WO2019200748A1 - 迁移学习方法、装置、计算机设备和存储介质 - Google Patents

迁移学习方法、装置、计算机设备和存储介质 Download PDF

Info

Publication number
WO2019200748A1
WO2019200748A1 PCT/CN2018/095497 CN2018095497W WO2019200748A1 WO 2019200748 A1 WO2019200748 A1 WO 2019200748A1 CN 2018095497 W CN2018095497 W CN 2018095497W WO 2019200748 A1 WO2019200748 A1 WO 2019200748A1
Authority
WO
WIPO (PCT)
Prior art keywords
parameter
data
feature extraction
training
extraction layer
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2018/095497
Other languages
English (en)
French (fr)
Inventor
韩茂琨
王健宗
肖京
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Ping An Technology Shenzhen Co Ltd
Original Assignee
Ping An Technology Shenzhen Co Ltd
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 Ping An Technology Shenzhen Co Ltd filed Critical Ping An Technology Shenzhen Co Ltd
Publication of WO2019200748A1 publication Critical patent/WO2019200748A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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

Definitions

  • the present application relates to the field of machine learning, and more particularly to migration learning methods, apparatus, computer equipment, and storage media.
  • the main purpose of the present application is to provide a migration learning method, which aims to solve the technical problem that the limited annotation data leads to poor generalization ability of the training model.
  • the present application provides a migration learning method applied to a migration learning network, the migration learning network includes a task training network and a domain classification network sharing the same feature extraction layer, the task training network further includes a task training model, and the domain classification network further includes a domain classification.
  • methods include:
  • the training result obtained by the feature extraction layer training according to the annotation data is migrated to the unlabeled data, and the unlabeled data is identified.
  • the application also provides a migration learning device, the migration learning network includes a task training network and a domain classification network that share the same feature extraction layer, the task training network further includes a task training model, and the domain classification network further includes a domain classifier, and the device includes:
  • a first input module configured to input the annotation data of the specified task into the task training model to obtain a first parameter of the feature extraction layer
  • a second input module configured to lock the first parameter, and input the unlabeled data and the annotation data into the domain classification network to obtain the domain classifier to distinguish the unlabeled data from the annotation data.
  • a training module configured to lock the second parameter, the domain classifier and the task training model jointly train the feature extraction layer to obtain the feature extraction layer without distinguishing the unlabeled data and the annotation data Third parameter;
  • a migration module configured to migrate the training result obtained by the feature extraction layer according to the annotation data to the unlabeled data based on the third parameter, and identify the unlabeled data.
  • the application also provides a computer device comprising a memory and a processor, the memory storing computer readable instructions, the processor implementing the steps of the method when the computer readable instructions are executed.
  • the present application also provides a computer non-transitory readable storage medium having stored thereon computer readable instructions that, when executed by a processor, implement the steps of the methods described above.
  • the present application realizes the effective identification of unlabeled data by expanding the migration learning data from the annotation data to the unlabeled data, expanding the data annotation, and improving the generalization ability of the existing model; and extracting the feature of the labeled data through the migration learning network.
  • the method migrates to the unlabeled data, performs automatic label recognition, improves the mapping scope of the label data, makes the data labeling more convenient, fast, accurate, and has a wider migration range, which is more conducive to the promotion and popularization of the model;
  • the migration learning network of the classification training network and the task training network does not distinguish between data types, expands the scope of application, and realizes effective migration from labeled data to unlabeled data, saving a lot of manpower, material resources and financial resources for data labeling.
  • FIG. 1 is a schematic flow chart of a migration learning method according to an embodiment of the present application.
  • FIG. 2 is a schematic structural diagram of a migration learning apparatus according to an embodiment of the present application.
  • FIG. 3 is a schematic structural diagram of a training module according to an embodiment of the present application.
  • FIG. 4 is a schematic structural diagram of a training unit according to an embodiment of the present application.
  • FIG. 5 is a schematic structural diagram of a migration learning apparatus according to another embodiment of the present application.
  • FIG. 6 is a schematic structural diagram of a migration learning apparatus according to still another embodiment of the present application.
  • FIG. 7 is a schematic structural diagram of a migration learning apparatus according to still another embodiment of the present application.
  • FIG. 8 is a schematic structural diagram of a migration learning apparatus according to still another embodiment of the present application.
  • FIG. 9 is a schematic diagram of the internal structure of a computer according to an embodiment of the present application.
  • a migration learning method applied to a migration learning network includes a task training network and a domain classification network sharing the same feature extraction layer, and the task training network further includes a task training model and a domain classification network. Also included is a domain classifier, the method comprising:
  • S1 Input the annotation data of the specified task into the task training model for model training to obtain the first parameter of the feature extraction layer.
  • the type and content of the annotation data in this step differ according to the specified task, and the content of the annotation data is several thousand, tens of thousands or hundreds of thousands of images.
  • the task training network of this step includes a feature extraction layer Mf and a task training model Mt.
  • the Mf of the embodiment mainly receives the picture input, and outputs the picture after the encoding, and the Mf uses the VGG-16 (Visual The Geometry Group, a convolutional neural network structure developed by the Visual Geometry Research Group at Oxford University, 16 indicates its depth.
  • VGG-16 Visual The Geometry Group, a convolutional neural network structure developed by the Visual Geometry Research Group at Oxford University, 16 indicates its depth.
  • the model and weight parameters that have been trained to predict the type of the picture, eliminating the need to train the neural network weight parameters to improve feature extraction.
  • the accuracy and the training cost are saved, and the migration capability evaluation of the migration learning network in the embodiment is further improved.
  • VGG-16 uses two convolutional layers, each with 64 filters, 2 convolutions -> Pools -> 2 convolutions -> Pool -> 3 convolutions - >Pool->3 convolutions->Pool->3 convolutions->Pool->full connection->full connection->Softmax (Softmax is a function, the main application is multi-classification, Softmax function is an extension of Sigmoid function The Sigmoid function can only be divided into two categories, and Softmax can be divided into multiple classes).
  • the filter settings for the convolutional layer and the pooled layer of VGG-16 are as follows: convolution kernel: 3*3; step size: 1; use maximum pooling kernel: 2*2; step size: 2.
  • the number of convolutional layers is from 64->128->256->512->512.
  • the number of convolution kernels increases exponentially.
  • the first step of the task training network of the present embodiment is to train the task training model.
  • the annotation data here, the image
  • the output or prediction result of the task training model is expected.
  • the result ie the label of the image
  • the weight of the task training model will be adjusted to reduce the error rate of the prediction. That is, by adjusting the weight, the matching rate between the prediction result of the task training model and the actual image label can be improved.
  • the task training model converges, the matching result of the task training model and the actual image label are the best.
  • the task training model can be used for inference, that is, predicting the image data read by it, and outputting: predicting the image (including class name, semantic concept, prediction probability, etc.).
  • Mt represents a task training model.
  • a full layer of 1024 nodes and a layer of Softmax are adopted.
  • the Mf and Mt are trained by labeling the data back-propagation.
  • the optimal weight of the Mf is obtained.
  • Mt has a good recognition effect on the eigenvectors of the VGG-16 convolutional layer and the pooled layer output during training convergence.
  • the convolutional layer weight and the pooling layer weight of the VGG-16 are the first parameters of the Mf. At this time, the eigenvectors of the Mf output have better performance in the feature recognition task.
  • S2 The first parameter is locked, and the unlabeled data and the label data are mixed into the domain classification network to obtain a second parameter that distinguishes the unlabeled data and the label data by the domain classifier.
  • the unlabeled data is B data (no detailed annotation information), and the annotation data is A data (with detailed annotation information).
  • the domain classification network in this embodiment includes a feature extraction layer Mf and a two-category domain classifier Md.
  • Md is a domain classifier, which is implemented by using a two-layer fully-connected layer and a layer of Softmax layer.
  • the migration learning capability of this embodiment lies in The domain classifier Md is added.
  • the domain classification network and the task training network of this embodiment share the same feature extraction layer Mf. In this step, when training the two-category domain classification network, the first parameter of the Mf is locked, and the locking keeps the parameter of the Mf unchanged.
  • the Mf By locking the first parameter of the Mf (ie, the weight parameter of the Mf identifying the labeled data), the Mf remains in the A. The eigenvector recognition of the data is best. Then, the A data and the B data are randomly mixed to jointly train the Md. When the two classification domain classification network trains to converge, it indicates that the categorical domain classifier Md can be based on the eigenvector of the current Mf output. A data and B data are optimally distinguished.
  • the A data and the B data are randomly mixed in the training process, and then input into the two-category domain classification network, and the A data and the B data simultaneously train the Md, and the weight parameter of the corresponding two-category domain classifier is recorded during the training convergence, that is, the second parameter Therefore, Md has the best effect on distinguishing between A data and B data under the second parameter.
  • the third parameter in this embodiment is a weight parameter when Mf does not distinguish between the labeled data and the unlabeled data.
  • the Mf can not distinguish between the A data and the B data, and directly migrate the training model to the unlabeled data of other application fields, thereby improving the generalization ability of the training model, so as to expand the application scope of the training model. Further promote the promotion and popularization of training models.
  • the Mf can be used to distinguish the data A and the data B from each other, and the annotation information feature of the A data is migrated to the B data, so that the B data has the same or similar data distribution as the A data.
  • step S3 includes:
  • Mf's cross-entropy LOSS function is combined with Mt's cross-entropy LOSS function to jointly back-train Mf, so as to use the parameters of the already trained domain classifier Md while training Mf, to reach the final Mf parameter (ie, the third
  • Mf's cross-entropy LOSS function is combined with Mt's cross-entropy LOSS function to jointly back-train Mf, so as to use the parameters of the already trained domain classifier Md while training Mf, to reach the final Mf parameter (ie, the third
  • the parameter can not make any distinction between the label data and the unlabeled data in different fields, and the feature recognition effect on the label data is also used for the unlabeled data.
  • step S30 includes:
  • the original LOSS function indicates that the original feature extraction layer has a better distinction between the two classification data.
  • the gradient entropy effect of the Md cross-entropy LOSS function is -2*log0.5, and the discrimination effect between the A data and the B data is the worst. The effect that the feature extraction layer Mf does not distinguish the two-category data is achieved, and the migration of the feature extraction layer Mf from the A data to the B data can be realized.
  • LOSS -(log2P) 2 -[log(2-2P)] 2 , where P is two
  • P is two
  • the calibration of a certain classification in the classification data reaches the minimum value of the modified LOSS function when the Softmax output of the model is [0.5, 0.5], and the feature extraction layer Mf does not make any distinction between the two classification data.
  • S302 Lock the second parameter, add the modified cross-entropy LOSS function of the domain classifier and the cross-entropy LOSS function of the task training model to obtain a sum of LOSS functions.
  • the cross-entropy LOSS function of the modified domain classifier and the cross-entropy LOSS function of the task training model are added to obtain the sum of the LOSS functions, so that the cross-entropy LOSS function of Md is shared with the cross-entropy LOSS function of Mt. Training the parameters of Mf, so that the cross-entropy LOSS function of Md and the cross-entropy LOSS function of Mt have the same trend of change to Mf.
  • the LOSS function of the domain classifier and the LOSS function of the task training model are added to obtain the sum of the LOSS functions, and then the feature extraction layer Mf is trained by gradient back propagation, and the LOSS function and task of the domain classifier are used.
  • the LOSS function of the training model approaches the minimum
  • the sum of the LOSS functions approaches the minimum to achieve the optimization purpose.
  • the training purpose of this step is to use the parameters of the already trained domain classifier while training the task model, so that the final Mf does not distinguish between the labeled data and the unlabeled data of different fields, so that the Mf labels data in different fields and Effective migration between unlabeled data.
  • S304 Determine to determine the third parameter when the sum of the LOSS functions reaches a preset value.
  • the preset value in this step is a certain value that approaches zero, such as 10 -6 , 10 -12, etc., so that the sum of the LOSS functions reaches a trend approaching zero, and the output result of the domain classifier Md at this time.
  • [0.5, 0.5] indicating that the feature extraction layer Mf has the weakest distinction between the labeled A data and the unlabeled B data in different fields, or does not distinguish between the A data and the B data, and trains the Mf to mark different data in different fields.
  • the weight parameter of the corresponding Mf is the third parameter.
  • step S3 the method includes:
  • the above steps S1, S2, and S3 are iteratively performed multiple times, and the learning rate and the gradient ratio are continuously adjusted to correct the weight of the LOSS function to implement Mf, Mt, and Mt in the migration learning network. Convergence of various parts of Md.
  • This embodiment uses a gradient descent algorithm to find a feasible solution or an optimal solution of the cross entropy function by iteration.
  • the convergence is very slow due to the difference between the dimensions, and the gradient descent algorithm is affected by the learning rate every iteration. The choice of learning rate is very important.
  • the learning rate is changed.
  • the learning rate is also large.
  • the learning rate is also small.
  • the learning rate is positively correlated with the gradient. There is a proportional coefficient, and a fixed proportional coefficient is used.
  • the values of the test learning rate are started from the following numerical values, 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, ....
  • the initial learning rate is 0.001, which is not expected to be multiplied by 3 times and replaced with 0.003. If it is not in line with expectations, then replace it with 0.01, so cycle until the most appropriate learning rate is found. That is, when selecting an appropriate learning rate for the gradient descent algorithm, a series of learning rate values can be selected by a multiple of 3 and then a multiple of 10 until a minimum value is found and another maximum value is found. The largest learning rate value, or a learning rate value slightly smaller than the maximum value, is the best learning rate value. Experiments have shown that the learning rate of this embodiment is better when it is set to be around 0.1 or 0.1.
  • S32 Iteratively iterate the training process specified times of the first parameter, the second parameter, and the third parameter to obtain an optimization parameter of the feature extraction layer of the migration learning network and an optimization parameter of the domain classifier.
  • the learning rate and the gradient ratio are continuously adjusted by the gradient descent calculation method to gradually optimize the Mf, Mt, and Md of the migration learning network through multiple iteration training.
  • the parameters of each part improve the migration learning effect. For example, after one step S1 to step S3, the learning learning effect of the learning network from A data to B data is migrated, and the recognition rate of the B data is 90%, and the steps S1 to S3 are performed after multiple iterations. After training, the learning and learning effect of the migration learning network from A data to B data is expressed as 98% or nearly 100% of the B data, which greatly improves the generalization ability of the migration learning network.
  • the face image data is marked as the tag data
  • the face data of the designated type of animal is the unlabeled data
  • the specified task is a face recognition task
  • S33 locking the optimization parameter of the feature extraction layer in the migration learning network, the optimization parameter of the domain classifier, and the recognition model parameter of the face recognition task, and inputting the labeled face data and the data of the unspecified type animal into the migration. Learn the network for migration learning.
  • This embodiment is an application verification of the migration learning network.
  • This embodiment is an example of migrating from annotation data of one domain to label recognition of unlabeled data of another domain.
  • the A data is the face image data for face recognition
  • the B data is the face data of other animals such as unlabeled pig face, cat face or face.
  • the parameters of each part of the Mf, Mt, and Md of the migration learning network are determined by labeling the face image data.
  • the Mf, Mt, and Md parameters of the determined migration learning network are used to migrate from the labeled face image data to the unlabeled pig face image data for feature labeling.
  • the labeled face image A data for face recognition and the unlabeled pig face image B data are input into the migration learning network optimized for each parameter, and the pig face data finally outputted by the learning network is migrated.
  • the correct labeling rate is as high as 99%, which indicates that the migration learning network of the present application achieves a good migration learning ability from label data to unlabeled data, and saves a lot of manpower, material resources and financial resources for data labeling.
  • Other embodiments of the present application can also be used for direct cross-domain migration of the migration learning network, so that migration learning achieves the same or similar recognition effect in the field of labeled data and the new business domain without data annotation, thereby eliminating the labeling of unlabeled data. jobs.
  • step S34 in this embodiment the method includes:
  • S35 Enter the specified type of animal annotation data into an animal category recognition model for model training.
  • the animal category recognition model in this step is trained by inputting the pig face image B data obtained by the above-mentioned migration learning information into the convolutional neural network to obtain a recognition model capable of accurately identifying the animal category, and can select according to different recognition targets. And pre-set the structure of the convolutional neural network.
  • the animal category recognition model of the present embodiment is a classification of local pig breeds in China, and the above animal category recognition model is used to effectively identify the classified information of existing farmed pigs, thereby providing powerful data analysis for the aquaculture industry.
  • China's local pig breeds can be divided into six types: North China, South China, Central China, Jianghai, Southwest, and Plateau. Each type has many unique types of pigs.
  • the original work of manually labeling data is extremely large and prone to errors.
  • the pig face data is marked, which greatly improves the business data processing capability.
  • S36 Input the pig face image to be classified into the animal category recognition model to obtain output data of the animal category classification in the animal category recognition model.
  • the pig face data to be classified is input into the above animal category recognition model to directly identify the species of the pig, and provide strong data support for the aquaculture industry, the pork market and the like, and improve productivity.
  • step S36 the method further includes:
  • the actual application capability of the annotation data obtained by the migration learning is further illustrated by the application details after the data annotation work.
  • a gender recognition model for effective classification of sows and boars is obtained.
  • the sows and boars have different meat quality.
  • the gender identification model the sex information of the pigs is obtained, which is more conducive to the convenience of the breeding owners according to the convenience of the sows and boars, so as to distinguish between the conservation feed and the breeding area in the process of large-scale breeding pig breeding. In order to effectively control the farming process according to the needs of the pork market.
  • the pig face labeling data obtained by the migration learning provides a more detailed difference to the pig user, so as to more specifically refine the work in the breeding process, including feed matching, breeding time control, etc., so that the breeding industry is more scientific. More data can be relied upon.
  • the present application realizes the effective identification of unlabeled data by expanding the migration learning data from the annotation data to the unlabeled data, expanding the data annotation, and improving the generalization ability of the existing model; and extracting the feature of the labeled data through the migration learning network.
  • Mode migration to unlabeled data automatic label recognition, improve the scope of data labeling, make data labeling more convenient, fast, accurate, and a wider range of migration, making it more conducive to the promotion and popularization of models; creating domain-based classification
  • the migration learning network of the training network and the task training network does not distinguish between data types, expands the scope of application, and realizes effective migration from labeled data to unlabeled data, saving a lot of manpower, material resources and financial resources for data labeling.
  • the present application further provides a migration learning device, where a migration learning network includes a task training network and a domain classification network sharing the same feature extraction layer, the task training network further includes a task training model, and the domain classification network further includes a domain classifier.
  • the device includes:
  • the first input module 1 is configured to input the annotation data of the specified task into the task training model for model training to obtain the first parameter of the feature extraction layer.
  • the type and content of the annotation data in this embodiment differ according to the specified task, and the content of the annotation data is several thousand, tens of thousands or hundreds of thousands of pictures.
  • the task training network of this embodiment includes a feature extraction layer Mf and a task training model Mt.
  • the Mf of the embodiment mainly receives the picture input, and outputs the picture after the encoding, and the Mf uses the VGG-16 (Visual The Geometry Group, a convolutional neural network structure developed by the Visual Geometry Research Group at Oxford University, 16 indicates its depth.
  • VGG-16 Visual The Geometry Group, a convolutional neural network structure developed by the Visual Geometry Research Group at Oxford University, 16 indicates its depth.
  • the model and weight parameters that have been trained to predict the type of the picture, eliminating the need to train the neural network weight parameters to improve feature extraction.
  • the accuracy and the training cost are saved, and the migration capability evaluation of the migration learning network in the embodiment is further improved.
  • VGG-16 uses two convolutional layers, each with 64 filters, 2 convolutions -> Pools -> 2 convolutions -> Pool -> 3 convolutions - >Pool->3 convolutions->Pool->3 convolutions->Pool->full connection->full connection->Softmax (Softmax is a function, the main application is multi-classification, Softmax function is an extension of Sigmoid function The Sigmoid function can only be divided into two categories, and Softmax can be divided into multiple classes).
  • the filter settings for the convolutional layer and the pooled layer of VGG-16 are as follows: convolution kernel: 3*3; step size: 1; use maximum pooling kernel: 2*2; step size: 2.
  • the number of convolutional layers is from 64->128->256->512->512.
  • the number of convolution kernels increases exponentially.
  • the first step of the task training network of the present embodiment is to train the task training model.
  • the annotation data here, the image
  • the output or prediction result of the task training model is expected.
  • the result ie the label of the image
  • the weight of the task training model will be adjusted to reduce the error rate of the prediction. That is, by adjusting the weight, the matching rate between the prediction result of the task training model and the actual image label can be improved.
  • the task training model converges, the matching result of the task training model and the actual image label are the best.
  • the task training model can be used for inference, that is, predicting the image data read by it, and outputting: predicting the image (including class name, semantic concept, prediction probability, etc.).
  • Mt represents a task training model.
  • a full layer of 1024 nodes and a layer of Softmax are adopted.
  • Mf and Mt are trained by labeling data normal back propagation, and when Mt training converges, the optimal weight of Mf is obtained.
  • Mt has a good recognition effect on the eigenvectors of the VGG-16 convolutional layer and the pooled layer output during training convergence.
  • the convolutional layer weight and the pooling layer weight of the VGG-16 are the first parameters of the Mf. At this time, the eigenvectors of the Mf output have better performance in the feature recognition task.
  • the second input module 2 is configured to lock the first parameter, and input the unlabeled data and the label data into the domain classification network to obtain a second parameter that the domain classifier distinguishes between the unlabeled data and the label data.
  • the unlabeled data is B data (no detailed annotation information), and the annotation data is A data (with detailed annotation information).
  • the domain classification network in this embodiment includes a feature extraction layer Mf and a two-category domain classifier Md.
  • Md is a domain classifier, which is implemented by using a two-layer fully-connected layer and a layer of Softmax layer.
  • the migration learning capability of this embodiment lies in The domain classifier Md is added.
  • the domain classification network and the task training network of this embodiment share the same feature extraction layer Mf. In the embodiment, when the two-category domain classification network is trained, the first parameter of the Mf is locked, and the lock retains the parameter of the Mf unchanged.
  • the Mf By locking the first parameter of the Mf (ie, the weight parameter of the Mf identifying the annotation data), the Mf remains in the pair.
  • the eigenvector recognition effect of A data is the best.
  • the A data and the B data are randomly mixed to jointly train Md.
  • the two classification domain classification network training converges, it indicates that the categorical domain classifier Md can be based on the eigenvector of the current Mf output.
  • the best effective distinction between A data and B data In this embodiment, the A data and the B data are randomly mixed in the training process, and then input into the two-category domain classification network.
  • the A data and the B data simultaneously train the Md, and the weight parameters of the corresponding two-category domain classifier are recorded during the training convergence, that is, the second The parameter, so Md has the best effect on distinguishing between A data and B data under the second parameter.
  • a training module 3 configured to lock the second parameter, the domain classifier and the task training model jointly train the feature extraction layer to obtain the feature extraction layer without distinguishing the unlabeled data and the label data The third parameter.
  • the third parameter in this embodiment is a weight parameter when Mf does not distinguish between the labeled data and the unlabeled data.
  • the migration module 4 is configured to migrate the training result obtained by the feature extraction layer according to the annotation data to the unlabeled data based on the third parameter, and identify the unlabeled data.
  • the Mf can not distinguish between the A data and the B data, and directly migrate the training model to the unlabeled data of other application fields, thereby improving the generalization ability of the training model, so as to expand the application scope of the training model. Further promote the promotion and popularization of training models.
  • the Mf can be used to distinguish the data A and the data B from each other, and the annotation information feature of the A data is migrated to the B data, so that the B data has the same or similar data distribution as the A data.
  • the training module 3 of this embodiment includes:
  • the training unit 30 is configured to jointly back-train the feature extraction layer by using a cross entropy LOSS function of the domain classifier and a cross entropy LOSS function of the task training model to obtain the feature extraction layer without distinguishing the Annotating data and a third parameter of the specified annotation data.
  • Mf's cross-entropy LOSS function is combined with Mt's cross-entropy LOSS function to jointly back-train Mf, so as to use the parameters of the already trained domain classifier Md while training Mf, to reach the final Mf parameter (ie, the third
  • Mf's cross-entropy LOSS function is combined with Mt's cross-entropy LOSS function to jointly back-train Mf, so as to use the parameters of the already trained domain classifier Md while training Mf, to reach the final Mf parameter (ie, the third
  • the parameter can not make any distinction between the label data and the unlabeled data in different fields, and the feature recognition effect on the label data is also used for the unlabeled data.
  • the training unit 30 of this embodiment includes:
  • the original LOSS function indicates that the original feature extraction layer has a better distinction between the two classification data.
  • the gradient entropy effect of the Md cross-entropy LOSS function is -2*log0.5, and the discrimination effect between the A data and the B data is the worst. The effect that the feature extraction layer Mf does not distinguish the two-category data is achieved, and the migration of the feature extraction layer Mf from the A data to the B data can be realized.
  • LOSS -(log2P) 2 -[log(2-2P)] 2 , where P is two
  • P is two
  • the calibration of a certain classification in the classification data reaches the minimum value of the modified LOSS function when the Softmax output of the model is [0.5, 0.5], and the feature extraction layer Mf does not make any distinction between the two classification data.
  • the adding sub-unit 302 is configured to lock the second parameter, and add the modified cross-entropy LOSS function of the domain classifier and the cross-entropy LOSS function of the task training model to obtain a sum of LOSS functions.
  • the cross-entropy LOSS function of the modified domain classifier and the cross-entropy LOSS function of the task training model are added to obtain the sum of the LOSS functions, so that the cross-entropy LOSS function of Md is shared with the cross-entropy LOSS function of Mt. Training the parameters of Mf, so that the cross-entropy LOSS function of Md and the cross-entropy LOSS function of Mt have the same trend of change to Mf.
  • the training sub-unit 303 is configured to train the feature extraction layer by the gradient back propagation of the LOSS function.
  • the LOSS function of the domain classifier and the LOSS function of the task training model are added to obtain the sum of the LOSS functions, and then the feature extraction layer Mf is trained by gradient back propagation, and the LOSS function and task of the domain classifier are used.
  • the LOSS function of the training model approaches the minimum
  • the sum of the LOSS functions approaches the minimum to achieve the optimization purpose.
  • the training objective of this embodiment is to use the parameters of the already trained domain classifier while training the task model, so that the final Mf does not distinguish between the labeled data and the unlabeled data of different domains, so that the Mf is labeled data in different fields. Effective migration between unlabeled data.
  • the determining subunit 304 is configured to determine the third parameter when the sum of the LOSS functions reaches a preset value.
  • the preset value in this embodiment is a certain value that approaches zero, such as 10 -6 , 10 -12 , etc., so that the sum of the LOSS functions reaches a trend of approaching zero, and the Md output of the domain classifier at this time
  • the result is [0.5, 0.5], which means that the feature extraction layer Mf has the weakest discrimination effect on the labeled data A and the unlabeled B data in different fields, or does not distinguish between the A data and the B data, and trains the Mf to mark the different fields.
  • the weight parameter of the corresponding Mf is the third parameter.
  • a migration learning apparatus includes:
  • the optimization module 31 is configured to optimize the learning rate and the gradient ratio during the training process.
  • the above steps S1, S2, and S3 are iteratively performed multiple times, and the learning rate and the gradient ratio are continuously adjusted to correct the weight of the LOSS function to implement Mf, Mt, and Mt in the migration learning network. Convergence of various parts of Md.
  • This embodiment uses a gradient descent algorithm to find a feasible solution or an optimal solution of the cross entropy function by iteration.
  • the convergence is very slow due to the difference between the dimensions, and the gradient descent algorithm is affected by the learning rate every iteration. The choice of learning rate is very important.
  • the learning rate is changed.
  • the learning rate is also large.
  • the learning rate is also small.
  • the learning rate is positively correlated with the gradient. There is a proportional coefficient, and a fixed proportional coefficient is used.
  • the values of the test learning rate are started from the following numerical values, 0.001, 0.003, 0.01, 0.03, 0.1, 0.3, 1, ....
  • the initial learning rate is 0.001, which is not expected to be multiplied by 3 times and replaced with 0.003. If it is not in line with expectations, then replace it with 0.01, so cycle until the most appropriate learning rate is found. That is, when selecting an appropriate learning rate for the gradient descent algorithm, a series of learning rate values can be selected by a multiple of 3 and then a multiple of 10 until a minimum value is found and another maximum value is found. The largest learning rate value, or a learning rate value slightly smaller than the maximum value, is the best learning rate value. Experiments have shown that the learning rate of this embodiment is better when it is set to be around 0.1 or 0.1.
  • the iterative module 32 is configured to sequentially iterate the training process specified times of the first parameter, the second parameter, and the third parameter to obtain an optimization parameter of the feature extraction layer and a domain classifier optimization parameter of the migration learning network.
  • the learning rate and the gradient ratio are continuously adjusted by the gradient descent calculation method to gradually optimize the Mf, Mt, and Md of the migration learning network through multiple iteration training.
  • the parameters of each part improve the migration learning effect. For example, after one step S1 to step S3, the learning learning effect of the learning network from A data to B data is migrated, and the recognition rate of the B data is 90%, and the steps S1 to S3 are performed after multiple iterations. After training, the learning and learning effect of the migration learning network from A data to B data is expressed as 98% or nearly 100% of the B data, which greatly improves the generalization ability of the migration learning network.
  • the face image data is labeled as the annotation data
  • the facial data of the designated animal is the unlabeled data
  • the designated task is a face recognition task
  • the migration learning module 33 is configured to lock optimization parameters of the feature extraction layer in the migration learning network, optimization parameters of the domain classifier, and recognition model parameters of the face recognition task, and the face data and the data of the specified species are not marked.
  • the migration learning network is mixedly input for migration learning.
  • This embodiment is an application verification of the migration learning network.
  • This embodiment is an example of migrating from annotation data of one domain to label recognition of unlabeled data of another domain.
  • the A data is the face image data for face recognition
  • the B data is the face data of other animals such as unlabeled pig face, cat face or face.
  • the parameters of each part of the Mf, Mt, and Md of the migration learning network are determined by labeling the face image data.
  • the Mf, Mt, and Md parameters of the determined migration learning network are used to migrate from the labeled face image data to the unlabeled pig face image data for feature labeling.
  • the output module 34 is configured to output the annotation data of the designated species of animals.
  • the labeled face image A data for face recognition and the unlabeled pig face image B data are input into the migration learning network optimized for each parameter, and the pig face data finally outputted by the learning network is migrated.
  • the correct labeling rate is as high as 99%, which indicates that the migration learning network of the present application achieves a good migration learning ability from label data to unlabeled data, and saves a lot of manpower, material resources and financial resources for data labeling.
  • Other embodiments of the present application can also be used for direct cross-domain migration of the migration learning network, so that migration learning achieves the same or similar recognition effect in the field of labeled data and the new business domain without data annotation, thereby eliminating the labeling of unlabeled data. jobs.
  • a migration learning apparatus further includes:
  • the third input module 35 is configured to input the specified type of animal annotation data into the animal category recognition model for model training.
  • the animal category recognition model in the embodiment is trained by inputting the pig face image B data obtained by the above-mentioned migration learning information into the convolutional neural network to obtain a recognition model capable of accurately identifying the animal category, and can perform the identification according to different recognition targets. Select and pre-set the composition of the convolutional neural network.
  • the animal category recognition model of the present embodiment is a classification of local pig breeds in China, and the above animal category recognition model is used to effectively identify the classified information of existing farmed pigs, thereby providing powerful data analysis for the aquaculture industry.
  • China's local pig breeds can be divided into six types: North China, South China, Central China, Jianghai, Southwest, and Plateau. Each type has many unique types of pigs. The original work of manually labeling data is extremely large and prone to errors.
  • the pig face data is marked, which greatly improves the business data processing capability.
  • the first obtaining module 36 is configured to input the image of the pig face to be classified into the animal category recognition model to obtain output data of the animal category classification in the animal category recognition model.
  • the pig face data to be classified is input into the above animal category recognition model to directly identify the species of the pig, and provide strong data support for the aquaculture industry, the pork market and the like, and improve productivity.
  • a migration learning apparatus further includes:
  • the fourth input module 37 is configured to input the specified type of animal annotation data into the animal gender recognition model for model training.
  • the actual application capability of the annotation data obtained by the migration learning is further illustrated by the application details after the data annotation work.
  • a gender recognition model for effective classification of sows and boars is obtained.
  • the sows and boars have different meat quality.
  • the gender identification model the sex information of the pigs is obtained, which is more conducive to the convenience of the breeding owners according to the convenience of the sows and boars, so as to distinguish between the conservation feed and the breeding area in the process of large-scale breeding pig breeding. In order to effectively control the farming process according to the needs of the pork market.
  • the second obtaining module 38 is configured to input the image of the pig face to be tested into the animal gender recognition model to obtain output data of the animal gender classification in the animal gender model.
  • the pig face labeling data obtained by the migration learning provides a more detailed difference to the pig user, so as to more specifically refine the work in the breeding process, including feed matching, breeding time control, etc., so that the breeding industry is more scientific. More data can be relied upon.
  • the computer device may be a server, and its internal structure may be as shown in FIG. 9.
  • the computer device includes a processor, memory, network interface, and database connected by a system bus. Among them, the computer designed processor is used to provide calculation and control capabilities.
  • the memory of the computer device includes a non-volatile storage medium, an internal memory.
  • the non-volatile storage medium stores an operating system, computer readable instructions, and a database.
  • the memory provides an environment for the operation of operating systems and computer readable instructions in a non-volatile storage medium.
  • the database of the computer device is used to store data such as migration learning.
  • the network interface of the computer device is used to communicate with an external terminal via a network connection.
  • the computer readable instructions when executed, perform the flow of an embodiment of the methods described above. It will be understood by those skilled in the art that the structure shown in FIG. 9 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation of the computer device to which the present application is applied.
  • An embodiment of the present application also provides a computer non-volatile readable storage medium having stored thereon computer readable instructions that, when executed, perform the processes of the embodiments of the methods described above.
  • the above description is only the preferred embodiment of the present application, and is not intended to limit the scope of the patent application, and the equivalent structure or equivalent process transformations made by the specification and the drawings of the present application, or directly or indirectly applied to other related The technical field is equally included in the scope of patent protection of the present application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Biology (AREA)
  • Evolutionary Computation (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Image Analysis (AREA)

Abstract

一种迁移学习方法，将指定任务的标注数据输入任务训练模型进行训练以获取第一参数；锁定第一参数，将未标注数据和标注数据混合输入域分类网络，以获取第二参数；锁定第二参数，域分类器联合任务训练模型共同训练特征提取层，以获取特征提取层不区分数据的第三参数；基于第三参数，对未标注数据进行识别。

Description

迁移学习方法、装置、计算机设备和存储介质
本申请要求于2018年4月17日提交中国专利局、申请号为2018103452545，发明名称为“迁移学习方法、装置、计算机设备和存储介质”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
本申请涉及到机器学习领域，特别是涉及到迁移学习方法、装置、计算机设备和存储介质。
 背景技术
目前的深度学习框架下，需要大量的标注数据来训练模型，对数据进行标注的工作耗时长、成本高，耗费大量人力、物力和财力，即便如此，通过有限的标注数据训练得到的模型的泛化能力提升空间有限，只能实现同类型数据迁移，数据类型跨度大、区别大时，在新的数据集上的反馈效果将会大打折扣，这极大的限制了深度学习方法在业界的应用范围和推广速度。
 技术问题
本申请的主要目的为提供一种迁移学习方法，旨在解决有限标注数据导致训练模型的泛化能力较差的技术问题。
 技术解决方案
本申请提供了一种应用于迁移学习网络的迁移学习方法，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，方法包括：
将指定任务的标注数据输入所述任务训练模型进行模型训练，以获取到特征提取层的第一参数；
锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
本申请还提供了一种迁移学习装置，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，装置包括：
第一输入模块，用于将指定任务的标注数据输入所述任务训练模型，以获取到特征提取层的第一参数；
第二输入模块，用于锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
训练模块，用于锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
迁移模块，用于基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
本申请还提供了一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，所述处理器执行所述计算机可读指令时实现上述方法的步骤。
本申请还提供了一种计算机非易失性可读存储介质，其上存储有计算机可读指令，所述计算机可读指令被处理器执行时实现上述的方法的步骤。
 有益效果
本申请通过开发从标注数据迁移到未标注数据的迁移学习网络，实现对未标注数据的有效识别，扩展数据标注，提高现有模型的泛化能力；通过迁移学习网络将已标注数据的特征提取方式迁移到未标注数据上，进行自动标注识别，提高标注数据的映射适用范围，使数据标注更方便、快捷、准确，迁移范围更广泛，使得更有利于模型的推广和普及应用；创建基于域分类训练网络和任务训练网络的迁移学习网络，对数据类型不作区分，扩大适用范围，实现从标注数据到未标注数据的有效迁移，为数据标注节省了大量的人力、物力和财力。
 附图说明
图1 本申请一实施例的迁移学习方法的流程示意图；
图2 本申请一实施例的迁移学习装置的结构示意图；
图3 本申请一实施例的训练模块的结构示意图；
图4 本申请一实施例的训练单元的结构示意图；
图5 本申请另一实施例的迁移学习装置的结构示意图；
图6 本申请再一实施例的迁移学习装置的结构示意图；
图7 本申请又一实施例的迁移学习装置的结构示意图；
图8 本申请又一实施例的迁移学习装置的结构示意图；
图9 本申请一实施例的计算机内部结构示意图。
 本发明的最佳实施方式
参照图1，本申请一实施例的应用于迁移学习网络的迁移学习方法，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，所述方法包括：
S1：将指定任务的标注数据输入任务训练模型进行模型训练，以获取到特征提取层的第一参数。
本步骤中的标注数据的种类和内容量依据指定任务的不同而不同，其中标注数据的内容量为几千张、几万张或十几万张的图片。本步骤的任务训练网络包括特征提取层Mf和任务训练模型Mt两部分。本实施例的Mf主要接收图片输入，输出编码之后的图片，Mf使用VGG-16（Visual Geometry Group，牛津大学的视觉几何研究组开发的卷积神经网络结构，16表示其深度）已经训练好的模型与权重参数来预测图片的类别,无需再训练神经网络权重参数，以提高特征提取的准确性并节约训练成本，进一步提高本实施例中迁移学习网络的迁移能力评估。VGG-16应用了两个卷积层，每个卷积层都是64个过滤器，2个卷积->Pool（池化）->2个卷积->Pool->3个卷积->Pool->3个卷积->Pool->3个卷积->Pool->全连接->全连接->Softmax（Softmax是一个函数，主要应用就是多分类，Softmax函数是Sigmoid函数的扩展，Sigmoid函数只能分两类，而Softmax能分多类）。VGG-16的卷积层和池化层的过滤器设置如下：卷积核：3*3；步长：1；使用最大池化核：2*2；步长：2。还有一个需要注意的是卷积层的过滤器即卷积核的个数，是从64->128->256->512->512。随着每层输出矩阵长宽成倍的缩小，卷积核的个数成倍的增长。
本实施例的任务训练网络满足应用的第一步是训练任务训练模型，在训练过程中首先输入标注数据（此处为图像）进入任务训练模型，随后任务训练模型的输出或预测结果会与预期结果（即图像的标签）进行比较。任务训练模型每次完成对数据的迭代后，任务训练模型的权重会得到调整，以减少预测的误差率，即通过调整权重，使任务训练模型的预测结果与实际图像标签的匹配率可得到改善。当任务训练模型收敛时，任务训练模型的预测结果与实际图像标签的匹配率最佳。训练完毕后，任务训练模型可以用于推断，即对其读取的图像数据进行预测，输出：预测图片 (包括类名, 语义概念, 预测概率等标注信息)。
本实施例中Mt代表任务训练模型，本实施例中采取两层1024个节点的全连接层加一层Softmax。本步骤通过标注数据正常反传训练Mf和Mt，Mt训练收敛时，获得Mf的最佳权重。通过将标注数据输入到Mf的VGG-16卷积层和池化层进行监督训练，训练收敛时，Mt对VGG-16的卷积层和池化层输出的特征向量具有良好的识别效果，此时的VGG-16的卷积层权重和池化层权重即Mf的第一参数，此时表示Mf输出的特征向量在特征识别任务上有较好的表现。
S2：锁定所述第一参数，将未标注数据和上述标注数据混合输入域分类网络，以获取到域分类器区分未标注数据和上述标注数据的第二参数。
本实施例中未标注数据为B数据（无详细标注信息），标注数据为A数据（具有详细标注信息）。本实施例的域分类网络包括特征提取层Mf和二分类域分类器Md两部分，Md是一个域分类器，使用两层全连接层加一层Softmax层实现，本实施例的迁移学习能力在于增加了域分类器Md，本实施例的域分类网络和任务训练网络共用同一特征提取层Mf。本步骤在训练二分类域分类网络时，锁定Mf的第一参数，锁定即保持Mf的参数不变，通过锁定Mf的第一参数（即Mf识别标注数据的权重参数），使得Mf保持对A数据的特征向量识别效果最佳，然后将A数据和B数据随机混合后共同训练Md，当二分类域分类网络训练收敛时，表示根据当前Mf输出的特征向量，二分类域分类器Md可将A数据和B数据进行最佳的有效区分。本步骤在训练过程中将A数据和B数据随机混合后随即输入二分类域分类网络，A数据和B数据同时训练Md，训练收敛时记录对应二分类域分类器的权重参数，即第二参数，所以Md在第二参数下对A数据和B数据区分效果最佳。
S3：锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和上述标注数据的第三参数。
本实施例的第三参数为Mf不区分标注数据和未标注数据时的权重参数。通过锁定步骤S2中的Md的第二参数，打开Mf，Md联合Mt共同反传训练Mf，以便Mf输出的特征向量同时满足任务需求，且对A数据和B数据的不进行区分识别，以便实现在A数据上训练的模型可以在B数据上达到同样的效果，即在A数据上训练的模型同样可识别并提取B数据的特征，并输出B数据的特征向量。使得Mf在标注数据A上训练的效果可以迁移到数据分布不同的B数据上，实际业务中B数据将会是新的业务数据，没有标注信息，但在业务B数据上达到一个接近A数据的业务指标效果，并且没有标注成本。
S4：基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
本实施例可通过Mf对A数据和B数据的不进行区分识别的作用，直接通过训练模型迁移到其他应用领域的无标注数据，提高训练模型的泛化能力，以便扩大训练模型的适用范围，进一步推广训练模型的推广和普及应用。本申请其他实施例也可通过Mf对数据A和数据B的不进行区分识别的作用，将A数据的标注信息特征迁移到B数据上，使B数据成为与A数据具有相同或近似的数据分布规律的新业务标注信息，通过在未标注数据上实现自动标注识别，提高数据标注的便利性和实用性，使数据标注更方便、快捷、准确，迁移范围更广泛，使得更有利于模型的推广和普及应用，同时也为数据标注节省了大量的人力、物力和财力。
进一步地，步骤S3，包括：
S30：通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
本实施例通过Md的交叉熵LOSS函数联合Mt的交叉熵LOSS函数共同反传训练Mf，以便在训练Mf的同时利用已经训练好的域分类器Md的参数，达到最终Mf的参数（即第三参数）可以对不同领域范围的标注数据和未标注数据不做任何区分的目的，即将对标注数据的特征识别效果同样也用于未标注数据上。
进一步地，上述步骤S30，包括：
S301：修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定。
本实施例中域分类器的原LOSS函数为：LOSS=-(logp)- [log(1-p)]，其中P为二分类数据中的某一分类的标定，比如P为A数据的标定，原LOSS函数表示原特征提取层对二分类数据有较好的区分。本实施例为了让特征提取层Mf对二分类数据不作区分，则Md的交叉熵LOSS函数为-2* log0.5时梯度消散效果最佳，则A数据和B数据的区分效果最不好，达到特征提取层Mf对二分类数据不作区分的效果，可实现特征提取层Mf从A数据到B数据的迁移学习。为了满足上述的对A数据和B数据不作区分的最佳条件，将域分类器的原LOSS函数替换为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定，则在模型的Softmax输出为[0.5,0.5]时，达到修改后的LOSS函数的最小值，此时特征提取层Mf对二分类数据不作任何区分。
S302：锁定所述第二参数，将修改后的所述域分类器的交叉熵LOSS函数和所述任务训练模型的交叉熵LOSS函数相加得到LOSS函数之和。
本实施例通过将修改后的域分类器的交叉熵LOSS函数和任务训练模型的交叉熵LOSS函数相加得到LOSS函数之和的方式，使Md的交叉熵LOSS函数跟Mt的交叉熵LOSS函数共同训练Mf的参数，使得Md的交叉熵LOSS函数跟Mt的交叉熵LOSS函数对Mf有相同的变化趋势的影响。当Mf的Softmax输出为[05, 0.5]时达到训练收敛,LOSS函数达到最小值，以便在训练Mf的同时利用已经训练好的域分类器Md的参数，达到最终Mf对不同领域范围的标注数据和未标注数据不做区分的目的。
S303：通过所述LOSS函数之和梯度反向传播训练所述特征提取层。
本实施例通过域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和，然后通过梯度反向传播训练所述特征提取层Mf，当域分类器的LOSS函数和任务训练模型的LOSS函数均趋近于最小时，LOSS函数之和趋近于最小，达到优化目的。本步骤的训练目的是在训练任务模型的同时，利用已经训练好的域分类器的参数，使得最终Mf对不同领域的标注数据和未标注数据不做区分，使Mf在不同领域的标注数据和未标注数据间进行有效迁移。
S304：训练至所述LOSS函数之和达到预设值时，确定第三参数。
本步骤中的预设值为趋近于零的某一确定值，比如10 -6，10 -12等，以便LOSS函数之和达到趋近于零的趋势，此时域分类器Md的输出结果为[0.5,0.5]，表示特征提取层Mf对不同领域的标注A数据和未标注B数据的区分效果最弱，或对A数据和B数据不作区分，训练到Mf对不同领域的标注A数据和未标注B数据的区分效果最弱时，对应的Mf的权重参数为第三参数。
本申请另一实施例中，步骤S3之后，包括：
S31：优化训练过程中的学习率和梯度比例。
本实施例中，为了达到最佳的迁移学习效果，将上述步骤S1、S2和S3多次迭代进行，不断调整学习率和梯度比例，来修正LOSS函数权重以实现迁移学习网络中Mf、Mt和Md各部分的收敛。本实施例利用梯度下降的算法，通过迭代找到交叉熵函数的一个可行解或最优解。但是在利用梯度下降算法求解的过程中，由于维度之间的差异使得收敛的很慢，梯度下降算法每次迭代，都会受到学习率的影响。学习率的选择非常重要，如果学习率过大，会导致求解的过程不稳定，导致每次迭代可能不会减小代价函数的结果，甚至会超过局部最小值导致无法收敛。如果学习率太小，导致求解的速度太慢，达到收敛所需要迭代的次数就会非常高，不利于算法的求解。本实施例采用变换的学习率，当梯度大的时候，学习率也大，梯度小的时候，学习率也小，学习率采用与梯度正相关的关系，存在一个比例系数，采用固定的比例系数，得到的结果就是w(t+1) = w(t) - 学习率*梯度变化率(w)，可缓解求解问题，优化训练过程。本实施例从以下几个数值开始试验学习率的值，0.001 ,0.003, 0.01, 0.03, 0.1, 0.3, 1, …。学习率初始值为0.001, 不符合预期乘以3倍用0.003代替，不符合预期再用0.01替代，如此循环直至找到最合适的学习率。即在为梯度下降算法选择合适的学习率时，可以大致按3的倍数再按10的倍数来选取一系列学习率值，直到找到一个最小值，同时找到另一最大值。其中最大的学习率值，或者比最大值略小一些的学习率值就是最佳的学习率值。实验证明本实施例的学习率设置为0.1或者0.1附近时效果较佳。
S32：依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得上述迁移学习网络的特征提取层的优化参数以及域分类器的优化参数。
本实施例中为了进一步优化从A数据到B数据的迁移学习效果，通过上述梯度下降计算法不断调整学习率和梯度比例，以便通过多次迭代训练，逐步优化迁移学习网络的Mf、Mt和Md各部分的参数，提高迁移学习效果。比如，经过一次步骤S1到步骤S3的训练，迁移学习网络从A数据到B数据的迁移学习效果，表示为对B数据的标注识别率为90%，经过多次迭代进行步骤S1到步骤S3的训练后，迁移学习网络从A数据到B数据的迁移学习效果，表示为对B数据的标注识别率为98%，或接近100%，大大提高了迁移学习网络的泛化能力。
进一步地，本申请再一实施例中，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；步骤S32之后，包括：
S33：锁定迁移学习网络中特征提取层的优化参数、域分类器的优化参数以及人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习。
本实施例为迁移学习网络的应用验证，本实施例通过从一领域的标注数据迁移到对另一领域的未标注数据完成标注识别为例。比如A数据为用于人脸识别的标注人脸图像数据，B数据为未标注的猪脸、猫脸或象脸等其他动物的脸部数据。比如，本实施例以从标注的人脸图像数据迁移至未标注的猪脸图像数据为例，通过标注人脸图像数据确定迁移学习网络的Mf、Mt和Md各部分参数。将已确定的迁移学习网络的Mf、Mt和Md各部分参数，用于从标注的人脸图像数据迁移至未标注的猪脸图像数据进行特征标注。
S34：输出指定种类动物标注数据。
本实施例通过将用于人脸识别的标注人脸图像A数据以及未标注的猪脸图像B数据，输入到上述优化好各参数的迁移学习网络中，迁移学习网络最后输出的猪脸数据的正确标注率高达99%，表明本申请的迁移学习网络实现了从标注数据到未标注数据的良好的迁移学习能力，为数据标注节省了大量的人力、物力和财力。本申请其他实施例也可通过迁移学习网络的直接跨领域迁移使用，使得迁移学习在已标注数据的领域和未标注数据的新业务领域达到相同或近似的识别效果，省却对未标注数据的标注工作。
进一步地，本实施例步骤S34之后，包括：
S35：将所述指定种类动物标注数据输入动物类别识别模型进行模型训练。
    本步骤中的动物类别识别模型通过将上述通过迁移学习获得的标注信息后的猪脸图像B数据输入卷积神经网络训练得到能精准识别动物类别的识别模型，并可根据不同的识别目标进行选择和预先设置卷积神经网络的组成结构。举例地，本实施例的动物类别识别模型为我国地方猪种的分类，通过上述动物类别识别模型，有效识别现有养殖猪的分类信息，为养殖业提供有力的数据分析。我国地方猪种可分为华北、华南、华中、江海、西南、高原六大类型，每一个类型又有许多独特的猪种类型，原有通过人工标注数据的工作量极大，且容易出现误差，通过上述的迁移学习网络进行猪脸数据标注，大大提高了业务数据处理能力。
S36：将待分类猪脸图像输入动物类别识别模型，以获取动物类别识别模型中动物类别分类的输出数据。
本实施例通过将待分类猪脸数据输入上述动物类别识别模型，直接识别猪种种类，为养殖业、猪肉市场等领域提供有力的数据支持，提高生产率。
进一步地，步骤S36之后，还包括：
S37：将所述指定种类动物标注数据输入动物性别识别模型进行模型训练。
本实施例通过对数据标注工作后的应用详述，进一步说明通过迁移学习获得的标注数据的实际应用能力。本实施例通过将新获得标注的猪脸图像B数据用于母猪和公猪的性别识别模型的卷积神经网络结构进行训练，获得了对母猪和公猪的有效分类的性别识别模型。母猪和公猪的肉质不同，根据性别识别模型获取猪的性别信息，更有利于养殖业主根据对母猪和公猪的方便，实现在大批量种猪养殖过程中，区分养护饲料、养殖区域等，以便根据猪肉市场的需求，有效把控养殖过程。
S38：将待测猪脸图像输入动物性别识别模型，以获取动物性别模型中动物性别分类的输出数据。
本实施例通过迁移学习获得的猪脸标注数据对养猪用户提供更细致的区别，以便更针对性地细化养殖过程中的工作，包括饲料匹配、养殖时间管控等，使养殖业更科学化，更有数据可依。
本申请通过开发从标注数据迁移到未标注数据的迁移学习网络，实现对未标注数据的有效识别，扩展数据标注，提高现有模型的泛化能力；通过迁移学习网络将已标注数据的特征提取方式迁移到未标注数据上，进行自动标注识别，提高数据标注的适用范围，使数据标注更方便、快捷、准确，迁移范围更广泛，使得更有利于模型的推广和普及应用；创建基于域分类训练网络和任务训练网络的迁移学习网络，对数据类型不作区分，扩大适用范围，实现从标注数据到未标注数据的有效迁移，为数据标注节省了大量的人力、物力和财力。
参照图2，本申请还提供了一种迁移学习装置，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，所述装置包括：
第一输入模块1，用于将指定任务的标注数据输入任务训练模型进行模型训练，以获取到特征提取层的第一参数。
本实施例中的标注数据的种类和内容量依据指定任务的不同而不同，其中标注数据的内容量为几千张、几万张或十几万张的图片。本实施例的任务训练网络包括特征提取层Mf和任务训练模型Mt两部分。本实施例的Mf主要接收图片输入，输出编码之后的图片，Mf使用VGG-16（Visual Geometry Group，牛津大学的视觉几何研究组开发的卷积神经网络结构，16表示其深度）已经训练好的模型与权重参数来预测图片的类别,无需再训练神经网络权重参数，以提高特征提取的准确性并节约训练成本，进一步提高本实施例中迁移学习网络的迁移能力评估。VGG-16应用了两个卷积层，每个卷积层都是64个过滤器，2个卷积->Pool（池化）->2个卷积->Pool->3个卷积->Pool->3个卷积->Pool->3个卷积->Pool->全连接->全连接->Softmax（Softmax是一个函数，主要应用就是多分类，Softmax函数是Sigmoid函数的扩展，Sigmoid函数只能分两类，而Softmax能分多类）。VGG-16的卷积层和池化层的过滤器设置如下：卷积核：3*3；步长：1；使用最大池化核：2*2；步长：2。还有一个需要注意的是卷积层的过滤器即卷积核的个数，是从64->128->256->512->512。随着每层输出矩阵长宽成倍的缩小，卷积核的个数成倍的增长。
本实施例的任务训练网络满足应用的第一步是训练任务训练模型，在训练过程中首先输入标注数据（此处为图像）进入任务训练模型，随后任务训练模型的输出或预测结果会与预期结果（即图像的标签）进行比较。任务训练模型每次完成对数据的迭代后，任务训练模型的权重会得到调整，以减少预测的误差率，即通过调整权重，使任务训练模型的预测结果与实际图像标签的匹配率可得到改善。当任务训练模型收敛时，任务训练模型的预测结果与实际图像标签的匹配率最佳。训练完毕后，任务训练模型可以用于推断，即对其读取的图像数据进行预测，输出：预测图片 (包括类名, 语义概念, 预测概率等标注信息)。
本实施例中Mt代表任务训练模型，本实施例中采取两层1024个节点的全连接层加一层Softmax。本实施例通过标注数据正常反传训练Mf和Mt，Mt训练收敛时，获得Mf的最佳权重。通过将标注数据输入到Mf的VGG-16卷积层和池化层进行监督训练，训练收敛时，Mt对VGG-16的卷积层和池化层输出的特征向量具有良好的识别效果，此时的VGG-16的卷积层权重和池化层权重即Mf的第一参数，此时表示Mf输出的特征向量在特征识别任务上有较好的表现。
第二输入模块2，用于锁定所述第一参数，将未标注数据和上述标注数据混合输入域分类网络，以获取到域分类器区分未标注数据和上述标注数据的第二参数。
本实施例中未标注数据为B数据（无详细标注信息），标注数据为A数据（具有详细标注信息）。本实施例的域分类网络包括特征提取层Mf和二分类域分类器Md两部分，Md是一个域分类器，使用两层全连接层加一层Softmax层实现，本实施例的迁移学习能力在于增加了域分类器Md，本实施例的域分类网络和任务训练网络共用同一特征提取层Mf。本实施例在训练二分类域分类网络时，锁定Mf的第一参数，锁定即保持Mf的参数不变，通过锁定Mf的第一参数（即Mf识别标注数据的权重参数），使得Mf保持对A数据的特征向量识别效果最佳，然后将A数据和B数据随机混合后共同训练Md，当二分类域分类网络训练收敛时，表示根据当前Mf输出的特征向量，二分类域分类器Md可将A数据和B数据进行最佳的有效区分。本实施例在训练过程中将A数据和B数据随机混合后随即输入二分类域分类网络，A数据和B数据同时训练Md，训练收敛时记录对应二分类域分类器的权重参数，即第二参数，所以Md在第二参数下对A数据和B数据区分效果最佳。
训练模块3，用于锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和上述标注数据的第三参数。
本实施例的第三参数为Mf不区分标注数据和未标注数据时的权重参数。通过锁定实施例第二输入模块2中的Md的第二参数，打开Mf，Md联合Mt共同训练Mf，以便Mf输出的特征向量同时满足任务需求，且对A数据和B数据的不进行区分识别，以便实现在A数据上训练的模型可以在B数据上达到同样的效果，即在A数据上训练的模型同样可识别并提取B数据的特征，并输出B数据的特征向量。使得Mf在标注数据A上训练的效果可以迁移到数据分布不同的B数据上，实际业务中B数据将会是新的业务数据，没有标注信息，但在业务B数据上达到一个接近A数据的业务指标效果，并且没有标注成本。
迁移模块4，用于基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
本实施例可通过Mf对A数据和B数据的不进行区分识别的作用，直接通过训练模型迁移到其他应用领域的无标注数据，提高训练模型的泛化能力，以便扩大训练模型的适用范围，进一步推广训练模型的推广和普及应用。本申请其他实施例也可通过Mf对数据A和数据B的不进行区分识别的作用，将A数据的标注信息特征迁移到B数据上，使B数据成为与A数据具有相同或近似的数据分布规律的新业务标注信息，通过在未标注数据上实现自动标注识别，提高数据标注的便利性和实用性，使数据标注更方便、快捷、准确，迁移范围更广泛，使得更有利于模型的推广和普及应用，同时也为数据标注节省了大量的人力、物力和财力。
参照图3，本实施例的训练模块3，包括：
训练单元30：用于通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
本实施例通过Md的交叉熵LOSS函数联合Mt的交叉熵LOSS函数共同反传训练Mf，以便在训练Mf的同时利用已经训练好的域分类器Md的参数，达到最终Mf的参数（即第三参数）可以对不同领域范围的标注数据和未标注数据不做任何区分的目的，即将对标注数据的特征识别效果同样也用于未标注数据上。
参照图4，本实施例的训练单元30，包括：
修改子单元301，用于修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定。
本实施例中域分类器的原LOSS函数为：LOSS=-(logp)- [log(1-p)]，其中P为二分类数据中的某一分类的标定，比如P为A数据的标定，原LOSS函数表示原特征提取层对二分类数据有较好的区分。本实施例为了让特征提取层Mf对二分类数据不作区分，则Md的交叉熵LOSS函数为-2* log0.5时梯度消散效果最佳，则A数据和B数据的区分效果最不好，达到特征提取层Mf对二分类数据不作区分的效果，可实现特征提取层Mf从A数据到B数据的迁移学习。为了满足上述的对A数据和B数据不作区分的最佳条件，将域分类器的原LOSS函数替换为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定，则在模型的Softmax输出为[0.5,0.5]时，达到修改后的LOSS函数的最小值，此时特征提取层Mf对二分类数据不作任何区分。
相加子单元302，用于锁定所述第二参数，将修改后的所述域分类器的交叉熵LOSS函数和所述任务训练模型的交叉熵LOSS函数相加得到LOSS函数之和。
本实施例通过将修改后的域分类器的交叉熵LOSS函数和任务训练模型的交叉熵LOSS函数相加得到LOSS函数之和的方式，使Md的交叉熵LOSS函数跟Mt的交叉熵LOSS函数共同训练Mf的参数，使得Md的交叉熵LOSS函数跟Mt的交叉熵LOSS函数对Mf有相同的变化趋势的影响。当Mf的Softmax输出为[05, 0.5]时达到训练收敛,LOSS函数达到最小值，以便在训练Mf的同时利用已经训练好的域分类器Md的参数，达到最终Mf对不同领域范围的标注数据和未标注数据不做区分的目的。
训练子单元303，用于通过所述LOSS函数之和梯度反向传播训练所述特征提取层。
本实施例通过域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和，然后通过梯度反向传播训练所述特征提取层Mf，当域分类器的LOSS函数和任务训练模型的LOSS函数均趋近于最小时，LOSS函数之和趋近于最小，达到优化目的。本实施例的训练目的是在训练任务模型的同时，利用已经训练好的域分类器的参数，使得最终Mf对不同领域的标注数据和未标注数据不做区分，使Mf在不同领域的标注数据和未标注数据间进行有效迁移。
确定子单元304，用于训练至所述LOSS函数之和达到预设值时，确定第三参数。
本实施例中的预设值为趋近于零的某一确定值，比如10 -6，10 -12等，以便LOSS函数之和达到趋近于零的趋势，此时域分类器的Md输出结果为[0.5,0.5]，表示特征提取层Mf对不同领域的标注A数据和未标注B数据的区分效果最弱，或对A数据和B数据不作区分，训练到Mf对不同领域的标注A数据和未标注B数据的区分效果最弱时，对应的Mf的权重参数为第三参数。
参照图5，本申请另一实施例的迁移学习装置，包括：
优化模块31，用于优化训练过程中的学习率和梯度比例。
本实施例中，为了达到最佳的迁移学习效果，将上述步骤S1、S2和S3多次迭代进行，不断调整学习率和梯度比例，来修正LOSS函数权重以实现迁移学习网络中Mf、Mt和Md各部分的收敛。本实施例利用梯度下降的算法，通过迭代找到交叉熵函数的一个可行解或最优解。但是在利用梯度下降算法求解的过程中，由于维度之间的差异使得收敛的很慢，梯度下降算法每次迭代，都会受到学习率的影响。学习率的选择非常重要，如果学习率过大，会导致求解的过程不稳定，导致每次迭代可能不会减小代价函数的结果，甚至会超过局部最小值导致无法收敛。如果学习率太小，导致求解的速度太慢，达到收敛所需要迭代的次数就会非常高，不利于算法的求解。本实施例采用变换的学习率，当梯度大的时候，学习率也大，梯度小的时候，学习率也小，学习率采用与梯度正相关的关系，存在一个比例系数，采用固定的比例系数，得到的结果就是w(t+1) = w(t) - 学习率*梯度变化率(w)，可缓解求解问题，优化训练过程。本实施例从以下几个数值开始试验学习率的值，0.001 ,0.003, 0.01, 0.03, 0.1, 0.3, 1, …。学习率初始值为0.001, 不符合预期乘以3倍用0.003代替，不符合预期再用0.01替代，如此循环直至找到最合适的学习率。即在为梯度下降算法选择合适的学习率时，可以大致按3的倍数再按10的倍数来选取一系列学习率值，直到找到一个最小值，同时找到另一最大值。其中最大的学习率值，或者比最大值略小一些的学习率值就是最佳的学习率值。实验证明本实施例的学习率设置为0.1或者0.1附近时效果较佳。
迭代模块32，用于依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得上述迁移学习网络的特征提取层的优化参数以及域分类器优化参数。
本实施例中为了进一步优化从A数据到B数据的迁移学习效果，通过上述梯度下降计算法不断调整学习率和梯度比例，以便通过多次迭代训练，逐步优化迁移学习网络的Mf、Mt和Md各部分的参数，提高迁移学习效果。比如，经过一次步骤S1到步骤S3的训练，迁移学习网络从A数据到B数据的迁移学习效果，表示为对B数据的标注识别率为90%，经过多次迭代进行步骤S1到步骤S3的训练后，迁移学习网络从A数据到B数据的迁移学习效果，表示为对B数据的标注识别率为98%，或接近100%，大大提高了迁移学习网络的泛化能力。
参照图6，本申请再一实施例的迁移学习装置，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；包括：
迁移学习模块33，用于锁定迁移学习网络中特征提取层的优化参数、域分类器的优化参数以及人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习。
本实施例为迁移学习网络的应用验证，本实施例通过从一领域的标注数据迁移到对另一领域的未标注数据完成标注识别为例。比如A数据为用于人脸识别的标注人脸图像数据，B数据为未标注的猪脸、猫脸或象脸等其他动物的脸部数据。比如，本实施例以从标注的人脸图像数据迁移至未标注的猪脸图像数据为例，通过标注人脸图像数据确定迁移学习网络的Mf、Mt和Md各部分参数。将已确定的迁移学习网络的Mf、Mt和Md各部分参数，用于从标注的人脸图像数据迁移至未标注的猪脸图像数据进行特征标注。
输出模块34，用于输出指定种类动物的标注数据。
本实施例通过将用于人脸识别的标注人脸图像A数据以及未标注的猪脸图像B数据，输入到上述优化好各参数的迁移学习网络中，迁移学习网络最后输出的猪脸数据的正确标注率高达99%，表明本申请的迁移学习网络实现了从标注数据到未标注数据的良好的迁移学习能力，为数据标注节省了大量的人力、物力和财力。本申请其他实施例也可通过迁移学习网络的直接跨领域迁移使用，使得迁移学习在已标注数据的领域和未标注数据的新业务领域达到相同或近似的识别效果，省却对未标注数据的标注工作。
参照图7，本申请又一实施例的迁移学习装置，还包括：
第三输入模块35，用于将所述指定种类动物标注数据输入动物类别识别模型进行模型训练。
本实施例中的动物类别识别模型通过将上述通过迁移学习获得的标注信息后的猪脸图像B数据输入卷积神经网络训练得到能精准识别动物类别的识别模型，并可根据不同的识别目标进行选择和预先设置卷积神经网络的组成结构。举例地，本实施例的动物类别识别模型为我国地方猪种的分类，通过上述动物类别识别模型，有效识别现有养殖猪的分类信息，为养殖业提供有力的数据分析。我国地方猪种可分为华北、华南、华中、江海、西南、高原六大类型，每一个类型又有许多独特的猪种类型，原有通过人工标注数据的工作量极大，且容易出现误差，通过上述的迁移学习网络进行猪脸数据标注，大大提高了业务数据处理能力。
第一获取模块36，用于将待分类猪脸图像输入动物类别识别模型，以获取动物类别识别模型中动物类别分类的输出数据。
本实施例通过将待分类猪脸数据输入上述动物类别识别模型，直接识别猪种种类，为养殖业、猪肉市场等领域提供有力的数据支持，提高生产率。
参照图8，本申请又一实施例的迁移学习装置，还包括：
第四输入模块37，用于将所述指定种类动物标注数据输入动物性别识别模型进行模型训练。
本实施例通过对数据标注工作后的应用详述，进一步说明通过迁移学习获得的标注数据的实际应用能力。本实施例通过将新获得标注的猪脸图像B数据用于母猪和公猪的性别识别模型的卷积神经网络结构进行训练，获得了对母猪和公猪的有效分类的性别识别模型。母猪和公猪的肉质不同，根据性别识别模型获取猪的性别信息，更有利于养殖业主根据对母猪和公猪的方便，实现在大批量种猪养殖过程中，区分养护饲料、养殖区域等，以便根据猪肉市场的需求，有效把控养殖过程。
第二获取模块38，用于将待测猪脸图像输入动物性别识别模型，以获取动物性别模型中动物性别分类的输出数据。
本实施例通过迁移学习获得的猪脸标注数据对养猪用户提供更细致的区别，以便更针对性地细化养殖过程中的工作，包括饲料匹配、养殖时间管控等，使养殖业更科学化，更有数据可依。
参照图9，本申请实施例中还提供一种计算机设备，该计算机设备可以是服务器，其内部结构可以如图9所示。该计算机设备包括通过系统总线连接的处理器、存储器、网络接口和数据库。其中，该计算机设计的处理器用于提供计算和控制能力。该计算机设备的存储器包括非易失性存储介质、内存储器。该非易失性存储介质存储有操作系统、计算机可读指令和数据库。该内存器为非易失性存储介质中的操作系统和计算机可读指令的运行提供环境。该计算机设备的数据库用于存储迁移学习等数据。该计算机设备的网络接口用于与外部的终端通过网络连接通信。该计算机可读指令在执行时，执行如上述各方法的实施例的流程。本领域技术人员可以理解，图9中示出的结构，仅仅是与本申请方案相关的部分结构的框图，并不构成对本申请方案所应用于其上的计算机设备的限定。
本申请一实施例还提供一种计算机非易失性可读存储介质，其上存储有计算机可读指令，该计算机可读指令在执行时，执行如上述各方法的实施例的流程。以上所述仅为本申请的优选实施例，并非因此限制本申请的专利范围，凡是利用本申请说明书及附图内容所作的等效结构或等效流程变换，或直接或间接运用在其他相关的技术领域，均同理包括在本申请的专利保护范围内。

Claims (20)

  1. 一种应用于迁移学习网络的迁移学习方法，其特征在于，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，所述方法包括：
    将指定任务的标注数据输入所述任务训练模型进行模型训练，以获取到特征提取层的第一参数；
    锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
    锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
    基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
  2. 根据权利要求1所述的迁移学习方法，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
  3. 根据权利要求2所述的迁移学习方法，其特征在于，所述通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定;
    锁定所述第二参数，将修改后的所述域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和；
    通过所述LOSS函数之和梯度反向传播训练所述特征提取层;
    训练至所述LOSS函数之和达到预设值时，确定所述第三参数。
  4. 根据权利要求1所述的迁移学习方法，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤之后，包括：
    优化训练过程中的学习率和梯度比例；
    依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数。
  5. 根据权利要求4所述的迁移学习方法，其特征在于，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；所述依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数的步骤之后，包括：
    锁定所述迁移学习网络中所述特征提取层的优化参数、所述域分类器的优化参数以及所述人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习；
    输出所述指定种类动物的标注数据。
  6. 一种迁移学习装置，其特征在于，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，装置包括：
    第一输入模块，用于将指定任务的标注数据输入所述任务训练模型，以获取到特征提取层的第一参数；
    第二输入模块，用于锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
    训练模块，用于锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
    迁移模块，用于基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
  7. 根据权利要求6所述的迁移学习装置，其特征在于，所述训练模块，包括：
    训练单元，用于通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
  8. 根据权利要求7所述的迁移学习装置，其特征在于，所述训练单元，包括：
    修改子单元，用于修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定;
    相加子单元，用于锁定所述第二参数，将修改后的所述域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和；
    训练子单元，用于通过所述LOSS函数之和梯度反向传播训练所述特征提取层;
    确定子单元，用于训练至所述LOSS函数之和达到预设值时，确定所述第三参数。
  9. 根据权利要求6所述的迁移学习装置，其特征在于，包括：
    优化模块，用于优化训练过程中的学习率和梯度比例；
    迭代模块，用于依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数。
  10. 根据权利要求9所述的迁移学习装置，其特征在于，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；迁移学习装置包括：
    迁移学习模块，用于锁定所述迁移学习网络中所述特征提取层的优化参数、所述域分类器的优化参数以及所述人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习；
    输出模块，用于输出所述指定种类动物的标注数据。
  11. 一种计算机设备，包括存储器和处理器，所述存储器存储有计算机可读指令，其特征在于，所述处理器执行所述计算机可读指令时实现迁移学习方法，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，方法包括：
    将指定任务的标注数据输入所述任务训练模型进行模型训练，以获取到特征提取层的第一参数；
    锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
    锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
    基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
  12. 根据权利要求11所述的计算机设备，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
  13. 根据权利要求12所述的计算机设备，其特征在于，所述通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定;
    锁定所述第二参数，将修改后的所述域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和；
    通过所述LOSS函数之和梯度反向传播训练所述特征提取层;
    训练至所述LOSS函数之和达到预设值时，确定所述第三参数。
  14. 根据权利要求11所述的计算机设备，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤之后，包括：
    优化训练过程中的学习率和梯度比例；
    依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数。
  15. 根据权利要求14所述的计算机设备，其特征在于，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；所述依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数的步骤之后，包括：
    锁定所述迁移学习网络中所述特征提取层的优化参数、所述域分类器的优化参数以及所述人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习；
    输出所述指定种类动物的标注数据。
  16. 一种计算机非易失性可读存储介质，其上存储有计算机可读指令，其特征在于，所述计算机可读指令被处理器执行时实现迁移学习方法，迁移学习网络包括共用同一特征提取层的任务训练网络和域分类网络，任务训练网络还包括任务训练模型，域分类网络还包括域分类器，方法包括：
    将指定任务的标注数据输入所述任务训练模型进行模型训练，以获取到特征提取层的第一参数；
    锁定所述第一参数，将未标注数据和所述标注数据混合输入所述域分类网络，以获取到所述域分类器区分所述未标注数据和所述标注数据的第二参数；
    锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述标注数据的第三参数；
    基于所述第三参数，将所述特征提取层根据所述标注数据训练得到的训练结果迁移至所述未标注数据，对所述未标注数据进行识别。
  17. 根据权利要求16所述的计算机非易失性可读存储介质，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数。
  18. 根据权利要求17所述的计算机非易失性可读存储介质，其特征在于，所述通过所述域分类器的交叉熵LOSS函数联合所述任务训练模型的交叉熵LOSS函数共同反传训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤，包括：
    修改所述域分类器的交叉熵LOSS函数，修改后的LOSS函数为：LOSS=-（log2P） 2-[log(2-2P)] 2，其中P为二分类数据中的某一分类的标定;
    锁定所述第二参数，将修改后的所述域分类器的LOSS函数和所述任务训练模型的LOSS函数相加得到LOSS函数之和；
    通过所述LOSS函数之和梯度反向传播训练所述特征提取层;
    训练至所述LOSS函数之和达到预设值时，确定所述第三参数。
  19. 根据权利要求16所述的计算机非易失性可读存储介质，其特征在于，所述锁定所述第二参数，所述域分类器联合所述任务训练模型共同训练所述特征提取层，以获取所述特征提取层不区分所述未标注数据和所述指定标注数据的第三参数的步骤之后，包括：
    优化训练过程中的学习率和梯度比例；
    依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数。
  20. 根据权利要求19所述的计算机非易失性可读存储介质，其特征在于，标注人脸图像数据为所述标注数据，指定种类动物的脸部数据为所述未标注数据，所述指定任务为人脸识别任务；所述依次迭代所述第一参数、第二参数以及第三参数的训练过程指定次数，以获得所述迁移学习网络的所述特征提取层的优化参数以及所述域分类器的优化参数的步骤之后，包括：
    锁定所述迁移学习网络中所述特征提取层的优化参数、所述域分类器的优化参数以及所述人脸识别任务的识别模型参数，将所述标注人脸数据和未标注指定种类动物的数据混合输入所述迁移学习网络进行迁移学习；
    输出所述指定种类动物的标注数据。
PCT/CN2018/095497 2018-04-17 2018-07-12 迁移学习方法、装置、计算机设备和存储介质 Ceased WO2019200748A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201810345254.5 2018-04-17
CN201810345254.5A CN108805160B (zh) 2018-04-17 2018-04-17 迁移学习方法、装置、计算机设备和存储介质

Publications (1)

Publication Number Publication Date
WO2019200748A1 true WO2019200748A1 (zh) 2019-10-24

Family

ID=64094805

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/095497 Ceased WO2019200748A1 (zh) 2018-04-17 2018-07-12 迁移学习方法、装置、计算机设备和存储介质

Country Status (2)

Country Link
CN (1) CN108805160B (zh)
WO (1) WO2019200748A1 (zh)

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110782043A (zh) * 2019-10-29 2020-02-11 腾讯科技（深圳）有限公司 模型优化方法、装置、存储介质及服务器
CN112016501A (zh) * 2020-09-04 2020-12-01 平安科技（深圳）有限公司 人脸识别模型的训练方法、装置和计算机设备
CN112016633A (zh) * 2020-09-25 2020-12-01 北京百度网讯科技有限公司 一种模型训练方法、装置、电子设备及存储介质
CN112183332A (zh) * 2020-09-28 2021-01-05 成都希盟泰克科技发展有限公司 一种基于迁移学习的牦牛脸识别方法
CN112215296A (zh) * 2020-10-21 2021-01-12 红相股份有限公司 基于迁移学习的红外图像识别方法及存储介质
CN112861926A (zh) * 2021-01-18 2021-05-28 平安科技（深圳）有限公司 耦合的多任务特征提取方法、装置、电子设备及存储介质
CN112926654A (zh) * 2021-02-25 2021-06-08 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
WO2021164292A1 (zh) * 2020-02-21 2021-08-26 平安科技（深圳）有限公司 基于大数据的阅读模型优化方法、装置、设备及介质
CN113627074A (zh) * 2021-07-13 2021-11-09 西安理工大学 一种基于迁移学习的地波传播时延预测方法
CN113761184A (zh) * 2020-09-29 2021-12-07 北京沃东天骏信息技术有限公司 文本数据的分类方法、设备及存储介质
CN113807150A (zh) * 2020-06-16 2021-12-17 阿里巴巴集团控股有限公司 数据处理、姿态预测方法、设备及存储介质
CN115115050A (zh) * 2022-06-29 2022-09-27 腾讯科技（深圳）有限公司 迁移学习模型的确定方法、装置、设备及存储介质
CN116310429A (zh) * 2022-12-02 2023-06-23 国网湖北省电力有限公司超高压公司 一种基于螺栓图像对比学习的输电线路螺栓异常分类方法
CN112116063B (zh) * 2020-08-11 2024-04-05 西安交通大学 一种基于元学习的特征偏移矫正方法

Families Citing this family (15)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110162981B (zh) * 2019-04-18 2020-10-02 阿里巴巴集团控股有限公司 数据处理方法及装置
US10846413B2 (en) 2019-04-18 2020-11-24 Advanced New Technologies Co., Ltd. Data processing method and device
CN110210355B (zh) * 2019-05-24 2021-11-26 华南农业大学 稻田杂草种类识别方法及系统、目标位置检测方法及系统
CN110377587B (zh) * 2019-07-15 2023-02-10 腾讯科技（深圳）有限公司 基于机器学习的迁移数据确定方法、装置、设备及介质
CN111079406B (zh) * 2019-12-13 2022-01-11 华中科技大学 自然语言处理模型训练方法、任务执行方法、设备及系统
CN111291802B (zh) * 2020-01-21 2023-12-12 华为技术有限公司 数据标注方法及装置
CN113496227A (zh) * 2020-04-08 2021-10-12 顺丰科技有限公司 一种字符识别模型的训练方法、装置、服务器及存储介质
CN111209901A (zh) * 2020-04-20 2020-05-29 湖南极点智能科技有限公司 一种面部识别方法、系统及相关装置
CN111753877B (zh) * 2020-05-19 2024-03-05 海克斯康制造智能技术(青岛)有限公司 一种基于深度神经网络迁移学习的产品质量检测方法
CN111582277A (zh) * 2020-06-15 2020-08-25 深圳天海宸光科技有限公司 一种基于迁移学习的车牌识别系统及方法
CN111931865B (zh) * 2020-09-17 2021-01-26 平安科技（深圳）有限公司 图像分类模型的训练方法、装置、计算机设备及存储介质
CN112163635B (zh) * 2020-10-14 2024-05-28 平安科技（深圳）有限公司 基于深度学习的图像分类方法、装置、服务器及介质
CN113128565B (zh) * 2021-03-25 2022-05-06 之江实验室 面向预训练标注数据不可知的图像自动标注系统和装置
CN115100706B (zh) * 2022-06-15 2025-07-01 阿里巴巴（中国）有限公司 图像处理方法、装置、存储介质和计算机设备
CN115171151A (zh) * 2022-06-16 2022-10-11 天津农学院 一种基于猪脸识别的猪只饮水行为检测方法

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102289686A (zh) * 2011-08-09 2011-12-21 北京航空航天大学 一种基于迁移学习的运动目标分类识别方法
CN106991439A (zh) * 2017-03-28 2017-07-28 南京天数信息科技有限公司 基于深度学习与迁移学习的图像识别方法
CN107316061A (zh) * 2017-06-22 2017-11-03 华南理工大学 一种深度迁移学习的不平衡分类集成方法
CN107527068A (zh) * 2017-08-07 2017-12-29 南京信息工程大学 基于cnn和域自适应学习的车型识别方法
CN107657279A (zh) * 2017-09-26 2018-02-02 中国科学院大学 一种基于少量样本的遥感目标检测方法

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20070233867A1 (en) * 2006-03-28 2007-10-04 Narasimhan Ganapathiraman Method and apparatus for preserving MAC addresses across a reboot
CN101794396B (zh) * 2010-03-25 2012-12-26 西安电子科技大学 基于迁移网络学习的遥感图像目标识别系统及方法
CN104680178B (zh) * 2015-03-06 2017-11-21 西安电子科技大学 基于迁移学习多吸引子细胞自动机的图像分类方法
CN106599863A (zh) * 2016-12-21 2017-04-26 中国科学院光电技术研究所 一种基于迁移学习技术的深度人脸识别方法
CN107133616B (zh) * 2017-04-02 2020-08-28 南京汇川图像视觉技术有限公司 一种基于深度学习的无分割字符定位与识别方法
CN107451616A (zh) * 2017-08-01 2017-12-08 西安电子科技大学 基于深度半监督迁移学习的多光谱遥感图像地物分类方法
CN107392321A (zh) * 2017-08-04 2017-11-24 北京奇艺世纪科技有限公司 一种应用迁移学习可行性度量方法及装置
CN107798349B (zh) * 2017-11-03 2020-07-14 合肥工业大学 一种基于深度稀疏自编码机的迁移学习方法
CN108399431A (zh) * 2018-02-28 2018-08-14 国信优易数据有限公司 分类模型训练方法以及分类方法

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102289686A (zh) * 2011-08-09 2011-12-21 北京航空航天大学 一种基于迁移学习的运动目标分类识别方法
CN106991439A (zh) * 2017-03-28 2017-07-28 南京天数信息科技有限公司 基于深度学习与迁移学习的图像识别方法
CN107316061A (zh) * 2017-06-22 2017-11-03 华南理工大学 一种深度迁移学习的不平衡分类集成方法
CN107527068A (zh) * 2017-08-07 2017-12-29 南京信息工程大学 基于cnn和域自适应学习的车型识别方法
CN107657279A (zh) * 2017-09-26 2018-02-02 中国科学院大学 一种基于少量样本的遥感目标检测方法

Cited By (20)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110782043B (zh) * 2019-10-29 2023-09-22 腾讯科技（深圳）有限公司 模型优化方法、装置、存储介质及服务器
CN110782043A (zh) * 2019-10-29 2020-02-11 腾讯科技（深圳）有限公司 模型优化方法、装置、存储介质及服务器
WO2021164292A1 (zh) * 2020-02-21 2021-08-26 平安科技（深圳）有限公司 基于大数据的阅读模型优化方法、装置、设备及介质
CN113807150A (zh) * 2020-06-16 2021-12-17 阿里巴巴集团控股有限公司 数据处理、姿态预测方法、设备及存储介质
CN112116063B (zh) * 2020-08-11 2024-04-05 西安交通大学 一种基于元学习的特征偏移矫正方法
CN112016501B (zh) * 2020-09-04 2023-08-29 平安科技（深圳）有限公司 人脸识别模型的训练方法、装置和计算机设备
CN112016501A (zh) * 2020-09-04 2020-12-01 平安科技（深圳）有限公司 人脸识别模型的训练方法、装置和计算机设备
CN112016633A (zh) * 2020-09-25 2020-12-01 北京百度网讯科技有限公司 一种模型训练方法、装置、电子设备及存储介质
CN112183332A (zh) * 2020-09-28 2021-01-05 成都希盟泰克科技发展有限公司 一种基于迁移学习的牦牛脸识别方法
CN113761184A (zh) * 2020-09-29 2021-12-07 北京沃东天骏信息技术有限公司 文本数据的分类方法、设备及存储介质
CN112215296A (zh) * 2020-10-21 2021-01-12 红相股份有限公司 基于迁移学习的红外图像识别方法及存储介质
CN112215296B (zh) * 2020-10-21 2023-05-05 红相股份有限公司 基于迁移学习的红外图像识别方法及存储介质
CN112861926B (zh) * 2021-01-18 2023-10-31 平安科技（深圳）有限公司 耦合的多任务特征提取方法、装置、电子设备及存储介质
CN112861926A (zh) * 2021-01-18 2021-05-28 平安科技（深圳）有限公司 耦合的多任务特征提取方法、装置、电子设备及存储介质
CN112926654B (zh) * 2021-02-25 2023-08-01 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
CN112926654A (zh) * 2021-02-25 2021-06-08 平安银行股份有限公司 预标注模型训练、证件预标注方法、装置、设备及介质
CN113627074A (zh) * 2021-07-13 2021-11-09 西安理工大学 一种基于迁移学习的地波传播时延预测方法
CN113627074B (zh) * 2021-07-13 2024-04-19 西安理工大学 一种基于迁移学习的地波传播时延预测方法
CN115115050A (zh) * 2022-06-29 2022-09-27 腾讯科技（深圳）有限公司 迁移学习模型的确定方法、装置、设备及存储介质
CN116310429A (zh) * 2022-12-02 2023-06-23 国网湖北省电力有限公司超高压公司 一种基于螺栓图像对比学习的输电线路螺栓异常分类方法

Also Published As

Publication number Publication date
CN108805160B (zh) 2020-03-24
CN108805160A (zh) 2018-11-13

Similar Documents

Publication Publication Date Title
WO2019200748A1 (zh) 迁移学习方法、装置、计算机设备和存储介质
CN114144770B (zh) 用于生成用于模型重新训练的数据集的系统和方法
US20230016082A1 (en) Targeted data acquisition for model training
CN108647691B (zh) 一种基于点击特征预测的图像分类方法
CN110377587B (zh) 基于机器学习的迁移数据确定方法、装置、设备及介质
WO2019227617A1 (zh) 动物种类的分类方法、装置、计算机设备和存储介质
CN108345702A (zh) 实体推荐方法和装置
WO2020232874A1 (zh) 基于迁移学习的建模方法、装置、计算机设备和存储介质
CN115906962A (zh) 用于通用硬件-神经网络体系结构搜索的系统
CN108228757A (zh) 图像搜索方法和装置、电子设备、存储介质、程序
CN110647904A (zh) 一种基于无标记数据迁移的跨模态检索方法及系统
CN110598869A (zh) 基于序列模型的分类方法、装置、电子设备
WO2023217222A1 (zh) 细胞信息统计方法、装置、设备及计算机可读存储介质
CN113627433B (zh) 基于数据扰动的跨域自适应语义分割方法及装置
Thompson et al. finFindR: Computer-assisted recognition and identification of bottlenose dolphin photos in r
CN114529191B (zh) 用于风险识别的方法和装置
CN104915436A (zh) 自适应多标签预测方法
CN116229196A (zh) 一种噪声样本的识别方法及装置
CN115130597A (zh) 模型训练方法、装置和存储介质
CN110162609A (zh) 用于向用户推荐咨询问题的方法及装置
CN115712751B (zh) 基于文本描述的跨领域人物搜索方法
CN111291803A (zh) 一种图像分级粒度迁移方法、系统、设备和介质
CN116304880A (zh) 一种基于特征学习的多标签分类模型训练方法及分类方法
CN111401551A (zh) 基于强化学习的弱监督自学习方法
CN111506832B (zh) 一种基于块矩阵补全的异构对象补全方法

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: 18915349

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: 18915349

Country of ref document: EP

Kind code of ref document: A1