CN116935160A - A training method, sample classification method, electronic device and medium - Google Patents

Abstract

The application provides a training method, a sample classification method, electronic equipment and a medium, comprising the following steps: acquiring training data, and acquiring characteristics of data samples in a support set and a query set by using a deep learning model; processing by utilizing the characteristics of the data samples of the support set, and obtaining the prototype vectors of each data sample of the support set and the difference values between the channels of the prototype vectors; processing by utilizing the characteristics of the data samples of the query set, and obtaining the difference value between all channels of the characteristics of the data samples of the query set; obtaining microKendall correlation coefficients between the characteristics of the data samples of the query set and the prototype vectors of each class by utilizing the differences between the channels of the prototype vectors of each class and the differences between the channels of the characteristics of the data samples of the query set; calculating a loss function according to the microKendell correlation coefficient; and training the deep learning model by using the loss function according to the training data. The training method can obtain a more elaborate deep learning model.

Description

A training method, sample classification method, electronic device and medium

Technical field

This application belongs to the field of machine learning technology and relates to a training method, especially a training method, a sample classification method, electronic equipment and media.

Background technique

As an important technology in the field of artificial intelligence, deep learning achieves data learning and processing by simulating the human brain neural network. Deep learning models have achieved success in many fields. However, when faced with a new task, a large amount of data needs to be collected to train the model to achieve better performance. Deep learning model training is a key step in deep learning. It improves the accuracy and performance of the model by training a large amount of data. Optimization algorithms based on gradient descent are mainly used in training.

Nowadays, deep learning model training still faces many challenges. For example, unbalanced training data will lead to model overfitting or underfitting; difficulty in selecting hyperparameters will affect model performance; long training time, high consumption of computing resources and other issues also restrict The development of deep learning model training. Therefore, there is still a lack of a training method to obtain more sophisticated deep learning models.

Contents of the invention

In view of the above shortcomings of the prior art, the purpose of this application is to provide a training method, sample classification method, electronic device and medium to solve the lack of a training method for obtaining a more sophisticated deep learning model in the prior art. The problem.

In the first aspect, this application provides a training method. The training method includes: obtaining training data, the training data includes a support set and a query set, the support set includes a plurality of images, and the query set includes a plurality of images; using a deep learning model to obtain the support set and the query set. The characteristics of the data samples in the query set are described; the characteristics of the data samples in the support set are used for processing, and the differences between the prototype vectors of each category of data samples in the support set and each channel of the prototype vector are obtained; the differences between the channels of the prototype vector are obtained Process the characteristics of the data samples to obtain the differences between the channels of the characteristics of the data samples of the query set; use the differences between the channels of the prototype vectors of each category and the differences of the data samples of the query set. The difference between each channel of the feature is used to obtain the differentiable Kendall correlation coefficient between the characteristics of the data sample of the query set and the prototype vector of each category; calculate the loss function according to the differentiable Kendall correlation coefficient; According to the training data, the deep learning model is trained using the loss function.

In this application, the differentiable Kendall correlation coefficient is obtained based on the difference between the channels of the prototype vector and the characteristics of the data samples in the query set, and then the data samples are used to train the deep learning model. During the training process, a loss function with differentiable Kendall correlation coefficients as parameters is used for adjustment. This training method can establish the correlation between the channel of the prototype vector and the channel of the query set sample feature, obtain the differentiable Kendall correlation coefficient, and further obtain a more sophisticated deep learning model.

In an implementation manner of the first aspect, using the characteristics of the data samples of the support set for processing, obtaining the prototype vector of each category of data samples in the support set and the difference between each channel of the prototype vector includes: according to According to the characteristics of the data samples in the support set, obtain the prototype vectors of each category of data samples in the support set; use the prototype vectors to obtain each channel value of the prototype vector; obtain all the channel values of the prototype vector according to the characteristics of the data samples in the support set. The difference between the channels of the prototype vector.

In an implementation manner of the first aspect, processing is performed using the characteristics of the data samples in the query set, and obtaining the differences between the channels of the characteristics of the data samples in the query set includes: according to the data samples in the query set The characteristics of the query set are used to obtain the channel values of the characteristics of the data samples of the query set; the differences between the channels of the characteristics of the data samples of the query set are obtained by using the channel values of the characteristics of the data samples of the query set.

In an implementation manner of the first aspect, obtaining the prototype vector of each category of the support set includes: obtaining the prototype of each category of data samples of the support set based on the average value of the characteristics of the data samples of each category of the support set. vector.

In an implementation of the first aspect, the calculation method of the loss function includes: obtaining the data sample of the query set to which the query set belongs based on the differentiable Kendall correlation coefficient between the characteristics of the query set data sample and the prototype vector of each category. The probability of each category; the loss function is obtained by using the probability that the data samples in the query set belong to each category and the number of data samples in the query set.

In an implementation of the first aspect, the loss function is: Among them, L is the loss function, |Q| is the number of samples in the query set, x is the input query set data sample, its category is y, c _y , c _j are the yth category and jth category of the support set Prototype vector, t is a constant, N is the total number of categories in the support set, /> is the differentiable Kendall correlation coefficient between the two input vectors, f _θ is the deep learning model, and θ is the parameter of the deep learning model.

In an implementation manner of the first aspect, using the loss function to train the deep learning model includes: when using the loss function to train the deep learning model, adjusting at least one parameter of the deep learning model. parameter value.

In an implementation manner of the first aspect, the deep learning model is a convolutional neural network ResNet or a Transformer-based neural network ViT.

In the second aspect, this application provides a sample classification method. The sample classification method includes: obtaining at least one reference sample of each category and a target sample to be classified; using a deep learning model to obtain characteristics of the reference sample and the target sample, the deep learning model using any one of the first aspects Obtained by training by the described training method; process according to the characteristics of the reference sample to obtain the prototype vector of each category; process according to the characteristics of the target sample and the prototype vector of each category to obtain the characteristics of the target sample The channel importance ranking results and the channel importance ranking results of the prototype vectors of each category; using the channel importance ranking results of the features of the target sample and the channel importance ranking results of the prototype vectors of each category , obtain the Kendall correlation coefficient between the characteristics of the target sample and the prototype vector of each category; use the Kendall correlation coefficient to obtain the classification result of the target sample.

In an implementation manner of the second aspect, obtaining the Kendall correlation coefficient includes: obtaining the channel importance ranking result of the characteristics of the target sample and the prototype vector; using the relationship between the characteristics of the target sample and the prototype vector. Match the importance ranking results of all corresponding channel pairs. The matching results include channel pairs with consistent importance ranking results and channel pairs with inconsistent importance ranking results; according to the number of channel pairs with consistent importance ranking results, importance ranking results The Kendall correlation coefficient is obtained for the number of inconsistent channel pairs and the number of total channel pairs.

In a third aspect, the present application provides an electronic device. The electronic device includes: a memory for storing a computer program; a processor for executing the computer program stored in the memory, so that the electronic device executes the training method according to any one of the first aspects. and/or the sample classification method described in the second aspect.

In a fourth aspect, this application provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the training method described in any one of the first aspects and/or the training method described in the second aspect is implemented. sample classification method.

Description of the drawings

FIG. 1A shows a schematic diagram of an application scenario of the training method described in the embodiment of the present application.

Figure 1B shows a schematic diagram of another application scenario of the training method described in the embodiment of the present application.

Figure 2 shows a schematic flow chart of the training method described in the embodiment of the present application.

Figure 3 shows a schematic flow chart of the training method described in the embodiment of the present application.

Figure 4 shows a schematic flow chart of the training method described in the embodiment of the present application.

Figure 5 shows a schematic flow chart of the calculation method of the loss function according to the embodiment of the present application.

Figure 6 shows a schematic flow chart of the sample classification method according to the embodiment of the present application.

Figure 7 shows a schematic flow chart of the sample classification method according to the embodiment of the present application.

FIG. 8 shows a schematic structural diagram of an electronic device according to an embodiment of the present application.

Component label description

1 Medical image classification system

11 Image acquisition equipment

12 local processors

13 display terminal

2 terminal-cloud interactive system

20 terminal

21 Cloud Server

800 Electronic equipment

810 memory

820 processor

830 monitor

Steps S11～S17

Steps S131～S133

Steps S141～S142

S21～S22 steps

Steps S31～S36

Steps S351～S353

Detailed ways

The following describes the implementation of the present application through specific examples. Those skilled in the art can easily understand other advantages and effects of the present application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments. Various details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, as long as there is no conflict, the following embodiments and the features in the embodiments can be combined with each other.

It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present application in a schematic manner, and the drawings only show the components related to the present application and do not follow the actual implementation of the component numbers, shapes and components. Dimension drawing, in actual implementation, the type, quantity and proportion of each component can be arbitrarily changed, and the component layout type may also be more complex.

Since in small-sample learning, the samples in the base data set used for model training and the samples in the small-sample task have no intersection in category, the model has not actually seen the category of the data in the small-sample task during training. This will lead to significant differences between the extracted features of the data samples on the small sample task and the sample features extracted by traditional supervised learning methods. An important characteristic of data sample features on small sample tasks is that the value distribution of the features is flatter. The values on most channels of the features are small and very close to each other. The negative Euclidean distance and cosine similarity used in existing methods are classified by directly calculating the geometric similarity between features, which will cause the classification results to be actually determined by a very small number of channels with large values. Dominant, and for the channels with smaller values that account for most of the features, it is difficult to distinguish their importance in classification. Among these channels, some channels may represent key distinguishing features of a certain category. If their importance cannot be distinguished well, incorrect classification results will be obtained.

At least to address the above problems, embodiments of the present application provide a training method. The training method includes: obtaining training data, the training data includes a support set and a query set, the support set includes a plurality of images, and the query set includes a plurality of images; using a deep learning model to obtain the support set and the query set. The characteristics of the data samples in the query set are described; the characteristics of the data samples in the support set are used for processing, and the differences between the prototype vectors of each category of data samples in the support set and each channel of the prototype vector are obtained; the differences between the channels of the prototype vector are obtained Process the characteristics of the data samples to obtain the differences between the channels of the characteristics of the data samples of the query set; use the differences between the channels of each of the prototype vectors and the characteristics of the data samples of the query set The difference between each channel is used to obtain the differentiable Kendall correlation coefficient between the characteristics of the data samples of the query set and each of the prototype vectors; calculate the loss function according to the differentiable Kendall correlation coefficient; according to the The training data is used to train the deep learning model using the loss function.

In the embodiment of the present application, the differentiable Kendall correlation coefficient is obtained based on the difference between the channels of the prototype vector and the difference between the characteristics of the data samples in the query set, and then the data samples are used to perform the deep learning model Training, the loss function with differentiable Kendall correlation coefficient as parameter is used for adjustment during the training process. This training method can establish the correlation between the channel of the prototype vector and the channel of the query set sample feature, obtain the differentiable Kendall correlation coefficient, and further obtain a more sophisticated deep learning model.

FIG. 1A shows a schematic diagram of an application scenario of the training method described in the embodiment of the present application. The medical image classification system 1 can be used to implement the training method provided by the embodiment of the present application, but the application scenarios of the training method provided by the embodiment of the present application are not limited to the medical image classification system 1 shown in Figure 1A. As shown in FIG. 1A , the medical image classification system 1 includes an image acquisition device 11 , a local processor 12 and a display terminal 13 . The training method provided by the embodiment of this application can be applied to the local processor 12 .

Wherein, the image acquisition device 11 is used to collect medical imaging data samples. Optionally, the image acquisition device may be a CT scanner, an X-ray machine, a magnetic resonance imaging device, etc., and this application is not limited thereto.

The local processor 12 is used to process and analyze the collected medical image data samples, and extract feature classification results of the medical image data samples. The local processor 12 may be one processor or a processor cluster composed of multiple processors, and the details are not limited here.

The display terminal 13 may be a device with a display function, and is used to display the processed classification results for analysis by doctors. Optionally, the display terminal 13 may be a medical display.

It should be noted that although FIG. 1A only shows an image acquisition device 11, a local processor 12 and a display terminal 13, it should be understood that the example in FIG. 1A is only used to understand this solution. Specifically, the display terminal and The number of servers should be flexibly determined based on the actual situation.

In other possible implementations, the training method described in this application can be applied to device-cloud interaction scenarios. Figure 1B shows a schematic diagram of another application scenario of the training method described in the embodiment of the present application. As shown in Figure 1B, the terminal-cloud interactive system 2 includes a terminal 20 and a cloud server 21. The terminal 20 and the cloud server 21 can communicate. The communication method is not limited to wired or wireless methods. The training method described in this application can Applied to cloud server 21.

The terminal 20 may be mobile or fixed. For example, the terminal 20 may be a wireless terminal or a wired terminal. The wireless terminal may refer to a device with wireless transceiver functions and may be deployed in a data processing center and a wired terminal. Medical laboratory. The terminal may be a medical processor, a mobile phone, a tablet, a notebook computer, etc., which are not limited here.

The cloud server 21 may include one or more servers, or one or more processing nodes, or one or more virtual machines running on the servers. The cloud server 21 may also be called a server cluster, a management platform, a data server, or a server cluster. Processing centers, etc. are not limited in the embodiments of this application.

The technical solutions in the embodiments of the present application will be described in detail below with reference to the drawings in the embodiments of the present application.

The following embodiments of the present application provide a training method, which can be implemented, for example, through the local processor 12 shown in Figure 1A or the cloud server 21 shown in Figure 1B. Figure 2 shows a schematic flow chart of the training method described in the embodiment of the present application. As shown in Figure 2, the training method includes steps S11 to S17.

Step S11: Obtain training data. The training data includes a support set and a query set. The support set includes multiple images, and the query set includes multiple images. The construction form of the support set is N-way, K-shot, that is, there are N categories of images in the support set, and each category contains K images.

Step S12: Use a deep learning model to obtain characteristics of the data samples in the support set and the query set.

Step S13: Use the characteristics of the data samples in the support set for processing, and obtain the differences between the prototype vectors of each category of data samples in the support set and each channel of the prototype vector.

Step S14: Use the characteristics of the data samples in the query set to perform processing, and obtain the differences between the channels of the characteristics of the data samples in the query set.

Step S15, using the difference between the channels of each prototype vector and the difference between the channels of the characteristics of the data samples of the query set, obtain the characteristics of the data samples of the query set and each of the Differentiable Kendall correlation coefficients between prototype vectors. Among them, the differentiable Kendall correlation coefficient is used to measure the degree of correlation between the strength of the monotonic relationship between two ordered variables. The greater the absolute value of the correlation coefficient, the stronger the correlation. The smaller the correlation coefficient, the degree of correlation. The weaker.

In some possible implementations, the calculation formula of the differentiable Kendall correlation coefficient is:

Among them, u=(u ₁ ,u ₂ ,...u _n ) is the characteristic of the data sample of the query set, v=(v ₁ ,v ₂ ,...v _n ) is the prototype vector, n is the total number of channels, u _i is the value of vector u on the i-th channel, vi is the value of vector _v on the i-th channel, and α is a parameter.

Step S16: Calculate a loss function based on the differentiable Kendall correlation coefficient.

Step S17: Use the loss function to train the deep learning model based on the training data.

In the embodiment of the present application, the differentiable Kendall correlation coefficient is obtained based on the difference between the channels of the prototype vector and the difference between the characteristics of the data samples in the query set, and then the data samples are used to perform the deep learning model Training, the loss function with differentiable Kendall correlation coefficient as parameter is used for adjustment during the training process. This training method can establish the correlation between the channel of the prototype vector and the channel of the query set sample feature, obtain the differentiable Kendall correlation coefficient, and further obtain a more sophisticated deep learning model.

Figure 3 shows a schematic flow chart of the training method described in the embodiment of the present application. As shown in Figure 3, using the characteristics of the data samples in the support set for processing, and obtaining the difference between the prototype vectors of each category of data samples in the support set and each channel of the prototype vector includes steps S131 to S133.

Step S131: According to the characteristics of the data samples in the support set, obtain prototype vectors of each category of data samples in the support set.

Step S132: Use the prototype vector to obtain each channel value of the prototype vector.

Step S133: Obtain the difference between each channel of the prototype vector according to the channel value of the prototype vector.

In some possible implementations, u _i is the value of vector u on the i-th channel, u _j is the value of vector u on the j-th channel, then the i-th and j-th prototype vector u The calculation formula for the difference between channels is: δ=u _i -u _j .

Figure 4 shows a schematic flow chart of the training method described in the embodiment of the present application. As shown in Figure 4, processing is performed using the characteristics of the data samples of the query set. Obtaining the difference between the channels of the characteristics of the data samples of the query set includes steps S141 to S142.

Step S141: Obtain the channel values of the characteristics of the data samples in the query set according to the characteristics of the data samples in the query set.

Step S142: Use the channel values of the characteristics of the data samples in the query set to obtain the differences between the channels of the characteristics of the data samples in the query set.

In one embodiment of the present application, obtaining the prototype vector of each category of the support set includes step S134: obtaining the prototype of each category of data samples in the support set based on the average value of the characteristics of the data samples of each category of the support set. vector.

Figure 5 shows a schematic flow chart of the calculation method of the loss function according to the embodiment of the present application. As shown in Figure 5, the calculation method of the loss function includes steps S21 to S22.

Step S21: Obtain the probability that the data sample of the query set belongs to each category based on the differentiable Kendall correlation coefficient between the characteristics of the query set data sample and the prototype vector of each category.

Step S22: Obtain the loss function using the probability that the data samples in the query set belong to each category and the number of data samples in the query set.

In some possible implementations, the support set is constructed in the form of 5-way, 5-shot. The probability that the data sample of the query set belongs to each category is obtained according to the differentiable Kendall correlation coefficient between the characteristics of the query set data sample and the prototype vector of each category. For the samples in the support set, all five samples in each category are averaged as the prototype vector of each category. The calculation formula for obtaining the probability that the data sample of the query set belongs to each category based on the differentiable Kendall correlation coefficient between the characteristics of the query set data sample and the prototype vector of each category is:

Among them, c _k is the class prototype vector of the k-th category, c _j is the class prototype vector of the j-th category, sim is the similarity measure, N is the total number of categories, t is a constant, x is the input query set data sample, y is the category to which the sample belongs, f _θ is the deep learning model, and θ is the parameter of the deep learning model.

In an embodiment of the present application, the calculation formula of the loss function is:

Among them, L is the loss function, |Q| is the number of samples in the query set, x is the input query set data sample, its category is y, c _y , c _j are the yth category and jth category of the support set Prototype vector, t is a constant, N is the total number of categories in the support set, is the differentiable Kendall correlation coefficient between the two input vectors, f _θ is the deep learning model, and θ is the parameter of the deep learning model.

In an embodiment of the present application, using the loss function to train the deep learning model includes: adjusting at least one parameter value of the deep learning model when using the loss function to train the deep learning model.

In one embodiment of the present application, the deep learning model is a convolutional neural network ResNet (Residual neural network, residual neural network) or a Transformer-based neural network ViT (Visiontransformer).

Figure 6 shows a schematic flow chart of the sample classification method according to the embodiment of the present application. As shown in Figure 6, the sample classification method includes steps S31 to S36.

Step S31: Obtain at least one reference sample of each category and the target sample to be classified. Optionally, the reference sample is a sample with existing classification results.

Step S32: Use a deep learning model to obtain the characteristics of the reference sample and the target sample. The deep learning model is trained using the training method described in any embodiment of this application. Optionally, use the trained deep learning model as a feature extractor to respectively extract features of the query set samples and support set samples of the target sample.

Step S33: Process according to the characteristics of the reference sample to obtain prototype vectors of each category.

Step S34: Process according to the characteristics of the target sample and the prototype vectors of each category to obtain the channel importance ranking results of the characteristics of the target sample and the channel importance ranking results of the prototype vectors of each category. Optionally, sort according to the value of each channel in the feature, and the importance of each channel is expressed by the sorting result.

In some possible implementations, for a feature with 5 channels, the values of each channel are (0.1, 0.11, 0.13, 0.12, 0.14). After sorting them, the importance ranking result of the features is: (1,2,4,3,5).

Step S35: Use the channel importance ranking results of the features of the target sample and the channel importance ranking results of the prototype vectors of each category to obtain the correlation between the features of the target sample and the prototype vectors of each category. Del correlation coefficient.

Step S36: Use the Kendall correlation coefficient to obtain the classification result of the target sample. Wherein, the classification result is the category corresponding to the prototype vector with the largest Kendall correlation coefficient of the characteristics of the target sample.

Figure 7 shows a schematic flow chart of the sample classification method according to the embodiment of the present application. As shown in Figure 7, obtaining the Kendall correlation coefficient includes steps S351 to S353.

Step S351: Obtain the characteristics of the target sample and the channel importance ranking result of the prototype vector.

Step S352, use the importance ranking results of all corresponding channel pairs between the characteristics of the target sample and the prototype vector to perform matching. The matching results include channel pairs with consistent importance ranking results and channel pairs with inconsistent importance ranking results. Optionally, any channel pair is a set of any two channels in each channel. The importance ranking result of any channel pair is the importance ranking result of any two channels. Optionally, if the matching fails, the obtained matching result is a channel pair with inconsistent importance ranking results. If the matching is successful, the obtained matching result is a channel pair with consistent importance ranking results.

Step S353: Obtain the Kendall correlation coefficient based on the number of channel pairs with consistent importance ranking results, the number of channel pairs with inconsistent importance ranking results, and the number of total channel pairs.

In some possible implementations, the calculation formula of the Kendall correlation coefficient is:

Among them, N _con is the number of channel pairs with consistent importance ranking, N _dis is the number of channel pairs with inconsistent importance ranking, and N _total is the number of total channel pairs.

Next, the calculation method of the above Kendall correlation coefficient will be introduced through a specific example. Among them, for any two features x and y, the channel importance ranking results of feature x and feature y are x=(1,2,3) and y=(1,3,2) respectively. Among them, channel 1 and channel 2 of feature x are (1,2), channel 1 and channel 2 of feature y are (1,3), and the importance ranking results of channel pairs (including channels 1 and 2) are consistent. Channel 1 and channel 3 of feature x are (1,3), channel 1 and channel 3 of feature y are (1,2), and the importance ranking results of channel pairs (including channels 1 and 3) are consistent. Channel 2 and channel 3 of feature x are (2,3), channel 2 and channel 3 of feature y are (3,2), and the importance ranking results of channel pairs (including channels 2 and 3) are inconsistent. The number of channel pairs with consistent importance ranking results is N _con =2, the number of channel pairs with inconsistent importance ranking results is N _dis =1, and the total number of channel pairs is N _total =3. The Kendall correlation coefficient obtained based on the number of channel pairs with consistent importance ranking results, the number of channel pairs with inconsistent importance ranking results, and the total number of channel pairs is 1/3.

It should be noted that the above are only possible implementation methods of the embodiments of the present application, and the present application is not limited thereto.

In the several embodiments provided in this application, it should be understood that the disclosed system, device or method can be implemented in other ways. For example, the device embodiments described above are only illustrative. For example, the division of modules/units is only a logical function division. In actual implementation, there may be other division methods, for example, multiple modules or units may be combined or can be integrated into another system, or some features can be ignored, or not implemented. On the other hand, the coupling or direct coupling or communication connection between each other shown or discussed may be indirect coupling or communication connection through some interfaces, devices or modules or units, which may be in electrical, mechanical or other forms.

Modules/units described as separate components may or may not be physically separate. Components shown as modules/units may or may not be physical modules, that is, they may be located in one place, or they may be distributed to multiple network units. superior. Some or all of the modules/units may be selected according to actual needs to achieve the purpose of the embodiments of the present application. For example, each functional module/unit in various embodiments of the present application can be integrated into a processing module, or each module/unit can exist physically alone, or two or more modules/units can be integrated into one module/unit. in the unit.

Those of ordinary skill in the art should further realize that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented with electronic hardware, computer software, or a combination of both. In order to clearly illustrate the hardware and software interchangeability. In the above description, the composition and steps of each example have been generally described according to functions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Skilled artisans may implement the described functionality using different methods for each specific application, but such implementations should not be considered beyond the scope of this application.

An embodiment of the present application also provides an electronic device. FIG. 8 shows a schematic structural diagram of an electronic device 800 according to an embodiment of the present application. As shown in FIG. 8 , the electronic device 800 in this embodiment includes a memory 810 and a processor 820 .

The memory 810 is used to store computer programs; optionally, the memory 810 includes various media that can store program codes, such as ROM, RAM, magnetic disks, USB disks, memory cards, or optical disks.

Specifically, memory 810 may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and/or cache memory. Electronic device 800 may further include other removable/non-removable, volatile/non-volatile computer system storage media. The memory 810 may include at least one program product having a set (eg, at least one) of program modules configured to perform the functions of various embodiments of the present application.

The processor 820 is connected to the memory 810 and is used to execute the computer program stored in the memory 810, so that the electronic device 800 executes the training method described in any embodiment of this application and/or the sample classification method described in the embodiment of this application.

Optionally, the processor 820 may be a general-purpose processor, including a central processing unit (CPU for short), a network processor (NP for short), etc.; it may also be a digital signal processor (Digital signal processor for short) DSP), Application specific integrated circuit (ASIC for short), Field programmable gate array (FPGA for short) or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.

Optionally, in this embodiment, the electronic device 800 may also include a display 830. The display 830 is communicatively connected to the memory 810 and the processor 820, and is used to display the relevant graphical user interface (Graphics user interface, abbreviation) of the training method described in any embodiment of the present application and/or the sample classification method described in the embodiment of the present application. GUI) interactive interface.

Embodiments of the present application also provide a computer-readable storage medium on which a computer program is stored. When the program is executed by the processor, the training method described in any embodiment of this application and/or the sample classification method described in the embodiment of this application is implemented.

The descriptions of the processes or structures corresponding to each of the above drawings have different emphasis. For parts that are not described in detail in a certain process or structure, please refer to the relevant descriptions of other processes or structures.

The above embodiments only illustrate the principles and effects of the present application, but are not used to limit the present application. Anyone familiar with this technology can modify or change the above embodiments without departing from the spirit and scope of the present application. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed in this application shall still be covered by the claims of this application.

Claims (12)

1. A training method, characterized in that it includes: Obtain training data, the training data includes a support set and a query set, the support set includes a plurality of images, and the query set includes a plurality of images; Using a deep learning model to obtain characteristics of the data samples in the support set and the query set; Utilize the characteristics of the data samples in the support set for processing, and obtain the difference between the prototype vector of each category of data samples in the support set and each channel of the prototype vector; Using the characteristics of the data samples of the query set for processing, obtaining the differences between the channels of the characteristics of the data samples of the query set; Using the difference between each channel of the prototype vector of each category and the difference between each channel of the characteristics of the data sample of the query set, obtain the characteristics of the data sample of the query set and the prototype of each category Differentiable Kendall correlation coefficient between vectors; Calculate a loss function based on the differentiable Kendall correlation coefficient; According to the training data, the deep learning model is trained using the loss function. 2. The training method according to claim 1, characterized in that the characteristics of the data samples of the support set are used for processing to obtain the prototype vectors of each category of data samples in the support set and the distance between each channel of the prototype vector. Differences include: According to the characteristics of the data samples in the support set, obtain the prototype vectors of each category of data samples in the support set; Using the prototype vector to obtain each channel value of the prototype vector; According to the value of each channel of the prototype vector, the difference between each channel of the prototype vector is obtained. 3. The training method according to claim 1, characterized in that, using the characteristics of the data samples of the query set for processing, obtaining the difference between the channels of the characteristics of the data samples of the query set includes: Obtain each channel value of the characteristics of the data samples in the query set according to the characteristics of the data samples in the query set; The difference between each channel of the characteristics of the data samples of the query set is obtained by using the channel values of the characteristics of the data samples of the query set. 4. The training method according to claim 1, wherein obtaining prototype vectors of each category of the support set includes: According to the average value of the characteristics of the data samples of each category of the support set, the prototype vector of the data samples of each category of the support set is obtained. 5. The training method according to claim 1, characterized in that the calculation method of the loss function includes: Obtain the probability that the data sample of the query set belongs to each category according to the differentiable Kendall correlation coefficient between the characteristics of the query set data sample and the prototype vector of each category; The loss function is obtained using the probability that the data samples in the query set belong to each category and the number of data samples in the query set. 6. The training method according to claim 5, characterized in that the loss function is: Among them, L is the loss function, |Q| is the number of samples in the query set, x is the input query set data sample, its category is y, c _y , c _j are the yth category and jth category of the support set Prototype vector, t is a constant, N is the total number of categories in the support set, is the differentiable Kendall correlation coefficient between the two input vectors, f _θ is the deep learning model, and θ is the parameter of the deep learning model. 7. The training method according to claim 1, wherein using the loss function to train the deep learning model includes: When the loss function is used to train the deep learning model, a parameter value of at least one of the deep learning models is adjusted. 8. The training method according to claim 1, characterized in that the deep learning model is a convolutional neural network ResNet or a Transformer-based neural network ViT. 9. A sample classification method, characterized by including: Obtain at least one reference sample for each category and the target sample to be classified; Using a deep learning model to obtain the characteristics of the reference sample and the target sample, the deep learning model is trained using the training method as described in any one of claims 1 to 8; Process according to the characteristics of the reference sample to obtain prototype vectors of each category; Process according to the characteristics of the target sample and the prototype vectors of each category to obtain the channel importance ranking results of the characteristics of the target sample and the channel importance ranking results of the prototype vectors of each category; Using the channel importance ranking results of the features of the target sample and the channel importance ranking results of the prototype vectors of each category, the Kendall correlation coefficient between the features of the target sample and the prototype vectors of each category is obtained ; The Kendall correlation coefficient is used to obtain the classification result of the target sample. 10. The sample classification method according to claim 9, characterized in that obtaining the Kendall correlation coefficient includes: Obtain the characteristics of the target sample and the channel importance ranking results of the prototype vector; Matching is performed using the importance ranking results of all corresponding channel pairs between the characteristics of the target sample and the prototype vector. The matching results include channel pairs with consistent importance ranking results and channel pairs with inconsistent importance ranking results; The Kendall correlation coefficient is obtained according to the number of channel pairs with consistent importance ranking results, the number of channel pairs with inconsistent importance ranking results, and the number of total channel pairs. 11. An electronic device, characterized in that the electronic device includes: Memory, used to store computer programs; Processor, the processor is used to execute the computer program stored in the memory, so that the electronic device executes the training method according to any one of claims 1 to 8 and/or any one of claims 9 to 10 The sample classification method described. 12. A computer-readable storage medium with a computer program stored thereon, characterized in that when the program is executed by a processor, the training method according to any one of claims 1 to 8 and/or claims 9 to 8 is implemented. The sample classification method described in any one of 10.