术领域
本申请涉及机器学习技术领域，特别是涉及一种机器学习系统及模型训练方法。
背景技术
目前，机器学习技术被广泛应用到各个技术领域中，例如：视频监控、行为分析、图像处理等等技术领域。
相关技术中，机器学习模型都是在应用软件的研发阶段，根据用户提供的目标场景的训练数据进行训练，并在训练好后集成到具体的应用软件中提供给用户进行使用。
因此，为了使得应用软件中集成的机器学习模型能够适应不同的应用场景，相关技术中，每更换一次目标场景，都需要根据用户提供的目标场景的训练数据进行重新训练。例如：对于集成了一个或多个机器学习模型的视频监控软件，如要应用到火车站，需要在研发阶段用火车站的视频样本进行训练；如要应用到飞机场，则需要在研发阶段用飞机场的视频样本进行训练。在训练完成后，才能将应用软件提供给用户。
可见，相关技术的机器学习模型都是在应用软件的研发阶段进行训练，无法实现对真实应用场景的自适应学习。
发明内容
本申请实施例的目的在于提供一种机器学习系统及模型训练方法，以实现对真实应用场景的自适应学习。具体技术方案如下：
在本申请实施的第一方面，提供了一种机器学习系统，所述系统包括：主控模块、原始数据获取模块、数据筛选模块、至少一种学习方式对应的学习模块和样本标注模块；
所述主控模块，用于基于用户确定的待训练机器学习模型及选择的至少一个目标学习方式，确定需要调用的目标模块；所述需要调用的目标模块中至少包含原始数据获取模块和各个目标学习方式对应的各个目标学习模块；基于预设的系统中各个模块的前后执行顺序，确定各个目标模块的目标调用顺序；按所述目标调用顺序，调用各个目标模块，对所述待训练机器学习模型进行训练；当前原始数据获取模块，用于获得数据采集设备捕获的当前应用场景的当前原始数据；所述数据筛选模块，用于在被调用时，基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；所述样本标注模块，用于在被调用时，将待标注数据输出给用户进行标注，得到当前新增样本；其中，待标注数据是用户从所述当前原始数据中选择的数据，或用户从所述数据筛选模块输出的目标原始数据中选择的数据；每个学习模块，用于在被调用时，按其对应的学习方式，基于预存的基线训练集中的训练样本和/或所述当前新增样本，对待训练机器学习模型进行训练，得到训练后的机器学习模型。
在本申请的一种实施例中，所述至少一种学习方式对应的学习模块中，包括：增量学习模块和/或半监督训练模块；所述增量学习模块，用于在被调用时，基于预存的基线训练集中的训练样本和所述当前新增样本，对待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；所述半监督训练模块，用于在被调用时，基于当前原始数据或目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
在本申请的一种实施例中，所述目标调用顺序包括以下顺序之一：先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块；先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块；先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用半监督训练模块；先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用半监督训练模块；先调用原始数据获取模块，再调用数据筛选模块，再调用半监督训练模块。
在本申请的一种实施例中，所述机器学习系统，还包括：离线编码模块；所述离线编码模块，用于预先用所述基线训练集中的训练样本对基础模型进行训练，并在基础模型训练完成后，将基线训练集中的训练样本输入到基础模型中，得到基础特征并保存；所述基础特征，用于在增量学习模块进行增量训练或所述半监督训练模块进行半监督训练中代表基础训练集的样本。
在本申请的一种实施例中，所述机器学习系统，还包括：核验模块；所述核验模块，用于在被调用时，基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对所述增量学习模块和/或所述半监督训练模块输出的训练后机器学习模型进行测试，将测试结果输出给用户进行核验。
在本申请的一种实施例中，所述目标调用顺序包括以下顺序之一：先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块，再调用半监督训练模块；先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块，再调用半监督训练模块；先调用原始数据获取模块，再调用数据筛选模块，再调用半监督训练模块，再调用核验模块。
在本申请的一种实施例中，所述机器学习系统，还包括：发布模块；所述发布模块，用于在被调用时，将最后一个学习模块输出的训练后机器学习模型，作为目标模型进行发布。
在本申请实施的第二方面，提供了一种模型训练方法，所述方法包括如下步骤：获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；获得数据采集设备捕获的当前应用场景的当前原始数据；基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。在本申请的一种实施例中，所述基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据的步骤，包括：基于公共数据集对所述待训练机器学习模型进行N次基 础训练，得到N个初始机器学习模型；其中，N大于或等于2；针对每个当前原始数据，将该当前原始数据分别输入N个初始机器学习模型，得到该当前原始数据的N个初始输出结果；将N个初始输出结果不同的当前原始数据，作为筛选出的有价值的目标原始数据。
在本申请的一种实施例中，所述目标学习方式，包括：增量学习；所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型。
在本申请的一种实施例中，所述目标学习方式，包括：半监督学习；所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
在本申请的一种实施例中，所述目标学习方式，包括：增量学习和半监督学习；所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
在本申请的一种实施例中，在所述将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型的步骤之前，该方法还包括：基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对所述增量训练后的机器学习模型进行测试，将测试结果输出给用户进行核验；在核验结果为符合要求的情况下，执行所述将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型的步骤。
在本申请的一种实施例中，所述方法还包括：将训练后机器学习模型，作为目标模型进行发布。
在本申请实施的又一方面，还提供了一种模型训练装置，所述装置包括：学习方式获得模块，用于获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；当前原始数据获得模块，用于获得数据采集设备捕获的当前应用场景的当前原始数据；原始数据筛选模块，用于基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；模型训练模块，用于按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
在本申请实施的又一方面，还提供了一种电子设备，包括处理器、通信接口、存储器和通信总 线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；存储器，用于存放计算机程序；处理器，用于执行存储器上所存放的程序时，实现任一所述的模型训练方法步骤。
本申请实施例还提供了一种计算机可读存储介质，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现任一所述的模型训练方法步骤。
本申请实施例还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述任一所述的模型训练方法。
本申请实施例有益效果：
本申请实施例提供的机器学习系统，包括主控模块、原始数据获取模块、数据筛选模块、样本标注模块以及至少一种学习方式对应的学习模块，主控模块基于用户的选择确定需要调用的目标模块，并基于预设的各模块的前后执行顺序确定目标调用顺序，以按照目标调用顺序调用目标模块对待训练的机器学习模型进行训练，目标模块中至少需要包含原始数据获取模块以及各目标学习方式对应的目标学习模块，原始数据获取模块获得数据采集设备捕获的当前应用场景的当前原始数据，数据筛选模块从当前原始数据中筛选出有价值的目标原始数据，样本标注模块将待标注的数据输出给用户进行标注，来得到新增样本，各学习模块在被调用时，按照其对应的学习方式，基于预存的基线训练集中的样本和/或新增样本，对待训练机器学习模型进行训练，得到训练后的机器学习模型。本申请实施例提供的机器学习系统，可以获取当前场景的数据，并且可以按照用户选择的不同学习方式对模型进行训练，来得到训练后的机器学习模型，也就是说，并不需要在模型应用于不同的应用场景时，由开发人员对模型重新设计，实现了对真实应用场景的自适应学习。
当然，实施本申请的任一产品或方法并不一定需要同时达到以上所述的所有优点。
附图说明
为了更清楚地说明本申请实施例和现有技术的技术方案，下面对实施例和现有技术中所需要使用的附图作简单地介绍，显而易见地，下面描述中的附图仅仅是本申请的一些实施例，对于本领域普通技术人员来讲，在不付出创造性劳动的前提下，还可以根据这些附图获得其他的附图。
图1为本申请实施例提供的机器学习系统的一种结构示意图；
图2a为本申请实施例提供的机器学习系统的第二种结构示意图；
图2b为基于图2a所示的机器学习系统的功能模块关系图；
图3a为图2a所示机器学习系统中的模块调用顺序的一种流程示意图；
图3b为图2a所示机器学习系统中的模块调用顺序的第二种流程示意图；
图3c为图2a所示机器学习系统中的模块调用顺序的第三种流程示意图；
图3d为图2a所示机器学习系统中的模块调用顺序的第四种流程示意图；
图3e为图2a所示机器学习系统中的模块调用顺序的第五种流程示意图；
图3f为图2a所示机器学习系统中的模块调用顺序的第六种流程示意图；
图4为本申请实施例提供的机器学习系统的第三种结构示意图；
图5a为图4所示机器学习系统中的模块调用顺序的一种流程示意图；
图5b为图4所示机器学习系统中的模块调用顺序的第二种流程示意图；
图5c为图4所示机器学习系统中的模块调用顺序的第三种流程示意图；
图6为本申请实施例提供的机器学习系统的第四种结构示意图；
图7为图6所示机器学习系统中的模块调用顺序的一种流程示意图；
图8为本申请实施例提供的模型训练方法的一种流程图；
图9为本申请实施例中筛选出有价值的目标原始数据的一种流程图；
图10为本申请实施例提供的模型训练方法的第二种流程图；
图11为本申请实施例提供的模型训练方法的第三种流程图；
图12为本申请实施例提供的一种模型训练装置的结构示意图；
图13为本申请实施例提供的电子设备的一种结构示意图。
具体实施方式
为使本申请的目的、技术方案、及优点更加清楚明白，以下参照附图并举实施例，对本申请进一步详细说明。显然，所描述的实施例仅仅是本申请一部分实施例，而不是全部的实施例。基于本申请中的实施例，本领域普通技术人员在没有作出创造性劳动前提下所获得的所有其他实施例，都属于本申请保护的范围。
为了实现对真实应用场景的自适应学习，本申请实施例提供了一种机器学习系统及模型训练方法，下面首先对本申请实施例提供的机器学习系统进行介绍。
本申请实施例提供的机器学习系统是一种面向用户的系统，可以由用户根据自己的需求基于该系统对模型进行训练，从而得到可以实现用户所需功能的机器学习模型。
如图1所示，图1为本申请实施例提供的机器学习系统的一种结构示意图，上述系统可以包括主控模块100、原始数据获取模块110、数据筛选模块120、样本标注模块130以及至少一个学习方式对应的学习模块140；
上述主控模块100，用于基于用户确定的待训练机器学习模型及选择的至少一个目标学习方式，确定需要调用的目标模块；需要调用的目标模块中至少包含原始数据获取模块和各个目标学习方式对应的各个目标学习模块；基于预设的系统中各个模块的前后执行顺序，确定各个目标模块的目标调用顺序；按所述目标调用顺序，调用各个目标模块，对所述待训练机器学习模型进行训练；
原始数据获取模块110，用于获得数据采集设备捕获的当前应用场景的当前原始数据；
数据筛选模块120，用于在被调用时，基于主动学习技术，从当前原始数据中，筛选出有价值的目标原始数据；
样本标注模块130，用于在被调用时，将待标注数据输出给用户进行标注，得到当前新增样本；其中，待标注数据是用户从当前原始数据中选择的数据，或用户从所述数据筛选模块输出的目标原始数据中选择的数据；
每个学习模块140，用于在被调用时，按其对应的学习方式，基于预存的基线训练集中的训练样本和/或当前新增样本，对待训练机器学习模型进行训练，得到训练后的机器学习模型。
本申请实施例提供的机器学习系统，包括主控模块、原始数据获取模块、数据筛选模块、样本标注模块以及至少一种学习方式对应的学习模块，主控模块基于用户的选择确定需要调用的目标模块，并基于预设的各模块的前后执行顺序确定目标调用顺序，以按照目标调用顺序调用目标模块对待训练的机器学习模型进行训练，目标模块中至少需要包含原始数据获取模块以及各目标学习方式对应的目标学习模块，原始数据获取模块获得数据采集设备捕获的当前应用场景的当前原始数据， 数据筛选模块从当前原始数据中筛选出有价值的目标原始数据，样本标注模块将待标注的数据输出给用户进行标注，来得到新增样本，各学习模块在被调用时，按照其对应的学习方式，基于预存的基线训练集中的样本和/或新增样本，对待训练机器学习模型进行训练，得到训练后的机器学习模型。本申请实施例提供的机器学习系统，通过主控模块100依次调用原始数据获取模块110、数据筛选模块120、样本标注模块130以及学习模块140，从而可以通过原始数据获取模块110获取当前场景的数据，并通过数据筛选模块120对获取到的当前场景的数据进行筛选，再通过样本标注模块130对获取到的当前场景的数据或经过筛选的数据进行标注得到增量样本，并且可以通过学习模块140基于增量样本和/或预存的训练样本按照用户选择的不同学习方式对模型进行训练，来得到训练后的机器学习模型，也就是说，并不需要在模型应用于不同的应用场景时，由开发人员对模型重新设计，实现了对真实应用场景的自适应学习。
作为本申请实施例的一种具体实施方式，上述用户可选的机器学习模型结构可以是任意结构的机器学习模型，如AlexNet网络、ResNet50网络等，并且该机器学习模型可以是用于进行检测的模型，也可以是用于识别的模型，还可以是用于进行图像修复的模型等，此处不作具体限定。
对于不同的具体应用场景来说，上述数据采集设备也可以是不同的。例如，当需要对车辆的类别进行识别时，即在识别道路上的车辆是机动车还是非机动车时，上述数据采集设备可以是路边架设的摄像头设备和/或射频读卡器。又例如，当需要对场景中是否发生火情进行检测时，上述数据采集设备可以是红外传感器、可见光传感器、烟雾传感器中的任意一种或多种传感器。
上述数据筛选模块120是基于主动学习技术对上述原始数据进行筛选的，本申请实施例中，上述主动学习技术可以是基于不确定度方法实现，也可以是基于特征分布方法实现，此处不作具体限定。在数据筛选模块120获取的有价值的数据数量达到预设的数量时，即可停止挑选。关于上述数据筛选模块如何获取有价值的数据，可以参见下面方法实施例部分的说明，此处暂不详述。
作为本申请实施例的一种具体实施方式，上述用户可以选择的学习方式可以包括增量学习，半监督学习等，相应的，基于图1，如图2a所示，上述学习模块140可以包括增量学习模块240和/或半监督训练模块250；
上述增量学习模块240，用于在被调用时，基于预存的基线训练集中的训练样本和当前新增样本，对待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；
上述基线训练集是在上述机器学习系统开放给用户进行使用之前的开发阶段中，开发人员对机器模型进行训练时构建的训练数据集，上述机器学习系统中包含的原始的机器模型，即机器学习系统初始时的待训练机器学习模型，在基线训练集中可以表现出较好的性能，上述性能可以是准确度、运行速度等。
可以理解的是，待训练机器学习模型可以预先经过基线训练集的训练，但是受限制于基线训练集中训练样本的标注方式，待训练机器学习模型能够从基线训练集中学习到的知识有限，可能无法适用于用户实际面临的真实应用场景。因此，需要使用增量学习模块240待训练机器学习模型进一步训练。
示例性的，基于上述识别车辆为机动车还是非机动车的应用场景，原始的训练集(基线训练集)中的数据标签只包含“机动车”和“非机动车”，因此待训练机器学习模型只能够学习到如何识别机动车和非机动车，而若用户需要机器学习模型在识别机动车与非机动车之外，还要能够实现对行 人的识别，则需要使用上述增量学习模块240对上述用户选择的待训练机器学习模型进行训练。
具体的，上述样本标注模块130将待标注的数据输出给用户进行标注，用户则可以将上述待标注数据标注为“机动车”、“非机动车”或“人”，标注后的上述待标注数据即为新增样本，增量学习模块240就可以基于上述新增样本或，新增样本和基线训练集对上述待训练机器学习模型进行训练，来得到增量学习训练后的机器学习模型，使得模型可以对机动车、非机动车、人进行识别。
上述进行增量学习的方式可以是蒸馏方式，也可以是采用动态网络的方式实现，本申请实施例中不作具体限定。
上述半监督训练模块250，用于在被调用时，基于当前原始数据或目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
基于上述识别机动车与非机动车的应用场景，基线训练集中的数据带有“机动车”或“非机动车”标签。通常，用户标注的数据(新增样本)不是所有的当前原始数据或目标原始数据，即用户往往仅标注当前原始数据或目标原始数据中的部分数据，因此，当前原始数据或目标原始数据中还有无标签数据，上述半监督训练模块就可以基于上述基线训练集以及上述原始数据或目标原始数据中的无标签数据，对待训练机器学习模型进行半监督训练，得到训练后的机器学习模型，以使得训练后的机器学习模型可以对当前应用场景中的车辆进行识别。
作为本申请实施例的一种具体实施方式，上述半监督学习可以是采用一致性约束方式、正则约束、Mean Teacher中的任一种方式或任意多种方式的组合来实现。
上述增量学习模块240以及半监督训练模块250均可基于上述基线训练集进行训练，保证了基础数据上的模型性能，即保证了训练得到的训练后的机器学习模型能够在基线训练集上有较好的模型性能。
在实际应用中，图2a所示的机器学习系统，还可以包括离线编码模块(图2a中未示出)；离线编码模块，用于预先用基线训练集中的训练样本对基础模型进行训练，并在基础模型训练完成后，将基线训练集中的训练样本输入到基础模型中，得到基础特征并保存；基础特征，用于在增量学习模块240进行增量训练或半监督训练模块250进行半监督训练中代表基础训练集的样本。一般情况下，基础模型与待训练机器学习模型的模型结构是相同的，例如：如果是基于蒸馏的增量训练，基础模型与待训练机器学习模型结构是相同的。而某些情况下，基础模型与待训练机器学习模型结构可以不同，例如：如果是基于动态网络的待训练机器学习模型，该模型的模型结构在训练过程中是会变化的，变化后就会与基础模型的结构有所不同。
这种情况下，基于图2a所示的机器学习系统的功能模块关系，参见图2b。如图2b所示：
上述离线编码模块基于上述基线训练集(基础数据)对基础模型进行训练，训练结束后即可得到上述训练后的基础模型的参数，将该参数进行特征压缩/加密，之后再对压缩/加密后的模型参数进行复原/解密，从而得到基础特征。
在运行上述机器学习系统以实现在线学习时，上述原始数据获取模块110获取到批数据(即大量数据，相当于上述当前原始数据)，并由数据筛选模块120对上述批数据，基于主动学习技术进行数据筛选，从而筛选出批数据中的有价值的数据，得到目标原始数据(有价值挑选)，由用户对上述目标原始数据中的待标记数据进行标记，从而得到少量有标签的数据(相当于上述新增样本) 以及其余大量的无标签数据。上述增量学习模块240可以基于上述新增样本以及上述基础特征，对上述待训练机器学习模型进行训练；上述半监督训练模块250可以基于上述无标签数据以及上述基础特征，对上述待训练机器学习模型进行训练。当然，上述增量训练后得到的模型可以作为半监督训练时的待训练机器学习模型，上述半监督训练得到的模型也可以作为增量训练的待训练机器学习模型，具体的模块调用情况可以参见下面图3a～图3f的实施例。
使用离线编码模块，在增量训练或半监督训练过程中，通过输入基础数据特征，能够保持该机器学习系统输出的训练后的机器学习模型的模型性能不低于原有模型(即待训练机器学习模型)性能。且通过使用离线编码模块，对于半监督训练模块，在适应新场景的同时，也能保持基础数据上的性能；对于增量学习模块，在能识别新增数据的同时，对基础数据也具备识别能力。
如上所述，上述主控模块100可以按照目标调用顺序依次调用各目标模块，作为本申请实施例的一种具体实施方式，如图3a所示，上述调用顺序可以是：
首先调用上述原始数据获取模块110，之后调用数据筛选模块120，再调用上述样本标注模块130，再调用增量学习模块240；
基于上述举例中的场景，上述原始数据获取模块110获取摄像头采集的原始图像数据，数据筛选模块120从上述各原始图像数据中筛选出有价值的目标原始数据，之后由样本标注模块将用户选择的目标原始数据中的数据输出给用户进行标注，得到新增样本，新增样本中的各数据被标记为“机动车”、“非机动车”或“人”，之后上述增量学习模块基于上述基线训练集与新增样本对待训练机器学习模型进行训练，得到目标模型(相当于上述训练后的机器学习模型)，该目标模型可以识别出原始数据中的机动车、非机动车以及人。
作为本申请实施例的又一种具体实施方式，如图3b所示，上述调用顺序可以是：首先调用原始数据获取模块110，之后调用样本标注模块130，再调用上述增量学习模块240。
与图3a中的目标调用顺序相比，图3b中所示的调用顺序中不调用上述数据筛选模块120，而是在获取到原始图像数据后，由用户选择原始图像数据中需要标注的数据，并由样本标注模块将上述待标注的数据输出给用户进行标注。其它模块执行内容与图3a中相同，此处不再赘述。
作为本申请实施例的又一种具体实施方式，如图3c所示，上述目标调用顺序可以是：首先调用原始数据获取模块110，之后调用数据筛选模块120，再调用上述样本标注模块130，再调用增量学习模块240，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，之后再调用半监督训练模块250。
基于上述图3a的举例，在使用增量学习模块240对待训练机器学习模型进行训练之后，得到了相应的增量训练后的机器学习模型，由于上述原始数据中还有很多无标签数据，因此，可以基于上述基线训练集与上述无标签数据对上述增量训练后的机器学习模型进行训练，使得最后得到的训练后的机器学习模型具有更好的性能，如鲁棒性。
作为本申请实施例的一种具体实施方式，如图3d所示，上述目标调用顺序可以是：首先调用原始数据获取模块110，之后调用上述样本标注模块130，之后调用增量学习模块240，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用半监督训练模块250。
与图3c所示实施例类似，此处不再赘述。
作为本申请实施例的又一种具体实施方式，如图3e所示，上述目标调用顺序可以是：首先调用原始数据获取模块110，之后调用数据筛选模块120，再调用半监督训练模块250。
基于上述举例中识别机动车与非机动车的场景，数据筛选模块120对当前原始数据进行筛选后，可以获取有价值的数据，上述有价值的数据是无标签数据，因此，可以基于上述基线训练集与上述数据筛选模块筛选出的无标签数据对上述待训练的机器模型进行半监督训练。
作为本申请实施例的又一种具体实施方式，如图3f所示，上述目标调用顺序可以是：首先调用原始数据获取模块110，再调用半监督训练模块250。
与图3e所示实施例类似，此处不再赘述。
当然，作为本申请实施例的一种具体实施方式，也可以在获取到上述数据筛选模块120输出的无标签数据后，由用户自己为上述无标签数据打上标签，并计算数据对应的输出的结果与上述用户设置的标签之间的损失函数值，并基于该损失函数值对待训练的机器学习模型进行参数调整，从而得到最终训练后的机器学习模型，此处不作具体限定。
作为本申请实施例的一种具体实施方式，基于图2a，如图4所示，上述机器学习系统还可以包括核验模块460；
上述核验模块460，用于在被调用时，基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对增量学习模块240和/或半监督训练模块250训练得到的训练后机器学习模型进行测试，将测试结果输出给用户进行核验。测试数据应当与训练得到训练后的机器学习模型时所使用的的数据不同。
实际应用中，核验模块460的数量可以是一个也可以是多个，可以在每个学习模块140后设置一个核验模块，例如：可以在所述增量学习模块240后设置一个核验模块460，也可以在半监督训练模块250后设置一个核验模块460。设置于学习模块140后的核验模块460用于对该学习模块140输出的训练后的机器学习模型的性能进行核验。
与上述基线训练集类似，上述基线测试集可以是在开发阶段由开发人员设置的，上述测试集可以由用户选择。用户进行核验的标准可以是，训练后的机器学习模型在上述基线测试集中的性能良好，且在上述原始数据或目标原始数据中的测试数据中性能良好，作为一种具体实施方式，上述性能良好的标准可以是在各数据集上的准确率大于预设的阈值，此处不作具体限定。
相应的，上述调用顺序可以包含以下顺序：
如图5a所示，先调用原始数据获取模块110，再调用数据筛选模块120，再调用样本标注模块130，再调用增量学习模块240，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块460，再调用半监督训练模块250；
基于上述图3c中的实施例，在得到增量训练后的机器学习模型后，可以使用核验模块460基于上述基线测试集以及从原始数据中选择的测试集对上述增量训练后的机器学习模型进行测试，若上述模型符合要求(如，准确率高于88％)，则调用半监督训练模块250对上述增量训练后的机器学习模型进行训练；若上述模型不符合要求，则可以重新选择待标记数据进行标记，并对待训练机器学习模型基于上述新的新增样本以及基线训练集进行训练。
如图5b所示，先调用原始数据获取模块110，再调用样本标注模块130，再调用增量学习模块240，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再 调用核验模块460，再调用半监督训练模块250；
与上述5a中的实施例类似，此处不再赘述。
如图5c所示，先调用原始数据获取模块110，再调用数据筛选模块120，再调用半监督训练模块250，再调用核验模块460。
基于上述3e中的实施例，在使用半监督训练模块250对待训练机器学习模型进行训练后即可得到半监督训练后的机器学习模型，则可以调用核验模块460对该半监督训练后的机器学习模型进行核验，核验过程与上述图5a中的过程类似，此处不再赘述。
作为本申请实施例的一种具体实施方式，基于图4，如图6所示，上述机器学习系统还可以包括发布模块670；
上述发布模块670，用于在被调用时，将最后一个学习模块140输出的训练后的机器学习模型，作为目标模型进行发布。
本申请实施例中，可以将最终得到的训练后的机器学习模型作为目标模型，该模型可以实现用户所需的功能，如进行车辆识别等。得到目标模型后，即可将目标模型发布至相应的平台中，其中，相应的平台是指适用于用户所面临的真实应用场景的平台，如真实应用场景为识别道路上的车辆，则相应平台车辆识别平台，又如真实应用场景为识别来访人员的身份，则相应平台为人脸识别平台等，以供相关人员使用。
相应的，如图7所示，上述目标调用顺序可以包括：先调用原始数据获取模块110，再调用数据筛选模块120，再调用样本标注模块130，再调用增量学习模块240，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块460，再调用半监督训练模块250，最后用半监督训练模块输出的半监督训练后的机器学习模型替换原始的机器学习模型，由发布模块670发布。
图7所示的实施例与上述图5a中的实施例类似，此处不再赘述。
由上述的实施例可见，本申请实施例提供的这种机器学习系统，可以由用户灵活地选择学习方式，进而进行相应的模型训练，能够适应不同的应用场景，具有很好的鲁棒性。
基于与上述机器学习系统相同的申请构思，本申请实施例还提供了一种模型训练方法，如图8所示，图8为本申请实施例提供的模型训练方法的流程图，上述方法具体可以包括以下步骤：
步骤800，获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；
步骤810，获得数据采集设备捕获的当前应用场景的当前原始数据；
步骤820，基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
如图9所示，上述步骤820具体可以包括以下步骤：
步骤901，基于公共数据集对所述待训练机器学习模型进行N次基础训练，得到N个初始机器学习模型；其中，N大于或等于2；
步骤902，针对每个当前原始数据，将该当前原始数据分别输入N个初始机器学习模型，得到该当前原始数据的N个初始输出结果；
步骤903，将N个初始输出结果不同的当前原始数据，作为筛选出的有价值的目标原始数据。
作为本申请的一种具体实施例，上述待训练机器学习模型可以是ResNet50分类网络，上述公共数据集可以是ImageNet数据集，得到有价值的目标原始数据的过程可以是：随机初始化上述 ResNet50分类网络，并基于上述ImageNet数据集对上述模型进行训练；以上过程重复执行N次(N≥2)，即可得到N个分类模型。对于一输入图像数据，用上述N个分类模型进行测试，如果输出结果不同，则认为不一致，即该图像数据就是有价值的目标原始数据。
当然，本申请实施例中还可以采用其他基于主动学习技术的方式来进行数据筛选，例如，可以采用置信度最低(Least Confident)方法从原始数据中，筛选出目标原始数据，即选择最大概率最小的数据为有价值的目标原始数据。以使用一个训练好的二分类模型为例，若输入两个图像数据，第一个图像数据的类别预测概率为(0.9,0.1)，第二个图像数据的类别预测结果为(0.51,0.49)，也就是说，第一个图像被判定为第一类的概率为0.9，而第二个图像被判定为第一类的概率是0.51，即第二个图像数据对于该模型来说是更难区分的，那么第二个图像数据就是有价值的目标原始数据。本申请实施例中，从原始数据中筛选出目标数据还可以采用熵方法(Entropy)、期望模型变化方法(Expected Model Change)等，此处不做具体限定。
如图8所示，步骤830，按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
由上述的实施例可见，本申请实施例提供的这种模型训练方法，可以由用户灵活地选择学习方式，进而进行相应的模型训练，能够适应不同的应用场景，具有很好的鲁棒性。
作为本申请实施例的一种具体实施方式，上述目标学习方式可以是增量学习方式，即可以按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型。
作为本申请实施例的另一种具体实施方式，上述目标学习方式可以是半监督学习，即可以按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
当然，在进行机器训练时，可以同时使用上述增量学习方式以及半监督学习方式。基于图8，如图10所示，图10为本申请实施例中提供的机器训练方法的第二种流程图，上述步骤830具体可以包括以下步骤：
步骤1010，按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；
步骤1011，将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
作为本申请实施例的一种具体实施方式，基于图10，如图11所示，在上述步骤1011之前，还可以包括：
步骤1110，基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对所述增量训练后的机器学习模型进行测试，将测试结果输出给用户进行核验；若符合要求，则执行步骤1011；
相应的，在步骤1011之后还可以包括：
步骤1111，将训练后机器学习模型，作为目标模型进行发布。
本申请实施例提供的模型训练方法中，首先获取用户确定的待训练机器学习模型以及至少一个目标学习方式，之后获取数据采集设备捕获的当前应用场景的当前原始数据，并基于主动学习技术，从当前原始数据中筛选出有价值的目标原始数据，之后按照用户选择目标学习方法，对待训练机器学习模型进行训练，来得到训练后的机器学习模型。本申请实施例提供的模型训练方法，可以获取当前场景的原始数据，并按照用户选择的学习方式，基于筛选后的当前场景的原始数据，对待训练的机器学习模型进行训练，也就是说，当将机器学习模型应用于不同的应用场景时，并不需要开发人员对其进行重新设计，实现了对真实场景的自适应学习。
上述模型训练方法已在上述机器学习系统实施例中进行详细叙述，此处不再赘述。
基于与上述模型训练方法相同的技术构思，本申请实施例还提供了一种模型训练装置，如图12所示，上述装置可以包括：
学习方式获得模块1200，用于获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；
当前原始数据获得模块1201，用于获得数据采集设备捕获的当前应用场景的当前原始数据；
原始数据筛选模块1202，用于基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
模型训练模块1203，用于按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
本申请实施例提供的模型训练方法中，首先获取用户确定的待训练机器学习模型以及至少一个目标学习方式，之后获取数据采集设备捕获的当前应用场景的当前原始数据，并基于主动学习技术，从当前原始数据中筛选出有价值的目标原始数据，之后按照用户选择目标学习方法，对待训练机器学习模型进行训练，来得到训练后的机器学习模型。本申请实施例提供的模型训练方法，可以获取当前场景的原始数据，并按照用户选择的学习方式，基于筛选后的当前场景的原始数据，对待训练的机器学习模型进行训练，也就是说，当将机器学习模型应用于不同的应用场景时，并不需要开发人员对其进行重新设计，实现了对真实场景的自适应学习。
本申请实施例还提供了一种电子设备，如图13所示，包括处理器1301、通信接口1302、存储器1303和通信总线1304，其中，处理器1301，通信接口1302，存储器1303通过通信总线1304完成相互间的通信，
存储器1303，用于存放计算机程序；
处理器1301，用于执行存储器1303上所存放的程序时，实现如下步骤：
获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；
获得数据采集设备捕获的当前应用场景的当前原始数据；
基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
本申请实施例提供的模型训练方法中，首先获取用户确定的待训练机器学习模型以及至少一个目标学习方式，之后获取数据采集设备捕获的当前应用场景的当前原始数据，并基于主动学习技术， 从当前原始数据中筛选出有价值的目标原始数据，之后按照用户选择目标学习方法，对待训练机器学习模型进行训练，来得到训练后的机器学习模型。本申请实施例提供的模型训练方法，可以获取当前场景的原始数据，并按照用户选择的学习方式，基于筛选后的当前场景的原始数据，对待训练的机器学习模型进行训练，也就是说，当将机器学习模型应用于不同的应用场景时，并不需要开发人员对其进行重新设计，实现了对真实场景的自适应学习。
上述电子设备提到的通信总线可以是外设部件互连标准(Peripheral Component Interconnect，PCI)总线或扩展工业标准结构(Extended Industry Standard Architecture，EISA)总线等。该通信总线可以分为地址总线、数据总线、控制总线等。为便于表示，图中仅用一条粗线表示，但并不表示仅有一根总线或一种类型的总线。
通信接口用于上述电子设备与其他设备之间的通信。
存储器可以包括随机存取存储器(Random Access Memory，RAM)，也可以包括非易失性存储器(Non-Volatile Memory，NVM)，例如至少一个磁盘存储器。可选的，存储器还可以是至少一个位于远离前述处理器的存储装置。
上述的处理器可以是通用处理器，包括中央处理器(Central Processing Unit，CPU)、网络处理器(Network Processor，NP)等；还可以是数字信号处理器(Digital Signal Processor，DSP)、专用集成电路(Application Specific Integrated Circuit，ASIC)、现场可编程门阵列(Field-Programmable Gate Array，FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
在本申请提供的又一实施例中，还提供了一种计算机可读存储介质，该计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现上述任一模型训练方法的步骤。
在本申请提供的又一实施例中，还提供了一种包含指令的计算机程序产品，当其在计算机上运行时，使得计算机执行上述实施例中任一模型训练方法。
在上述实施例中，可以全部或部分地通过软件、硬件、固件或者其任意组合来实现。当使用软件实现时，可以全部或部分地以计算机程序产品的形式实现。所述计算机程序产品包括一个或多个计算机指令。在计算机上加载和执行所述计算机程序指令时，全部或部分地产生按照本申请实施例所述的流程或功能。所述计算机可以是通用计算机、专用计算机、计算机网络、或者其他可编程装置。所述计算机指令可以存储在计算机可读存储介质中，或者从一个计算机可读存储介质向另一个计算机可读存储介质传输，例如，所述计算机指令可以从一个网站站点、计算机、服务器或数据中心通过有线(例如同轴电缆、光纤、数字用户线(DSL))或无线(例如红外、无线、微波等)方式向另一个网站站点、计算机、服务器或数据中心进行传输。所述计算机可读存储介质可以是计算机能够存取的任何可用介质或者是包含一个或多个可用介质集成的服务器、数据中心等数据存储设备。所述可用介质可以是磁性介质，(例如，软盘、硬盘、磁带)、光介质(例如，DVD)、或者半导体介质(例如固态硬盘Solid State Disk(SSD))等。
需要说明的是，在本文中，诸如第一和第二等之类的关系术语仅仅用来将一个实体或者操作与另一个实体或操作区分开来，而不一定要求或者暗示这些实体或操作之间存在任何这种实际的关系或者顺序。而且，术语“包括”、“包含”或者其任何其他变体意在涵盖非排他性的包含，从而使得包括一系列要素的过程、方法、物品或者设备不仅包括那些要素，而且还包括没有明确列出的其他要素，或者是还包括为这种过程、方法、物品或者设备所固有的要素。在没有更多限制的情况下，由 语句“包括一个……”限定的要素，并不排除在包括所述要素的过程、方法、物品或者设备中还存在另外的相同要素。
本说明书中的各个实施例均采用相关的方式描述，各个实施例之间相同相似的部分互相参见即可，每个实施例重点说明的都是与其他实施例的不同之处。尤其，对于方法、装置、电子设备实施例而言，由于其基本相似于系统实施例，所以描述的比较简单，相关之处参见方法实施例的部分说明即可。
以上所述仅为本申请的较佳实施例，并不用以限制本申请，凡在本申请的精神和原则之内，所做的任何修改、等同替换、改进等，均应包含在本申请保护的范围之内。

Claims (17)

  1. 一种机器学习系统，其特征在于，所述系统包括：主控模块、原始数据获取模块、数据筛选模块、至少一种学习方式对应的学习模块和样本标注模块；
    所述主控模块，用于基于用户确定的待训练机器学习模型及选择的至少一个目标学习方式，确定需要调用的目标模块；所述需要调用的目标模块中至少包含原始数据获取模块和各个目标学习方式对应的各个目标学习模块；基于预设的系统中各个模块的前后执行顺序，确定各个目标模块的目标调用顺序；按所述目标调用顺序，调用各个目标模块，对所述待训练机器学习模型进行训练；
    所述原始数据获取模块，用于获得数据采集设备捕获的当前应用场景的当前原始数据；
    所述数据筛选模块，用于在被调用时，基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
    所述样本标注模块，用于在被调用时，将待标注数据输出给用户进行标注，得到当前新增样本；其中，待标注数据是用户从所述当前原始数据中选择的数据，或用户从所述数据筛选模块输出的目标原始数据中选择的数据；
    每个学习模块，用于在被调用时，按其对应的学习方式，基于预存的基线训练集中的训练样本和/或所述当前新增样本，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
  2. 根据权利要求1所述的机器学习系统，其特征在于，
    所述至少一种学习方式对应的学习模块中，包括：增量学习模块和/或半监督训练模块；
    所述增量学习模块，用于在被调用时，基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；
    所述半监督训练模块，用于在被调用时，基于当前原始数据或目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
  3. 根据权利要求2所述的机器学习系统，其特征在于，
    所述目标调用顺序包括以下顺序之一：
    先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块；
    先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块；
    先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用半监督训练模块；
    先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用半监督训练模块；
    先调用原始数据获取模块，再调用数据筛选模块，再调用半监督训练模块。
  4. 根据权利要求2所述的机器学习系统，其特征在于，所述机器学习系统，还包括：离线编码模块；
    所述离线编码模块，用于预先用所述基线训练集中的训练样本对基础模型进行训练，并在基础模型训练完成后，将基线训练集中的训练样本输入到基础模型中，得到基础特征并保存；所述基础特征，用于在增量学习模块进行增量训练或所述半监督训练模块进行半监督训练中代表基础训练集的样本。
  5. 根据权利要求2所述的机器学习系统，其特征在于，所述机器学习系统，还包括：核验模块；
    所述核验模块，用于在被调用时，基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对所述增量学习模块和/或所述半监督训练模块输出的训练后机器学习模型进行测试，将测试结果输出给用户进行核验。
  6. 根据权利要求5所述的机器学习系统，其特征在于，
    所述目标调用顺序包括以下顺序之一：
    先调用原始数据获取模块，再调用数据筛选模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块，再调用半监督训练模块；
    先调用原始数据获取模块，再调用样本标注模块，再调用增量学习模块，将增量学习模块输出的增量训练后的机器学习模型，作为更新后的待训练机器学习模型，再调用核验模块，再调用半监督训练模块；
    先调用原始数据获取模块，再调用数据筛选模块，再调用半监督训练模块，再调用核验模块。
  7. 根据权利要求1所述的机器学习系统，其特征在于，所述机器学习系统，还包括：发布模块；
    所述发布模块，用于在被调用时，将最后一个学习模块输出的训练后机器学习模型，作为升级后的模型进行发布。
  8. 一种模型训练方法，其特征在于，所述方法包括如下步骤：
    获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；
    获得数据采集设备捕获的当前应用场景的当前原始数据；
    基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
    按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
  9. 根据权利要求8所述的方法，其特征在于，
    所述基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据的步骤，包括：
    基于公共数据集对所述待训练机器学习模型进行N次基础训练，得到N个初始机器学习模型；其中，N大于或等于2；
    针对每个当前原始数据，将该当前原始数据分别输入N个初始机器学习模型，得到该当前原始数据的N个初始输出结果；
    将N个初始输出结果不同的当前原始数据，作为筛选出的有价值的目标原始数据。
  10. 根据权利要求8所述的方法，其特征在于，
    所述目标学习方式，包括：增量学习；
    所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：
    按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型。
  11. 根据权利要求8所述的方法，其特征在于，
    所述目标学习方式，包括：半监督学习；
    所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：
    按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
  12. 根据权利要求8所述的方法，其特征在于，
    所述目标学习方式，包括：增量学习和半监督学习；
    所述按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型的步骤，包括：
    按照增量学习方式，对所述目标原始数据进行标注，得到当前新增样本；基于预存的基线训练集中的训练样本和所述当前新增样本，对所述待训练机器学习模型进行增量训练，得到增量训练后的机器学习模型；
    将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型。
  13. 根据权利要求12所述的方法，其特征在于，
    在所述将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型的步骤之前，该方法还包括：
    基于预存的基线测试集和从当前原始数据或目标原始数据中选择的测试数据，对所述增量训练后的机器学习模型进行测试，将测试结果输出给用户进行核验；
    在核验结果为符合要求的情况下，执行所述将所述增量训练后的机器学习模型，作为更新后的待训练机器学习模型，按照半监督学习方式，基于所述目标原始数据中未被标注的数据，和预存的基线训练集中的训练样本，对更新后的待训练机器学习模型进行半监督训练，得到半监督训练后的机器学习模型的步骤。
  14. 根据权利要求8所述的方法，其特征在于，所述方法还包括：
    将训练后机器学习模型，作为升级后的模型进行发布。
  15. 一种模型训练装置，其特征在于，所述装置包括：
    学习方式获得模块，用于获得用户确定的待训练机器学习模型及选择的至少一个目标学习方式；
    当前原始数据获得模块，用于获得数据采集设备捕获的当前应用场景的当前原始数据；
    原始数据筛选模块，用于基于主动学习技术，从所述当前原始数据中，筛选出有价值的目标原始数据；
    模型训练模块，用于按照所述目标学习方式，基于所述目标原始数据，对所述待训练机器学习模型进行训练，得到训练后的机器学习模型。
  16. 一种电子设备，其特征在于，包括处理器、通信接口、存储器和通信总线，其中，处理器，通信接口，存储器通过通信总线完成相互间的通信；
    存储器，用于存放计算机程序；
    处理器，用于执行存储器上所存放的程序时，实现权利要求8-14任一所述的方法步骤。
  17. 一种计算机可读存储介质，其特征在于，所述计算机可读存储介质内存储有计算机程序，所述计算机程序被处理器执行时实现权利要求8-14任一所述的方法步骤。
PCT/CN2022/109701 2021-08-05 2022-08-02 一种机器学习系统及模型训练方法 WO2023011470A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202110898727.6A CN113516251B (zh) 2021-08-05 2021-08-05 一种机器学习系统及模型训练方法
CN202110898727.6 2021-08-05

Publications (1)

Publication Number Publication Date
WO2023011470A1 true WO2023011470A1 (zh) 2023-02-09

Family

ID=78066916

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/109701 WO2023011470A1 (zh) 2021-08-05 2022-08-02 一种机器学习系统及模型训练方法

Country Status (2)

Country Link
CN (1) CN113516251B (zh)
WO (1) WO2023011470A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113516251B (zh) * 2021-08-05 2023-06-06 上海高德威智能交通系统有限公司 一种机器学习系统及模型训练方法
CN114332873A (zh) * 2021-12-22 2022-04-12 北京三快在线科技有限公司 一种识别模型的训练方法及装置
CN115796272B (zh) * 2022-11-24 2024-03-12 北京百度网讯科技有限公司 基于深度学习平台的模型训练方法、数据处理方法和装置

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104318242A (zh) * 2014-10-08 2015-01-28 中国人民解放军空军工程大学 一种高效的svm主动半监督学习算法
CN108833409A (zh) * 2018-06-15 2018-11-16 北京网思科平科技有限公司 基于深度学习和半监督学习的webshell检测方法及装置
CN110610197A (zh) * 2019-08-19 2019-12-24 北京迈格威科技有限公司 一种困难样本挖掘和模型训练方法、装置和电子设备
US20200125897A1 (en) * 2018-10-18 2020-04-23 Deepnorth Inc. Semi-Supervised Person Re-Identification Using Multi-View Clustering
CN112347261A (zh) * 2020-12-07 2021-02-09 携程计算机技术（上海）有限公司 分类模型训练方法、系统、设备及存储介质
CN112559784A (zh) * 2020-11-02 2021-03-26 浙江智慧视频安防创新中心有限公司 基于增量学习的图像分类方法及系统
CN112784095A (zh) * 2021-01-18 2021-05-11 北京洛塔信息技术有限公司 困难样本数据挖掘方法、装置、设备和存储介质
US11043208B1 (en) * 2020-02-20 2021-06-22 Clinc, Inc. Systems and methods for mixed setting training for slot filling machine learning tasks in a machine learning task-oriented dialogue system
CN113516251A (zh) * 2021-08-05 2021-10-19 上海高德威智能交通系统有限公司 一种机器学习系统及模型训练方法

Family Cites Families (10)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108304936B (zh) * 2017-07-12 2021-11-16 腾讯科技（深圳）有限公司 机器学习模型训练方法和装置、表情图像分类方法和装置
CN107885762B (zh) * 2017-09-19 2021-06-11 北京百度网讯科技有限公司 智能大数据系统、提供智能大数据服务的方法和设备
CN109034188B (zh) * 2018-06-15 2021-11-05 北京金山云网络技术有限公司 机器学习模型的获取方法、获取装置、设备及存储介质
CN109409528A (zh) * 2018-09-10 2019-03-01 平安科技（深圳）有限公司 模型生成方法、装置、计算机设备及存储介质
CN109325541A (zh) * 2018-09-30 2019-02-12 北京字节跳动网络技术有限公司 用于训练模型的方法和装置
CN110705717B (zh) * 2019-09-30 2022-05-17 支付宝(杭州)信息技术有限公司 计算机执行的机器学习模型的训练方法、装置及设备
CN110968695A (zh) * 2019-11-18 2020-04-07 罗彤 基于弱监督技术主动学习的智能标注方法、装置及平台
CN111222648B (zh) * 2020-01-15 2023-09-26 深圳前海微众银行股份有限公司 半监督机器学习优化方法、装置、设备及存储介质
CN111783867A (zh) * 2020-06-28 2020-10-16 京东数字科技控股有限公司 机器学习算法选择方法及装置
CN112183577A (zh) * 2020-08-31 2021-01-05 华为技术有限公司 一种半监督学习模型的训练方法、图像处理方法及设备

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104318242A (zh) * 2014-10-08 2015-01-28 中国人民解放军空军工程大学 一种高效的svm主动半监督学习算法
CN108833409A (zh) * 2018-06-15 2018-11-16 北京网思科平科技有限公司 基于深度学习和半监督学习的webshell检测方法及装置
US20200125897A1 (en) * 2018-10-18 2020-04-23 Deepnorth Inc. Semi-Supervised Person Re-Identification Using Multi-View Clustering
CN110610197A (zh) * 2019-08-19 2019-12-24 北京迈格威科技有限公司 一种困难样本挖掘和模型训练方法、装置和电子设备
US11043208B1 (en) * 2020-02-20 2021-06-22 Clinc, Inc. Systems and methods for mixed setting training for slot filling machine learning tasks in a machine learning task-oriented dialogue system
CN112559784A (zh) * 2020-11-02 2021-03-26 浙江智慧视频安防创新中心有限公司 基于增量学习的图像分类方法及系统
CN112347261A (zh) * 2020-12-07 2021-02-09 携程计算机技术（上海）有限公司 分类模型训练方法、系统、设备及存储介质
CN112784095A (zh) * 2021-01-18 2021-05-11 北京洛塔信息技术有限公司 困难样本数据挖掘方法、装置、设备和存储介质
CN113516251A (zh) * 2021-08-05 2021-10-19 上海高德威智能交通系统有限公司 一种机器学习系统及模型训练方法

Also Published As

Publication number Publication date
CN113516251A (zh) 2021-10-19
CN113516251B (zh) 2023-06-06

Similar Documents

Publication Publication Date Title
WO2023011470A1 (zh) 一种机器学习系统及模型训练方法
EP3637310A1 (en) Method and apparatus for generating vehicle damage information
CN108197652B (zh) 用于生成信息的方法和装置
CN109345553B (zh) 一种手掌及其关键点检测方法、装置和终端设备
CN110363220B (zh) 行为类别检测方法、装置、电子设备和计算机可读介质
KR102002024B1 (ko) 객체 라벨링 처리 방법 및 객체 관리 서버
CN110084317B (zh) 用于识别图像的方法和装置
CN113140012B (zh) 图像处理方法、装置、介质及电子设备
CN113778864A (zh) 一种测试用例的生成方法和装置、电子设备和存储介质
US11115338B2 (en) Intelligent conversion of internet domain names to vector embeddings
CN112259078A (zh) 一种音频识别模型的训练和非正常音频识别的方法和装置
CN114064872B (zh) 对话数据信息的智能存储方法、装置、设备及介质
CN114638304A (zh) 图像识别模型的训练方法、图像识别方法及装置
CN112434717A (zh) 一种模型训练方法及装置
CN117475364A (zh) 图片检测方法、设备及计算机存储介质
CN114065858A (zh) 一种模型训练方法、装置、电子设备及存储介质
CN116188887A (zh) 属性识别预训练模型生成方法、属性识别模型生成方法
CN116189109A (zh) 模型训练方法、道路事件检测方法、装置及相关设备
CN112083974B (zh) 一种广告窗口关闭方法、装置及电子设备
US20200073891A1 (en) Systems and methods for classifying data in high volume data streams
CN110827261B (zh) 图像质量检测方法及装置、存储介质及电子设备
CN110334763B (zh) 模型数据文件生成、图像识别方法、装置、设备及介质
CN113609018A (zh) 测试方法、训练方法、装置、设备、介质和程序产品
CN113569879A (zh) 异常识别模型的训练方法、异常账号识别方法及相关装置
CN118172547B (zh) 图像目标识别方法、装置、电子设备和计算机可读介质

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: 22852182

Country of ref document: EP

Kind code of ref document: A1

WD Withdrawal of designations after international publication
NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 22852182

Country of ref document: EP

Kind code of ref document: A1