CN112633517A - Training method of machine learning model, computer equipment and storage medium - Google Patents
Training method of machine learning model, computer equipment and storage medium Download PDFInfo
- Publication number
- CN112633517A CN112633517A CN202011591857.7A CN202011591857A CN112633517A CN 112633517 A CN112633517 A CN 112633517A CN 202011591857 A CN202011591857 A CN 202011591857A CN 112633517 A CN112633517 A CN 112633517A
- Authority
- CN
- China
- Prior art keywords
- loss function
- sample data
- function value
- sample
- twin
- 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.)
- Granted
Links
- 238000012549 training Methods 0.000 title claims abstract description 82
- 238000010801 machine learning Methods 0.000 title claims abstract description 68
- 238000000034 method Methods 0.000 title claims abstract description 66
- 230000006870 function Effects 0.000 claims abstract description 286
- 238000004364 calculation method Methods 0.000 claims abstract description 90
- 238000000605 extraction Methods 0.000 claims description 34
- 238000004590 computer program Methods 0.000 claims description 3
- 238000011161 development Methods 0.000 description 9
- 238000004422 calculation algorithm Methods 0.000 description 7
- 238000012545 processing Methods 0.000 description 7
- 238000010586 diagram Methods 0.000 description 6
- 230000006978 adaptation Effects 0.000 description 4
- 238000013145 classification model Methods 0.000 description 3
- 238000001514 detection method Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 3
- 238000013473 artificial intelligence Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 238000010276 construction Methods 0.000 description 1
- 238000012937 correction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013500 data storage Methods 0.000 description 1
- 238000007636 ensemble learning method Methods 0.000 description 1
- 238000005286 illumination Methods 0.000 description 1
- 238000010295 mobile communication Methods 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 230000008707 rearrangement Effects 0.000 description 1
- 238000012827 research and development Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
- 239000007787 solid Substances 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
- 230000009466 transformation Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N20/00—Machine learning
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Software Systems (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Physics & Mathematics (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Artificial Intelligence (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a training method of a machine learning model, computer equipment and a storage medium, wherein the method comprises the following steps: acquiring a plurality of first sample data and second sample data from a plurality of sample domains; inputting each first sample data into a target model to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of judgment network parameters; inputting second sample data into a target model loaded with each judgment network parameter to obtain a plurality of second loss function values; and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met. The technical scheme of the embodiment of the invention can realize the automation of the training process of the machine learning model and effectively improve the robustness of the machine learning model in different scenes.
Description
Technical Field
The embodiment of the invention relates to the technical field of artificial intelligence, in particular to a training method of a machine learning model, computer equipment and a storage medium.
Background
With research and development of data analysis technology and artificial intelligence technology, people pay more and more attention to robustness of a machine learning model in different application scenes. The machine learning model with high robustness can maintain certain performance in a new scene, and is high in effectiveness.
In order to improve the robustness of a machine learning model, the existing method introduces algorithms such as regularization or dropout and the like in the model training process; or, fine-tuning the machine learning model, namely using the training samples of the new application scene to continue training on the basis of the pre-training model.
Although the robustness of the machine learning model can be improved to a certain extent by introducing algorithms such as regularization or dropout, the stability of the machine learning model is easy to reduce, and the final effect of the model often depends on the experience of a developer and a large number of trial and error; after the machine learning model is finely tuned, the final effect of the model is usually not as good as the re-training effect of the model, and in any of the above methods, a developer needs to perform adaptation work of the model to various scenes and maintain numerous model versions.
Disclosure of Invention
The embodiment of the invention provides a training method of a machine learning model, computer equipment and a storage medium, which can realize automation of a training process of the machine learning model and effectively improve the robustness of the machine learning model in different scenes.
In a first aspect, an embodiment of the present invention provides a training method for a machine learning model, including:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
In a second aspect, embodiments of the present invention also provide a computer device, including a processor and a memory, the memory storing instructions that, when executed, cause the processor to:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
In a third aspect, an embodiment of the present invention further provides a storage medium, where the storage medium is configured to store instructions for performing:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
The embodiment of the invention obtains a plurality of first sample data and second sample data from a plurality of sample domains, inputs each first sample data into a target model, respectively uses a feature extraction network and a discrimination network in the target model to perform forward calculation on each first sample data to obtain a plurality of first loss function values, and performs backward calculation to obtain a plurality of discrimination network parameters; inputting each second sample data into a target model loaded with the corresponding judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values; and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the technical means of finishing the training condition is met, so that the automation of the training process of the machine learning model can be realized, the robustness of the machine learning model in different scenes can be effectively improved, and the development cost of the machine learning model can be saved.
Drawings
Fig. 1 is a flowchart of a training method of a machine learning model according to an embodiment of the present invention;
FIG. 2a is a flowchart of a training method of a machine learning model according to a second embodiment of the present invention;
FIG. 2b is a diagram illustrating an implementation of a training method for a machine learning model according to a second embodiment of the present invention;
fig. 3 is a schematic structural diagram of a training apparatus for machine learning models according to a third embodiment of the present invention;
fig. 4 is a schematic structural diagram of a computer device according to a fourth embodiment of the present invention.
Detailed Description
The present invention will be described in further detail with reference to the accompanying drawings and examples. It is to be understood that the specific embodiments described herein are merely illustrative of the invention and are not limiting of the invention. It should be further noted that, for the convenience of description, only some of the structures related to the present invention are shown in the drawings, not all of the structures.
Before discussing exemplary embodiments in more detail, it should be noted that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although a flowchart may describe the operations (or steps) as a sequential process, many of the operations can be performed in parallel, concurrently or simultaneously. In addition, the order of the operations may be re-arranged. The process may be terminated when its operations are completed, but may have additional steps not included in the figure. The processes may correspond to methods, functions, procedures, subroutines, and the like.
The term "target model" as used herein refers to a machine learning model to be trained that is the result of the machine learning algorithm output, which can be considered as outputting a result given an input. For example, when the machine learning model is a specific classification model, the output result may be category information corresponding to the input data.
The term "feature extraction network" is used herein to extract features corresponding to input data, the "discrimination network" is used to discriminate the input data according to features of the input data, and the "feature extraction network" and the "discrimination network" together constitute the "target model".
For ease of understanding, the main inventive concepts of the embodiments of the present invention are briefly described.
Robustness of the machine learning model in different application scenes is particularly important for the model in practical engineering application. In the prior art, in order to improve the robustness of a machine learning model, algorithms such as regularization or dropout are generally introduced; or, fine-tuning the machine learning model, namely using the training samples of the new application scene to continue training on the basis of the pre-training model. However, the existing training method of the machine learning model generally requires a developer to perform adaptation work of the model to various scenes, and maintains numerous model versions, which results in higher model development cost.
The inventor considers whether the robustness of the machine learning model in different scenes can be improved while the automatic training of the machine learning model is realized by a method aiming at the problem that the adaptation work of the machine learning model needs to be finished manually in the prior art, thereby improving the effectiveness of the machine learning model and saving the development cost of the model.
Based on the above thought, the inventor creatively proposes that a plurality of first sample data and a plurality of second sample data are obtained from a plurality of sample domains, each first sample data is input into a target model, each first sample data is subjected to forward calculation by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and backward calculation is performed to obtain a plurality of discrimination network parameters; inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values; and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the technical means of finishing the training condition is met, so that the automation of the training process of the machine learning model can be realized, the robustness of the machine learning model in different scenes can be effectively improved, and the development cost of the machine learning model can be saved.
Example one
Fig. 1 is a flowchart of a training method of a machine learning model according to an embodiment of the present invention. The embodiment of the present invention is applicable to the case of training a machine learning model, and the method can be executed by the training apparatus for a machine learning model provided in the embodiment of the present invention, and the apparatus can be implemented in a software and/or hardware manner, and can be generally integrated in a computer device. As shown in fig. 1, the method of the embodiment of the present invention specifically includes:
In this embodiment, the sample domain is a set including a plurality of samples, each sample domain corresponds to a specific scene, each sample domain includes a plurality of sample data corresponding to the scene, and different sample domains correspond to different scenes. Specifically, the scene may be source information of the sample data in the sample domain, such as a type of a device that acquires the sample, a background condition for generating the sample (for example, a lighting condition when the sample of the picture is generated, an indoor scene, an outdoor scene, and the like).
In this step, optionally, sample data corresponding to each sample domain may be obtained in different sample domains, sample data corresponding to several sample domains is used as first sample data, and sample data corresponding to other several sample domains is used as second sample data. The first sample data and the second sample data are jointly used for iterative training of a target model.
And 120, inputting the first sample data into the target model, performing forward calculation on the first sample data by using the feature extraction network and the discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters.
The target model is a machine learning model to be trained, and specifically, the target model may be a classification model, a regression model, a data clustering model, or the like. The target model comprises a feature extraction network and a discrimination network.
In this embodiment, the training process of the machine learning model can be divided into two stages: an exploration phase and a deviation rectifying phase. The first sample data is sample data participating in an exploration phase, and the second sample data is sample data participating in a deviation rectifying phase.
The searching stage is that after each first sample data is input into the target model, each first sample data is subjected to forward calculation through the feature extraction network and the discrimination network in the target model to obtain a first loss function value corresponding to each first sample data, and then each first sample data is subjected to backward calculation through the discrimination network in the target model to obtain a discrimination network parameter corresponding to the first loss function value.
The first sample data can be input into the target model in parallel, so that the training time of the target model can be reduced, and the training efficiency of the target model can be improved.
And step 130, inputting each second sample data into the target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values.
In this embodiment, the deviation rectifying stage is to calculate the discriminant network parameters, load the discriminant network parameters in turn on the target model, input the second sample data into the target model loaded with different discriminant network parameters, extract the network according to the features in the target model, and perform forward calculation on each pair of second sample data of the discriminant network to obtain the second loss function values. It should be noted that, unlike the ensemble learning method, only one target model always exists in this embodiment.
And step 140, calculating to obtain a total loss function value according to each first loss function value and each second loss function value, performing backward calculation according to the total loss function value, and returning to execute the operation of obtaining a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
In this step, the first loss function values and the second loss function values may be tensor-added to obtain a total loss function value, or the first loss function values and the second loss function values may be tensor-added according to a preset weight ratio to obtain a total loss function value.
In this embodiment, after the total loss function value is calculated, the total loss function value may be calculated backward through the target model, parameters of the target model (for example, a weight and a bias of the target model) are adjusted according to the calculation result, then the first sample data and the second sample data are continuously obtained from the plurality of sample domains, and the total loss function value is calculated again until the target model meets the training termination condition. When the total loss function value is smaller than the preset threshold, it may be determined that the target model converges, that is, the target model satisfies the condition for ending the training.
In this embodiment, by acquiring first sample data and second sample data corresponding to different scenes, multi-scene sample data can be used for training a target model, so that the robustness of the target model in different scenes can be improved, a developer does not need to perform adaptation work of the target model for various scenes, the automation of the training process of the target model can be further realized, and the development cost of the target model can be saved; secondly, forward calculation is carried out on the first sample data through the feature extraction network and the discrimination network to obtain a first loss function value and discrimination network parameters, then under the condition that the discrimination network parameters are loaded on the target model, forward calculation is carried out on the second sample data to obtain second loss function values, a total loss function value is obtained through calculation according to the first loss function value and the second loss function values, the result of counterstudy of the feature extraction network and the discrimination network can be combined, parameter adjustment is effectively carried out on the target model, and therefore the training efficiency of the target model can be improved.
The embodiment of the invention obtains a plurality of first sample data and second sample data from a plurality of sample domains, inputs each first sample data into a target model, respectively uses a feature extraction network and a discrimination network in the target model to perform forward calculation on each first sample data to obtain a plurality of first loss function values, and performs backward calculation to obtain a plurality of discrimination network parameters; inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values; and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the technical means of finishing the training condition is met, so that the automation of the training process of the machine learning model can be realized, the robustness of the machine learning model in different scenes can be effectively improved, and the development cost of the machine learning model can be saved.
Example two
This embodiment is a further refinement of the first embodiment, and the same or corresponding terms as those in the first embodiment are explained, and this embodiment is not repeated. Fig. 2a is a flowchart of a training method for a machine learning model according to a second embodiment of the present invention, in this embodiment, the technical solution of this embodiment may be combined with one or more methods in the solutions of the foregoing embodiments, and in this embodiment, as shown in fig. 2a, the method according to the second embodiment of the present invention may further include:
Wherein each sample domain includes a plurality of sample data corresponding to the same scene.
In one implementation of the embodiments of the present invention, obtaining a plurality of sample fields includes: obtaining a plurality of sample data, and dividing each sample data into a plurality of sample sets according to a preset classification rule; and adjusting the proportion of different types of sample data in each sample set and the proportion of the sample data among the sample sets to obtain a plurality of sample domains.
In this embodiment, the classification rule may be constructed according to the data source attribute corresponding to the sample data and the tag attribute of the sample data. The data source attribute may be a type of device generating the sample data, a size of the sample data, a background condition for forming the sample data, a field to which the sample data belongs, and the like. The tag attribute may be tag information carried by the sample data, for example, when the sample data is face image data, the tag attribute may be information such as corresponding age, gender, and race.
In a specific embodiment, it is assumed that the sample data is picture data, and the background condition for forming the sample data may be an illumination condition, an indoor scene, an outdoor scene, and the like; assuming that the sample data is a photograph containing a helmet, the field to which the sample data belongs may be a chemical plant or a construction site.
In this embodiment, after a plurality of sample data are acquired, each sample data is first allocated to a corresponding sample set according to the established classification rule, and each sample set corresponds to a specific data source attribute or a label attribute. In order to ensure that the sample data in each sample set reaches the balance, the proportion between different types of sample data in each sample set and the proportion of the sample data between sample sets can be adjusted by using a preset sample balancing method according to the requirement of the hyper-parameter, so as to obtain a plurality of sample domains. The sample equalization method may include a sampling method and a data enhancement method.
In a specific embodiment, assuming that the target model is a classification model, and a sample set includes five sample pictures of a type a and ten sample pictures of a type B, the sample pictures of a type a may be geometrically transformed (for example, flipped, rotated, cropped, deformed, scaled, etc.) by using a data enhancement method, so that the number of the sample pictures of a type a is equal to that of the sample pictures of B type. Assuming that a total of six sample pictures are included in a sample set and a total of ten sample pictures are included in another sample set, geometric transformation can be performed on the pictures in the previous sample set by using a data enhancement method, so that the number of sample pictures in the two sample sets is equal.
Therefore, by establishing a classification rule, each sample data is distributed to a corresponding sample set, and samples in each sample set can be ensured to correspond to the same scene; by adjusting the proportion of different types of sample data in each sample set and the proportion of sample data among the sample sets, the sample data in each sample set can be balanced, so that the target model can be fairly trained by the multi-scene samples, and the training efficiency of the target model and the effectiveness of the training result of the target model can be improved.
In an implementation manner of the embodiment of the present invention, acquiring corresponding batch sample data from each sample domain includes: and acquiring batch sample data matched with the batch threshold value from each sample domain according to a preset batch threshold value through a data loader corresponding to each sample domain.
Each sample domain corresponds to one data loader, batch sample data is obtained through the data loaders, automation of data obtaining can be achieved, and accordingly training efficiency of the target model is improved.
And step 220, randomly selecting batch sample data corresponding to the multiple sample domains from all the batch sample data as first sample data, and taking the batch sample data corresponding to the remaining multiple sample domains as second sample data.
In this embodiment, in all the batch sample data, batch sample data corresponding to a plurality of sample domains may be randomly selected, and each batch sample data is used as first sample data, and then the batch sample data corresponding to each remaining sample domain is used as second sample data.
The number of the first sample data may be equal to the number of the second sample data, or when the number of the first sample data is not equal to the number of the second sample data, a difference between the number of the first sample data and the number of the second sample data needs to be smaller than a preset threshold. Therefore, the phenomenon that the difference between the quantity of the first sample data and the quantity of the second sample data is too large can be avoided, the quantity of the sample data participating in the exploration stage and the quantity of the sample data participating in the deviation correction stage are close, and therefore the target model can be trained fairly by the multi-scene sample data, the training efficiency of the target model is improved, and the effectiveness of the training result of the target model is improved.
And step 230, inputting the first sample data into the target model, performing forward calculation on the first sample data by using the feature extraction network and the discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters.
The first sample data can be input into the target model in parallel, so that the training time of the target model can be reduced, and the training efficiency of the target model can be improved.
In an implementation manner of the embodiment of the present invention, after obtaining the plurality of discriminant network parameters, the method further includes: and obtaining the characteristics which are output by the characteristic extraction network in the target model and correspond to the first sample data, and calculating a first twin loss function value through a twin network loss function according to the characteristics which correspond to the first sample data.
And 240, inputting each second sample data into the target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values.
In this embodiment, optionally, after obtaining a plurality of second loss function values, the method further includes: and acquiring the characteristics, corresponding to each second sample data, output by the characteristic extraction network in the target model, and calculating a second twin loss function value through a twin network loss function according to the characteristics corresponding to each second sample data.
And step 250, calculating to obtain a total loss function value according to each first loss function value and each second loss function value, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
In this embodiment, optionally, the calculating, according to each first loss function value and each second loss function value, to obtain a total loss function value, and performing backward calculation according to the total loss function value, includes: and calculating to obtain a total loss function value according to each first loss function value, each second loss function value, each first twin loss function value and each second twin loss function value, performing backward calculation according to the total loss function value, and updating the parameters of the target model according to the calculation result.
Therefore, the convergence speed of the target model can be increased and the training efficiency of the target model can be improved by calculating the first twin loss function value and the second twin loss function value and combining the first twin loss function value and the second twin loss function value to calculate the total loss function value.
In a specific embodiment, calculating a total loss function value based on each of the first loss function value, the second loss function value, the first twin loss function value, and the second twin loss function value includes: tensor addition is carried out on each first loss function value and each second loss function value to obtain a first function value; tensor addition is carried out on the first twin loss function value and the second twin loss function value to obtain a second function value; and calculating to obtain the total loss function value according to the first function value, the second function value and the hyper-parameter corresponding to the target model.
The second function value may be multiplied by a hyper-parameter corresponding to the target model to obtain a third function value, and finally, the third function value and the first function value are subjected to tensor addition to obtain a total loss function value.
The embodiment of the invention obtains a plurality of second loss function values by obtaining a plurality of sample fields, obtaining corresponding batch sample data from each sample field, randomly selecting the batch sample data corresponding to the plurality of sample fields as first sample data from all the batch sample data, taking the batch sample data corresponding to the remaining plurality of sample fields as second sample data, then inputting each first sample data into a target model, respectively using a feature extraction network and a discrimination network to carry out forward calculation on each first sample data to obtain a plurality of first loss function values, carrying out backward calculation to obtain a plurality of discrimination network parameters, then inputting each second sample data into the target model loading each discrimination network parameter, carrying out forward calculation on each second sample data to obtain a plurality of second loss function values, and finally obtaining each first loss function value and each second loss function value according to each first loss function value, and calculating to obtain a total loss function value, carrying out backward calculation according to the total loss function value, returning to execute the operation of obtaining a plurality of first sample data and second sample data from a plurality of sample domains until the technical means of finishing the training condition is met, realizing the automation of the training process of the machine learning model, effectively improving the robustness of the machine learning model in different scenes and saving the development cost of the machine learning model.
For better describing the technical solution provided by the embodiment of the present invention, fig. 2b is a schematic diagram of an implementation manner of a training method of a machine learning model in the embodiment, and the following implementation manners may be referred to in the embodiment of the present invention:
assuming that the target model is a living body detection model, firstly, dividing a plurality of sample data into living body samples and non-living body samples according to the label attributes; then, according to the data source attribute of each sample data, dividing the sample data into four sample sets, such as a sample set a, a sample set B, a sample set C and a sample set D in fig. 2B, and then adjusting the proportion of different types of sample data in each sample set and the proportion of sample data between each sample set by using a data enhancement method, so as to realize that living samples and non-living samples in each sample set keep 1: 1, the number of samples between each sample set keeps 1: 1, the adjusted sample sets are used as sample fields, such as sample field a, sample field B, sample field C, and sample field D in fig. 2B.
The living body detection model in this example is divided into two parts, namely a feature Extraction Network (ClueExtr) and a living body discrimination Network FasCls, and in one iteration training, the specific implementation steps are as follows:
step 1: the method comprises the steps of obtaining sample data participating in small-batch training in each sample domain from data loaders corresponding to different sample domains according to configuration of hyper-parameters, wherein the sample data are batch sample data D1, D2, D3 and D4, randomly disordering the batch sample data by using a random ordering algorithm, selecting the batch sample data ranked at the first two bits to participate in learning of an exploration phase, and marking the batch sample data as A1 and A2 (namely first sample data), and then selecting the batch sample data ranked at the last two bits to participate in learning of a deviation rectification phase, and marking the batch sample data as B1 and B2 (namely second sample data).
Step 2: in the exploration learning phase, the following operations are performed on the first sample data a1, a2 participating in the phase: forward calculation is carried out on A1 and A2 through a feature extraction network, and corresponding features Feat _ A1 and Feat _ A2 are output; then, according to the respective corresponding characteristics of A1 and A2, respectively carrying out forward calculation on A1 and A2 through a living body discrimination network, and outputting respective classification information false _ A1 and false _ A2; because the living body detection task is a binary model, the first Loss function values Cls _ Loss _ A1 and Cls _ Loss _ A2 corresponding to A1 and A2 can be calculated by a cross-entropy algorithm according to the classification information of A1 and A2; then, a1 and a2 are respectively calculated backwards through the discriminant network, and the discriminant network parameters Cls _ parameters _ a1 and Cls _ parameters _ a2 corresponding to a1 and a2 are obtained.
After the above processing, the Cls _ Loss _ a1 and the Cls _ Loss _ a2 are combined into Cls _ Loss _ a by tensor addition. And then, aiming at the characteristics Feat _ A1 and Feat _ A2 corresponding to A1 and A2 respectively, calculating a first twin Loss function value Contrast _ Loss _ A through a twin network Loss function.
And step 3: in the deskewing phase, the following operations are performed on the second sample data B1 and B2 participating in the phase: b1 and B2 are respectively calculated in a forward direction through a feature extraction network, and the corresponding features Feat _ B1 and Feat _ B2 are output. The living body discrimination network loads the above discrimination network parameters respectively, and performs forward calculation on B1 and B2, and outputs classification information corresponding to different discrimination network parameters, namely, false info _ B1_ a1, false info _ B1_ a2, false info _ B2_ a1, and false info _ B2_ a 2. Wherein, the false information _ B1_ a1 indicates the classification information corresponding to B1 when the living body judgment network loads the judgment network parameters corresponding to a 1.
Then, according to the classification information corresponding to each of B1 and B2, a cross entropy algorithm is used for calculating second Loss function values Cls _ Loss _ B1_ A1, Cls _ Loss _ B1_ A2, Cls _ Loss _ B2_ A1 and Cls _ Loss _ B2_ A2 corresponding to each of B1 and B2. Wherein, Cls _ Loss _ B1_ A1 represents a second Loss function value corresponding to B1 when the living body discrimination network loads the discrimination network parameters corresponding to A1.
After the processing, the Cls _ Loss _ B1_ A1 and the Cls _ Loss _ B1_ A2 are combined into Cls _ Loss _ B1 through tensor addition; the Cls _ Loss _ B2_ A1 and the Cls _ Loss _ B2_ A2 are combined into Cls _ Loss _ B2 through tensor addition; the Cls _ Loss _ B1 and the Cls _ Loss _ B2 are combined into Cls _ Loss _ B by tensor addition. And then aiming at the characteristics Feat _ B1 and Feat _ B2 corresponding to B1 and B2 respectively, calculating a second twin Loss function value Contrast _ Loss _ B through a twin network Loss function.
And 4, step 4: combining the Cls _ Loss _ A, Cls _ Loss _ B into the Cls _ Loss (namely a first function value) through tensor addition, combining the Contrast _ Loss _ A, Contrast _ Loss _ B into the Contrast _ Loss (namely a second function value) through tensor addition, and calculating the total Loss function value TotalLoss of the current iteration according to the corresponding super-parameter w _ Cls _ Contrast of the target model, wherein:
TotalLoss＝Cls_Loss+w_Cls_Contrast*Contrast_Loss
the overall loss function value is then back-calculated for this iteration.
And 5: and repeating the steps 1-4 until the target model converges. The condition for ending the training of the target model may be: and the precision of the target model training result is greater than a preset precision threshold value, or the total loss function value is smaller than a preset loss threshold value.
The preset accuracy threshold may be set to 99.9999%, the preset loss threshold may be set to 0.000001, and the specific value is preset in actual conditions, which is not limited in this embodiment.
The method provided by the embodiment of the invention can realize the automation of the training process of the machine learning model, effectively improve the robustness of the machine learning model in different scenes and save the development cost of the machine learning model.
EXAMPLE III
Fig. 3 is a schematic structural diagram of a training apparatus for machine learning models according to a third embodiment of the present invention. The apparatus may be implemented in software and/or hardware and may generally be integrated in a computer device. As shown in fig. 3, the apparatus includes: a sample data acquisition module 310, a first sample data input module 320, a second sample data input module 330, and a total loss function value calculation module 340.
The sample data obtaining module 310 is configured to obtain a plurality of first sample data and second sample data from a plurality of sample domains, where each sample domain includes a plurality of sample data corresponding to the same scene; a first sample data input module 320, configured to input each of the first sample data into a target model, perform forward calculation on each of the first sample data by using a feature extraction network and a discrimination network in the target model, respectively, to obtain a plurality of first loss function values, and perform backward calculation, to obtain a plurality of discrimination network parameters; a second sample data input module 330, configured to input each piece of second sample data into a target model in which each discrimination network parameter is loaded, and perform forward calculation on each piece of second sample data to obtain a plurality of second loss function values; and the total loss function value calculating module 340 is configured to calculate a total loss function value according to each first loss function value and each second loss function value, perform backward calculation according to the total loss function value, and return to perform an operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until a training termination condition is met.
The embodiment of the invention obtains a plurality of first sample data and second sample data from a plurality of sample domains, inputs each first sample data into a target model, respectively uses a feature extraction network and a discrimination network in the target model to perform forward calculation on each first sample data to obtain a plurality of first loss function values, and performs backward calculation to obtain a plurality of discrimination network parameters; inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values; and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the technical means of finishing the training condition is met, so that the automation of the training process of the machine learning model can be realized, the robustness of the machine learning model in different scenes can be effectively improved, and the development cost of the machine learning model can be saved.
On the basis of the foregoing embodiments, the sample data obtaining module 310 may include:
the system comprises a batch sample data acquisition unit, a data processing unit and a data processing unit, wherein the batch sample data acquisition unit is used for acquiring a plurality of sample domains and acquiring corresponding batch sample data from each sample domain;
the sample data selecting unit is used for randomly selecting batch sample data corresponding to the plurality of sample domains from all the batch sample data as first sample data, and using the batch sample data corresponding to the remaining plurality of sample domains as second sample data;
the system comprises a sample data dividing unit, a data processing unit and a data processing unit, wherein the sample data dividing unit is used for acquiring a plurality of sample data and dividing each sample data into a plurality of sample sets according to a preset classification rule;
the proportion adjustment unit is used for adjusting the proportion between different types of sample data in each sample set and the proportion of the sample data between the sample sets to obtain a plurality of sample domains;
and the data acquisition unit is used for acquiring batch sample data matched with the batch threshold value from each sample domain according to a preset batch threshold value through the data loader corresponding to each sample domain.
On the basis of the above embodiments, the first sample data input module 320 may include:
and the first twin loss function value calculating unit is used for acquiring the features, corresponding to the first sample data, output by the feature extraction network in the target model, and calculating a first twin loss function value through a twin network loss function according to the features, corresponding to the first sample data.
The second sample data input module 330 may include:
and the second twin loss function value calculating unit is used for acquiring the characteristics, corresponding to each second sample data, output by the characteristic extraction network in the target model, and calculating a second twin loss function value through a twin network loss function according to the characteristics corresponding to each second sample data.
The total loss function value calculation module 340 may include:
the parameter updating unit is used for calculating to obtain a total loss function value according to each first loss function value, each second loss function value, each first twin loss function value and each second twin loss function value, carrying out backward calculation according to the total loss function value, and updating the parameters of the target model according to the calculation results;
the first function value calculating unit is used for carrying out tensor addition on each first loss function value and each second loss function value to obtain a first function value;
the second function value calculating unit is used for carrying out tensor addition on the first twin loss function value and the second twin loss function value to obtain a second function value;
and the loss function value calculation unit is used for calculating the total loss function value according to the first function value, the second function value and the hyper-parameter corresponding to the target model.
The training device of the machine learning model can execute the training method of the machine learning model provided by any embodiment of the invention, and has the corresponding functional modules and beneficial effects of executing the training method of the machine learning model.
Example four
Fig. 4 is a schematic structural diagram of a computer apparatus according to a fourth embodiment of the present invention, as shown in fig. 4, the computer apparatus includes a processor 410, a memory 420, an input device 430, and an output device 440; the number of the processors 410 in the computer device may be one or more, and one processor 410 is taken as an example in fig. 4; the processor 410, the memory 420, the input device 430 and the output device 440 in the computer apparatus may be connected by a bus or other means, and the connection by the bus is exemplified in fig. 4.
The memory 420 serves as a computer-readable storage medium for storing software programs, computer-executable programs, and modules, such as program instructions/modules corresponding to a training method of a machine learning model in the embodiments of the present invention (for example, the sample data acquisition module 310, the first sample data input module 320, the second sample data input module 330, and the total loss function value calculation module 340 in a training apparatus of a machine learning model). The processor 410 executes various functional applications and data processing of the computer device by executing software programs, instructions and modules stored in the memory 420, so as to implement the training method of the machine learning model described above. That is, the program when executed by the processor implements:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
The memory 420 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system, an application program required for at least one function; the storage data area may store data created according to the use of the terminal, and the like. Further, the memory 420 may include high speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other non-volatile solid state storage device. In some examples, memory 420 may further include memory located remotely from processor 410, which may be connected to a computer device through a network. Examples of such networks include, but are not limited to, the internet, intranets, local area networks, mobile communication networks, and combinations thereof.
The input device 430 may be used to receive input numeric or character information and generate key signal inputs related to user settings and function control of the computer apparatus, and may include a keyboard and a mouse, etc. The output device 440 may include a display device such as a display screen.
On the basis of the above embodiments, the processor 410 is configured to obtain a plurality of first sample data and second sample data from a plurality of sample domains by:
acquiring a plurality of sample domains, and acquiring corresponding batch sample data from each sample domain;
in all the batch sample data, the batch sample data corresponding to the multiple sample fields is randomly selected as first sample data, and the batch sample data corresponding to the remaining multiple sample fields is used as second sample data.
On the basis of the above embodiments, the processor 410 is configured to obtain a plurality of sample fields by:
obtaining a plurality of sample data, and dividing each sample data into a plurality of sample sets according to a preset classification rule;
and adjusting the proportion of different types of sample data in each sample set and the proportion of the sample data among the sample sets to obtain a plurality of sample domains.
On the basis of the foregoing embodiments, the processor 410 is configured to obtain corresponding batch sample data from each sample domain by:
and acquiring batch sample data matched with the batch threshold value from each sample domain according to a preset batch threshold value through a data loader corresponding to each sample domain.
The processor 410, after obtaining the plurality of discriminant network parameters, further performs the following operations:
acquiring features which are output by a feature extraction network and correspond to the first sample data in the target model, and calculating a first twin loss function value through a twin network loss function according to the features corresponding to the first sample data;
the processor 410, after obtaining the plurality of second loss function values, further performs the following operations:
acquiring features, corresponding to each second sample data, output by a feature extraction network in the target model, and calculating a second twin loss function value through a twin network loss function according to the features corresponding to each second sample data;
the processor 410 is configured to calculate a total loss function value from each first loss function value and each second loss function value, and perform backward calculation from the total loss function value by:
and calculating to obtain a total loss function value according to each first loss function value, each second loss function value, each first twin loss function value and each second twin loss function value, performing backward calculation according to the total loss function value, and updating the parameters of the target model according to the calculation result.
The processor 410 is arranged to calculate a total loss function value from each of the first loss function value, the second loss function value, the first twin loss function value and the second twin loss function value by:
tensor addition is carried out on each first loss function value and each second loss function value to obtain a first function value;
tensor addition is carried out on the first twin loss function value and the second twin loss function value to obtain a second function value;
and calculating to obtain the total loss function value according to the first function value, the second function value and the hyper-parameter corresponding to the target model.
EXAMPLE five
Fifth, an embodiment of the present invention further provides a computer-readable storage medium, on which a computer program is stored, where the computer program, when executed by a processor, implements the method according to any embodiment of the present invention. Of course, the embodiment of the present invention provides a computer-readable storage medium, which can perform related operations in a training method of a machine learning model according to any embodiment of the present invention. That is, the program when executed by the processor implements:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
From the above description of the embodiments, it is obvious for those skilled in the art that the present invention can be implemented by software and necessary general hardware, and certainly, can also be implemented by hardware, but the former is a better embodiment in many cases. Based on such understanding, the technical solutions of the present invention may be embodied in the form of a software product, which can be stored in a computer-readable storage medium, such as a floppy disk, a Read-Only Memory (ROM), a Random Access Memory (RAM), a FLASH Memory (FLASH), a hard disk or an optical disk of a computer, and includes several instructions for enabling a computer device (which may be a personal computer, a server, or a network device) to execute the methods according to the embodiments of the present invention.
It should be noted that, in the embodiment of the training apparatus for machine learning models, the units and modules included in the training apparatus are only divided according to the functional logic, but are not limited to the above division as long as the corresponding functions can be implemented; in addition, specific names of the functional units are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present invention.
It is to be noted that the foregoing is only illustrative of the preferred embodiments of the present invention and the technical principles employed. It will be understood by those skilled in the art that the present invention is not limited to the particular embodiments described herein, but is capable of various obvious changes, rearrangements and substitutions as will now become apparent to those skilled in the art without departing from the scope of the invention. Therefore, although the present invention has been described in greater detail by the above embodiments, the present invention is not limited to the above embodiments, and may include other equivalent embodiments without departing from the spirit of the present invention, and the scope of the present invention is determined by the scope of the appended claims.
Claims (13)
1. A method for training a machine learning model, comprising:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
2. The method of claim 1, wherein obtaining a plurality of first sample data and second sample data from a plurality of sample domains comprises:
acquiring a plurality of sample domains, and acquiring corresponding batch sample data from each sample domain;
in all the batch sample data, the batch sample data corresponding to the multiple sample fields is randomly selected as first sample data, and the batch sample data corresponding to the remaining multiple sample fields is used as second sample data.
3. The method of claim 2, wherein obtaining a plurality of sample fields comprises:
obtaining a plurality of sample data, and dividing each sample data into a plurality of sample sets according to a preset classification rule;
and adjusting the proportion of different types of sample data in each sample set and the proportion of the sample data among the sample sets to obtain a plurality of sample domains.
4. The method of claim 2, wherein obtaining a corresponding batch of sample data from each sample domain comprises:
and acquiring batch sample data matched with the batch threshold value from each sample domain according to a preset batch threshold value through a data loader corresponding to each sample domain.
5. The method of claim 1, after obtaining the plurality of discriminative network parameters, further comprising:
acquiring features which are output by a feature extraction network and correspond to the first sample data in the target model, and calculating a first twin loss function value through a twin network loss function according to the features corresponding to the first sample data;
after obtaining a plurality of second loss function values, the method further comprises:
acquiring features, corresponding to each second sample data, output by a feature extraction network in the target model, and calculating a second twin loss function value through a twin network loss function according to the features corresponding to each second sample data;
calculating to obtain a total loss function value according to each first loss function value and each second loss function value, and performing backward calculation according to the total loss function value, wherein the method comprises the following steps:
and calculating to obtain a total loss function value according to each first loss function value, each second loss function value, each first twin loss function value and each second twin loss function value, performing backward calculation according to the total loss function value, and updating the parameters of the target model according to the calculation result.
6. The method of claim 5, wherein calculating a total loss function value based on each of the first loss function value, the second loss function value, the first twin loss function value, and the second twin loss function value comprises:
tensor addition is carried out on each first loss function value and each second loss function value to obtain a first function value;
tensor addition is carried out on the first twin loss function value and the second twin loss function value to obtain a second function value;
and calculating to obtain the total loss function value according to the first function value, the second function value and the hyper-parameter corresponding to the target model.
7. A computer device comprising a processor and a memory, the memory to store instructions that, when executed, cause the processor to:
acquiring a plurality of first sample data and second sample data from a plurality of sample domains, wherein each sample domain comprises a plurality of sample data corresponding to the same scene;
inputting the first sample data into a target model, performing forward calculation on the first sample data by using a feature extraction network and a discrimination network in the target model respectively to obtain a plurality of first loss function values, and performing backward calculation to obtain a plurality of discrimination network parameters;
inputting each second sample data into a target model loaded with each judgment network parameter, and performing forward calculation on each second sample data to obtain a plurality of second loss function values;
and calculating to obtain a total loss function value according to the first loss function values and the second loss function values, performing backward calculation according to the total loss function value, and returning to execute the operation of acquiring a plurality of first sample data and second sample data from a plurality of sample domains until the training termination condition is met.
8. The computer device of claim 7, wherein the processor is configured to obtain a plurality of first sample data and second sample data from a plurality of sample fields by:
acquiring a plurality of sample domains, and acquiring corresponding batch sample data from each sample domain;
in all the batch sample data, the batch sample data corresponding to the multiple sample fields is randomly selected as first sample data, and the batch sample data corresponding to the remaining multiple sample fields is used as second sample data.
9. The computer device of claim 8, wherein the processor is configured to obtain the plurality of sample fields by:
obtaining a plurality of sample data, and dividing each sample data into a plurality of sample sets according to a preset classification rule;
and adjusting the proportion of different types of sample data in each sample set and the proportion of the sample data among the sample sets to obtain a plurality of sample domains.
10. The apparatus of claim 8, wherein the processor is arranged to obtain a corresponding batch of sample data from each sample field by:
and acquiring batch sample data matched with the batch threshold value from each sample domain according to a preset batch threshold value through a data loader corresponding to each sample domain.
11. The apparatus of claim 7, wherein the processor, after obtaining the plurality of discriminative network parameters, further performs the following:
acquiring features which are output by a feature extraction network and correspond to the first sample data in the target model, and calculating a first twin loss function value through a twin network loss function according to the features corresponding to the first sample data;
the processor, after obtaining a plurality of second loss function values, further performs the following:
acquiring features, corresponding to each second sample data, output by a feature extraction network in the target model, and calculating a second twin loss function value through a twin network loss function according to the features corresponding to each second sample data;
the processor is configured to calculate a total loss function value from each first loss function value and each second loss function value, and perform backward calculation from the total loss function value by:
and calculating to obtain a total loss function value according to each first loss function value, each second loss function value, each first twin loss function value and each second twin loss function value, performing backward calculation according to the total loss function value, and updating the parameters of the target model according to the calculation result.
12. The apparatus of claim 11 wherein the processor is configured to calculate the total loss function value from each of the first loss function value, the second loss function value, the first twin loss function value, and the second twin loss function value by:
tensor addition is carried out on each first loss function value and each second loss function value to obtain a first function value;
tensor addition is carried out on the first twin loss function value and the second twin loss function value to obtain a second function value;
and calculating to obtain the total loss function value according to the first function value, the second function value and the hyper-parameter corresponding to the target model.
13. A computer-readable storage medium, on which a computer program is stored which, when being executed by a processor, carries out a method of training a machine learning model according to any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011591857.7A CN112633517B (en) | 2020-12-29 | 2020-12-29 | Training method of machine learning model, computer equipment and storage medium |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202011591857.7A CN112633517B (en) | 2020-12-29 | 2020-12-29 | Training method of machine learning model, computer equipment and storage medium |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112633517A true CN112633517A (en) | 2021-04-09 |
| CN112633517B CN112633517B (en) | 2024-02-02 |
Family
ID=75286086
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202011591857.7A Active CN112633517B (en) | 2020-12-29 | 2020-12-29 | Training method of machine learning model, computer equipment and storage medium |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112633517B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114283307A (en) * | 2021-12-24 | 2022-04-05 | 中国科学技术大学 | Network training method based on resampling strategy |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110210503A (en) * | 2019-06-14 | 2019-09-06 | 厦门历思科技服务有限公司 | A kind of seal recognition methods and device and equipment |
| CN111340850A (en) * | 2020-03-20 | 2020-06-26 | 军事科学院***工程研究院***总体研究所 | Ground target tracking method of unmanned aerial vehicle based on twin network and central logic loss |
| CN111462173A (en) * | 2020-02-28 | 2020-07-28 | 大连理工大学人工智能大连研究院 | Visual tracking method based on twin network discriminant feature learning |
| CN111612281A (en) * | 2020-06-23 | 2020-09-01 | 中国人民解放军国防科技大学 | Method and device for predicting pedestrian flow peak value of subway station and computer equipment |
| US20200342328A1 (en) * | 2019-04-26 | 2020-10-29 | Naver Corporation | Training a convolutional neural network for image retrieval with a listwise ranking loss function |
-
2020
- 2020-12-29 CN CN202011591857.7A patent/CN112633517B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200342328A1 (en) * | 2019-04-26 | 2020-10-29 | Naver Corporation | Training a convolutional neural network for image retrieval with a listwise ranking loss function |
| CN110210503A (en) * | 2019-06-14 | 2019-09-06 | 厦门历思科技服务有限公司 | A kind of seal recognition methods and device and equipment |
| CN111462173A (en) * | 2020-02-28 | 2020-07-28 | 大连理工大学人工智能大连研究院 | Visual tracking method based on twin network discriminant feature learning |
| CN111340850A (en) * | 2020-03-20 | 2020-06-26 | 军事科学院***工程研究院***总体研究所 | Ground target tracking method of unmanned aerial vehicle based on twin network and central logic loss |
| CN111612281A (en) * | 2020-06-23 | 2020-09-01 | 中国人民解放军国防科技大学 | Method and device for predicting pedestrian flow peak value of subway station and computer equipment |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN114283307A (en) * | 2021-12-24 | 2022-04-05 | 中国科学技术大学 | Network training method based on resampling strategy |
| CN114283307B (en) * | 2021-12-24 | 2023-10-27 | 中国科学技术大学 | Network training method based on resampling strategy |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112633517B (en) | 2024-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Xu et al. | Learning low-rank label correlations for multi-label classification with missing labels | |
| WO2022042123A1 (en) | Image recognition model generation method and apparatus, computer device and storage medium | |
| CN110232373A (en) | Face cluster method, apparatus, equipment and storage medium | |
| WO2020253127A1 (en) | Facial feature extraction model training method and apparatus, facial feature extraction method and apparatus, device, and storage medium | |
| CN111881714A (en) | Unsupervised cross-domain pedestrian re-identification method | |
| CN111581414B (en) | Method, device, equipment and storage medium for identifying, classifying and searching clothes | |
| WO2023103864A1 (en) | Node model updating method for resisting bias transfer in federated learning | |
| CN112906865B (en) | Neural network architecture searching method and device, electronic equipment and storage medium | |
| CN112418320A (en) | Enterprise association relation identification method and device and storage medium | |
| Wu et al. | Evolutionary multitasking with solution space cutting for point cloud registration | |
| CN111079930A (en) | Method and device for determining quality parameters of data set and electronic equipment | |
| CN112633517A (en) | Training method of machine learning model, computer equipment and storage medium | |
| CN107193979B (en) | Method for searching homologous images | |
| CN112966054A (en) | Enterprise graph node relation-based ethnic group division method and computer equipment | |
| Pan et al. | A deep learning based fast image saliency detection algorithm | |
| CN110765917A (en) | Active learning method, device, terminal and medium suitable for face recognition model training | |
| CN109493279B (en) | Large-scale unmanned aerial vehicle image parallel splicing method | |
| CN115878891A (en) | Live content generation method, device, equipment and computer storage medium | |
| CN111695526B (en) | Network model generation method, pedestrian re-recognition method and device | |
| CN114462526A (en) | Classification model training method and device, computer equipment and storage medium | |
| CN115587297A (en) | Method, apparatus, device and medium for constructing image recognition model and image recognition | |
| Sato et al. | Affine template matching by differential evolution with adaptive two‐part search | |
| CN113537270A (en) | Data classification method, multi-classification model training method, device, equipment and medium | |
| Korablyov et al. | Dendritic Artificial Immune Network Model for Computing. | |
| Li et al. | Improved algorithm for adversarial linear mixture MDPs with bandit feedback and unknown transition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |