A model startup method and device

Abstract

The present invention discloses a model startup method and device, the method comprising: obtaining original data for model startup; obtaining the number of enhanced data according to the original data; obtaining the enhancement strategy corresponding to the original data according to the number of enhanced data, the enhancement strategy comprising a data enhancement sub-strategy and a feature enhancement sub-strategy; based on the original data, enhancing according to the enhancement strategy to obtain incremental data; screening the incremental data to obtain target incremental data; combining the target incremental data and the original data to obtain startup data, and starting the model with the startup data. The present invention performs data enhancement processing in the processes of original data processing and feature data processing, and screens the processing results, thereby improving the accuracy of the processing results, solving the cold start problem, and reducing the data volume demand and time cost in the early stage of the model being put into use.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a model startup method and device. Background Art

[0002] Model startup refers to the process of building a model from scratch, putting it into use, and finally entering online iterative updates without obtaining sufficient data. For a multi-object (user) classification / inference problem, existing cold start technologies focus more on how to obtain additional information to assist in training the model in a way that is different from the normal model process. A common method is to introduce additional information and additional logic for assistance. For example, in the absence of user behavior data, the recommendation system will make inferences based on some of the user's inherent registration information, such as age and gender; for example, when the user uses it for the first time, the user is asked to make some additional choices to obtain strong information for verification or screening.

[0003] It can be seen that the cold start technology in the prior art has at least one of the following defects:

[0004] Use additional analysis logic: This requires the use of additional information beyond the normal model data flow, which will have additional requirements for the data flow of the business scenario;

[0005] Designing additional user interaction flows: This will introduce new variations in model performance and incur additional costs and the problems and risks they may cause;

[0006] Lack of utilization of small amounts of startup data: The room for improving model performance in the original data during the data collection process is ignored, i.e., there is no ability to perform online iterations in the early stages of the model. Summary of the invention

[0007] The invention provides a model starting method and device.

[0008] In one aspect, the present invention provides a model startup method, the method comprising:

[0009] Get the raw data for model initiation;

[0010] Acquire the amount of enhanced data according to the original data;

[0011] Acquire an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy;

[0012] Based on the original data, enhancement is performed according to the enhancement strategy to obtain incremental data;

[0013] Screen the incremental data to obtain target incremental data;

[0014] The target incremental data and the original data are combined to obtain startup data, and the model is started with the startup data.

[0015] Another aspect provides a model starting device, the device comprising:

[0016] A raw data acquisition module, used to acquire raw data for model startup;

[0017] An enhanced data acquisition module, used to acquire the enhanced data quantity according to the original data;

[0018] A strategy acquisition module, used to acquire an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy;

[0019] An enhancement module, configured to enhance the original data according to the enhancement strategy to obtain incremental data;

[0020] A screening module is used to screen the incremental data to obtain target incremental data;

[0021] A combining module is used to combine the target incremental data and the original data to obtain startup data, and to start the model with the startup data.

[0022] Another aspect provides a computer-readable storage medium for storing a program, wherein the program implements a model startup method when executed.

[0023] The startup method and device provided by the present invention can integrate multiple enhancement methods, perform data enhancement processing in the processes of original data processing and feature data processing, and screen the processing results, thereby improving the accuracy of the processing results. This solution can significantly improve the performance of the model on a small amount of data while slightly improving the overall effect of the model, thereby reducing the data volume demand and time cost in the early stage of the model being put into use, and to a certain extent solves the problem of model cold start. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0025] Figure 1 It is a flow chart of a model startup method provided by the present invention;

[0026] Figure 2This is a comparison chart of the recognition recall rate effect provided by the present invention;

[0027] Figure 3 is a schematic diagram of picture data enhancement provided by the present invention;

[0028] Figure 4 It is a flow chart of a feature enhancement method provided by the present invention;

[0029] Figure 5 is another flow chart of a feature enhancement method provided by the present invention;

[0030] Figure 6 is another flow chart of a feature enhancement method provided by the present invention;

[0031] Figure 7 It is a flow chart of a method for screening incremental data to obtain target incremental data provided by the present invention;

[0032] Figure 8 It is a block diagram of a model starting device provided by the present invention;

[0033] Fig. 9 is a block diagram of an enhancement module provided by the present invention;

[0034] Fig.10 is a block diagram of a screening module provided by the present invention;

[0035] Fig.11 It is a schematic diagram of the hardware structure of a device provided by the present invention for implementing the method provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0037] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or devices.

[0038] Cold start solutions in the prior art can be roughly divided into three types.

[0039] The first type is to assist the model in making analysis and decisions by obtaining other dimensional information of the object (i.e., non-model input). The other dimensional information can also be called additional information. For example, the cold start strategy of a music recommendation model. By allowing new users to perform some operations with strong identification to perform the initial category division and requiring users to select the song type once, a large amount of data collection and analysis costs can be reduced for the subsequent work of the recommendation model. After the user has completed a series of simple but carefully designed input logic, the recommendation model can already establish a rough user portrait for the user, so that it can be continuously iterated and updated based on the user data generated later.

[0040] The second type is to use the Bandit algorithm and its related derivative methods in various fields. The Bandit algorithm uses categories or topics to represent the interests of each user, and through several experiments, it characterizes the probability of interest in each topic in the minds of new users. If the user is interested in a topic (providing explicit feedback or implicit feedback), it means that the user has gained benefits. If the user is not interested in a topic, no benefits are gained. In this way, the cycle of "select-observe-update-select" is experienced to approach the topics that the user is really interested in.

[0041] The third type is to design a corresponding startup mechanism based on the relevant model of information entropy or specific domain information, and perform a round of sniffing on the user's behavior or attributes that is different from the model logic to obtain key information.

[0042] The above three types require obtaining additional information, and the introduction of additional information will increase the design complexity of the model, reduce the space for improving the model effect by the original data, and the introduction of new changes may also induce other risks.

[0043] In view of this, the embodiment of the present invention discloses a model startup method, such as Figure 1As shown, the method includes:

[0044] S101. Obtaining original data for model startup.

[0045] S102. Obtain the amount of enhanced data according to the original data.

[0046] In the embodiment of the present invention, the cold start problem is alleviated by starting by enhancing the original data, and the data enhancement is entirely based on the characteristics of the original data, thereby avoiding the risk of introducing additional information. However, when the amount of original data is small, excessive enhancement may make it difficult to ensure the quality of the comprehensive data obtained after enhancement. Therefore, in the embodiment of the present invention, when the amount of original data is at different stages, the amount of enhanced data after enhancement is first controlled, thereby reducing the deviation between the distribution of comprehensive data and real data after the original data and enhanced data are merged to a certain extent.

[0047] In a feasible embodiment of the present invention, the data volume of the enhanced data has a positive correlation with the data volume of the original data.

[0048] S103 acquires an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy.

[0049] Specifically, the enhancement strategy can be set according to the characteristics, format, applicable scenario, and activated model profile of the original data, and the embodiment of the present invention does not limit the specific acquisition method. Of course, the enhancement strategy can also be triggered in response to the user's selection.

[0050] Specifically, the data enhancement sub-strategy may include whether to perform data enhancement, the method of data increment, and the number of incremental data obtained by each data enhancement method; the feature enhancement sub-strategy may include whether to perform feature enhancement, the method of feature enhancement, and the number of incremental data obtained by each feature enhancement method.

[0051] In the embodiment of the present invention, data enhancement refers to enhancing the data before feature extraction, and may also be enhancing the original data; feature enhancement refers to enhancing the data after feature processing.

[0052] S104. Based on the original data, enhancement is performed according to the enhancement strategy to obtain incremental data.

[0053] In the embodiment of the present invention, the enhancement strategy can be executed to generate more equivalent / similar data based on limited original data, thereby reducing model overfitting and playing a role in supplementing data.

[0054] During the enhancement process, more than one data enhancement method and more than one feature enhancement method may be used in combination, and the order of using each selected enhancement method and the number of incremental data obtained may be set in the enhancement strategy.

[0055] In the process of enhancement, the degree of coupling between various enhancement methods is low, the calling method is flexible, and the degree of freedom is high. Compared with the enhancement method for a specific format / field, the applicable scenarios of data enhancement are expanded to a certain extent. The embodiment of the present invention performs enhancement completely based on the original data and does not rely on additional information, thereby avoiding additional user interaction.

[0056] S105. Screen the incremental data to obtain target incremental data.

[0057] Since the newly generated incremental data may deviate from the distribution of the original real data during the execution of the enhancement strategy in the embodiment of the present invention, it may cause the change of the overall data distribution, affecting the effect and stability of the model. On the other hand, the incremental data is more sufficient in quantity than the original data, so the embodiment of the present invention reduces the negative impact of the enhancement step as much as possible by adding a round of processing mechanisms that can be controlled by rules and standards, and the added processing mechanisms that can be controlled by rules and standards are executed by the data screening step.

[0058] S106. Combine the target incremental data and the original data to obtain startup data, and start the model with the startup data.

[0059] In most cases, data is the most important factor that determines the machine learning model. In specific applications, especially in the early stages of modeling, insufficient data is an extremely common problem. Reducing the data volume required for the model can often bring huge benefits. A model startup method provided in an embodiment of the present invention can solve the cold start problem caused by insufficient data, reduce the model's demand for data and the cost of acquiring raw data, and does not require the introduction of additional information, thereby avoiding the potential risks caused by the introduction of additional information. It can be widely used in various recommendation systems and scenarios that rely on machine learning. Users can call and fine-tune enhancement strategies based on data and model characteristics. Taking the owner identification scenario as an example, in the owner identification scenario, the enhancement strategy can be opened and closed and adjusted through flags and hyperparameter settings in the process, and its recognition recall rate is as follows: Figure 2 As shown, Figure 2The two polylines marked as 1 and 2 are the recall rate curves corresponding to the start-up method in the embodiment of the present invention, and the polyline marked as 3 is the recall rate curve using the original data. It can be seen that after using the start-up method provided by the embodiment of the present invention, the recall rate of the model is significantly improved, and when the amount of original data is small, the start-up method provided by the embodiment of the present invention is more obvious for improving the recall rate.

[0060] It can be seen that the startup method provided by the embodiment of the present invention has at least the following beneficial effects:

[0061] (1) Reduce the amount of data collection required before the model is put into use, thereby improving the model experience when the data volume is not saturated, allowing the model to be put into online use more quickly;

[0062] (2) The overall performance of the model is improved to varying degrees in all stages;

[0063] (3) It provides more flexible control and usage methods, which can be adjusted and deployed as needed.

[0064] In a feasible embodiment of the present invention, a data enhancement method is provided. The data enhancement method may be a noise adding method, specifically, mixing Gaussian noise or other forms of noise with original data.

[0065] In another feasible embodiment of the present invention, in an image processing scenario, a data enhancement method is also provided, which can perform data enhancement based on image continuity characteristics such as rotation, flipping, translation, and cropping on the original data. Figure 3 As shown, it shows a schematic diagram of an original picture becoming multiple pictures after data enhancement, thereby obtaining a picture set.

[0066] In another feasible embodiment of the present invention, data enhancement can also be performed based on an incompletely trained GAN network. Generative Adversarial Networks (GAN) is a deep learning model and one of the most promising methods for unsupervised learning on complex distributions in recent years. The model produces a fairly good output through the mutual game learning of (at least) two modules in the framework: the generative model and the discriminative model. Specifically, other data with original sample category information can be generated, and a first label is set for the other data. An incompletely trained GAN network is used, and other data is generated with the original sample as input, and the other data is set as a second label. The data with the first label and the data with the second label are used as incremental data to start the model.

[0067] The embodiment of the present invention may provide more than one feature enhancement method, and the selection of the feature enhancement method may be set in a feature enhancement sub-strategy.

[0068] In a feasible implementation, a feature enhancement method is provided, such as Figure 4 Said method comprises:

[0069] S1. Obtain the data to be enhanced.

[0070] Specifically, the data to be enhanced is input data for feature enhancement, which can be data directly subjected to feature processing of original data, or data subjected to other enhancement processing and feature processing.

[0071] In a feasible implementation manner, the format of each piece of data to be enhanced can be represented by an n*1 vector.

[0072] S2. Randomly extract N pieces of data with replacement from the data to be enhanced to obtain data to be mixed.

[0073] Specifically, if the data to be enhanced includes multiple categories, the extraction sources of the data to be mixed should be of the same category.

[0074] S3. Interpolate the data to be mixed according to random weights to obtain incremental data.

[0075] Taking N=2 as an example, the target incremental data=kx+(1-k)y, where x and y are the data to be mixed, and k is the random weight. Specifically, the random weight can be obtained using a Gaussian distribution or a symmetric Beta distribution. Among them, Beta Distribution is a density function that is a conjugate prior distribution of Bernoulli distribution and binomial distribution, and has important applications in machine learning and mathematical statistics.

[0076] The incremental data is the new data obtained by feature enhancement. By looping steps S2-S3, multiple incremental data can be obtained, thereby achieving the purpose of expanding the data volume.

[0077] In another feasible embodiment, a feature enhancement method is provided, such as Figure 5 Said method comprises:

[0078] S10. Obtain data to be enhanced.

[0079] Specifically, the data to be enhanced is input data for feature enhancement, which can be data directly subjected to feature processing of original data, or data subjected to other enhancement processing and feature processing.

[0080] In a feasible implementation manner, the format of each piece of data to be enhanced can be represented by an n*1 vector.

[0081] S20. Randomly extract 3 pieces of data with replacement from the data to be enhanced to obtain data to be mixed.

[0082] Specifically, if the data to be enhanced includes multiple categories, the extraction sources of the data to be mixed should be of the same category.

[0083] S30. Randomly select one of the data to be mixed as a reference, and the other two as variables, and obtain incremental data according to the formula z=x1+x2-x0, where x0 is the reference, and x1 and x2 are variables.

[0084] The incremental data is the new data obtained by feature enhancement. By looping steps S2-S3, multiple incremental data can be obtained, thereby achieving the purpose of expanding the data volume.

[0085] In another feasible embodiment, a feature enhancement method is provided, such as Figure 6 Said method comprises:

[0086] S100. Obtain data to be enhanced.

[0087] Specifically, the data to be enhanced is input data for feature enhancement, which can be data directly subjected to feature processing of original data, or data subjected to other enhancement processing and feature processing.

[0088] In a feasible implementation manner, the format of each piece of data to be enhanced can be represented by an n*1 vector.

[0089] S200. Obtain statistical features of the data to be enhanced.

[0090] Specifically, the mean and / or standard deviation of the data to be enhanced may be calculated.

[0091] S300. Obtain noise distribution according to the statistical characteristics.

[0092] Specifically, the noise distribution can be a random distribution, the mean of the random distribution conforms to the statistical characteristics of the data to be enhanced; the noise can also be a distribution with a mean of 0, and the standard deviation of the distribution is obtained by the standard deviation of the coefficient to be enhanced, specifically, the standard deviation of the coefficient to be enhanced multiplied by a scaling factor, and the scaling factor can be set according to actual needs.

[0093] S400. Randomly select noise according to the noise distribution.

[0094] S500. Obtain incremental data by accumulating the data to be enhanced and the noise.

[0095] The incremental data is the new data obtained by feature enhancement. By looping steps S400-S500, a plurality of incremental data can be obtained, thereby achieving the purpose of expanding the data volume.

[0096] The embodiments of the present invention disclose a variety of data enhancement methods and feature enhancement methods, each of which can be used independently and decoupled from each other. The mixed use of the methods can significantly improve the enhancement effect, thereby obtaining a large amount of incremental data. However, each enhancement method has advantages and disadvantages. For example, the method of introducing random noise may cause the fluctuation of the model effect due to the introduction of hyperparameters of the noise distribution, some image enhancement methods have limited applicable scenarios, and the enhancement effect of the enhancement method based on the adversarial network is limited, etc. Therefore, the enhancement strategy should be flexibly designed in combination with the actual scenario to achieve the maximum data enhancement effect.

[0097] In order to improve the accuracy of incremental data, the embodiment of the present invention further discloses a method for filtering incremental data to obtain target incremental data, such as Figure 7 As shown, the method includes:

[0098] S1051. Obtain a first probability distribution and a first digital feature based on all original data.

[0099] S1052. Obtain the screening ratio coefficient.

[0100] The screening scale factor is an adjustable parameter. The numerical features used in the screening scale factor can be adjusted according to the distribution of the original data, such as using the mean or median.

[0101] S1053. Obtain current data, where the current data includes all original data, incremental data determined to be retained, and current incremental data.

[0102] S1054. Obtain target data whose current data falls into a preset interval in the first probability distribution.

[0103] Specifically, the preset interval can be set according to actual needs. In a feasible implementation, the preset interval can be an interval outside 3σ, that is, the target data is data that does not meet the 3σ criterion. The 3σ criterion is also called the Laida criterion. It first assumes that a set of test data contains only random errors, calculates and processes them to obtain the standard deviation, and determines an interval according to a certain probability. It is believed that any error exceeding this interval is not a random error but a gross error, and the data containing this error should be eliminated.

[0104] S1055. Obtain a second digital feature of the target data.

[0105] S1056. Determine whether the current incremental data is determined to be retained based on the second digital feature, the proportional coefficient and the first digital feature, and perform corresponding operations based on the determination result.

[0106] The first digital feature and the second digital feature are obtained using the same method, and may be a certain feature value that identifies a probability distribution, which is not specifically limited in the embodiment of the present invention.

[0107] In a feasible implementation, if the product of the second digital feature and the screening ratio coefficient is smaller than the first digital feature, the current incremental data is retained, otherwise, it is discarded.

[0108] S1057. Determine whether there is the next incremental data.

[0109] S1058. If yes, take the next incremental data as the current incremental data and execute step S1053.

[0110] The startup method provided in the embodiment of the present invention can integrate multiple enhancement methods, perform data enhancement processing in the processes of original data processing and feature data processing, and screen the processing results, thereby improving the accuracy of the processing results. This solution significantly improves the performance of the model on a small amount of data on the basis of slightly improving the overall effect of the model, thereby reducing the data volume demand and time cost in the early stage of the model being put into use, and to a certain extent solves the model startup problem encountered in the project.

[0111] The embodiment of the present invention also provides a model starting device, such as Figure 8 As shown, the device comprises:

[0112] The original data acquisition module 201 is used to acquire the original data used to start the model;

[0113] An enhanced data quantity acquisition module 202 is used to acquire the enhanced data quantity according to the original data;

[0114] A strategy acquisition module 203 is used to acquire an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy;

[0115] An enhancement module 204 is used to enhance the original data according to the enhancement strategy to obtain incremental data;

[0116] A screening module 205 is used to screen the incremental data to obtain target incremental data;

[0117] The combining module 206 is used to combine the target incremental data and the original data to obtain the starting data, and start the model with the starting data.

[0118] Furthermore, if Fig. 9 As shown, the enhancement module 204 includes:

[0119] The first feature enhancement unit 2041 is used to obtain data to be enhanced; randomly extract N pieces of data with replacement from the data to be enhanced to obtain data to be mixed; and interpolate the data to be mixed according to random weights to obtain incremental data;

[0120] The second feature enhancement unit 2042 is used to obtain data to be enhanced; randomly extract 3 data with replacement from the data to be enhanced to obtain data to be mixed; randomly select one data from the data to be mixed as a reference and the other two data as variables, and obtain incremental data according to the formula z=x1+x2-x0, where x0 is the reference and x1 and x2 are variables;

[0121] The third feature enhancement unit 2043 is used to obtain the data to be enhanced; obtain the statistical features of the data to be enhanced; obtain the noise distribution according to the statistical features; randomly select noise according to the noise distribution; and obtain incremental data by accumulating the data to be enhanced and the noise.

[0122] like Fig.10 As shown, the screening module 205 includes:

[0123] A first feature acquisition unit 2051 is used to acquire a first probability distribution and a first digital feature according to all original data;

[0124] A screening ratio coefficient obtaining unit 2052 is used to obtain a screening ratio coefficient;

[0125] The current data acquisition unit 2053 is used to acquire the current data, where the current data includes all the original data, the incremental data determined to be retained, and the current incremental data;

[0126] A target data acquisition unit 2054 is used to acquire target data whose current data falls into a preset interval in the first probability distribution;

[0127] A second feature acquisition unit 2055, used to acquire a second digital feature of the target data;

[0128] The screening unit 2056 is used to determine whether the current incremental data is determined to be retained according to the second digital feature, the proportional coefficient and the first digital feature, and perform corresponding operations according to the determination result.

[0129] A model starting device and method embodiment disclosed in the embodiment of the present invention are both based on the same inventive concept.

[0130] An embodiment of the present invention also provides a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded by a processor and executing the steps of a model startup method described in an embodiment of the present invention. The specific execution process can be described in detail in the method embodiment, and will not be repeated here.

[0131] Further, Fig.11 A schematic diagram of the hardware structure of a device for implementing the method provided by an embodiment of the present invention is shown. The device may be a computer terminal, a mobile terminal or a server. The device may also participate in constituting the apparatus or recommendation system provided by an embodiment of the present invention. Fig.11 As shown, the computer terminal 10 (or mobile device 10 or server 10) may include one or more (102a, 102b, ..., 102n are used to illustrate) processors 102 (the processor 102 may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission device 106 for communication functions. In addition, it may also include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, a power supply and / or a camera. It can be understood by those skilled in the art that Fig.11 The structure shown is only for illustration and does not limit the structure of the above electronic device. Fig.11 More or fewer components as shown, or with Fig.11 Different configurations are shown.

[0132] It should be noted that the one or more processors 102 and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10 (or mobile device). As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0133] The memory 104 can be used to store software programs and modules of application software, such as program instructions / data storage devices corresponding to the method in the embodiment of the present invention. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, to implement the above-mentioned model startup method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, the memory 104 may further include a memory remotely arranged relative to the processor 102, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0134] The transmission device 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission device 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission device 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0135] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 (or mobile device).

[0136] It should be noted that the sequence of the embodiments of the present invention described above is for description only and does not represent the advantages and disadvantages of the embodiments. The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0137] Each embodiment in this specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the device and server embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiments.

[0138] A person skilled in the art will understand that all or part of the steps to implement the above embodiments may be accomplished by hardware or by instructing related hardware through a program, and the program may be stored in a computer-readable storage medium, and the above-mentioned storage medium may be a read-only memory, a disk or an optical disk, etc.

[0139] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A model startup method, characterized in that: The method comprises: Acquire original data for starting the model, wherein the original data is image data; Acquire the amount of enhanced data according to the original data; Acquire an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy, wherein the data enhancement sub-strategy includes an enhancement method for the original data, and the feature enhancement sub-strategy includes an enhancement method for the feature-processed data; Based on the original data, enhancement is performed according to the enhancement strategy to obtain incremental data; Obtain a first probability distribution and a first digital feature according to all the original data; Get the screening ratio coefficient; Acquire current data, where the current data includes all original data, incremental data determined to be retained, and current incremental data; Acquire target data where the current data falls within a preset interval in the first probability distribution; Acquire a second digital feature of the target data; Determining whether the current incremental data is determined to be retained according to the second digital feature, the proportional coefficient and the first digital feature, and performing corresponding operations according to the determination result to obtain target incremental data; The target incremental data and the original data are combined to obtain startup data, and the model is started with the startup data.

2. The method according to claim 1, characterized in that: The step of performing enhancement according to the enhancement strategy to obtain incremental data includes performing enhancement using a feature enhancement method, and the step of performing enhancement using the feature enhancement method includes: Obtain data to be enhanced; Randomly extracting N pieces of data with replacement from the data to be enhanced to obtain data to be mixed; The data to be mixed are interpolated according to random weights to obtain incremental data.

3. The method according to claim 1, characterized in that: The step of performing enhancement according to the enhancement strategy to obtain incremental data includes performing enhancement using a feature enhancement method, and the step of performing enhancement using the feature enhancement method includes: Obtain data to be enhanced; Randomly extracting 3 pieces of data with replacement from the data to be enhanced to obtain data to be mixed; One piece of data to be mixed is randomly selected as a reference, and the other two pieces are selected as variables, and incremental data is obtained according to the formula z=x1+x2-x0, wherein x0 is the reference, and x1 and x2 are variables.

4. The method according to claim 1, characterized in that: The step of performing enhancement according to the enhancement strategy to obtain incremental data includes performing enhancement using a feature enhancement method, and the step of performing enhancement using the feature enhancement method includes: Obtain data to be enhanced; Acquire statistical features of the data to be enhanced; Acquire noise distribution according to the statistical characteristics; randomly selecting noise according to the noise distribution; The incremental data is obtained by accumulating the data to be enhanced and the noise.

5. The method according to claim 1, characterized in that Also includes: Determine whether there is the next incremental data; The next piece of incremental data is used as the current incremental data, and the execution returns to the step of obtaining the current data, wherein the current data includes all the original data, the incremental data determined to be retained, and the current incremental data.

6. The method according to claim 1, characterized in that Also includes: The first digital feature and the second digital feature are obtained using the same method; The first digital feature and the second digital feature are adjusted according to the distribution of the original data.

7. A model starting device, characterized in that: The device comprises: The original data acquisition module is used to acquire the original data used to start the model, wherein the original data is image data; An enhanced data acquisition module, used to acquire the enhanced data quantity according to the original data; A strategy acquisition module, used for acquiring an enhancement strategy corresponding to the original data according to the amount of enhanced data, wherein the enhancement strategy includes a data enhancement sub-strategy and a feature enhancement sub-strategy, wherein the data enhancement sub-strategy includes an enhancement method for the original data, and the feature enhancement sub-strategy includes an enhancement method for the feature-processed data; An enhancement module, configured to enhance the original data according to the enhancement strategy to obtain incremental data; A screening module is used to screen the incremental data to obtain target incremental data; A combining module, used for combining the target incremental data and the original data to obtain startup data, and starting the model with the startup data; The screening module comprises: A first feature acquisition unit, used for acquiring a first probability distribution and a first digital feature according to all original data; A screening ratio coefficient obtaining unit, used to obtain the screening ratio coefficient; A current data acquisition unit, used to acquire current data, wherein the current data includes all original data, incremental data determined to be retained, and current incremental data; A target data acquisition unit, used to acquire target data where the current data falls within a preset interval in the first probability distribution; A second feature acquisition unit, used to acquire a second digital feature of the target data; A screening unit is used to determine whether the current incremental data is determined to be retained according to the second digital feature, the proportional coefficient and the first digital feature, and perform corresponding operations according to the determination result.

8. The device according to claim 7, characterized in that: The enhancement module comprises: A first feature enhancement unit is used to obtain data to be enhanced; randomly extract N pieces of data with replacement from the data to be enhanced to obtain data to be mixed; and interpolate the data to be mixed according to random weights to obtain incremental data; The second feature enhancement unit is used to obtain data to be enhanced; randomly extract 3 data with replacement from the data to be enhanced to obtain data to be mixed; randomly select one data from the data to be mixed as a reference and the other two data as variables, and obtain incremental data according to the formula z=x1+x2-x0, where x0 is the reference and x1 and x2 are variables; The third feature enhancement unit is used to obtain data to be enhanced; obtain statistical features of the data to be enhanced; obtain noise distribution according to the statistical features; randomly select noise according to the noise distribution; and obtain incremental data by accumulating the data to be enhanced and the noise.

9. A computer-readable storage medium for storing a program, characterized in that: When the program is executed, the model startup method according to any one of claims 1 to 6 is implemented.

Images