Training device, method, equipment and computer-readable medium

Abstract

To provide a training device, a training method, a training apparatus, and a computer readable medium that can acquire a desired target model with a small number of training data sets, reduce complexity and cost of processing, and shorten a development cycle.SOLUTION: A training device 1000 includes: an initialization module 1001 that initializes an initial model based on a basic category sample to acquire an intermediate model including a first parameter set related to a basic category; and a training module 1002 that trains the intermediate model by using a training data set to acquire a target model, and while training the intermediate model, the training device makes the first parameter set constant. The target model acquired is a neural network model for detecting cracks in a surface of an object from an image.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present application relates to the field of training models, and in particular to a training device, a training method, a training apparatus and a computer-readable medium. [Background technology]

[0002] Deep neural networks are tools for large-scale, multi-parameter optimization. Through large amounts of training data, deep neural networks can learn to identify hidden features that are difficult to summarize in the data, and accomplish many complex tasks, such as face detection, image semantic segmentation, object detection, motion tracking, and natural language translation. Summary of the Invention [Problem to be solved by the invention]

[0003] Currently, to obtain a neural network model for target detection, the neural network must be trained using a large amount of pre-tagged training data. Therefore, the performance and effectiveness of target detection based on deep neural networks are highly dependent on the training data provided. For example, for complex tasks such as detecting cracks on the surface of an object, it is common for staff to artificially tag cracks in images to obtain training data. However, training data obtained by artificially tagging images may contain theoretical cognitive and visual errors, which can lead to inaccurate training data. In addition, handling large amounts of training data requires a long time and places a heavy mental and visual burden on staff, making it prone to incorrect tagging and unsuitable for industrial production.

[0004] In view of the above problems, the present disclosure provides a training device, a training method, a training apparatus, and a computer-readable medium. [Means for solving the problem]

[0005] According to one aspect of the present disclosure, an image processing device is provided, which includes: an initialization module that initializes an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with the base category; and a training module that trains the intermediate model using a training dataset to obtain a target model, wherein the first parameter set is constant during training of the intermediate model.

[0006] According to one example of the present disclosure, the initial model is obtained based on a plurality of sub-models related to the target model, and the distance from the initial model to the plurality of sub-models is the shortest, so that the initial model can quickly converge to the plurality of sub-models.

[0007] According to one example of the present disclosure, the initial model is obtained based on a plurality of sub-models related to the target model, and the sub-models have different weights in the initial model.

[0008] According to one example of the present disclosure, the initial model is a randomly generated model.

[0009] According to one example of the present disclosure, the basic category samples include: positive Includes sample only.

[0010] According to one example of the present disclosure, the initialization module obtains the first parameter set through a memory network.

[0011] According to one example of the present disclosure, the first parameter set is a parameter set whose variation during initialization exceeds a first threshold.

[0012] According to one example of the present disclosure, the base category samples are obtained based on existing information including one or more of images, text, video, and audio.

[0013] According to one aspect of the present disclosure, there is provided a training method, comprising: initializing an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with the base category; and training the intermediate model using a training dataset to obtain a target model, wherein the first parameter set is constant while training the intermediate model.

[0014] According to one aspect of the present disclosure, there is provided a training device that includes a processor and a memory having computer-readable program instructions stored therein, the training device performing a training method, when the computer-readable program instructions are run by the processor, including the steps of initializing an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with the base category, and training the intermediate model using a training dataset to obtain a target model, wherein the first parameter set is constant during training of the intermediate model.

[0015] According to one aspect of the present disclosure, there is provided a computer readable storage medium for storing computer readable instructions for causing a computer to perform the training method as described above by said program.

[0016] In the above-mentioned aspect of the present disclosure, an initial model is first initialized based on basic category samples to obtain an intermediate model, a first parameter set in the intermediate model is fixed, and then the intermediate model is trained using a training dataset to obtain a target model, thereby enabling a desired target model to be obtained with a small training dataset, reducing processing complexity and cost and shortening the development cycle. [Brief explanation of the drawings]

[0017] The above and other objects, features, and advantages of the present disclosure will become more apparent when the embodiments of the present disclosure are described in more detail with reference to the drawings. The drawings are provided for a better understanding of the embodiments of the present disclosure, constitute a part of the specification, and are intended to explain the present disclosure together with the embodiments of the present disclosure, and are not intended to constitute limitations on the present disclosure. In the drawings, the same reference numerals generally represent the same components or steps. [Figure 1] FIG. 1 is a schematic diagram illustrating an application scenario of a neural network model training method according to an embodiment of the present disclosure. [Figure 2] FIG. 1 is a functional block diagram illustrating a training device according to an embodiment of the present disclosure. [Figure 3] 1A to 1C are schematic diagrams illustrating examples of models according to an embodiment of the present disclosure. [Figure 4] FIG. 10 is a schematic diagram illustrating an example of an image converted from a knowledge definition of a crack according to an embodiment of the present disclosure. [Figure 5] FIG. 1 is a schematic diagram illustrating obtaining a target model used for crack detection according to an embodiment of the present disclosure. [Figure 6] 1 is a flowchart illustrating a training method according to an embodiment of the present disclosure. [Figure 7] FIG. 1 is a functional block diagram illustrating a training device according to an embodiment of the present disclosure. [Figure 8] FIG. 1 is a schematic diagram illustrating a computer-readable storage medium according to an embodiment of the present disclosure. [Figure 9] FIG. 1 is a diagram illustrating an example of a hardware structure of an electronic device according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE INVENTION

[0018] The technical means in the embodiments of the present disclosure will be clearly and completely described below with reference to the drawings in the embodiments of the present disclosure. It is clear that the described embodiments are only some embodiments of the present disclosure, and are not all embodiments. Based on the embodiments in the present disclosure, all other embodiments that a person skilled in the art can obtain without inventive work belong to the scope of the claims of the present disclosure.

[0019] Flowcharts are used herein to explain steps according to the methods of the present application. It should be understood that the preceding or following steps do not necessarily have to be performed in a precise order. Conversely, the order may be reversed, or various steps may be processed simultaneously. At the same time, other operations may be added to these processes, or one or more steps may be removed from these processes.

[0020] The present disclosure proposes a training device that initializes an initial model based on basic category samples to obtain an intermediate model, and then trains the intermediate model using a training dataset to obtain a target model, thereby obtaining a desired target model with a small training dataset, reducing processing complexity and cost and shortening the development cycle.

[0021] Referring to FIG. 1, an application scenario of the neural network model (target model) training method according to an embodiment of the present disclosure is schematically described.

[0022] As shown in Fig. 1, a neural network model 20 according to an embodiment of the present disclosure receives an input 10, performs feature extraction processing on the input 10, and generates an output 30 based on the extracted features. In an embodiment of the present disclosure, the input 10 may be an object to be processed, such as an image, a video, or a natural language text. The neural network model 20 performs processing on the input 10, such as image semantic segmentation, object detection, motion tracking, and natural language translation, to generate the output 30. The neural network model 20 can be incorporated into a terminal device or a server and can process the input.

[0023] The neural network model 20 may include various neural network models, such as, but not limited to, a convolutional neural network (CNN) (including GoogLeNet, AlexNet, VGG network, etc.), a convolutional neural network with area (R-CNN), a region proposal network (RPN), a recurrent neural network (RNN), a stack-based deep neural network (S-DNN), a deep belief network (DBN), a restricted Boltzmann machine (RBM), a fully convolutional network, a long short-term memory (LSTM) network, and a classification network. Note that a neural network model performing a task may include a sub-neural network, and the sub-neural network may include a heterogeneous neural network and may be realized by a heterogeneous neural network model.

[0024] In one embodiment of the present disclosure, the neural network model 20 is trained in a training device 1000 shown in FIG.

[0025] 2 is a functional block diagram of a training device 1000 according to an embodiment of the present disclosure. The training device 1000 according to an embodiment of the present disclosure includes an initialization module 1001 and a training module 1002. It will be understood by those skilled in the art that these unit modules can be implemented in various forms, such as solely hardware, solely software, or a combination thereof, and the present disclosure is not limited to any of them. For example, these units can be implemented by a central processing unit (CPU), a text processor (GPU), a tensor processing unit (TPU), a field programmable gate array (FPGA), or other form of processing unit having data processing capabilities and / or instruction execution capabilities, and corresponding computer instructions.

[0026] 3 is a schematic diagram of an example of each model according to an embodiment of the present disclosure. A training device 1000 according to an embodiment of the present disclosure will be described below with reference to FIGS.

[0027] For example, as shown in FIG. 3, the initialization module 1001 may initialize the initial model 24 based on the base category samples to obtain an intermediate model 25 including a first set of parameters associated with the base category.

[0028] For example, the basic category sample is obtained based on existing information (e.g., common knowledge information) including one or more of images, text, video, audio, and prior knowledge. Note that the basic category sample may be any appropriate existing information, and is not limited here.

[0029] For example, when the target model obtained by the training device 1000 is used for crack detection, the basic category sample may be a conventional knowledge definition of a crack, i.e., "a line on the surface of an object that splits along that line," or an image converted from that knowledge definition.

[0030] For example, the basic category sample is positive For example, when the basic category sample is an image converted from a conventional crack knowledge definition, the image may contain only cracks but no background information. Figure 4 is an example of an image converted from a crack knowledge definition according to an embodiment of the present disclosure.

[0031] The initialization module 1001 positive The initial model 24 is initialized with the basic category samples, which only contain samples, so that the initial model 24 can quickly converge to the intermediate model 25, and the initialization complexity can be reduced.

[0032] For example, the initial model 24 may be obtained based on optimized-based meta learning, learning with a small amount of image data (few shot learning), or the like.

[0033] For example, the initial model 24 may be obtained based on multiple sub-models related to the target model 26. By learning the multiple sub-models related to the target model 26, the distance from the initial model 24 to the multiple sub-models is minimized, and the initial model 24 can quickly converge to the multiple sub-models.

[0034] For example, when the target model 26 is used for crack detection, multiple sub-models related to the crack detection model can be obtained, and the initial model 24 can be obtained from the multiple sub-models using conventional techniques, so that the initial model 24 can quickly converge to the multiple sub-models.

[0035] Alternatively, the initial model 24 may be derived based on multiple sub-models associated with the target model 26, with the multiple sub-models having different weights in the initial model 24.

[0036] Alternatively, the initial model 24 may be a randomly generated model.

[0037] For example, the initialization module 1001 may extract features from the initial model 24 using a memory network to obtain the first parameter set. For example, the first parameter set is a parameter set whose change amount during initialization exceeds a first threshold. For example, the first threshold can be obtained by statistical analysis, and is not limited thereto.

[0038] For example, the basic category sample includes positiveBased on the sample, an initial model 24 is initialized in a memory network to obtain a response matrix of the initialized initial model (i.e., intermediate model 25) for the basic category sample, and a set of elements in the response matrix whose transformation amount exceeds a first threshold during initialization can be set as a first parameter set.

[0039] For example, the initialization module 1001 can initialize the initial model 24 once based on the base category samples to obtain the intermediate model 25, so as to reduce processing complexity.

[0040] Next, for example, the training module 1002 can train the intermediate model 25 with a training dataset (e.g., input 27 in FIG. 3) to obtain a target model 26, wherein the first parameter set is kept constant based on a memory network while training the intermediate model.

[0041] For example, as described above, the initialization module 1001 may: positive An intermediate model including a first parameter set can be obtained by initializing an initial model with a base category sample including only samples. Then, the training module 1002 can refine or fine-tune the training for the target model by keeping the first parameter set constant during the training phase to achieve an initial positioning for the target object's features, and then training the intermediate model with a training data set. The training data set can include: positive It may also be a data set that contains both positive and negative samples.

[0042] Note that, as shown in FIG. 3, during the training phase, the intermediate model may be trained with a loss function 29 that includes a second set of parameters that is different from the first set of parameters to obtain a target model 26.

[0043] As an example, the response matrix obtained by initialization based on the base category samples using the initialization module 1001 is assumed to be [a1, a2, a3, a4, a5, a6, a7] = [0, 0.1, 0.3, 0.6, 0.7, 0.9, 0.4], and each element in the response matrix represents the change in each position of the intermediate model 25 relative to the base category sample. The loss values ​​at each position obtained by training the intermediate model 25 with the training dataset are assumed to be [a1, a2, a3, a4, a5, a6, a7] = [0.3, 0.1, 0.7, 0.5, 0.2, 0.6, 0.1]. The criteria for the target model 26 are that the set of elements whose change in the response matrix during initialization exceeds 0.5 (i.e., the first threshold) is set as the first parameter set, and the final loss value of the loss function during training is assumed not to exceed 0.2. Then, while training the intermediate model 24, the elements in the response matrix whose change amount at each position exceeds 0.5 (e.g., a4, a5, a6) are kept constant during training (i.e., their final loss value exceeds 0.2), while the intermediate model is trained with the loss function at other positions (e.g., a1, a2, a3, a7) to obtain the target model 26.

[0044] 5 is a schematic diagram illustrating the acquisition of a target model used for crack detection according to an embodiment of the present disclosure. For example, if the target model is a crack detection model, an initialization module 1001 may initialize an initial model with an image 51 converted from a crack knowledge definition to acquire an intermediate model including a first parameter set (as shown in FIG. 5, positions 52 corresponding to the first parameter set). Then, a training module 1002 may train the intermediate model with a training dataset 54 including, for example, artificially tagged cracks and background images to acquire a target model for detecting cracks 55. During training, the intermediate model is trained with a loss function including a second dataset while keeping the first parameter set (as shown in FIG. 5, positions 52 corresponding to the first parameter set) constant, thereby acquiring a target model used for crack detection.

[0045] It should be noted that the training device 1000 according to the present disclosure may be applied to any suitable field, including but not limited to target detection, iconographic semantic segmentation, natural language translation, and the like, without any limitation herein.

[0046] It can be seen that the training device 1000 according to the present disclosure can perform training with a relatively small training data set in the training stage, rather than directly using a large amount of training data set to train an initial model and obtain a target model, thereby reducing the processing complexity.

[0047] The training device 1000 according to an embodiment of the present disclosure has been described above with reference to Figures 1 to 5. A training method 100 according to an embodiment of the present disclosure will now be described.

[0048] 6 is a flowchart of a training method according to an embodiment of the present disclosure. The method can be automatically completed by a computer or the like. For example, the method can be applied to face detection, iconographic semantic segmentation, object detection, motion tracking, natural language translation, etc. For example, the training method can be implemented in software, hardware, firmware, or any combination thereof, and can be downloaded to and executed by a processor in a device such as a mobile phone, tablet, laptop, desktop computer, or network server.

[0049] As shown in FIG. 6, the training method includes the following steps S101 to S102.

[0050] In step S101, an initial model is initialized based on the base category samples to obtain an intermediate model including a first parameter set associated with the base category.

[0051] In step S102, the intermediate model is trained using a training data set to obtain a target model, and the first parameter set is constant during training of the intermediate model.

[0052] For example, in step S101, the basic category sample is obtained based on existing information (e.g., common knowledge), and the existing information includes one or more of images, text, video, audio, and prior knowledge. Note that the basic category sample may be any appropriate existing information, and is not limited here.

[0053] For example, when the target model is used for crack detection, the base category sample may be a conventional knowledge definition of a crack, i.e., "a line on the surface of an object that splits along that line," or an image converted from that knowledge definition.

[0054] For example, the basic category sample is positive For example, when the basic category sample is an image converted from a conventional crack knowledge definition, the image may contain only cracks but no background information.

[0055] In addition, positive The initial model is initialized with basic category samples containing only samples, so that the initial model can quickly converge to the intermediate model and reduce the initialization complexity.

[0056] For example, the initial model may be obtained based on optimized-based meta learning, learning with a small amount of image data (few shot learning), or the like.

[0057] For example, an initial model may be obtained based on a plurality of sub-models associated with the target model, and by learning the plurality of sub-models associated with the target model, the distance between the initial model and the plurality of sub-models is minimized, allowing the initial model to converge to the plurality of sub-models quickly.

[0058] For example, when a target model is used for crack detection, multiple sub-models related to the crack detection model can be obtained, and an initial model can be obtained from the multiple sub-models using conventional techniques, so that the initial model can quickly converge to the multiple sub-models.

[0059] Alternatively, the initial model may be derived based on multiple sub-models associated with a target model, with the multiple sub-models having different weights in the initial model.

[0060] Alternatively, the initial model may be a randomly generated model.

[0061] For example, the first parameter set can be obtained by extracting features from an initial model using a memory network. For example, the first parameter set is a parameter set whose change amount during initialization exceeds a first threshold. For example, the first threshold can be obtained by statistical analysis, and is not limited thereto.

[0062] For example, the basic category sample includes positive Based on the sample, an initial model can be initialized in a memory network to obtain a response matrix of the initialized initial model (i.e., intermediate model) for the basic category sample, and a set of elements in the response matrix whose transformation amount exceeds a first threshold during initialization can be set as a first parameter set.

[0063] For example, to reduce processing complexity, the initial model can be initialized once based on the base category samples to obtain the intermediate model.

[0064] Regarding step S102, for example, the intermediate model can be trained using a training dataset to obtain a target model, wherein the first parameter set is kept constant based on a memory network during training of the intermediate model.

[0065] For example, as above, positive An initial model can be initialized with a base category sample containing only samples to obtain an intermediate model with a first parameter set. Then, in the training phase, the first parameter set can be kept constant to achieve an initial positioning relative to the target object's features, and then the training relative to the target model can be refined or fine-tuned by continuing to train the intermediate model with a training data set, which includes: positive It may also be a data set that contains both positive and negative samples.

[0066] It should be noted that in the training stage, the intermediate model can be trained with a loss function that includes a second set of parameters different from the first set of parameters to obtain the target model.

[0067] As an example, assume that the response matrix obtained by initialization based on the base category samples is [a1, a2, a3, a4, a5, a6, a7] = [0, 0.1, 0.3, 0.6, 0.7, 0.9, 0.4], and each element in the response matrix represents the change in each position of the intermediate model relative to the base category samples. Assume that the loss values ​​at each position obtained by training the intermediate model on the training dataset are [a1, a2, a3, a4, a5, a6, a7] = [0.3, 0.1, 0.7, 0.5, 0.2, 0.6, 0.1]. The criteria for the target model are that the set of elements in the response matrix whose change during initialization exceeds 0.5 (i.e., the first threshold) is set as the first parameter set, and that the final loss value of the loss function during training does not exceed 0.2. Then, while training the intermediate model, the elements in the response matrix whose change amount at each position exceeds 0.5 (e.g., a4, a5, a6) are kept constant during training (i.e., their loss values ​​exceed 0.2), while the intermediate model is trained with the loss function at other positions (e.g., a1, a2, a3, a7) to obtain the target model.

[0068] According to the method of the present disclosure, first, an initial model is initialized based on basic category samples to obtain an intermediate model, a first parameter set in the intermediate model is fixed, and then the intermediate model is trained with a training dataset to obtain a target model. This shows that a desired target model can be obtained with a small training dataset, thereby reducing processing complexity and cost and shortening the development cycle.

[0069] Hereinafter, a training device 1100 according to an embodiment of the present disclosure will be described with reference to Fig. 7. Fig. 7 is a schematic diagram of a training device according to an embodiment of the present disclosure. The function of the training device according to this embodiment is similar to the details of the method described with reference to Fig. 6 above, and therefore, for the sake of brevity, a detailed description of the similar content will be omitted here.

[0070] The training device according to the present disclosure includes a processor 1102 and a memory 1101 having computer-readable instructions stored therein. When the computer-readable instructions are executed by the processor, a training method is performed. The training method includes initializing an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with a base category, and training the intermediate model using a training dataset to obtain a target model, wherein the first parameter set is constant during training of the intermediate model.

[0071] For the technical effects of the training device 1000 and the training equipment 1100 in different embodiments, please refer to the technical effects of the training method provided by the embodiments of the present disclosure, and detailed descriptions will be omitted here.

[0072] Training device 1000 and training device 1100 can be used with a variety of suitable electronic devices.

[0073] FIG. 8 is a schematic diagram of a computer-readable storage medium 1200 according to an embodiment of the present disclosure.

[0074] As shown in FIG. 8 , the present disclosure further includes a computer-readable storage medium 1200 for storing computer-readable instructions 1201, which, when executed by a computer, cause the computer to perform a training method including: initializing an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with the base category; and training the intermediate model using a training dataset to obtain a target model, wherein the first parameter set is constant while training the intermediate model.

[0075] <Hardware structure> The block diagrams used to explain the above embodiments show functional blocks. These functional blocks (components) are realized by any combination of hardware and / or software. Furthermore, the means for realizing each functional block is not particularly limited. That is, each functional block may be realized by a single device that is physically and / or logically coupled, or may be realized by two or more physically and / or logically separated devices that are directly and / or indirectly (e.g., wired and / or wirelessly) connected to the above multiple devices.

[0076] For example, an electronic device according to an embodiment of the present disclosure can function as a computer that executes the processing of the training method according to the present disclosure. Fig. 9 is a diagram showing an example of the hardware structure of an electronic device 60 according to an embodiment of the present disclosure. The electronic device 60 can be configured as a computer device that physically includes a processor 1010, a memory 1020, a storage 1030, a communication device 1040, an input device 1050, an output device 1060, a bus 1070, and the like.

[0077] In the following description, the word "device" may be replaced with "circuit," "instrument," "unit," etc. The hardware structure of electronic device 60 may include each device shown in one or more of the figures, or may not include some of the devices.

[0078] For example, although only one processor 1010 is shown, multiple processors may be used. Furthermore, processing may be performed by a single processor, or by one or more processors simultaneously, sequentially, or otherwise. Furthermore, processor 1010 may be implemented by one or more chips.

[0079] Each function in electronic device 60 is realized, for example, by loading predetermined software (programs) into hardware such as processor 1010 and memory 1020, causing processor 1010 to perform calculations, controlling communication via communication device 1040, and controlling the reading and / or writing of data in memory 1020 and storage 1030.

[0080] The processor 1010 controls the entire computer by running, for example, an operating system, and may be configured as a central processing unit (CPU) including an interface with peripheral devices, a control device, an arithmetic unit, registers, etc.

[0081] The processor 1010 also reads programs (program codes), software modules, data, etc. from the storage 1030 and / or the communication device 1040 into the memory 1020 and executes various processes based on the programs. The programs may be programs that cause a computer to execute at least some of the operations described in the above-described embodiments. For example, the control unit 401 of the electronic device 60 may be implemented by a control program stored in the memory 1020 and executed by the processor 1010, and similar implementations may be implemented for the other functional blocks.

[0082] The memory 1020 is a computer-readable recording medium and may be composed of at least one of, for example, read-only memory (ROM), programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), random access memory (RAM), or other suitable storage medium. The memory 1020 may also be referred to as a register, cache, main memory, etc. The memory 1020 may store executable programs (program codes), software modules, etc. for implementing a wireless communication method according to an embodiment of the present disclosure.

[0083] Storage 1030 is a computer-readable recording medium, and may be, for example, at least one of a flexible disk, a floppy disk, a magneto-optical disk (e.g., a read-only disk (e.g., a Compact Disc ROM (CD-ROM)), a digital versatile disk, a Blu-ray (registered trademark) disk), a removable disk, a hard disk drive, a smart card, flash (e.g., a card, a stick, a key drive), a magnetic stripe, a database, a server, or other suitable storage medium. Storage 1030 may also be referred to as an auxiliary storage device.

[0084] The communication device 1040 is hardware (transmission and reception equipment) for performing communication between computers via a wired and / or wireless network, and is also called, for example, a network device, a network controller, a network card, or a communication module.

[0085] The input device 1050 is an input device (for example, a keyboard, a mouse, a microphone, a switch, a button, a sensor, etc.) that receives input from the outside. The output device 1060 is an output device (for example, a display, a speaker, a light emitting diode (LED) lamp, etc.) that outputs to the outside. The input device 1050 and the output device 1060 may be an integrated structure (for example, a touch panel).

[0086] Furthermore, each device, such as the processor 1010 and the memory 1020, is connected via a bus 1070 for communicating information. The bus 1070 may be configured as a single bus or may be configured as different buses between the devices.

[0087] The electronic device 60 may also include hardware such as a microprocessor, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a programmable logic device (PLD), or a field programmable gate array (FPGA), and some or all of the functional blocks may be realized by the hardware. For example, the processor 1010 may be implemented by at least one of these pieces of hardware.

[0088] Software, whether referred to as software, firmware, middleware, microcode, hardware description language, or otherwise, shall be construed broadly to refer to instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, threads of execution, steps, functions, and the like.

[0089] Additionally, software, instructions, information, etc. may be transmitted or received over a transmission medium. For example, when transmitting the software from a website, server, or other remote resource using wired technologies (such as coaxial cable, fiber optic cable, twisted pair, Digital Subscriber Line (DSL)), and / or wireless technologies (such as infrared, microwave), these wired and / or wireless technologies are included within the definition of transmission media.

[0090] Each aspect / embodiment described herein may be used alone or in combination, or may be switched during execution. Furthermore, the order of processing steps, sequences, flowcharts, etc., of each aspect / embodiment described herein may be changed if there is no contradiction. For example, with respect to the methods described herein, various step units are shown in an exemplary order, and are not limited to the specific order shown.

[0091] As used herein, a phrase such as "based on" does not mean "based only on," unless expressly stated otherwise. In other words, a phrase such as "based on" refers to both "based only on" and "based at least on."

[0092] Any reference herein to units using names such as "first," "second," etc., is not intended to exhaustively limit the number or order of those units. These names may be used herein as a convenient way to distinguish between two or more units. Thus, a reference to a first unit and a second unit does not imply that only two units may be employed or that the first unit must precede the second unit in some way.

[0093] When used in this specification or claims, the terms "including," "comprising," and variations thereof are open in the same manner as the term "comprising." Furthermore, the term "or" as used in this specification or claims does not mean an exclusive or.

[0094] Those skilled in the art will appreciate that aspects of the present application may be illustrated and described in terms of any patentable class or context, including any new and useful process, machine, manufacture, or combination of matter, or any new and useful improvement thereof. Accordingly, aspects of the present application may be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or by a combination of hardware and software. Such hardware or software may be referred to as a "data block," "module," "engine," "unit," "component," or "system." Additionally, aspects of the present application may be represented as a computer product residing on one or more computer-readable mediums containing computer-readable program code.

[0095] This application uses specific language to describe embodiments of the application. For example, "one embodiment," "one embodiment," and / or "some embodiments" refer to certain components, structures, or features associated with at least one embodiment of the application. Therefore, it should be emphasized and noted that "one embodiment," "one embodiment," or "one alternative embodiment" referenced two or more times in different places in this specification do not necessarily refer to the same embodiment. Also, certain components, structures, or features of one or more embodiments of the application may be combined as appropriate.

[0096] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs. Furthermore, it should be understood that terms, for example, as defined in conventional dictionaries, unless expressly defined herein, should be interpreted to have a meaning consistent with their meaning in the context of the relevant art, without the application of idealized or overly formalized meanings.

[0097] Although the present disclosure has been described in detail above, it is clear to those skilled in the art that the present disclosure is not limited to the embodiments described in the present specification. The present disclosure can be implemented in amended and modified forms without departing from the spirit and scope of the present disclosure defined by the claims. Therefore, the description in this specification is intended to be illustrative and explanatory and does not have any limiting meaning on the present disclosure.

Claims

1. an initialization module for initializing an initial model based on the base category samples to obtain an intermediate model including a first parameter set associated with the base category; a training module for training the intermediate model using a training dataset to obtain a target model; the first parameter set is constant during training of the intermediate model; The first parameter set is a parameter set whose change amount exceeds a first threshold during initialization. Training equipment.

2. the initial model is obtained based on a plurality of sub-models associated with the target model; when acquiring the initial model, the training device acquires the initial model by learning the plurality of sub-models so that the distance from the initial model to the plurality of sub-models is shortest, and the initial model quickly converges to the plurality of sub-models.

10. The training device of claim 1.

3. The initial model is obtained based on a plurality of sub-models related to the target model, and weights of the plurality of sub-models in the initial model are different.

10. The training device of claim 1.

4. The base category samples include only positive samples. The training device according to any one of claims 1 to 3.

5. the initialization module obtains the first parameter set via a memory network; The training device according to any one of claims 1 to 3.

6. The basic category samples are obtained based on existing information including one or more of images, text, video, and audio. The training device according to any one of claims 1 to 3.

7. A training method for use with a training device, comprising: the training device initializing an initial model based on base category samples to obtain an intermediate model including a first parameter set associated with the base category; the training device training the intermediate model using a training dataset to obtain a target model; the first parameter set is constant while the training device trains the intermediate model; The first parameter set is a parameter set whose change amount exceeds a first threshold during initialization. Training methods.

8. a processor; a memory having computer readable program instructions stored thereon; The computer-readable program instructions, when executed by the processor, perform a training method; The training method is initializing an initial model based on the base category samples to obtain an intermediate model including a first parameter set associated with the base category; training the intermediate model using a training dataset to obtain a target model; the first parameter set is constant during training of the intermediate model; The first parameter set is a parameter set whose change amount exceeds a first threshold during initialization. Training equipment.

9. A computer-readable storage medium for storing computer-readable instructions, comprising: A computer-readable storage medium that causes a computer to execute the training method of claim 7 in accordance with the computer-readable instructions.

Images