WO2020098121A1

WO2020098121A1 - Method and device for training fast model, computer apparatus, and storage medium

Info

Publication number: WO2020098121A1
Application number: PCT/CN2018/125592
Authority: WO
Inventors: 徐玲玲
Original assignee: 平安科技（深圳）有限公司
Priority date: 2018-11-13
Filing date: 2018-12-29
Publication date: 2020-05-22
Also published as: CN109614989A; CN109614989B

Abstract

Embodiments of the present application disclose a method and device for training a fast model, a computer apparatus, and a storage medium. The method comprises: acquiring a pre-configured training sample image; inputting the training sample image into a pre-configured auxiliary training model and an initial fast model; calculating a feature distance between a feature vector of the training sample image extracted by the auxiliary training model and a feature vector of the training sample image extracted by the fast model; and performing back propagation on the feature distance to correct a weight parameter in the fast model. When the fast model is trained, a sample image involved in training does not need to be marked, thereby reducing the time and effort required for marking, and increasing training speed. A distance between a feature vector representing a sample image feature and output by the auxiliary model and a feature vector representing a sample image feature and output by the fast model is directly calculated, and back propagation is performed, thereby maximally shortening the training time.

Description

Training method, device, computer equipment and storage medium of rapid model

This application requires the priority of the Chinese patent application submitted to the China Patent Office on November 13, 2018, with the application number 201811348231.6 and the invention titled "fast model training method, device, computer equipment and storage medium", all of which are approved by The reference is incorporated in this application.

Technical field

The embodiments of the present application relate to the field of model training, and in particular, to a rapid model training method, device, computer equipment, and storage medium.

Background technique

Since the advent of mathematical methods for simulating human actual neural networks, people have gradually become accustomed to calling this artificial neural network directly as a neural network. Neural networks have broad and attractive prospects in the fields of system identification, pattern recognition, and intelligent control. Especially in intelligent control, people are particularly interested in the self-learning function of neural networks, and regard this important feature of neural networks as One of the key keys to solve the problem of controller adaptability in automatic control.

In the prior art, in order to make the neural network model have the ability to accurately classify a certain type or multiple types of images, it is necessary to train the initialized neural network model. Among them, the training method is to collect a certain level of sample images. In order to make the neural network model trained to convergence more robust, the training samples required are often massive. During training, the training samples are manually calibrated, and then the sample images are input into the neural network model to obtain the classification results output by the neural network model. Compare the classification results with the artificial calibration. If they are inconsistent, correct the weights of the neural network model through the reverse algorithm to make the neural network model gradually converge. Due to the strong randomness of the output during model training, the training process is extremely long .

The inventor of the present application found in the research that in order to improve the accuracy of the neural network model in the prior art, it often takes a lot of sample images to spend a lot of time to train the neural network model to the convergence state, but does not have the above Under the condition of conditions, the neural network model is often unable to be trained, or the neural network model obtained by training has poor stability and low accuracy.

Summary of the invention

Embodiments of the present application provide a training method, device, computer equipment, and storage medium that can train a model to convergence with a small number of samples and a short time through an auxiliary training model.

In order to solve the above technical problems, a technical solution adopted by the embodiment created by the present application is: to provide a rapid model training method, including: acquiring a preset training sample image; inputting the training sample image into a preset assistant In the training model and the initial fast model, the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast model is a neural network model to be trained; calculating the auxiliary training The feature distance between the feature vector of the training sample image extracted by the model and the feature vector of the training sample image extracted by the fast model; by back-propagating the feature distance, the weight parameter in the fast model is corrected.

To solve the above technical problems, embodiments of the present application also provide a rapid model training device, including: an acquisition module for acquiring a preset training sample image; a processing module for inputting the training sample image into a preset In the auxiliary training model and the initial fast model, the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast model is a neural network model to be trained; a calculation module, Used to calculate the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the fast model; the execution module is used to back-propagate the feature distance To correct the weight parameters in the fast model.

To solve the above technical problems, the embodiments of the present application further provide a computer device, including a memory and a processor, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, The processor executes the steps of the fast model training method described above.

To solve the above technical problems, the embodiments of the present application further provide a storage medium storing computer-readable instructions, which when executed by one or more processors cause the one or more processors to execute the above The steps of the rapid model training method are described.

By training the fast model, there is no need to mark the sample images participating in the training, which saves the time and effort required for the marking, and improves the training speed. At the same time, directly calculate the distance between the feature vector output from the auxiliary model that characterizes the sample image and the feature vector output from the fast model that characterizes the sample image (Euclidean distance and / or cosine distance) and perform back propagation. The method transforms the training of the fast model into a simple regression algorithm, which can shorten the training time to the maximum extent, and can guarantee the accuracy rate of the fast model output when the training is completed.

BRIEF DESCRIPTION

FIG. 1 is a schematic diagram of a basic process of a rapid model training method according to an embodiment of this application;

FIG. 2 is a schematic flowchart of determining whether to perform back propagation through a threshold according to an embodiment of the present application;

FIG. 3 is a schematic flowchart of calculating a feature distance between two model extraction feature vectors according to an embodiment of the present application;

4 is a schematic flowchart of obtaining training sample images according to specific application scenarios according to an embodiment of the present application;

5 is a schematic flowchart of a training sample image acquired by a database inspection station according to an embodiment of the present application;

6 is a schematic flowchart of a method for generating a derived sample image according to an embodiment of this application;

7 is a schematic diagram of a basic structure of a training device for a rapid model according to an embodiment of the present application;

8 is a block diagram of a basic structure of a computer device according to an embodiment of the present application.

detailed description

In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application.

Some processes described in the specification and claims of this application and the above drawings include multiple operations in a specific order, but it should be clearly understood that these operations may not be in the order in which they appear in this document Execution or parallel execution. The sequence numbers of operations such as 101 and 102 are only used to distinguish different operations. The sequence number itself does not represent any execution sequence. In addition, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent a sequence, nor limit "first" and "second". Are different types.

The technical solutions in the embodiments of the present application will be described clearly and completely in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without making creative work fall within the protection scope of the present application.

Those skilled in the art can understand that the "terminal" and "terminal device" used herein include both wireless signal receiver devices, which only have wireless signal receiver devices without transmitting capabilities, and also include hardware for receiving and transmitting hardware. A device having a device capable of performing receiving and transmitting hardware for bidirectional communication on a bidirectional communication link. Such devices may include: cellular or other communication devices with single-line displays or multi-line displays or cellular or other communication devices without multi-line displays; PCS (Personal Communications Services), which can combine voice and data Processing, fax and / or data communication capabilities; PDA (Personal Digital Assistant), which can include radio frequency receivers, pagers, Internet / Intranet access, web browsers, notepads, calendars and / or GPS (Global Positioning System (Global Positioning System) receiver; conventional laptop and / or palmtop computer or other device that has and / or includes a conventional radio frequency receiver and / or palmtop computer or other device. As used herein, "terminal" and "terminal equipment" may be portable, transportable, installed in a vehicle (aeronautical, maritime, and / or terrestrial), or adapted and / or configured to operate locally, and / or In a distributed form, it operates at any other location on the earth and / or space. The "terminal" and "terminal device" used herein may also be a communication terminal, an Internet terminal, a music / video playback terminal, for example, may be a PDA, MID (Mobile Internet Device), and / or have music / video playback Functional mobile phones can also be smart TVs, set-top boxes and other devices.

Please refer to FIG. 1 for details. FIG. 1 is a schematic diagram of a basic process of the fast model training method of this embodiment.

As shown in Figure 1, a fast model training method includes:

S1100. Acquire a preset training sample image;

Before training the fast model, you need to prepare training sample images. The number of training sample images corresponds to the specific requirements for the fast model. For example, the number of training sample images for application scenarios that require high generalization capabilities of the fast model More, otherwise, less.

For the acquisition of training sample images, the public image database, its own image database, or data crawling can be used to crawl from the Internet.

Fast model means that compared with the existing neural network model model, the number of free quantities (weights) that need to be determined through training is small, the model size is small, or only a small number of training sample images are needed to train to convergence Neural network model.

Due to its small size, the fast model has low requirements for computer processing power and a slightly single function. Therefore, it is often used in (not limited to) mobile smart terminals or environmental factors (the image to be processed changes more single). Application scenarios.

S1200. Input the training sample image into a preset auxiliary training model and an initial fast model, where the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast The model is the neural network model to be trained;

In this embodiment, in order to implement rapid model training, an auxiliary training model needs to be used for training. In addition, in this embodiment, the auxiliary training model participating in training is pre-trained to a convergence state, that is, a neural network model that can accurately classify training sample images. In this embodiment, the auxiliary training model can be a convolutional neural network model (CNN) that has been trained to a convergence state, but the auxiliary training model can also be: a deep neural network model (DNN), a recurrent neural network model (RNN), or the above Deformation model of three network models. The fast model can also be any one of the above three models or a deformed model, but the fast model is smaller than the auxiliary training model.

It should be pointed out that the training direction of the fast model and the auxiliary training model must be the same, or the purpose of the fast model training is one of the functions of the auxiliary training model. For example, the purpose of fast model training is to recognize the face images of yellow people. The function of the auxiliary training model should also be to recognize the face images, but compared to the fast model, the auxiliary training model has more powerful recognition. Capabilities, such as being able to recognize face images of all races, or not only the face images but also the user's age, gender, or face value represented by the face images.

After acquiring the training sample images, after performing image processing on the training sample images (zooming or cutting the training sample images into pictures of fixed specifications), the training sample images are respectively input into the auxiliary training model and the fast model. So far, the auxiliary training model and the fast model separately extract the feature vectors of the training sample images.

S1300. Calculate the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the rapid model;

Obtain the feature vectors of the training sample images extracted by the auxiliary training model and the fast model respectively. Then calculate the feature distance between the two model feature vectors. The feature distance includes the Euclidean distance and / or cosine distance between the two feature vectors. The calculation of the characteristic distance is performed through the loss function. In some embodiments, the loss function can not only calculate the Euclidean distance between the two model feature vectors, but also calculate the cosine distance between the two model feature vectors, and through the combination of Euclidean distance and cosine distance from Dimension, correct the weight of the fast model, accelerate the training of the fast model, and improve the accuracy of the fast model.

Since the auxiliary training model has been trained to a converged state in advance, the feature vectors of the training sample images extracted by the auxiliary training model are accurate feature vectors, and during the training process, the feature vectors output by the fast model are less accurate. Calculating the distance between the two model feature vectors is actually calculating the distance between the feature vectors output by the fast model and the standard feature vectors.

S1400. Correct the weight parameters in the rapid model by backpropagating the feature distance.

After calculating the feature distance of the feature vectors of the same training sample image extracted by the two models, the feature vector is back-propagated through the loss function to correct the weight of the image filter (convolutional layer) in the fast model. As a result, the extracted feature vectors of the fast model are brought closer to the extracted feature vectors of the auxiliary training model. So far, for the completion of a fast model training, it should be pointed out that steps S1100-S1200 are only a step in the rapid model training process. Due to the number of training sample images, the training of steps S1100-S1200 is continuously looped until The fast model is trained until it converges.

The above embodiment does not need to mark the sample images participating in the training when training the fast model, which saves the time and effort required for the marking, and improves the training speed. At the same time, directly calculate the distance between the feature vector output from the auxiliary model that characterizes the sample image and the feature vector output from the fast model that characterizes the sample image (Euclidean distance and / or cosine distance) and perform back propagation. The method transforms the training of the fast model into a simple regression algorithm, which can shorten the training time to the maximum extent, and can guarantee the accuracy rate of the fast model output when the training is completed.

In some embodiments, the fast model training process does not need to be backpropagated every time. It is only necessary to perform backpropagation when the feature distance between the two models is greater than the set threshold. Please refer to FIG. 2, which is a schematic flowchart of determining whether to perform back propagation through a threshold in this embodiment.

As shown in Figure 2, before S1400, it also includes:

S1311: Compare the feature distance with a preset first threshold;

In order to verify whether the distance between the feature vectors extracted by the two models meets the needs of model training, a test threshold is set, that is, the first threshold. The setting of the first threshold can be set according to the accuracy requirement for the fast model. That is, the higher the accuracy requirement for the fast model, the smaller the set value of the first threshold, and conversely, the lower the accuracy requirement for the fast model, the larger the set value of the first threshold.

S1312. When the feature distance is greater than the first threshold, confirm to backpropagate the feature distance.

By comparing and determining that the feature distance is greater than the first threshold, it is determined that the distance between the feature vector extracted by the fast model and the feature vector extracted by the auxiliary training model is large, which does not meet the needs of rapid model training, and the fast propagation needs to be adjusted by back propagation The weight of the model, so that when the training sample image is input again, the distance between the feature vector extracted by the fast model and the feature vector extracted by the auxiliary training model tends to become smaller. When it is determined by comparison that the feature distance is less than or equal to the first threshold, it proves that the fast model's image understanding ability of the currently input training sample image meets the set requirements without back propagation, but the next training sample image is used to quickly The model is trained.

In this embodiment, it should be noted that steps S1100-S1200 and steps S1311-S1312 are performed cyclically. Also included after S1400

S1411: Iteratively and iteratively input the training sample image into the auxiliary training model and the rapid model, and when the feature distance is less than or equal to the first threshold, confirm that the training of the training sample image ends.

When it is determined by comparison that the feature distance is greater than the first threshold, the auxiliary training model and the fast model are input again to the same training sample image, and the steps of S1200-S1400 are repeatedly performed until when the feature distance is determined to be less than or equal to the first by comparison At the threshold, the training for the training sample image ends, and the training for other training sample images continues.

When the distance between the feature vector extracted by the fast model and the feature vector extracted by the auxiliary training model is less than the first threshold, or the proportion less than the first threshold is greater than the set value, such as 99% or other ratio set value At this time, the training for the fast samples ends.

In some embodiments, the feature distance includes the Euclidean distance and / or cosine distance, and calculating the feature distance between the extracted feature vectors of the two models is to calculate the Euclidean distance and / or cosine distance between the feature vectors. Please refer to FIG. 3, which is a schematic flowchart of calculating the feature distance between two model extraction feature vectors in this embodiment.

As shown in FIG. 3, S1300 also includes:

S1321: Acquire a first feature vector of the training sample image extracted by the auxiliary training model and a second feature vector of the training sample image extracted by the rapid model;

Obtain the feature vectors of the training sample images extracted by the auxiliary training model and the fast model respectively. Among them, the feature vector of the training sample image extracted by the auxiliary training model is the first feature vector, and the feature vector of the training sample image extracted by the fast model is the second feature vector.

S1322. Compare the Euclidean distance and / or the cosine distance of the first feature vector and the second feature vector.

The feature distance between the first feature vector and the second feature vector is calculated. The feature distance includes the Euclidean distance and / or cosine distance between the two feature vectors. The calculation of the characteristic distance is performed through the loss function. In some embodiments, the loss function can not only calculate the Euclidean distance between the two model feature vectors, but also calculate the cosine distance between the two model feature vectors, and through the combination of Euclidean distance and cosine distance from Dimension, correct the weight of the fast model, accelerate the training of the fast model, and improve the accuracy of the fast model.

In some embodiments, in order to improve the accuracy of the rapid model in a specific application environment, it is necessary to obtain targeted training sample images to train the rapid model. Please refer to FIG. 4, which is a schematic flowchart of obtaining training sample images according to specific application scenarios in this embodiment.

As shown in Figure 4, before S1100, it also includes:

S1011: Acquire image information of a target user or a target scene that needs to be identified, where the image information includes the image category of each identified image;

Before training the fast model, you first need to collect suitable materials, and in order to enhance the accuracy of the fast model in practical applications, you need to identify the specific users who use the fast model or the category pictures in a specific scene. For example, the user who uses the quick model is a senior pet lover. The user often uploads various pet pictures for species identification, and then collects the identification images uploaded by the user.

Each identified image corresponds to one piece of image information, and the image information includes the image type of each identified image. For example, when the recognition image is a pet photo, the image type is the species name of the pet.

S1012. Cluster the recognized images according to the image category in the image information to obtain the image type that the target user or target scene prefers to identify;

According to the acquired image category of each recognition image, the user image is clustered and counted. Clustering statistics is to classify the same type of image in the identified image into one type. Then, according to the number of recognized images in the same image type, the image type preferred for recognition in the user or application scene is determined. For example, the user uploads various animal images for identification, but because the user is an avid canine animal enthusiast, most of the uploaded identification images are images of the whole family. Therefore, the user who prefers to identify the images by statistical clustering The type is canine image. Or, if the user is a traveler, the identification images uploaded are all animal and plant images during the trip, and it is confirmed by statistics that the most recent images uploaded by the user are all animal and plant images of the African continent, then it is determined that the user is in Africa or African animals and plants are of great interest, and the image type of preference identification is determined to be African animals and plants images.

S1013: Collect training sample images of the training fast model according to the image type.

Collect training sample images for training fast models by image type. Training sample images can be crawled from the Internet using public image databases, own image databases, or through data crawling. For example, through the database of the African National Museum, obtain pictures of African animal and plant samples as training sample images, or use web crawlers to crawl pictures of pet dogs as training sample images.

Because when used in a special application scenario, the fast model does not need too high generalization ability, but the accuracy of image recognition for specific things is indeed very high. It is determined by identifying the application scenario or user preference. Collecting training sample images and training the fast model can make the fast model have higher accuracy in the targeted field and meet the needs of users.

In some embodiments, the training sample images are all stored in the image database. When a certain fast model needs to be trained, the training sample images need to be obtained through retrieval. Please refer to FIG. 5. FIG. 5 is a schematic flowchart of a training sample image acquired by the database inspection in this embodiment.

As shown in Figure 5, after S1013 includes:

S1021: Retrieve in a preset image database using the image type as a limiting condition;

After obtaining the image type preferred by the user or the application scene, the image type is used as a search condition to search in a preset image database. Among them, all the images in the image database are set according to the image content when they are put into the library. For example, when the image is an animal, the image label includes the animal's class, order, family, genus and name. .

Retrieval under the condition that the image type is not retrieved can retrieve images with the same or similar image tags as the image type.

S1022: Confirm the retrieved image as the training sample image.

The image recalled by retrieval is confirmed as the training sample image corresponding to the fast model.

In some embodiments, in order to make up for the shortcomings of the training sample image, and also to enhance the stability and anti-interference ability of the fast model, after acquiring the original sample image, the sample image is image processed to generate a derivative derived from the original sample image Sample image. Please refer to FIG. 6, which is a schematic flowchart of a method for generating a derived sample image according to this embodiment.

As shown in FIG. 6, after S1022, it also includes:

S1031: Perform image processing on the original sample image to generate a derived sample image derived from the original sample image;

The training sample image crawled through the database or through the web crawler technology is defined as the original sample image. A new sample image generated by performing image processing on the original sample image is defined as a derived sample image.

In this embodiment, the method for performing image processing on the original sample image includes (but is not limited to): combining the original sample image with one or more processing methods among image cropping, image rotation, or noise interference.

Since the actual content expressed by the original sample image and the derived sample image is completely the same, the feature vectors of the auxiliary training model to extract the original sample image and the derived sample image should be identical or similar. Therefore, it is necessary to further confirm whether the derived sample image after image processing is deformed to such an extent that the model cannot be recognized, and if so, the derived sample image cannot be used as a training sample.

S1032: Extract feature vectors of the original sample image and feature vectors of the derived sample image;

The original sample image and the derived sample image are successively input into the auxiliary training model to extract the feature vector of the original sample image and the feature vector of the derived sample image.

The model for extracting feature vectors is not limited to the auxiliary training model. In some embodiments, the model for extracting feature vectors can also be a neural network model that has been trained to a convergence state and has the same training direction as the fast model.

S1033: Calculate the feature difference between the feature vector of the original sample image and the feature vector of the derived sample image;

After calculating the feature vector of the original sample image and the feature vector of the derived sample image, the feature difference between the feature vector of the original sample image and the feature vector of the derived sample image is calculated. The feature difference is calculated by a loss function, and the feature difference can be the Euclidean distance and / or cosine distance of the two feature vectors.

S1034. When the feature difference value is less than or equal to a preset second threshold, confirm that the original sample image and the derived sample image are the training sample image.

The calculated feature difference value is compared with a preset second threshold, wherein, in order to verify whether the distance between the two feature vectors meets the needs of model training, a test threshold is set, that is, the second threshold. The setting of the second threshold can be set according to the interference requirement for the fast model. That is, the higher the anti-interference requirement for the fast model, the smaller the set value of the second threshold value, and conversely, the lower the anti-interference requirement for the fast model, the larger the set value of the second threshold value.

When the feature difference is less than or equal to the preset second threshold, it indicates that the derived sample image meets the training requirements, confirm that the original sample image and the derived sample image are the training sample image; otherwise, it is confirmed that the derived sample image is not Training requirements are discarded.

Since the derived sample image in the training sample image is obtained from the original sample image after image processing, the content of the image contained therein is substantially the same, and through the verification of the feature difference, it can ensure that the training sample image is trained to convergence quickly The model can accurately extract the correct feature vector even when the classified images have interference, which improves the anti-interference and stability of the model. At the same time, because the training sample image does not need to be calibrated during the training process of the fast model, by calculating the feature difference between the original sample image and the derived sample image, the fast model can accurately extract the feature vectors with interference images.

To solve the above technical problems, the embodiments of the present application also provide a rapid model training device.

Please refer to FIG. 7 for details. FIG. 7 is a schematic diagram of the basic structure of the training device of the rapid model in this embodiment.

As shown in FIG. 7, a rapid model training device includes: an acquisition module 2100, a processing module 2200, a calculation module 2300, and an execution module 2400. Among them, the obtaining module 2100 is used to obtain a preset training sample image; the processing module 2200 is used to input the training sample image into a preset auxiliary training model and an initial rapid model, wherein the auxiliary training model is pre-trained to a converged state A neural network model used to extract image feature vectors, the fast model is the neural network model to be trained; the calculation module 2300 is used to calculate the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the fast model Feature distance; the execution module 2400 is used to correct the weight parameters in the fast model by backpropagating the feature distance.

The training device of the fast model does not need to mark the sample images participating in the training when training the fast model, which saves the time and effort required for the marking, and improves the speed of training. At the same time, directly calculate the distance between the feature vector output from the auxiliary model that characterizes the sample image and the feature vector output from the fast model that characterizes the sample image (Euclidean distance and / or cosine distance) and perform back propagation. The method transforms the training of the fast model into a simple regression algorithm, which can shorten the training time to the maximum extent, and can guarantee the accuracy rate of the fast model output when the training is completed.

In some embodiments, the rapid model training device further includes: a first comparison submodule and a first execution submodule. The first comparison sub-module is used to compare the feature distance with a preset first threshold; the first execution sub-module is used to confirm back propagation of the feature distance when the feature distance is greater than the first threshold.

In some embodiments, the rapid model training device further includes: a second execution submodule for repeatedly iteratively inputting the training sample images into the auxiliary training model and the rapid model, and when the feature distance is less than or equal to the first threshold, the training is confirmed The training of the sample image ends.

In some embodiments, the feature distance includes the Euclidean distance and / or the cosine distance, and the rapid model training device further includes: a first acquisition submodule and a second comparison submodule. The first acquisition submodule is used to acquire the first feature vector of the training sample image extracted by the auxiliary training model and the second feature vector of the training sample image extracted by the fast model; the second comparison submodule is used to compare the first feature The Euclidean distance and / or cosine distance of the vector and the second feature vector.

In some embodiments, the rapid model training device further includes: a second acquisition submodule, a first processing submodule, and a third execution submodule. Among them, the second acquisition submodule is used to acquire image information of the target user or the target scene that needs to be identified, wherein the image information includes the image category of each identified image; the first processing submodule is used to match the image category in the image information Recognize the images and perform clustering to obtain the image type that the target user or target scene prefers to identify; the third execution submodule is used to collect training sample images for training the fast model according to the image type.

In some embodiments, the rapid model training device further includes: a second processing submodule and a first confirmation submodule. Wherein, the second processing sub-module is used for searching in a preset image database with the image type as the limiting condition; the first confirmation sub-module is used for confirming the image recalled by the retrieval as the training sample image.

In some embodiments, the rapid model training device further includes: a third processing submodule, a first extraction submodule, a first calculation submodule, and a fourth execution submodule. Among them, the third processing sub-module is used to perform image processing on the original sample image to generate a derived sample image derived from the original sample image; the first extraction sub-module is used to extract the feature vector of the original sample image and the feature vector of the derived sample image; The first calculation submodule is used to calculate the feature difference between the feature vector of the original sample image and the feature vector of the derived sample image; the fourth execution submodule is used to confirm when the feature difference is less than or equal to the preset second threshold The original sample image and the derived sample image are training sample images.

To solve the above technical problems, embodiments of the present application also provide computer equipment. For details, please refer to FIG. 8, which is a block diagram of the basic structure of the computer device of this embodiment.

As shown in FIG. 8, a schematic diagram of the internal structure of the computer device. The computer device includes a processor, a non-volatile storage medium, a memory, and a network interface connected through a system bus. The non-volatile storage medium of the computer device stores an operating system, a database, and computer-readable instructions. The database may store a sequence of control information. When the computer-readable instructions are executed by the processor, the processor may implement a A fast model training method. The processor of the computer device is used to provide calculation and control capabilities, and support the operation of the entire computer device. The memory of the computer device may store computer readable instructions. When the computer readable instructions are executed by the processor, the processor may cause the processor to execute a rapid model training method. The network interface of the computer device is used to connect and communicate with the terminal. Those skilled in the art may understand that the structure shown in FIG. 8 is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer equipment to which the solution of the present application is applied. The specific computer equipment may It includes more or fewer components than shown in the figure, or some components are combined, or have a different component arrangement.

In this embodiment, the processor is used to perform specific functions of the acquisition module 2100, the processing module 2200, the calculation module 2300, and the execution module 2400 in FIG. 7, and the memory stores program codes and various types of data required to execute the above modules. The network interface is used for data transmission between user terminals or servers. The memory in this embodiment stores the program codes and data required to execute all submodules in the face image key point detection device, and the server can call the server program codes and data to execute the functions of all submodules.

The computer equipment does not need to mark the sample images participating in the training when training the fast model, which saves the time and effort required for the marking and increases the training speed. At the same time, directly calculate the distance between the feature vector output from the auxiliary model that characterizes the sample image and the feature vector output from the fast model that characterizes the sample image (Euclidean distance and / or cosine distance) and perform back propagation. The method transforms the training of the fast model into a simple regression algorithm, which can shorten the training time to the maximum extent, and can guarantee the accuracy rate of the fast model output when the training is completed.

The present application also provides a storage medium storing computer-readable instructions. When the computer-readable instructions are executed by one or more processors, the one or more processors execute the steps of the rapid model training method in any of the foregoing embodiments .

A person of ordinary skill in the art may understand that all or part of the processes in the method of the above embodiments may be completed by instructing relevant hardware through a computer program. The computer program may be stored in a computer-readable storage medium. When executed, it may include the processes of the foregoing method embodiments. Wherein, the aforementioned storage medium may be a non-volatile storage medium such as a magnetic disk, an optical disc, a read-only memory (Read-Only Memory, ROM), or a random access memory (Random Access Memory, RAM), etc.

Claims

A fast model training method, including:

Obtain preset training sample images;

The training sample image is input into a preset auxiliary training model and an initial fast model, wherein the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast model is The neural network model to be trained;

Calculating the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the rapid model;

By backpropagating the feature distance, the weight parameters in the fast model are corrected.
The method for training a fast model according to claim 1, before the backpropagating the feature distance to correct the weight parameter in the fast model further includes:

Compare the feature distance with a preset first threshold;

When the feature distance is greater than the first threshold, it is confirmed that the feature distance is back propagated.
The method for training a fast model according to claim 2, after correcting the weight parameter in the fast model by backpropagating the feature distance, further comprising:

Iteratively and iteratively input the training sample image into the auxiliary training model and the rapid model, and when the feature distance is less than or equal to the first threshold, confirm that the training of the training sample image is completed.
The method for training a fast model according to claim 1, wherein the feature distance includes an Euclidean distance and / or a cosine distance, and calculating the feature vector of the training sample image extracted by the auxiliary training model and the extracted by the fast model The feature distance between feature vectors of the training sample image includes:

Acquiring a first feature vector of the training sample image extracted by the auxiliary training model and a second feature vector of the training sample image extracted by the fast model;

Compare the Euclidean distance and / or the cosine distance of the first feature vector and the second feature vector.
According to the rapid model training method of claim 1, before acquiring the preset training sample image further comprises:

Acquiring image information of a target user or a target scene that needs to be identified, wherein the image information includes the image category of each identified image;

Clustering the identified images according to the image category in the image information to obtain the image type that the target user or target scene prefers to identify;

Collect training sample images of the training fast model according to the image type.
The method for training a rapid model according to claim 5, the collecting the training sample images for training the rapid model according to the image type includes:

Search in a preset image database using the image type as a limiting condition;

The image recalled by the retrieval is confirmed as the training sample image.
The rapid model training method according to claim 6, wherein the training sample image includes an original sample image and a derived sample image, and after confirming the retrieved image as the training sample image, the method further includes:

Performing image processing on the original sample image to generate a derived sample image derived from the original sample image;

Extract the feature vector of the original sample image and the feature vector of the derived sample image;

Calculating the feature difference between the feature vector of the original sample image and the feature vector of the derived sample image;

When the feature difference value is less than or equal to a preset second threshold, it is confirmed that the original sample image and the derived sample image are the training sample image.
A rapid model training device, including:

The acquisition module is used to acquire preset training sample images;

A processing module, configured to input the training sample image into a preset auxiliary training model and an initial fast model, wherein the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, The fast model is a neural network model to be trained;

A calculation module, configured to calculate the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the rapid model;

An execution module is used to correct the weight parameters in the fast model by backpropagating the feature distance.
A computer device includes a memory and a processor, and the memory stores computer-readable instructions. When the computer-readable instructions are executed by the processor, the processor executes a fast model training method, The method includes the following steps:

Obtain preset training sample images;

The training sample image is input into a preset auxiliary training model and an initial fast model, wherein the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast model is The neural network model to be trained;

Calculating the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the rapid model;

By backpropagating the feature distance, the weight parameters in the fast model are corrected.
The computer device according to claim 9, before the backpropagating the feature distance to correct the weight parameter in the fast model further comprises:

Compare the feature distance with a preset first threshold;

When the feature distance is greater than the first threshold, it is confirmed that the feature distance is back propagated.
The computer device according to claim 10, wherein after backpropagating the feature distance to correct the weight parameter in the fast model, the method further comprises:

Iteratively and iteratively input the training sample image into the auxiliary training model and the rapid model, and when the feature distance is less than or equal to the first threshold, confirm that the training of the training sample image is completed.
The computer device according to claim 9, wherein the feature distance includes an Euclidean distance and / or a cosine distance, and calculating the feature vector of the training sample image extracted by the auxiliary training model and the training sample image extracted by the rapid model The feature distance between the feature vectors of includes:

Acquiring a first feature vector of the training sample image extracted by the auxiliary training model and a second feature vector of the training sample image extracted by the fast model;

Compare the Euclidean distance and / or the cosine distance of the first feature vector and the second feature vector.
The computer device according to claim 9, before the acquiring the preset training sample image further comprises:

Acquiring image information of a target user or a target scene that needs to be identified, wherein the image information includes the image category of each identified image;

Clustering the identified images according to the image category in the image information to obtain the image type that the target user or target scene prefers to identify;

Collect training sample images of the training fast model according to the image type.
The computer device according to claim 13, the collecting training sample images of the training fast model according to the image type comprises:

Search in a preset image database using the image type as a limiting condition;

The image recalled by the retrieval is confirmed as the training sample image.
The computer device according to claim 14, wherein the training sample image includes an original sample image and a derived sample image, and after confirming the retrieved image as the training sample image, further includes:

Performing image processing on the original sample image to generate a derived sample image derived from the original sample image;

Extract the feature vector of the original sample image and the feature vector of the derived sample image;

Calculating the feature difference between the feature vector of the original sample image and the feature vector of the derived sample image;

When the feature difference value is less than or equal to a preset second threshold, it is confirmed that the original sample image and the derived sample image are the training sample images.
A non-volatile storage medium storing computer-readable instructions, which when executed by one or more processors, causes the one or more processors to execute a rapid model training method, The method includes the following steps:

Obtain preset training sample images;

The training sample image is input into a preset auxiliary training model and an initial fast model, wherein the auxiliary training model is a neural network model that is pre-trained to a convergence state for extracting image feature vectors, and the fast model is The neural network model to be trained;

Calculating the feature distance between the feature vector of the training sample image extracted by the auxiliary training model and the feature vector of the training sample image extracted by the rapid model;

By backpropagating the feature distance, the weight parameters in the fast model are corrected.
The non-volatile storage medium according to claim 16, before the backpropagating the characteristic distance to correct the weight parameter in the fast model further includes:

Compare the feature distance with a preset first threshold;

When the feature distance is greater than the first threshold, it is confirmed that the feature distance is back propagated.
The non-volatile storage medium according to claim 17, after the back propagation of the characteristic distance to correct the weight parameter in the fast model, further comprising:

Iteratively and iteratively input the training sample image into the auxiliary training model and the rapid model, and when the feature distance is less than or equal to the first threshold, confirm that the training of the training sample image is completed.
The non-volatile storage medium according to claim 16, wherein the feature distance includes an Euclidean distance and / or a cosine distance, and calculating the feature vector of the training sample image extracted by the auxiliary training model and the fast model extraction The feature distances between the feature vectors of the training sample images include:

Acquiring a first feature vector of the training sample image extracted by the auxiliary training model and a second feature vector of the training sample image extracted by the fast model;

Compare the Euclidean distance and / or the cosine distance of the first feature vector and the second feature vector.
The non-volatile storage medium according to claim 16, before the acquiring the preset training sample image further comprises:

Acquiring image information of a target user or a target scene that needs to be identified, wherein the image information includes the image category of each identified image;

Clustering the identified images according to the image category in the image information to obtain the image type that the target user or target scene prefers to identify;

Collect training sample images of the training fast model according to the image type.